fix(sandbox): make missing sandbox.mounts host_path a loud ERROR (#3244) (#3250)

In Docker production deployments, LocalSandboxProvider runs inside the
deer-flow-gateway container, so any `sandbox.mounts[].host_path` from
config.yaml is resolved against the gateway container's filesystem — not
the host machine. When the path isn't also bind-mounted into the gateway
service, the mount was silently dropped with only a WARNING log, leaving
agents reading an empty directory in production while the same config
worked under `make dev`.

Escalate the missing-host_path branch to logger.error with explicit
guidance about Docker bind mounts and docker-compose, so the failure is
hard to miss in default log configurations. Skip behaviour is preserved
to avoid breaking existing deployments.

Also clarify the misleading `VolumeMountConfig.host_path` field
description so it documents reality for both providers:

  - LocalSandboxProvider checks host_path from inside the gateway process
    (host in `make dev`, container in `make up`).
  - AioSandboxProvider (DooD) passes host_path straight to `docker -v`
    for the sandbox container, where the host Docker daemon resolves it
    from the host machine's perspective.

config.example.yaml's `sandbox.mounts` comment gets a Note: block
pointing operators at the docker-compose bind-mount requirement so the
Docker-mode gotcha is discoverable from the canonical template.

Adds a regression test that:
  - confirms missing host_path is still skipped (no behaviour break);
  - asserts an ERROR record is emitted referencing the offending paths;
  - asserts the message contains actionable Docker/gateway/docker-compose
    keywords so future refactors can't quietly downgrade it.

Refs: https://github.com/bytedance/deer-flow/issues/3244

.agent/skills/smoke-test/SKILL.md

@@ -59,7 +59,7 @@ smoke-test/
 2. **Check pnpm** - Package manager
 3. **Check uv** - Python package manager
 4. **Check nginx** - Reverse proxy
-5. **Check required ports** - Confirm that ports 2026, 3000, 8001, and 2024 are not occupied
+5. **Check required ports** - Confirm that ports 2026, 3000, and 8001 are not occupied
 
 **Docker mode environment check** (if Docker is selected):
 1. **Check whether Docker is installed** - Run `docker --version`
@@ -93,17 +93,17 @@ smoke-test/
 ### Phase 5: Service Health Check
 
 **Local mode health check**:
-1. **Check process status** - Confirm that LangGraph, Gateway, Frontend, and Nginx processes are all running
+1. **Check process status** - Confirm that Gateway, Frontend, and Nginx processes are all running
 2. **Check frontend service** - Visit `http://localhost:2026` and verify that the page loads
 3. **Check API Gateway** - Verify the `http://localhost:2026/health` endpoint
-4. **Check LangGraph service** - Verify the availability of relevant endpoints
+4. **Check LangGraph-compatible API** - Verify the `/api/langgraph/*` route exposed by Gateway
 5. **Frontend route smoke check** - Run `bash .agent/skills/smoke-test/scripts/frontend_check.sh` to verify key routes under `/workspace`
 
 **Docker mode health check** (when using Docker):
 1. **Check container status** - Run `docker ps` and confirm that all containers are running
 2. **Check frontend service** - Visit `http://localhost:2026` and verify that the page loads
 3. **Check API Gateway** - Verify the `http://localhost:2026/health` endpoint
-4. **Check LangGraph service** - Verify the availability of relevant endpoints
+4. **Check LangGraph-compatible API** - Verify the `/api/langgraph/*` route exposed by Gateway
 5. **Frontend route smoke check** - Run `bash .agent/skills/smoke-test/scripts/frontend_check.sh` to verify key routes under `/workspace`
 
 ### Optional Functional Verification
@@ -135,7 +135,7 @@ smoke-test/
 
 The following warnings can appear during smoke testing and do not block a successful result:
 - Feishu/Lark SSL errors in Gateway logs (certificate verification failure) can be ignored if that channel is not enabled
-- Warnings in LangGraph logs about missing methods in the custom checkpointer, such as `adelete_for_runs` or `aprune`, do not affect the core functionality
+- Warnings in Gateway logs about missing methods in the custom checkpointer, such as `adelete_for_runs` or `aprune`, do not affect the core functionality
 
 ## Key Tools
 

.agent/skills/smoke-test/references/SOP.md

@@ -138,7 +138,6 @@ This document describes the detailed operating steps for each phase of the DeerF
    lsof -i :2026  # Main port
    lsof -i :3000  # Frontend
    lsof -i :8001  # Gateway
-   lsof -i :2024  # LangGraph
    ```
 
 **Success Criteria**: All ports are free, or they are occupied only by DeerFlow-related processes.
@@ -258,7 +257,7 @@ This document describes the detailed operating steps for each phase of the DeerF
 **Steps**:
 1. Run `make dev-daemon` (background mode)
 
-**Description**: This command starts all services (LangGraph, Gateway, Frontend, Nginx).
+**Description**: This command starts all services (Gateway embedded runtime, Frontend, Nginx).
 
 **Notes**:
 - `make dev` runs in the foreground and stops with Ctrl+C
@@ -272,7 +271,6 @@ This document describes the detailed operating steps for each phase of the DeerF
 **Steps**:
 1. Wait 90-120 seconds for all services to start completely
 2. You can monitor startup progress by checking these log files:
-   - `logs/langgraph.log`
    - `logs/gateway.log`
    - `logs/frontend.log`
    - `logs/nginx.log`
@@ -316,11 +314,10 @@ This document describes the detailed operating steps for each phase of the DeerF
 **Steps**:
 1. Run the following command to check processes:
    ```bash
-   ps aux | grep -E "(langgraph|uvicorn|next|nginx)" | grep -v grep
+   ps aux | grep -E "(uvicorn|next|nginx)" | grep -v grep
    ```
 
 **Success Criteria**: Confirm that the following processes are running:
-- LangGraph (`langgraph dev`)
 - Gateway (`uvicorn app.gateway.app:app`)
 - Frontend (`next dev` or `next start`)
 - Nginx (`nginx`)
@@ -356,10 +353,11 @@ curl http://localhost:2026/health
 
 ---
 
-#### 5.1.4 Check LangGraph Service
+#### 5.1.4 Check LangGraph-compatible API
 
 **Steps**:
-1. Visit relevant LangGraph endpoints to verify availability
+1. Visit `http://localhost:2026/api/langgraph/assistants/lead_agent` to verify Gateway's LangGraph-compatible API route is reachable.
+2. A `401` response is acceptable when authentication is enabled and no session cookie is provided.
 
 ---
 
@@ -373,7 +371,6 @@ curl http://localhost:2026/health
    - `deer-flow-nginx`
    - `deer-flow-frontend`
    - `deer-flow-gateway`
-   - `deer-flow-langgraph` (if not in gateway mode)
 
 ---
 
@@ -406,10 +403,11 @@ curl http://localhost:2026/health
 
 ---
 
-#### 5.2.4 Check LangGraph Service
+#### 5.2.4 Check LangGraph-compatible API
 
 **Steps**:
-1. Visit relevant LangGraph endpoints to verify availability
+1. Visit `http://localhost:2026/api/langgraph/assistants/lead_agent` to verify Gateway's LangGraph-compatible API route is reachable.
+2. A `401` response is acceptable when authentication is enabled and no session cookie is provided.
 
 ---
 

.agent/skills/smoke-test/references/troubleshooting.md

@@ -254,7 +254,6 @@ Processes exit quickly after running `make dev-daemon`.
 **Solutions**:
 1. Check log files:
    ```bash
-   tail -f logs/langgraph.log
    tail -f logs/gateway.log
    tail -f logs/frontend.log
    tail -f logs/nginx.log
@@ -367,24 +366,7 @@ Errors appear in `gateway.log`.
    uv sync
    ```
 
-4. Confirm that the LangGraph service is running normally (if not in gateway mode)
-
----
-
-### Issue: LangGraph Fails to Start
-
-**Symptoms**:
-Errors appear in `langgraph.log`.
-
-**Solutions**:
-1. Check LangGraph logs:
-   ```bash
-   tail -f logs/langgraph.log
-   ```
-
-2. Check config.yaml
-3. Check whether Python dependencies are complete
-4. Confirm that port 2024 is not occupied
+4. Confirm that the Gateway process is running normally.
 
 ---
 
@@ -519,7 +501,7 @@ Accessing `/health` returns an error or times out.
 
 2. Confirm that config.yaml exists and has valid formatting
 3. Check whether Python dependencies are complete
-4. Confirm that the LangGraph service is running normally
+4. Confirm that the Gateway process is running normally.
 
 **Solutions** (Docker mode):
 1. Check gateway container logs:
@@ -529,7 +511,7 @@ Accessing `/health` returns an error or times out.
 
 2. Confirm that config.yaml is mounted correctly
 3. Check whether Python dependencies are complete
-4. Confirm that the LangGraph service is running normally
+4. Confirm that the Gateway process is running normally.
 
 ---
 
@@ -539,7 +521,7 @@ Accessing `/health` returns an error or times out.
 
 #### View All Service Processes
 ```bash
-ps aux | grep -E "(langgraph|uvicorn|next|nginx)" | grep -v grep
+ps aux | grep -E "(uvicorn|next|nginx)" | grep -v grep
 ```
 
 #### View Service Logs
@@ -548,7 +530,6 @@ ps aux | grep -E "(langgraph|uvicorn|next|nginx)" | grep -v grep
 tail -f logs/*.log
 
 # View specific service logs
-tail -f logs/langgraph.log
 tail -f logs/gateway.log
 tail -f logs/frontend.log
 tail -f logs/nginx.log

.agent/skills/smoke-test/scripts/check_local_env.sh

@@ -65,7 +65,7 @@ if ! command -v lsof >/dev/null 2>&1; then
     echo "  Install lsof and rerun this check"
     all_passed=false
 else
-    for port in 2026 3000 8001 2024; do
+    for port in 2026 3000 8001; do
         if lsof -i :$port >/dev/null 2>&1; then
             echo "⚠  Port $port is already in use:"
             lsof -i :$port | head -2

.agent/skills/smoke-test/scripts/deploy_local.sh

@@ -54,7 +54,6 @@ echo "=========================================="
 echo ""
 echo "🌐 Access URL: http://localhost:2026"
 echo "📋 View logs:"
-echo "   - logs/langgraph.log"
 echo "   - logs/gateway.log"
 echo "   - logs/frontend.log"
 echo "   - logs/nginx.log"

.agent/skills/smoke-test/scripts/health_check.sh

@@ -76,12 +76,11 @@ if [ "$mode" = "docker" ]; then
         all_passed=false
     fi
 else
-    summary_hint="logs/{langgraph,gateway,frontend,nginx}.log"
+    summary_hint="logs/{gateway,frontend,nginx}.log"
     print_step "1. Checking local service ports..."
     check_listen_port "Nginx" 2026
     check_listen_port "Frontend" 3000
     check_listen_port "Gateway" 8001
-    check_listen_port "LangGraph" 2024
 fi
 echo ""
 
@@ -104,8 +103,8 @@ else
 fi
 echo ""
 
-echo "5. Checking LangGraph service..."
-check_http_status "LangGraph service" "http://localhost:2024/" "200|301|302|307|308|404"
+echo "5. Checking LangGraph-compatible Gateway API..."
+check_http_status "LangGraph-compatible Gateway API" "http://localhost:2026/api/langgraph/assistants/lead_agent" "200|401"
 echo ""
 
 echo "=========================================="

.agent/skills/smoke-test/templates/report.docker.template.md

@@ -78,7 +78,7 @@
 - [x] Container status - {{status_containers}}
 - [x] Frontend service - {{status_frontend}}
 - [x] API Gateway - {{status_api_gateway}}
-- [x] LangGraph service - {{status_langgraph}}
+- [x] LangGraph-compatible Gateway API - {{status_langgraph}}
 
 **Phase Status**: {{stage5_status}}
 
@@ -147,7 +147,6 @@ Commit Message: {{git_commit_message}}
 | deer-flow-nginx | {{nginx_status}} | {{nginx_uptime}} |
 | deer-flow-frontend | {{frontend_status}} | {{frontend_uptime}} |
 | deer-flow-gateway | {{gateway_status}} | {{gateway_uptime}} |
-| deer-flow-langgraph | {{langgraph_status}} | {{langgraph_uptime}} |
 
 ---
 

.agent/skills/smoke-test/templates/report.local.template.md

@@ -80,7 +80,7 @@
 - [x] Process status - {{status_processes}}
 - [x] Frontend service - {{status_frontend}}
 - [x] API Gateway - {{status_api_gateway}}
-- [x] LangGraph service - {{status_langgraph}}
+- [x] LangGraph-compatible Gateway API - {{status_langgraph}}
 
 **Phase Status**: {{stage5_status}}
 
@@ -152,7 +152,7 @@ Commit Message: {{git_commit_message}}
 | Nginx | {{nginx_status}} | {{nginx_endpoint}} |
 | Frontend | {{frontend_status}} | {{frontend_endpoint}} |
 | Gateway | {{gateway_status}} | {{gateway_endpoint}} |
-| LangGraph | {{langgraph_status}} | {{langgraph_endpoint}} |
+| Gateway LangGraph API | {{langgraph_status}} | {{langgraph_endpoint}} |
 
 ---
 
@@ -166,7 +166,7 @@ Commit Message: {{git_commit_message}}
 
 ### If the Test Fails
 1. [ ] Review references/troubleshooting.md for common solutions
-2. [ ] Check local logs: `logs/{langgraph,gateway,frontend,nginx}.log`
+2. [ ] Check local logs: `logs/{gateway,frontend,nginx}.log`
 3. [ ] Verify configuration file format and content
 4. [ ] If needed, fully reset the environment: `make stop && make clean && make install && make dev-daemon`
 

.env.example

@@ -21,6 +21,7 @@ INFOQUEST_API_KEY=your-infoquest-api-key
 # DEEPSEEK_API_KEY=your-deepseek-api-key
 # NOVITA_API_KEY=your-novita-api-key  # OpenAI-compatible, see https://novita.ai
 # MINIMAX_API_KEY=your-minimax-api-key  # OpenAI-compatible, see https://platform.minimax.io
+# STEPFUN_API_KEY=your-stepfun-api-key  # OpenAI-compatible, see https://platform.stepfun.com
 # VLLM_API_KEY=your-vllm-api-key  # OpenAI-compatible
 # FEISHU_APP_ID=your-feishu-app-id
 # FEISHU_APP_SECRET=your-feishu-app-secret
@@ -50,6 +51,11 @@ INFOQUEST_API_KEY=your-infoquest-api-key
 # Set to "false" to disable Swagger UI, ReDoc, and OpenAPI schema in production
 # GATEWAY_ENABLE_DOCS=false
 
+# Shared internal Gateway auth token for multi-worker deployments.
+# `make up` generates and persists this automatically; set it manually only
+# when you run Gateway workers outside the bundled deploy script.
+# DEER_FLOW_INTERNAL_AUTH_TOKEN=your-shared-internal-token
+
 # ── Frontend SSR → Gateway wiring ─────────────────────────────────────────────
 # The Next.js server uses these to reach the Gateway during SSR (auth checks,
 # /api/* rewrites). They default to localhost values that match `make dev` and

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,159 @@
+name: 🐛 Bug report
+description: Report something that isn't working so maintainers can reproduce and fix it.
+title: "[bug] "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to file a bug. A clear, reproducible report is the
+        single biggest factor in how fast it gets fixed.
+
+        Please fill in every required field — especially **reproduction steps** and **logs**.
+
+  - type: checkboxes
+    id: preflight
+    attributes:
+      label: Before you start
+      options:
+        - label: I searched [existing issues](https://github.com/bytedance/deer-flow/issues?q=is%3Aissue) and this is not a duplicate.
+          required: true
+        - label: I can reproduce this on the latest `main`.
+          required: false
+
+  - type: input
+    id: summary
+    attributes:
+      label: Problem summary
+      description: One sentence describing the bug.
+      placeholder: e.g. make dev fails to start the gateway service
+    validations:
+      required: true
+
+  - type: dropdown
+    id: areas
+    attributes:
+      label: Affected area(s)
+      description: Which part of DeerFlow does this touch? Select all that apply.
+      multiple: true
+      options:
+        - Frontend (UI / Next.js)
+        - Backend API (gateway / endpoints / SSE)
+        - Agents / LangGraph (graph, prompts, langgraph.json)
+        - Sandbox / Docker
+        - Skills
+        - MCP
+        - Config / setup (make, config.yaml, env)
+        - Docs
+        - Not sure
+    validations:
+      required: true
+
+  - type: textarea
+    id: actual
+    attributes:
+      label: What happened?
+      description: The actual behavior. Include the key error lines verbatim.
+      placeholder: When I do X, I expected Y but I got Z.
+    validations:
+      required: true
+
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected behavior
+      placeholder: What did you expect to happen instead?
+    validations:
+      required: true
+
+  - type: textarea
+    id: reproduce
+    attributes:
+      label: Steps to reproduce
+      description: Exact commands and sequence. Minimal steps that reliably reproduce the problem.
+      placeholder: |
+        1. make check
+        2. make install
+        3. make dev
+        4. ...
+    validations:
+      required: true
+
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant logs
+      description: Paste key lines from logs (for example `logs/gateway.log`, `logs/frontend.log`). Redact secrets.
+      render: shell
+    validations:
+      required: true
+
+  - type: dropdown
+    id: run_mode
+    attributes:
+      label: How are you running DeerFlow?
+      options:
+        - Local (make dev)
+        - Docker (make docker-start)
+        - CI
+        - Other
+    validations:
+      required: true
+
+  - type: dropdown
+    id: os
+    attributes:
+      label: Operating system
+      options:
+        - macOS
+        - Linux
+        - Windows
+        - Other
+    validations:
+      required: true
+
+  - type: input
+    id: platform_details
+    attributes:
+      label: Platform details
+      description: Architecture and shell, if relevant.
+      placeholder: e.g. arm64, zsh
+
+  - type: input
+    id: python_version
+    attributes:
+      label: Python version
+      placeholder: e.g. Python 3.12.9
+
+  - type: input
+    id: node_version
+    attributes:
+      label: Node.js version
+      placeholder: e.g. v22.11.0
+
+  - type: input
+    id: pnpm_version
+    attributes:
+      label: pnpm version
+      placeholder: e.g. 10.26.2
+
+  - type: input
+    id: uv_version
+    attributes:
+      label: uv version
+      placeholder: e.g. 0.7.20
+
+  - type: textarea
+    id: git_info
+    attributes:
+      label: Git state
+      description: Output of `git branch --show-current` and the latest commit SHA.
+      placeholder: |
+        branch: feature/my-branch
+        commit: abcdef1
+
+  - type: textarea
+    id: additional
+    attributes:
+      label: Additional context
+      description: Screenshots, related issues, config snippets (redacted), or anything else that helps triage.

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,11 @@
+blank_issues_enabled: false
+contact_links:
+  - name: 💬 Questions & usage help
+    url: https://github.com/bytedance/deer-flow/discussions/categories/q-a
+    about: "How do I use X? Why does Y behave like that? Ask in Discussions — it gets answered faster and stays searchable."
+  - name: 💡 Ideas & proposals
+    url: https://github.com/bytedance/deer-flow/discussions/categories/ideas
+    about: Have a half-formed idea? Float it in Discussions before opening a formal feature request.
+  - name: 🔒 Report a security vulnerability
+    url: https://github.com/bytedance/deer-flow/security/policy
+    about: Do not open a public issue for security problems. Follow the security policy instead.

.github/ISSUE_TEMPLATE/feature-request.yml

@@ -0,0 +1,67 @@
+name: 💡 Feature request
+description: Propose a new capability or an improvement to an existing one.
+title: "[feat] "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for the suggestion. For non-trivial features, please open a
+        [Discussion](https://github.com/bytedance/deer-flow/discussions/categories/ideas)
+        first to align on scope before writing code.
+
+  - type: checkboxes
+    id: preflight
+    attributes:
+      label: Before you start
+      options:
+        - label: I searched [existing issues](https://github.com/bytedance/deer-flow/issues?q=is%3Aissue) and this is not a duplicate.
+          required: true
+
+  - type: textarea
+    id: problem
+    attributes:
+      label: Problem / motivation
+      description: What problem does this solve? What is painful today, or what does it unblock?
+      placeholder: "I'm always frustrated when ..."
+    validations:
+      required: true
+
+  - type: textarea
+    id: solution
+    attributes:
+      label: Proposed solution
+      description: Describe the change from a user's / caller's perspective.
+    validations:
+      required: true
+
+  - type: dropdown
+    id: areas
+    attributes:
+      label: Affected area(s)
+      description: Which part of DeerFlow would this touch? Select all that apply.
+      multiple: true
+      options:
+        - Frontend (UI / Next.js)
+        - Backend API (gateway / endpoints / SSE)
+        - Agents / LangGraph (graph, prompts, langgraph.json)
+        - Sandbox / Docker
+        - Skills
+        - MCP
+        - Config / setup
+        - Docs
+        - Not sure
+    validations:
+      required: true
+
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives considered
+      description: Other approaches you weighed and why you discarded them.
+
+  - type: textarea
+    id: additional
+    attributes:
+      label: Additional context
+      description: Mockups, links, related issues, or anything else that helps.

.github/ISSUE_TEMPLATE/runtime-information.yml

@@ -1,128 +0,0 @@
-name: Runtime Information
-description: Report runtime/environment details to help reproduce an issue.
-title: "[runtime] "
-labels:
-  - needs-triage
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for sharing runtime details.
-        Complete this form so maintainers can quickly reproduce and diagnose the problem.
-
-  - type: input
-    id: summary
-    attributes:
-      label: Problem summary
-      description: Short summary of the issue.
-      placeholder: e.g. make dev fails to start gateway service
-    validations:
-      required: true
-
-  - type: textarea
-    id: expected
-    attributes:
-      label: Expected behavior
-      placeholder: What did you expect to happen?
-    validations:
-      required: true
-
-  - type: textarea
-    id: actual
-    attributes:
-      label: Actual behavior
-      placeholder: What happened instead? Include key error lines.
-    validations:
-      required: true
-
-  - type: dropdown
-    id: os
-    attributes:
-      label: Operating system
-      options:
-        - macOS
-        - Linux
-        - Windows
-        - Other
-    validations:
-      required: true
-
-  - type: input
-    id: platform_details
-    attributes:
-      label: Platform details
-      description: Add architecture and shell if relevant.
-      placeholder: e.g. arm64, zsh
-
-  - type: input
-    id: python_version
-    attributes:
-      label: Python version
-      placeholder: e.g. Python 3.12.9
-
-  - type: input
-    id: node_version
-    attributes:
-      label: Node.js version
-      placeholder: e.g. v23.11.0
-
-  - type: input
-    id: pnpm_version
-    attributes:
-      label: pnpm version
-      placeholder: e.g. 10.26.2
-
-  - type: input
-    id: uv_version
-    attributes:
-      label: uv version
-      placeholder: e.g. 0.7.20
-
-  - type: dropdown
-    id: run_mode
-    attributes:
-      label: How are you running DeerFlow?
-      options:
-        - Local (make dev)
-        - Docker (make docker-dev)
-        - CI
-        - Other
-    validations:
-      required: true
-
-  - type: textarea
-    id: reproduce
-    attributes:
-      label: Reproduction steps
-      description: Provide exact commands and sequence.
-      placeholder: |
-        1. make check
-        2. make install
-        3. make dev
-        4. ...
-    validations:
-      required: true
-
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant logs
-      description: Paste key lines from logs (for example logs/gateway.log, logs/frontend.log).
-      render: shell
-    validations:
-      required: true
-
-  - type: textarea
-    id: git_info
-    attributes:
-      label: Git state
-      description: Share output of git branch and latest commit SHA.
-      placeholder: |
-        branch: feature/my-branch
-        commit: abcdef1
-
-  - type: textarea
-    id: additional
-    attributes:
-      label: Additional context
-      description: Add anything else that might help triage.

.github/labels.yml

@@ -0,0 +1,119 @@
+# Declarative label source of truth for DeerFlow.
+#
+# This file is the single source of truth for repository labels used by the
+# auto-labeling workflows (.github/workflows/pr-labeler.yml, pr-triage.yml,
+# issue-triage.yml). Auto-labelers can only apply labels that already exist,
+# so every label referenced by a workflow MUST be declared here.
+#
+# Apply with:  uv run --with pyyaml python scripts/sync_labels.py [--repo OWNER/NAME]
+# CI keeps it in sync via .github/workflows/label-sync.yml (runs on changes here).
+#
+# Sync is additive/update-only: it creates or updates the labels listed below
+# and never deletes labels that are not listed.
+#
+# Color = 6-digit hex without the leading '#'.
+
+labels:
+  # ── Type ─────────────────────────────────────────────────────────────────
+  # Mostly GitHub defaults; declared here so colors/descriptions stay stable
+  # and so issue templates can rely on them existing.
+  - name: bug
+    color: d73a4a
+    description: Something isn't working
+  - name: enhancement
+    color: a2eeef
+    description: New feature or request
+  - name: documentation
+    color: 0075ca
+    description: Improvements or additions to documentation
+  - name: question
+    color: d876e3
+    description: Further information is requested
+
+  # ── Area (auto, by changed paths — see .github/labeler.yml) ───────────────
+  # Mirrors the "Surface area" section of the pull request template.
+  - name: "area:frontend"
+    color: c5def5
+    description: Next.js frontend under frontend/
+  - name: "area:backend"
+    color: c5def5
+    description: Gateway / runtime / core backend under backend/
+  - name: "area:agents"
+    color: c5def5
+    description: Agents, subagents, graph wiring, prompts, langgraph.json
+  - name: "area:sandbox"
+    color: c5def5
+    description: Sandboxed execution and docker/
+  - name: "area:skills"
+    color: c5def5
+    description: Skills under skills/ or the skills harness
+  - name: "area:mcp"
+    color: c5def5
+    description: Model Context Protocol integration
+  - name: "area:ci"
+    color: c5def5
+    description: GitHub Actions, CI config, repo tooling
+  - name: "area:docs"
+    color: c5def5
+    description: Documentation and Markdown only
+  - name: "area:deps"
+    color: c5def5
+    description: Dependency manifests / lockfiles
+
+  # ── Size (auto, by additions + deletions — see pr-triage.yml) ─────────────
+  - name: "size/XS"
+    color: "009900"
+    description: PR changes < 20 lines
+  - name: "size/S"
+    color: 77bb00
+    description: PR changes 20-100 lines
+  - name: "size/M"
+    color: eebb00
+    description: PR changes 100-300 lines
+  - name: "size/L"
+    color: ee9900
+    description: PR changes 300-700 lines
+  - name: "size/XL"
+    color: ee5500
+    description: PR changes 700+ lines
+
+  # ── Risk (auto, by changed paths — see pr-triage.yml) ─────────────────────
+  - name: "risk:low"
+    color: 0e8a16
+    description: "Low risk: docs / i18n / assets only"
+  - name: "risk:medium"
+    color: fbca04
+    description: "Medium risk: regular code changes"
+  - name: "risk:high"
+    color: b60205
+    description: "High risk: backend API, agents, sandbox, auth, deps, CI"
+
+  # ── Priority (manual) ─────────────────────────────────────────────────────
+  - name: P0
+    color: b60205
+    description: Critical priority
+  - name: P1
+    color: d93f0b
+    description: Major priority
+  - name: P2
+    color: e99695
+    description: Normal priority
+
+  # ── Status (auto + manual) ────────────────────────────────────────────────
+  - name: needs-triage
+    color: fef2c0
+    description: Awaiting maintainer triage
+  - name: needs-validation
+    color: d4c5f9
+    description: Touches front/back contract surface; needs real-path validation
+  - name: skip-validation
+    color: cccccc
+    description: "Maintainer override: do not auto-add needs-validation on this PR"
+  - name: reviewing
+    color: 5319e7
+    description: A maintainer is reviewing this PR
+
+  # ── Contributor ───────────────────────────────────────────────────────────
+  - name: first-time-contributor
+    color: c2e0c6
+    description: First contribution to this repository — be welcoming

.github/pull_request_template.md

@@ -0,0 +1,75 @@
+<!-- Reference a related issue with #123. Use Fixes / Closes / Resolves to
+     auto-close it on merge. Delete this line if the PR doesn't reference an issue. -->
+Fixes #
+
+## Why
+
+<!-- Why are you opening this PR? Cover two things:
+       - The trigger — what made you write this? A bug you hit, a feature you need,
+         tech debt, or a prod issue?
+       - The pain being addressed — user-facing problem, or what it unblocks.
+     For non-trivial features, please open an issue/discussion first to align on
+     scope before writing code. -->
+
+
+## What changed
+
+<!-- Describe the change from a user's / caller's perspective, not as a code diff. e.g.:
+       - "Settings now has a 'Custom endpoint' field, off by default"
+       - "Backend /api/chat gains a `stream` flag, defaults to false"
+       - "Default model changed from X to Y — existing users notice on first run" -->
+
+
+## Surface area
+
+<!-- Check every box that applies. Reviewers use this to scope the review. -->
+
+- [ ] **Frontend UI** — page / component / setting / interaction under `frontend/`
+- [ ] **Backend API** — endpoint / SSE event / request-response shape under `backend/app`
+- [ ] **Agents / LangGraph** — agent node, graph wiring, `langgraph.json`, or prompt change
+- [ ] **Sandbox** — `docker/` or sandboxed execution
+- [ ] **Skills** — change under `skills/`
+- [ ] **Dependencies** — new/upgraded entry in `backend/pyproject.toml` or `frontend/package.json` (say what it buys us)
+- [ ] **Default behavior change** — changes existing behavior without the user opting in (default model, default setting, data shape)
+- [ ] **Docs / tests / CI only** — no runtime behavior change
+
+
+## Screenshots / Recording
+
+<!-- If you checked "Frontend UI", attach screenshots showing the entry point —
+     where users discover the change — not just the feature in isolation.
+     Before/after is best for behavior changes. Short GIFs welcome. -->
+
+
+## Bug fix verification
+
+<!-- Skip (delete) this section if this PR is not a bug fix.
+
+     Bugs should be encoded as a failing test that goes red before the fix.
+     Confirm:
+       - Test path that reproduces the bug:
+       - Did it go red on `main` and green on this branch? (yes / no)
+       - If a red test wasn't cheap to write, explain why and what you did instead. -->
+
+
+## Validation
+
+<!-- What you actually ran. Run at least the checks for the area you changed:
+       Backend:   cd backend  && make lint && make test
+       Frontend:  cd frontend && pnpm format && pnpm lint && pnpm typecheck && BETTER_AUTH_SECRET=local-dev-secret pnpm build && make test
+       Frontend E2E (if you touched frontend/): cd frontend && make test-e2e -->
+
+
+## AI assistance
+
+<!-- DeerFlow is an AI project — most PRs here use AI coding tools, and that's
+     welcome. Disclosing it just helps reviewers calibrate how closely to read the
+     diff. Please fill all three; don't delete the section. -->
+
+**Tool(s) used:** <!-- e.g. Claude Code, Cursor, GitHub Copilot, Codex, Windsurf, or "none" -->
+
+**How you used it:** <!-- e.g. "generated the module from a spec", "autocomplete only",
+     "AI wrote tests, I wrote the impl". A prompt or conversation link is great too. -->
+
+- [ ] I've read and understand every line of this change and take responsibility for it — it's not unreviewed AI output.
+

.github/workflows/backend-blocking-io-tests.yml

@@ -0,0 +1,46 @@
+name: Backend Blocking IO
+
+on:
+  push:
+    branches: ["main"]
+    paths:
+      - "backend/**"
+      - ".github/workflows/backend-blocking-io-tests.yml"
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+    paths:
+      - "backend/**"
+      - ".github/workflows/backend-blocking-io-tests.yml"
+
+concurrency:
+  group: blocking-io-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  backend-blocking-io:
+    if: github.event_name != 'pull_request' || github.event.pull_request.draft == false
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
+
+      - name: Install backend dependencies
+        working-directory: backend
+        run: uv sync --group dev
+
+      - name: Run blocking IO regression tests
+        working-directory: backend
+        run: make test-blocking-io

.github/workflows/label-sync.yml

@@ -0,0 +1,38 @@
+name: Label Sync
+
+# Keeps repository labels in sync with the declarative source of truth
+# (.github/labels.yml). Runs whenever that file changes on main, and can be
+# triggered manually. Additive/update-only — never deletes labels.
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - ".github/labels.yml"
+      - "scripts/sync_labels.py"
+      - ".github/workflows/label-sync.yml"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  issues: write
+
+concurrency:
+  group: label-sync
+  cancel-in-progress: false
+
+jobs:
+  sync:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+
+      - name: Sync labels
+        run: uv run --with pyyaml python scripts/sync_labels.py
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GH_REPO: ${{ github.repository }}

.github/workflows/replay-e2e.yml

@@ -0,0 +1,108 @@
+name: Replay E2E (front-back contract)
+
+# Guards the front-back contract via record/replay (no API key in CI):
+#   Layer 1 — backend golden: replay a recorded trace through the real gateway,
+#             assert the SSE event sequence matches the committed golden.
+#   Layer 2 — full-stack render: real Next.js frontend + real gateway (replay
+#             model) + Chromium; assert the replayed turns render in the browser.
+# Triggered by changes on EITHER side of the contract so a backend change can no
+# longer pass without the frontend-facing checks running.
+
+on:
+  push:
+    branches: ["main"]
+    paths:
+      - "frontend/**"
+      - "backend/app/gateway/**"
+      - "backend/packages/harness/**"
+      - "backend/tests/fixtures/replay/**"
+      - "backend/tests/replay_provider.py"
+      - "backend/tests/_replay_fixture.py"
+      - "backend/tests/seed_runs_router.py"
+      - "backend/tests/test_replay_golden.py"
+      - "backend/scripts/run_replay_gateway.py"
+      - ".github/workflows/replay-e2e.yml"
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+    paths:
+      - "frontend/**"
+      - "backend/app/gateway/**"
+      - "backend/packages/harness/**"
+      - "backend/tests/fixtures/replay/**"
+      - "backend/tests/replay_provider.py"
+      - "backend/tests/_replay_fixture.py"
+      - "backend/tests/seed_runs_router.py"
+      - "backend/tests/test_replay_golden.py"
+      - "backend/scripts/run_replay_gateway.py"
+      - ".github/workflows/replay-e2e.yml"
+
+concurrency:
+  group: replay-e2e-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  backend-replay-golden:
+    name: Layer 1 — backend golden (no API key)
+    if: github.event_name != 'pull_request' || github.event.pull_request.draft == false
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v6
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+      - name: Install backend dependencies
+        working-directory: backend
+        run: uv sync --group dev
+      - name: Replay golden (backend SSE contract)
+        working-directory: backend
+        run: PYTHONPATH=. uv run pytest tests/test_replay_golden.py -v
+
+  fullstack-replay-render:
+    name: Layer 2 — full-stack render (no API key)
+    if: github.event_name != 'pull_request' || github.event.pull_request.draft == false
+    runs-on: ubuntu-latest
+    timeout-minutes: 25
+    steps:
+      - uses: actions/checkout@v6
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+      - name: Install backend dependencies (replay gateway)
+        working-directory: backend
+        run: uv sync --group dev
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+      - name: Enable Corepack
+        run: corepack enable
+      - name: Use pinned pnpm version
+        run: corepack prepare pnpm@10.26.2 --activate
+      - name: Install frontend dependencies
+        working-directory: frontend
+        run: pnpm install --frozen-lockfile
+      - name: Install Playwright Chromium
+        working-directory: frontend
+        run: npx playwright install chromium --with-deps
+      - name: Full-stack replay render (DOM assertions are the gate)
+        working-directory: frontend
+        run: pnpm exec playwright test -c playwright.real-backend.config.ts
+      - name: Upload report + render artifact
+        uses: actions/upload-artifact@v4
+        if: ${{ !cancelled() }}
+        with:
+          name: replay-render
+          path: |
+            frontend/playwright-report/
+            frontend/test-results/
+          retention-days: 7

.github/workflows/triage.yml

@@ -0,0 +1,223 @@
+name: Triage
+
+# One workflow for all event-driven PR/issue labeling. Replaces the former
+# pr-labeler / pr-triage / issue-triage workflows (and drops actions/labeler).
+#
+# Design notes:
+#   * All jobs are pure-metadata: they read changed-file lists / PR fields / the
+#     review payload via the API and write labels. PR code is NEVER checked out
+#     or executed, so pull_request_target is safe here.
+#   * Each job only reconciles labels in namespaces IT owns
+#     (area:* / size/* / risk:* / needs-validation). It never touches labels
+#     applied by maintainers or other tools (bug, priority, etc.). first-time-
+#     contributor and reviewing are add-only.
+#   * State is read LIVE (listFiles + listLabelsOnIssue) at run time, not from
+#     the (stale) event payload, so rapid synchronize events converge instead
+#     of thrashing.
+
+on:
+  pull_request_target:
+    types: [opened, synchronize, reopened, ready_for_review]
+  pull_request_review:
+    types: [submitted]
+  issues:
+    types: [opened]
+
+permissions:
+  contents: read
+  pull-requests: write
+  issues: write
+
+jobs:
+  # ── PR: area / size / risk / needs-validation / first-time ─────────────────
+  pr-labels:
+    if: github.event_name == 'pull_request_target' && github.event.pull_request.draft == false
+    runs-on: ubuntu-latest
+    concurrency:
+      group: triage-pr-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Apply PR labels from live state
+        uses: actions/github-script@v8
+        with:
+          script: |
+            const pr = context.payload.pull_request;
+            const { owner, repo } = context.repo;
+            const num = pr.number;
+
+            // ---- live changed files ----
+            const files = await github.paginate(github.rest.pulls.listFiles, {
+              owner, repo, pull_number: num, per_page: 100,
+            });
+            const paths = files.map(f => f.filename);
+            const m = (re) => paths.some(p => re.test(p));
+
+            // ---- area: replaces .github/labeler.yml (path -> area) ----
+            const AREA_RULES = [
+              ['area:frontend', [/^frontend\//]],
+              ['area:backend',  [/^backend\/app\//, /^backend\/packages\/harness\/deerflow\/(runtime|persistence|config|tools|guardrails|tracing|models|utils|uploads)\//]],
+              ['area:agents',   [/^backend\/packages\/harness\/deerflow\/(agents|subagents|reflection)\//, /(^|\/)langgraph\.json$/, /^backend\/.*\/prompts\//]],
+              ['area:sandbox',  [/^docker\//, /^backend\/packages\/harness\/deerflow\/sandbox\//, /(^|\/)Dockerfile$/]],
+              ['area:skills',   [/^skills\//, /^backend\/packages\/harness\/deerflow\/skills\//, /^frontend\/src\/core\/skills\//]],
+              ['area:mcp',      [/^backend\/packages\/harness\/deerflow\/mcp\//, /^frontend\/src\/core\/mcp\//]],
+              ['area:ci',       [/^\.github\//, /^scripts\//]],
+              ['area:docs',     [/^docs\//, /\.mdx?$/]],
+              ['area:deps',     [/(^|\/)(pyproject\.toml|uv\.lock|package\.json|pnpm-lock\.yaml)$/]],
+            ];
+            const areaLabels = AREA_RULES
+              .filter(([, res]) => res.some(re => m(re)))
+              .map(([label]) => label);
+
+            // ---- size: additions+deletions, excluding lockfiles/snapshots ----
+            const EXCLUDE_SIZE = /(^|\/)(uv\.lock|pnpm-lock\.yaml|package-lock\.json)$|\.snap$/;
+            const churn = files
+              .filter(f => !EXCLUDE_SIZE.test(f.filename))
+              .reduce((s, f) => s + (f.additions || 0) + (f.deletions || 0), 0);
+            const sizeLabel =
+              churn < 20 ? 'size/XS' :
+              churn < 100 ? 'size/S' :
+              churn < 300 ? 'size/M' :
+              churn < 700 ? 'size/L' : 'size/XL';
+
+            // ---- risk ----
+            const docsOnly = paths.length > 0 && paths.every(p =>
+              /\.(md|mdx|txt)$/i.test(p) || p.startsWith('docs/') ||
+              /\.(png|jpe?g|gif|svg|webp|ico)$/i.test(p));
+            const highRisk =
+              m(/^backend\/app\/gateway\//) ||
+              m(/^backend\/packages\/harness\/deerflow\/(agents|subagents|sandbox)\//) ||
+              m(/(^|\/)langgraph\.json$/) ||
+              m(/(^|\/)(auth|authz|security)/i) ||
+              m(/(pyproject\.toml|uv\.lock|package\.json|pnpm-lock\.yaml)$/) ||
+              m(/^docker\//) ||
+              m(/^\.github\/workflows\//);
+            const riskLabel = docsOnly ? 'risk:low' : (highRisk ? 'risk:high' : 'risk:medium');
+
+            // ---- needs-validation: front/back contract surface ----
+            const contract =
+              m(/^backend\/app\/gateway\//) ||
+              m(/^backend\/packages\/harness\/deerflow\/(agents|subagents)\//) ||
+              m(/(^|\/)langgraph\.json$/) ||
+              m(/^frontend\/src\/core\/(api|threads|messages)\//);
+
+            // ---- live current labels (NOT the stale event payload) ----
+            const current = (await github.paginate(github.rest.issues.listLabelsOnIssue, {
+              owner, repo, issue_number: num, per_page: 100,
+            })).map(l => l.name);
+            const hasSkip = current.includes('skip-validation');
+
+            // Reconcile ONLY namespaces we own; never touch others.
+            const owned = (n) =>
+              n.startsWith('area:') || n.startsWith('size/') ||
+              n.startsWith('risk:') || n === 'needs-validation';
+            const desired = new Set([...areaLabels, sizeLabel, riskLabel]);
+            if (contract && !hasSkip) desired.add('needs-validation');
+
+            const toRemove = current.filter(n => owned(n) && !desired.has(n));
+            const toAdd = [...desired].filter(n => !current.includes(n));
+
+            // first-time-contributor: add-only, on opened, real users only.
+            if (context.payload.action === 'opened' &&
+                pr.user.type === 'User' &&
+                ['FIRST_TIME_CONTRIBUTOR', 'FIRST_TIMER'].includes(pr.author_association) &&
+                !current.includes('first-time-contributor')) {
+              toAdd.push('first-time-contributor');
+            }
+
+            for (const name of toRemove) {
+              try {
+                await github.rest.issues.removeLabel({ owner, repo, issue_number: num, name });
+              } catch (e) {
+                if (e.status !== 404) throw e;
+              }
+            }
+            if (toAdd.length) {
+              await github.rest.issues.addLabels({ owner, repo, issue_number: num, labels: toAdd });
+            }
+            core.info(`area=[${areaLabels.join(',')}] ${sizeLabel} ${riskLabel} churn=${churn} ` +
+              `validation=${desired.has('needs-validation')} ` +
+              `(+${toAdd.join(',') || '-'} / -${toRemove.join(',') || '-'})`);
+
+  # ── PR: reviewing label on a maintainer's human review ─────────────────────
+  reviewing:
+    if: github.event_name == 'pull_request_review'
+    runs-on: ubuntu-latest
+    concurrency:
+      group: triage-review-${{ github.event.pull_request.number }}
+      cancel-in-progress: false
+    steps:
+      - name: Add reviewing label for maintainer reviews
+        uses: actions/github-script@v8
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const num = context.payload.pull_request.number;
+            const review = context.payload.review;
+            const assoc = review.author_association;     // payload field; no API call
+            const type = review.user && review.user.type;
+
+            // author_association is NONE for every automated reviewer
+            // (Copilot, CodeRabbit, Codex, Sourcery, ...), so this allowlist
+            // drops them all without a denylist — and never calls the
+            // collaborators API that 404s on "Copilot is not a user".
+            // user.type === 'User' guards the rare bot-added-as-collaborator case.
+            if (!['OWNER', 'MEMBER', 'COLLABORATOR'].includes(assoc) || type !== 'User') {
+              core.info(`reviewer ${review.user && review.user.login} assoc=${assoc} type=${type}; skipping.`);
+              return;
+            }
+
+            const labels = (await github.paginate(github.rest.issues.listLabelsOnIssue, {
+              owner, repo, issue_number: num, per_page: 100,
+            })).map(l => l.name);
+            if (labels.includes('reviewing')) {
+              core.info('Already labeled reviewing; skipping.');
+              return;
+            }
+            try {
+              await github.rest.issues.addLabels({
+                owner, repo, issue_number: num, labels: ['reviewing'],
+              });
+              core.info('Added "reviewing".');
+            } catch (e) {
+              if (e.status === 403) core.info('No permission to label (expected on some fork PRs).');
+              else throw e;
+            }
+
+  # ── Issue: needs-triage on every new issue ────────────────────────────────
+  issue-triage:
+    if: github.event_name == 'issues'
+    runs-on: ubuntu-latest
+    concurrency:
+      group: triage-issue-${{ github.event.issue.number }}
+      cancel-in-progress: false
+    steps:
+      - name: Add needs-triage label
+        uses: actions/github-script@v8
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const issue_number = context.payload.issue.number;
+
+            // Read live labels (not the event payload) so labels added at creation
+            // time via the API or by another automation are seen — consistent with
+            // the live-state reads in the PR jobs above.
+            const current = (await github.paginate(github.rest.issues.listLabelsOnIssue, {
+              owner, repo, issue_number, per_page: 100,
+            })).map(l => l.name);
+            if (current.includes('needs-triage')) {
+              core.info('Issue already has needs-triage; nothing to do.');
+              return;
+            }
+            // Self-heal: create the label if it does not exist yet.
+            try {
+              await github.rest.issues.createLabel({
+                owner, repo, name: 'needs-triage', color: 'fef2c0',
+                description: 'Awaiting maintainer triage',
+              });
+            } catch (e) {
+              if (e.status !== 422) throw e; // 422 = already exists
+            }
+            await github.rest.issues.addLabels({
+              owner, repo, issue_number, labels: ['needs-triage'],
+            });
+            core.info(`Added needs-triage to #${issue_number}.`);

CONTRIBUTING.md

@@ -287,6 +287,21 @@ Nginx (port 2026) ← Unified entry point
    git push origin feature/your-feature-name
    ```
 
+## AI assistance disclosure
+
+DeerFlow is an AI project and we welcome AI-assisted contributions. To help
+reviewers calibrate how closely to read a change, **every pull request must
+complete the "AI assistance" section of the
+[PR template](.github/pull_request_template.md)**:
+
+- which tool(s) you used (or `none`),
+- how you used them, and
+- a confirmation that a human has read, understands, and takes responsibility
+  for the change.
+
+Please don't delete the section. PRs that ignore it may be asked to fill it in
+before review.
+
 ## Testing
 
 ```bash

Makefile

@@ -1,6 +1,6 @@
 # DeerFlow - Unified Development Environment
 
-.PHONY: help config config-upgrade check install setup doctor dev dev-daemon start start-daemon stop up down clean docker-init docker-start docker-stop docker-logs docker-logs-frontend docker-logs-gateway
+.PHONY: help config config-upgrade check install setup doctor detect-thread-boundaries detect-blocking-io dev dev-daemon start start-daemon stop up down clean docker-init docker-start docker-stop docker-logs docker-logs-frontend docker-logs-gateway
 
 BASH ?= bash
 BACKEND_UV_RUN = cd backend && uv run
@@ -23,6 +23,8 @@ help:
 	@echo "  make config          - Generate local config files (aborts if config already exists)"
 	@echo "  make config-upgrade  - Merge new fields from config.example.yaml into config.yaml"
 	@echo "  make check           - Check if all required tools are installed"
+	@echo "  make detect-thread-boundaries - Inventory async/thread boundary points"
+	@echo "  make detect-blocking-io        - Inventory blocking IO that may block the backend event loop"
 	@echo "  make install         - Install all dependencies (frontend + backend + pre-commit hooks)"
 	@echo "  make setup-sandbox   - Pre-pull sandbox container image (recommended)"
 	@echo "  make dev             - Start all services in development mode (with hot-reloading)"
@@ -51,6 +53,12 @@ setup:
 doctor:
 	@$(BACKEND_UV_RUN) python ../scripts/doctor.py
 
+detect-thread-boundaries:
+	@$(PYTHON) ./scripts/detect_thread_boundaries.py
+
+detect-blocking-io:
+	@$(MAKE) -C backend detect-blocking-io
+
 config:
 	@$(PYTHON) ./scripts/configure.py
 
@@ -81,36 +89,7 @@ install:
 
 # Pre-pull sandbox Docker image (optional but recommended)
 setup-sandbox:
-	@echo "=========================================="
-	@echo "  Pre-pulling Sandbox Container Image"
-	@echo "=========================================="
-	@echo ""
-	@IMAGE=$$(grep -A 20 "# sandbox:" config.yaml 2>/dev/null | grep "image:" | awk '{print $$2}' | head -1); \
-	if [ -z "$$IMAGE" ]; then \
-		IMAGE="enterprise-public-cn-beijing.cr.volces.com/vefaas-public/all-in-one-sandbox:latest"; \
-		echo "Using default image: $$IMAGE"; \
-	else \
-		echo "Using configured image: $$IMAGE"; \
-	fi; \
-	echo ""; \
-	if command -v container >/dev/null 2>&1 && [ "$$(uname)" = "Darwin" ]; then \
-		echo "Detected Apple Container on macOS, pulling image..."; \
-		container image pull "$$IMAGE" || echo "⚠ Apple Container pull failed, will try Docker"; \
-	fi; \
-	if command -v docker >/dev/null 2>&1; then \
-		echo "Pulling image using Docker..."; \
-		if docker pull "$$IMAGE"; then \
-			echo ""; \
-			echo "✓ Sandbox image pulled successfully"; \
-		else \
-			echo ""; \
-			echo "⚠ Failed to pull sandbox image (this is OK for local sandbox mode)"; \
-		fi; \
-	else \
-		echo "✗ Neither Docker nor Apple Container is available"; \
-		echo "  Please install Docker: https://docs.docker.com/get-docker/"; \
-		exit 1; \
-	fi
+	@$(RUN_WITH_GIT_BASH) ./scripts/setup-sandbox.sh
 
 # Start all services in development mode (with hot-reloading)
 dev:
@@ -140,7 +119,6 @@ stop:
 clean: stop
 	@echo "Cleaning up..."
 	@-rm -rf backend/.deer-flow 2>/dev/null || true
-	@-rm -rf backend/.langgraph_api 2>/dev/null || true
 	@-rm -rf logs/*.log 2>/dev/null || true
 	@echo "✓ Cleanup complete"
 

README.md

@@ -546,6 +546,15 @@ LANGFUSE_BASE_URL=https://cloud.langfuse.com
 
 If you are using a self-hosted Langfuse instance, set `LANGFUSE_BASE_URL` to your deployment URL.
 
+**Trace correlation fields.** Every agent run is annotated with Langfuse's reserved trace attributes so the Sessions and Users pages light up automatically:
+
+- `session_id` = LangGraph `thread_id` — groups every trace of the same conversation
+- `user_id` = effective user from `get_effective_user_id()` (falls back to `default` in no-auth mode)
+- `trace_name` = assistant id (defaults to `lead-agent`)
+- `tags` = `[env:<DEER_FLOW_ENV>, model:<model_name>]` (omitted when not set)
+
+These are injected into `RunnableConfig.metadata` at the graph invocation root for both the gateway path (`runtime/runs/worker.py::run_agent`) and the embedded path (`client.py::DeerFlowClient.stream`), so any LangChain-compatible callback can read them. Set `DEER_FLOW_ENV` (or `ENVIRONMENT`) to tag traces by deployment environment.
+
 #### Using Both Providers
 
 If both LangSmith and Langfuse are enabled, DeerFlow attaches both tracing callbacks and reports the same model activity to both systems.
@@ -576,6 +585,8 @@ A standard Agent Skill is a structured capability module — a Markdown file tha
 
 Skills are loaded progressively — only when the task needs them, not all at once. This keeps the context window lean and makes DeerFlow work well even with token-sensitive models.
 
+Users can explicitly activate an enabled skill for a single turn by starting the request with `/skill-name`, for example `/data-analysis analyze uploads/foo.csv`. DeerFlow loads that skill's `SKILL.md` as hidden current-turn context while leaving the base prompt limited to skill metadata. Slash activation respects disabled skills, custom-agent skill whitelists, and existing channel commands such as `/new` and `/help`.
+
 When you install `.skill` archives through the Gateway, DeerFlow accepts standard optional frontmatter metadata such as `version`, `author`, and `compatibility` instead of rejecting otherwise valid external skills.
 
 Tools follow the same philosophy. DeerFlow comes with a core toolset — web search, web fetch, file operations, bash execution — and supports custom tools via MCP servers and Python functions. Swap anything. Add anything.
@@ -731,6 +742,12 @@ DeerFlow has key high-privilege capabilities including **system command executio
 We welcome contributions! Please see [CONTRIBUTING.md](CONTRIBUTING.md) for development setup, workflow, and guidelines.
 
 Regression coverage includes Docker sandbox mode detection and provisioner kubeconfig-path handling tests in `backend/tests/`.
+Backend blocking-IO diagnostics are available from the repository root with
+`make detect-blocking-io`: it statically scans backend business code for
+blocking IO that may run on the backend event loop, prints a concise summary,
+and writes complete JSON findings to `.deer-flow/blocking-io-findings.json`.
+The JSON includes compact review records with `priority`, `location`,
+`blocking_call`, `event_loop_exposure`, `reason`, and `code`.
 Gateway artifact serving now forces active web content types (`text/html`, `application/xhtml+xml`, `image/svg+xml`) to download as attachments instead of inline rendering, reducing XSS risk for generated artifacts.
 
 ## License

backend/.gitignore

@@ -24,5 +24,10 @@ config.yaml
 # Langgraph
 .langgraph_api
 
+# Sandbox runtime working dir — pre-created and excluded from uvicorn reload
+# (scripts/serve.sh, docker/dev-entrypoint.sh). Anchored so it does not match
+# the source package backend/packages/harness/deerflow/sandbox/.
+/sandbox/
+
 # Claude Code settings
 .claude/settings.local.json

backend/CLAUDE.md

@@ -88,18 +88,57 @@ make stop       # Stop all services
 
 **Backend directory** (for backend development only):
 ```bash
-make install    # Install backend dependencies
-make dev        # Run Gateway API with reload (port 8001)
-make gateway    # Run Gateway API only (port 8001)
-make test       # Run all backend tests
-make lint       # Lint with ruff
-make format     # Format code with ruff
+make install            # Install backend dependencies
+make dev                # Run Gateway API with reload (port 8001)
+make gateway            # Run Gateway API only (port 8001)
+make test               # Run all backend tests
+make test-blocking-io   # Run strict Blockbuster runtime gate on tests/blocking_io/
+make lint               # Lint with ruff
+make format             # Format code with ruff
 ```
 
+The `detect-blocking-io` target parses `app/`, `packages/harness/deerflow/`,
+and `scripts/` with AST. By default it reports only blocking IO candidates that
+are inside async code, reachable from async code in the same file, or reachable
+from sync-only `AgentMiddleware` before/after hooks that LangGraph can execute
+on the async graph path. It prints a concise summary and writes complete JSON
+findings to `.deer-flow/blocking-io-findings.json` at the repository root
+(both `make detect-blocking-io` from the repo root and `cd backend && make
+detect-blocking-io` resolve to the same repo-root path). JSON findings include
+`priority`, `location`, `blocking_call`, `event_loop_exposure`, `reason`, and
+`code` for model-assisted or manual review. `priority` is a deterministic
+review ordering from operation type, not proof of a bug. Bare-name same-file
+calls are resolved by function name, so duplicate helper names in one file can
+conservatively over-report async reachability. It is intentionally
+informational and is not run from CI in this round.
+
 Regression tests related to Docker/provisioner behavior:
 - `tests/test_docker_sandbox_mode_detection.py` (mode detection from `config.yaml`)
 - `tests/test_provisioner_kubeconfig.py` (kubeconfig file/directory handling)
 
+Blocking-IO runtime gate (`tests/blocking_io/`):
+- Wraps every item under `tests/blocking_io/` with a strict Blockbuster
+  context scoped to `app.*` and `deerflow.*` (see
+  `tests/support/detectors/blocking_io_runtime.py`). Any sync blocking IO
+  call whose stack passes through DeerFlow business code while running on
+  the asyncio event loop raises `BlockingError` and fails the test.
+- Regression anchors live there: `test_skills_load.py` (locks the
+  `asyncio.to_thread` offload around `LocalSkillStorage.load_skills`, fix
+  for #1917); `test_sqlite_lifespan.py` (locks the offload around
+  SQLite path resolution plus `ensure_sqlite_parent_dir`, fix for #1912);
+  `test_jsonl_run_event_store.py` (locks `JsonlRunEventStore`'s async
+  API offloading its file IO via `asyncio.to_thread`, fix #3084); and
+  `test_uploads_middleware.py` (locks `UploadsMiddleware.abefore_agent`
+  offloading the uploads-directory scan off the event loop).
+- `test_gate_smoke.py` is a meta-test asserting the gate actually catches
+  unoffloaded blocking IO and that the `@pytest.mark.allow_blocking_io`
+  opt-out works.
+- Coverage boundary: the gate only sees code that test execution actually
+  touches. Static AST coverage is a separate concern (out of scope for
+  this PR).
+- CI: runs on every PR via `.github/workflows/backend-blocking-io-tests.yml`,
+  hard-fail.
+
 Boundary check (harness → app import firewall):
 - `tests/test_harness_boundary.py` — ensures `packages/harness/deerflow/` never imports from `app.*`
 
@@ -153,7 +192,7 @@ from deerflow.config import get_app_config
 
 ### Middleware Chain
 
-Lead-agent middlewares are assembled in strict append order across `packages/harness/deerflow/agents/middlewares/tool_error_handling_middleware.py` (`build_lead_runtime_middlewares`) and `packages/harness/deerflow/agents/lead_agent/agent.py` (`_build_middlewares`):
+Lead-agent middlewares are assembled in strict append order across `packages/harness/deerflow/agents/middlewares/tool_error_handling_middleware.py` (`build_lead_runtime_middlewares`) and `packages/harness/deerflow/agents/lead_agent/agent.py` (`build_middlewares`):
 
 1. **ThreadDataMiddleware** - Creates per-thread directories under the user's isolation scope (`backend/.deer-flow/users/{user_id}/threads/{thread_id}/user-data/{workspace,uploads,outputs}`); resolves `user_id` via `get_effective_user_id()` (falls back to `"default"` in no-auth mode); Web UI thread deletion now follows LangGraph thread removal with Gateway cleanup of the local thread directory
 2. **UploadsMiddleware** - Tracks and injects newly uploaded files into conversation
@@ -163,16 +202,17 @@ Lead-agent middlewares are assembled in strict append order across `packages/har
 6. **GuardrailMiddleware** - Pre-tool-call authorization via pluggable `GuardrailProvider` protocol (optional, if `guardrails.enabled` in config). Evaluates each tool call and returns error ToolMessage on deny. Three provider options: built-in `AllowlistProvider` (zero deps), OAP policy providers (e.g. `aport-agent-guardrails`), or custom providers. See [docs/GUARDRAILS.md](docs/GUARDRAILS.md) for setup, usage, and how to implement a provider.
 7. **SandboxAuditMiddleware** - Audits sandboxed shell/file operations for security logging before tool execution continues
 8. **ToolErrorHandlingMiddleware** - Converts tool exceptions into error `ToolMessage`s so the run can continue instead of aborting
-9. **SummarizationMiddleware** - Context reduction when approaching token limits (optional, if enabled)
-10. **TodoListMiddleware** - Task tracking with `write_todos` tool (optional, if plan_mode)
-11. **TokenUsageMiddleware** - Records token usage metrics when token tracking is enabled (optional); subagent usage is cached by `tool_call_id` only while token usage is enabled and merged back into the dispatching AIMessage by message position rather than message id
-12. **TitleMiddleware** - Auto-generates thread title after first complete exchange and normalizes structured message content before prompting the title model
-13. **MemoryMiddleware** - Queues conversations for async memory update (filters to user + final AI responses)
-14. **ViewImageMiddleware** - Injects base64 image data before LLM call (conditional on vision support)
-15. **DeferredToolFilterMiddleware** - Hides deferred tool schemas from the bound model until tool search is enabled (optional)
-16. **SubagentLimitMiddleware** - Truncates excess `task` tool calls from model response to enforce `MAX_CONCURRENT_SUBAGENTS` limit (optional, if `subagent_enabled`)
-17. **LoopDetectionMiddleware** - Detects repeated tool-call loops; hard-stop responses clear both structured `tool_calls` and raw provider tool-call metadata before forcing a final text answer
-18. **ClarificationMiddleware** - Intercepts `ask_clarification` tool calls, interrupts via `Command(goto=END)` (must be last)
+9. **SkillActivationMiddleware** - Detects strict `/skill-name task` syntax on the latest real user message, resolves only enabled and runtime-allowed skills, reads `SKILL.md` from trusted skill storage, injects the skill body as hidden current-turn model context, and records a `middleware:skill_activation` audit event with skill name, category, path, and content hash
+10. **SummarizationMiddleware** - Context reduction when approaching token limits (optional, if enabled)
+11. **TodoListMiddleware** - Task tracking with `write_todos` tool (optional, if plan_mode)
+12. **TokenUsageMiddleware** - Records token usage metrics when token tracking is enabled (optional); subagent usage is cached by `tool_call_id` only while token usage is enabled and merged back into the dispatching AIMessage by message position rather than message id
+13. **TitleMiddleware** - Auto-generates thread title after first complete exchange and normalizes structured message content before prompting the title model
+14. **MemoryMiddleware** - Queues conversations for async memory update (filters to user + final AI responses)
+15. **ViewImageMiddleware** - Injects base64 image data before LLM call (conditional on vision support)
+16. **DeferredToolFilterMiddleware** - Hides deferred (MCP) tool schemas from the bound model using a build-time deferred-name set + catalog hash, reading per-thread promotions from `ThreadState.promoted` (hash-scoped, no ContextVar); a tool becomes bound on subsequent turns after `tool_search` returns its schema (optional, if `tool_search.enabled`)
+17. **SubagentLimitMiddleware** - Truncates excess `task` tool calls from model response to enforce `MAX_CONCURRENT_SUBAGENTS` limit (optional, if `subagent_enabled`)
+18. **LoopDetectionMiddleware** - Detects repeated tool-call loops; hard-stop responses clear both structured `tool_calls` and raw provider tool-call metadata before forcing a final text answer
+19. **ClarificationMiddleware** - Intercepts `ask_clarification` tool calls, interrupts via `Command(goto=END)` (must be last)
 
 ### Configuration System
 
@@ -184,6 +224,10 @@ Setup: Copy `config.example.yaml` to `config.yaml` in the **project root** direc
 
 **Config Caching**: `get_app_config()` caches the parsed config, but automatically reloads it when the resolved config path changes or the file's mtime increases. This keeps Gateway and LangGraph reads aligned with `config.yaml` edits without requiring a manual process restart.
 
+**Config Hot-Reload Boundary**: Gateway dependencies route through `get_app_config()` on every request, so per-run fields like `models[*].max_tokens`, `summarization.*`, `title.*`, `memory.*`, `subagents.*`, `tools[*]`, and the agent system prompt pick up `config.yaml` edits on the next message. `AppConfig` is intentionally **not** cached on `app.state` — `lifespan()` keeps a local `startup_config` variable for one-shot bootstrap work and passes it to `langgraph_runtime(app, startup_config)`.
+
+Infrastructure fields are **restart-required**. The authoritative list lives in `packages/harness/deerflow/config/reload_boundary.py::STARTUP_ONLY_FIELDS` and is mirrored by the standardised `"startup-only:"` prefix on the corresponding `Field(description=...)` in `AppConfig`, so IDE hover on those fields surfaces the reason inline (no need to context-switch into this table). Currently registered: `database`, `checkpointer`, `run_events`, `stream_bridge`, `sandbox`, `log_level`, `channels`. Adding a new restart-required field requires updating the registry; drift is pinned by `tests/test_reload_boundary.py`.
+
 Configuration priority:
 1. Explicit `config_path` argument
 2. `DEER_FLOW_CONFIG_PATH` environment variable
@@ -220,7 +264,7 @@ CORS is same-origin by default when requests enter through nginx on port 2026. S
 | **Uploads** (`/api/threads/{id}/uploads`) | `POST /` - upload files (auto-converts PDF/PPT/Excel/Word); `GET /list` - list; `DELETE /{filename}` - delete |
 | **Threads** (`/api/threads/{id}`) | `DELETE /` - remove DeerFlow-managed local thread data after LangGraph thread deletion; unexpected failures are logged server-side and return a generic 500 detail |
 | **Artifacts** (`/api/threads/{id}/artifacts`) | `GET /{path}` - serve artifacts; active content types (`text/html`, `application/xhtml+xml`, `image/svg+xml`) are always forced as download attachments to reduce XSS risk; `?download=true` still forces download for other file types |
-| **Suggestions** (`/api/threads/{id}/suggestions`) | `POST /` - generate follow-up questions; rich list/block model content is normalized before JSON parsing |
+| **Suggestions** (`/api/threads/{id}/suggestions`) | `POST /` - generate follow-up questions; rich list/block model content is normalized and inline reasoning (`<think>...</think>`, including unclosed/truncated blocks from reasoning models like MiniMax-M3) is stripped before JSON parsing |
 | **Thread Runs** (`/api/threads/{id}/runs`) | `POST /` - create background run; `POST /stream` - create + SSE stream; `POST /wait` - create + block; `GET /` - list runs; `GET /{rid}` - run details; `POST /{rid}/cancel` - cancel; `GET /{rid}/join` - join SSE; `GET /{rid}/messages` - paginated messages `{data, has_more}`; `GET /{rid}/events` - full event stream; `GET /../messages` - thread messages with feedback; `GET /../token-usage` - aggregate tokens |
 | **Feedback** (`/api/threads/{id}/runs/{rid}/feedback`) | `PUT /` - upsert feedback; `DELETE /` - delete user feedback; `POST /` - create feedback; `GET /` - list feedback; `GET /stats` - aggregate stats; `DELETE /{fid}` - delete specific |
 | **Runs** (`/api/runs`) | `POST /stream` - stateless run + SSE; `POST /wait` - stateless run + block; `GET /{rid}/messages` - paginated messages by run_id `{data, has_more}` (cursor: `after_seq`/`before_seq`); `GET /{rid}/feedback` - list feedback by run_id |
@@ -230,13 +274,14 @@ CORS is same-origin by default when requests enter through nginx on port 2026. S
 - When a persistent `RunStore` is configured, `get()` and `list_by_thread()` hydrate historical runs from the store. In-memory records win for the same `run_id` so task, abort, and stream-control state stays attached to active local runs.
 - `cancel()` and `create_or_reject(..., multitask_strategy="interrupt"|"rollback")` persist interrupted status through `RunStore.update_status()`, matching normal `set_status()` transitions.
 - Store-only hydrated runs are readable history. If the current worker has no in-memory task/control state for that run, cancellation APIs can return 409 because this worker cannot stop the task.
+- `POST /wait` (both thread-scoped and `/api/runs/wait`) drains the stream bridge via `wait_for_run_completion()` instead of bare `await record.task`, so it honours the run's `on_disconnect` setting and cancels the background run on real client disconnect rather than returning a stale checkpoint (issue #3265).
 
 Proxied through nginx: `/api/langgraph/*` → Gateway LangGraph-compatible runtime, all other `/api/*` → Gateway REST APIs.
 
 ### Sandbox System (`packages/harness/deerflow/sandbox/`)
 
 **Interface**: Abstract `Sandbox` with `execute_command`, `read_file`, `write_file`, `list_dir`
-**Provider Pattern**: `SandboxProvider` with `acquire`, `get`, `release` lifecycle
+**Provider Pattern**: `SandboxProvider` with `acquire`, `acquire_async`, `get`, `release` lifecycle. Async agent/tool paths call async sandbox lifecycle hooks so Docker sandbox creation, discovery, cross-process locking, readiness polling, and release stay off the event loop.
 **Implementations**:
 - `LocalSandboxProvider` - Local filesystem execution. `acquire(thread_id)` returns a per-thread `LocalSandbox` (id `local:{thread_id}`) whose `path_mappings` resolve `/mnt/user-data/{workspace,uploads,outputs}` and `/mnt/acp-workspace` to that thread's host directories, so the public `Sandbox` API honours the `/mnt/user-data` contract uniformly with AIO. `acquire()` / `acquire(None)` keeps the legacy generic singleton (id `local`) for callers without a thread context. Per-thread sandboxes are held in an LRU cache (default 256 entries) guarded by a `threading.Lock`.
 - `AioSandboxProvider` (`packages/harness/deerflow/community/`) - Docker-based isolation
@@ -261,6 +306,7 @@ Proxied through nginx: `/api/langgraph/*` → Gateway LangGraph-compatible runti
 **Concurrency**: `MAX_CONCURRENT_SUBAGENTS = 3` enforced by `SubagentLimitMiddleware` (truncates excess tool calls in `after_model`), 15-minute timeout
 **Flow**: `task()` tool → `SubagentExecutor` → background thread → poll 5s → SSE events → result
 **Events**: `task_started`, `task_running`, `task_completed`/`task_failed`/`task_timed_out`
+**Deferred MCP tools** (if `tool_search.enabled`): `SubagentExecutor._build_initial_state` assembles deferral after policy filtering via the shared `assemble_deferred_tools` (fail-closed), appends the `tool_search` tool, injects the `<available-deferred-tools>` section into the subagent's `SystemMessage`, and threads the setup to `_create_agent`, which attaches `DeferredToolFilterMiddleware` through `build_subagent_runtime_middlewares(deferred_setup=...)`. Subagents thus withhold full MCP schemas until promotion, same as the lead agent; each task run gets a fresh `ThreadState` so promotion is isolated per run
 
 ### Tool System (`packages/harness/deerflow/tools/`)
 
@@ -295,7 +341,7 @@ Proxied through nginx: `/api/langgraph/*` → Gateway LangGraph-compatible runti
 - **Cache invalidation**: Detects config file changes via mtime comparison
 - **Transports**: stdio (command-based), SSE, HTTP
 - **OAuth (HTTP/SSE)**: Supports token endpoint flows (`client_credentials`, `refresh_token`) with automatic token refresh + Authorization header injection
-- **Runtime updates**: Gateway API saves to extensions_config.json; LangGraph detects via mtime
+- **Runtime updates**: Gateway API saves to extensions_config.json; the Gateway-embedded runtime detects changes via mtime
 
 ### Skills System (`packages/harness/deerflow/skills/`)
 
@@ -303,6 +349,7 @@ Proxied through nginx: `/api/langgraph/*` → Gateway LangGraph-compatible runti
 - **Format**: Directory with `SKILL.md` (YAML frontmatter: name, description, license, allowed-tools)
 - **Loading**: `load_skills()` recursively scans `skills/{public,custom}` for `SKILL.md`, parses metadata, and reads enabled state from extensions_config.json
 - **Injection**: Enabled skills listed in agent system prompt with container paths
+- **Slash activation**: `/skill-name task` loads that enabled skill's `SKILL.md` for the current model call only. The resolver rejects leading whitespace, missing separators, reserved channel commands (`/new`, `/help`, `/bootstrap`, `/status`, `/models`, `/memory`), disabled skills, and skills outside a custom agent's whitelist.
 - **Installation**: `POST /api/skills/install` extracts .skill ZIP archive to custom/ directory
 
 ### Model Factory (`packages/harness/deerflow/models/factory.py`)
@@ -322,7 +369,7 @@ Proxied through nginx: `/api/langgraph/*` → Gateway LangGraph-compatible runti
 
 ### IM Channels System (`app/channels/`)
 
-Bridges external messaging platforms (Feishu, Slack, Telegram, DingTalk) to the DeerFlow agent via the LangGraph Server.
+Bridges external messaging platforms (Feishu, Slack, Telegram, DingTalk) to the DeerFlow agent via Gateway's LangGraph-compatible API.
 
 
 **Architecture**: Channels communicate with Gateway through the `langgraph-sdk` HTTP client (same as the frontend), ensuring threads are created and managed server-side. The internal SDK client injects process-local internal auth plus a matching CSRF cookie/header pair so Gateway accepts state-changing thread/run requests from channel workers without relying on browser session cookies.
@@ -397,6 +444,24 @@ Focused regression coverage for the updater lives in `backend/tests/test_memory_
 - `resolve_variable(path)` - Import module and return variable (e.g., `module.path:variable_name`)
 - `resolve_class(path, base_class)` - Import and validate class against base class
 
+### Tracing System (`packages/harness/deerflow/tracing/`)
+
+LangSmith and Langfuse are both supported. The wiring lives in two layers:
+
+- `factory.py::build_tracing_callbacks()` — returns the LangChain `CallbackHandler` list for the providers currently enabled via env vars (`LANGSMITH_TRACING`, `LANGFUSE_TRACING`, etc.). The handlers are attached at the **graph invocation root** for in-graph runs (`make_lead_agent` and `DeerFlowClient.stream` both append them to `config["callbacks"]` before invoking the graph) so a single run produces one trace with all node / LLM / tool calls as child spans. Standalone callers — anything that invokes a model outside such a graph (e.g. `MemoryUpdater`) — keep `create_chat_model`'s default `attach_tracing=True`, which falls back to model-level callback attachment.
+- `metadata.py::build_langfuse_trace_metadata()` — builds the Langfuse-reserved trace attributes for `RunnableConfig.metadata`. The Langfuse v4 `langchain.CallbackHandler` lifts these onto the root trace (see its `_parse_langfuse_trace_attributes`), but only when it sees `on_chain_start(parent_run_id=None)` — which is why the callbacks have to live at the graph root, not the model.
+
+**Trace-attribute injection points**: both `runtime/runs/worker.py::run_agent` (gateway path) and `client.py::DeerFlowClient.stream` (embedded path) merge the metadata into `config["metadata"]` right before constructing the graph. Caller-supplied keys win via `setdefault`, so an external `session_id` override is preserved. Field mapping:
+
+| Langfuse field         | Source                                       |
+|-----------------------|----------------------------------------------|
+| `langfuse_session_id` | LangGraph `thread_id`                         |
+| `langfuse_user_id`    | `get_effective_user_id()` (`default` in no-auth) |
+| `langfuse_trace_name` | `RunRecord.assistant_id` / client `agent_name` (defaults to `lead-agent`) |
+| `langfuse_tags`       | `env:<DEER_FLOW_ENV>` + `model:<model_name>`  |
+
+Returns `{}` when Langfuse is not in the enabled providers — LangSmith-only deployments are unaffected. Set `DEER_FLOW_ENV` (or `ENVIRONMENT`) to tag traces by deployment environment. Tests live in `tests/test_tracing_factory.py`, `tests/test_tracing_metadata.py`, `tests/test_worker_langfuse_metadata.py`, and `tests/test_client_langfuse_metadata.py`.
+
 ### Config Schema
 
 **`config.yaml`** key sections:
@@ -430,7 +495,7 @@ Both can be modified at runtime via Gateway API endpoints or `DeerFlowClient` me
   - `"messages-tuple"` — per-chunk update: for AI text this is a **delta** (concat per `id` to rebuild the full message); tool calls and tool results are emitted once each
   - `"custom"` — forwarded from `StreamWriter`
   - `"end"` — stream finished (carries cumulative `usage` counted once per message id)
-- Agent created lazily via `create_agent()` + `_build_middlewares()`, same as `make_lead_agent`
+- Agent created lazily via `create_agent()` + `build_middlewares()`, same as `make_lead_agent`
 - Supports `checkpointer` parameter for state persistence across turns
 - `reset_agent()` forces agent recreation (e.g. after memory or skill changes)
 - See [docs/STREAMING.md](docs/STREAMING.md) for the full design: why Gateway and DeerFlowClient are parallel paths, LangGraph's `stream_mode` semantics, the per-id dedup invariants, and regression testing strategy

backend/Dockerfile

@@ -64,7 +64,7 @@ FROM builder AS dev
 # Install Docker CLI (for DooD: allows starting sandbox containers via host Docker socket)
 COPY --from=docker:cli /usr/local/bin/docker /usr/local/bin/docker
 
-EXPOSE 8001 2024
+EXPOSE 8001
 
 CMD ["sh", "-c", "cd backend && PYTHONPATH=. uv run uvicorn app.gateway.app:app --host 0.0.0.0 --port 8001"]
 
@@ -94,8 +94,8 @@ WORKDIR /app
 # Copy backend with pre-built virtualenv from builder
 COPY --from=builder /app/backend ./backend
 
-# Expose ports (gateway: 8001, langgraph: 2024)
-EXPOSE 8001 2024
+# Expose Gateway API port.
+EXPOSE 8001
 
 # Default command (can be overridden in docker-compose)
 CMD ["sh", "-c", "cd backend && PYTHONPATH=. uv run --no-sync uvicorn app.gateway.app:app --host 0.0.0.0 --port 8001"]

backend/Makefile

@@ -2,13 +2,16 @@ install:
 	uv sync
 
 dev:
-	PYTHONPATH=. uv run uvicorn app.gateway.app:app --host 0.0.0.0 --port 8001 --reload
+	PYTHONPATH=. PYTHONIOENCODING=utf-8 PYTHONUTF8=1 uv run uvicorn app.gateway.app:app --host 0.0.0.0 --port 8001 --reload
 
 gateway:
-	PYTHONPATH=. uv run uvicorn app.gateway.app:app --host 0.0.0.0 --port 8001
+	PYTHONPATH=. PYTHONIOENCODING=utf-8 PYTHONUTF8=1 uv run uvicorn app.gateway.app:app --host 0.0.0.0 --port 8001
 
 test:
-	PYTHONPATH=. uv run pytest tests/ -v
+	PYTHONPATH=. PYTHONIOENCODING=utf-8 PYTHONUTF8=1 uv run pytest tests/ -v
+
+test-blocking-io:
+	PYTHONPATH=. PYTHONIOENCODING=utf-8 PYTHONUTF8=1 uv run pytest tests/blocking_io -q --tb=short
 
 lint:
 	uvx ruff check .
@@ -16,3 +19,6 @@ lint:
 
 format:
 	uvx ruff check . --fix && uvx ruff format .
+
+detect-blocking-io:
+	@PYTHONPATH=. PYTHONIOENCODING=utf-8 PYTHONUTF8=1 uv run python ../scripts/detect_blocking_io_static.py --output ../.deer-flow/blocking-io-findings.json

backend/README.md

@@ -69,7 +69,7 @@ Middlewares execute in strict order, each handling a specific concern:
 Per-thread isolated execution with virtual path translation:
 
 - **Abstract interface**: `execute_command`, `read_file`, `write_file`, `list_dir`
-- **Providers**: `LocalSandboxProvider` (filesystem) and `AioSandboxProvider` (Docker, in community/)
+- **Providers**: `LocalSandboxProvider` (filesystem) and `AioSandboxProvider` (Docker, in community/). Async runtime paths use async sandbox lifecycle hooks so startup, readiness polling, and release do not block the event loop.
 - **Virtual paths**: `/mnt/user-data/{workspace,uploads,outputs}` → thread-specific physical directories
 - **Skills path**: `/mnt/skills` → `deer-flow/skills/` directory
 - **Skills loading**: Recursively discovers nested `SKILL.md` files under `skills/{public,custom}` and preserves nested container paths
@@ -362,6 +362,7 @@ make dev        # Run Gateway API + embedded agent runtime (port 8001)
 make gateway    # Run Gateway API without reload (port 8001)
 make lint       # Run linter (ruff)
 make format     # Format code (ruff)
+make detect-blocking-io  # Inventory blocking IO that may block the backend event loop
 ```
 
 ### Code Style
@@ -378,6 +379,18 @@ make format     # Format code (ruff)
 uv run pytest
 ```
 
+`make detect-blocking-io` statically scans backend business code for blocking
+IO that may run on the backend event loop and is not test-coverage-bound. It
+prints a concise summary for human review and writes complete JSON findings to
+`.deer-flow/blocking-io-findings.json` at the repository root (regardless of
+whether the target is invoked from the repo root or from `backend/`). JSON
+findings include both broad IO category and review-oriented fields such as
+`priority`, `location`, `blocking_call`, `event_loop_exposure`, `reason`, and
+`code`. `priority` is a deterministic review ordering from the operation type,
+not proof of a bug. Bare-name same-file calls are resolved by function name,
+so duplicate helper names in one file can conservatively over-report async
+reachability.
+
 ---
 
 ## Technology Stack

backend/app/channels/commands.py

@@ -18,3 +18,10 @@ KNOWN_CHANNEL_COMMANDS: frozenset[str] = frozenset(
         "/help",
     }
 )
+
+
+def is_known_channel_command(text: str) -> bool:
+    """Return whether text starts with a registered channel control command."""
+    if not text.startswith("/"):
+        return False
+    return text.split(maxsplit=1)[0].lower() in KNOWN_CHANNEL_COMMANDS

backend/app/channels/dingtalk.py

@@ -14,7 +14,7 @@ from typing import Any
 import httpx
 
 from app.channels.base import Channel
-from app.channels.commands import KNOWN_CHANNEL_COMMANDS
+from app.channels.commands import is_known_channel_command
 from app.channels.message_bus import InboundMessage, InboundMessageType, MessageBus, OutboundMessage, ResolvedAttachment
 
 logger = logging.getLogger(__name__)
@@ -59,9 +59,7 @@ def _normalize_allowed_users(allowed_users: Any) -> set[str]:
 
 
 def _is_dingtalk_command(text: str) -> bool:
-    if not text.startswith("/"):
-        return False
-    return text.split(maxsplit=1)[0].lower() in KNOWN_CHANNEL_COMMANDS
+    return is_known_channel_command(text)
 
 
 def _extract_text_from_rich_text(rich_text_list: list) -> str:

backend/app/channels/discord.py

@@ -10,6 +10,7 @@ from pathlib import Path
 from typing import Any
 
 from app.channels.base import Channel
+from app.channels.commands import is_known_channel_command
 from app.channels.message_bus import InboundMessageType, MessageBus, OutboundMessage, ResolvedAttachment
 
 logger = logging.getLogger(__name__)
@@ -300,7 +301,7 @@ class DiscordChannel(Channel):
 
             # If this is a known active thread, process normally
             if thread_id in self._active_thread_ids:
-                msg_type = InboundMessageType.COMMAND if text.startswith("/") else InboundMessageType.CHAT
+                msg_type = InboundMessageType.COMMAND if is_known_channel_command(text) else InboundMessageType.CHAT
                 inbound = self._make_inbound(
                     chat_id=chat_id,
                     user_id=str(message.author.id),
@@ -407,7 +408,7 @@ class DiscordChannel(Channel):
             chat_id = channel_id
             typing_target = message.channel  # Type into the channel
 
-        msg_type = InboundMessageType.COMMAND if text.startswith("/") else InboundMessageType.CHAT
+        msg_type = InboundMessageType.COMMAND if is_known_channel_command(text) else InboundMessageType.CHAT
         inbound = self._make_inbound(
             chat_id=chat_id,
             user_id=str(message.author.id),

backend/app/channels/feishu.py

@@ -7,22 +7,30 @@ import json
 import logging
 import re
 import threading
+import time
 from typing import Any, Literal
 
 from app.channels.base import Channel
-from app.channels.commands import KNOWN_CHANNEL_COMMANDS
-from app.channels.message_bus import InboundMessage, InboundMessageType, MessageBus, OutboundMessage, ResolvedAttachment
+from app.channels.commands import is_known_channel_command
+from app.channels.message_bus import (
+    PENDING_CLARIFICATION_METADATA_KEY,
+    RESOLVED_FROM_PENDING_CLARIFICATION_METADATA_KEY,
+    InboundMessage,
+    InboundMessageType,
+    MessageBus,
+    OutboundMessage,
+    ResolvedAttachment,
+)
 from deerflow.config.paths import VIRTUAL_PATH_PREFIX, get_paths
 from deerflow.runtime.user_context import get_effective_user_id
 from deerflow.sandbox.sandbox_provider import get_sandbox_provider
 
 logger = logging.getLogger(__name__)
+PENDING_CLARIFICATION_TTL_SECONDS = 30 * 60
 
 
 def _is_feishu_command(text: str) -> bool:
-    if not text.startswith("/"):
-        return False
-    return text.split(maxsplit=1)[0].lower() in KNOWN_CHANNEL_COMMANDS
+    return is_known_channel_command(text)
 
 
 class FeishuChannel(Channel):
@@ -56,6 +64,7 @@ class FeishuChannel(Channel):
         self._background_tasks: set[asyncio.Task] = set()
         self._running_card_ids: dict[str, str] = {}
         self._running_card_tasks: dict[str, asyncio.Task] = {}
+        self._pending_clarifications: dict[tuple[str, str], list[dict[str, Any]]] = {}
         self._CreateFileRequest = None
         self._CreateFileRequestBody = None
         self._CreateImageRequest = None
@@ -63,6 +72,16 @@ class FeishuChannel(Channel):
         self._GetMessageResourceRequest = None
         self._thread_lock = threading.Lock()
 
+    @staticmethod
+    def _non_empty_str(value: Any) -> str | None:
+        if isinstance(value, str) and value.strip():
+            return value.strip()
+        return None
+
+    @staticmethod
+    def _pending_key(chat_id: str, user_id: str) -> tuple[str, str]:
+        return (chat_id, user_id)
+
     @property
     def supports_streaming(self) -> bool:
         return True
@@ -531,18 +550,25 @@ class FeishuChannel(Channel):
                         "[Feishu] failed to patch running card %s, falling back to final reply",
                         running_card_id,
                     )
-                    await self._reply_card(source_message_id, msg.text)
+                    fallback_card_id = await self._reply_card(source_message_id, msg.text)
+                    self._remember_thread_mapping(msg, source_message_id, fallback_card_id)
+                    self._remember_pending_clarification(msg, fallback_card_id)
                 else:
+                    self._remember_thread_mapping(msg, source_message_id, running_card_id)
+                    self._remember_pending_clarification(msg, running_card_id)
                     logger.info("[Feishu] running card updated: source=%s card=%s", source_message_id, running_card_id)
             elif msg.is_final:
-                await self._reply_card(source_message_id, msg.text)
+                final_card_id = await self._reply_card(source_message_id, msg.text)
+                self._remember_thread_mapping(msg, source_message_id, final_card_id)
+                self._remember_pending_clarification(msg, final_card_id)
             elif awaited_running_card_task:
                 logger.warning(
                     "[Feishu] running card task finished without message_id for source=%s, skipping duplicate non-final creation",
                     source_message_id,
                 )
             else:
-                await self._ensure_running_card(source_message_id, msg.text)
+                created_card_id = await self._ensure_running_card(source_message_id, msg.text)
+                self._remember_thread_mapping(msg, source_message_id, created_card_id)
 
             if msg.is_final:
                 self._running_card_ids.pop(source_message_id, None)
@@ -553,6 +579,129 @@ class FeishuChannel(Channel):
 
     # -- internal ----------------------------------------------------------
 
+    def _remember_thread_mapping(self, msg: OutboundMessage, *topic_ids: str | None) -> None:
+        store = self.config.get("channel_store")
+        if store is None or not msg.thread_id:
+            return
+
+        metadata_topic_ids = [
+            msg.metadata.get("message_id"),
+            msg.metadata.get("root_id"),
+            msg.metadata.get("parent_id"),
+            msg.metadata.get("thread_id"),
+            msg.metadata.get("topic_id"),
+        ]
+        user_id = ""
+        raw_user_id = msg.metadata.get("user_id")
+        if isinstance(raw_user_id, str):
+            user_id = raw_user_id
+
+        seen: set[str] = set()
+        for topic_id in [*topic_ids, *metadata_topic_ids]:
+            topic_id = self._non_empty_str(topic_id)
+            if not topic_id or topic_id in seen:
+                continue
+            seen.add(topic_id)
+            try:
+                store.set_thread_id(
+                    self.name,
+                    msg.chat_id,
+                    msg.thread_id,
+                    topic_id=topic_id,
+                    user_id=user_id,
+                )
+            except Exception:
+                logger.exception("[Feishu] failed to remember thread mapping for topic_id=%s", topic_id)
+
+    def _remember_pending_clarification(self, msg: OutboundMessage, card_message_id: str | None) -> None:
+        if not msg.is_final or msg.metadata.get(PENDING_CLARIFICATION_METADATA_KEY) is not True:
+            return
+
+        user_id = self._non_empty_str(msg.metadata.get("user_id"))
+        topic_id = self._non_empty_str(msg.metadata.get("topic_id"))
+        source_message_id = self._non_empty_str(msg.thread_ts) or self._non_empty_str(msg.metadata.get("message_id"))
+        if not (user_id and topic_id and msg.thread_id and source_message_id and card_message_id):
+            return
+
+        key = self._pending_key(msg.chat_id, user_id)
+        pending = {
+            "thread_id": msg.thread_id,
+            "topic_id": topic_id,
+            "source_message_id": source_message_id,
+            "card_message_id": card_message_id,
+            "created_at": time.time(),
+        }
+        with self._thread_lock:
+            # Plain-message clarification continuity is a short-lived in-memory
+            # hint; explicit Feishu replies are still covered by persisted
+            # message-id mappings.
+            self._pending_clarifications.setdefault(key, []).append(pending)
+        logger.info(
+            "[Feishu] pending clarification remembered: chat_id=%s user_id=%s topic_id=%s thread_id=%s",
+            msg.chat_id,
+            user_id,
+            topic_id,
+            msg.thread_id,
+        )
+
+    def _consume_pending_clarification(self, chat_id: str, user_id: str) -> dict[str, Any] | None:
+        key = self._pending_key(chat_id, user_id)
+        with self._thread_lock:
+            pending_items = self._pending_clarifications.get(key)
+            if not pending_items:
+                return None
+
+            now = time.time()
+            while pending_items:
+                pending = pending_items.pop(0)
+                created_at = pending.get("created_at")
+                if isinstance(created_at, (int, float)) and now - created_at <= PENDING_CLARIFICATION_TTL_SECONDS:
+                    if pending_items:
+                        self._pending_clarifications[key] = pending_items
+                    else:
+                        self._pending_clarifications.pop(key, None)
+                    return pending
+                logger.info("[Feishu] pending clarification expired: chat_id=%s user_id=%s", chat_id, user_id)
+
+            self._pending_clarifications.pop(key, None)
+            return None
+
+    def _ensure_pending_thread_mapping(self, chat_id: str, user_id: str, pending: dict[str, Any]) -> None:
+        store = self.config.get("channel_store")
+        topic_id = self._non_empty_str(pending.get("topic_id"))
+        thread_id = self._non_empty_str(pending.get("thread_id"))
+        if store is None or not topic_id or not thread_id:
+            return
+        try:
+            store.set_thread_id(self.name, chat_id, thread_id, topic_id=topic_id, user_id=user_id)
+        except Exception:
+            logger.exception("[Feishu] failed to restore pending clarification mapping for topic_id=%s", topic_id)
+
+    def _resolve_topic_id(
+        self,
+        chat_id: str,
+        msg_id: str,
+        *,
+        root_id: str | None,
+        parent_id: str | None,
+        thread_id: str | None,
+    ) -> tuple[str, bool]:
+        store = self.config.get("channel_store")
+        candidates = [root_id, parent_id, thread_id]
+
+        if store is not None:
+            for candidate in candidates:
+                candidate = self._non_empty_str(candidate)
+                if not candidate:
+                    continue
+                try:
+                    if store.get_thread_id(self.name, chat_id, topic_id=candidate):
+                        return candidate, True
+                except Exception:
+                    logger.exception("[Feishu] failed to resolve stored topic mapping for topic_id=%s", candidate)
+
+        return root_id or msg_id, False
+
     @staticmethod
     def _log_future_error(fut, name: str, msg_id: str) -> None:
         """Callback for run_coroutine_threadsafe futures to surface errors."""
@@ -593,7 +742,9 @@ class FeishuChannel(Channel):
 
             # root_id is set when the message is a reply within a Feishu thread.
             # Use it as topic_id so all replies share the same DeerFlow thread.
-            root_id = getattr(message, "root_id", None) or None
+            root_id = self._non_empty_str(getattr(message, "root_id", None))
+            parent_id = self._non_empty_str(getattr(message, "parent_id", None))
+            feishu_thread_id = self._non_empty_str(getattr(message, "thread_id", None))
 
             # Parse message content
             content = json.loads(message.content)
@@ -654,10 +805,12 @@ class FeishuChannel(Channel):
             text = text.strip()
 
             logger.info(
-                "[Feishu] parsed message: chat_id=%s, msg_id=%s, root_id=%s, sender=%s, text=%r",
+                "[Feishu] parsed message: chat_id=%s, msg_id=%s, root_id=%s, parent_id=%s, thread_id=%s, sender=%s, text=%r",
                 chat_id,
                 msg_id,
                 root_id,
+                parent_id,
+                feishu_thread_id,
                 sender_id,
                 text[:100] if text else "",
             )
@@ -673,8 +826,24 @@ class FeishuChannel(Channel):
             else:
                 msg_type = InboundMessageType.CHAT
 
-            # topic_id: use root_id for replies (same topic), msg_id for new messages (new topic)
-            topic_id = root_id or msg_id
+            # Prefer any platform message id that already maps to a DeerFlow
+            # thread. This keeps replies to bot clarification cards in the
+            # original conversation even when Feishu reports the card as root.
+            topic_id, resolved_from_stored_mapping = self._resolve_topic_id(
+                chat_id,
+                msg_id,
+                root_id=root_id,
+                parent_id=parent_id,
+                thread_id=feishu_thread_id,
+            )
+            resolved_from_pending = False
+            if msg_type == InboundMessageType.CHAT and not resolved_from_stored_mapping:
+                pending = self._consume_pending_clarification(chat_id, sender_id)
+                pending_topic_id = self._non_empty_str(pending.get("topic_id")) if pending else None
+                if pending_topic_id:
+                    topic_id = pending_topic_id
+                    self._ensure_pending_thread_mapping(chat_id, sender_id, pending)
+                    resolved_from_pending = True
 
             inbound = self._make_inbound(
                 chat_id=chat_id,
@@ -683,7 +852,15 @@ class FeishuChannel(Channel):
                 msg_type=msg_type,
                 thread_ts=msg_id,
                 files=files_list,
-                metadata={"message_id": msg_id, "root_id": root_id},
+                metadata={
+                    "message_id": msg_id,
+                    "root_id": root_id,
+                    "parent_id": parent_id,
+                    "thread_id": feishu_thread_id,
+                    "topic_id": topic_id,
+                    "user_id": sender_id,
+                    RESOLVED_FROM_PENDING_CLARIFICATION_METADATA_KEY: resolved_from_pending,
+                },
             )
             inbound.topic_id = topic_id
 

backend/app/channels/manager.py

@@ -8,6 +8,7 @@ import mimetypes
 import re
 import time
 from collections.abc import Awaitable, Callable, Mapping
+from dataclasses import dataclass
 from pathlib import Path
 from typing import Any
 
@@ -15,11 +16,24 @@ import httpx
 from langgraph_sdk.errors import ConflictError
 
 from app.channels.commands import KNOWN_CHANNEL_COMMANDS
-from app.channels.message_bus import InboundMessage, InboundMessageType, MessageBus, OutboundMessage, ResolvedAttachment
+from app.channels.message_bus import (
+    PENDING_CLARIFICATION_METADATA_KEY,
+    InboundMessage,
+    InboundMessageType,
+    MessageBus,
+    OutboundMessage,
+    ResolvedAttachment,
+)
 from app.channels.store import ChannelStore
 from app.gateway.csrf_middleware import CSRF_COOKIE_NAME, CSRF_HEADER_NAME, generate_csrf_token
 from app.gateway.internal_auth import create_internal_auth_headers
+from deerflow.config.agents_config import load_agent_config
+from deerflow.config.paths import make_safe_user_id
 from deerflow.runtime.user_context import get_effective_user_id
+from deerflow.skills.slash import parse_slash_skill_reference
+from deerflow.skills.storage import get_or_new_skill_storage
+from deerflow.skills.storage.skill_storage import SkillStorage
+from deerflow.utils.messages import ORIGINAL_USER_CONTENT_KEY
 
 logger = logging.getLogger(__name__)
 
@@ -116,6 +130,16 @@ class InvalidChannelSessionConfigError(ValueError):
     """Raised when IM channel session overrides contain invalid agent config."""
 
 
+class SlashSkillCommandResolutionError(RuntimeError):
+    """Raised when IM slash-skill command resolution cannot complete safely."""
+
+
+@dataclass(frozen=True, slots=True)
+class _SlashSkillCommandResolution:
+    route_to_chat: bool = False
+    failure_message: str | None = None
+
+
 def _is_thread_busy_error(exc: BaseException | None) -> bool:
     if exc is None:
         return False
@@ -146,13 +170,6 @@ def _normalize_custom_agent_name(raw_value: str) -> str:
     return normalized
 
 
-def _strip_loop_warning_text(text: str) -> str:
-    """Remove middleware-authored loop warning lines from display text."""
-    if "[LOOP DETECTED]" not in text:
-        return text
-    return "\n".join(line for line in text.splitlines() if "[LOOP DETECTED]" not in line).strip()
-
-
 def _extract_response_text(result: dict | list) -> str:
     """Extract the last AI message text from a LangGraph runs.wait result.
 
@@ -162,7 +179,6 @@ def _extract_response_text(result: dict | list) -> str:
     Handles special cases:
     - Regular AI text responses
     - Clarification interrupts (``ask_clarification`` tool messages)
-    - Strips loop-detection warnings attached to tool-call AI messages
     """
     if isinstance(result, list):
         messages = result
@@ -181,6 +197,8 @@ def _extract_response_text(result: dict | list) -> str:
 
         # Stop at the last human message — anything before it is a previous turn
         if msg_type == "human":
+            if _is_hidden_human_control_message(msg):
+                continue
             break
 
         # Check for tool messages from ask_clarification (interrupt case)
@@ -192,12 +210,7 @@ def _extract_response_text(result: dict | list) -> str:
         # Regular AI message with text content
         if msg_type == "ai":
             content = msg.get("content", "")
-            has_tool_calls = bool(msg.get("tool_calls"))
             if isinstance(content, str) and content:
-                if has_tool_calls:
-                    content = _strip_loop_warning_text(content)
-                    if not content:
-                        continue
                 return content
             # content can be a list of content blocks
             if isinstance(content, list):
@@ -208,13 +221,59 @@ def _extract_response_text(result: dict | list) -> str:
                     elif isinstance(block, str):
                         parts.append(block)
                 text = "".join(parts)
-                if has_tool_calls:
-                    text = _strip_loop_warning_text(text)
                 if text:
                     return text
     return ""
 
 
+def _messages_from_result(result: dict | list) -> list[Any]:
+    if isinstance(result, list):
+        return result
+    if isinstance(result, dict):
+        messages = result.get("messages", [])
+        if isinstance(messages, list):
+            return messages
+    return []
+
+
+def _current_turn_messages(result: dict | list) -> list[dict[str, Any]]:
+    messages = _messages_from_result(result)
+    current_turn: list[dict[str, Any]] = []
+    for msg in reversed(messages):
+        if not isinstance(msg, dict):
+            continue
+        if msg.get("type") == "human":
+            break
+        current_turn.append(msg)
+    current_turn.reverse()
+    return current_turn
+
+
+def _has_current_turn_clarification(result: dict | list) -> bool:
+    """Return True only when the current turn's final result is clarification."""
+    for msg in reversed(_current_turn_messages(result)):
+        msg_type = msg.get("type")
+        if msg_type == "tool":
+            return msg.get("name") == "ask_clarification"
+        if msg_type == "ai":
+            content = msg.get("content")
+            if isinstance(content, str):
+                if content:
+                    return False
+            elif content:
+                return False
+            if msg.get("tool_calls"):
+                return False
+    return False
+
+
+def _response_metadata(base_metadata: dict[str, Any], *, pending_clarification: bool = False) -> dict[str, Any]:
+    metadata = _slim_metadata(base_metadata)
+    if pending_clarification:
+        metadata[PENDING_CLARIFICATION_METADATA_KEY] = True
+    return metadata
+
+
 def _extract_text_content(content: Any) -> str:
     """Extract text from a streaming payload content field."""
     if isinstance(content, str):
@@ -328,6 +387,8 @@ def _extract_artifacts(result: dict | list) -> list[str]:
             continue
         # Stop at the last human message — anything before it is a previous turn
         if msg.get("type") == "human":
+            if _is_hidden_human_control_message(msg):
+                continue
             break
         # Look for AI messages with present_files tool calls
         if msg.get("type") == "ai":
@@ -340,6 +401,18 @@ def _extract_artifacts(result: dict | list) -> list[str]:
     return artifacts
 
 
+def _is_hidden_human_control_message(msg: Mapping[str, Any]) -> bool:
+    """Return whether a human message is an internal control message hidden from UI."""
+    if msg.get("type") != "human":
+        return False
+
+    additional_kwargs = msg.get("additional_kwargs")
+    if not isinstance(additional_kwargs, Mapping):
+        return False
+
+    return additional_kwargs.get("hide_from_ui") is True
+
+
 def _format_artifact_text(artifacts: list[str]) -> str:
     """Format artifact paths into a human-readable text block listing filenames."""
     import posixpath
@@ -353,6 +426,46 @@ def _format_artifact_text(artifacts: list[str]) -> str:
 _OUTPUTS_VIRTUAL_PREFIX = "/mnt/user-data/outputs/"
 
 
+def _unknown_command_reply(command: str | None = None) -> str:
+    available = " | ".join(sorted(KNOWN_CHANNEL_COMMANDS))
+    if command:
+        return f"Unknown command: /{command}. Available commands: {available}"
+    return f"Unknown command. Available commands: {available}"
+
+
+def _human_input_message(content: str, *, original_content: str | None = None) -> dict[str, Any]:
+    message: dict[str, Any] = {"role": "human", "content": content}
+    if original_content is not None and original_content != content:
+        message["additional_kwargs"] = {ORIGINAL_USER_CONTENT_KEY: original_content}
+    return message
+
+
+def _resolve_slash_skill_command(
+    text: str,
+    available_skills: set[str] | None = None,
+    storage: SkillStorage | Callable[[], SkillStorage] | None = None,
+) -> _SlashSkillCommandResolution | None:
+    reference = parse_slash_skill_reference(text)
+    if reference is None:
+        return None
+    try:
+        resolved_storage = storage() if callable(storage) else storage or get_or_new_skill_storage()
+        skills = resolved_storage.load_skills(enabled_only=False)
+
+        skill = next((candidate for candidate in skills if candidate.name == reference.name), None)
+        if skill is None:
+            return None
+        if not skill.enabled:
+            return _SlashSkillCommandResolution(failure_message=f"Skill `/{reference.name}` is installed but disabled. Enable it before using slash activation.")
+        if available_skills is not None and reference.name not in available_skills:
+            return _SlashSkillCommandResolution(failure_message=f"Skill `/{reference.name}` is not available for this agent.")
+
+        return _SlashSkillCommandResolution(route_to_chat=True)
+    except Exception as exc:
+        logger.exception("[Manager] failed to resolve slash skill command")
+        raise SlashSkillCommandResolutionError("Failed to resolve slash skill command. Please check the skill configuration.") from exc
+
+
 def _resolve_attachments(thread_id: str, artifacts: list[str]) -> list[ResolvedAttachment]:
     """Resolve virtual artifact paths to host filesystem paths with metadata.
 
@@ -567,6 +680,7 @@ class ChannelManager:
         self._default_session = _as_dict(default_session)
         self._channel_sessions = dict(channel_sessions or {})
         self._client = None  # lazy init — langgraph_sdk async client
+        self._skill_storage: SkillStorage | None = None
         self._csrf_token = generate_csrf_token()
         self._semaphore: asyncio.Semaphore | None = None
         self._running = False
@@ -614,12 +728,20 @@ class ChannelManager:
         configurable["checkpoint_ns"] = ""
         configurable["thread_id"] = thread_id
 
+        # ``user_id`` drives user-scoped filesystem buckets that only accept
+        # ``[A-Za-z0-9_-]``, so normalize the channel id and keep the raw value
+        # under ``channel_user_id`` for platform-facing lookups.
+        run_context_identity: dict[str, Any] = {"thread_id": thread_id}
+        if msg.user_id:
+            run_context_identity["user_id"] = make_safe_user_id(msg.user_id)
+            run_context_identity["channel_user_id"] = msg.user_id
+
         run_context = _merge_dicts(
             DEFAULT_RUN_CONTEXT,
             self._default_session.get("context"),
             channel_layer.get("context"),
             user_layer.get("context"),
-            {"thread_id": thread_id},
+            run_context_identity,
         )
 
         # Custom agents are implemented as lead_agent + agent_name context.
@@ -631,6 +753,21 @@ class ChannelManager:
 
         return assistant_id, run_config, run_context
 
+    def _resolve_available_skill_names(self, msg: InboundMessage) -> set[str] | None:
+        thread_id = self.store.get_thread_id(msg.channel_name, msg.chat_id, topic_id=msg.topic_id) or ""
+        _, _, run_context = self._resolve_run_params(msg, thread_id)
+        if run_context.get("is_bootstrap"):
+            return {"bootstrap"}
+
+        agent_name = run_context.get("agent_name")
+        if not isinstance(agent_name, str) or not agent_name.strip():
+            return None
+
+        agent_config = load_agent_config(_normalize_custom_agent_name(agent_name))
+        if agent_config and agent_config.skills is not None:
+            return set(agent_config.skills)
+        return None
+
     # -- LangGraph SDK client (lazy) ----------------------------------------
 
     def _get_client(self):
@@ -648,6 +785,11 @@ class ChannelManager:
             )
         return self._client
 
+    def _get_skill_storage(self) -> SkillStorage:
+        if self._skill_storage is None:
+            self._skill_storage = get_or_new_skill_storage()
+        return self._skill_storage
+
     # -- lifecycle ---------------------------------------------------------
 
     async def start(self) -> None:
@@ -717,6 +859,14 @@ class ChannelManager:
                     exc,
                 )
                 await self._send_error(msg, str(exc))
+            except SlashSkillCommandResolutionError as exc:
+                logger.warning(
+                    "Slash skill command resolution failed for %s (chat=%s): %s",
+                    msg.channel_name,
+                    msg.chat_id,
+                    exc,
+                )
+                await self._send_error(msg, str(exc))
             except Exception:
                 logger.exception(
                     "Error handling message from %s (chat=%s)",
@@ -771,9 +921,11 @@ class ChannelManager:
         if extra_context:
             run_context.update(extra_context)
 
+        original_text = msg.text
         uploaded = await _ingest_inbound_files(thread_id, msg)
         if uploaded:
             msg.text = f"{_format_uploaded_files_block(uploaded)}\n\n{msg.text}".strip()
+        human_message = _human_input_message(msg.text, original_content=original_text)
 
         if self._channel_supports_streaming(msg.channel_name):
             await self._handle_streaming_chat(
@@ -783,6 +935,7 @@ class ChannelManager:
                 assistant_id,
                 run_config,
                 run_context,
+                human_message,
             )
             return
 
@@ -791,7 +944,7 @@ class ChannelManager:
             result = await client.runs.wait(
                 thread_id,
                 assistant_id,
-                input={"messages": [{"role": "human", "content": msg.text}]},
+                input={"messages": [human_message]},
                 config=run_config,
                 context=run_context,
                 multitask_strategy="reject",
@@ -805,6 +958,7 @@ class ChannelManager:
                 raise
 
         response_text = _extract_response_text(result)
+        pending_clarification = _has_current_turn_clarification(result)
         artifacts = _extract_artifacts(result)
 
         logger.info(
@@ -830,7 +984,7 @@ class ChannelManager:
             artifacts=artifacts,
             attachments=attachments,
             thread_ts=msg.thread_ts,
-            metadata=_slim_metadata(msg.metadata),
+            metadata=_response_metadata(msg.metadata, pending_clarification=pending_clarification),
         )
         logger.info("[Manager] publishing outbound message to bus: channel=%s, chat_id=%s", msg.channel_name, msg.chat_id)
         await self.bus.publish_outbound(outbound)
@@ -843,6 +997,7 @@ class ChannelManager:
         assistant_id: str,
         run_config: dict[str, Any],
         run_context: dict[str, Any],
+        human_message: dict[str, Any],
     ) -> None:
         logger.info("[Manager] invoking runs.stream(thread_id=%s, text=%r)", thread_id, msg.text[:100])
 
@@ -858,7 +1013,7 @@ class ChannelManager:
             async for chunk in client.runs.stream(
                 thread_id,
                 assistant_id,
-                input={"messages": [{"role": "human", "content": msg.text}]},
+                input={"messages": [human_message]},
                 config=run_config,
                 context=run_context,
                 stream_mode=["messages-tuple", "values"],
@@ -892,7 +1047,7 @@ class ChannelManager:
                         text=latest_text,
                         is_final=False,
                         thread_ts=msg.thread_ts,
-                        metadata=_slim_metadata(msg.metadata),
+                        metadata=_response_metadata(msg.metadata),
                     )
                 )
                 last_published_text = latest_text
@@ -906,6 +1061,7 @@ class ChannelManager:
         finally:
             result = last_values if last_values is not None else {"messages": [{"type": "ai", "content": latest_text}]}
             response_text = _extract_response_text(result)
+            pending_clarification = _has_current_turn_clarification(result)
             artifacts = _extract_artifacts(result)
             response_text, attachments = _prepare_artifact_delivery(thread_id, response_text, artifacts)
 
@@ -937,18 +1093,27 @@ class ChannelManager:
                     attachments=attachments,
                     is_final=True,
                     thread_ts=msg.thread_ts,
-                    metadata=_slim_metadata(msg.metadata),
+                    metadata=_response_metadata(msg.metadata, pending_clarification=pending_clarification),
                 )
             )
 
     # -- command handling --------------------------------------------------
 
     async def _handle_command(self, msg: InboundMessage) -> None:
-        text = msg.text.strip()
+        raw_text = msg.text
+        text = raw_text.strip()
         parts = text.split(maxsplit=1)
-        command = parts[0].lower().lstrip("/")
+        reply: str | None = None
+        if not parts:
+            command = None
+            reply = _unknown_command_reply()
+        else:
+            command = parts[0].lower().removeprefix("/")
 
-        if command == "bootstrap":
+        if reply is None and not raw_text.startswith("/"):
+            reply = _unknown_command_reply(command)
+
+        if reply is None and command == "bootstrap":
             from dataclasses import replace as _dc_replace
 
             chat_text = parts[1] if len(parts) > 1 else "Initialize workspace"
@@ -956,7 +1121,7 @@ class ChannelManager:
             await self._handle_chat(chat_msg, extra_context={"is_bootstrap": True})
             return
 
-        if command == "new":
+        if reply is None and command == "new":
             # Create a new thread through Gateway
             client = self._get_client()
             thread = await client.threads.create()
@@ -969,14 +1134,14 @@ class ChannelManager:
                 user_id=msg.user_id,
             )
             reply = "New conversation started."
-        elif command == "status":
+        elif reply is None and command == "status":
             thread_id = self.store.get_thread_id(msg.channel_name, msg.chat_id, topic_id=msg.topic_id)
             reply = f"Active thread: {thread_id}" if thread_id else "No active conversation."
-        elif command == "models":
+        elif reply is None and command == "models":
             reply = await self._fetch_gateway("/api/models", "models")
-        elif command == "memory":
+        elif reply is None and command == "memory":
             reply = await self._fetch_gateway("/api/memory", "memory")
-        elif command == "help":
+        elif reply is None and command == "help":
             reply = (
                 "Available commands:\n"
                 "/bootstrap — Start a bootstrap session (enables agent setup)\n"
@@ -984,16 +1149,32 @@ class ChannelManager:
                 "/status — Show current thread info\n"
                 "/models — List available models\n"
                 "/memory — Show memory status\n"
+                "/<skill-name> <task> — Activate an enabled skill for one turn\n"
                 "/help — Show this help"
             )
-        else:
-            available = " | ".join(sorted(KNOWN_CHANNEL_COMMANDS))
-            reply = f"Unknown command: /{command}. Available commands: {available}"
+        elif reply is None:
+            slash_resolution = await asyncio.to_thread(
+                lambda: _resolve_slash_skill_command(
+                    raw_text,
+                    self._resolve_available_skill_names(msg),
+                    self._get_skill_storage,
+                )
+            )
+            if slash_resolution and slash_resolution.failure_message:
+                reply = slash_resolution.failure_message
+            elif slash_resolution and slash_resolution.route_to_chat:
+                from dataclasses import replace as _dc_replace
+
+                chat_msg = _dc_replace(msg, msg_type=InboundMessageType.CHAT)
+                await self._handle_chat(chat_msg)
+                return
+            else:
+                reply = _unknown_command_reply(command)
 
         outbound = OutboundMessage(
             channel_name=msg.channel_name,
             chat_id=msg.chat_id,
-            thread_id=self.store.get_thread_id(msg.channel_name, msg.chat_id) or "",
+            thread_id=self.store.get_thread_id(msg.channel_name, msg.chat_id, topic_id=msg.topic_id) or "",
             text=reply,
             thread_ts=msg.thread_ts,
             metadata=_slim_metadata(msg.metadata),
@@ -1031,7 +1212,7 @@ class ChannelManager:
         outbound = OutboundMessage(
             channel_name=msg.channel_name,
             chat_id=msg.chat_id,
-            thread_id=self.store.get_thread_id(msg.channel_name, msg.chat_id) or "",
+            thread_id=self.store.get_thread_id(msg.channel_name, msg.chat_id, topic_id=msg.topic_id) or "",
             text=error_text,
             thread_ts=msg.thread_ts,
             metadata=_slim_metadata(msg.metadata),

backend/app/channels/message_bus.py

@@ -13,6 +13,9 @@ from typing import Any
 
 logger = logging.getLogger(__name__)
 
+PENDING_CLARIFICATION_METADATA_KEY = "pending_clarification"
+RESOLVED_FROM_PENDING_CLARIFICATION_METADATA_KEY = "resolved_from_pending_clarification"
+
 
 # ---------------------------------------------------------------------------
 # Message types

backend/app/channels/slack.py

@@ -9,6 +9,7 @@ from typing import Any
 from markdown_to_mrkdwn import SlackMarkdownConverter
 
 from app.channels.base import Channel
+from app.channels.commands import is_known_channel_command
 from app.channels.message_bus import InboundMessageType, MessageBus, OutboundMessage, ResolvedAttachment
 
 logger = logging.getLogger(__name__)
@@ -32,6 +33,20 @@ def _normalize_allowed_users(allowed_users: Any) -> set[str]:
     return {str(user_id) for user_id in values if str(user_id)}
 
 
+def _strip_leading_slack_bot_mention(text: str, bot_user_id: str | None) -> str:
+    if not bot_user_id:
+        return text
+    if not text.startswith("<@"):
+        return text
+    end = text.find(">")
+    if end <= 2:
+        return text
+    mentioned_user_id = text[2:end].split("|", 1)[0].lstrip("!")
+    if mentioned_user_id != bot_user_id:
+        return text
+    return text[end + 1 :].lstrip()
+
+
 class SlackChannel(Channel):
     """Slack IM channel using Socket Mode (WebSocket, no public IP).
 
@@ -49,6 +64,8 @@ class SlackChannel(Channel):
         self._web_client = None
         self._loop: asyncio.AbstractEventLoop | None = None
         self._allowed_users = _normalize_allowed_users(config.get("allowed_users", []))
+        configured_bot_user_id = config.get("bot_user_id")
+        self._bot_user_id = str(configured_bot_user_id).lstrip("@") if configured_bot_user_id else None
 
     async def start(self) -> None:
         if self._running:
@@ -72,6 +89,17 @@ class SlackChannel(Channel):
             return
 
         self._web_client = WebClient(token=bot_token)
+        if self._bot_user_id is None:
+            try:
+                auth_info = await asyncio.to_thread(self._web_client.auth_test)
+                user_id = auth_info.get("user_id") if isinstance(auth_info, dict) else None
+                if user_id is None:
+                    auth_get = getattr(auth_info, "get", None)
+                    user_id = auth_get("user_id") if callable(auth_get) else None
+                if isinstance(user_id, str) and user_id:
+                    self._bot_user_id = user_id
+            except Exception:
+                logger.warning("[Slack] failed to resolve bot user id; app mention text may include the bot mention", exc_info=True)
         self._socket_client = SocketModeClient(
             app_token=app_token,
             web_client=self._web_client,
@@ -210,6 +238,12 @@ class SlackChannel(Channel):
             if event_type != "events_api":
                 return
 
+            if self._bot_user_id is None:
+                authorization = next((item for item in req.payload.get("authorizations", []) if isinstance(item, dict)), None)
+                user_id = authorization.get("user_id") if authorization else None
+                if isinstance(user_id, str) and user_id:
+                    self._bot_user_id = user_id
+
             event = req.payload.get("event", {})
             etype = event.get("type", "")
 
@@ -233,13 +267,15 @@ class SlackChannel(Channel):
             return
 
         text = event.get("text", "").strip()
+        if event.get("type") == "app_mention":
+            text = _strip_leading_slack_bot_mention(text, self._bot_user_id)
         if not text:
             return
 
         channel_id = event.get("channel", "")
         thread_ts = event.get("thread_ts") or event.get("ts", "")
 
-        if text.startswith("/"):
+        if is_known_channel_command(text):
             msg_type = InboundMessageType.COMMAND
         else:
             msg_type = InboundMessageType.CHAT

backend/app/channels/telegram.py

@@ -60,12 +60,17 @@ class TelegramChannel(Channel):
 
         # Command handlers
         app.add_handler(CommandHandler("start", self._cmd_start))
+        app.add_handler(CommandHandler("bootstrap", self._cmd_generic))
         app.add_handler(CommandHandler("new", self._cmd_generic))
         app.add_handler(CommandHandler("status", self._cmd_generic))
         app.add_handler(CommandHandler("models", self._cmd_generic))
         app.add_handler(CommandHandler("memory", self._cmd_generic))
         app.add_handler(CommandHandler("help", self._cmd_generic))
 
+        # Slash skill commands are dynamic and cannot all be pre-registered
+        # with Telegram, so route unknown slash commands through chat handling.
+        app.add_handler(MessageHandler(filters.TEXT & filters.COMMAND, self._on_text))
+
         # General message handler
         app.add_handler(MessageHandler(filters.TEXT & ~filters.COMMAND, self._on_text))
 
@@ -228,6 +233,33 @@ class TelegramChannel(Channel):
             return True
         return user_id in self._allowed_users
 
+    def _get_bot_username(self, context) -> str | None:
+        bot = getattr(context, "bot", None)
+        username = getattr(bot, "username", None)
+        if not username and self._application is not None:
+            username = getattr(getattr(self._application, "bot", None), "username", None)
+        return str(username) if username else None
+
+    @staticmethod
+    def _strip_bot_username_from_leading_command(text: str, bot_username: str | None) -> str:
+        username = (bot_username or "").lstrip("@").lower()
+        if not username or not text.startswith("/"):
+            return text
+
+        parts = text.split(maxsplit=1)
+        command_token = parts[0]
+        if "@" not in command_token:
+            return text
+
+        command_name, addressed_username = command_token[1:].rsplit("@", 1)
+        if not command_name or addressed_username.lower() != username:
+            return text
+
+        normalized = f"/{command_name}"
+        if len(parts) > 1:
+            normalized = f"{normalized} {parts[1]}"
+        return normalized
+
     async def _cmd_start(self, update, context) -> None:
         """Handle /start command."""
         if not self._check_user(update.effective_user.id):
@@ -243,7 +275,7 @@ class TelegramChannel(Channel):
         if not self._check_user(update.effective_user.id):
             return
 
-        text = update.message.text
+        text = self._strip_bot_username_from_leading_command(update.message.text.strip(), self._get_bot_username(context))
         chat_id = str(update.effective_chat.id)
         user_id = str(update.effective_user.id)
         msg_id = str(update.message.message_id)
@@ -279,7 +311,7 @@ class TelegramChannel(Channel):
         if not self._check_user(update.effective_user.id):
             return
 
-        text = update.message.text.strip()
+        text = self._strip_bot_username_from_leading_command(update.message.text.strip(), self._get_bot_username(context))
         if not text:
             return
 

backend/app/channels/wechat.py

@@ -22,6 +22,7 @@ from cryptography.hazmat.primitives import padding
 from cryptography.hazmat.primitives.ciphers import Cipher, algorithms, modes
 
 from app.channels.base import Channel
+from app.channels.commands import is_known_channel_command
 from app.channels.message_bus import InboundMessageType, MessageBus, OutboundMessage, ResolvedAttachment
 
 logger = logging.getLogger(__name__)
@@ -620,7 +621,7 @@ class WechatChannel(Channel):
             chat_id=chat_id,
             user_id=chat_id,
             text=text,
-            msg_type=InboundMessageType.COMMAND if text.startswith("/") else InboundMessageType.CHAT,
+            msg_type=InboundMessageType.COMMAND if is_known_channel_command(text) else InboundMessageType.CHAT,
             thread_ts=thread_ts,
             files=files,
             metadata={

backend/app/channels/wecom.py

@@ -8,6 +8,7 @@ from collections.abc import Awaitable, Callable
 from typing import Any, cast
 
 from app.channels.base import Channel
+from app.channels.commands import is_known_channel_command
 from app.channels.message_bus import (
     InboundMessageType,
     MessageBus,
@@ -270,7 +271,7 @@ class WeComChannel(Channel):
 
         user_id = (body.get("from") or {}).get("userid")
 
-        inbound_type = InboundMessageType.COMMAND if text.startswith("/") else InboundMessageType.CHAT
+        inbound_type = InboundMessageType.COMMAND if is_known_channel_command(text) else InboundMessageType.CHAT
         inbound = self._make_inbound(
             chat_id=user_id,  # keep user's conversation in memory
             user_id=user_id,

backend/app/gateway/app.py

@@ -161,10 +161,16 @@ async def _migrate_orphaned_threads(store, admin_user_id: str) -> int:
 async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
     """Application lifespan handler."""
 
-    # Load config and check necessary environment variables at startup
+    # Load config and check necessary environment variables at startup.
+    # `startup_config` is a local snapshot used only for one-shot bootstrap
+    # work (logging level, langgraph_runtime engines, channels). Request-time
+    # config resolution always routes through `get_app_config()` in
+    # `app/gateway/deps.py::get_config()` so `config.yaml` edits become
+    # visible without a process restart. We deliberately do NOT cache this
+    # snapshot on `app.state` to keep that contract enforceable.
     try:
-        app.state.config = get_app_config()
-        apply_logging_level(app.state.config.log_level)
+        startup_config = get_app_config()
+        apply_logging_level(startup_config.log_level)
         logger.info("Configuration loaded successfully")
     except Exception as e:
         error_msg = f"Failed to load configuration during gateway startup: {e}"
@@ -173,8 +179,27 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
     config = get_gateway_config()
     logger.info(f"Starting API Gateway on {config.host}:{config.port}")
 
+    # Pre-warm tiktoken encoding cache so the first memory-injection request
+    # never blocks on the BPE data download (which hits an OpenAI/Azure URL
+    # that may be unreachable in restricted networks — see issue #3402).
+    try:
+        from deerflow.agents.memory.prompt import warm_tiktoken_cache
+
+        warmed = await asyncio.wait_for(
+            asyncio.to_thread(warm_tiktoken_cache),
+            timeout=5,
+        )
+        if warmed:
+            logger.info("tiktoken encoding cache warmed successfully")
+        else:
+            logger.warning("tiktoken encoding cache warm-up failed; token counting will use character-based fallback")
+    except TimeoutError:
+        logger.warning("tiktoken encoding cache warm-up timed out; token counting will use character-based fallback")
+    except Exception:
+        logger.warning("tiktoken warm-up skipped", exc_info=True)
+
     # Initialize LangGraph runtime components (StreamBridge, RunManager, checkpointer, store)
-    async with langgraph_runtime(app):
+    async with langgraph_runtime(app, startup_config):
         logger.info("LangGraph runtime initialised")
 
         # Check admin bootstrap state and migrate orphan threads after admin exists.
@@ -185,7 +210,7 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
         try:
             from app.channels.service import start_channel_service
 
-            channel_service = await start_channel_service(app.state.config)
+            channel_service = await start_channel_service(startup_config)
             logger.info("Channel service started: %s", channel_service.get_status())
         except Exception:
             logger.exception("No IM channels configured or channel service failed to start")

backend/app/gateway/deps.py

@@ -3,11 +3,22 @@
 **Getters** (used by routers): raise 503 when a required dependency is
 missing, except ``get_store`` which returns ``None``.
 
+``AppConfig`` is intentionally *not* cached on ``app.state``. Routers and the
+run path resolve it through :func:`deerflow.config.app_config.get_app_config`,
+which performs mtime-based hot reload, so edits to ``config.yaml`` take
+effect on the next request without a process restart. The engines created in
+:func:`langgraph_runtime` (stream bridge, persistence, checkpointer, store,
+run-event store) accept a ``startup_config`` snapshot — they are
+restart-required by design and stay bound to that snapshot to keep the live
+process consistent with itself.
+
 Initialization is handled directly in ``app.py`` via :class:`AsyncExitStack`.
 """
 
 from __future__ import annotations
 
+import asyncio
+import logging
 from collections.abc import AsyncGenerator, Callable
 from contextlib import AsyncExitStack, asynccontextmanager
 from typing import TYPE_CHECKING, TypeVar, cast
@@ -15,36 +26,144 @@ from typing import TYPE_CHECKING, TypeVar, cast
 from fastapi import FastAPI, HTTPException, Request
 from langgraph.types import Checkpointer
 
-from deerflow.config.app_config import AppConfig
+from deerflow.config.app_config import AppConfig, get_app_config
 from deerflow.persistence.feedback import FeedbackRepository
 from deerflow.runtime import RunContext, RunManager, StreamBridge
 from deerflow.runtime.events.store.base import RunEventStore
 from deerflow.runtime.runs.store.base import RunStore
 
+logger = logging.getLogger(__name__)
+
+# Upper bound (seconds) for draining in-flight runs during shutdown, before the
+# AsyncExitStack tears down the checkpointer (and its connection pool). Kept
+# local to avoid an app -> deps -> app import cycle. This is a *separate* budget
+# from ``app.gateway.app._SHUTDOWN_HOOK_TIMEOUT_SECONDS`` (currently also 5.0s,
+# which bounds channel-service stop): the two govern independent teardown steps
+# and may diverge, but both count toward the lifespan shutdown window — revisit
+# them together if their sum must stay within the server's graceful-shutdown
+# timeout.
+_RUN_DRAIN_TIMEOUT_SECONDS = 5.0
+
+
+async def _drain_inflight_runs(run_manager: RunManager) -> None:
+    """Drain in-flight runs before the checkpointer is torn down (issue #3373).
+
+    Shields the (internally-bounded) drain so that even if the lifespan
+    coroutine is itself cancelled mid-shutdown — a second SIGINT or the server's
+    graceful-shutdown timeout, i.e. the same signal storm behind #3373 — the
+    checkpointer pool is not closed while run tasks are still writing
+    checkpoints. On such a cancellation we let the already-running drain finish
+    (it is bounded by ``RunManager.shutdown``'s own timeout) and then propagate
+    the cancellation.
+    """
+    drain = asyncio.create_task(run_manager.shutdown(timeout=_RUN_DRAIN_TIMEOUT_SECONDS))
+    try:
+        await asyncio.shield(drain)
+    except asyncio.CancelledError:
+        # Re-shield so this second wait does not abandon the in-flight drain;
+        # it is bounded, so this cannot hang. Then re-raise to honour shutdown.
+        try:
+            await asyncio.shield(drain)
+        except Exception:
+            logger.exception("In-flight run drain failed after shutdown cancellation")
+        raise
+    except Exception:
+        logger.exception("Failed to drain in-flight runs during shutdown")
+
+
 if TYPE_CHECKING:
     from app.gateway.auth.local_provider import LocalAuthProvider
     from app.gateway.auth.repositories.sqlite import SQLiteUserRepository
     from deerflow.persistence.thread_meta.base import ThreadMetaStore
+    from deerflow.runtime import RunRecord
 
 
 T = TypeVar("T")
 
 
-def get_config(request: Request) -> AppConfig:
-    """Return the app-scoped ``AppConfig`` stored on ``app.state``."""
-    config = getattr(request.app.state, "config", None)
-    if config is None:
-        raise HTTPException(status_code=503, detail="Configuration not available")
-    return config
+async def _mark_latest_recovered_threads_error(
+    run_manager: RunManager,
+    thread_store: ThreadMetaStore,
+    recovered_runs: list[RunRecord],
+) -> None:
+    """Mark thread status as error only when its newest run was recovered."""
+    recovered_by_thread: dict[str, set[str]] = {}
+    for record in recovered_runs:
+        recovered_by_thread.setdefault(record.thread_id, set()).add(record.run_id)
+
+    for thread_id, recovered_run_ids in recovered_by_thread.items():
+        try:
+            latest_runs = await run_manager.list_by_thread(thread_id, user_id=None, limit=1)
+        except Exception:
+            logger.warning("Failed to find latest run for thread %s during run reconciliation", thread_id, exc_info=True)
+            continue
+        if not latest_runs or latest_runs[0].run_id not in recovered_run_ids:
+            continue
+        try:
+            await thread_store.update_status(thread_id, "error", user_id=None)
+        except Exception:
+            logger.warning("Failed to mark thread %s as error during run reconciliation", thread_id, exc_info=True)
+
+
+def get_config() -> AppConfig:
+    """Return the freshest ``AppConfig`` for the current request.
+
+    Routes through :func:`deerflow.config.app_config.get_app_config`, which
+    honours runtime ``ContextVar`` overrides and reloads ``config.yaml`` from
+    disk when its mtime changes. ``AppConfig`` is not cached on ``app.state``
+    at all — the only startup-time snapshot lives as a local
+    ``startup_config`` variable inside ``lifespan()`` and is passed
+    explicitly into :func:`langgraph_runtime` for the engines that are
+    restart-required by design. Routing every request through
+    :func:`get_app_config` closes the bytedance/deer-flow issue #3107 BUG-001
+    split-brain where the worker / lead-agent thread saw a stale startup
+    snapshot.
+
+    Hot-reload boundary: fields backed by startup-time singletons
+    (engines, sandbox provider, IM channels, logging handler) require a
+    process restart to change at runtime. The authoritative list lives in
+    :mod:`deerflow.config.reload_boundary` and is mirrored by the
+    standardised ``"startup-only:"`` prefix on the matching
+    ``Field(description=...)`` in :class:`AppConfig` — IDE hover on those
+    fields will surface the boundary inline. See
+    ``backend/CLAUDE.md`` "Config Hot-Reload Boundary" for the operator
+    summary.
+
+    Any failure to materialise the config (missing file, permission denied,
+    YAML parse error, validation error) is reported as 503 — semantically
+    "the gateway cannot serve requests without a usable configuration" — and
+    logged with the original exception so operators have something to debug.
+    """
+    try:
+        return get_app_config()
+    except Exception as exc:  # noqa: BLE001 - request boundary: log and degrade gracefully
+        logger.exception("Failed to load AppConfig at request time")
+        raise HTTPException(status_code=503, detail="Configuration not available") from exc
 
 
 @asynccontextmanager
-async def langgraph_runtime(app: FastAPI) -> AsyncGenerator[None, None]:
+async def langgraph_runtime(app: FastAPI, startup_config: AppConfig) -> AsyncGenerator[None, None]:
     """Bootstrap and tear down all LangGraph runtime singletons.
 
+    ``startup_config`` is the ``AppConfig`` snapshot taken once during
+    ``lifespan()`` for one-shot infrastructure bootstrap. The engines and
+    stores constructed here (stream bridge, persistence engine, checkpointer,
+    store, run-event store) are restart-required by design — they hold live
+    connections, file handles, or singleton providers — so they bind to this
+    snapshot and survive across `config.yaml` edits. Request-time consumers
+    must still go through :func:`get_config` for any field that should be
+    hot-reloadable. See ``backend/CLAUDE.md`` "Config Hot-Reload Boundary".
+
+    The matching ``run_events_config`` is frozen onto ``app.state`` so
+    :func:`get_run_context` pairs a freshly-loaded ``AppConfig`` with the
+    *startup-time* run-events configuration the underlying ``event_store``
+    was built from — otherwise the runtime could end up combining a live
+    new ``run_events_config`` with an event store still bound to the
+    previous backend.
+
     Usage in ``app.py``::
 
-        async with langgraph_runtime(app):
+        async with langgraph_runtime(app, startup_config):
             yield
     """
     from deerflow.persistence.engine import close_engine, get_session_factory, init_engine_from_config
@@ -53,9 +172,7 @@ async def langgraph_runtime(app: FastAPI) -> AsyncGenerator[None, None]:
     from deerflow.runtime.events.store import make_run_event_store
 
     async with AsyncExitStack() as stack:
-        config = getattr(app.state, "config", None)
-        if config is None:
-            raise RuntimeError("langgraph_runtime() requires app.state.config to be initialized")
+        config = startup_config
 
         app.state.stream_bridge = await stack.enter_async_context(make_stream_bridge(config))
 
@@ -84,16 +201,38 @@ async def langgraph_runtime(app: FastAPI) -> AsyncGenerator[None, None]:
 
         app.state.thread_store = make_thread_store(sf, app.state.store)
 
-        # Run event store (has its own factory with config-driven backend selection)
+        # Run event store. The store and the matching ``run_events_config`` are
+        # both frozen at startup so ``get_run_context`` does not combine a
+        # freshly-reloaded ``AppConfig.run_events`` with a store still bound to
+        # the previous backend.
         run_events_config = getattr(config, "run_events", None)
+        app.state.run_events_config = run_events_config
         app.state.run_event_store = make_run_event_store(run_events_config)
 
         # RunManager with store backing for persistence
         app.state.run_manager = RunManager(store=app.state.run_store)
+        if getattr(config.database, "backend", None) == "sqlite":
+            from deerflow.utils.time import now_iso
+
+            # Startup-only recovery: clean shutdowns return no active rows and
+            # the thread-status update below becomes a no-op.
+            recovered_runs = await app.state.run_manager.reconcile_orphaned_inflight_runs(
+                error="Gateway restarted before this run reached a durable final state.",
+                before=now_iso(),
+            )
+            await _mark_latest_recovered_threads_error(app.state.run_manager, app.state.thread_store, recovered_runs)
 
         try:
             yield
         finally:
+            # Drain in-flight run tasks BEFORE the AsyncExitStack tears down the
+            # checkpointer (and its connection pool). A run still mid-graph would
+            # otherwise leak into asyncio.run() shutdown, where langgraph's
+            # _checkpointer_put_after_previous aput races the closed pool and
+            # raises PoolClosed (issue #3373).
+            run_manager = getattr(app.state, "run_manager", None)
+            if run_manager is not None:
+                await _drain_inflight_runs(run_manager)
             await close_engine()
 
 
@@ -139,16 +278,20 @@ def get_thread_store(request: Request) -> ThreadMetaStore:
 def get_run_context(request: Request) -> RunContext:
     """Build a :class:`RunContext` from ``app.state`` singletons.
 
-    Returns a *base* context with infrastructure dependencies.
+    Returns a *base* context with infrastructure dependencies. The
+    ``app_config`` field is resolved live so per-run fields (e.g.
+    ``models[*].max_tokens``) follow ``config.yaml`` edits; the
+    ``event_store`` / ``run_events_config`` pair stays frozen to the snapshot
+    captured in :func:`langgraph_runtime` so callers never see a store bound
+    to one backend paired with a config pointing at another.
     """
-    config = get_config(request)
     return RunContext(
         checkpointer=get_checkpointer(request),
         store=get_store(request),
         event_store=get_run_event_store(request),
-        run_events_config=getattr(config, "run_events", None),
+        run_events_config=getattr(request.app.state, "run_events_config", None),
         thread_store=get_thread_store(request),
-        app_config=config,
+        app_config=get_config(),
     )
 
 

backend/app/gateway/internal_auth.py

@@ -1,26 +1,38 @@
-"""Process-local authentication for Gateway internal callers."""
+"""Authentication for trusted Gateway internal callers."""
 
 from __future__ import annotations
 
+import os
 import secrets
 from types import SimpleNamespace
 
 from deerflow.runtime.user_context import DEFAULT_USER_ID
 
 INTERNAL_AUTH_HEADER_NAME = "X-DeerFlow-Internal-Token"
-_INTERNAL_AUTH_TOKEN = secrets.token_urlsafe(32)
+INTERNAL_AUTH_ENV_VAR = "DEER_FLOW_INTERNAL_AUTH_TOKEN"
+INTERNAL_SYSTEM_ROLE = "internal"
+
+
+def _load_internal_auth_token() -> str:
+    token = os.environ.get(INTERNAL_AUTH_ENV_VAR)
+    if token:
+        return token
+    return secrets.token_urlsafe(32)
+
+
+_INTERNAL_AUTH_TOKEN = _load_internal_auth_token()
 
 
 def create_internal_auth_headers() -> dict[str, str]:
-    """Return headers that authenticate same-process Gateway internal calls."""
+    """Return headers that authenticate trusted Gateway internal calls."""
     return {INTERNAL_AUTH_HEADER_NAME: _INTERNAL_AUTH_TOKEN}
 
 
 def is_valid_internal_auth_token(token: str | None) -> bool:
-    """Return True when *token* matches the process-local internal token."""
+    """Return True when *token* matches this Gateway worker's internal token."""
     return bool(token) and secrets.compare_digest(token, _INTERNAL_AUTH_TOKEN)
 
 
 def get_internal_user():
     """Return the synthetic user used for trusted internal channel calls."""
-    return SimpleNamespace(id=DEFAULT_USER_ID, system_role="internal")
+    return SimpleNamespace(id=DEFAULT_USER_ID, system_role=INTERNAL_SYSTEM_ROLE)

backend/app/gateway/pagination.py

@@ -0,0 +1,15 @@
+"""Shared pagination helpers for gateway routers."""
+
+from __future__ import annotations
+
+
+def trim_run_message_page(rows: list[dict], *, limit: int, after_seq: int | None) -> tuple[list[dict], bool]:
+    """Trim a ``limit + 1`` run-message page while preserving page boundaries."""
+    has_more = len(rows) > limit
+    if not has_more:
+        return rows, False
+
+    if after_seq is not None:
+        return rows[:limit], True
+
+    return rows[-limit:], True

backend/app/gateway/routers/agents.py

@@ -1,5 +1,6 @@
 """CRUD API for custom agents."""
 
+import asyncio
 import logging
 import re
 import shutil
@@ -213,48 +214,61 @@ async def create_agent_endpoint(request: AgentCreateRequest) -> AgentResponse:
     user_id = get_effective_user_id()
     paths = get_paths()
 
-    agent_dir = paths.user_agent_dir(user_id, normalized_name)
-    legacy_dir = paths.agent_dir(normalized_name)
+    def _create_agent() -> AgentResponse | None:
+        # Worker thread: base-dir resolution, existence checks, directory/file
+        # creation, read-back, and failure cleanup are all blocking filesystem
+        # IO that must stay off the event loop.
+        agent_dir = paths.user_agent_dir(user_id, normalized_name)
+        legacy_dir = paths.agent_dir(normalized_name)
 
-    if agent_dir.exists() or legacy_dir.exists():
-        raise HTTPException(status_code=409, detail=f"Agent '{normalized_name}' already exists")
+        if legacy_dir.exists():
+            return None  # signals 409 to the caller
+
+        try:
+            try:
+                agent_dir.mkdir(parents=True, exist_ok=False)
+            except FileExistsError:
+                return None  # signals 409 to the caller
+            # Write config.yaml
+            config_data: dict = {"name": normalized_name}
+            if request.description:
+                config_data["description"] = request.description
+            if request.model is not None:
+                config_data["model"] = request.model
+            if request.tool_groups is not None:
+                config_data["tool_groups"] = request.tool_groups
+            if request.skills is not None:
+                config_data["skills"] = request.skills
+
+            config_file = agent_dir / "config.yaml"
+            with open(config_file, "w", encoding="utf-8") as f:
+                yaml.dump(config_data, f, default_flow_style=False, allow_unicode=True)
+
+            # Write SOUL.md
+            soul_file = agent_dir / "SOUL.md"
+            soul_file.write_text(request.soul, encoding="utf-8")
+
+            logger.info(f"Created agent '{normalized_name}' at {agent_dir}")
+
+            agent_cfg = load_agent_config(normalized_name, user_id=user_id)
+            return _agent_config_to_response(agent_cfg, include_soul=True, user_id=user_id)
+        except Exception:
+            # Clean up partial state on failure before surfacing the error.
+            if agent_dir.exists():
+                shutil.rmtree(agent_dir)
+            raise
 
     try:
-        agent_dir.mkdir(parents=True, exist_ok=True)
-
-        # Write config.yaml
-        config_data: dict = {"name": normalized_name}
-        if request.description:
-            config_data["description"] = request.description
-        if request.model is not None:
-            config_data["model"] = request.model
-        if request.tool_groups is not None:
-            config_data["tool_groups"] = request.tool_groups
-        if request.skills is not None:
-            config_data["skills"] = request.skills
-
-        config_file = agent_dir / "config.yaml"
-        with open(config_file, "w", encoding="utf-8") as f:
-            yaml.dump(config_data, f, default_flow_style=False, allow_unicode=True)
-
-        # Write SOUL.md
-        soul_file = agent_dir / "SOUL.md"
-        soul_file.write_text(request.soul, encoding="utf-8")
-
-        logger.info(f"Created agent '{normalized_name}' at {agent_dir}")
-
-        agent_cfg = load_agent_config(normalized_name, user_id=user_id)
-        return _agent_config_to_response(agent_cfg, include_soul=True, user_id=user_id)
-
-    except HTTPException:
-        raise
+        response = await asyncio.to_thread(_create_agent)
     except Exception as e:
-        # Clean up on failure
-        if agent_dir.exists():
-            shutil.rmtree(agent_dir)
         logger.error(f"Failed to create agent '{request.name}': {e}", exc_info=True)
         raise HTTPException(status_code=500, detail=f"Failed to create agent: {str(e)}")
 
+    if response is None:
+        raise HTTPException(status_code=409, detail=f"Agent '{normalized_name}' already exists")
+
+    return response
+
 
 @router.put(
     "/agents/{name}",
@@ -428,19 +442,30 @@ async def delete_agent(name: str) -> None:
     name = _normalize_agent_name(name)
     user_id = get_effective_user_id()
     paths = get_paths()
-    agent_dir = paths.user_agent_dir(user_id, name)
 
-    if not agent_dir.exists():
-        if paths.agent_dir(name).exists():
-            raise HTTPException(
-                status_code=409,
-                detail=(f"Agent '{name}' only exists in the legacy shared layout and is not scoped to a user. Run scripts/migrate_user_isolation.py to move legacy agents into the per-user layout before deleting."),
-            )
-        raise HTTPException(status_code=404, detail=f"Agent '{name}' not found")
+    def _remove_agent_dir() -> tuple[str, str]:
+        # Runs in a worker thread: resolving the base dir, probing the directory
+        # (`exists`), and removing it (`rmtree`) are all blocking filesystem IO
+        # that must stay off the event loop.
+        agent_dir = paths.user_agent_dir(user_id, name)
+        if not agent_dir.exists():
+            outcome = "legacy" if paths.agent_dir(name).exists() else "missing"
+            return outcome, str(agent_dir)
+        shutil.rmtree(agent_dir)
+        return "deleted", str(agent_dir)
 
     try:
-        shutil.rmtree(agent_dir)
-        logger.info(f"Deleted agent '{name}' from {agent_dir}")
+        outcome, agent_dir = await asyncio.to_thread(_remove_agent_dir)
     except Exception as e:
         logger.error(f"Failed to delete agent '{name}': {e}", exc_info=True)
         raise HTTPException(status_code=500, detail=f"Failed to delete agent: {str(e)}")
+
+    if outcome == "legacy":
+        raise HTTPException(
+            status_code=409,
+            detail=(f"Agent '{name}' only exists in the legacy shared layout and is not scoped to a user. Run scripts/migrate_user_isolation.py to move legacy agents into the per-user layout before deleting."),
+        )
+    if outcome == "missing":
+        raise HTTPException(status_code=404, detail=f"Agent '{name}' not found")
+
+    logger.info(f"Deleted agent '{name}' from {agent_dir}")

backend/app/gateway/routers/mcp.py

@@ -1,9 +1,10 @@
 import json
 import logging
+import os
 from pathlib import Path
 from typing import Literal
 
-from fastapi import APIRouter, HTTPException
+from fastapi import APIRouter, HTTPException, Request, status
 from pydantic import BaseModel, Field
 
 from deerflow.config.extensions_config import ExtensionsConfig, get_extensions_config, reload_extensions_config
@@ -12,6 +13,11 @@ logger = logging.getLogger(__name__)
 router = APIRouter(prefix="/api", tags=["mcp"])
 
 
+_MCP_STDIO_COMMAND_ALLOWLIST_ENV = "DEER_FLOW_MCP_STDIO_COMMAND_ALLOWLIST"
+_DEFAULT_MCP_STDIO_COMMAND_ALLOWLIST = frozenset({"npx", "uvx"})
+_SHELL_METACHARS = frozenset(";|&`$<>\n\r")
+
+
 class McpOAuthConfigResponse(BaseModel):
     """OAuth configuration for an MCP server."""
 
@@ -63,13 +69,178 @@ class McpConfigUpdateRequest(BaseModel):
     )
 
 
+_MASKED_VALUE = "***"
+
+
+async def _require_admin_user(request: Request) -> None:
+    """Require the authenticated caller to be an admin user.
+
+    ``AuthMiddleware`` normally stamps ``request.state.user`` before the
+    request reaches this router. Falling back to the strict dependency keeps
+    this route safe even in tests or alternative ASGI compositions that mount
+    the router without the global middleware.
+    """
+    user = getattr(request.state, "user", None)
+    if user is None:
+        from app.gateway.deps import get_current_user_from_request
+
+        user = await get_current_user_from_request(request)
+
+    if getattr(user, "system_role", None) != "admin":
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="Admin privileges required to manage MCP configuration.",
+        )
+
+
+def _allowed_stdio_commands() -> set[str]:
+    """Return executable names allowed for API-managed stdio MCP servers."""
+    raw = os.environ.get(_MCP_STDIO_COMMAND_ALLOWLIST_ENV)
+    base = set(_DEFAULT_MCP_STDIO_COMMAND_ALLOWLIST)
+    if raw is None:
+        return base
+    extra = {item.strip() for item in raw.split(",") if item.strip()}
+    return base | extra
+
+
+def _stdio_command_name(command: str | None, *, server_name: str) -> str:
+    """Normalize and validate a stdio command field from the API boundary."""
+    if command is None or not command.strip():
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail=f"MCP server '{server_name}' with stdio transport requires a command.",
+        )
+
+    stripped = command.strip()
+    has_path_separator = "/" in stripped or "\\" in stripped
+    if stripped != command or has_path_separator or any(ch.isspace() for ch in stripped) or any(ch in stripped for ch in _SHELL_METACHARS):
+        raise HTTPException(
+            status_code=status.HTTP_400_BAD_REQUEST,
+            detail=(f"MCP server '{server_name}' command must be a single executable name; put parameters in args instead."),
+        )
+
+    return stripped
+
+
+def _validate_mcp_update_request(request: McpConfigUpdateRequest) -> None:
+    """Validate API-submitted MCP config before it is persisted.
+
+    Local config files can still express arbitrary advanced setups, but the
+    HTTP API is an untrusted boundary. Restricting stdio commands here reduces
+    the blast radius of a compromised authenticated browser session.
+    """
+    allowed_commands = _allowed_stdio_commands()
+    for name, server in request.mcp_servers.items():
+        transport_type = (server.type or "stdio").lower()
+        if transport_type != "stdio":
+            continue
+
+        command_name = _stdio_command_name(server.command, server_name=name)
+        if command_name not in allowed_commands:
+            allowed = ", ".join(sorted(allowed_commands)) or "<none>"
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail=(f"MCP server '{name}' uses disallowed stdio command '{command_name}'. Allowed commands: {allowed}. Configure {_MCP_STDIO_COMMAND_ALLOWLIST_ENV} to extend this list."),
+            )
+
+
+def _mask_server_config(server: McpServerConfigResponse) -> McpServerConfigResponse:
+    """Return a copy of server config with sensitive fields masked.
+
+    Masks env values, header values, and removes OAuth secrets so they
+    are not exposed through the GET API endpoint.
+    """
+    masked_env = {k: _MASKED_VALUE for k in server.env}
+    masked_headers = {k: _MASKED_VALUE for k in server.headers}
+    masked_oauth = None
+    if server.oauth is not None:
+        masked_oauth = server.oauth.model_copy(
+            update={
+                "client_secret": None,
+                "refresh_token": None,
+            }
+        )
+    return server.model_copy(
+        update={
+            "env": masked_env,
+            "headers": masked_headers,
+            "oauth": masked_oauth,
+        }
+    )
+
+
+def _merge_preserving_secrets(
+    incoming: McpServerConfigResponse,
+    existing: McpServerConfigResponse,
+) -> McpServerConfigResponse:
+    """Merge incoming config with existing, preserving secrets masked by GET.
+
+    When the frontend toggles ``enabled`` it round-trips the full config:
+    GET (masked) → modify enabled → PUT (masked values sent back).
+    This function ensures masked values (``***``) are replaced with the
+    real secrets from the current on-disk config.
+
+    ``***`` is only accepted for keys that already exist in *existing*.
+    New keys must provide a real value.
+
+    For OAuth secrets, ``None`` means "preserve the existing stored value"
+    so masked GET responses can be safely round-tripped. To explicitly clear
+    a stored secret, clients may send an empty string, which is converted
+    to ``None`` before persisting.
+    """
+    merged_env = {}
+    for k, v in incoming.env.items():
+        if v == _MASKED_VALUE:
+            if k in existing.env:
+                merged_env[k] = existing.env[k]
+            else:
+                raise HTTPException(
+                    status_code=400,
+                    detail=f"Cannot set env key '{k}' to masked value '***'; provide a real value.",
+                )
+        else:
+            merged_env[k] = v
+
+    merged_headers = {}
+    for k, v in incoming.headers.items():
+        if v == _MASKED_VALUE:
+            if k in existing.headers:
+                merged_headers[k] = existing.headers[k]
+            else:
+                raise HTTPException(
+                    status_code=400,
+                    detail=f"Cannot set header '{k}' to masked value '***'; provide a real value.",
+                )
+        else:
+            merged_headers[k] = v
+
+    merged_oauth = incoming.oauth
+    if incoming.oauth is not None and existing.oauth is not None:
+        # None = preserve (masked round-trip), "" = explicitly clear, else = new value
+        merged_client_secret = existing.oauth.client_secret if incoming.oauth.client_secret is None else (None if incoming.oauth.client_secret == "" else incoming.oauth.client_secret)
+        merged_refresh_token = existing.oauth.refresh_token if incoming.oauth.refresh_token is None else (None if incoming.oauth.refresh_token == "" else incoming.oauth.refresh_token)
+        merged_oauth = incoming.oauth.model_copy(
+            update={
+                "client_secret": merged_client_secret,
+                "refresh_token": merged_refresh_token,
+            }
+        )
+    return incoming.model_copy(
+        update={
+            "env": merged_env,
+            "headers": merged_headers,
+            "oauth": merged_oauth,
+        }
+    )
+
+
 @router.get(
     "/mcp/config",
     response_model=McpConfigResponse,
     summary="Get MCP Configuration",
     description="Retrieve the current Model Context Protocol (MCP) server configurations.",
 )
-async def get_mcp_configuration() -> McpConfigResponse:
+async def get_mcp_configuration(request: Request) -> McpConfigResponse:
     """Get the current MCP configuration.
 
     Returns:
@@ -83,16 +254,19 @@ async def get_mcp_configuration() -> McpConfigResponse:
                     "enabled": true,
                     "command": "npx",
                     "args": ["-y", "@modelcontextprotocol/server-github"],
-                    "env": {"GITHUB_TOKEN": "ghp_xxx"},
+                    "env": {"GITHUB_TOKEN": "***"},
                     "description": "GitHub MCP server for repository operations"
                 }
             }
         }
         ```
     """
+    await _require_admin_user(request)
+
     config = get_extensions_config()
 
-    return McpConfigResponse(mcp_servers={name: McpServerConfigResponse(**server.model_dump()) for name, server in config.mcp_servers.items()})
+    servers = {name: _mask_server_config(McpServerConfigResponse(**server.model_dump())) for name, server in config.mcp_servers.items()}
+    return McpConfigResponse(mcp_servers=servers)
 
 
 @router.put(
@@ -101,7 +275,7 @@ async def get_mcp_configuration() -> McpConfigResponse:
     summary="Update MCP Configuration",
     description="Update Model Context Protocol (MCP) server configurations and save to file.",
 )
-async def update_mcp_configuration(request: McpConfigUpdateRequest) -> McpConfigResponse:
+async def update_mcp_configuration(request: Request, body: McpConfigUpdateRequest) -> McpConfigResponse:
     """Update the MCP configuration.
 
     This will:
@@ -134,6 +308,9 @@ async def update_mcp_configuration(request: McpConfigUpdateRequest) -> McpConfig
         ```
     """
     try:
+        await _require_admin_user(request)
+        _validate_mcp_update_request(body)
+
         # Get the current config path (or determine where to save it)
         config_path = ExtensionsConfig.resolve_config_path()
 
@@ -142,14 +319,39 @@ async def update_mcp_configuration(request: McpConfigUpdateRequest) -> McpConfig
             config_path = Path.cwd().parent / "extensions_config.json"
             logger.info(f"No existing extensions config found. Creating new config at: {config_path}")
 
-        # Load current config to preserve skills configuration
+        # Load current config to preserve skills
         current_config = get_extensions_config()
 
-        # Convert request to dict format for JSON serialization
-        config_data = {
-            "mcpServers": {name: server.model_dump() for name, server in request.mcp_servers.items()},
-            "skills": {name: {"enabled": skill.enabled} for name, skill in current_config.skills.items()},
-        }
+        # Load raw (un-resolved) JSON from disk to use as the merge source.
+        # This preserves $VAR placeholders in env values and top-level keys
+        # like mcpInterceptors that would otherwise be lost.
+        raw_servers: dict[str, dict] = {}
+        raw_other_keys: dict = {}
+        if config_path is not None and config_path.exists():
+            with open(config_path, encoding="utf-8") as f:
+                raw_data = json.load(f)
+            raw_servers = raw_data.get("mcpServers", {})
+            # Preserve any top-level keys beyond mcpServers/skills
+            for key, value in raw_data.items():
+                if key not in ("mcpServers", "skills"):
+                    raw_other_keys[key] = value
+
+        # Merge incoming server configs with raw on-disk secrets
+        merged_servers: dict[str, McpServerConfigResponse] = {}
+        for name, incoming in body.mcp_servers.items():
+            raw_server = raw_servers.get(name)
+            if raw_server is not None:
+                merged_servers[name] = _merge_preserving_secrets(
+                    incoming,
+                    McpServerConfigResponse(**raw_server),
+                )
+            else:
+                merged_servers[name] = incoming
+
+        # Build config data preserving all top-level keys from the original file
+        config_data = dict(raw_other_keys)
+        config_data["mcpServers"] = {name: server.model_dump() for name, server in merged_servers.items()}
+        config_data["skills"] = {name: {"enabled": skill.enabled} for name, skill in current_config.skills.items()}
 
         # Write the configuration to file
         with open(config_path, "w", encoding="utf-8") as f:
@@ -157,13 +359,15 @@ async def update_mcp_configuration(request: McpConfigUpdateRequest) -> McpConfig
 
         logger.info(f"MCP configuration updated and saved to: {config_path}")
 
-        # NOTE: No need to reload/reset cache here - LangGraph Server (separate process)
-        # will detect config file changes via mtime and reinitialize MCP tools automatically
-
-        # Reload the configuration and update the global cache
+        # Reload the Gateway configuration and update the global cache. The
+        # agent runtime lives in Gateway, so this keeps API reads and tool
+        # execution aligned after extensions_config.json changes.
         reloaded_config = reload_extensions_config()
-        return McpConfigResponse(mcp_servers={name: McpServerConfigResponse(**server.model_dump()) for name, server in reloaded_config.mcp_servers.items()})
+        servers = {name: _mask_server_config(McpServerConfigResponse(**server.model_dump())) for name, server in reloaded_config.mcp_servers.items()}
+        return McpConfigResponse(mcp_servers=servers)
 
+    except HTTPException:
+        raise
     except Exception as e:
         logger.error(f"Failed to update MCP configuration: {e}", exc_info=True)
         raise HTTPException(status_code=500, detail=f"Failed to update MCP configuration: {str(e)}")

backend/app/gateway/routers/runs.py

@@ -7,7 +7,6 @@ is reused so that conversation history is preserved across calls.
 
 from __future__ import annotations
 
-import asyncio
 import logging
 import uuid
 
@@ -16,8 +15,9 @@ from fastapi.responses import StreamingResponse
 
 from app.gateway.authz import require_permission
 from app.gateway.deps import get_checkpointer, get_feedback_repo, get_run_event_store, get_run_manager, get_run_store, get_stream_bridge
+from app.gateway.pagination import trim_run_message_page
 from app.gateway.routers.thread_runs import RunCreateRequest
-from app.gateway.services import sse_consumer, start_run
+from app.gateway.services import sse_consumer, start_run, wait_for_run_completion
 from deerflow.runtime import serialize_channel_values
 
 logger = logging.getLogger(__name__)
@@ -66,24 +66,25 @@ async def stateless_wait(body: RunCreateRequest, request: Request) -> dict:
     Otherwise a new temporary thread is created.
     """
     thread_id = _resolve_thread_id(body)
+    bridge = get_stream_bridge(request)
+    run_mgr = get_run_manager(request)
     record = await start_run(body, thread_id, request)
 
+    completed = True
     if record.task is not None:
-        try:
-            await record.task
-        except asyncio.CancelledError:
-            pass
+        completed = await wait_for_run_completion(bridge, record, request, run_mgr)
 
-    checkpointer = get_checkpointer(request)
-    config = {"configurable": {"thread_id": thread_id}}
-    try:
-        checkpoint_tuple = await checkpointer.aget_tuple(config)
-        if checkpoint_tuple is not None:
-            checkpoint = getattr(checkpoint_tuple, "checkpoint", {}) or {}
-            channel_values = checkpoint.get("channel_values", {})
-            return serialize_channel_values(channel_values)
-    except Exception:
-        logger.exception("Failed to fetch final state for run %s", record.run_id)
+    if completed:
+        checkpointer = get_checkpointer(request)
+        config = {"configurable": {"thread_id": thread_id}}
+        try:
+            checkpoint_tuple = await checkpointer.aget_tuple(config)
+            if checkpoint_tuple is not None:
+                checkpoint = getattr(checkpoint_tuple, "checkpoint", {}) or {}
+                channel_values = checkpoint.get("channel_values", {})
+                return serialize_channel_values(channel_values)
+        except Exception:
+            logger.exception("Failed to fetch final state for run %s", record.run_id)
 
     return {"status": record.status.value, "error": record.error}
 
@@ -129,8 +130,7 @@ async def run_messages(
         before_seq=before_seq,
         after_seq=after_seq,
     )
-    has_more = len(rows) > limit
-    data = rows[:limit] if has_more else rows
+    data, has_more = trim_run_message_page(rows, limit=limit, after_seq=after_seq)
     return {"data": data, "has_more": has_more}
 
 

backend/app/gateway/routers/suggestions.py

@@ -1,5 +1,6 @@
 import json
 import logging
+import re
 
 from fastapi import APIRouter, Depends, Request
 from langchain_core.messages import HumanMessage, SystemMessage
@@ -30,6 +31,31 @@ class SuggestionsResponse(BaseModel):
     suggestions: list[str] = Field(default_factory=list, description="Suggested follow-up questions")
 
 
+# Matches a complete <think>...</think> block (case-insensitive, spans newlines).
+_THINK_BLOCK_RE = re.compile(r"<think\b[^>]*>.*?</think\s*>", re.IGNORECASE | re.DOTALL)
+# Matches a dangling, unclosed <think> (model truncated at max_tokens mid-thought).
+_OPEN_THINK_RE = re.compile(r"<think\b[^>]*>", re.IGNORECASE)
+
+
+def _strip_think_blocks(text: str) -> str:
+    """Remove reasoning-model ``<think>...</think>`` blocks from the response.
+
+    Reasoning models such as MiniMax-M3 inline their chain-of-thought into the
+    message ``content`` wrapped in ``<think>...</think>`` (``reasoning_split``
+    defaults to false), rather than exposing a separate ``reasoning_content``
+    field. The thinking text frequently contains ``[`` / ``]`` characters, which
+    corrupted the downstream ``find('[')`` / ``rfind(']')`` JSON extraction and
+    produced empty suggestions. We strip the reasoning before parsing so only
+    the actual answer remains.
+    """
+    text = _THINK_BLOCK_RE.sub("", text)
+    # Drop any unclosed <think> (and everything after it) left by truncation.
+    open_match = _OPEN_THINK_RE.search(text)
+    if open_match:
+        text = text[: open_match.start()]
+    return text.strip()
+
+
 def _strip_markdown_code_fence(text: str) -> str:
     stripped = text.strip()
     if not stripped.startswith("```"):
@@ -41,7 +67,8 @@ def _strip_markdown_code_fence(text: str) -> str:
 
 
 def _parse_json_string_list(text: str) -> list[str] | None:
-    candidate = _strip_markdown_code_fence(text)
+    candidate = _strip_think_blocks(text)
+    candidate = _strip_markdown_code_fence(candidate)
     start = candidate.find("[")
     end = candidate.rfind("]")
     if start == -1 or end == -1 or end <= start:

backend/app/gateway/routers/thread_runs.py

@@ -21,7 +21,8 @@ from pydantic import BaseModel, Field
 
 from app.gateway.authz import require_permission
 from app.gateway.deps import get_checkpointer, get_current_user, get_feedback_repo, get_run_event_store, get_run_manager, get_run_store, get_stream_bridge
-from app.gateway.services import sse_consumer, start_run
+from app.gateway.pagination import trim_run_message_page
+from app.gateway.services import sse_consumer, start_run, wait_for_run_completion
 from deerflow.runtime import RunRecord, RunStatus, serialize_channel_values
 
 logger = logging.getLogger(__name__)
@@ -66,6 +67,14 @@ class RunResponse(BaseModel):
     multitask_strategy: str = "reject"
     created_at: str = ""
     updated_at: str = ""
+    total_input_tokens: int = 0
+    total_output_tokens: int = 0
+    total_tokens: int = 0
+    llm_call_count: int = 0
+    lead_agent_tokens: int = 0
+    subagent_tokens: int = 0
+    middleware_tokens: int = 0
+    message_count: int = 0
 
 
 class ThreadTokenUsageModelBreakdown(BaseModel):
@@ -111,6 +120,14 @@ def _record_to_response(record: RunRecord) -> RunResponse:
         multitask_strategy=record.multitask_strategy,
         created_at=record.created_at,
         updated_at=record.updated_at,
+        total_input_tokens=record.total_input_tokens,
+        total_output_tokens=record.total_output_tokens,
+        total_tokens=record.total_tokens,
+        llm_call_count=record.llm_call_count,
+        lead_agent_tokens=record.lead_agent_tokens,
+        subagent_tokens=record.subagent_tokens,
+        middleware_tokens=record.middleware_tokens,
+        message_count=record.message_count,
     )
 
 
@@ -159,24 +176,25 @@ async def stream_run(thread_id: str, body: RunCreateRequest, request: Request) -
 @require_permission("runs", "create", owner_check=True, require_existing=True)
 async def wait_run(thread_id: str, body: RunCreateRequest, request: Request) -> dict:
     """Create a run and block until it completes, returning the final state."""
+    bridge = get_stream_bridge(request)
+    run_mgr = get_run_manager(request)
     record = await start_run(body, thread_id, request)
 
+    completed = True
     if record.task is not None:
-        try:
-            await record.task
-        except asyncio.CancelledError:
-            pass
+        completed = await wait_for_run_completion(bridge, record, request, run_mgr)
 
-    checkpointer = get_checkpointer(request)
-    config = {"configurable": {"thread_id": thread_id}}
-    try:
-        checkpoint_tuple = await checkpointer.aget_tuple(config)
-        if checkpoint_tuple is not None:
-            checkpoint = getattr(checkpoint_tuple, "checkpoint", {}) or {}
-            channel_values = checkpoint.get("channel_values", {})
-            return serialize_channel_values(channel_values)
-    except Exception:
-        logger.exception("Failed to fetch final state for run %s", record.run_id)
+    if completed:
+        checkpointer = get_checkpointer(request)
+        config = {"configurable": {"thread_id": thread_id}}
+        try:
+            checkpoint_tuple = await checkpointer.aget_tuple(config)
+            if checkpoint_tuple is not None:
+                checkpoint = getattr(checkpoint_tuple, "checkpoint", {}) or {}
+                channel_values = checkpoint.get("channel_values", {})
+                return serialize_channel_values(channel_values)
+        except Exception:
+            logger.exception("Failed to fetch final state for run %s", record.run_id)
 
     return {"status": record.status.value, "error": record.error}
 
@@ -261,7 +279,12 @@ async def join_run(thread_id: str, run_id: str, request: Request) -> StreamingRe
     )
 
 
-@router.api_route("/{thread_id}/runs/{run_id}/stream", methods=["GET", "POST"], response_model=None)
+# Register GET and POST as separate routes so each method gets a unique OpenAPI
+# operationId. ``api_route(methods=["GET", "POST"])`` shares one route registration
+# across both methods, which makes FastAPI emit the same ``operationId`` twice and
+# warn about a duplicate operation id during OpenAPI generation.
+@router.get("/{thread_id}/runs/{run_id}/stream", response_model=None)
+@router.post("/{thread_id}/runs/{run_id}/stream", response_model=None)
 @require_permission("runs", "read", owner_check=True)
 async def stream_existing_run(
     thread_id: str,
@@ -380,8 +403,7 @@ async def list_run_messages(
         before_seq=before_seq,
         after_seq=after_seq,
     )
-    has_more = len(rows) > limit
-    data = rows[:limit] if has_more else rows
+    data, has_more = trim_run_message_page(rows, limit=limit, after_seq=after_seq)
     return {"data": data, "has_more": has_more}
 
 
@@ -402,8 +424,15 @@ async def list_run_events(
 
 @router.get("/{thread_id}/token-usage", response_model=ThreadTokenUsageResponse)
 @require_permission("threads", "read", owner_check=True)
-async def thread_token_usage(thread_id: str, request: Request) -> ThreadTokenUsageResponse:
+async def thread_token_usage(
+    thread_id: str,
+    request: Request,
+    include_active: bool = Query(default=False, description="Include running run progress snapshots"),
+) -> ThreadTokenUsageResponse:
     """Thread-level token usage aggregation."""
     run_store = get_run_store(request)
-    agg = await run_store.aggregate_tokens_by_thread(thread_id)
+    if include_active:
+        agg = await run_store.aggregate_tokens_by_thread(thread_id, include_active=True)
+    else:
+        agg = await run_store.aggregate_tokens_by_thread(thread_id)
     return ThreadTokenUsageResponse(thread_id=thread_id, **agg)

backend/app/gateway/routers/threads.py

@@ -17,7 +17,7 @@ import uuid
 from typing import Any
 
 from fastapi import APIRouter, HTTPException, Request
-from langgraph.checkpoint.base import empty_checkpoint
+from langgraph.checkpoint.base import empty_checkpoint, uuid6
 from pydantic import BaseModel, Field, field_validator
 
 from app.gateway.authz import require_permission
@@ -536,9 +536,21 @@ async def update_thread_state(thread_id: str, body: ThreadStateUpdateRequest, re
         metadata["step"] = metadata.get("step", 0) + 1
         metadata["writes"] = {body.as_node: body.values}
 
+    # Assign a new checkpoint ID so aput performs an INSERT rather than an
+    # in-place REPLACE of the existing row.  Use uuid6 (time-ordered) rather
+    # than uuid4 (random) so the new ID is always lexicographically greater
+    # than the previous one — LangGraph's checkpointers determine the "latest"
+    # checkpoint by max(checkpoint_ids) string order, matching the uuid6 epoch.
+    checkpoint["id"] = str(uuid6())
+
     # aput requires checkpoint_ns in the config — use the same config used for the
-    # read (which always includes checkpoint_ns="").  Do NOT include checkpoint_id
-    # so that aput generates a fresh checkpoint ID for the new snapshot.
+    # read (which always includes checkpoint_ns=""). The fresh checkpoint ID is
+    # assigned above via checkpoint["id"]; keep checkpoint_id out of the config so
+    # the write is keyed by the new checkpoint payload rather than the prior read.
+    # All supported savers (InMemorySaver, AsyncSqliteSaver, AsyncPostgresSaver)
+    # persist and echo back checkpoint["id"] verbatim — none mint their own — so
+    # the new_config below carries the uuid6 we assigned here. (Regression-locked
+    # by test_update_thread_state_inserts_new_checkpoint_each_call.)
     write_config: dict[str, Any] = {
         "configurable": {
             "thread_id": thread_id,
@@ -557,7 +569,7 @@ async def update_thread_state(thread_id: str, body: ThreadStateUpdateRequest, re
 
     # Sync title changes through the ThreadMetaStore abstraction so /threads/search
     # reflects them immediately in both sqlite and memory backends.
-    if body.values and "title" in body.values:
+    if thread_store and body.values and "title" in body.values:
         new_title = body.values["title"]
         if new_title:  # Skip empty strings and None
             try:

backend/app/gateway/routers/uploads.py

@@ -39,15 +39,39 @@ DEFAULT_MAX_FILE_SIZE = 50 * 1024 * 1024
 DEFAULT_MAX_TOTAL_SIZE = 100 * 1024 * 1024
 
 
+class UploadedFileInfo(BaseModel):
+    """Uploaded file metadata exposed by upload and list APIs."""
+
+    filename: str
+    size: int
+    path: str
+    virtual_path: str
+    artifact_url: str
+    extension: str | None = None
+    modified: float | None = None
+    original_filename: str | None = None
+    markdown_file: str | None = None
+    markdown_path: str | None = None
+    markdown_virtual_path: str | None = None
+    markdown_artifact_url: str | None = None
+
+
 class UploadResponse(BaseModel):
     """Response model for file upload."""
 
     success: bool
-    files: list[dict[str, str]]
+    files: list[UploadedFileInfo]
     message: str
     skipped_files: list[str] = Field(default_factory=list)
 
 
+class UploadListResponse(BaseModel):
+    """Response model for uploaded file listing."""
+
+    files: list[UploadedFileInfo]
+    count: int
+
+
 class UploadLimits(BaseModel):
     """Application-level upload limits exposed to clients."""
 
@@ -69,11 +93,30 @@ def _make_file_sandbox_writable(file_path: os.PathLike[str] | str) -> None:
         logger.warning("Skipping sandbox chmod for symlinked upload path: %s", file_path)
         return
 
-    writable_mode = stat.S_IMODE(file_stat.st_mode) | stat.S_IWUSR | stat.S_IWGRP | stat.S_IWOTH
+    writable_mode = stat.S_IMODE(file_stat.st_mode) | stat.S_IWUSR | stat.S_IWGRP | stat.S_IWOTH | stat.S_IRGRP | stat.S_IROTH
     chmod_kwargs = {"follow_symlinks": False} if os.chmod in os.supports_follow_symlinks else {}
     os.chmod(file_path, writable_mode, **chmod_kwargs)
 
 
+def _make_file_sandbox_readable(file_path: os.PathLike[str] | str) -> None:
+    """Ensure uploaded files are readable by the sandbox process.
+
+    For Docker sandboxes (AIO), the gateway writes files as root with 0o600
+    permissions, then bind-mounts the host directory into the container. The
+    sandbox process inside the container runs as a non-root user and cannot
+    read those files without group/other read bits. This function adds
+    ``S_IRGRP | S_IROTH`` so the sandbox can read the uploaded content.
+    """
+    file_stat = os.lstat(file_path)
+    if stat.S_ISLNK(file_stat.st_mode):
+        logger.warning("Skipping sandbox chmod for symlinked upload path: %s", file_path)
+        return
+
+    readable_mode = stat.S_IMODE(file_stat.st_mode) | stat.S_IRGRP | stat.S_IROTH
+    chmod_kwargs = {"follow_symlinks": False} if os.chmod in os.supports_follow_symlinks else {}
+    os.chmod(file_path, readable_mode, **chmod_kwargs)
+
+
 def _uses_thread_data_mounts(sandbox_provider: SandboxProvider) -> bool:
     return bool(getattr(sandbox_provider, "uses_thread_data_mounts", False))
 
@@ -237,7 +280,7 @@ async def upload_files(
 
             file_info = {
                 "filename": safe_filename,
-                "size": str(file_size),
+                "size": file_size,
                 "path": str(sandbox_uploads / safe_filename),
                 "virtual_path": virtual_path,
                 "artifact_url": upload_artifact_url(thread_id, safe_filename),
@@ -276,6 +319,16 @@ async def upload_files(
             _cleanup_uploaded_paths(written_paths)
             raise HTTPException(status_code=500, detail=f"Failed to upload {file.filename}: {str(e)}")
 
+    # Uploaded files are created with 0o600 permissions (owner read/write only).
+    # In Docker sandbox deployments the gateway writes as root but the sandbox
+    # process runs as a non-root user (typically UID 1000).  Without group/other
+    # read bits the sandbox cannot access the files — whether the uploads
+    # directory is bind-mounted into the container or synced via
+    # sandbox.update_file.  Always add group/other read bits so every sandbox
+    # configuration can read the uploaded content.
+    for file_path in written_paths:
+        _make_file_sandbox_readable(file_path)
+
     if sync_to_sandbox:
         for file_path, virtual_path in sandbox_sync_targets:
             _make_file_sandbox_writable(file_path)
@@ -304,9 +357,9 @@ async def get_upload_limits(
     return _get_upload_limits(config)
 
 
-@router.get("/list", response_model=dict)
+@router.get("/list", response_model=UploadListResponse)
 @require_permission("threads", "read", owner_check=True)
-async def list_uploaded_files(thread_id: str, request: Request) -> dict:
+async def list_uploaded_files(thread_id: str, request: Request) -> UploadListResponse:
     """List all files in a thread's uploads directory."""
     try:
         uploads_dir = get_uploads_dir(thread_id)
@@ -320,7 +373,7 @@ async def list_uploaded_files(thread_id: str, request: Request) -> dict:
     for f in result["files"]:
         f["path"] = str(sandbox_uploads / f["filename"])
 
-    return result
+    return UploadListResponse(**result)
 
 
 @router.delete("/{filename}")

backend/app/gateway/services.py

@@ -15,9 +15,11 @@ from collections.abc import Mapping
 from typing import Any
 
 from fastapi import HTTPException, Request
-from langchain_core.messages import HumanMessage
+from langchain_core.messages import BaseMessage
+from langchain_core.messages.utils import convert_to_messages
 
 from app.gateway.deps import get_run_context, get_run_manager, get_stream_bridge
+from app.gateway.internal_auth import INTERNAL_SYSTEM_ROLE
 from app.gateway.utils import sanitize_log_param
 from deerflow.config.app_config import get_app_config
 from deerflow.runtime import (
@@ -32,6 +34,7 @@ from deerflow.runtime import (
     UnsupportedStrategyError,
     run_agent,
 )
+from deerflow.runtime.runs.naming import resolve_root_run_name
 
 logger = logging.getLogger(__name__)
 
@@ -75,21 +78,35 @@ def normalize_stream_modes(raw: list[str] | str | None) -> list[str]:
 
 
 def normalize_input(raw_input: dict[str, Any] | None) -> dict[str, Any]:
-    """Convert LangGraph Platform input format to LangChain state dict."""
+    """Convert LangGraph Platform input format to LangChain state dict.
+
+    Delegates dict→message coercion to ``langchain_core.messages.utils.convert_to_messages``
+    so that ``additional_kwargs`` (e.g. uploaded-file metadata — gh #3132), ``id``,
+    ``name``, and non-human roles (ai/system/tool) survive unchanged.  An earlier
+    hand-rolled version only forwarded ``content`` and collapsed every role to
+    ``HumanMessage``, which silently stripped frontend-supplied attachments.
+
+    Malformed message dicts (missing ``role``/``type``/``content``, unsupported
+    role, etc.) raise ``HTTPException(400)`` with the offending index, instead
+    of bubbling up as a 500.  The gateway is a system boundary, so per-entry
+    validation errors are the right shape for clients to retry against.
+    """
     if raw_input is None:
         return {}
     messages = raw_input.get("messages")
     if messages and isinstance(messages, list):
-        converted = []
-        for msg in messages:
-            if isinstance(msg, dict):
-                role = msg.get("role", msg.get("type", "user"))
-                content = msg.get("content", "")
-                if role in ("user", "human"):
-                    converted.append(HumanMessage(content=content))
-                else:
-                    # TODO: handle other message types (system, ai, tool)
-                    converted.append(HumanMessage(content=content))
+        converted: list[Any] = []
+        for index, msg in enumerate(messages):
+            if isinstance(msg, BaseMessage):
+                converted.append(msg)
+            elif isinstance(msg, dict):
+                try:
+                    converted.extend(convert_to_messages([msg]))
+                except (ValueError, TypeError, NotImplementedError) as exc:
+                    raise HTTPException(
+                        status_code=400,
+                        detail=f"Invalid message at input.messages[{index}]: {exc}",
+                    ) from exc
             else:
                 converted.append(msg)
         return {**raw_input, "messages": converted}
@@ -124,7 +141,14 @@ def merge_run_context_overrides(config: dict[str, Any], context: Mapping[str, An
     """Merge whitelisted keys from ``body.context`` into both ``config['configurable']``
     and ``config['context']`` so they are visible to legacy configurable readers and
     to LangGraph ``ToolRuntime.context`` consumers (e.g. the ``setup_agent`` tool —
-    see issue #2677)."""
+    see issue #2677).
+
+    ``user_id`` is intentionally propagated into ``config['context']`` in addition to
+    the whitelisted keys, so non-web callers (e.g. IM channels) that supply identity in
+    ``body.context`` keep it on ``ToolRuntime.context``. It is merged with
+    ``setdefault`` so a server-authenticated id stamped by
+    :func:`inject_authenticated_user_context` always wins over the client-supplied one.
+    """
     if not context:
         return
     configurable = config.setdefault("configurable", {})
@@ -135,6 +159,8 @@ def merge_run_context_overrides(config: dict[str, Any], context: Mapping[str, An
                 configurable.setdefault(key, context[key])
             if isinstance(runtime_context, dict):
                 runtime_context.setdefault(key, context[key])
+    if "user_id" in context and isinstance(runtime_context, dict):
+        runtime_context.setdefault("user_id", context["user_id"])
 
 
 def inject_authenticated_user_context(config: dict[str, Any], request: Request) -> None:
@@ -150,6 +176,9 @@ def inject_authenticated_user_context(config: dict[str, Any], request: Request)
     if user_id is None:
         return
 
+    if getattr(user, "system_role", None) == INTERNAL_SYSTEM_ROLE:
+        return
+
     runtime_context = config.setdefault("context", {})
     if isinstance(runtime_context, dict):
         runtime_context["user_id"] = str(user_id)
@@ -235,6 +264,7 @@ def build_run_config(
             target = config.setdefault("configurable", {})
         if target is not None and "agent_name" not in target:
             target["agent_name"] = normalized
+        config.setdefault("run_name", resolve_root_run_name(config, normalized))
     if metadata:
         config.setdefault("metadata", {}).update(metadata)
     return config
@@ -385,3 +415,51 @@ async def sse_consumer(
         if record.status in (RunStatus.pending, RunStatus.running):
             if record.on_disconnect == DisconnectMode.cancel:
                 await run_mgr.cancel(record.run_id)
+
+
+async def wait_for_run_completion(
+    bridge: StreamBridge,
+    record: RunRecord,
+    request: Request,
+    run_mgr: RunManager,
+) -> bool:
+    """Block until the run publishes ``END_SENTINEL``, honouring on_disconnect.
+
+    The non-streaming ``/wait`` endpoints used to ``await record.task``
+    directly with no disconnect handling.  When the client (or an
+    intermediate HTTP proxy) timed out during a long tool call such as
+    ``pip install``, the handler would swallow ``CancelledError`` and
+    serialize whatever checkpoint happened to exist — masking a half-finished
+    run as a normal completion (issue #3265).
+
+    This helper consumes the same bridge that ``sse_consumer`` does so the
+    wait path shares its disconnect semantics: each wake-up polls
+    ``request.is_disconnected()``; on a real disconnect it cancels the
+    background run when ``record.on_disconnect`` is ``cancel``.  The bridge's
+    heartbeat sentinels guarantee at least one wake-up per
+    ``heartbeat_interval`` even when the agent emits no events for a while.
+
+    Returns:
+        ``True`` when ``END_SENTINEL`` was observed (run reached a terminal
+        state), ``False`` when the loop exited because the client
+        disconnected.  Callers must skip checkpoint serialization on
+        ``False`` so a partial checkpoint is not returned as a normal
+        response.
+    """
+    completed = False
+    try:
+        async for entry in bridge.subscribe(record.run_id):
+            # END_SENTINEL means the run reached a terminal state; honour it
+            # even if the client just disconnected so the caller still serializes
+            # the real final checkpoint.
+            if entry is END_SENTINEL:
+                completed = True
+                return True
+            if await request.is_disconnected():
+                break
+            # Heartbeats and regular events: keep waiting for END_SENTINEL.
+        return completed
+    finally:
+        if not completed and record.status in (RunStatus.pending, RunStatus.running):
+            if record.on_disconnect == DisconnectMode.cancel:
+                await run_mgr.cancel(record.run_id)

backend/docs/API.md

@@ -228,10 +228,13 @@ Get current MCP server configurations.
 GET /api/mcp/config
 ```
 
+Requires an authenticated admin session. Sensitive env/header/OAuth secret
+values are masked in the response.
+
 **Response:**
 ```json
 {
-  "mcpServers": {
+  "mcp_servers": {
     "github": {
       "enabled": true,
       "type": "stdio",
@@ -241,13 +244,6 @@ GET /api/mcp/config
         "GITHUB_TOKEN": "***"
       },
       "description": "GitHub operations"
-    },
-    "filesystem": {
-      "enabled": false,
-      "type": "stdio",
-      "command": "npx",
-      "args": ["-y", "@modelcontextprotocol/server-filesystem"],
-      "description": "File system access"
     }
   }
 }
@@ -262,10 +258,15 @@ PUT /api/mcp/config
 Content-Type: application/json
 ```
 
+Requires an authenticated admin session. API-managed `stdio` MCP servers may
+only use allowed executable names for `command` (default: `npx`, `uvx`). Set
+`DEER_FLOW_MCP_STDIO_COMMAND_ALLOWLIST` to a comma-separated list when a
+deployment needs additional trusted launchers.
+
 **Request Body:**
 ```json
 {
-  "mcpServers": {
+  "mcp_servers": {
     "github": {
       "enabled": true,
       "type": "stdio",
@@ -283,8 +284,18 @@ Content-Type: application/json
 **Response:**
 ```json
 {
-  "success": true,
-  "message": "MCP configuration updated"
+  "mcp_servers": {
+    "github": {
+      "enabled": true,
+      "type": "stdio",
+      "command": "npx",
+      "args": ["-y", "@modelcontextprotocol/server-github"],
+      "env": {
+        "GITHUB_TOKEN": "***"
+      },
+      "description": "GitHub operations"
+    }
+  }
 }
 ```
 

backend/docs/AUTH_TEST_DOCKER_GAP.md

@@ -29,7 +29,7 @@ All other test plan sections were executed against either:
 | TC-DOCKER-03 | Per-worker rate limiter divergence | Confirms in-process `_login_attempts` dict doesn't share state across `gunicorn` workers (4 by default in the compose file); known limitation, documented | needs multi-worker container |
 | TC-DOCKER-04 | IM channels use internal Gateway auth | Verify Feishu/Slack/Telegram dispatchers attach the process-local internal auth header plus CSRF cookie/header when calling Gateway-compatible LangGraph APIs | needs `docker logs` |
 | TC-DOCKER-05 | Reset credentials surfacing | `reset_admin` writes a 0600 credential file in `DEER_FLOW_HOME` instead of logging plaintext. The file-based behavior is validated by non-Docker reset tests, so the only Docker-specific gap is verifying the volume mount carries the file out to the host | needs container + host volume |
-| TC-DOCKER-06 | Gateway-mode Docker deploy | `./scripts/deploy.sh --gateway` produces a 3-container topology (no `langgraph` container); same auth flow as standard mode | needs `docker compose --profile gateway` |
+| TC-DOCKER-06 | Docker deploy uses Gateway embedded runtime | `./scripts/deploy.sh` produces a Gateway + frontend + nginx topology (no `langgraph` container); same auth flow as local `make dev` | needs `docker compose up` |
 
 ## Coverage already provided by non-Docker tests
 
@@ -43,7 +43,7 @@ the test cases that ran on sg_dev or local:
 | TC-DOCKER-03 (per-worker rate limit) | TC-GW-04 + TC-REENT-09 (single-worker rate limit + 5min expiry). The cross-worker divergence is an architectural property of the in-memory dict; no auth code path differs |
 | TC-DOCKER-04 (IM channels use internal auth) | Code-level: `app/channels/manager.py` creates the `langgraph_sdk` client with `create_internal_auth_headers()` plus CSRF cookie/header, so channel workers do not rely on browser cookies |
 | TC-DOCKER-05 (credential surfacing) | `reset_admin` writes `.deer-flow/admin_initial_credentials.txt` with mode 0600 and logs only the path — the only Docker-unique step is whether the bind mount projects this path onto the host, which is a `docker compose` config check, not a runtime behavior change |
-| TC-DOCKER-06 (gateway-mode container) | Section 七 7.2 covered by TC-GW-01..05 + Section 二 (gateway-mode auth flow on sg_dev) — same Gateway code, container is just a packaging change |
+| TC-DOCKER-06 (Gateway embedded runtime container) | Section 七 7.2 covered by TC-GW-01..05 + Section 二 (Gateway auth flow on sg_dev) — same Gateway code, container is just a packaging change |
 
 ## Reproduction steps when Docker becomes available
 

backend/docs/AUTH_TEST_PLAN.md

@@ -4,10 +4,12 @@
 
 | 模式 | 启动命令 | Auth 层 | 端口 |
 |------|---------|---------|------|
-| 标准模式 | `make dev` | Gateway AuthMiddleware + LangGraph auth | 2026 (nginx) |
-| Gateway 模式 | `make dev-pro` | Gateway AuthMiddleware（全量） | 2026 (nginx) |
+| 标准模式 | `make dev` | Gateway AuthMiddleware（全量） | 2026 (nginx) |
 | 直连 Gateway | `cd backend && make gateway` | Gateway AuthMiddleware | 8001 |
-| 直连 LangGraph | `cd backend && make dev` | LangGraph auth | 2024 |
+| 直连 LangGraph 兼容性 | 手动运行 LangGraph 工具链时使用 | LangGraph auth | 2024 |
+
+`make dev`、Docker dev 和生产部署默认都运行 Gateway embedded runtime。
+`app.gateway.langgraph_auth` 仅用于保留的直连 LangGraph 工具链 / Studio 兼容性测试，不是标准服务启动路径。
 
 每种模式下都需执行以下测试。
 
@@ -21,10 +23,8 @@
 # 清除已有数据
 rm -f backend/.deer-flow/data/deerflow.db
 
-# 选择模式启动
-make dev          # 标准模式
-# 或
-make dev-pro      # Gateway 模式
+# 启动标准模式（Gateway embedded runtime）
+make dev
 ```
 
 **验证点：**
@@ -57,7 +57,7 @@ make dev
 
 ## 二、接口流程测试
 
-> 以下用 `BASE=http://localhost:2026` 为例。标准模式和 Gateway 模式都用此地址。
+> 以下用 `BASE=http://localhost:2026` 为例。标准模式经 nginx 暴露此地址。
 > 直连测试替换为对应端口。
 >
 > **CSRF token 提取**：多处用到从 cookie jar 提取 CSRF token，统一使用：
@@ -211,20 +211,18 @@ curl -s -X POST $BASE/api/threads/search \
 
 **预期：** 返回 0 或仅包含 user2 自己的 thread
 
-### 2.3 标准模式 LangGraph Server 隔离
+### 2.3 LangGraph-compatible Gateway 路由隔离
 
-> 仅在标准模式下测试。Gateway 模式不跑 LangGraph Server。
-
-#### TC-API-10: LangGraph 端点需要 cookie
+#### TC-API-10: LangGraph-compatible 端点需要 cookie
 
 ```bash
-# 不带 cookie 访问 LangGraph 接口
+# 不带 cookie 访问 LangGraph-compatible 接口
 curl -s -w "%{http_code}" $BASE/api/langgraph/threads
 ```
 
 **预期：** 401
 
-#### TC-API-11: LangGraph 带 cookie 可访问
+#### TC-API-11: LangGraph-compatible 路由带 cookie 可访问
 
 ```bash
 curl -s $BASE/api/langgraph/threads -b user1.txt | jq length
@@ -232,10 +230,10 @@ curl -s $BASE/api/langgraph/threads -b user1.txt | jq length
 
 **预期：** 200，返回 user1 的 thread 列表
 
-#### TC-API-12: LangGraph 隔离 — 用户只看到自己的
+#### TC-API-12: LangGraph-compatible 路由隔离 — 用户只看到自己的
 
 ```bash
-# user2 查 LangGraph threads
+# user2 查 threads
 curl -s $BASE/api/langgraph/threads -b user2.txt | jq length
 ```
 
@@ -1234,21 +1232,11 @@ P2=$(awk -F': ' '/^password:/ {print $2}' /tmp/deerflow-reset-p2.txt)
 ## 七、模式差异测试
 
 > 以下用 `GW=http://localhost:8001` 表示直连 Gateway，`BASE=http://localhost:2026` 表示经 nginx。
-> Gateway 模式启动命令：`make dev-pro`（或 `./scripts/serve.sh --dev --gateway`）。
+> 标准启动命令：`make dev`（或 `./scripts/serve.sh --dev`）。
 
-### 7.1 标准模式独有
+### 7.1 标准启动模式
 
-> 启动命令：`make dev`（或 `./scripts/serve.sh --dev`）
-
-#### TC-MODE-01: LangGraph Server 独立运行，需 cookie
-
-```bash
-# 无 cookie 访问 LangGraph
-curl -s -w "%{http_code}" -o /dev/null $BASE/api/langgraph/threads/search
-# 预期: 403（LangGraph auth handler 拒绝）
-```
-
-#### TC-MODE-02: LangGraph auth 的 token_version 检查
+#### TC-MODE-01: Gateway AuthMiddleware 的 token_version 检查
 
 ```bash
 # 登录拿 cookie
@@ -1261,9 +1249,9 @@ curl -s -X POST $BASE/api/v1/auth/change-password \
   -b cookies.txt -H "Content-Type: application/json" -H "X-CSRF-Token: $CSRF" \
   -d '{"current_password":"正确密码","new_password":"NewPass1!"}' -c new_cookies.txt
 
-# 用旧 cookie 访问 LangGraph
+# 用旧 cookie 访问 LangGraph-compatible 路由
 curl -s -w "%{http_code}" $BASE/api/langgraph/threads/search -b cookies.txt
-# 预期: 403（token_version 不匹配）
+# 预期: 401（token_version 不匹配）
 
 # 用新 cookie 访问
 CSRF2=$(grep csrf_token new_cookies.txt | awk '{print $NF}')
@@ -1272,7 +1260,7 @@ curl -s -w "%{http_code}" -X POST $BASE/api/langgraph/threads/search \
 # 预期: 200
 ```
 
-#### TC-MODE-03: LangGraph auth 的 owner filter 隔离
+#### TC-MODE-02: Gateway owner filter 隔离
 
 ```bash
 # user1 创建 thread
@@ -1297,18 +1285,9 @@ print('OK: user2 sees', len(threads), 'threads, none belong to user1')
 "
 ```
 
-### 7.2 Gateway 模式独有
-
-> 启动命令：`make dev-pro`（或 `./scripts/serve.sh --dev --gateway`）
-> 无 LangGraph Server 进程，agent runtime 嵌入 Gateway。
-
-#### TC-MODE-04: 所有请求经 AuthMiddleware
+#### TC-MODE-03: 所有请求经 AuthMiddleware
 
 ```bash
-# 确认 LangGraph Server 未运行
-curl -s -w "%{http_code}" -o /dev/null http://localhost:2024/ok
-# 预期: 000（连接被拒）
-
 # Gateway API 受保护
 curl -s -w "%{http_code}" -o /dev/null $BASE/api/models
 # 预期: 401
@@ -1319,7 +1298,7 @@ curl -s -w "%{http_code}" -o /dev/null -X POST $BASE/api/langgraph/threads/searc
 # 预期: 401
 ```
 
-#### TC-MODE-05: Gateway 模式下完整 auth 流程
+#### TC-MODE-04: 标准模式下完整 auth 流程
 
 ```bash
 # 登录
@@ -1334,7 +1313,7 @@ curl -s -X POST $BASE/api/langgraph/threads \
   -d '{"metadata":{}}' | python3 -c "import sys,json; print(json.load(sys.stdin)['thread_id'])"
 # 预期: 返回 thread_id
 
-# CSRF 保护（Gateway 模式下 CSRFMiddleware 直接覆盖所有路由）
+# CSRF 保护（CSRFMiddleware 覆盖所有 Gateway 路由）
 curl -s -w "%{http_code}" -o /dev/null -X POST $BASE/api/langgraph/threads \
   -b cookies.txt -H "Content-Type: application/json" -d '{"metadata":{}}'
 # 预期: 403（CSRF token missing）
@@ -1433,7 +1412,7 @@ done
 
 ### 7.4 Docker 部署
 
-> 启动命令：`./scripts/deploy.sh`（标准）或 `./scripts/deploy.sh --gateway`（Gateway 模式）
+> 启动命令：`./scripts/deploy.sh`
 > Docker Compose 文件：`docker/docker-compose.yaml`
 >
 > 前置条件：
@@ -1542,16 +1521,16 @@ docker logs deer-flow-gateway 2>&1 | grep -iE "Password: .{15,}" && echo "FAIL:
 - 容器日志输出**路径**（不是密码本身），符合 CodeQL `py/clear-text-logging-sensitive-data` 规则
 - `grep "Password:"` 在日志中**应当无匹配**（旧行为已废弃，simplify pass 移除了日志泄露路径）
 
-#### TC-DOCKER-06: Gateway 模式 Docker 部署
+#### TC-DOCKER-06: Docker 部署
 
 ```bash
-# Gateway 模式：无 langgraph 容器
-./scripts/deploy.sh --gateway
+# 标准 Docker 模式：runtime 嵌入 gateway 容器
+./scripts/deploy.sh
 sleep 15
 
-# 确认 langgraph 容器不存在
-docker ps --filter name=deer-flow-langgraph --format '{{.Names}}' | wc -l
-# 预期: 0
+# 确认 gateway 容器存在
+docker ps --filter name=deer-flow-gateway --format '{{.Names}}'
+# 预期: deer-flow-gateway
 
 # auth 流程正常：未登录受保护接口返回 401
 curl -s -w "%{http_code}" -o /dev/null $BASE/api/models

backend/docs/AUTH_UPGRADE.md

@@ -124,8 +124,8 @@ python -c "import secrets; print(secrets.token_urlsafe(32))"
 
 ## 兼容性
 
-- **标准模式**（`make dev`）：完全兼容；无 admin 时访问 `/setup` 初始化
-- **Gateway 模式**（`make dev-pro`）：完全兼容
+- **本地开发**（`make dev`）：Gateway embedded runtime 完全兼容；无 admin 时访问 `/setup` 初始化
+- **Gateway embedded runtime**：标准脚本、Docker dev 和生产部署均通过 Gateway 提供认证与 LangGraph-compatible API
 - **Docker 部署**：完全兼容，`.deer-flow/data/deerflow.db` 需持久化卷挂载
 - **IM 渠道**（Feishu/Slack/Telegram）：通过 Gateway 内部认证通信，使用 `default` 用户桶
 - **DeerFlowClient**（嵌入式）：不经过 HTTP，不受认证影响

backend/docs/BLOCKING_IO_DETECTION.md

@@ -0,0 +1,154 @@
+# Blocking IO detection usage and maintenance
+
+This document describes how to use and maintain DeerFlow backend blocking-IO
+detection for async event-loop safety.
+
+The goal is narrow: find and prevent synchronous IO from blocking backend
+async event-loop paths. Static and runtime detection are complementary, but
+they have different jobs.
+
+## Static detector
+
+The static detector is the discovery tool. It scans backend source code and
+reports candidate blocking-IO call sites that may need human review.
+
+Run it from the repository root:
+
+```bash
+make detect-blocking-io
+```
+
+Or from `backend/`:
+
+```bash
+make detect-blocking-io
+```
+
+The report is written to:
+
+```text
+.deer-flow/blocking-io-findings.json
+```
+
+Use this output for review and triage. A static finding is a candidate, not
+proof that production blocks the event loop at runtime. The current static
+rules are intentionally broad; prefer triaging existing output before adding
+new static rules.
+
+Add a static rule only when review finds a recurring high-risk blocking
+pattern that is invisible to the current detector.
+
+## Runtime detector
+
+The runtime detector is the CI regression guard. It uses Blockbuster to fail a
+focused test when code under `app.*` or `deerflow.*` performs blocking IO on
+the asyncio event-loop thread.
+
+Run it from `backend/`:
+
+```bash
+make test-blocking-io
+```
+
+The runtime gate starts from confirmed production bugs and protects those
+paths from regressing. It does not prove that the entire backend is free of
+blocking IO; it only covers the production paths exercised by
+`backend/tests/blocking_io/`.
+
+## Maintenance workflow
+
+Use the static detector to find candidates, then use review to decide which
+async production paths are worth protecting in CI.
+
+The normal workflow is:
+
+1. Run the static detector to find backend blocking-IO candidates.
+2. Use human review to pick high-risk production async paths.
+3. Add or update a focused runtime anchor in `backend/tests/blocking_io/`.
+4. Let CI prevent that path from regressing.
+
+Runtime detection has two maintenance paths.
+
+### Add a runtime rule
+
+Add a runtime rule when Blockbuster's default rules do not cover a generic
+blocking primitive used by production code.
+
+Rules belong in:
+
+```text
+backend/tests/support/detectors/blocking_io_runtime.py
+```
+
+Add them to `_PROJECT_BLOCKING_RULES`, not directly inside individual tests.
+Keeping rules centralized makes it clear which extra primitives DeerFlow
+expects Blockbuster to catch.
+
+Example shape:
+
+```python
+import subprocess
+
+from blockbuster import BlockBusterFunction
+
+_PROJECT_BLOCKING_RULES = (
+    (
+        "subprocess.Popen.__init__",
+        BlockBusterFunction(
+            subprocess.Popen,
+            "__init__",
+            scanned_modules=["app", "deerflow"],
+        ),
+    ),
+)
+```
+
+Do not add a runtime rule just because a business path is not tested. A rule
+only expands what Blockbuster can intercept after code runs.
+
+### Add a runtime anchor
+
+Add a runtime anchor when a high-risk async production path should be protected
+by CI but no existing `backend/tests/blocking_io/` test executes it.
+
+Anchors belong in:
+
+```text
+backend/tests/blocking_io/
+```
+
+A good anchor should:
+
+- Call the real production async entry point.
+- Avoid bypassing the blocking surface with test-only `asyncio.to_thread`
+  wrappers.
+- Use real local filesystem inputs when the bug shape is filesystem IO.
+- Mock only the external dependency boundary, such as a network service or
+  third-party saver class.
+- Fail if a future change moves the blocking operation back onto the event
+  loop.
+
+Avoid testing only the low-level helper unless that helper is the production
+async entry point. The runtime gate is most useful when it protects the caller
+that production actually executes.
+
+## Current runtime coverage
+
+The runtime anchors protect confirmed blocking-IO bug shapes:
+
+- SQLite checkpointer setup, including path resolution and parent-directory
+  creation.
+- Subagent skill metadata loading through `SubagentExecutor._load_skills()`.
+- `JsonlRunEventStore` async API (`put` / `list_*` / `delete_*`): the JSONL
+  run-event backend offloads its synchronous file IO via `asyncio.to_thread`
+  (fix #3084); this anchor drives the real async API under the gate so any
+  blocking IO reintroduced on the loop fails, not only removal of one
+  `to_thread` call.
+- `UploadsMiddleware.before_agent` uploads-directory scan: a sync-only middleware
+  hook runs on the event loop under async graph execution, so the scan is
+  offloaded via `abefore_agent` + `run_in_executor`.
+- Gate health checks: Blockbuster catches unoffloaded calls, opt-out works, and
+  patches are restored after exceptions.
+
+As static detection and review identify more high-risk async paths, add new
+runtime anchors incrementally.

backend/docs/CONFIGURATION.md

@@ -36,6 +36,7 @@ models:
 - OpenAI (`langchain_openai:ChatOpenAI`)
 - Anthropic (`langchain_anthropic:ChatAnthropic`)
 - DeepSeek (`langchain_deepseek:ChatDeepSeek`)
+- Xiaomi MiMo (`deerflow.models.patched_mimo:PatchedChatMiMo`)
 - Claude Code OAuth (`deerflow.models.claude_provider:ClaudeChatModel`)
 - Codex CLI (`deerflow.models.openai_codex_provider:CodexChatModel`)
 - Any LangChain-compatible provider
@@ -94,25 +95,35 @@ models:
         thinking:
           type: enabled
 
-  - name: minimax-m2.5
-    display_name: MiniMax M2.5
+  - name: minimax-m3
+    display_name: MiniMax M3
     use: langchain_openai:ChatOpenAI
-    model: MiniMax-M2.5
+    model: MiniMax-M3
     api_key: $MINIMAX_API_KEY
     base_url: https://api.minimax.io/v1
     max_tokens: 4096
     temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
     supports_vision: true
 
-  - name: minimax-m2.5-highspeed
-    display_name: MiniMax M2.5 Highspeed
+  - name: minimax-m2.7
+    display_name: MiniMax M2.7
     use: langchain_openai:ChatOpenAI
-    model: MiniMax-M2.5-highspeed
+    model: MiniMax-M2.7
     api_key: $MINIMAX_API_KEY
     base_url: https://api.minimax.io/v1
     max_tokens: 4096
     temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
-    supports_vision: true
+    supports_vision: false  # M2.7 is text-only; M3 supports vision
+
+  - name: minimax-m2.7-highspeed
+    display_name: MiniMax M2.7 Highspeed
+    use: langchain_openai:ChatOpenAI
+    model: MiniMax-M2.7-highspeed
+    api_key: $MINIMAX_API_KEY
+    base_url: https://api.minimax.io/v1
+    max_tokens: 4096
+    temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
+    supports_vision: false  # M2.7 is text-only; M3 supports vision
   - name: openrouter-gemini-2.5-flash
     display_name: Gemini 2.5 Flash (OpenRouter)
     use: langchain_openai:ChatOpenAI
@@ -166,6 +177,37 @@ models:
 
 For Gemini accessed **without** thinking (e.g. via OpenRouter where thinking is not activated), the plain `langchain_openai:ChatOpenAI` with `supports_thinking: false` is sufficient and no patch is needed.
 
+**MiMo with thinking via OpenAI-compatible API**:
+
+MiMo returns `reasoning_content` on assistant messages in thinking mode. In multi-turn agent conversations with tool calls, subsequent requests must preserve that historical `reasoning_content` on assistant messages or the MiMo API can return HTTP 400. Standard `langchain_openai:ChatOpenAI` drops this provider-specific field, so use `deerflow.models.patched_mimo:PatchedChatMiMo`:
+
+For pay-as-you-go API keys (`sk-...`), use `https://api.xiaomimimo.com/v1`. For Token Plan keys (`tp-...`), use the regional Token Plan Base URL shown in the MiMo console, such as `https://token-plan-cn.xiaomimimo.com/v1`. MiMo documents these key types as separate and non-interchangeable.
+
+`PatchedChatMiMo` is model-id agnostic. Use it for every MiMo thinking model entry you configure, including model entries referenced by `subagents.*.model` overrides (for example `mimo-v2.5-pro`, `mimo-v2.5`, `mimo-v2-pro`, `mimo-v2-omni`, or `mimo-v2-flash`).
+
+```yaml
+models:
+  - name: mimo-v2.5-pro
+    display_name: MiMo V2.5 Pro
+    use: deerflow.models.patched_mimo:PatchedChatMiMo
+    model: mimo-v2.5-pro
+    api_key: $MIMO_API_KEY
+    base_url: https://api.xiaomimimo.com/v1
+    max_tokens: 8192
+    supports_thinking: true
+    supports_vision: false
+    when_thinking_enabled:
+      extra_body:
+        thinking:
+          type: enabled
+    when_thinking_disabled:
+      extra_body:
+        thinking:
+          type: disabled
+```
+
+`PatchedChatMiMo` preserves MiMo's `choices[].message.reasoning_content`, streaming `delta.reasoning_content`, and request-history assistant `reasoning_content` fields. It does not reuse the DeepSeek provider.
+
 ### Tool Groups
 
 Organize tools into logical groups:
@@ -319,6 +361,7 @@ models:
 - `OPENAI_API_KEY` - OpenAI API key
 - `ANTHROPIC_API_KEY` - Anthropic API key
 - `DEEPSEEK_API_KEY` - DeepSeek API key
+- `MIMO_API_KEY` - Xiaomi MiMo API key
 - `NOVITA_API_KEY` - Novita API key (OpenAI-compatible endpoint)
 - `TAVILY_API_KEY` - Tavily search API key
 - `DEER_FLOW_PROJECT_ROOT` - Project root for relative runtime paths

backend/docs/MCP_SERVER.md

@@ -14,6 +14,19 @@ DeerFlow supports configurable MCP servers and skills to extend its capabilities
 3. Configure each server’s command, arguments, and environment variables as needed.
 4. Restart the application to load and register MCP tools.
 
+## Filesystem MCP Servers
+
+DeerFlow already provides built-in file tools for thread-scoped workspace access.
+Do not add an MCP filesystem server for the same DeerFlow workspace. The
+overlapping file tools use different path semantics, which can make LLM tool
+selection and file access behavior unstable.
+
+DeerFlow does not currently adapt the MCP Roots mode for filesystem servers. In
+particular, it does not publish per-thread MCP roots or map DeerFlow sandbox
+paths such as `/mnt/user-data/...` to paths accepted by
+`@modelcontextprotocol/server-filesystem`. Use DeerFlow's built-in file tools
+for DeerFlow workspace files.
+
 ## OAuth Support (HTTP/SSE MCP Servers)
 
 For `http` and `sse` MCP servers, DeerFlow supports OAuth token acquisition and automatic token refresh.
@@ -88,7 +101,6 @@ MCP servers expose tools that are automatically discovered and integrated into D
 
 MCP servers can provide access to:
 
-- **File systems**
 - **Databases** (e.g., PostgreSQL)
 - **External APIs** (e.g., GitHub, Brave Search)
 - **Browser automation** (e.g., Puppeteer)
@@ -97,4 +109,4 @@ MCP servers can provide access to:
 ## Learn More
 
 For detailed documentation about the Model Context Protocol, visit:  
-https://modelcontextprotocol.io
+https://modelcontextprotocol.io

backend/docs/README.md

@@ -19,6 +19,7 @@ This directory contains detailed documentation for the DeerFlow backend.
 | [STREAMING.md](STREAMING.md) | Token-level streaming design: Gateway vs DeerFlowClient paths, `stream_mode` semantics, per-id dedup |
 | [FILE_UPLOAD.md](FILE_UPLOAD.md) | File upload functionality |
 | [PATH_EXAMPLES.md](PATH_EXAMPLES.md) | Path types and usage examples |
+| [SANDBOX_MEMORY_PROFILING.md](SANDBOX_MEMORY_PROFILING.md) | Sandbox memory baseline and runtime comparison guide |
 | [summarization.md](summarization.md) | Context summarization feature |
 | [plan_mode_usage.md](plan_mode_usage.md) | Plan mode with TodoList |
 | [AUTO_TITLE_GENERATION.md](AUTO_TITLE_GENERATION.md) | Automatic title generation |

backend/docs/REPLAY_E2E.md

@@ -0,0 +1,120 @@
+# Record/Replay E2E — front-back contract verification
+
+Deterministic, **key-free** end-to-end checks that a backend change can't
+silently break the frontend (and vice-versa). Two complementary layers, fed by a
+single recording.
+
+## Why
+
+The mock-based frontend e2e hand-writes the backend's JSON/SSE, so a backend
+schema or SSE change passes green ("fake green"). These layers replay a recorded
+**real** run against the **real** backend (and, for Layer 2, the real frontend),
+so contract drift turns the build red instead.
+
+## The two layers
+
+- **Layer 1 — backend golden** (`tests/test_replay_golden.py`): replays a fixture
+  through the real FastAPI gateway with `ReplayChatModel` and asserts the streamed
+  SSE event sequence equals a committed golden. Fast, no browser. Guards protocol
+  *shape*.
+- **Layer 2 — full-stack render** (`frontend/tests/e2e-real-backend/`): real
+  Next.js + real gateway (replay model) + Chromium; asserts the replayed
+  auto-title and a follow-up suggestion render in the browser. Guards semantic
+  *render*. (Complementary to Layer 1 — neither subsumes the other.)
+
+Layer 2 also hosts **cross-stack contract scenarios** — the dangerous class
+where a backend change silently breaks a frontend assumption and *both sides'
+unit tests stay green*. See below.
+
+## Cross-stack scenario: multi-run render order (`multi-run-order.spec.ts`)
+
+Regression guard for issue **#3352** (after context compression, refreshing a
+thread rendered history out of order). Root cause was a front-back desync:
+backend `RunManager.list_by_thread` returns runs **newest-first** (PR #2932),
+while the frontend (`core/threads/hooks.ts`) iterated runs and **prepended** each
+loaded page — inverting chronological order once the checkpoint no longer held
+the older messages. The backend ordering test was green throughout, and the
+frontend regression unit test hardcodes "backend returns newest-first" in a mock,
+so only a *real frontend against a real backend* catches the desync.
+
+This scenario does **not** record a conversation. It uses a **test-only seeder**
+(`tests/seed_runs_router.py`, mounted on the replay gateway only when
+`DEERFLOW_ENABLE_TEST_SEED=1`) to stand up a thread with ≥2 runs and per-run
+message events — and deliberately **no checkpoint**, which is the #3352
+precondition: it forces the frontend's per-run reload path to be the sole source
+of truth so the ordering bug becomes observable. The seeder writes through the
+gateway's own run/event stores using the request's auth context, so the real
+`list_by_thread` → `/runs/{id}/messages` → prepend path runs live. Reverting the
+#3354 frontend fix turns this spec red.
+
+## How replay works
+
+`tests/replay_provider.py::ReplayChatModel` returns recorded assistant turns keyed
+by a **normalized hash of the model caller + conversation**. The conversation is
+human / ai / tool messages — role, text, tool-call name+args; with
+`<system-reminder>`, dates, UUIDs, tmp paths stripped. The caller is the stable
+source of the model call (`lead_agent`, `middleware:title`, `suggest_agent`,
+`subagent:*`, etc.). A miss raises loudly rather than passing silently.
+
+**The system prompt is excluded from the match key.** The lead-agent system
+prompt is a living, frequently-edited implementation detail — its wording changes
+across PRs (e.g. #3195 added a "File Editing Workflow" section). Hashing it would
+make every fixture go stale and red-fail unrelated PRs the moment anyone edits the
+prompt. The conversation flow (user input → tool calls → results → answer) is the
+stable contract that identifies a recorded turn. The caller still stays in the
+key so two different model users with identical conversation text do not compete
+for the same replay bucket. (This mirrors how open-design's mock picker keys on
+the user prompt, not the system internals.) Combined with pinning skills +
+extensions empty and disabling memory/summarization
+(`tests/_replay_fixture.py::build_config_yaml`), a fixture replays the same across
+machines, days, prompt edits, and CI. Replaying needs **no API key**.
+
+A swallowed hash-miss keeps the SSE *event shapes* identical (the gateway wraps it
+into a normal assistant error message), so the Layer-1 golden can't catch a miss
+by shape alone — it inspects `replay_provider.replay_misses()` and fails loud
+instead. Layer-2 already fails on a miss (the recorded turns never render).
+
+## Record a new scenario (needs a real key — dev machine only)
+
+Recording drives the **real frontend** so captured inputs match exactly what the
+browser sends; fixtures contain no API key.
+
+```bash
+# 1. drive the real frontend against a real-model gateway, capturing model calls
+OPENAI_API_KEY=... OPENAI_API_BASE=<openai-compatible-endpoint>/v1 \
+  DEERFLOW_RECORD_OUT=/tmp/rec/turns.jsonl RECORD_MODEL=<model> \
+  bash -c 'cd frontend && pnpm exec playwright test -c playwright.record.config.ts'
+
+# 2. stitch the capture into a fixture
+cd backend && uv run python scripts/build_fixture_from_jsonl.py \
+  --jsonl /tmp/rec/turns.jsonl --meta /tmp/rec/turns.jsonl.meta.json \
+  --out tests/fixtures/replay/<scenario>.<mode>.json --model <model>
+
+# 3. regenerate the committed golden
+DEERFLOW_WRITE_GOLDEN=1 PYTHONPATH=. uv run pytest tests/test_replay_golden.py
+```
+
+## Run (no key)
+
+```bash
+cd backend  && PYTHONPATH=. uv run pytest tests/test_replay_golden.py          # Layer 1
+cd frontend && pnpm exec playwright test -c playwright.real-backend.config.ts  # Layer 2
+```
+
+## CI
+
+`.github/workflows/replay-e2e.yml` runs both layers on changes to **either** side
+of the contract (`frontend/**`, `backend/app/gateway/**`,
+`backend/packages/harness/**`, fixtures). DOM assertions are the gate; the rendered
+screenshot + Playwright HTML report are uploaded as a CI artifact.
+
+## Known limitations
+
+- Visual regression baselines are OS-specific, so they are a **local dev gate
+  only** (gitignored); CI uploads the render as an artifact for human review
+  instead of hard-asserting a cross-OS baseline.
+- Fixtures are coupled to the recording-time prompt; if new
+  environment-dependent content enters the system prompt, extend the
+  normalization in `replay_provider.py` (or pin it in `build_config_yaml`).
+- Re-record a scenario if the agent graph changes how many model calls it makes
+  — the replay raises loudly on a hash miss pointing at the divergence.

backend/docs/SANDBOX_MEMORY_PROFILING.md

@@ -0,0 +1,81 @@
+# Sandbox Memory Profiling
+
+This guide records a repeatable baseline before changing the sandbox runtime.
+Issue #3213 reports per-sandbox memory near 1 GiB in Kubernetes. Before adding
+or recommending a new provider, capture the current AIO sandbox baseline and
+compare candidates with the same DeerFlow workload.
+
+## What to Measure
+
+Measure at least these samples:
+
+1. Empty sandbox after it becomes ready.
+2. After a simple bash command.
+3. After a Python task that imports common packages.
+4. After a Node task when Node-based workloads are expected.
+5. After generating files under `/mnt/user-data/outputs`.
+6. After release and warm reuse.
+7. At the target concurrency level, for example 10, 50, or 100 sandboxes.
+
+`kubectl top` reports Kubernetes/container working set memory. Treat it as a
+capacity signal, not exclusive RSS/PSS. Pod-level memory includes every
+container in the Pod and may include cache charged to the cgroup. If a result
+looks surprising, inspect the sandbox processes and cgroup metrics on the node
+before drawing conclusions.
+
+## Capture a Snapshot
+
+Run this from the repository root:
+
+```bash
+python scripts/sandbox_memory_profile.py \
+  --namespace deer-flow \
+  --selector app=deer-flow-sandbox \
+  --sample empty \
+  --include-processes \
+  --format markdown
+```
+
+Use a descriptive `--sample` value for each phase:
+
+```bash
+python scripts/sandbox_memory_profile.py --sample after-bash --format json
+python scripts/sandbox_memory_profile.py --sample after-python --format json
+python scripts/sandbox_memory_profile.py --sample after-artifact --format json
+```
+
+`--include-processes` runs `kubectl exec ... ps` in each sandbox Pod and adds
+the highest-RSS processes to the report. This helps distinguish Pod-level cgroup
+memory from process RSS. The two numbers will not match exactly because cgroup
+memory can include cache and other kernel-accounted memory.
+
+Save the raw JSON when comparing backends so totals, pod names, images,
+requests, limits, and timestamps can be audited later.
+
+## Candidate Runtime Matrix
+
+For AIO, CubeSandbox, OpenSandbox, gVisor, Kata, or another candidate, compare
+the same workload and record:
+
+| Area | Required Evidence |
+| --- | --- |
+| Capacity | Pod or instance count, total memory, average memory, max memory |
+| Startup | Ready latency at 1, 10, 50, and 100 concurrent sandboxes |
+| Commands | Bash output, timeout behavior, failure shape |
+| Files | `read_file`, `write_file`, binary `update_file`, `list_dir`, `glob`, `grep` |
+| Uploads | Files uploaded by the gateway are visible inside the sandbox |
+| Artifacts | Files written to `/mnt/user-data/outputs` are readable by the backend artifact API |
+| Paths | `/mnt/user-data/workspace`, `/mnt/user-data/uploads`, `/mnt/user-data/outputs`, `/mnt/acp-workspace`, and skills paths keep their expected semantics |
+| Isolation | Different users and threads cannot read each other's data |
+| Cleanup | Release, idle timeout, process restart, and orphan cleanup free resources |
+| Operations | Deployment prerequisites, privileged components, networking, storage, and upgrade path |
+
+## PR Guidance
+
+Do not claim that a new provider fixes high-concurrency memory usage until the
+same DeerFlow workload has been measured on both the current AIO sandbox and the
+candidate backend.
+
+For an experimental provider PR, prefer `Related to #3213` unless the PR also
+includes reproducible DeerFlow workload data that demonstrates the target memory
+reduction and preserves uploads, outputs, artifacts, and isolation behavior.

backend/docs/TODO.md

@@ -26,7 +26,7 @@
   - Replace sync `requests` with `httpx.AsyncClient` in community tools (tavily, jina_ai, firecrawl, infoquest, image_search)
   - [x] Replace sync `model.invoke()` with async `model.ainvoke()` in title_middleware and memory updater
   - Consider `asyncio.to_thread()` wrapper for remaining blocking file I/O
-  - For production: use `langgraph up` (multi-worker) instead of `langgraph dev` (single-worker)
+  - For production: tune Gateway worker/runtime settings for long-running agent workloads
 
 ## Resolved Issues
 

backend/docs/middleware-execution-flow.md

@@ -4,22 +4,22 @@
 
 `create_deerflow_agent` 通过 `RuntimeFeatures` 组装的完整 middleware 链（默认全开时）：
 
-| # | Middleware | `before_agent` | `before_model` | `after_model` | `after_agent` | `wrap_tool_call` | 主 Agent | Subagent | 来源 |
-|---|-----------|:-:|:-:|:-:|:-:|:-:|:-:|:-:|------|
-| 0 | ThreadDataMiddleware | ✓ | | | | | ✓ | ✓ | `sandbox` |
-| 1 | UploadsMiddleware | ✓ | | | | | ✓ | ✗ | `sandbox` |
-| 2 | SandboxMiddleware | ✓ | | | ✓ | | ✓ | ✓ | `sandbox` |
-| 3 | DanglingToolCallMiddleware | | | ✓ | | | ✓ | ✗ | 始终开启 |
-| 4 | GuardrailMiddleware | | | | | ✓ | ✓ | ✓ | *Phase 2 纳入* |
-| 5 | ToolErrorHandlingMiddleware | | | | | ✓ | ✓ | ✓ | 始终开启 |
-| 6 | SummarizationMiddleware | | | ✓ | | | ✓ | ✗ | `summarization` |
-| 7 | TodoMiddleware | | | ✓ | | | ✓ | ✗ | `plan_mode` 参数 |
-| 8 | TitleMiddleware | | | ✓ | | | ✓ | ✗ | `auto_title` |
-| 9 | MemoryMiddleware | | | | ✓ | | ✓ | ✗ | `memory` |
-| 10 | ViewImageMiddleware | | ✓ | | | | ✓ | ✗ | `vision` |
-| 11 | SubagentLimitMiddleware | | | ✓ | | | ✓ | ✗ | `subagent` |
-| 12 | LoopDetectionMiddleware | | | ✓ | | | ✓ | ✗ | 始终开启 |
-| 13 | ClarificationMiddleware | | | ✓ | | | ✓ | ✗ | 始终最后 |
+| # | Middleware | `before_agent` | `before_model` | `after_model` | `after_agent` | `wrap_model_call` | `wrap_tool_call` | 主 Agent | Subagent | 来源 |
+|---|-----------|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|------|
+| 0 | ThreadDataMiddleware | ✓ | | | | | | ✓ | ✓ | `sandbox` |
+| 1 | UploadsMiddleware | ✓ | | | | | | ✓ | ✗ | `sandbox` |
+| 2 | SandboxMiddleware | ✓ | | | ✓ | | | ✓ | ✓ | `sandbox` |
+| 3 | DanglingToolCallMiddleware | | | | | ✓ | | ✓ | ✗ | 始终开启 |
+| 4 | GuardrailMiddleware | | | | | | ✓ | ✓ | ✓ | *Phase 2 纳入* |
+| 5 | ToolErrorHandlingMiddleware | | | | | | ✓ | ✓ | ✓ | 始终开启 |
+| 6 | SummarizationMiddleware | | ✓ | | | | | ✓ | ✗ | `summarization` |
+| 7 | TodoMiddleware | | ✓ | ✓ | | ✓ | | ✓ | ✗ | `plan_mode` 参数 |
+| 8 | TitleMiddleware | | | ✓ | | | | ✓ | ✗ | `auto_title` |
+| 9 | MemoryMiddleware | | | | ✓ | | | ✓ | ✗ | `memory` |
+| 10 | ViewImageMiddleware | | ✓ | | | | | ✓ | ✗ | `vision` |
+| 11 | SubagentLimitMiddleware | | | ✓ | | | | ✓ | ✗ | `subagent` |
+| 12 | LoopDetectionMiddleware | ✓ | | ✓ | ✓ | ✓ | | ✓ | ✗ | 始终开启 |
+| 13 | ClarificationMiddleware | | | | | | ✓ | ✓ | ✗ | 始终最后 |
 
 主 agent **14 个** middleware（`make_lead_agent`），subagent **4 个**（ThreadData、Sandbox、Guardrail、ToolErrorHandling）。`create_deerflow_agent` Phase 1 实现 **13 个**（Guardrail 仅支持自定义实例，无内置默认）。
 
@@ -35,7 +35,7 @@ graph TB
 
     subgraph BA ["<b>before_agent</b> 正序 0→N"]
         direction TB
-        TD["[0] ThreadData<br/>创建线程目录"] --> UL["[1] Uploads<br/>扫描上传文件"] --> SB["[2] Sandbox<br/>获取沙箱"]
+        TD["[0] ThreadData<br/>创建线程目录"] --> UL["[1] Uploads<br/>扫描上传文件"] --> SB["[2] Sandbox<br/>获取沙箱"] --> LD_BA["[12] LoopDetection<br/>清理 stale warning"]
     end
 
     subgraph BM ["<b>before_model</b> 正序 0→N"]
@@ -43,34 +43,42 @@ graph TB
         VI["[10] ViewImage<br/>注入图片 base64"]
     end
 
-    SB --> VI
-    VI --> M["<b>MODEL</b>"]
+    subgraph WM ["<b>wrap_model_call</b>"]
+        direction TB
+        DTC_WM["[3] DanglingToolCall<br/>补悬空 ToolMessage"] --> LD_WM["[12] LoopDetection<br/>注入当前 run warning"]
+    end
+
+    LD_BA --> VI
+    VI --> DTC_WM
+    LD_WM --> M["<b>MODEL</b>"]
 
     subgraph AM ["<b>after_model</b> 反序 N→0"]
         direction TB
-        CL["[13] Clarification<br/>拦截 ask_clarification"] --> LD["[12] LoopDetection<br/>检测循环"] --> SL["[11] SubagentLimit<br/>截断多余 task"] --> TI["[8] Title<br/>生成标题"] --> SM["[6] Summarization<br/>上下文压缩"] --> DTC["[3] DanglingToolCall<br/>补缺失 ToolMessage"]
+        LD["[12] LoopDetection<br/>检测循环/排队 warning"] --> SL["[11] SubagentLimit<br/>截断多余 task"] --> TI["[8] Title<br/>生成标题"]
     end
 
-    M --> CL
+    M --> LD
 
     subgraph AA ["<b>after_agent</b> 反序 N→0"]
         direction TB
-        SBR["[2] Sandbox<br/>释放沙箱"] --> MEM["[9] Memory<br/>入队记忆"]
+        LD_CLEAN["[12] LoopDetection<br/>清理 pending warning"] --> MEM["[9] Memory<br/>入队记忆"] --> SBR["[2] Sandbox<br/>释放沙箱"]
     end
 
-    DTC --> SBR
-    MEM --> END(["response"])
+    TI --> LD_CLEAN
+    SBR --> END(["response"])
 
     classDef beforeNode fill:#a0a8b5,stroke:#636b7a,color:#2d3239
     classDef modelNode fill:#b5a8a0,stroke:#7a6b63,color:#2d3239
+    classDef wrapModelNode fill:#a8a0b5,stroke:#6b637a,color:#2d3239
     classDef afterModelNode fill:#b5a0a8,stroke:#7a636b,color:#2d3239
     classDef afterAgentNode fill:#a0b5a8,stroke:#637a6b,color:#2d3239
     classDef terminalNode fill:#a8b5a0,stroke:#6b7a63,color:#2d3239
 
-    class TD,UL,SB,VI beforeNode
+    class TD,UL,SB,LD_BA,VI beforeNode
+    class DTC_WM,LD_WM wrapModelNode
     class M modelNode
-    class CL,LD,SL,TI,SM,DTC afterModelNode
-    class SBR,MEM afterAgentNode
+    class LD,SL,TI afterModelNode
+    class LD_CLEAN,SBR,MEM afterAgentNode
     class START,END terminalNode
 ```
 
@@ -82,13 +90,12 @@ sequenceDiagram
     participant TD as ThreadDataMiddleware
     participant UL as UploadsMiddleware
     participant SB as SandboxMiddleware
+    participant LD as LoopDetectionMiddleware
     participant VI as ViewImageMiddleware
+    participant DTC as DanglingToolCallMiddleware
     participant M as MODEL
-    participant CL as ClarificationMiddleware
     participant SL as SubagentLimitMiddleware
     participant TI as TitleMiddleware
-    participant SM as SummarizationMiddleware
-    participant DTC as DanglingToolCallMiddleware
     participant MEM as MemoryMiddleware
 
     U ->> TD: invoke
@@ -103,19 +110,26 @@ sequenceDiagram
     activate SB
     Note right of SB: before_agent 获取沙箱
 
-    SB ->> VI: before_model
+    SB ->> LD: before_agent
+    activate LD
+    Note right of LD: before_agent 清理同 thread 旧 run 的 pending warning
+    LD ->> VI: before_model
     activate VI
     Note right of VI: before_model 注入图片 base64
 
-    VI ->> M: messages + tools
+    VI ->> DTC: wrap_model_call
+    activate DTC
+    Note right of DTC: wrap_model_call 补悬空 ToolMessage
+    DTC ->> LD: wrap_model_call
+    Note right of LD: wrap_model_call drain 当前 run warning 并追加到末尾
+    LD ->> M: messages + tools
     activate M
-    M -->> CL: AI response
+    M -->> LD: AI response
     deactivate M
 
-    activate CL
-    Note right of CL: after_model 拦截 ask_clarification
-    CL -->> SL: after_model
-    deactivate CL
+    Note right of LD: after_model 检测循环；warning 入队，hard-stop 清 tool_calls
+    LD -->> SL: after_model
+    deactivate LD
 
     activate SL
     Note right of SL: after_model 截断多余 task
@@ -124,22 +138,18 @@ sequenceDiagram
 
     activate TI
     Note right of TI: after_model 生成标题
-    TI -->> SM: after_model
+    TI -->> DTC: done
     deactivate TI
 
-    activate SM
-    Note right of SM: after_model 上下文压缩
-    SM -->> DTC: after_model
-    deactivate SM
-
-    activate DTC
-    Note right of DTC: after_model 补缺失 ToolMessage
-    DTC -->> VI: done
     deactivate DTC
 
     VI -->> SB: done
     deactivate VI
 
+    Note right of LD: after_agent 清理当前 run 未消费 warning
+
+    Note right of MEM: after_agent 入队记忆
+
     Note right of SB: after_agent 释放沙箱
     SB -->> UL: done
     deactivate SB
@@ -147,8 +157,6 @@ sequenceDiagram
     UL -->> TD: done
     deactivate UL
 
-    Note right of MEM: after_agent 入队记忆
-
     TD -->> U: response
     deactivate TD
 ```
@@ -224,12 +232,12 @@ sequenceDiagram
     participant TD as ThreadData
     participant UL as Uploads
     participant SB as Sandbox
+    participant LD as LoopDetection
     participant VI as ViewImage
+    participant DTC as DanglingToolCall
     participant M as MODEL
-    participant CL as Clarification
     participant SL as SubagentLimit
     participant TI as Title
-    participant SM as Summarization
     participant MEM as Memory
 
     U ->> TD: invoke
@@ -238,34 +246,40 @@ sequenceDiagram
     Note right of UL: before_agent 扫描文件
     UL ->> SB: .
     Note right of SB: before_agent 获取沙箱
+    SB ->> LD: .
+    Note right of LD: before_agent 清理 stale pending warning
 
     loop 每轮对话（tool call 循环）
         SB ->> VI: .
         Note right of VI: before_model 注入图片
-        VI ->> M: messages + tools
-        M -->> CL: AI response
-        Note right of CL: after_model 拦截 ask_clarification
-        CL -->> SL: .
+        VI ->> DTC: .
+        Note right of DTC: wrap_model_call 补悬空工具结果
+        DTC ->> LD: .
+        Note right of LD: wrap_model_call 注入当前 run warning
+        LD ->> M: messages + tools
+        M -->> LD: AI response
+        Note right of LD: after_model 检测循环/排队 warning
+        LD -->> SL: .
         Note right of SL: after_model 截断多余 task
         SL -->> TI: .
         Note right of TI: after_model 生成标题
-        TI -->> SM: .
-        Note right of SM: after_model 上下文压缩
     end
 
-    Note right of SB: after_agent 释放沙箱
-    SB -->> MEM: .
+    Note right of LD: after_agent 清理当前 run pending warning
+    LD -->> MEM: .
     Note right of MEM: after_agent 入队记忆
-    MEM -->> U: response
+    MEM -->> SB: .
+    Note right of SB: after_agent 释放沙箱
+    SB -->> U: response
 ```
 
 > [!warning] 不是洋葱
-> 14 个 middleware 中只有 SandboxMiddleware 有 before/after 对称（获取/释放）。其余都是单向的：要么只在 `before_*` 做事，要么只在 `after_*` 做事。`before_agent` / `after_agent` 只跑一次，`before_model` / `after_model` 每轮循环都跑。
+> 大部分 middleware 只用一个阶段。SandboxMiddleware 使用 `before_agent`/`after_agent` 做资源获取/释放；LoopDetectionMiddleware 也使用这两个钩子，但用途是清理 run-scoped pending warnings，不是资源生命周期对称。`before_agent` / `after_agent` 只跑一次，`before_model` / `after_model` / `wrap_model_call` 每轮循环都跑。
 
 硬依赖只有 2 处：
 
 1. **ThreadData 在 Sandbox 之前** — sandbox 需要线程目录
-2. **Clarification 在列表最后** — `after_model` 反序时最先执行，第一个拦截 `ask_clarification`
+2. **Clarification 在列表最后** — `wrap_tool_call` 处理 `ask_clarification` 时优先拦截，并通过 `Command(goto=END)` 中断执行
 
 ### 结论
 
@@ -273,19 +287,19 @@ sequenceDiagram
 |---|---|---|
 | 每个 middleware | before + after 对称 | 大多只用一个钩子 |
 | 激活条 | 嵌套（外长内短） | 不嵌套（串行） |
-| 反序的意义 | 清理与初始化配对 | 仅影响 after_model 的执行优先级 |
+| 反序的意义 | 清理与初始化配对 | 影响 `after_model` / `after_agent` 的执行优先级 |
 | 典型例子 | Auth: 校验 token / 清理上下文 | ThreadData: 只创建目录，没有清理 |
 
 ## 关键设计点
 
 ### ClarificationMiddleware 为什么在列表最后？
 
-位置最后 = `after_model` 最先执行。它需要**第一个**看到 model 输出，检查是否有 `ask_clarification` tool call。如果有，立即中断（`Command(goto=END)`），后续 middleware 的 `after_model` 不再执行。
+位置最后使它在工具调用包装链中优先拦截 `ask_clarification`。如果命中，它返回 `Command(goto=END)`，把格式化后的澄清问题写成 `ToolMessage` 并中断执行。
 
 ### SandboxMiddleware 的对称性
 
 `before_agent`（正序第 3 个）获取沙箱，`after_agent`（反序第 1 个）释放沙箱。外层进入 → 外层退出，天然的洋葱对称。
 
-### 大部分 middleware 只用一个钩子
+### LoopDetectionMiddleware 为什么同时用多个钩子？
 
-14 个 middleware 中，只有 SandboxMiddleware 同时用了 `before_agent` + `after_agent`（获取/释放）。其余都只在一个阶段执行。洋葱模型的反序特性主要影响 `after_model` 阶段的执行顺序。
+`after_model` 只做检测：重复工具调用达到 warning 阈值时，把 warning 放入 `(thread_id, run_id)` 作用域的 pending 队列。真正注入发生在下一次 `wrap_model_call`：此时上一轮 `AIMessage(tool_calls)` 对应的 `ToolMessage` 已经在请求里，warning 追加在末尾，不会破坏 OpenAI/Moonshot 的 tool-call pairing。`before_agent` 清理同一 thread 下旧 run 的残留 warning，`after_agent` 清理当前 run 没被消费的 warning。

backend/docs/plan_mode_usage.md

@@ -127,8 +127,8 @@ complex_agent = create_agent_for_task("high")
 ## How It Works
 
 1. When `make_lead_agent(config)` is called, it extracts `is_plan_mode` from `config.configurable`
-2. The config is passed to `_build_middlewares(config)`
-3. `_build_middlewares()` reads `is_plan_mode` and calls `_create_todo_list_middleware(is_plan_mode)`
+2. The config is passed to `build_middlewares(config)`
+3. `build_middlewares()` reads `is_plan_mode` and calls `_create_todo_list_middleware(is_plan_mode)`
 4. If `is_plan_mode=True`, a `TodoListMiddleware` instance is created and added to the middleware chain
 5. The middleware automatically adds a `write_todos` tool to the agent's toolset
 6. The agent can use this tool to manage tasks during execution
@@ -141,7 +141,7 @@ make_lead_agent(config)
   │
   ├─> Extracts: is_plan_mode = config.configurable.get("is_plan_mode", False)
   │
-  └─> _build_middlewares(config)
+  └─> build_middlewares(config)
         │
         ├─> ThreadDataMiddleware
         ├─> SandboxMiddleware
@@ -156,7 +156,7 @@ make_lead_agent(config)
 ### Agent Module
 - **Location**: `packages/harness/deerflow/agents/lead_agent/agent.py`
 - **Function**: `_create_todo_list_middleware(is_plan_mode: bool)` - Creates TodoListMiddleware if plan mode is enabled
-- **Function**: `_build_middlewares(config: RunnableConfig)` - Builds middleware chain based on runtime config
+- **Function**: `build_middlewares(config: RunnableConfig)` - Builds middleware chain based on runtime config
 - **Function**: `make_lead_agent(config: RunnableConfig)` - Creates agent with appropriate middlewares
 
 ### Runtime Configuration

backend/packages/harness/deerflow/agents/lead_agent/agent.py

@@ -1,3 +1,25 @@
+"""Lead agent factory.
+
+INVARIANT — tracing callback placement
+======================================
+
+Tracing callbacks (Langfuse, LangSmith) are attached at the **graph
+invocation root** in :func:`_make_lead_agent` (see the
+``build_tracing_callbacks()`` block that appends to ``config["callbacks"]``).
+Every ``create_chat_model(...)`` call inside this module — and inside any
+middleware reachable from this graph (e.g. ``TitleMiddleware``) — MUST pass
+``attach_tracing=False``.
+
+Forgetting that flag emits duplicate spans (one rooted at the graph, one at
+the model) AND prevents the Langfuse handler's ``propagate_attributes``
+path from firing, so ``session_id`` / ``user_id`` never reach the trace.
+The four current sites are: bootstrap agent, default agent, summarization
+middleware, and the async path inside ``TitleMiddleware``. Any new in-graph
+``create_chat_model`` call must add to this list and pass the flag.
+"""
+
+from __future__ import annotations
+
 import logging
 
 from langchain.agents import create_agent
@@ -9,6 +31,7 @@ from deerflow.agents.memory.summarization_hook import memory_flush_hook
 from deerflow.agents.middlewares.clarification_middleware import ClarificationMiddleware
 from deerflow.agents.middlewares.loop_detection_middleware import LoopDetectionMiddleware
 from deerflow.agents.middlewares.memory_middleware import MemoryMiddleware
+from deerflow.agents.middlewares.safety_finish_reason_middleware import SafetyFinishReasonMiddleware
 from deerflow.agents.middlewares.subagent_limit_middleware import SubagentLimitMiddleware
 from deerflow.agents.middlewares.summarization_middleware import BeforeSummarizationHook, DeerFlowSummarizationMiddleware
 from deerflow.agents.middlewares.title_middleware import TitleMiddleware
@@ -22,9 +45,12 @@ from deerflow.config.app_config import AppConfig, get_app_config
 from deerflow.models import create_chat_model
 from deerflow.skills.tool_policy import filter_tools_by_skill_allowed_tools
 from deerflow.skills.types import Skill
+from deerflow.tracing import build_tracing_callbacks
 
 logger = logging.getLogger(__name__)
 
+_BOOTSTRAP_SKILL_NAMES = {"bootstrap"}
+
 
 def _get_runtime_config(config: RunnableConfig) -> dict:
     """Merge legacy configurable options with LangGraph runtime context."""
@@ -73,10 +99,14 @@ def _create_summarization_middleware(*, app_config: AppConfig | None = None) ->
     # Bind "middleware:summarize" tag so RunJournal identifies these LLM calls
     # as middleware rather than lead_agent (SummarizationMiddleware is a
     # LangChain built-in, so we tag the model at creation time).
+    # attach_tracing=False because the graph-level RunnableConfig (set in
+    # ``_make_lead_agent``) already carries tracing callbacks; binding them
+    # again at the model level would emit duplicate spans and break
+    # ``session_id`` / ``user_id`` propagation.
     if config.model_name:
-        model = create_chat_model(name=config.model_name, thinking_enabled=False, app_config=resolved_app_config)
+        model = create_chat_model(name=config.model_name, thinking_enabled=False, app_config=resolved_app_config, attach_tracing=False)
     else:
-        model = create_chat_model(thinking_enabled=False, app_config=resolved_app_config)
+        model = create_chat_model(thinking_enabled=False, app_config=resolved_app_config, attach_tracing=False)
     model = model.with_config(tags=["middleware:summarize"])
 
     # Prepare kwargs
@@ -237,20 +267,31 @@ Being proactive with task management demonstrates thoroughness and ensures all r
 # ViewImageMiddleware should be before ClarificationMiddleware to inject image details before LLM
 # ToolErrorHandlingMiddleware should be before ClarificationMiddleware to convert tool exceptions to ToolMessages
 # ClarificationMiddleware should be last to intercept clarification requests after model calls
-def _build_middlewares(
+def build_middlewares(
     config: RunnableConfig,
     model_name: str | None,
     agent_name: str | None = None,
     custom_middlewares: list[AgentMiddleware] | None = None,
     *,
+    available_skills: set[str] | None = None,
     app_config: AppConfig | None = None,
+    deferred_setup=None,
 ):
-    """Build middleware chain based on runtime configuration.
+    """Build the lead-agent middleware chain based on runtime configuration.
+
+    Public entry point for the lead agent's full middleware composition. Used by
+    ``make_lead_agent`` and by the embedded ``DeerFlowClient`` (a lead-agent variant
+    that needs the identical chain). Keep this name stable: it is imported across a
+    module boundary, so renames/signature changes ripple into ``client.py``.
 
     Args:
         config: Runtime configuration containing configurable options like is_plan_mode.
+        model_name: Resolved runtime model name; gates vision-only middleware.
         agent_name: If provided, MemoryMiddleware will use per-agent memory storage.
         custom_middlewares: Optional list of custom middlewares to inject into the chain.
+        app_config: Explicit AppConfig; falls back to ``get_app_config()`` when omitted.
+        deferred_setup: Optional deferred-MCP-tool setup that attaches
+            ``DeferredToolFilterMiddleware`` when ``tool_search`` is enabled.
 
     Returns:
         List of middleware instances.
@@ -264,6 +305,13 @@ def _build_middlewares(
 
     middlewares.append(DynamicContextMiddleware(agent_name=agent_name, app_config=resolved_app_config))
 
+    # Deterministically load a full SKILL.md when the user starts the turn with
+    # /skill-name. This keeps the base system prompt metadata-only while giving
+    # explicit user activation priority over model-side relevance guessing.
+    from deerflow.agents.middlewares.skill_activation_middleware import SkillActivationMiddleware
+
+    middlewares.append(SkillActivationMiddleware(available_skills=available_skills, app_config=resolved_app_config))
+
     # Add summarization middleware if enabled
     summarization_middleware = _create_summarization_middleware(app_config=resolved_app_config)
     if summarization_middleware is not None:
@@ -292,11 +340,13 @@ def _build_middlewares(
     if model_config is not None and model_config.supports_vision:
         middlewares.append(ViewImageMiddleware())
 
-    # Add DeferredToolFilterMiddleware to hide deferred tool schemas from model binding
-    if resolved_app_config.tool_search.enabled:
+    # Hide deferred tool schemas from model binding until tool_search promotes them.
+    # The deferred set + catalog hash come from the build-time setup (assembled
+    # after tool-policy filtering); promotion is read from graph state.
+    if deferred_setup is not None and deferred_setup.deferred_names:
         from deerflow.agents.middlewares.deferred_tool_filter_middleware import DeferredToolFilterMiddleware
 
-        middlewares.append(DeferredToolFilterMiddleware())
+        middlewares.append(DeferredToolFilterMiddleware(deferred_setup.deferred_names, deferred_setup.catalog_hash))
 
     # Add SubagentLimitMiddleware to truncate excess parallel task calls
     subagent_enabled = cfg.get("subagent_enabled", False)
@@ -313,6 +363,15 @@ def _build_middlewares(
     if custom_middlewares:
         middlewares.extend(custom_middlewares)
 
+    # SafetyFinishReasonMiddleware — suppress tool execution when the provider
+    # safety-terminated the response. Registered after custom middlewares so
+    # that LangChain's reverse-order after_model dispatch runs Safety first;
+    # cleared tool_calls then flow through Loop/Subagent accounting without
+    # firing extra alarms. See safety_finish_reason_middleware.py docstring.
+    safety_config = resolved_app_config.safety_finish_reason
+    if safety_config.enabled:
+        middlewares.append(SafetyFinishReasonMiddleware.from_config(safety_config))
+
     # ClarificationMiddleware should always be last
     middlewares.append(ClarificationMiddleware())
     return middlewares
@@ -320,7 +379,7 @@ def _build_middlewares(
 
 def _available_skill_names(agent_config, is_bootstrap: bool) -> set[str] | None:
     if is_bootstrap:
-        return {"bootstrap"}
+        return set(_BOOTSTRAP_SKILL_NAMES)
     if agent_config and agent_config.skills is not None:
         return set(agent_config.skills)
     return None
@@ -351,6 +410,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
     # Lazy import to avoid circular dependency
     from deerflow.tools import get_available_tools
     from deerflow.tools.builtins import setup_agent, update_agent
+    from deerflow.tools.builtins.tool_search import assemble_deferred_tools
 
     cfg = _get_runtime_config(config)
     resolved_app_config = app_config
@@ -408,20 +468,44 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
         }
     )
 
+    # Inject tracing callbacks at the graph invocation root so a single LangGraph
+    # run produces one trace with all node / LLM / tool calls as child spans,
+    # AND so the Langfuse handler sees ``on_chain_start(parent_run_id=None)`` and
+    # actually propagates ``langfuse_session_id`` / ``langfuse_user_id`` from
+    # ``config["metadata"]`` onto the trace. Without root-level attachment the
+    # model is a nested observation and the handler strips ``langfuse_*`` keys.
+    tracing_callbacks = build_tracing_callbacks()
+    if tracing_callbacks:
+        existing = config.get("callbacks") or []
+        if not isinstance(existing, list):
+            existing = list(existing)
+        config["callbacks"] = [*existing, *tracing_callbacks]
+
     skills_for_tool_policy = _load_enabled_skills_for_tool_policy(available_skills, app_config=resolved_app_config)
 
     if is_bootstrap:
         # Special bootstrap agent with minimal prompt for initial custom agent creation flow
-        tools = get_available_tools(model_name=model_name, subagent_enabled=subagent_enabled, app_config=resolved_app_config) + [setup_agent]
+        # Keep the bootstrap skill set intentionally narrow so agent creation
+        # remains deterministic before the custom agent's own config exists.
+        raw_tools = get_available_tools(model_name=model_name, subagent_enabled=subagent_enabled, app_config=resolved_app_config) + [setup_agent]
+        filtered = filter_tools_by_skill_allowed_tools(raw_tools, skills_for_tool_policy)
+        final_tools, setup = assemble_deferred_tools(filtered, enabled=resolved_app_config.tool_search.enabled)
         return create_agent(
-            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, app_config=resolved_app_config),
-            tools=filter_tools_by_skill_allowed_tools(tools, skills_for_tool_policy),
-            middleware=_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
+            model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, app_config=resolved_app_config, attach_tracing=False),
+            tools=final_tools,
+            middleware=build_middlewares(
+                config,
+                model_name=model_name,
+                available_skills=set(_BOOTSTRAP_SKILL_NAMES),
+                app_config=resolved_app_config,
+                deferred_setup=setup,
+            ),
             system_prompt=apply_prompt_template(
                 subagent_enabled=subagent_enabled,
                 max_concurrent_subagents=max_concurrent_subagents,
-                available_skills=set(["bootstrap"]),
+                available_skills=set(_BOOTSTRAP_SKILL_NAMES),
                 app_config=resolved_app_config,
+                deferred_names=setup.deferred_names,
             ),
             state_schema=ThreadState,
         )
@@ -430,17 +514,27 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
     # The default agent (no agent_name) does not see this tool.
     extra_tools = [update_agent] if agent_name else []
     # Default lead agent (unchanged behavior)
-    tools = get_available_tools(model_name=model_name, groups=agent_config.tool_groups if agent_config else None, subagent_enabled=subagent_enabled, app_config=resolved_app_config)
+    raw_tools = get_available_tools(model_name=model_name, groups=agent_config.tool_groups if agent_config else None, subagent_enabled=subagent_enabled, app_config=resolved_app_config)
+    filtered = filter_tools_by_skill_allowed_tools(raw_tools + extra_tools, skills_for_tool_policy)
+    final_tools, setup = assemble_deferred_tools(filtered, enabled=resolved_app_config.tool_search.enabled)
     return create_agent(
-        model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, reasoning_effort=reasoning_effort, app_config=resolved_app_config),
-        tools=filter_tools_by_skill_allowed_tools(tools + extra_tools, skills_for_tool_policy),
-        middleware=_build_middlewares(config, model_name=model_name, agent_name=agent_name, app_config=resolved_app_config),
+        model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, reasoning_effort=reasoning_effort, app_config=resolved_app_config, attach_tracing=False),
+        tools=final_tools,
+        middleware=build_middlewares(
+            config,
+            model_name=model_name,
+            agent_name=agent_name,
+            available_skills=available_skills,
+            app_config=resolved_app_config,
+            deferred_setup=setup,
+        ),
         system_prompt=apply_prompt_template(
             subagent_enabled=subagent_enabled,
             max_concurrent_subagents=max_concurrent_subagents,
             agent_name=agent_name,
-            available_skills=set(agent_config.skills) if agent_config and agent_config.skills is not None else None,
+            available_skills=available_skills,
             app_config=resolved_app_config,
+            deferred_names=setup.deferred_names,
         ),
         state_schema=ThreadState,
     )

backend/packages/harness/deerflow/agents/lead_agent/prompt.py

@@ -10,6 +10,7 @@ from deerflow.config.agents_config import load_agent_soul
 from deerflow.skills.storage import get_or_new_skill_storage
 from deerflow.skills.types import Skill, SkillCategory
 from deerflow.subagents import get_available_subagent_names
+from deerflow.tools.builtins.tool_search import get_deferred_tools_prompt_section
 
 if TYPE_CHECKING:
     from deerflow.config.app_config import AppConfig
@@ -542,6 +543,14 @@ combined with a FastAPI gateway for REST API access [citation:FastAPI](https://f
 {subagent_reminder}- Skill First: Always load the relevant skill before starting **complex** tasks.
 - Progressive Loading: Load resources incrementally as referenced in skills
 - Output Files: Final deliverables must be in `/mnt/user-data/outputs`
+- File Editing Workflow: When revising an existing file, prefer
+  `str_replace` over `write_file` — it sends only the diff and avoids
+  re-emitting the whole file (mirrors Claude Code's Edit and Codex's
+  apply_patch). When writing long new content from scratch, split it
+  into sections: the first `write_file` call creates the file, then use
+  `write_file` with append=True to extend it section by section. This
+  keeps each tool call small and avoids mid-stream chunk-gap timeouts
+  on oversized single-shot writes. (See issue #3189.)  
 - Clarity: Be direct and helpful, avoid unnecessary meta-commentary
 - Including Images and Mermaid: Images and Mermaid diagrams are always welcomed in the Markdown format, and you're encouraged to use `![Image Description](image_path)\n\n` or "```mermaid" to display images in response or Markdown files
 - Multi-task: Better utilize parallel tool calling to call multiple tools at one time for better performance
@@ -616,6 +625,11 @@ You have access to skills that provide optimized workflows for specific tasks. E
 4. Load referenced resources only when needed during execution
 5. Follow the skill's instructions precisely
 
+**Explicit Slash Skill Activation:**
+- If the user starts a request with `/<skill-name>`, that skill was explicitly requested for the current turn.
+- Follow the activated skill before choosing a general workflow.
+- The runtime injects the activated skill content for explicit slash activations; do not call `read_file` for that SKILL.md again unless the injected skill references supporting resources you need.
+
 **Skills are located at:** {container_base_path}
 {skill_evolution_section}
 {skills_list}
@@ -678,42 +692,13 @@ SOUL.md or config.yaml — those write into a temporary sandbox/tool workspace a
 Rules:
 - Always pass the FULL replacement text for `soul` (no patch semantics). Start from your current SOUL above and apply the user's edits.
 - Only pass the fields that should change. Omit the others to preserve them.
+- Never pass literal strings like `"null"`, `"none"`, or `"undefined"` for unchanged fields.
 - Pass `skills=[]` to disable all skills, or omit `skills` to keep the existing whitelist.
 - After `update_agent` returns successfully, tell the user the change is persisted and will take effect on the next turn.
 </self_update>
 """
 
 
-def get_deferred_tools_prompt_section(*, app_config: AppConfig | None = None) -> str:
-    """Generate <available-deferred-tools> block for the system prompt.
-
-    Lists only deferred tool names so the agent knows what exists
-    and can use tool_search to load them.
-    Returns empty string when tool_search is disabled or no tools are deferred.
-    """
-    from deerflow.tools.builtins.tool_search import get_deferred_registry
-
-    if app_config is None:
-        try:
-            from deerflow.config import get_app_config
-
-            config = get_app_config()
-        except Exception:
-            return ""
-    else:
-        config = app_config
-
-    if not config.tool_search.enabled:
-        return ""
-
-    registry = get_deferred_registry()
-    if not registry:
-        return ""
-
-    names = "\n".join(e.name for e in registry.entries)
-    return f"<available-deferred-tools>\n{names}\n</available-deferred-tools>"
-
-
 def _build_acp_section(*, app_config: AppConfig | None = None) -> str:
     """Build the ACP agent prompt section, only if ACP agents are configured."""
     if app_config is None:
@@ -772,6 +757,7 @@ def apply_prompt_template(
     agent_name: str | None = None,
     available_skills: set[str] | None = None,
     app_config: AppConfig | None = None,
+    deferred_names: frozenset[str] = frozenset(),
 ) -> str:
     # Include subagent section only if enabled (from runtime parameter)
     n = max_concurrent_subagents
@@ -799,7 +785,7 @@ def apply_prompt_template(
     skills_section = get_skills_prompt_section(available_skills, app_config=app_config)
 
     # Get deferred tools section (tool_search)
-    deferred_tools_section = get_deferred_tools_prompt_section(app_config=app_config)
+    deferred_tools_section = get_deferred_tools_prompt_section(deferred_names=deferred_names)
 
     # Build ACP agent section only if ACP agents are configured
     acp_section = _build_acp_section(app_config=app_config)

backend/packages/harness/deerflow/agents/memory/prompt.py

@@ -1,9 +1,14 @@
 """Prompt templates for memory update and injection."""
 
+from __future__ import annotations
+
+import logging
 import math
 import re
 from typing import Any
 
+logger = logging.getLogger(__name__)
+
 try:
     import tiktoken
 
@@ -160,6 +165,39 @@ Rules:
 Return ONLY valid JSON."""
 
 
+# Module-level tiktoken encoding cache.  Populated lazily on first use;
+# subsequent calls are a dict lookup (no network I/O).  Pre-warming at
+# startup via :func:`warm_tiktoken_cache` avoids blocking a request on the
+# (potentially slow) first ``get_encoding`` call.
+_tiktoken_encoding_cache: dict[str, tiktoken.Encoding] = {}
+
+
+def _get_tiktoken_encoding(encoding_name: str = "cl100k_base") -> tiktoken.Encoding | None:
+    """Return a cached tiktoken encoding, or ``None`` on failure / unavailability.
+
+    On the very first call for a given *encoding_name*, tiktoken may need to
+    download the BPE data from ``openaipublic.blob.core.windows.net``.  In
+    network-restricted environments (e.g. deployments behind the GFW) this
+    download can block for tens of minutes before the OS TCP timeout kicks in.
+    The caller must therefore be prepared for this to block and should run it
+    off the event loop (e.g. via ``asyncio.to_thread``).
+    """
+    if not TIKTOKEN_AVAILABLE:
+        return None
+
+    cached = _tiktoken_encoding_cache.get(encoding_name)
+    if cached is not None:
+        return cached
+
+    try:
+        encoding = tiktoken.get_encoding(encoding_name)
+        _tiktoken_encoding_cache[encoding_name] = encoding
+        return encoding
+    except Exception:
+        logger.warning("Failed to load tiktoken encoding %r; falling back to char-based estimation", encoding_name, exc_info=True)
+        return None
+
+
 def _count_tokens(text: str, encoding_name: str = "cl100k_base") -> int:
     """Count tokens in text using tiktoken.
 
@@ -170,18 +208,30 @@ def _count_tokens(text: str, encoding_name: str = "cl100k_base") -> int:
     Returns:
         The number of tokens in the text.
     """
-    if not TIKTOKEN_AVAILABLE:
+    encoding = _get_tiktoken_encoding(encoding_name)
+    if encoding is None:
         # Fallback to character-based estimation if tiktoken is not available
+        # or the encoding failed to load.
         return len(text) // 4
 
     try:
-        encoding = tiktoken.get_encoding(encoding_name)
         return len(encoding.encode(text))
     except Exception:
         # Fallback to character-based estimation on error
         return len(text) // 4
 
 
+def warm_tiktoken_cache() -> bool:
+    """Pre-warm the tiktoken encoding cache.
+
+    Call at startup (off the event loop) so the first request never blocks
+    on the BPE download.  Returns ``True`` if the encoding was loaded
+    successfully (or was already cached), ``False`` if tiktoken is
+    unavailable or the download failed.
+    """
+    return _get_tiktoken_encoding("cl100k_base") is not None
+
+
 def _coerce_confidence(value: Any, default: float = 0.0) -> float:
     """Coerce a confidence-like value to a bounded float in [0, 1].
 

backend/packages/harness/deerflow/agents/memory/updater.py

@@ -227,6 +227,110 @@ def _extract_text(content: Any) -> str:
     return str(content)
 
 
+_REQUIRED_MEMORY_UPDATE_TOP_LEVEL_KEYS = frozenset({"user", "history", "newFacts", "factsToRemove"})
+
+
+def _normalize_memory_update_fact(fact: Any) -> dict[str, Any] | None:
+    """Normalize a single fact entry from a model-produced memory update."""
+    if not isinstance(fact, dict):
+        return None
+
+    raw_content = fact.get("content")
+    if not isinstance(raw_content, str):
+        return None
+    content = raw_content.strip()
+    if not content:
+        return None
+
+    raw_category = fact.get("category")
+    category = raw_category.strip() if isinstance(raw_category, str) and raw_category.strip() else "context"
+
+    raw_confidence = fact.get("confidence", 0.5)
+    if isinstance(raw_confidence, bool):
+        return None
+    if isinstance(raw_confidence, str):
+        raw_confidence = raw_confidence.strip()
+        if not raw_confidence:
+            return None
+        try:
+            raw_confidence = float(raw_confidence)
+        except ValueError:
+            return None
+    elif isinstance(raw_confidence, (int, float)):
+        raw_confidence = float(raw_confidence)
+    else:
+        return None
+
+    if not math.isfinite(raw_confidence):
+        return None
+
+    normalized_fact = {
+        "content": content,
+        "category": category,
+        "confidence": raw_confidence,
+    }
+    source_error = fact.get("sourceError")
+    if isinstance(source_error, str):
+        normalized_source_error = source_error.strip()
+        if normalized_source_error:
+            normalized_fact["sourceError"] = normalized_source_error
+
+    return normalized_fact
+
+
+def _normalize_memory_update_data(update_data: dict[str, Any]) -> dict[str, Any]:
+    """Coerce parsed memory update data into the shape consumed by _apply_updates."""
+    user = update_data.get("user")
+    history = update_data.get("history")
+    new_facts = update_data.get("newFacts")
+    facts_to_remove = update_data.get("factsToRemove")
+    normalized_facts_to_remove = [fact_id for fact_id in facts_to_remove if isinstance(fact_id, str)] if isinstance(facts_to_remove, list) else []
+    normalized_new_facts = []
+    dropped_new_fact = not isinstance(new_facts, list)
+    if isinstance(new_facts, list):
+        for fact in new_facts:
+            normalized_fact = _normalize_memory_update_fact(fact)
+            if normalized_fact is not None:
+                normalized_new_facts.append(normalized_fact)
+            else:
+                dropped_new_fact = True
+
+    if normalized_facts_to_remove and dropped_new_fact:
+        raise json.JSONDecodeError(
+            "Unsafe partial memory update: factsToRemove with malformed newFacts",
+            json.dumps(update_data, ensure_ascii=False),
+            0,
+        )
+
+    return {
+        "user": user if isinstance(user, dict) else {},
+        "history": history if isinstance(history, dict) else {},
+        "newFacts": normalized_new_facts,
+        "factsToRemove": normalized_facts_to_remove,
+    }
+
+
+def _parse_memory_update_response(response_content: Any) -> dict[str, Any]:
+    """Parse the first valid memory-update JSON object from an LLM response.
+
+    Some providers may wrap JSON in thinking traces, prose, or markdown fences
+    even when prompted to return JSON only. This parser accepts safely
+    extractable JSON objects but does not repair truncated or malformed JSON.
+    """
+    response_text = _extract_text(response_content).strip()
+    decoder = json.JSONDecoder()
+
+    for match in re.finditer(r"\{", response_text):
+        try:
+            parsed, _end = decoder.raw_decode(response_text[match.start() :])
+        except json.JSONDecodeError:
+            continue
+        if isinstance(parsed, dict) and _REQUIRED_MEMORY_UPDATE_TOP_LEVEL_KEYS.issubset(parsed):
+            return _normalize_memory_update_data(parsed)
+
+    raise json.JSONDecodeError("No valid memory update JSON object found", response_text, 0)
+
+
 # Matches sentences that describe a file-upload *event* rather than general
 # file-related work.  Deliberately narrow to avoid removing legitimate facts
 # such as "User works with CSV files" or "prefers PDF export".
@@ -338,7 +442,7 @@ class MemoryUpdater:
             reinforcement_detected=reinforcement_detected,
         )
         prompt = MEMORY_UPDATE_PROMPT.format(
-            current_memory=json.dumps(current_memory, indent=2),
+            current_memory=json.dumps(current_memory, indent=2, ensure_ascii=False),
             conversation=conversation_text,
             correction_hint=correction_hint,
         )
@@ -353,13 +457,7 @@ class MemoryUpdater:
         user_id: str | None = None,
     ) -> bool:
         """Parse the model response, apply updates, and persist memory."""
-        response_text = _extract_text(response_content).strip()
-
-        if response_text.startswith("```"):
-            lines = response_text.split("\n")
-            response_text = "\n".join(lines[1:-1] if lines[-1] == "```" else lines[1:])
-
-        update_data = json.loads(response_text)
+        update_data = _parse_memory_update_response(response_content)
         # Deep-copy before in-place mutation so a subsequent save() failure
         # cannot corrupt the still-cached original object reference.
         updated_memory = self._apply_updates(copy.deepcopy(current_memory), update_data, thread_id)

backend/packages/harness/deerflow/agents/middlewares/dangling_tool_call_middleware.py

@@ -15,6 +15,7 @@ to the end of the message list as before_model + add_messages reducer would do.
 
 import json
 import logging
+from collections import defaultdict, deque
 from collections.abc import Awaitable, Callable
 from typing import override
 
@@ -25,6 +26,11 @@ from langchain_core.messages import ToolMessage
 
 logger = logging.getLogger(__name__)
 
+# Workaround for issue #2894: malformed write_file calls can carry huge Markdown
+# payloads in invalid tool-call args. Keep recovery error details short so the
+# synthetic ToolMessage does not echo large or malformed content back to the model.
+_MAX_RECOVERY_ERROR_DETAIL_LEN = 500
+
 
 class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
     """Inserts placeholder ToolMessages for dangling tool calls before model invocation.
@@ -97,9 +103,25 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
     @staticmethod
     def _synthetic_tool_message_content(tool_call: dict) -> str:
         if tool_call.get("invalid"):
+            name = tool_call.get("name")
             error = tool_call.get("error")
-            if isinstance(error, str) and error:
-                return f"[Tool call could not be executed because its arguments were invalid: {error}]"
+            error_text = error[:_MAX_RECOVERY_ERROR_DETAIL_LEN] if isinstance(error, str) and error else ""
+            # Workaround for issue #2894: malformed write_file calls can carry huge Markdown
+            # payloads in invalid tool-call args. Keep recovery guidance actionable without
+            # echoing large or malformed content back to the model.
+            if name == "write_file":
+                details = f" Parser error: {error_text}" if error_text else ""
+                return (
+                    "[write_file failed before execution: the tool-call arguments were not valid JSON, "
+                    "so no file was written. This often happens when the model tries to write a very "
+                    "large Markdown file in a single tool call, especially when `content` contains "
+                    "unescaped quotes, inline JSON, backslashes, or code fences. Do not retry the same "
+                    "large `write_file` payload for this artifact; provide the report/content directly "
+                    "as normal assistant text in your next response. If a file write is still needed "
+                    f"later, split the file into smaller sections instead of one large payload.{details}]"
+                )
+            if error_text:
+                return f"[Tool call could not be executed because its arguments were invalid: {error_text}]"
             return "[Tool call could not be executed because its arguments were invalid.]"
         return "[Tool call was interrupted and did not return a result.]"
 
@@ -109,10 +131,10 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
         This normalizes model-bound causal order before provider serialization while
         preserving already-valid transcripts unchanged.
         """
-        tool_messages_by_id: dict[str, ToolMessage] = {}
+        tool_messages_by_id: dict[str, deque[ToolMessage]] = defaultdict(deque)
         for msg in messages:
             if isinstance(msg, ToolMessage):
-                tool_messages_by_id.setdefault(msg.tool_call_id, msg)
+                tool_messages_by_id[msg.tool_call_id].append(msg)
 
         tool_call_ids: set[str] = set()
         for msg in messages:
@@ -124,7 +146,6 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
                     tool_call_ids.add(tc_id)
 
         patched: list = []
-        consumed_tool_msg_ids: set[str] = set()
         patch_count = 0
         for msg in messages:
             if isinstance(msg, ToolMessage) and msg.tool_call_id in tool_call_ids:
@@ -136,13 +157,13 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
 
             for tc in self._message_tool_calls(msg):
                 tc_id = tc.get("id")
-                if not tc_id or tc_id in consumed_tool_msg_ids:
+                if not tc_id:
                     continue
 
-                existing_tool_msg = tool_messages_by_id.get(tc_id)
+                tool_msg_queue = tool_messages_by_id.get(tc_id)
+                existing_tool_msg = tool_msg_queue.popleft() if tool_msg_queue else None
                 if existing_tool_msg is not None:
                     patched.append(existing_tool_msg)
-                    consumed_tool_msg_ids.add(tc_id)
                 else:
                     patched.append(
                         ToolMessage(
@@ -152,7 +173,6 @@ class DanglingToolCallMiddleware(AgentMiddleware[AgentState]):
                             status="error",
                         )
                     )
-                    consumed_tool_msg_ids.add(tc_id)
                     patch_count += 1
 
         if patched == messages:

backend/packages/harness/deerflow/agents/middlewares/deferred_tool_filter_middleware.py

@@ -1,12 +1,15 @@
 """Middleware to filter deferred tool schemas from model binding.
 
-When tool_search is enabled, MCP tools are registered in the DeferredToolRegistry
-and passed to ToolNode for execution, but their schemas should NOT be sent to the
-LLM via bind_tools (that's the whole point of deferral — saving context tokens).
+When tool_search is enabled, MCP tools are still passed to ToolNode for
+execution, but their schemas must NOT be sent to the LLM via bind_tools until
+the model has discovered them via tool_search. This middleware removes the
+still-deferred tools from request.tools before model binding, and blocks tool
+calls to tools that have not been promoted yet.
 
-This middleware intercepts wrap_model_call and removes deferred tools from
-request.tools so that model.bind_tools only receives active tool schemas.
-The agent discovers deferred tools at runtime via the tool_search tool.
+The deferred name set and the catalog hash are injected at construction time
+(no ContextVar). Promotion state is read from graph state (``state["promoted"]``),
+scoped by catalog hash so a stale persisted promotion cannot expose a renamed
+or drifted tool.
 """
 
 import logging
@@ -24,47 +27,49 @@ logger = logging.getLogger(__name__)
 
 
 class DeferredToolFilterMiddleware(AgentMiddleware[AgentState]):
-    """Remove deferred tools from request.tools before model binding.
+    """Hide deferred tool schemas from the bound model until promoted.
 
     ToolNode still holds all tools (including deferred) for execution routing,
-    but the LLM only sees active tool schemas — deferred tools are discoverable
-    via tool_search at runtime.
+    but the LLM only sees active tool schemas plus tools that have already been
+    promoted (recorded in ``state["promoted"]`` under the current catalog hash).
     """
 
+    def __init__(self, deferred_names: frozenset[str], catalog_hash: str | None):
+        super().__init__()
+        self._deferred = deferred_names
+        self._catalog_hash = catalog_hash
+
+    def _promoted(self, state) -> set[str]:
+        promoted = (state or {}).get("promoted")
+        if promoted and promoted.get("catalog_hash") == self._catalog_hash:
+            return set(promoted.get("names") or [])
+        return set()
+
+    def _hidden(self, state) -> set[str]:
+        return set(self._deferred) - self._promoted(state)
+
     def _filter_tools(self, request: ModelRequest) -> ModelRequest:
-        from deerflow.tools.builtins.tool_search import get_deferred_registry
-
-        registry = get_deferred_registry()
-        if not registry:
+        if not self._deferred:
             return request
-
-        deferred_names = registry.deferred_names
-        active_tools = [t for t in request.tools if getattr(t, "name", None) not in deferred_names]
-
-        if len(active_tools) < len(request.tools):
-            logger.debug(f"Filtered {len(request.tools) - len(active_tools)} deferred tool schema(s) from model binding")
-
-        return request.override(tools=active_tools)
+        hide = self._hidden(request.state)
+        if not hide:
+            return request
+        active = [t for t in request.tools if getattr(t, "name", None) not in hide]
+        if len(active) < len(request.tools):
+            logger.debug("Filtered %d deferred tool schema(s) from model binding", len(request.tools) - len(active))
+        return request.override(tools=active)
 
     def _blocked_tool_message(self, request: ToolCallRequest) -> ToolMessage | None:
-        from deerflow.tools.builtins.tool_search import get_deferred_registry
-
-        registry = get_deferred_registry()
-        if not registry:
+        if not self._deferred:
             return None
-
-        tool_name = str(request.tool_call.get("name") or "")
-        if not tool_name:
+        name = str(request.tool_call.get("name") or "")
+        if not name or name not in self._hidden(request.state):
             return None
-
-        if not registry.contains(tool_name):
-            return None
-
         tool_call_id = str(request.tool_call.get("id") or "missing_tool_call_id")
         return ToolMessage(
-            content=(f"Error: Tool '{tool_name}' is deferred and has not been promoted yet. Call tool_search first to expose and promote this tool's schema, then retry."),
+            content=(f"Error: Tool '{name}' is deferred and has not been promoted yet. Call tool_search first to expose and promote this tool's schema, then retry."),
             tool_call_id=tool_call_id,
-            name=tool_name,
+            name=name,
             status="error",
         )
 

backend/packages/harness/deerflow/agents/middlewares/dynamic_context_middleware.py

@@ -28,6 +28,7 @@ Date-update format:
 
 from __future__ import annotations
 
+import asyncio
 import logging
 import re
 import uuid
@@ -43,6 +44,12 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+# Upper bound (seconds) for a single _inject() offload.  If the warm-up at
+# gateway startup failed silently, the first request may still hit a cold
+# tiktoken BPE download that blocks until the OS TCP timeout (~26 min).
+# This cap ensures the request degrades gracefully instead of hanging.
+_INJECT_TIMEOUT_SECONDS = 5.0
+
 _DATE_RE = re.compile(r"<current_date>([^<]+)</current_date>")
 _DYNAMIC_CONTEXT_REMINDER_KEY = "dynamic_context_reminder"
 _SUMMARY_MESSAGE_NAME = "summary"
@@ -201,4 +208,25 @@ class DynamicContextMiddleware(AgentMiddleware):
 
     @override
     async def abefore_agent(self, state, runtime: Runtime) -> dict | None:
-        return self._inject(state)
+        # _inject() performs synchronous file I/O (memory JSON loading) and
+        # potentially blocking network calls (tiktoken encoding download on
+        # first use).  Offload to a thread so the event loop is never blocked
+        # — a blocking call here starves all concurrent HTTP handlers (auth,
+        # SSE heartbeats, etc.).  See issue #3402.
+        #
+        # Bounded timeout: if startup warm-up failed silently (e.g. network
+        # blip during deploy), the first request's cold tiktoken download can
+        # block for tens of minutes (OS TCP timeout).  Time-box injection so
+        # the request degrades gracefully (no memory context) rather than
+        # hanging.
+        try:
+            return await asyncio.wait_for(
+                asyncio.to_thread(self._inject, state),
+                timeout=_INJECT_TIMEOUT_SECONDS,
+            )
+        except TimeoutError:
+            logger.warning(
+                "DynamicContextMiddleware: injection timed out (%.1fs); skipping memory/date injection for this turn",
+                _INJECT_TIMEOUT_SECONDS,
+            )
+            return None

backend/packages/harness/deerflow/agents/middlewares/llm_error_handling_middleware.py

@@ -62,6 +62,41 @@ _AUTH_PATTERNS = (
     "未授权",
 )
 
+# Per-exception retry budget overrides.
+#
+# Some transient errors are retriable in principle but expensive to retry at
+# the default budget. StreamChunkTimeoutError in particular fires after the
+# upstream provider has already stalled for `stream_chunk_timeout` seconds
+# (typically 120-240s); a full 3-attempt loop can therefore stack 6-12 minutes
+# of dead air before surfacing the failure to the user. We keep exactly one
+# retry (cheap reconnect that catches genuine transient TCP blips) and then
+# fail fast — the same buffered payload is overwhelmingly likely to fail
+# again at the upstream provider for the same reason.
+#
+# Keys are exception class *names* (not classes) so we don't introduce
+# import-time coupling on optional dependencies like langchain-openai. The
+# value is the absolute max attempt count, NOT additional retries — so a
+# value of 2 means "1 first attempt + 1 retry" (the CR-requested
+# "keep one retry" behavior).
+_RETRY_BUDGET_OVERRIDES: dict[str, int] = {
+    "StreamChunkTimeoutError": 2,
+}
+
+# Exception class names that indicate the upstream stream-chunk watchdog
+# fired because the model stalled mid-flight. These deserve a more specific
+# user-facing message than the generic "temporarily unavailable" copy,
+# because the typical root cause is a long tool-call serialization stalling
+# the upstream stream — and the most actionable advice we can give the user
+# is "ask for a shorter / split output" rather than "wait and retry".
+# Generic connection drops (httpx RemoteProtocolError / ReadError) are
+# intentionally excluded: they routinely fire on transient network blips
+# with normal payloads, where the "split the work" guidance is misleading.
+_STREAM_DROP_EXCEPTIONS: frozenset[str] = frozenset(
+    {
+        "StreamChunkTimeoutError",
+    }
+)
+
 
 class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
     """Retry transient LLM errors and surface graceful assistant messages."""
@@ -83,6 +118,18 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
         self._circuit_state = "closed"
         self._circuit_probe_in_flight = False
 
+    def _max_attempts_for(self, exc: BaseException) -> int:
+        """Return the effective max attempt count for this exception.
+
+        Falls back to `self.retry_max_attempts` unless the exception class name
+        appears in the per-exception override table.
+        """
+        override = _RETRY_BUDGET_OVERRIDES.get(type(exc).__name__)
+        if override is None:
+            return self.retry_max_attempts
+
+        return min(override, self.retry_max_attempts)
+
     def _check_circuit(self) -> bool:
         """Returns True if circuit is OPEN (fast fail), False otherwise."""
         with self._circuit_lock:
@@ -153,6 +200,7 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
             "InternalServerError",
             "ReadError",  # httpx.ReadError: connection dropped mid-stream
             "RemoteProtocolError",  # httpx: server closed connection unexpectedly
+            "StreamChunkTimeoutError",  # langchain-openai: chunk gap exceeded stream_chunk_timeout
         }:
             return True, "transient"
         if status_code in _RETRIABLE_STATUS_CODES:
@@ -177,6 +225,24 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
     def _build_circuit_breaker_message(self) -> str:
         return "The configured LLM provider is currently unavailable due to continuous failures. Circuit breaker is engaged to protect the system. Please wait a moment before trying again."
 
+    def _build_error_fallback_message(
+        self,
+        content: str,
+        *,
+        error_type: str,
+        reason: str,
+        detail: str,
+    ) -> AIMessage:
+        return AIMessage(
+            content=content,
+            additional_kwargs={
+                "deerflow_error_fallback": True,
+                "error_type": error_type,
+                "error_reason": reason,
+                "error_detail": detail,
+            },
+        )
+
     def _build_user_message(self, exc: BaseException, reason: str) -> str:
         detail = _extract_error_detail(exc)
         if reason == "quota":
@@ -184,9 +250,31 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
         if reason == "auth":
             return "The configured LLM provider rejected the request because authentication or access is invalid. Please check the provider credentials and try again."
         if reason in {"busy", "transient"}:
+            # Stream-drop failures (chunk-gap timeout, peer-closed connection,
+            # raw read error) almost always point at a single oversized
+            # tool-call payload — the model spent so long serializing JSON
+            # arguments that the upstream provider buffered and the stream
+            # gap exceeded `stream_chunk_timeout`. Surfacing this distinct
+            # cause lets the user split or shorten their next request
+            # instead of helplessly retrying the same prompt.
+            if type(exc).__name__ in _STREAM_DROP_EXCEPTIONS:
+                return (
+                    "The model's streaming response was interrupted before it could "
+                    "finish. This usually happens when a single response or tool call "
+                    "is very large — please ask the assistant to split the work into "
+                    "smaller steps, or shorten the requested output, and try again."
+                )
             return "The configured LLM provider is temporarily unavailable after multiple retries. Please wait a moment and continue the conversation."
         return f"LLM request failed: {detail}"
 
+    def _build_user_fallback_message(self, exc: BaseException, reason: str) -> AIMessage:
+        return self._build_error_fallback_message(
+            self._build_user_message(exc, reason),
+            error_type=type(exc).__name__,
+            reason=reason,
+            detail=_extract_error_detail(exc),
+        )
+
     def _emit_retry_event(self, attempt: int, wait_ms: int, reason: str) -> None:
         try:
             from langgraph.config import get_stream_writer
@@ -212,7 +300,12 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
         handler: Callable[[ModelRequest], ModelResponse],
     ) -> ModelCallResult:
         if self._check_circuit():
-            return AIMessage(content=self._build_circuit_breaker_message())
+            return self._build_error_fallback_message(
+                self._build_circuit_breaker_message(),
+                error_type="CircuitBreakerOpen",
+                reason="circuit_open",
+                detail="LLM circuit breaker is open",
+            )
 
         attempt = 1
         while True:
@@ -228,7 +321,8 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
                 raise
             except Exception as exc:
                 retriable, reason = self._classify_error(exc)
-                if retriable and attempt < self.retry_max_attempts:
+                max_attempts = self._max_attempts_for(exc)
+                if retriable and attempt < max_attempts:
                     wait_ms = self._build_retry_delay_ms(attempt, exc)
                     logger.warning(
                         "Transient LLM error on attempt %d/%d; retrying in %dms: %s",
@@ -249,7 +343,7 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
                 )
                 if retriable:
                     self._record_failure()
-                return AIMessage(content=self._build_user_message(exc, reason))
+                return self._build_user_fallback_message(exc, reason)
 
     @override
     async def awrap_model_call(
@@ -258,7 +352,12 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
         handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
     ) -> ModelCallResult:
         if self._check_circuit():
-            return AIMessage(content=self._build_circuit_breaker_message())
+            return self._build_error_fallback_message(
+                self._build_circuit_breaker_message(),
+                error_type="CircuitBreakerOpen",
+                reason="circuit_open",
+                detail="LLM circuit breaker is open",
+            )
 
         attempt = 1
         while True:
@@ -274,7 +373,8 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
                 raise
             except Exception as exc:
                 retriable, reason = self._classify_error(exc)
-                if retriable and attempt < self.retry_max_attempts:
+                max_attempts = self._max_attempts_for(exc)
+                if retriable and attempt < max_attempts:
                     wait_ms = self._build_retry_delay_ms(attempt, exc)
                     logger.warning(
                         "Transient LLM error on attempt %d/%d; retrying in %dms: %s",
@@ -295,7 +395,7 @@ class LLMErrorHandlingMiddleware(AgentMiddleware[AgentState]):
                 )
                 if retriable:
                     self._record_failure()
-                return AIMessage(content=self._build_user_message(exc, reason))
+                return self._build_user_fallback_message(exc, reason)
 
 
 def _matches_any(detail: str, patterns: tuple[str, ...]) -> bool:

backend/packages/harness/deerflow/agents/middlewares/loop_detection_middleware.py

@@ -6,10 +6,36 @@ arguments indefinitely until the recursion limit kills the run.
 Detection strategy:
   1. After each model response, hash the tool calls (name + args).
   2. Track recent hashes in a sliding window.
-  3. If the same hash appears >= warn_threshold times, inject a
-     "you are repeating yourself — wrap up" system message (once per hash).
+  3. If the same hash appears >= warn_threshold times, queue a
+     "you are repeating yourself — wrap up" warning for the current
+     thread/run. The warning is **injected at the next model call** (in
+     ``wrap_model_call``) as a ``HumanMessage`` appended to the message
+     list, *after* all ToolMessage responses to the previous
+     AIMessage(tool_calls).
   4. If it appears >= hard_limit times, strip all tool_calls from the
      response so the agent is forced to produce a final text answer.
+
+Why the warning is injected at ``wrap_model_call`` instead of
+``after_model``:
+
+  ``after_model`` fires immediately after the model emits an
+  ``AIMessage`` that may carry ``tool_calls``. The tools node has not
+  run yet, so no matching ``ToolMessage`` exists in the history. Any
+  message we add here lands *between* the assistant's tool_calls and
+  their responses. OpenAI/Moonshot reject the next request with
+  ``"tool_call_ids did not have response messages"`` because their
+  validators require the assistant's tool_calls to be followed
+  immediately by tool messages. Anthropic also disallows mid-stream
+  ``SystemMessage``. By deferring the warning to ``wrap_model_call``,
+  every prior ToolMessage is already present in the request's message
+  list and the warning is appended at the end — pairing intact, no
+  ``AIMessage`` semantics are mutated.
+
+Queued warnings are intentionally transient. If a run ends before the
+next model request drains a queued warning, ``after_agent`` drops it
+instead of carrying it into a later invocation for the same thread. The
+hard-stop path still forces termination when the configured safety limit
+is reached.
 """
 
 from __future__ import annotations
@@ -19,11 +45,14 @@ import json
 import logging
 import threading
 from collections import OrderedDict, defaultdict
+from collections.abc import Awaitable, Callable
 from copy import deepcopy
 from typing import TYPE_CHECKING, override
 
 from langchain.agents import AgentState
 from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
+from langchain_core.messages import HumanMessage
 from langgraph.runtime import Runtime
 
 if TYPE_CHECKING:
@@ -38,6 +67,7 @@ _DEFAULT_WINDOW_SIZE = 20  # track last N tool calls
 _DEFAULT_MAX_TRACKED_THREADS = 100  # LRU eviction limit
 _DEFAULT_TOOL_FREQ_WARN = 30  # warn after 30 calls to the same tool type
 _DEFAULT_TOOL_FREQ_HARD_LIMIT = 50  # force-stop after 50 calls to the same tool type
+_MAX_PENDING_WARNINGS_PER_RUN = 4
 
 
 def _normalize_tool_call_args(raw_args: object) -> tuple[dict, str | None]:
@@ -195,6 +225,12 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
         self._warned: dict[str, set[str]] = defaultdict(set)
         self._tool_freq: dict[str, dict[str, int]] = defaultdict(lambda: defaultdict(int))
         self._tool_freq_warned: dict[str, set[str]] = defaultdict(set)
+        # Per-thread/run queue of warnings to inject at the next model call.
+        # Populated by ``after_model`` (detection) and drained by
+        # ``wrap_model_call`` (injection); see module docstring.
+        self._pending_warnings: dict[tuple[str, str], list[str]] = defaultdict(list)
+        self._pending_warning_touch_order: OrderedDict[tuple[str, str], None] = OrderedDict()
+        self._max_pending_warning_keys = max(1, self.max_tracked_threads * 2)
 
     @classmethod
     def from_config(cls, config: LoopDetectionConfig) -> LoopDetectionMiddleware:
@@ -213,9 +249,20 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
         """Extract thread_id from runtime context for per-thread tracking."""
         thread_id = runtime.context.get("thread_id") if runtime.context else None
         if thread_id:
-            return thread_id
+            return str(thread_id)
         return "default"
 
+    def _get_run_id(self, runtime: Runtime) -> str:
+        """Extract run_id from runtime context for per-run warning scoping."""
+        run_id = runtime.context.get("run_id") if runtime.context else None
+        if run_id:
+            return str(run_id)
+        return "default"
+
+    def _pending_key(self, runtime: Runtime) -> tuple[str, str]:
+        """Return the pending-warning key for the current thread/run."""
+        return self._get_thread_id(runtime), self._get_run_id(runtime)
+
     def _evict_if_needed(self) -> None:
         """Evict least recently used threads if over the limit.
 
@@ -226,8 +273,52 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
             self._warned.pop(evicted_id, None)
             self._tool_freq.pop(evicted_id, None)
             self._tool_freq_warned.pop(evicted_id, None)
+            for key in list(self._pending_warnings):
+                if key[0] == evicted_id:
+                    self._drop_pending_warning_key_locked(key)
             logger.debug("Evicted loop tracking for thread %s (LRU)", evicted_id)
 
+    def _drop_pending_warning_key_locked(self, key: tuple[str, str]) -> None:
+        """Drop all pending-warning bookkeeping for one thread/run key.
+
+        Must be called while holding self._lock.
+        """
+        self._pending_warnings.pop(key, None)
+        self._pending_warning_touch_order.pop(key, None)
+
+    def _touch_pending_warning_key_locked(self, key: tuple[str, str]) -> None:
+        """Mark a pending-warning key as recently used.
+
+        Must be called while holding self._lock.
+        """
+        self._pending_warning_touch_order[key] = None
+        self._pending_warning_touch_order.move_to_end(key)
+
+    def _prune_pending_warning_state_locked(self, protected_key: tuple[str, str]) -> None:
+        """Cap pending-warning state across abnormal or concurrent runs.
+
+        Must be called while holding self._lock.
+        """
+        overflow = len(self._pending_warning_touch_order) - self._max_pending_warning_keys
+        if overflow <= 0:
+            return
+
+        candidates = [key for key in self._pending_warning_touch_order if key != protected_key]
+        for key in candidates[:overflow]:
+            self._drop_pending_warning_key_locked(key)
+
+    def _queue_pending_warning(self, runtime: Runtime, warning: str) -> None:
+        """Queue one transient warning for the current thread/run with caps."""
+        pending_key = self._pending_key(runtime)
+        with self._lock:
+            warnings = self._pending_warnings[pending_key]
+            if warning not in warnings:
+                warnings.append(warning)
+            if len(warnings) > _MAX_PENDING_WARNINGS_PER_RUN:
+                del warnings[: len(warnings) - _MAX_PENDING_WARNINGS_PER_RUN]
+            self._touch_pending_warning_key_locked(pending_key)
+            self._prune_pending_warning_state_locked(protected_key=pending_key)
+
     def _track_and_check(self, state: AgentState, runtime: Runtime) -> tuple[str | None, bool]:
         """Track tool calls and check for loops.
 
@@ -268,6 +359,12 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
             if len(history) > self.window_size:
                 history[:] = history[-self.window_size :]
 
+            warned_hashes = self._warned.get(thread_id)
+            if warned_hashes is not None:
+                warned_hashes.intersection_update(history)
+                if not warned_hashes:
+                    self._warned.pop(thread_id, None)
+
             count = history.count(call_hash)
             tool_names = [tc.get("name", "?") for tc in tool_calls]
 
@@ -381,7 +478,10 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
         warning, hard_stop = self._track_and_check(state, runtime)
 
         if hard_stop:
-            # Strip tool_calls from the last AIMessage to force text output
+            # Strip tool_calls from the last AIMessage to force text output.
+            # Once tool_calls are stripped, the AIMessage no longer requires
+            # matching ToolMessage responses, so mutating it in place here
+            # is safe for OpenAI/Moonshot pairing validators.
             messages = state.get("messages", [])
             last_msg = messages[-1]
             content = self._append_text(last_msg.content, warning or _HARD_STOP_MSG)
@@ -389,33 +489,48 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
             return {"messages": [stripped_msg]}
 
         if warning:
-            # WORKAROUND for v2.0-m1 — see #2724.
-            #
-            # Append the warning to the AIMessage content instead of
-            # injecting a separate HumanMessage. Inserting any non-tool
-            # message between an AIMessage(tool_calls=...) and its
-            # ToolMessage responses breaks OpenAI/Moonshot strict pairing
-            # validation ("tool_call_ids did not have response messages")
-            # because the tools node has not run yet at after_model time.
-            # tool_calls are preserved so the tools node still executes.
-            #
-            # This is a temporary mitigation: mutating an existing
-            # AIMessage to carry framework-authored text leaks loop-warning
-            # text into downstream consumers (MemoryMiddleware fact
-            # extraction, TitleMiddleware, telemetry, model replay) as if
-            # the model said it. The proper fix is to defer warning
-            # injection from after_model to wrap_model_call so every prior
-            # ToolMessage is already in the request — see RFC #2517 (which
-            # lists "loop intervention does not leave invalid
-            # tool-call/tool-message state" as acceptance criteria) and
-            # the prototype on `fix/loop-detection-tool-call-pairing`.
-            messages = state.get("messages", [])
-            last_msg = messages[-1]
-            patched_msg = last_msg.model_copy(update={"content": self._append_text(last_msg.content, warning)})
-            return {"messages": [patched_msg]}
+            # Defer injection to the next model call. We must NOT alter the
+            # AIMessage(tool_calls=...) here (would put framework words in
+            # the model's mouth, polluting downstream consumers like
+            # MemoryMiddleware), nor insert a separate non-tool message
+            # (would break OpenAI/Moonshot tool-call pairing because the
+            # tools node has not produced ToolMessage responses yet). The
+            # warning is delivered via ``wrap_model_call`` below.
+            self._queue_pending_warning(runtime, warning)
+            return None
 
         return None
 
+    def _clear_other_run_pending_warnings(self, runtime: Runtime) -> None:
+        """Drop stale pending warnings for previous runs in this thread."""
+        thread_id, current_run_id = self._pending_key(runtime)
+        with self._lock:
+            for key in list(self._pending_warnings):
+                if key[0] == thread_id and key[1] != current_run_id:
+                    self._drop_pending_warning_key_locked(key)
+
+    def _clear_current_run_pending_warnings(self, runtime: Runtime) -> None:
+        """Drop pending warnings owned by the current thread/run."""
+        pending_key = self._pending_key(runtime)
+        with self._lock:
+            self._drop_pending_warning_key_locked(pending_key)
+
+    @staticmethod
+    def _format_warning_message(warnings: list[str]) -> str:
+        """Merge pending warnings into one prompt message."""
+        deduped = list(dict.fromkeys(warnings))
+        return "\n\n".join(deduped)
+
+    @override
+    def before_agent(self, state: AgentState, runtime: Runtime) -> dict | None:
+        self._clear_other_run_pending_warnings(runtime)
+        return None
+
+    @override
+    async def abefore_agent(self, state: AgentState, runtime: Runtime) -> dict | None:
+        self._clear_other_run_pending_warnings(runtime)
+        return None
+
     @override
     def after_model(self, state: AgentState, runtime: Runtime) -> dict | None:
         return self._apply(state, runtime)
@@ -424,6 +539,59 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
     async def aafter_model(self, state: AgentState, runtime: Runtime) -> dict | None:
         return self._apply(state, runtime)
 
+    @override
+    def after_agent(self, state: AgentState, runtime: Runtime) -> dict | None:
+        self._clear_current_run_pending_warnings(runtime)
+        return None
+
+    @override
+    async def aafter_agent(self, state: AgentState, runtime: Runtime) -> dict | None:
+        self._clear_current_run_pending_warnings(runtime)
+        return None
+
+    def _drain_pending_warnings(self, runtime: Runtime) -> list[str]:
+        """Pop and return all queued warnings for *runtime*'s thread/run."""
+        pending_key = self._pending_key(runtime)
+        with self._lock:
+            warnings = self._pending_warnings.pop(pending_key, [])
+            self._pending_warning_touch_order.pop(pending_key, None)
+        return warnings
+
+    def _augment_request(self, request: ModelRequest) -> ModelRequest:
+        """Append queued loop warnings (if any) to the outgoing message list.
+
+        The warning is placed *after* every existing message, including the
+        ToolMessage responses to the previous AIMessage(tool_calls). This
+        keeps ``assistant tool_calls -> tool_messages`` pairing intact for
+        OpenAI/Moonshot, avoids the Anthropic mid-stream SystemMessage
+        restriction (we use HumanMessage), and never mutates an existing
+        AIMessage.
+        """
+        warnings = self._drain_pending_warnings(request.runtime)
+        if not warnings:
+            return request
+        new_messages = [
+            *request.messages,
+            HumanMessage(content=self._format_warning_message(warnings), name="loop_warning"),
+        ]
+        return request.override(messages=new_messages)
+
+    @override
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelCallResult:
+        return handler(self._augment_request(request))
+
+    @override
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> ModelCallResult:
+        return await handler(self._augment_request(request))
+
     def reset(self, thread_id: str | None = None) -> None:
         """Clear tracking state. If thread_id given, clear only that thread."""
         with self._lock:
@@ -432,8 +600,13 @@ class LoopDetectionMiddleware(AgentMiddleware[AgentState]):
                 self._warned.pop(thread_id, None)
                 self._tool_freq.pop(thread_id, None)
                 self._tool_freq_warned.pop(thread_id, None)
+                for key in list(self._pending_warnings):
+                    if key[0] == thread_id:
+                        self._drop_pending_warning_key_locked(key)
             else:
                 self._history.clear()
                 self._warned.clear()
                 self._tool_freq.clear()
                 self._tool_freq_warned.clear()
+                self._pending_warnings.clear()
+                self._pending_warning_touch_order.clear()

backend/packages/harness/deerflow/agents/middlewares/safety_finish_reason_middleware.py

@@ -0,0 +1,317 @@
+"""Suppress tool execution when the provider safety-terminated the response.
+
+Background — see issue bytedance/deer-flow#3028.
+
+Some providers (OpenAI ``finish_reason='content_filter'``, Anthropic
+``stop_reason='refusal'``, Gemini ``finish_reason='SAFETY'`` ...) can stop
+generation mid-stream while still returning partially-formed ``tool_calls``.
+LangChain's tool router treats any AIMessage with a non-empty ``tool_calls``
+field as "go execute these", so half-truncated arguments — e.g. a markdown
+``write_file`` that stops in the middle of a sentence — get dispatched as if
+they were complete. The agent then sees the truncated file, tries to fix it,
+gets filtered again, and loops.
+
+This middleware sits at ``after_model`` and gates that behaviour: when a
+configured ``SafetyTerminationDetector`` fires *and* the AIMessage carries
+tool calls, we strip the tool calls (both structured and raw provider
+payloads), append a user-facing explanation, and stash observability fields
+in ``additional_kwargs.safety_termination`` so logs, traces, and SSE
+consumers can see what happened.
+
+Hook choice: ``after_model`` (not ``wrap_model_call``) because the response
+is a *normal* return — not an exception — and we want to participate in the
+same after-model chain as ``LoopDetectionMiddleware``, with which we share
+the same tool-call-suppression mechanic but a different trigger.
+
+Placement: register *after* ``LoopDetectionMiddleware`` in the middleware
+list. LangChain factory wires ``after_model`` edges in reverse list order
+(``langchain/agents/factory.py:add_edge("model", middleware_w_after_model[-1])``,
+then walks ``range(len-1, 0, -1)``), so the *last* registered middleware is
+the *first* to observe the model output. Registering Safety after Loop
+means Safety sees the raw response first, clears tool calls if it fires,
+and Loop then accounts against the cleaned message.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, override
+
+from langchain.agents import AgentState
+from langchain.agents.middleware import AgentMiddleware
+from langchain_core.messages import AIMessage
+from langgraph.runtime import Runtime
+
+from deerflow.agents.middlewares.safety_termination_detectors import (
+    SafetyTermination,
+    SafetyTerminationDetector,
+    default_detectors,
+)
+from deerflow.agents.middlewares.tool_call_metadata import clone_ai_message_with_tool_calls
+
+if TYPE_CHECKING:
+    from deerflow.config.safety_finish_reason_config import SafetyFinishReasonConfig
+
+logger = logging.getLogger(__name__)
+
+
+_USER_FACING_MESSAGE = (
+    "The model provider stopped this response with a safety-related signal "
+    "({reason_field}={reason_value!r}, detector={detector!r}). Any tool "
+    "calls produced in this turn were suppressed because their arguments "
+    "may be truncated and unsafe to execute. Please rephrase the request "
+    "or ask for a narrower output."
+)
+
+
+class SafetyFinishReasonMiddleware(AgentMiddleware[AgentState]):
+    """Strip tool_calls from AIMessages flagged by a SafetyTerminationDetector."""
+
+    def __init__(self, detectors: list[SafetyTerminationDetector] | None = None) -> None:
+        super().__init__()
+        # Copy so caller mutations after construction don't leak into us.
+        self._detectors: list[SafetyTerminationDetector] = list(detectors) if detectors else default_detectors()
+
+    @classmethod
+    def from_config(cls, config: SafetyFinishReasonConfig) -> SafetyFinishReasonMiddleware:
+        """Construct from validated Pydantic config, honouring the
+        reflection-loaded detector list when provided.
+
+        An explicit empty list is intentionally rejected — it would silently
+        disable detection while leaving the middleware in the chain, which
+        is the worst of both worlds. Use ``enabled: false`` instead.
+        """
+        if config.detectors is None:
+            return cls()
+
+        if not config.detectors:
+            raise ValueError("safety_finish_reason.detectors must be omitted (use built-ins) or contain at least one entry; use enabled=false to disable the middleware entirely.")
+
+        from deerflow.reflection import resolve_variable
+
+        detectors: list[SafetyTerminationDetector] = []
+        for entry in config.detectors:
+            detector_cls = resolve_variable(entry.use)
+            kwargs = dict(entry.config) if entry.config else {}
+            detector = detector_cls(**kwargs)
+            if not isinstance(detector, SafetyTerminationDetector):
+                raise TypeError(f"{entry.use} did not produce a SafetyTerminationDetector (got {type(detector).__name__}); ensure it has a `name` attribute and a `detect(message)` method")
+            detectors.append(detector)
+        return cls(detectors=detectors)
+
+    # ----- detection -------------------------------------------------------
+
+    def _detect(self, message: AIMessage) -> SafetyTermination | None:
+        for detector in self._detectors:
+            try:
+                hit = detector.detect(message)
+            except Exception:  # noqa: BLE001 - never let a buggy detector break the agent run
+                logger.exception("SafetyTerminationDetector %r raised; treating as no-match", getattr(detector, "name", type(detector).__name__))
+                continue
+            if hit is not None:
+                return hit
+        return None
+
+    # ----- message rewriting ----------------------------------------------
+
+    @staticmethod
+    def _append_user_message(content: object, text: str) -> str | list:
+        """Append a plain-text explanation to AIMessage content.
+
+        Mirrors ``LoopDetectionMiddleware._append_text`` so list-content
+        responses (Anthropic thinking blocks, vLLM reasoning splits) keep
+        their structure instead of being string-coerced into a TypeError.
+        """
+        if content is None or content == "":
+            return text
+        if isinstance(content, list):
+            return [*content, {"type": "text", "text": f"\n\n{text}"}]
+        if isinstance(content, str):
+            return content + f"\n\n{text}"
+        return str(content) + f"\n\n{text}"
+
+    def _build_suppressed_message(
+        self,
+        message: AIMessage,
+        termination: SafetyTermination,
+    ) -> AIMessage:
+        suppressed_names = [tc.get("name") or "unknown" for tc in (message.tool_calls or [])]
+        explanation = _USER_FACING_MESSAGE.format(
+            reason_field=termination.reason_field,
+            reason_value=termination.reason_value,
+            detector=termination.detector,
+        )
+        new_content = self._append_user_message(message.content, explanation)
+
+        # clone_ai_message_with_tool_calls handles structured tool_calls,
+        # raw additional_kwargs.tool_calls, and function_call in one shot.
+        # It only rewrites finish_reason when the old value was "tool_calls",
+        # which is not our case — content_filter / refusal / SAFETY stay put
+        # so downstream SSE / converters keep seeing the real provider reason.
+        cleared = clone_ai_message_with_tool_calls(message, [], content=new_content)
+
+        # Re-clone additional_kwargs so we don't accidentally mutate the
+        # dict returned by clone_ai_message_with_tool_calls (which already
+        # made a shallow copy, but downstream model_copy still references
+        # it). Then stamp the observability record.
+        kwargs = dict(getattr(cleared, "additional_kwargs", None) or {})
+        kwargs["safety_termination"] = {
+            "detector": termination.detector,
+            "reason_field": termination.reason_field,
+            "reason_value": termination.reason_value,
+            "suppressed_tool_call_count": len(suppressed_names),
+            "suppressed_tool_call_names": suppressed_names,
+            "extras": dict(termination.extras) if termination.extras else {},
+        }
+        return cleared.model_copy(update={"additional_kwargs": kwargs})
+
+    # ----- observability ---------------------------------------------------
+
+    def _emit_event(
+        self,
+        termination: SafetyTermination,
+        suppressed_names: list[str],
+        runtime: Runtime,
+    ) -> None:
+        """Notify SSE consumers (e.g. the web UI) that a tool turn was
+        suppressed so they can reconcile any "tool starting..." placeholders
+        already streamed to the user. Failures are logged at debug and
+        ignored — this is a best-effort signal."""
+        try:
+            from langgraph.config import get_stream_writer
+
+            writer = get_stream_writer()
+        except Exception:  # noqa: BLE001
+            logger.debug("get_stream_writer unavailable; skipping safety_termination event", exc_info=True)
+            return
+
+        thread_id = None
+        if runtime is not None and getattr(runtime, "context", None):
+            thread_id = runtime.context.get("thread_id") if isinstance(runtime.context, dict) else None
+
+        try:
+            writer(
+                {
+                    "type": "safety_termination",
+                    "detector": termination.detector,
+                    "reason_field": termination.reason_field,
+                    "reason_value": termination.reason_value,
+                    "suppressed_tool_call_count": len(suppressed_names),
+                    "suppressed_tool_call_names": suppressed_names,
+                    "thread_id": thread_id,
+                }
+            )
+        except Exception:  # noqa: BLE001
+            logger.debug("Failed to emit safety_termination stream event", exc_info=True)
+
+    def _record_audit_event(
+        self,
+        termination: SafetyTermination,
+        message,
+        tool_calls: list[dict],
+        runtime: Runtime,
+    ) -> None:
+        """Write a ``middleware:safety_termination`` record to RunEventStore
+        for post-run auditability.
+
+        The custom stream event in ``_emit_event`` is consumed by live SSE
+        clients and disappears after the run; this event is persisted so an
+        operator can answer "which runs were safety-suppressed today?" from
+        a single SQL query without joining the message body. Worker exposes
+        the run-scoped ``RunJournal`` via ``runtime.context["__run_journal"]``;
+        absent in unit-test / subagent / no-event-store paths, in which case
+        we silently skip.
+
+        Tool **arguments** are deliberately **not** recorded — those are the
+        very content the provider filtered; persisting them would defeat the
+        purpose of the safety filter. Names / count / ids are sufficient for
+        audit and debugging (issue #3028 review).
+        """
+        journal = None
+        if runtime is not None and getattr(runtime, "context", None):
+            context = runtime.context
+            if isinstance(context, dict):
+                journal = context.get("__run_journal")
+        if journal is None:
+            return
+
+        suppressed_names = [tc.get("name") or "unknown" for tc in tool_calls]
+        suppressed_ids = [tc.get("id") for tc in tool_calls if tc.get("id")]
+
+        changes = {
+            "detector": termination.detector,
+            "reason_field": termination.reason_field,
+            "reason_value": termination.reason_value,
+            "suppressed_tool_call_count": len(tool_calls),
+            "suppressed_tool_call_names": suppressed_names,
+            "suppressed_tool_call_ids": suppressed_ids,
+            "message_id": getattr(message, "id", None),
+            "extras": dict(termination.extras) if termination.extras else {},
+        }
+
+        try:
+            journal.record_middleware(
+                tag="safety_termination",
+                name=type(self).__name__,
+                hook="after_model",
+                action="suppress_tool_calls",
+                changes=changes,
+            )
+        except Exception:  # noqa: BLE001
+            # Audit-event persistence must never break agent execution.
+            logger.debug("Failed to record middleware:safety_termination event", exc_info=True)
+
+    # ----- main apply ------------------------------------------------------
+
+    def _apply(self, state: AgentState, runtime: Runtime) -> dict | None:
+        messages = state.get("messages", [])
+        if not messages:
+            return None
+
+        last = messages[-1]
+        if not isinstance(last, AIMessage):
+            return None
+
+        # Issue scope: only intervene when there's something to suppress.
+        # ``content_filter`` without tool_calls is allowed through unchanged
+        # so the partial text response (if any) reaches the user naturally.
+        tool_calls = last.tool_calls
+        if not tool_calls:
+            return None
+
+        termination = self._detect(last)
+        if termination is None:
+            return None
+
+        patched = self._build_suppressed_message(last, termination)
+
+        thread_id = None
+        if runtime is not None and getattr(runtime, "context", None):
+            thread_id = runtime.context.get("thread_id") if isinstance(runtime.context, dict) else None
+
+        logger.warning(
+            "Provider safety termination detected — suppressed %d tool call(s)",
+            len(tool_calls),
+            extra={
+                "thread_id": thread_id,
+                "detector": termination.detector,
+                "reason_field": termination.reason_field,
+                "reason_value": termination.reason_value,
+                "suppressed_tool_call_names": [tc.get("name") for tc in tool_calls],
+            },
+        )
+
+        self._emit_event(termination, [tc.get("name") or "unknown" for tc in tool_calls], runtime)
+        self._record_audit_event(termination, last, list(tool_calls), runtime)
+
+        return {"messages": [patched]}
+
+    # ----- hooks -----------------------------------------------------------
+
+    @override
+    def after_model(self, state: AgentState, runtime: Runtime) -> dict | None:
+        return self._apply(state, runtime)
+
+    @override
+    async def aafter_model(self, state: AgentState, runtime: Runtime) -> dict | None:
+        return self._apply(state, runtime)

backend/packages/harness/deerflow/agents/middlewares/safety_termination_detectors.py

@@ -0,0 +1,237 @@
+"""Detectors for provider-side safety termination signals.
+
+Different LLM providers signal "I stopped this response for safety reasons"
+through different fields with different values. This module defines a small
+strategy interface and three built-in detectors that cover the major
+providers DeerFlow supports today. New providers (Wenxin, Hunyuan, Bedrock
+adapters, in-house gateways, ...) can be added by implementing
+``SafetyTerminationDetector`` and wiring it through
+``config.yaml: safety_finish_reason.detectors``.
+
+The middleware that consumes these detectors lives in
+``safety_finish_reason_middleware.py``.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Protocol, runtime_checkable
+
+from langchain_core.messages import AIMessage
+
+
+@dataclass(frozen=True)
+class SafetyTermination:
+    """A detected safety-related termination signal.
+
+    Attributes:
+        detector: Name of the detector that produced this result. Used for
+            observability so operators can see which provider rule fired.
+        reason_field: The message metadata field that carried the signal
+            (e.g. ``finish_reason``, ``stop_reason``).
+        reason_value: The actual value of that field
+            (e.g. ``content_filter``, ``refusal``, ``SAFETY``).
+        extras: Provider-specific metadata that may help downstream
+            consumers (e.g. Azure OpenAI content_filter_results, Gemini
+            safety_ratings). Detectors are free to populate or skip this.
+    """
+
+    detector: str
+    reason_field: str
+    reason_value: str
+    extras: dict[str, Any] = field(default_factory=dict)
+
+
+@runtime_checkable
+class SafetyTerminationDetector(Protocol):
+    """Strategy interface for provider safety termination detection."""
+
+    name: str
+
+    def detect(self, message: AIMessage) -> SafetyTermination | None:
+        """Return a SafetyTermination if *message* indicates provider safety
+        termination, otherwise return ``None``.
+
+        Implementations must be side-effect free and tolerant of missing or
+        oddly-typed metadata — detectors run on every model response.
+        """
+        ...
+
+
+def _get_metadata_value(message: AIMessage, field_name: str) -> str | None:
+    """Read a string-typed value from either ``response_metadata`` or
+    ``additional_kwargs``.
+
+    LangChain provider adapters are inconsistent about where they stash
+    provider stop signals. Most modern adapters use ``response_metadata``,
+    but some legacy / passthrough paths still surface them via
+    ``additional_kwargs``. We check both, in that order, and only accept
+    string values — Pydantic enums or dicts are ignored so we never raise
+    on malformed inputs.
+    """
+    for container_name in ("response_metadata", "additional_kwargs"):
+        container = getattr(message, container_name, None) or {}
+        if not isinstance(container, dict):
+            continue
+        value = container.get(field_name)
+        if isinstance(value, str) and value:
+            return value
+    return None
+
+
+class OpenAICompatibleContentFilterDetector:
+    """OpenAI-compatible content_filter signal.
+
+    Covers OpenAI, Azure OpenAI, Moonshot/Kimi, DeepSeek, Mistral, vLLM,
+    Qwen (OpenAI-compatible mode), and any other adapter that follows the
+    OpenAI ``finish_reason`` convention.
+
+    Some Chinese providers ship custom OpenAI-compatible gateways that use
+    alternative tokens like ``sensitive`` or ``violation``. Extend the set
+    via the ``finish_reasons`` kwarg in config.
+    """
+
+    name = "openai_compatible_content_filter"
+
+    def __init__(self, finish_reasons: list[str] | tuple[str, ...] | None = None) -> None:
+        configured = finish_reasons if finish_reasons is not None else ("content_filter",)
+        self._finish_reasons: frozenset[str] = frozenset(r.lower() for r in configured)
+
+    def detect(self, message: AIMessage) -> SafetyTermination | None:
+        value = _get_metadata_value(message, "finish_reason")
+        if value is None or value.lower() not in self._finish_reasons:
+            return None
+
+        extras: dict[str, Any] = {}
+        # Azure OpenAI ships a structured content_filter_results block; carry it
+        # through so operators can see *what* was filtered without re-tracing.
+        response_metadata = getattr(message, "response_metadata", None) or {}
+        if isinstance(response_metadata, dict):
+            filter_results = response_metadata.get("content_filter_results")
+            if filter_results:
+                extras["content_filter_results"] = filter_results
+
+        return SafetyTermination(
+            detector=self.name,
+            reason_field="finish_reason",
+            reason_value=value,
+            extras=extras,
+        )
+
+
+class AnthropicRefusalDetector:
+    """Anthropic ``stop_reason == "refusal"`` signal.
+
+    Anthropic models surface safety refusals via a dedicated ``stop_reason``
+    rather than ``finish_reason``. See:
+    https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/handle-streaming-refusals
+    """
+
+    name = "anthropic_refusal"
+
+    def __init__(self, stop_reasons: list[str] | tuple[str, ...] | None = None) -> None:
+        configured = stop_reasons if stop_reasons is not None else ("refusal",)
+        self._stop_reasons: frozenset[str] = frozenset(r.lower() for r in configured)
+
+    def detect(self, message: AIMessage) -> SafetyTermination | None:
+        value = _get_metadata_value(message, "stop_reason")
+        if value is None or value.lower() not in self._stop_reasons:
+            return None
+        return SafetyTermination(
+            detector=self.name,
+            reason_field="stop_reason",
+            reason_value=value,
+        )
+
+
+class GeminiSafetyDetector:
+    """Gemini / Vertex AI safety-related finish reasons.
+
+    Gemini uses the same ``finish_reason`` field as OpenAI but with an
+    enumerated upper-case taxonomy. The default set covers every Gemini
+    finish_reason that means "the model stopped because the content/image
+    tripped a safety, blocklist, recitation, or PII filter" — i.e. cases
+    where any tool_calls returned alongside are likely truncated/
+    unreliable. Full enum:
+    https://docs.cloud.google.com/python/docs/reference/aiplatform/latest/google.cloud.aiplatform_v1.types.Candidate.FinishReason
+
+    Intentionally **excluded** from the default set:
+    - ``STOP``                       — normal termination.
+    - ``MAX_TOKENS``                 — output length truncation, not safety
+                                       (same root failure mode as
+                                       content_filter, but issue #3028
+                                       scopes it out; expose separately if
+                                       desired).
+    - ``LANGUAGE`` / ``NO_IMAGE``    — capability mismatches, unrelated to
+                                       safety; tool_calls would be absent
+                                       anyway.
+    - ``MALFORMED_FUNCTION_CALL`` /
+      ``UNEXPECTED_TOOL_CALL``       — tool-call protocol errors. The
+                                       tool_calls are *also* unreliable
+                                       here, but the failure category is
+                                       distinct from safety filtering;
+                                       handle in a dedicated detector to
+                                       keep observability records honest.
+    - ``OTHER`` / ``IMAGE_OTHER`` /
+      ``FINISH_REASON_UNSPECIFIED``  — too broad to enable by default;
+                                       opt in via ``finish_reasons=`` if
+                                       your provider abuses these.
+    """
+
+    name = "gemini_safety"
+
+    _DEFAULT_FINISH_REASONS = (
+        # Text safety
+        "SAFETY",
+        "BLOCKLIST",
+        "PROHIBITED_CONTENT",
+        "SPII",
+        "RECITATION",
+        # Image safety (multimodal generation)
+        "IMAGE_SAFETY",
+        "IMAGE_PROHIBITED_CONTENT",
+        "IMAGE_RECITATION",
+    )
+
+    def __init__(self, finish_reasons: list[str] | tuple[str, ...] | None = None) -> None:
+        configured = finish_reasons if finish_reasons is not None else self._DEFAULT_FINISH_REASONS
+        self._finish_reasons: frozenset[str] = frozenset(r.upper() for r in configured)
+
+    def detect(self, message: AIMessage) -> SafetyTermination | None:
+        value = _get_metadata_value(message, "finish_reason")
+        if value is None or value.upper() not in self._finish_reasons:
+            return None
+
+        extras: dict[str, Any] = {}
+        response_metadata = getattr(message, "response_metadata", None) or {}
+        if isinstance(response_metadata, dict):
+            # Gemini surfaces per-category scoring under safety_ratings.
+            ratings = response_metadata.get("safety_ratings")
+            if ratings:
+                extras["safety_ratings"] = ratings
+
+        return SafetyTermination(
+            detector=self.name,
+            reason_field="finish_reason",
+            reason_value=value,
+            extras=extras,
+        )
+
+
+def default_detectors() -> list[SafetyTerminationDetector]:
+    """Built-in detector set used when no custom detectors are configured."""
+    return [
+        OpenAICompatibleContentFilterDetector(),
+        AnthropicRefusalDetector(),
+        GeminiSafetyDetector(),
+    ]
+
+
+__all__ = [
+    "AnthropicRefusalDetector",
+    "GeminiSafetyDetector",
+    "OpenAICompatibleContentFilterDetector",
+    "SafetyTermination",
+    "SafetyTerminationDetector",
+    "default_detectors",
+]

backend/packages/harness/deerflow/agents/middlewares/skill_activation_middleware.py

@@ -0,0 +1,289 @@
+"""Middleware for explicit slash skill activation."""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import html
+import logging
+import uuid
+from collections.abc import Awaitable, Callable
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, override
+
+from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.types import ModelRequest, ModelResponse
+from langchain_core.messages import AIMessage, HumanMessage
+
+from deerflow.skills.slash import parse_slash_skill_reference, resolve_slash_skill
+from deerflow.skills.storage import get_or_new_skill_storage
+from deerflow.skills.storage.skill_storage import SkillStorage
+from deerflow.skills.types import SKILL_MD_FILE
+from deerflow.utils.messages import get_original_user_content_text
+
+if TYPE_CHECKING:
+    from deerflow.config.app_config import AppConfig
+
+logger = logging.getLogger(__name__)
+
+_SLASH_SKILL_ACTIVATION_KEY = "slash_skill_activation"
+_SLASH_SKILL_ACTIVATION_TARGET_ID_KEY = "slash_skill_activation_target_id"
+_SUMMARY_MESSAGE_NAME = "summary"
+
+
+@dataclass(frozen=True, slots=True)
+class _Activation:
+    skill_name: str
+    category: str
+    container_file_path: str
+    skill_content: str
+    content_hash: str
+    remaining_text: str
+
+
+@dataclass(frozen=True, slots=True)
+class _ActivationResolution:
+    activation: _Activation | None = None
+    failure_message: str | None = None
+
+
+def is_slash_skill_activation_reminder(message: object) -> bool:
+    """Return whether a message is hidden slash-skill activation context."""
+    return isinstance(message, HumanMessage) and bool(message.additional_kwargs.get(_SLASH_SKILL_ACTIVATION_KEY))
+
+
+def _is_user_activation_target(message: object) -> bool:
+    if not isinstance(message, HumanMessage):
+        return False
+    if message.name == _SUMMARY_MESSAGE_NAME:
+        return False
+    if message.additional_kwargs.get("hide_from_ui"):
+        return False
+    return True
+
+
+class SkillActivationMiddleware(AgentMiddleware):
+    """Inject full SKILL.md content when the user explicitly types /skill-name."""
+
+    def __init__(
+        self,
+        *,
+        available_skills: set[str] | None = None,
+        app_config: AppConfig | None = None,
+    ) -> None:
+        super().__init__()
+        self._available_skills = set(available_skills) if available_skills is not None else None
+        self._app_config = app_config
+
+    def _storage(self) -> SkillStorage:
+        if self._app_config is not None:
+            return get_or_new_skill_storage(app_config=self._app_config)
+        return get_or_new_skill_storage()
+
+    @staticmethod
+    def _read_skill_content(skill_file: Path, skills_root: Path) -> str:
+        if skill_file.name != SKILL_MD_FILE:
+            raise ValueError(f"Expected {SKILL_MD_FILE}, got {skill_file.name}")
+        resolved_root = skills_root.resolve()
+        resolved_file = skill_file.resolve()
+        try:
+            resolved_file.relative_to(resolved_root)
+        except ValueError as exc:
+            raise ValueError("Resolved skill file must stay within the configured skills root.") from exc
+        if not resolved_file.is_file():
+            raise FileNotFoundError(resolved_file)
+        return resolved_file.read_text(encoding="utf-8")
+
+    def _resolve_activation(self, text: str) -> _ActivationResolution | None:
+        reference = parse_slash_skill_reference(text)
+        if reference is None:
+            return None
+
+        storage = self._storage()
+        skills = storage.load_skills(enabled_only=False)
+        skill = next((candidate for candidate in skills if candidate.name == reference.name), None)
+        if skill is None:
+            return _ActivationResolution(failure_message=f"Skill `/{reference.name}` is not installed.")
+        if not skill.enabled:
+            return _ActivationResolution(failure_message=f"Skill `/{reference.name}` is installed but disabled. Enable it before using slash activation.")
+        if self._available_skills is not None and reference.name not in self._available_skills:
+            return _ActivationResolution(failure_message=f"Skill `/{reference.name}` is not available for this agent.")
+
+        resolved = resolve_slash_skill(
+            text,
+            skills,
+            available_skills=self._available_skills,
+            container_base_path=storage.get_container_root(),
+        )
+        if resolved is None:
+            return _ActivationResolution(failure_message=f"Skill `/{reference.name}` could not be resolved.")
+
+        try:
+            skill_content = self._read_skill_content(resolved.skill.skill_file, storage.get_skills_root_path())
+        except (OSError, ValueError):
+            logger.exception("Failed to read slash-activated skill %s", resolved.skill.name)
+            return _ActivationResolution(failure_message=f"Skill `/{reference.name}` could not be loaded safely. Please check the skill installation.")
+
+        content_hash = hashlib.sha256(skill_content.encode("utf-8")).hexdigest()
+        return _ActivationResolution(
+            activation=_Activation(
+                skill_name=resolved.skill.name,
+                category=str(resolved.skill.category),
+                container_file_path=resolved.container_file_path,
+                skill_content=skill_content,
+                content_hash=content_hash,
+                remaining_text=resolved.remaining_text,
+            )
+        )
+
+    @staticmethod
+    def _build_activation_reminder(activation: _Activation) -> str:
+        user_request = activation.remaining_text or ("No additional task text was provided after the slash skill command. Ask the user what they want to do with this skill if the next step is unclear.")
+        escaped_user_request = html.escape(user_request, quote=False)
+        escaped_skill_content = html.escape(activation.skill_content, quote=False)
+        escaped_skill_name = html.escape(activation.skill_name, quote=True)
+        escaped_category = html.escape(activation.category, quote=True)
+        escaped_path = html.escape(activation.container_file_path, quote=True)
+        escaped_content_hash = html.escape(activation.content_hash, quote=True)
+        return f"""<slash_skill_activation>
+The user explicitly activated the `{activation.skill_name}` skill for this turn.
+Treat the task text as:
+<user_request>
+{escaped_user_request}
+</user_request>
+
+Follow this skill before choosing a general workflow. Load supporting resources from the same skill directory only when needed.
+
+<skill name="{escaped_skill_name}" category="{escaped_category}" path="{escaped_path}" sha256="{escaped_content_hash}">
+<skill_content encoding="xml-escaped">
+{escaped_skill_content}
+</skill_content>
+</skill>
+</slash_skill_activation>"""
+
+    @staticmethod
+    def _has_existing_activation_for_target(messages: list, target_index: int, target: HumanMessage) -> bool:
+        if target_index <= 0:
+            return False
+
+        if target.id:
+            for previous in messages[:target_index]:
+                if not is_slash_skill_activation_reminder(previous):
+                    continue
+                target_id = previous.additional_kwargs.get(_SLASH_SKILL_ACTIVATION_TARGET_ID_KEY)
+                if target_id == target.id or previous.id == f"{target.id}__slash_activation":
+                    return True
+
+        previous = messages[target_index - 1]
+        return is_slash_skill_activation_reminder(previous)
+
+    def _find_activation_target(self, messages: list) -> tuple[int, HumanMessage, _ActivationResolution] | None:
+        if not messages:
+            return None
+
+        target_index = next((idx for idx in range(len(messages) - 1, -1, -1) if _is_user_activation_target(messages[idx])), None)
+        if target_index is None:
+            return None
+
+        target = messages[target_index]
+        if target is None:
+            return None
+        if self._has_existing_activation_for_target(messages, target_index, target):
+            return None
+
+        content = get_original_user_content_text(target.content, target.additional_kwargs)
+        resolution = self._resolve_activation(content)
+        if resolution is None:
+            return None
+        return target_index, target, resolution
+
+    @staticmethod
+    def _record_activation(request: ModelRequest, activation: _Activation, *, hook: str) -> None:
+        runtime = getattr(request, "runtime", None)
+        context = getattr(runtime, "context", None)
+        journal = context.get("__run_journal") if isinstance(context, dict) else None
+        if journal is None:
+            return
+        try:
+            journal.record_middleware(
+                "skill_activation",
+                name="SkillActivationMiddleware",
+                hook=hook,
+                action="activate",
+                changes={
+                    "skill_name": activation.skill_name,
+                    "category": activation.category,
+                    "path": activation.container_file_path,
+                    "content_hash": activation.content_hash,
+                },
+            )
+        except Exception:
+            logger.debug("Failed to record slash skill activation audit event", exc_info=True)
+
+    def _prepare_model_request(self, request: ModelRequest, *, hook: str) -> ModelRequest | AIMessage | None:
+        target_and_resolution = self._find_activation_target(list(request.messages))
+        if target_and_resolution is None:
+            return None
+
+        target_index, target, resolution = target_and_resolution
+        if resolution.failure_message:
+            return AIMessage(content=resolution.failure_message)
+
+        activation = resolution.activation
+        if activation is None:
+            return None
+
+        logger.info(
+            "SkillActivationMiddleware: activating slash skill %s category=%s path=%s hash=%s",
+            activation.skill_name,
+            activation.category,
+            activation.container_file_path,
+            activation.content_hash,
+        )
+        self._record_activation(request, activation, hook=hook)
+        activation_msg = self._make_activation_message(target, self._build_activation_reminder(activation))
+        messages = list(request.messages)
+        messages.insert(target_index, activation_msg)
+        return request.override(messages=messages)
+
+    @staticmethod
+    def _make_activation_message(target: HumanMessage, activation_content: str) -> HumanMessage:
+        stable_id = target.id or str(uuid.uuid4())
+        additional_kwargs = {
+            "hide_from_ui": True,
+            _SLASH_SKILL_ACTIVATION_KEY: True,
+        }
+        if target.id:
+            additional_kwargs[_SLASH_SKILL_ACTIVATION_TARGET_ID_KEY] = target.id
+        return HumanMessage(
+            content=activation_content,
+            id=f"{stable_id}__slash_activation",
+            additional_kwargs=additional_kwargs,
+        )
+
+    @override
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelResponse | AIMessage:
+        prepared = self._prepare_model_request(request, hook="wrap_model_call")
+        if prepared is None:
+            return handler(request)
+        if isinstance(prepared, AIMessage):
+            return prepared
+        return handler(prepared)
+
+    @override
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> ModelResponse | AIMessage:
+        prepared = await asyncio.to_thread(self._prepare_model_request, request, hook="awrap_model_call")
+        if prepared is None:
+            return await handler(request)
+        if isinstance(prepared, AIMessage):
+            return prepared
+        return await handler(prepared)

backend/packages/harness/deerflow/agents/middlewares/summarization_middleware.py

@@ -9,8 +9,9 @@ from typing import Any, Protocol, override, runtime_checkable
 
 from langchain.agents import AgentState
 from langchain.agents.middleware import SummarizationMiddleware
-from langchain_core.messages import AIMessage, AnyMessage, HumanMessage, RemoveMessage, ToolMessage
+from langchain_core.messages import AIMessage, AnyMessage, HumanMessage, RemoveMessage, ToolMessage, get_buffer_string
 from langgraph.config import get_config
+from langgraph.constants import TAG_NOSTREAM
 from langgraph.graph.message import REMOVE_ALL_MESSAGES
 from langgraph.runtime import Runtime
 
@@ -116,6 +117,74 @@ class DeerFlowSummarizationMiddleware(SummarizationMiddleware):
         self._preserve_recent_skill_count = max(0, preserve_recent_skill_count)
         self._preserve_recent_skill_tokens = max(0, preserve_recent_skill_tokens)
         self._preserve_recent_skill_tokens_per_skill = max(0, preserve_recent_skill_tokens_per_skill)
+        # The summary LLM call runs inside a LangGraph middleware hook, so its token
+        # stream would otherwise be captured by the messages-tuple stream callback and
+        # broadcast to the frontend as a phantom AI message. Tag a dedicated model copy
+        # with TAG_NOSTREAM so the streaming handler skips it.
+        # Keep self.model untagged so the parent's profile / ls_params inspection still works.
+        #
+        # Preserve any tags already bound on the model (e.g. "middleware:summarize" set in
+        # lead_agent/agent.py for RunJournal attribution): RunnableBinding.with_config does a
+        # shallow merge that would otherwise overwrite the existing tags list entirely.
+        existing_tags = list((getattr(self.model, "config", None) or {}).get("tags") or [])
+        merged_tags = [*existing_tags, TAG_NOSTREAM] if TAG_NOSTREAM not in existing_tags else existing_tags
+        self._summary_model = self.model.with_config(tags=merged_tags)
+
+    @override
+    def _create_summary(self, messages_to_summarize: list[AnyMessage]) -> str:
+        return self._summarize_with(messages_to_summarize)
+
+    @override
+    async def _acreate_summary(self, messages_to_summarize: list[AnyMessage]) -> str:
+        return await self._asummarize_with(messages_to_summarize)
+
+    def _summarize_with(self, messages_to_summarize: list[AnyMessage]) -> str:
+        """Mirror the parent ``_create_summary`` but invoke the nostream-tagged model.
+
+        We do not swap ``self.model`` at the instance level: the agent/middleware is
+        cached and reused across concurrent runs, so a temporary swap would leak the
+        ``RunnableBinding`` to other coroutines during ``await`` and break parent logic
+        that inspects the raw model (``profile`` / ``_get_ls_params``).
+        """
+        if not messages_to_summarize:
+            return "No previous conversation history."
+        prompt = self._build_summary_prompt(messages_to_summarize)
+        if prompt is None:
+            return "Previous conversation was too long to summarize."
+        try:
+            response = self._summary_model.invoke(
+                prompt,
+                config={"metadata": {"lc_source": "summarization"}},
+            )
+            return response.text.strip()
+        except Exception as e:
+            return f"Error generating summary: {e!s}"
+
+    async def _asummarize_with(self, messages_to_summarize: list[AnyMessage]) -> str:
+        """Async counterpart of :meth:`_summarize_with` using the nostream model."""
+        if not messages_to_summarize:
+            return "No previous conversation history."
+        prompt = self._build_summary_prompt(messages_to_summarize)
+        if prompt is None:
+            return "Previous conversation was too long to summarize."
+        try:
+            response = await self._summary_model.ainvoke(
+                prompt,
+                config={"metadata": {"lc_source": "summarization"}},
+            )
+            return response.text.strip()
+        except Exception as e:
+            return f"Error generating summary: {e!s}"
+
+    def _build_summary_prompt(self, messages_to_summarize: list[AnyMessage]) -> str | None:
+        """Build the summary prompt, returning ``None`` when trimming leaves nothing."""
+        trimmed_messages = self._trim_messages_for_summary(messages_to_summarize)
+        if not trimmed_messages:
+            return None
+        # Format messages to avoid token inflation from metadata when str() is called on
+        # message objects.
+        formatted_messages = get_buffer_string(trimmed_messages)
+        return self.summary_prompt.format(messages=formatted_messages).rstrip()
 
     def before_model(self, state: AgentState, runtime: Runtime) -> dict | None:
         return self._maybe_summarize(state, runtime)

backend/packages/harness/deerflow/agents/middlewares/title_middleware.py

@@ -160,7 +160,11 @@ class TitleMiddleware(AgentMiddleware[TitleMiddlewareState]):
         prompt, user_msg = self._build_title_prompt(state)
 
         try:
-            model_kwargs = {"thinking_enabled": False}
+            # attach_tracing=False because ``_get_runnable_config()`` inherits
+            # the graph-level RunnableConfig (set in ``_make_lead_agent``) whose
+            # callbacks already carry tracing handlers; binding them again at
+            # the model level would emit duplicate spans.
+            model_kwargs = {"thinking_enabled": False, "attach_tracing": False}
             if self._app_config is not None:
                 model_kwargs["app_config"] = self._app_config
             if config.model_name:

backend/packages/harness/deerflow/agents/middlewares/todo_middleware.py

@@ -20,11 +20,13 @@ from collections.abc import Awaitable, Callable
 from typing import Any, override
 
 from langchain.agents.middleware import TodoListMiddleware
-from langchain.agents.middleware.todo import PlanningState, Todo
+from langchain.agents.middleware.todo import Todo
 from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse, hook_config
 from langchain_core.messages import AIMessage, HumanMessage
 from langgraph.runtime import Runtime
 
+from deerflow.agents.thread_state import ThreadState
+
 
 def _todos_in_messages(messages: list[Any]) -> bool:
     """Return True if any AIMessage in *messages* contains a write_todos tool call."""
@@ -113,10 +115,12 @@ class TodoMiddleware(TodoListMiddleware):
     and injects a reminder message so the model can continue tracking progress.
     """
 
+    state_schema = ThreadState
+
     @override
     def before_model(
         self,
-        state: PlanningState,
+        state: ThreadState,
         runtime: Runtime,
     ) -> dict[str, Any] | None:
         """Inject a todo-list reminder when write_todos has left the context window."""
@@ -154,7 +158,7 @@ class TodoMiddleware(TodoListMiddleware):
     @override
     async def abefore_model(
         self,
-        state: PlanningState,
+        state: ThreadState,
         runtime: Runtime,
     ) -> dict[str, Any] | None:
         """Async version of before_model."""
@@ -249,12 +253,12 @@ class TodoMiddleware(TodoListMiddleware):
             self._drop_completion_reminder_key_locked(key)
 
     @override
-    def before_agent(self, state: PlanningState, runtime: Runtime) -> dict[str, Any] | None:
+    def before_agent(self, state: ThreadState, runtime: Runtime) -> dict[str, Any] | None:
         self._clear_other_run_completion_reminders(runtime)
         return None
 
     @override
-    async def abefore_agent(self, state: PlanningState, runtime: Runtime) -> dict[str, Any] | None:
+    async def abefore_agent(self, state: ThreadState, runtime: Runtime) -> dict[str, Any] | None:
         self._clear_other_run_completion_reminders(runtime)
         return None
 
@@ -262,7 +266,7 @@ class TodoMiddleware(TodoListMiddleware):
     @override
     def after_model(
         self,
-        state: PlanningState,
+        state: ThreadState,
         runtime: Runtime,
     ) -> dict[str, Any] | None:
         """Prevent premature agent exit when todo items are still incomplete.
@@ -308,7 +312,7 @@ class TodoMiddleware(TodoListMiddleware):
     @hook_config(can_jump_to=["model"])
     async def aafter_model(
         self,
-        state: PlanningState,
+        state: ThreadState,
         runtime: Runtime,
     ) -> dict[str, Any] | None:
         """Async version of after_model."""
@@ -349,11 +353,11 @@ class TodoMiddleware(TodoListMiddleware):
         return await handler(self._augment_request(request))
 
     @override
-    def after_agent(self, state: PlanningState, runtime: Runtime) -> dict[str, Any] | None:
+    def after_agent(self, state: ThreadState, runtime: Runtime) -> dict[str, Any] | None:
         self._clear_current_run_completion_reminders(runtime)
         return None
 
     @override
-    async def aafter_agent(self, state: PlanningState, runtime: Runtime) -> dict[str, Any] | None:
+    async def aafter_agent(self, state: ThreadState, runtime: Runtime) -> dict[str, Any] | None:
         self._clear_current_run_completion_reminders(runtime)
         return None

backend/packages/harness/deerflow/agents/middlewares/tool_error_handling_middleware.py

@@ -2,7 +2,7 @@
 
 import logging
 from collections.abc import Awaitable, Callable
-from typing import override
+from typing import TYPE_CHECKING, override
 
 from langchain.agents import AgentState
 from langchain.agents.middleware import AgentMiddleware
@@ -12,10 +12,48 @@ from langgraph.prebuilt.tool_node import ToolCallRequest
 from langgraph.types import Command
 
 from deerflow.config.app_config import AppConfig
+from deerflow.subagents.status_contract import (
+    extract_subagent_status,
+    make_subagent_additional_kwargs,
+)
+
+if TYPE_CHECKING:
+    from deerflow.tools.builtins.tool_search import DeferredToolSetup
 
 logger = logging.getLogger(__name__)
 
 _MISSING_TOOL_CALL_ID = "missing_tool_call_id"
+_TASK_TOOL_NAME = "task"
+
+
+def _stamp_task_subagent_status(message: ToolMessage, *, tool_name: str, error: str | None = None) -> ToolMessage:
+    """Centralised stamping of ``additional_kwargs.subagent_status``.
+
+    Bytedance/deer-flow issue #3146: the frontend now reads the subagent
+    status from a structured field instead of parsing the leading text of
+    the task tool's return string. That contract is enforced here, in the
+    one place every task tool result flows through, rather than at the 5
+    normal-return + 3 ``Error:`` pre-execution branches inside
+    ``task_tool.py``. Centralisation prevents the "added a new return
+    path, forgot the stamp" drift mode.
+
+    For non-``task`` tools this is a no-op so other tools' additional_kwargs
+    conventions are untouched.
+    """
+    if tool_name != _TASK_TOOL_NAME:
+        return message
+    content = message.content if isinstance(message.content, str) else ""
+    status = extract_subagent_status(content)
+    if status is None:
+        # Non-terminal streaming chunks or unrecognised shapes leave the
+        # field unset so the frontend can keep the card on its in-progress
+        # placeholder until a real terminal frame arrives.
+        return message
+    stamp = make_subagent_additional_kwargs(status, error=error)
+    existing = dict(message.additional_kwargs or {})
+    existing.update(stamp)
+    message.additional_kwargs = existing
+    return message
 
 
 class ToolErrorHandlingMiddleware(AgentMiddleware[AgentState]):
@@ -29,12 +67,31 @@ class ToolErrorHandlingMiddleware(AgentMiddleware[AgentState]):
             detail = detail[:497] + "..."
 
         content = f"Error: Tool '{tool_name}' failed with {exc.__class__.__name__}: {detail}. Continue with available context, or choose an alternative tool."
-        return ToolMessage(
+        message = ToolMessage(
             content=content,
             tool_call_id=tool_call_id,
             name=tool_name,
             status="error",
         )
+        # Stamp the structured subagent status on the wrapper too: the
+        # frontend would otherwise have to fall back to prefix-matching
+        # ``Error: Tool 'task' failed ...`` on the wire. The ``subagent_error``
+        # carries the same ``ExcClass: detail`` shape the wrapper string
+        # uses so debugging artifacts stay aligned.
+        structured_error = f"{exc.__class__.__name__}: {detail}"
+        return _stamp_task_subagent_status(message, tool_name=tool_name, error=structured_error)
+
+    @staticmethod
+    def _maybe_stamp(result: ToolMessage | Command, request: ToolCallRequest) -> ToolMessage | Command:
+        """Apply the subagent stamp to successful task tool returns.
+
+        ``Command`` results bypass the stamp — they encode LangGraph
+        control flow rather than user-facing tool output.
+        """
+        if not isinstance(result, ToolMessage):
+            return result
+        tool_name = str(request.tool_call.get("name") or "")
+        return _stamp_task_subagent_status(result, tool_name=tool_name)
 
     @override
     def wrap_tool_call(
@@ -43,13 +100,14 @@ class ToolErrorHandlingMiddleware(AgentMiddleware[AgentState]):
         handler: Callable[[ToolCallRequest], ToolMessage | Command],
     ) -> ToolMessage | Command:
         try:
-            return handler(request)
+            result = handler(request)
         except GraphBubbleUp:
             # Preserve LangGraph control-flow signals (interrupt/pause/resume).
             raise
         except Exception as exc:
             logger.exception("Tool execution failed (sync): name=%s id=%s", request.tool_call.get("name"), request.tool_call.get("id"))
             return self._build_error_message(request, exc)
+        return self._maybe_stamp(result, request)
 
     @override
     async def awrap_tool_call(
@@ -58,13 +116,14 @@ class ToolErrorHandlingMiddleware(AgentMiddleware[AgentState]):
         handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command]],
     ) -> ToolMessage | Command:
         try:
-            return await handler(request)
+            result = await handler(request)
         except GraphBubbleUp:
             # Preserve LangGraph control-flow signals (interrupt/pause/resume).
             raise
         except Exception as exc:
             logger.exception("Tool execution failed (async): name=%s id=%s", request.tool_call.get("name"), request.tool_call.get("id"))
             return self._build_error_message(request, exc)
+        return self._maybe_stamp(result, request)
 
 
 def _build_runtime_middlewares(
@@ -77,9 +136,11 @@ def _build_runtime_middlewares(
     """Build shared base middlewares for agent execution."""
     from deerflow.agents.middlewares.llm_error_handling_middleware import LLMErrorHandlingMiddleware
     from deerflow.agents.middlewares.thread_data_middleware import ThreadDataMiddleware
+    from deerflow.agents.middlewares.tool_output_budget_middleware import ToolOutputBudgetMiddleware
     from deerflow.sandbox.middleware import SandboxMiddleware
 
     middlewares: list[AgentMiddleware] = [
+        ToolOutputBudgetMiddleware.from_app_config(app_config),
         ThreadDataMiddleware(lazy_init=lazy_init),
         SandboxMiddleware(lazy_init=lazy_init),
     ]
@@ -87,7 +148,7 @@ def _build_runtime_middlewares(
     if include_uploads:
         from deerflow.agents.middlewares.uploads_middleware import UploadsMiddleware
 
-        middlewares.insert(1, UploadsMiddleware())
+        middlewares.insert(2, UploadsMiddleware())
 
     if include_dangling_tool_call_patch:
         from deerflow.agents.middlewares.dangling_tool_call_middleware import DanglingToolCallMiddleware
@@ -141,6 +202,7 @@ def build_subagent_runtime_middlewares(
     app_config: AppConfig | None = None,
     model_name: str | None = None,
     lazy_init: bool = True,
+    deferred_setup: "DeferredToolSetup | None" = None,
 ) -> list[AgentMiddleware]:
     """Middlewares shared by subagent runtime before subagent-only middlewares."""
     if app_config is None:
@@ -164,4 +226,24 @@ def build_subagent_runtime_middlewares(
 
         middlewares.append(ViewImageMiddleware())
 
+    # Hide deferred (MCP) tool schemas from the subagent's model binding until
+    # tool_search promotes them. This is the same wiring the lead agent gets. The deferred
+    # set + catalog hash come from the build-time setup (assembled after
+    # tool-policy filtering); promotion is read from graph state. Empty/None
+    # setup (deferral disabled or no MCP tool survived) is a pure no-op.
+    if deferred_setup is not None and deferred_setup.deferred_names:
+        from deerflow.agents.middlewares.deferred_tool_filter_middleware import DeferredToolFilterMiddleware
+
+        middlewares.append(DeferredToolFilterMiddleware(deferred_setup.deferred_names, deferred_setup.catalog_hash))
+
+    # Same provider safety-termination guard the lead agent uses — subagents
+    # are equally exposed to truncated tool_calls returned with
+    # finish_reason=content_filter (and friends), and the bad call would then
+    # propagate back to the lead agent via the task tool result.
+    safety_config = app_config.safety_finish_reason
+    if safety_config.enabled:
+        from deerflow.agents.middlewares.safety_finish_reason_middleware import SafetyFinishReasonMiddleware
+
+        middlewares.append(SafetyFinishReasonMiddleware.from_config(safety_config))
+
     return middlewares

backend/packages/harness/deerflow/agents/middlewares/tool_output_budget_middleware.py

@@ -0,0 +1,643 @@
+"""Middleware that enforces a per-result budget on tool outputs.
+
+Oversized tool results are persisted to disk and replaced with a compact
+preview containing a file reference.  When disk persistence is
+unavailable the middleware falls back to head+tail truncation so the
+model context is never blown by a single large tool return.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+import shlex
+import uuid
+from collections.abc import Awaitable, Callable
+from dataclasses import replace as dc_replace
+from typing import TYPE_CHECKING, Any, override
+
+from langchain.agents import AgentState
+from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
+from langchain_core.messages import ToolMessage
+from langgraph.prebuilt.tool_node import ToolCallRequest
+from langgraph.types import Command
+
+from deerflow.config.tool_output_config import ToolOutputConfig
+from deerflow.sandbox.sandbox_provider import get_sandbox_provider
+
+if TYPE_CHECKING:
+    from deerflow.sandbox.sandbox import Sandbox
+
+logger = logging.getLogger(__name__)
+
+# Virtual outputs root inside the sandbox. Host-mounted sandboxes map this to
+# the thread outputs dir on the host; for non-mounted (remote) sandboxes the
+# same path is written directly into the sandbox filesystem so the model's
+# ``read_file`` tool can read it back (issue #3416).
+_VIRTUAL_OUTPUTS_BASE = "/mnt/user-data/outputs"
+
+
+def _default_config() -> ToolOutputConfig:
+    return ToolOutputConfig()
+
+
+# ---------------------------------------------------------------------------
+# Text helpers
+# ---------------------------------------------------------------------------
+
+
+def _message_text(content: Any) -> str | None:
+    """Extract a plain-text representation from a ToolMessage content field.
+
+    Returns ``None`` for non-string / multimodal content so the caller
+    can skip budget enforcement (images, structured blocks, etc.).
+    """
+    if isinstance(content, str):
+        return content
+    if content is None:
+        return None
+    if isinstance(content, list):
+        pieces: list[str] = []
+        for part in content:
+            if isinstance(part, str):
+                pieces.append(part)
+            elif isinstance(part, dict) and isinstance(part.get("text"), str):
+                pieces.append(part["text"])
+            else:
+                return None
+        return "\n".join(pieces) if pieces else None
+    return None
+
+
+def _snap_to_line_boundary(text: str, pos: int) -> int:
+    """Return *pos* or the nearest preceding newline+1, whichever is closer.
+
+    Used so that previews and truncations end on a complete line when
+    possible.  If no newline exists in the second half of ``text[:pos]``
+    the original *pos* is returned unchanged.
+    """
+    if pos <= 0 or pos >= len(text):
+        return pos
+    half = pos // 2
+    nl = text.rfind("\n", half, pos)
+    if nl >= 0:
+        return nl + 1
+    return pos
+
+
+# ---------------------------------------------------------------------------
+# Disk persistence
+# ---------------------------------------------------------------------------
+
+_EXT_MAP: dict[str, str] = {
+    "bash": "log",
+    "bash_tool": "log",
+    "web_fetch": "log",
+}
+
+
+def _sanitize_tool_name(name: str) -> str:
+    """Strip path separators and traversal components from a tool name."""
+    base = os.path.basename(name)
+    safe = base.replace("..", "").replace("/", "_").replace("\\", "_")
+    return safe or "unknown"
+
+
+def _build_externalized_filename(*, tool_name: str, tool_call_id: str) -> str:
+    """Build the on-disk filename for an externalized tool output.
+
+    Shared by the host-disk and sandbox externalization paths so both
+    produce the identical naming scheme.
+    """
+    safe_name = _sanitize_tool_name(tool_name)
+    ext = _EXT_MAP.get(tool_name, "txt")
+    short_id = uuid.uuid4().hex[:12]
+    return f"{safe_name}-{short_id}.{ext}"
+
+
+def _externalize(
+    content: str,
+    *,
+    tool_name: str,
+    tool_call_id: str,
+    outputs_path: str,
+    storage_subdir: str,
+) -> str | None:
+    """Write *content* to disk and return the virtual path, or ``None`` on failure."""
+    if os.path.isabs(storage_subdir) or ".." in storage_subdir:
+        return None
+    storage_dir = os.path.join(outputs_path, storage_subdir)
+    try:
+        os.makedirs(storage_dir, exist_ok=True)
+    except OSError:
+        return None
+
+    filename = _build_externalized_filename(tool_name=tool_name, tool_call_id=tool_call_id)
+    filepath = os.path.join(storage_dir, filename)
+
+    if not os.path.abspath(filepath).startswith(os.path.abspath(storage_dir)):
+        return None
+
+    try:
+        with open(filepath, "w", encoding="utf-8") as f:
+            f.write(content)
+    except OSError:
+        return None
+
+    return f"{_VIRTUAL_OUTPUTS_BASE}/{storage_subdir}/{filename}"
+
+
+def _externalize_to_sandbox(
+    content: str,
+    *,
+    tool_name: str,
+    tool_call_id: str,
+    storage_subdir: str,
+    sandbox: Sandbox,
+) -> str | None:
+    """Write *content* into the sandbox filesystem and return the virtual path.
+
+    Used when the sandbox does not use thread-data mounts (e.g. a remote AIO
+    sandbox): the host-side :func:`_externalize` virtual path would not exist
+    inside the sandbox, so the model's ``read_file`` tool could not read it
+    back (issue #3416). Returns the same virtual-path contract on success, or
+    ``None`` to signal the caller to fall back to inline truncation.
+    """
+    if os.path.isabs(storage_subdir) or ".." in storage_subdir:
+        return None
+    filename = _build_externalized_filename(tool_name=tool_name, tool_call_id=tool_call_id)
+    virtual_dir = f"{_VIRTUAL_OUTPUTS_BASE}/{storage_subdir}"
+    virtual_path = f"{virtual_dir}/{filename}"
+    try:
+        # AIO sandbox write_file does NOT create parent directories, so create
+        # them explicitly before writing. execute_command returns its stdout
+        # verbatim (including an "Error: ..." string on failure) rather than
+        # raising, so we cannot rely on exception propagation here.
+        sandbox.execute_command(f"mkdir -p {shlex.quote(virtual_dir)}")
+        sandbox.write_file(virtual_path, content)
+        # Validate the file landed: execute_command may have silently failed
+        # to create the directory, and write_file backends differ. Refuse to
+        # hand the model an unreadable read_file path.
+        check = sandbox.execute_command(f"test -s {shlex.quote(virtual_path)} && echo OK || echo MISSING")
+        if not isinstance(check, str) or check.strip() != "OK":
+            logger.warning(
+                "Sandbox externalize validation failed: path=%s, check=%r",
+                virtual_path,
+                check,
+            )
+            return None
+    except Exception:
+        logger.exception(
+            "Failed to externalize %s output to sandbox (call_id=%s)",
+            tool_name,
+            tool_call_id,
+        )
+        return None
+    return virtual_path
+
+
+# ---------------------------------------------------------------------------
+# Preview / fallback builders
+# ---------------------------------------------------------------------------
+
+
+def _build_preview(
+    content: str,
+    *,
+    tool_name: str,
+    virtual_path: str,
+    head_chars: int,
+    tail_chars: int,
+) -> str:
+    """Build a preview with a file reference for externalized output."""
+    total = len(content)
+    head_end = _snap_to_line_boundary(content, min(head_chars, total))
+    tail_start = max(head_end, total - tail_chars)
+    tail_start_snapped = _snap_to_line_boundary(content, tail_start)
+    if tail_start_snapped > head_end:
+        tail_start = tail_start_snapped
+
+    head = content[:head_end]
+    tail = content[tail_start:] if tail_start < total else ""
+
+    omitted = total - len(head) - len(tail)
+    ref = f"\n\n[Full {tool_name} output saved to {virtual_path} ({total} chars, ~{total // 4} tokens). Use read_file with start_line and end_line to access specific sections. {omitted} chars omitted from this preview.]\n\n"
+
+    parts = [head, ref]
+    if tail:
+        parts.append(tail)
+    return "".join(parts)
+
+
+def _build_fallback(
+    content: str,
+    *,
+    tool_name: str,
+    max_chars: int,
+    head_chars: int,
+    tail_chars: int,
+) -> str:
+    """Build a head+tail truncation when disk persistence is unavailable.
+
+    The returned string is guaranteed to be no longer than *max_chars*.
+    """
+    total = len(content)
+    if max_chars <= 0 or total <= max_chars:
+        return content
+
+    marker_template = "\n\n[... {n} chars omitted from {tn} output. Persistent storage unavailable. Consider narrowing the query or using more specific parameters.]\n\n"
+    marker_overhead = len(marker_template.format(n=total, tn=tool_name))
+
+    if marker_overhead >= max_chars:
+        return content[:max_chars]
+
+    budget = max_chars - marker_overhead
+    effective_head = min(head_chars, budget)
+    effective_tail = min(tail_chars, max(0, budget - effective_head))
+
+    head_end = _snap_to_line_boundary(content, min(effective_head, total))
+    tail_start = max(head_end, total - effective_tail)
+    tail_start_snapped = _snap_to_line_boundary(content, tail_start)
+    if tail_start_snapped > head_end:
+        tail_start = tail_start_snapped
+
+    head = content[:head_end]
+    tail = content[tail_start:] if tail_start < total else ""
+    omitted = total - len(head) - len(tail)
+
+    marker = marker_template.format(n=omitted, tn=tool_name)
+
+    parts = [head, marker]
+    if tail:
+        parts.append(tail)
+    return "".join(parts)
+
+
+# ---------------------------------------------------------------------------
+# Core budget logic
+# ---------------------------------------------------------------------------
+
+
+def _resolve_outputs_path(request: ToolCallRequest) -> str | None:
+    """Best-effort extraction of the thread outputs path."""
+    runtime = getattr(request, "runtime", None)
+    if runtime is None:
+        return None
+    state = getattr(runtime, "state", None)
+    if state is None:
+        return None
+    thread_data = state.get("thread_data")
+    if not isinstance(thread_data, dict):
+        return None
+    outputs_path = thread_data.get("outputs_path")
+    return outputs_path if isinstance(outputs_path, str) else None
+
+
+def _resolve_sandbox(request: ToolCallRequest) -> Sandbox | None:
+    """Resolve the active sandbox for the current tool call, or ``None``.
+
+    Reads the sandbox_id that ``SandboxMiddleware`` (and the sandbox tools
+    themselves) write into ``runtime.state["sandbox"]``. We intentionally do
+    NOT call ``provider.acquire`` here: acquiring a sandbox can trigger
+    blocking remote I/O, and this resolver runs on every tool call. Tools
+    that do not use a sandbox (``web_search``, MCP, ...) will return ``None``
+    here, which is fine -- the caller falls back to inline truncation.
+    """
+    runtime = getattr(request, "runtime", None)
+    state = getattr(runtime, "state", None)
+    if not isinstance(state, dict):
+        return None
+    sandbox_state = state.get("sandbox")
+    if not isinstance(sandbox_state, dict):
+        return None
+    sandbox_id = sandbox_state.get("sandbox_id")
+    if not sandbox_id:
+        return None
+    try:
+        return get_sandbox_provider().get(sandbox_id)
+    except Exception:
+        logger.exception("Failed to look up sandbox %s for tool-output externalization", sandbox_id)
+        return None
+
+
+def _budget_content(
+    content: str,
+    *,
+    tool_name: str,
+    tool_call_id: str,
+    outputs_path: str | None,
+    config: ToolOutputConfig,
+    sandbox: Sandbox | None = None,
+) -> str | None:
+    """Apply budget to *content*. Returns ``None`` if no change needed."""
+    threshold = config.tool_overrides.get(tool_name, config.externalize_min_chars)
+    if threshold <= 0 and config.fallback_max_chars <= 0:
+        return None
+    if len(content) <= threshold and len(content) <= config.fallback_max_chars:
+        return None
+
+    if threshold > 0 and len(content) > threshold:
+        virtual_path: str | None = None
+        # Decide persistence target based on what's available, without touching
+        # the sandbox provider unless a sandbox was actually resolved for this
+        # call. This keeps the legacy host-disk path provider-free, so callers
+        # without a configured sandbox (and CI environments without a
+        # config.yaml) continue to externalize to the host as before.
+        if sandbox is not None:
+            provider = None
+            try:
+                provider = get_sandbox_provider()
+            except Exception:
+                logger.exception("Failed to get sandbox provider for tool-output externalization; falling back to inline truncation")
+            if provider is not None and getattr(provider, "uses_thread_data_mounts", False):
+                # Host-mounted sandbox: host outputs path is bind-mounted into
+                # the sandbox at the same virtual path, so writing host-side is
+                # equivalent. Preserve the original behavior to avoid extra
+                # sandbox round-trips.
+                if outputs_path:
+                    virtual_path = _externalize(
+                        content,
+                        tool_name=tool_name,
+                        tool_call_id=tool_call_id,
+                        outputs_path=outputs_path,
+                        storage_subdir=config.storage_subdir,
+                    )
+            else:
+                virtual_path = _externalize_to_sandbox(
+                    content,
+                    tool_name=tool_name,
+                    tool_call_id=tool_call_id,
+                    storage_subdir=config.storage_subdir,
+                    sandbox=sandbox,
+                )
+        elif outputs_path:
+            # No sandbox in this call (legacy / non-sandbox tools): write to
+            # host outputs path directly, no provider needed.
+            virtual_path = _externalize(
+                content,
+                tool_name=tool_name,
+                tool_call_id=tool_call_id,
+                outputs_path=outputs_path,
+                storage_subdir=config.storage_subdir,
+            )
+        if virtual_path is not None:
+            logger.info(
+                "Externalized %s output (%d chars) to %s",
+                tool_name,
+                len(content),
+                virtual_path,
+            )
+            return _build_preview(
+                content,
+                tool_name=tool_name,
+                virtual_path=virtual_path,
+                head_chars=config.preview_head_chars,
+                tail_chars=config.preview_tail_chars,
+            )
+
+    if config.fallback_max_chars > 0 and len(content) > config.fallback_max_chars:
+        logger.warning(
+            "Fallback-truncating %s output: %d chars → %d max",
+            tool_name,
+            len(content),
+            config.fallback_max_chars,
+        )
+        return _build_fallback(
+            content,
+            tool_name=tool_name,
+            max_chars=config.fallback_max_chars,
+            head_chars=config.fallback_head_chars,
+            tail_chars=config.fallback_tail_chars,
+        )
+
+    return None
+
+
+# ---------------------------------------------------------------------------
+# Result patchers
+# ---------------------------------------------------------------------------
+
+
+def _patch_tool_message(
+    msg: ToolMessage,
+    config: ToolOutputConfig,
+    outputs_path: str | None,
+    sandbox: Sandbox | None = None,
+) -> ToolMessage:
+    """Apply budget to a single ToolMessage. Returns the original if unchanged."""
+    tool_name = msg.name or "unknown"
+    if tool_name in config.exempt_tools:
+        return msg
+
+    text = _message_text(msg.content)
+    if text is None:
+        return msg
+
+    replacement = _budget_content(
+        text,
+        tool_name=tool_name,
+        tool_call_id=msg.tool_call_id or "",
+        outputs_path=outputs_path,
+        config=config,
+        sandbox=sandbox,
+    )
+    if replacement is None:
+        return msg
+
+    update: dict[str, Any] = {"content": replacement}
+    if getattr(msg, "response_metadata", None):
+        update["response_metadata"] = dict(msg.response_metadata)
+    if getattr(msg, "additional_kwargs", None):
+        update["additional_kwargs"] = dict(msg.additional_kwargs)
+    return msg.model_copy(update=update)
+
+
+def _effective_trigger(tool_name: str, config: ToolOutputConfig) -> int:
+    """Smallest content length that could trigger budgeting for *tool_name*.
+
+    Mirrors the trigger conditions in :func:`_budget_content` (per-tool
+    externalize threshold OR global fallback), so the pre-scan never produces
+    a false negative. Returns ``-1`` when nothing could ever trigger.
+    """
+    candidates: list[int] = []
+    externalize = config.tool_overrides.get(tool_name, config.externalize_min_chars)
+    if externalize > 0:
+        candidates.append(externalize)
+    if config.fallback_max_chars > 0:
+        candidates.append(config.fallback_max_chars)
+    return min(candidates) if candidates else -1
+
+
+def _tool_message_over_budget(msg: ToolMessage, config: ToolOutputConfig) -> bool:
+    """Cheap, per-tool-aware check: is this ToolMessage non-exempt and over its trigger?"""
+    if (msg.name or "") in config.exempt_tools:
+        return False
+    trigger = _effective_trigger(msg.name or "", config)
+    if trigger < 0:
+        return False
+    text = _message_text(msg.content)
+    return text is not None and len(text) > trigger
+
+
+def _needs_budget(result: ToolMessage | Command, config: ToolOutputConfig) -> bool:
+    """Fast check whether *result* could need budgeting (avoids thread offload for small outputs)."""
+    if isinstance(result, ToolMessage):
+        return _tool_message_over_budget(result, config)
+    update = getattr(result, "update", None)
+    if isinstance(update, dict):
+        for msg in update.get("messages", []):
+            if isinstance(msg, ToolMessage) and _tool_message_over_budget(msg, config):
+                return True
+    return False
+
+
+def _patch_result(
+    result: ToolMessage | Command,
+    config: ToolOutputConfig,
+    outputs_path: str | None,
+    sandbox: Sandbox | None = None,
+) -> ToolMessage | Command:
+    """Apply budget to a tool call result (ToolMessage or Command)."""
+    if isinstance(result, ToolMessage):
+        return _patch_tool_message(result, config, outputs_path, sandbox)
+
+    update = getattr(result, "update", None)
+    if not isinstance(update, dict):
+        return result
+
+    messages = update.get("messages")
+    if not isinstance(messages, list):
+        return result
+
+    new_messages: list[Any] = []
+    changed = False
+    for msg in messages:
+        if isinstance(msg, ToolMessage):
+            patched = _patch_tool_message(msg, config, outputs_path, sandbox)
+            if patched is not msg:
+                changed = True
+            new_messages.append(patched)
+        else:
+            new_messages.append(msg)
+
+    if not changed:
+        return result
+
+    return dc_replace(result, update={**update, "messages": new_messages})
+
+
+def _patch_model_messages(messages: list[Any], config: ToolOutputConfig) -> list[Any] | None:
+    """Apply budget to historical ToolMessages in a model request. Returns ``None`` if unchanged.
+
+    A cheap pre-scan bails out before allocating a new list when no historical
+    ToolMessage exceeds the budget — the common case once every result has
+    already been budgeted at tool-call time, so a long history is not rebuilt
+    on every model call.
+
+    Historical messages do not get a ``sandbox`` argument: any oversized tool
+    message in history was already budgeted (and possibly externalized) at
+    tool-call time, so the only thing left for the history path to do is
+    inline fallback truncation, which needs no sandbox.
+    """
+    if not any(isinstance(msg, ToolMessage) and _tool_message_over_budget(msg, config) for msg in messages):
+        return None
+
+    updated: list[Any] = []
+    changed = False
+    for msg in messages:
+        if isinstance(msg, ToolMessage):
+            patched = _patch_tool_message(msg, config, outputs_path=None)
+            if patched is not msg:
+                changed = True
+            updated.append(patched)
+        else:
+            updated.append(msg)
+    return updated if changed else None
+
+
+# ---------------------------------------------------------------------------
+# Middleware class
+# ---------------------------------------------------------------------------
+
+
+class ToolOutputBudgetMiddleware(AgentMiddleware[AgentState]):
+    """Enforce per-result budget on tool outputs via externalization or truncation."""
+
+    def __init__(self, config: ToolOutputConfig | None = None) -> None:
+        super().__init__()
+        self._config = config if config is not None else _default_config()
+
+    @classmethod
+    def from_app_config(cls, app_config: Any) -> ToolOutputBudgetMiddleware:
+        tool_output = getattr(app_config, "tool_output", None)
+        if isinstance(tool_output, ToolOutputConfig):
+            return cls(config=tool_output)
+        return cls()
+
+    # -- tool call hooks ---------------------------------------------------
+
+    @override
+    def wrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], ToolMessage | Command],
+    ) -> ToolMessage | Command:
+        result = handler(request)
+        if not self._config.enabled:
+            return result
+        if not _needs_budget(result, self._config):
+            return result
+        outputs_path = _resolve_outputs_path(request)
+        sandbox = _resolve_sandbox(request)
+        return _patch_result(result, self._config, outputs_path, sandbox)
+
+    @override
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command]],
+    ) -> ToolMessage | Command:
+        result = await handler(request)
+        if not self._config.enabled:
+            return result
+        if not _needs_budget(result, self._config):
+            return result
+        outputs_path = _resolve_outputs_path(request)
+        # _resolve_sandbox only touches runtime.state and the provider's
+        # in-memory sandbox registry, so it is safe to call on the event
+        # loop. The actual sandbox I/O (mkdir/write/test) happens inside
+        # _patch_result, which is offloaded to a worker thread below.
+        sandbox = _resolve_sandbox(request)
+        return await asyncio.to_thread(_patch_result, result, self._config, outputs_path, sandbox)
+
+    # -- model call hooks (historical message truncation) ------------------
+
+    @override
+    def wrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], ModelResponse],
+    ) -> ModelCallResult:
+        if self._config.enabled:
+            messages = getattr(request, "messages", None)
+            if isinstance(messages, list):
+                patched = _patch_model_messages(messages, self._config)
+                if patched is not None:
+                    request = request.override(messages=patched)
+        return handler(request)
+
+    @override
+    async def awrap_model_call(
+        self,
+        request: ModelRequest,
+        handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+    ) -> ModelCallResult:
+        if self._config.enabled:
+            messages = getattr(request, "messages", None)
+            if isinstance(messages, list):
+                patched = _patch_model_messages(messages, self._config)
+                if patched is not None:
+                    request = request.override(messages=patched)
+        return await handler(request)

backend/packages/harness/deerflow/agents/middlewares/uploads_middleware.py

@@ -7,11 +7,13 @@ from typing import NotRequired, override
 from langchain.agents import AgentState
 from langchain.agents.middleware import AgentMiddleware
 from langchain_core.messages import HumanMessage
+from langchain_core.runnables import run_in_executor
 from langgraph.runtime import Runtime
 
 from deerflow.config.paths import Paths, get_paths
 from deerflow.runtime.user_context import get_effective_user_id
 from deerflow.utils.file_conversion import extract_outline
+from deerflow.utils.messages import ORIGINAL_USER_CONTENT_KEY, message_content_to_text
 
 logger = logging.getLogger(__name__)
 
@@ -264,6 +266,8 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
 
         # Extract original content - handle both string and list formats
         original_content = last_message.content
+        additional_kwargs = dict(last_message.additional_kwargs or {})
+        additional_kwargs.setdefault(ORIGINAL_USER_CONTENT_KEY, message_content_to_text(original_content))
         if isinstance(original_content, str):
             # Simple case: string content, just prepend files message
             updated_content = f"{files_message}\n\n{original_content}"
@@ -284,7 +288,7 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
             content=updated_content,
             id=last_message.id,
             name=last_message.name,
-            additional_kwargs=last_message.additional_kwargs,
+            additional_kwargs=additional_kwargs,
         )
 
         messages[last_message_index] = updated_message
@@ -293,3 +297,16 @@ class UploadsMiddleware(AgentMiddleware[UploadsMiddlewareState]):
             "uploaded_files": new_files,
             "messages": messages,
         }
+
+    @override
+    async def abefore_agent(self, state: UploadsMiddlewareState, runtime: Runtime) -> dict | None:
+        """Async hook that offloads the synchronous uploads scan off the event loop.
+
+        ``before_agent`` performs blocking filesystem IO (directory enumeration,
+        ``stat``, reading sibling ``.md`` outlines). When the graph runs async,
+        langgraph would otherwise execute the sync hook directly on the event
+        loop, so it is dispatched to a worker thread via ``run_in_executor``.
+        ``run_in_executor`` copies the current context, so the ``user_id``
+        contextvar read by ``get_effective_user_id()`` is preserved.
+        """
+        return await run_in_executor(None, self.before_agent, state, runtime)

backend/packages/harness/deerflow/agents/middlewares/view_image_middleware.py

@@ -179,8 +179,10 @@ class ViewImageMiddleware(AgentMiddleware[ViewImageMiddlewareState]):
         # Create the image details message with text and image content
         image_content = self._create_image_details_message(state)
 
-        # Create a new human message with mixed content (text + images)
-        human_msg = HumanMessage(content=image_content)
+        # Create a new human message with mixed content (text + images). This is
+        # internal context for the model only, so hide it from the chat UI and IM
+        # channels (matches the other middleware-injected context messages).
+        human_msg = HumanMessage(content=image_content, additional_kwargs={"hide_from_ui": True})
 
         logger.debug("Injecting image details message with images before LLM call")
 

backend/packages/harness/deerflow/agents/thread_state.py

@@ -45,11 +45,51 @@ def merge_viewed_images(existing: dict[str, ViewedImageData] | None, new: dict[s
     return {**existing, **new}
 
 
+def merge_todos(existing: list | None, new: list | None) -> list | None:
+    """Reducer for todos list - keeps the last non-None value.
+
+    Semantics:
+    - If `new` is None (node didn't touch todos), preserve `existing`.
+    - If `new` is provided (even empty list), it represents an explicit
+      update and wins over `existing`.
+    """
+    if new is None:
+        return existing
+    return new
+
+
+class PromotedTools(TypedDict):
+    catalog_hash: str
+    names: list[str]
+
+
+def merge_promoted(existing: PromotedTools | None, new: PromotedTools | None) -> PromotedTools | None:
+    """Reducer for deferred-tool promotions, scoped by catalog hash.
+
+    - new None/empty -> preserve existing (node didn't touch promotions).
+    - catalog_hash changed -> replace wholesale, dropping stale names (prevents a
+      persisted bare name from exposing a different tool after catalog drift).
+    - same catalog_hash -> union names, dedupe, preserve order.
+    """
+    if not new:
+        return existing
+    if existing is None or existing.get("catalog_hash") != new["catalog_hash"]:
+        return {
+            "catalog_hash": new["catalog_hash"],
+            "names": list(dict.fromkeys(new["names"])),
+        }
+    return {
+        "catalog_hash": existing["catalog_hash"],
+        "names": list(dict.fromkeys(existing["names"] + new["names"])),
+    }
+
+
 class ThreadState(AgentState):
     sandbox: NotRequired[SandboxState | None]
     thread_data: NotRequired[ThreadDataState | None]
     title: NotRequired[str | None]
     artifacts: Annotated[list[str], merge_artifacts]
-    todos: NotRequired[list | None]
+    todos: Annotated[list | None, merge_todos]
     uploaded_files: NotRequired[list[dict] | None]
     viewed_images: Annotated[dict[str, ViewedImageData], merge_viewed_images]  # image_path -> {base64, mime_type}
+    promoted: Annotated[PromotedTools | None, merge_promoted]

backend/packages/harness/deerflow/client.py

@@ -19,6 +19,7 @@ import asyncio
 import json
 import logging
 import mimetypes
+import os
 import shutil
 import tempfile
 import uuid
@@ -32,7 +33,7 @@ from langchain.agents.middleware import AgentMiddleware
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 from langchain_core.runnables import RunnableConfig
 
-from deerflow.agents.lead_agent.agent import _build_middlewares
+from deerflow.agents.lead_agent.agent import build_middlewares
 from deerflow.agents.lead_agent.prompt import apply_prompt_template
 from deerflow.agents.thread_state import ThreadState
 from deerflow.config.agents_config import AGENT_NAME_PATTERN
@@ -42,6 +43,8 @@ from deerflow.config.paths import get_paths
 from deerflow.models import create_chat_model
 from deerflow.runtime.user_context import get_effective_user_id
 from deerflow.skills.storage import get_or_new_skill_storage
+from deerflow.tools.builtins.tool_search import assemble_deferred_tools
+from deerflow.tracing import build_tracing_callbacks, inject_langfuse_metadata
 from deerflow.uploads.manager import (
     claim_unique_filename,
     delete_file_safe,
@@ -123,6 +126,7 @@ class DeerFlowClient:
         agent_name: str | None = None,
         available_skills: set[str] | None = None,
         middlewares: Sequence[AgentMiddleware] | None = None,
+        environment: str | None = None,
     ):
         """Initialize the client.
 
@@ -140,6 +144,12 @@ class DeerFlowClient:
             agent_name: Name of the agent to use.
             available_skills: Optional set of skill names to make available. If None (default), all scanned skills are available.
             middlewares: Optional list of custom middlewares to inject into the agent.
+            environment: Deployment environment label that ends up in
+                ``langfuse_tags`` (e.g. ``"production"`` / ``"staging"``).
+                When ``None`` the worker/client falls back to the
+                ``DEER_FLOW_ENV`` or ``ENVIRONMENT`` env vars. Pass an
+                explicit value for programmatic callers that do not want
+                env-var coupling.
         """
         if config_path is not None:
             reload_app_config(config_path)
@@ -156,6 +166,7 @@ class DeerFlowClient:
         self._agent_name = agent_name
         self._available_skills = set(available_skills) if available_skills is not None else None
         self._middlewares = list(middlewares) if middlewares else []
+        self._environment = environment
 
         # Lazy agent — created on first call, recreated when config changes.
         self._agent = None
@@ -227,15 +238,30 @@ class DeerFlowClient:
         subagent_enabled = cfg.get("subagent_enabled", False)
         max_concurrent_subagents = cfg.get("max_concurrent_subagents", 3)
 
+        tools = self._get_tools(model_name=model_name, subagent_enabled=subagent_enabled)
+        final_tools, deferred_setup = assemble_deferred_tools(tools, enabled=self._app_config.tool_search.enabled)
         kwargs: dict[str, Any] = {
-            "model": create_chat_model(name=model_name, thinking_enabled=thinking_enabled),
-            "tools": self._get_tools(model_name=model_name, subagent_enabled=subagent_enabled),
-            "middleware": _build_middlewares(config, model_name=model_name, agent_name=self._agent_name, custom_middlewares=self._middlewares),
+            # attach_tracing=False because ``stream()`` injects tracing
+            # callbacks at the graph invocation root so a single embedded run
+            # produces one trace with correct session_id / user_id propagation.
+            # Attaching them again on the model would emit duplicate spans.
+            "model": create_chat_model(name=model_name, thinking_enabled=thinking_enabled, attach_tracing=False),
+            "tools": final_tools,
+            "middleware": build_middlewares(
+                config,
+                model_name=model_name,
+                agent_name=self._agent_name,
+                available_skills=self._available_skills,
+                custom_middlewares=self._middlewares,
+                app_config=self._app_config,
+                deferred_setup=deferred_setup,
+            ),
             "system_prompt": apply_prompt_template(
                 subagent_enabled=subagent_enabled,
                 max_concurrent_subagents=max_concurrent_subagents,
                 agent_name=self._agent_name,
                 available_skills=self._available_skills,
+                deferred_names=deferred_setup.deferred_names,
             ),
             "state_schema": ThreadState,
         }
@@ -571,6 +597,28 @@ class DeerFlowClient:
             thread_id = str(uuid.uuid4())
 
         config = self._get_runnable_config(thread_id, **kwargs)
+
+        # Inject tracing callbacks and Langfuse trace metadata at the graph
+        # invocation root so the embedded client matches the gateway worker's
+        # behaviour: a single ``stream()`` produces one trace with all node /
+        # LLM / tool calls nested under it, and the trace carries the reserved
+        # ``langfuse_session_id`` / ``langfuse_user_id`` keys that the Langfuse
+        # CallbackHandler lifts onto the root trace's ``sessionId`` / ``userId``.
+        tracing_callbacks = build_tracing_callbacks()
+        if tracing_callbacks:
+            existing_callbacks = list(config.get("callbacks") or [])
+            config["callbacks"] = [*existing_callbacks, *tracing_callbacks]
+
+        configurable = config.get("configurable") or {}
+        inject_langfuse_metadata(
+            config,
+            thread_id=thread_id,
+            user_id=get_effective_user_id(),
+            assistant_id=self._agent_name or "lead-agent",
+            model_name=configurable.get("model_name") or self._model_name,
+            environment=self._environment or os.environ.get("DEER_FLOW_ENV") or os.environ.get("ENVIRONMENT"),
+        )
+
         self._ensure_agent(config)
 
         state: dict[str, Any] = {"messages": [HumanMessage(content=message)]}
@@ -1170,7 +1218,7 @@ class DeerFlowClient:
 
                 info: dict[str, Any] = {
                     "filename": dest_name,
-                    "size": str(dest.stat().st_size),
+                    "size": dest.stat().st_size,
                     "path": str(dest),
                     "virtual_path": upload_virtual_path(dest_name),
                     "artifact_url": upload_artifact_url(thread_id, dest_name),

backend/packages/harness/deerflow/community/aio_sandbox/aio_sandbox.py

@@ -1,4 +1,5 @@
 import base64
+import errno
 import logging
 import shlex
 import threading
@@ -6,11 +7,14 @@ import uuid
 
 from agent_sandbox import Sandbox as AioSandboxClient
 
+from deerflow.config.paths import VIRTUAL_PATH_PREFIX
 from deerflow.sandbox.sandbox import Sandbox
 from deerflow.sandbox.search import GrepMatch, path_matches, should_ignore_path, truncate_line
 
 logger = logging.getLogger(__name__)
 
+_MAX_DOWNLOAD_SIZE = 100 * 1024 * 1024  # 100 MB
+
 _ERROR_OBSERVATION_SIGNATURE = "'ErrorObservation' object has no attribute 'exit_code'"
 
 
@@ -35,11 +39,63 @@ class AioSandbox(Sandbox):
         self._client = AioSandboxClient(base_url=base_url, timeout=600)
         self._home_dir = home_dir
         self._lock = threading.Lock()
+        self._closed = False
 
     @property
     def base_url(self) -> str:
         return self._base_url
 
+    def close(self) -> None:
+        """Best-effort close of the host-side HTTP client owned by this sandbox.
+
+        The agent_sandbox SDK is Fern-generated and exposes no ``close()`` /
+        ``__exit__``, so we reach the socket-owning ``httpx.Client`` explicitly
+        through its attribute chain::
+
+            Sandbox._client_wrapper        -> SyncClientWrapper
+                .httpx_client              -> Fern HttpClient (a wrapper, NOT httpx.Client)
+                    .httpx_client          -> httpx.Client     <- the real socket owner
+
+        Closing it releases pooled sockets so long-running provider lifecycles
+        do not accumulate unreclaimed host-side resources (#2872).
+
+        Resolution is most-specific-first with graceful degradation: if a future
+        SDK adds a top-level ``Sandbox.close()`` it is picked up automatically
+        without changing this code. Idempotent, thread-safe, and non-fatal:
+        failures during teardown are logged and swallowed so provider/backend
+        cleanup is never blocked.
+        """
+        with self._lock:
+            if self._closed:
+                return
+            self._closed = True
+            client = self._client
+            # Drop the reference under the lock for use-after-close safety: any
+            # later command on this instance fails loudly instead of reusing a
+            # half-closed client.
+            self._client = None
+
+        if client is None:
+            return
+
+        # Walk from the real httpx.Client up to the top-level client, picking the
+        # first object that actually exposes close().
+        wrapper = getattr(client, "_client_wrapper", None)
+        fern_http = getattr(wrapper, "httpx_client", None)
+        real_httpx = getattr(fern_http, "httpx_client", None)
+        target = next(
+            (c for c in (real_httpx, fern_http, client) if c is not None and hasattr(c, "close")),
+            None,
+        )
+        if target is None:
+            logger.debug("AioSandbox %s: no closable client found, nothing to release", self.id)
+            return
+
+        try:
+            target.close()
+        except Exception as e:
+            logger.warning(f"Error closing AioSandbox client for {self.id}: {e}")
+
     @property
     def home_dir(self) -> str:
         """Get the home directory inside the sandbox."""
@@ -102,6 +158,49 @@ class AioSandbox(Sandbox):
             logger.error(f"Failed to read file in sandbox: {e}")
             return f"Error: {e}"
 
+    def download_file(self, path: str) -> bytes:
+        """Download file bytes from the sandbox.
+
+        Raises:
+            PermissionError: If the path contains '..' traversal segments or is
+                outside ``VIRTUAL_PATH_PREFIX``.
+            OSError: If the file cannot be retrieved from the sandbox.
+        """
+        # Reject path traversal before sending to the container API.
+        # LocalSandbox gets this implicitly via _resolve_path;
+        # here the path is forwarded verbatim so we must check explicitly.
+        normalised = path.replace("\\", "/")
+        for segment in normalised.split("/"):
+            if segment == "..":
+                logger.error(f"Refused download due to path traversal: {path}")
+                raise PermissionError(f"Access denied: path traversal detected in '{path}'")
+
+        stripped_path = normalised.lstrip("/")
+        allowed_prefix = VIRTUAL_PATH_PREFIX.lstrip("/")
+        if stripped_path != allowed_prefix and not stripped_path.startswith(f"{allowed_prefix}/"):
+            logger.error("Refused download outside allowed directory: path=%s, allowed_prefix=%s", path, VIRTUAL_PATH_PREFIX)
+            raise PermissionError(f"Access denied: path must be under '{VIRTUAL_PATH_PREFIX}': '{path}'")
+
+        with self._lock:
+            try:
+                chunks: list[bytes] = []
+                total = 0
+                for chunk in self._client.file.download_file(path=path):
+                    total += len(chunk)
+                    if total > _MAX_DOWNLOAD_SIZE:
+                        raise OSError(
+                            errno.EFBIG,
+                            f"File exceeds maximum download size of {_MAX_DOWNLOAD_SIZE} bytes",
+                            path,
+                        )
+                    chunks.append(chunk)
+                return b"".join(chunks)
+            except OSError:
+                raise
+            except Exception as e:
+                logger.error(f"Failed to download file in sandbox: {e}")
+                raise OSError(f"Failed to download file '{path}' from sandbox: {e}") from e
+
     def list_dir(self, path: str, max_depth: int = 2) -> list[str]:
         """List the contents of a directory in the sandbox.
 

backend/packages/harness/deerflow/community/aio_sandbox/aio_sandbox_provider.py

@@ -10,6 +10,7 @@ The provider itself handles:
 - Mount computation (thread-specific, skills)
 """
 
+import asyncio
 import atexit
 import hashlib
 import logging
@@ -18,6 +19,7 @@ import signal
 import threading
 import time
 import uuid
+from concurrent.futures import ThreadPoolExecutor
 
 try:
     import fcntl
@@ -32,7 +34,7 @@ from deerflow.sandbox.sandbox import Sandbox
 from deerflow.sandbox.sandbox_provider import SandboxProvider
 
 from .aio_sandbox import AioSandbox
-from .backend import SandboxBackend, wait_for_sandbox_ready
+from .backend import SandboxBackend, wait_for_sandbox_ready, wait_for_sandbox_ready_async
 from .local_backend import LocalContainerBackend
 from .remote_backend import RemoteSandboxBackend
 from .sandbox_info import SandboxInfo
@@ -46,6 +48,9 @@ DEFAULT_CONTAINER_PREFIX = "deer-flow-sandbox"
 DEFAULT_IDLE_TIMEOUT = 600  # 10 minutes in seconds
 DEFAULT_REPLICAS = 3  # Maximum concurrent sandbox containers
 IDLE_CHECK_INTERVAL = 60  # Check every 60 seconds
+THREAD_LOCK_EXECUTOR_WORKERS = min(32, (os.cpu_count() or 1) + 4)
+_THREAD_LOCK_EXECUTOR = ThreadPoolExecutor(max_workers=THREAD_LOCK_EXECUTOR_WORKERS, thread_name_prefix="sandbox-lock-wait")
+atexit.register(_THREAD_LOCK_EXECUTOR.shutdown, wait=False, cancel_futures=True)
 
 
 def _lock_file_exclusive(lock_file) -> None:
@@ -66,6 +71,40 @@ def _unlock_file(lock_file) -> None:
     msvcrt.locking(lock_file.fileno(), msvcrt.LK_UNLCK, 1)
 
 
+def _open_lock_file(lock_path):
+    return open(lock_path, "a", encoding="utf-8")
+
+
+async def _acquire_thread_lock_async(lock: threading.Lock) -> None:
+    """Acquire a threading.Lock without polling or using the default executor."""
+    loop = asyncio.get_running_loop()
+    acquire_future = loop.run_in_executor(_THREAD_LOCK_EXECUTOR, lock.acquire, True)
+
+    try:
+        acquired = await asyncio.shield(acquire_future)
+    except asyncio.CancelledError:
+        acquire_future.add_done_callback(lambda task: _release_cancelled_lock_acquire(lock, task))
+        raise
+
+    if not acquired:
+        raise RuntimeError("Failed to acquire sandbox thread lock")
+
+
+def _release_cancelled_lock_acquire(lock: threading.Lock, task: asyncio.Future[bool]) -> None:
+    """Release a lock acquired after its awaiting coroutine was cancelled."""
+    if task.cancelled():
+        return
+
+    try:
+        acquired = task.result()
+    except Exception as e:
+        logger.warning(f"Cancelled sandbox lock acquisition finished with error: {e}")
+        return
+
+    if acquired:
+        lock.release()
+
+
 class AioSandboxProvider(SandboxProvider):
     """Sandbox provider that manages containers running the AIO sandbox.
 
@@ -416,6 +455,96 @@ class AioSandboxProvider(SandboxProvider):
                 self._thread_locks[thread_id] = threading.Lock()
             return self._thread_locks[thread_id]
 
+    def _sandbox_id_for_thread(self, thread_id: str | None) -> str:
+        """Return deterministic IDs for thread sandboxes and random IDs otherwise."""
+        return self._deterministic_sandbox_id(thread_id) if thread_id else str(uuid.uuid4())[:8]
+
+    def _reuse_in_process_sandbox(self, thread_id: str | None, *, post_lock: bool = False) -> str | None:
+        """Reuse an active in-process sandbox for a thread if one is still tracked."""
+        if thread_id is None:
+            return None
+
+        with self._lock:
+            if thread_id not in self._thread_sandboxes:
+                return None
+
+            existing_id = self._thread_sandboxes[thread_id]
+            if existing_id in self._sandboxes:
+                suffix = " (post-lock check)" if post_lock else ""
+                logger.info(f"Reusing in-process sandbox {existing_id} for thread {thread_id}{suffix}")
+                self._last_activity[existing_id] = time.time()
+                return existing_id
+
+            del self._thread_sandboxes[thread_id]
+            return None
+
+    def _reclaim_warm_pool_sandbox(self, thread_id: str | None, sandbox_id: str, *, post_lock: bool = False) -> str | None:
+        """Promote a warm-pool sandbox back to active tracking if available."""
+        if thread_id is None:
+            return None
+
+        with self._lock:
+            if sandbox_id not in self._warm_pool:
+                return None
+
+            info, _ = self._warm_pool.pop(sandbox_id)
+            sandbox = AioSandbox(id=sandbox_id, base_url=info.sandbox_url)
+            self._sandboxes[sandbox_id] = sandbox
+            self._sandbox_infos[sandbox_id] = info
+            self._last_activity[sandbox_id] = time.time()
+            self._thread_sandboxes[thread_id] = sandbox_id
+
+        suffix = " (post-lock check)" if post_lock else f" at {info.sandbox_url}"
+        logger.info(f"Reclaimed warm-pool sandbox {sandbox_id} for thread {thread_id}{suffix}")
+        return sandbox_id
+
+    def _recheck_cached_sandbox(self, thread_id: str, sandbox_id: str) -> str | None:
+        """Re-check in-memory caches after acquiring the cross-process file lock."""
+        return self._reuse_in_process_sandbox(thread_id, post_lock=True) or self._reclaim_warm_pool_sandbox(thread_id, sandbox_id, post_lock=True)
+
+    def _register_discovered_sandbox(self, thread_id: str, info: SandboxInfo) -> str:
+        """Track a sandbox discovered through the backend."""
+        sandbox = AioSandbox(id=info.sandbox_id, base_url=info.sandbox_url)
+        with self._lock:
+            self._sandboxes[info.sandbox_id] = sandbox
+            self._sandbox_infos[info.sandbox_id] = info
+            self._last_activity[info.sandbox_id] = time.time()
+            self._thread_sandboxes[thread_id] = info.sandbox_id
+
+        logger.info(f"Discovered existing sandbox {info.sandbox_id} for thread {thread_id} at {info.sandbox_url}")
+        return info.sandbox_id
+
+    def _register_created_sandbox(self, thread_id: str | None, sandbox_id: str, info: SandboxInfo) -> str:
+        """Track a newly-created sandbox in the active maps."""
+        sandbox = AioSandbox(id=sandbox_id, base_url=info.sandbox_url)
+        with self._lock:
+            self._sandboxes[sandbox_id] = sandbox
+            self._sandbox_infos[sandbox_id] = info
+            self._last_activity[sandbox_id] = time.time()
+            if thread_id:
+                self._thread_sandboxes[thread_id] = sandbox_id
+
+        logger.info(f"Created sandbox {sandbox_id} for thread {thread_id} at {info.sandbox_url}")
+        return sandbox_id
+
+    def _replica_count(self) -> tuple[int, int]:
+        """Return configured replicas and currently tracked sandbox count."""
+        replicas = self._config.get("replicas", DEFAULT_REPLICAS)
+        with self._lock:
+            total = len(self._sandboxes) + len(self._warm_pool)
+        return replicas, total
+
+    def _log_replicas_soft_cap(self, replicas: int, sandbox_id: str, evicted: str | None) -> None:
+        """Log the result of enforcing the warm-pool replica budget."""
+        if evicted:
+            logger.info(f"Evicted warm-pool sandbox {evicted} to stay within replicas={replicas}")
+            return
+
+        # All slots are occupied by active sandboxes — proceed anyway and log.
+        # The replicas limit is a soft cap; we never forcibly stop a container
+        # that is actively serving a thread.
+        logger.warning(f"All {replicas} replica slots are in active use; creating sandbox {sandbox_id} beyond the soft limit")
+
     # ── Core: acquire / get / release / shutdown ─────────────────────────
 
     def acquire(self, thread_id: str | None = None) -> str:
@@ -440,6 +569,23 @@ class AioSandboxProvider(SandboxProvider):
         else:
             return self._acquire_internal(thread_id)
 
+    async def acquire_async(self, thread_id: str | None = None) -> str:
+        """Acquire a sandbox environment without blocking the event loop.
+
+        Mirrors ``acquire()`` while keeping blocking backend operations off the
+        event loop and using async-native readiness polling for newly created
+        sandboxes.
+        """
+        if thread_id:
+            thread_lock = self._get_thread_lock(thread_id)
+            await _acquire_thread_lock_async(thread_lock)
+            try:
+                return await self._acquire_internal_async(thread_id)
+            finally:
+                thread_lock.release()
+
+        return await self._acquire_internal_async(thread_id)
+
     def _acquire_internal(self, thread_id: str | None) -> str:
         """Internal sandbox acquisition with two-layer consistency.
 
@@ -448,33 +594,17 @@ class AioSandboxProvider(SandboxProvider):
                  sandbox_id is deterministic from thread_id so no shared state file
                  is needed — any process can derive the same container name)
         """
-        # ── Layer 1: In-process cache (fast path) ──
-        if thread_id:
-            with self._lock:
-                if thread_id in self._thread_sandboxes:
-                    existing_id = self._thread_sandboxes[thread_id]
-                    if existing_id in self._sandboxes:
-                        logger.info(f"Reusing in-process sandbox {existing_id} for thread {thread_id}")
-                        self._last_activity[existing_id] = time.time()
-                        return existing_id
-                    else:
-                        del self._thread_sandboxes[thread_id]
+        cached_id = self._reuse_in_process_sandbox(thread_id)
+        if cached_id is not None:
+            return cached_id
 
         # Deterministic ID for thread-specific, random for anonymous
-        sandbox_id = self._deterministic_sandbox_id(thread_id) if thread_id else str(uuid.uuid4())[:8]
+        sandbox_id = self._sandbox_id_for_thread(thread_id)
 
         # ── Layer 1.5: Warm pool (container still running, no cold-start) ──
-        if thread_id:
-            with self._lock:
-                if sandbox_id in self._warm_pool:
-                    info, _ = self._warm_pool.pop(sandbox_id)
-                    sandbox = AioSandbox(id=sandbox_id, base_url=info.sandbox_url)
-                    self._sandboxes[sandbox_id] = sandbox
-                    self._sandbox_infos[sandbox_id] = info
-                    self._last_activity[sandbox_id] = time.time()
-                    self._thread_sandboxes[thread_id] = sandbox_id
-                    logger.info(f"Reclaimed warm-pool sandbox {sandbox_id} for thread {thread_id} at {info.sandbox_url}")
-                    return sandbox_id
+        reclaimed_id = self._reclaim_warm_pool_sandbox(thread_id, sandbox_id)
+        if reclaimed_id is not None:
+            return reclaimed_id
 
         # ── Layer 2: Backend discovery + create (protected by cross-process lock) ──
         # Use a file lock so that two processes racing to create the same sandbox
@@ -485,6 +615,26 @@ class AioSandboxProvider(SandboxProvider):
 
         return self._create_sandbox(thread_id, sandbox_id)
 
+    async def _acquire_internal_async(self, thread_id: str | None) -> str:
+        """Async counterpart to ``_acquire_internal``."""
+        cached_id = self._reuse_in_process_sandbox(thread_id)
+        if cached_id is not None:
+            return cached_id
+
+        # Deterministic ID for thread-specific, random for anonymous
+        sandbox_id = self._sandbox_id_for_thread(thread_id)
+
+        # ── Layer 1.5: Warm pool (container still running, no cold-start) ──
+        reclaimed_id = self._reclaim_warm_pool_sandbox(thread_id, sandbox_id)
+        if reclaimed_id is not None:
+            return reclaimed_id
+
+        # ── Layer 2: Backend discovery + create (protected by cross-process lock) ──
+        if thread_id:
+            return await self._discover_or_create_with_lock_async(thread_id, sandbox_id)
+
+        return await self._create_sandbox_async(thread_id, sandbox_id)
+
     def _discover_or_create_with_lock(self, thread_id: str, sandbox_id: str) -> str:
         """Discover an existing sandbox or create a new one under a cross-process file lock.
 
@@ -503,40 +653,50 @@ class AioSandboxProvider(SandboxProvider):
                 locked = True
                 # Re-check in-process caches under the file lock in case another
                 # thread in this process won the race while we were waiting.
-                with self._lock:
-                    if thread_id in self._thread_sandboxes:
-                        existing_id = self._thread_sandboxes[thread_id]
-                        if existing_id in self._sandboxes:
-                            logger.info(f"Reusing in-process sandbox {existing_id} for thread {thread_id} (post-lock check)")
-                            self._last_activity[existing_id] = time.time()
-                            return existing_id
-                    if sandbox_id in self._warm_pool:
-                        info, _ = self._warm_pool.pop(sandbox_id)
-                        sandbox = AioSandbox(id=sandbox_id, base_url=info.sandbox_url)
-                        self._sandboxes[sandbox_id] = sandbox
-                        self._sandbox_infos[sandbox_id] = info
-                        self._last_activity[sandbox_id] = time.time()
-                        self._thread_sandboxes[thread_id] = sandbox_id
-                        logger.info(f"Reclaimed warm-pool sandbox {sandbox_id} for thread {thread_id} (post-lock check)")
-                        return sandbox_id
+                cached_id = self._recheck_cached_sandbox(thread_id, sandbox_id)
+                if cached_id is not None:
+                    return cached_id
 
                 # Backend discovery: another process may have created the container.
                 discovered = self._backend.discover(sandbox_id)
                 if discovered is not None:
-                    sandbox = AioSandbox(id=discovered.sandbox_id, base_url=discovered.sandbox_url)
-                    with self._lock:
-                        self._sandboxes[discovered.sandbox_id] = sandbox
-                        self._sandbox_infos[discovered.sandbox_id] = discovered
-                        self._last_activity[discovered.sandbox_id] = time.time()
-                        self._thread_sandboxes[thread_id] = discovered.sandbox_id
-                    logger.info(f"Discovered existing sandbox {discovered.sandbox_id} for thread {thread_id} at {discovered.sandbox_url}")
-                    return discovered.sandbox_id
+                    return self._register_discovered_sandbox(thread_id, discovered)
 
                 return self._create_sandbox(thread_id, sandbox_id)
             finally:
                 if locked:
                     _unlock_file(lock_file)
 
+    async def _discover_or_create_with_lock_async(self, thread_id: str, sandbox_id: str) -> str:
+        """Async counterpart to ``_discover_or_create_with_lock``."""
+        paths = get_paths()
+        user_id = get_effective_user_id()
+        await asyncio.to_thread(paths.ensure_thread_dirs, thread_id, user_id=user_id)
+        lock_path = paths.thread_dir(thread_id, user_id=user_id) / f"{sandbox_id}.lock"
+
+        lock_file = await asyncio.to_thread(_open_lock_file, lock_path)
+        locked = False
+        try:
+            await asyncio.to_thread(_lock_file_exclusive, lock_file)
+            locked = True
+            # Re-check in-process caches under the file lock in case another
+            # thread in this process won the race while we were waiting.
+            cached_id = self._recheck_cached_sandbox(thread_id, sandbox_id)
+            if cached_id is not None:
+                return cached_id
+
+            # Backend discovery is sync because local discovery may inspect
+            # Docker and perform a health check; keep it off the event loop.
+            discovered = await asyncio.to_thread(self._backend.discover, sandbox_id)
+            if discovered is not None:
+                return self._register_discovered_sandbox(thread_id, discovered)
+
+            return await self._create_sandbox_async(thread_id, sandbox_id)
+        finally:
+            if locked:
+                await asyncio.to_thread(_unlock_file, lock_file)
+            await asyncio.to_thread(lock_file.close)
+
     def _evict_oldest_warm(self) -> str | None:
         """Destroy the oldest container in the warm pool to free capacity.
 
@@ -574,18 +734,10 @@ class AioSandboxProvider(SandboxProvider):
 
         # Enforce replicas: only warm-pool containers count toward eviction budget.
         # Active sandboxes are in use by live threads and must not be forcibly stopped.
-        replicas = self._config.get("replicas", DEFAULT_REPLICAS)
-        with self._lock:
-            total = len(self._sandboxes) + len(self._warm_pool)
+        replicas, total = self._replica_count()
         if total >= replicas:
             evicted = self._evict_oldest_warm()
-            if evicted:
-                logger.info(f"Evicted warm-pool sandbox {evicted} to stay within replicas={replicas}")
-            else:
-                # All slots are occupied by active sandboxes — proceed anyway and log.
-                # The replicas limit is a soft cap; we never forcibly stop a container
-                # that is actively serving a thread.
-                logger.warning(f"All {replicas} replica slots are in active use; creating sandbox {sandbox_id} beyond the soft limit")
+            self._log_replicas_soft_cap(replicas, sandbox_id, evicted)
 
         info = self._backend.create(thread_id, sandbox_id, extra_mounts=extra_mounts or None)
 
@@ -594,16 +746,27 @@ class AioSandboxProvider(SandboxProvider):
             self._backend.destroy(info)
             raise RuntimeError(f"Sandbox {sandbox_id} failed to become ready within timeout at {info.sandbox_url}")
 
-        sandbox = AioSandbox(id=sandbox_id, base_url=info.sandbox_url)
-        with self._lock:
-            self._sandboxes[sandbox_id] = sandbox
-            self._sandbox_infos[sandbox_id] = info
-            self._last_activity[sandbox_id] = time.time()
-            if thread_id:
-                self._thread_sandboxes[thread_id] = sandbox_id
+        return self._register_created_sandbox(thread_id, sandbox_id, info)
 
-        logger.info(f"Created sandbox {sandbox_id} for thread {thread_id} at {info.sandbox_url}")
-        return sandbox_id
+    async def _create_sandbox_async(self, thread_id: str | None, sandbox_id: str) -> str:
+        """Async counterpart to ``_create_sandbox``."""
+        extra_mounts = await asyncio.to_thread(self._get_extra_mounts, thread_id)
+
+        # Enforce replicas: only warm-pool containers count toward eviction budget.
+        # Active sandboxes are in use by live threads and must not be forcibly stopped.
+        replicas, total = self._replica_count()
+        if total >= replicas:
+            evicted = await asyncio.to_thread(self._evict_oldest_warm)
+            self._log_replicas_soft_cap(replicas, sandbox_id, evicted)
+
+        info = await asyncio.to_thread(self._backend.create, thread_id, sandbox_id, extra_mounts=extra_mounts or None)
+
+        # Wait for sandbox to be ready without blocking the event loop.
+        if not await wait_for_sandbox_ready_async(info.sandbox_url, timeout=60):
+            await asyncio.to_thread(self._backend.destroy, info)
+            raise RuntimeError(f"Sandbox {sandbox_id} failed to become ready within timeout at {info.sandbox_url}")
+
+        return self._register_created_sandbox(thread_id, sandbox_id, info)
 
     def get(self, sandbox_id: str) -> Sandbox | None:
         """Get a sandbox by ID. Updates last activity timestamp.
@@ -627,14 +790,20 @@ class AioSandboxProvider(SandboxProvider):
         thread on its next turn without a cold-start.  The container will only be
         stopped when the replicas limit forces eviction or during shutdown.
 
+        The host-side HTTP client owned by the cached ``AioSandbox`` instance is
+        closed before the instance is dropped (#2872). The warm-pool entry only
+        stores ``SandboxInfo``, so a fresh ``AioSandbox`` (and a fresh client)
+        is constructed if the container is later reclaimed.
+
         Args:
             sandbox_id: The ID of the sandbox to release.
         """
         info = None
+        sandbox = None
         thread_ids_to_remove: list[str] = []
 
         with self._lock:
-            self._sandboxes.pop(sandbox_id, None)
+            sandbox = self._sandboxes.pop(sandbox_id, None)
             info = self._sandbox_infos.pop(sandbox_id, None)
             thread_ids_to_remove = [tid for tid, sid in self._thread_sandboxes.items() if sid == sandbox_id]
             for tid in thread_ids_to_remove:
@@ -644,6 +813,15 @@ class AioSandboxProvider(SandboxProvider):
             if info and sandbox_id not in self._warm_pool:
                 self._warm_pool[sandbox_id] = (info, time.time())
 
+        if sandbox is not None:
+            # Defense-in-depth: close() already swallows its own errors; this
+            # guard only protects against a future close() that misbehaves, so
+            # host-side client cleanup can never block parking in the warm pool.
+            try:
+                sandbox.close()
+            except Exception as e:
+                logger.warning(f"Error closing sandbox {sandbox_id} during release: {e}")
+
         logger.info(f"Released sandbox {sandbox_id} to warm pool (container still running)")
 
     def destroy(self, sandbox_id: str) -> None:
@@ -652,14 +830,19 @@ class AioSandboxProvider(SandboxProvider):
         Unlike release(), this actually stops the container.  Use this for
         explicit cleanup, capacity-driven eviction, or shutdown.
 
+        The host-side HTTP client owned by the cached ``AioSandbox`` instance is
+        closed alongside backend/container destruction so no client/socket
+        resources leak (#2872).
+
         Args:
             sandbox_id: The ID of the sandbox to destroy.
         """
         info = None
+        sandbox = None
         thread_ids_to_remove: list[str] = []
 
         with self._lock:
-            self._sandboxes.pop(sandbox_id, None)
+            sandbox = self._sandboxes.pop(sandbox_id, None)
             info = self._sandbox_infos.pop(sandbox_id, None)
             thread_ids_to_remove = [tid for tid, sid in self._thread_sandboxes.items() if sid == sandbox_id]
             for tid in thread_ids_to_remove:
@@ -671,6 +854,15 @@ class AioSandboxProvider(SandboxProvider):
             else:
                 self._warm_pool.pop(sandbox_id, None)
 
+        if sandbox is not None:
+            # Defense-in-depth: close() already swallows its own errors; this
+            # guard only protects against a future close() that misbehaves, so
+            # host-side client cleanup can never block container destruction.
+            try:
+                sandbox.close()
+            except Exception as e:
+                logger.warning(f"Error closing sandbox {sandbox_id} during destroy: {e}")
+
         if info:
             self._backend.destroy(info)
             logger.info(f"Destroyed sandbox {sandbox_id}")

backend/packages/harness/deerflow/community/aio_sandbox/backend.py

@@ -2,10 +2,12 @@
 
 from __future__ import annotations
 
+import asyncio
 import logging
 import time
 from abc import ABC, abstractmethod
 
+import httpx
 import requests
 
 from .sandbox_info import SandboxInfo
@@ -35,6 +37,34 @@ def wait_for_sandbox_ready(sandbox_url: str, timeout: int = 30) -> bool:
     return False
 
 
+async def wait_for_sandbox_ready_async(sandbox_url: str, timeout: int = 30, poll_interval: float = 1.0) -> bool:
+    """Async variant of sandbox readiness polling.
+
+    Use this from async runtime paths so sandbox startup waits do not block the
+    event loop. The synchronous ``wait_for_sandbox_ready`` function remains for
+    existing synchronous backend/provider call sites.
+    """
+    loop = asyncio.get_running_loop()
+    deadline = loop.time() + timeout
+
+    async with httpx.AsyncClient(timeout=5) as client:
+        while True:
+            remaining = deadline - loop.time()
+            if remaining <= 0:
+                break
+            try:
+                response = await client.get(f"{sandbox_url}/v1/sandbox", timeout=min(5.0, remaining))
+                if response.status_code == 200:
+                    return True
+            except httpx.RequestError:
+                pass
+            remaining = deadline - loop.time()
+            if remaining <= 0:
+                break
+            await asyncio.sleep(min(poll_interval, remaining))
+    return False
+
+
 class SandboxBackend(ABC):
     """Abstract base for sandbox provisioning backends.
 
@@ -44,7 +74,7 @@ class SandboxBackend(ABC):
     """
 
     @abstractmethod
-    def create(self, thread_id: str, sandbox_id: str, extra_mounts: list[tuple[str, str, bool]] | None = None) -> SandboxInfo:
+    def create(self, thread_id: str | None, sandbox_id: str, extra_mounts: list[tuple[str, str, bool]] | None = None) -> SandboxInfo:
         """Create/provision a new sandbox.
 
         Args:

backend/packages/harness/deerflow/community/aio_sandbox/local_backend.py

@@ -241,7 +241,7 @@ class LocalContainerBackend(SandboxBackend):
 
     # ── SandboxBackend interface ──────────────────────────────────────────
 
-    def create(self, thread_id: str, sandbox_id: str, extra_mounts: list[tuple[str, str, bool]] | None = None) -> SandboxInfo:
+    def create(self, thread_id: str | None, sandbox_id: str, extra_mounts: list[tuple[str, str, bool]] | None = None) -> SandboxInfo:
         """Start a new container and return its connection info.
 
         Args:

backend/packages/harness/deerflow/community/aio_sandbox/remote_backend.py

@@ -59,7 +59,7 @@ class RemoteSandboxBackend(SandboxBackend):
 
     def create(
         self,
-        thread_id: str,
+        thread_id: str | None,
         sandbox_id: str,
         extra_mounts: list[tuple[str, str, bool]] | None = None,
     ) -> SandboxInfo:
@@ -132,7 +132,7 @@ class RemoteSandboxBackend(SandboxBackend):
             logger.warning("Provisioner list_running failed: %s", exc)
             return []
 
-    def _provisioner_create(self, thread_id: str, sandbox_id: str, extra_mounts: list[tuple[str, str, bool]] | None = None) -> SandboxInfo:
+    def _provisioner_create(self, thread_id: str | None, sandbox_id: str, extra_mounts: list[tuple[str, str, bool]] | None = None) -> SandboxInfo:
         """POST /api/sandboxes → create Pod + Service."""
         try:
             resp = requests.post(

backend/packages/harness/deerflow/community/ddg_search/tools.py

@@ -11,12 +11,85 @@ from deerflow.config import get_app_config
 
 logger = logging.getLogger(__name__)
 
+DEFAULT_BACKEND = "auto"
+DEFAULT_REGION = "wt-wt"
+DEFAULT_SAFESEARCH = "moderate"
+DEFAULT_WIKIPEDIA_REGION = "us-en"
+
+WIKIPEDIA_BACKENDS = {"auto", "all", "wikipedia"}
+WIKIPEDIA_LANGUAGE_ALIASES = {
+    "jp": "ja",
+    "kr": "ko",
+    "tzh": "zh",
+    "wt": "en",
+}
+
+
+def _normalize_backend(backend: str | list[str] | tuple[str, ...] | None) -> str:
+    if backend is None:
+        return DEFAULT_BACKEND
+    if isinstance(backend, (list, tuple)):
+        return ",".join(str(part).strip() for part in backend if str(part).strip()) or DEFAULT_BACKEND
+    return str(backend).strip() or DEFAULT_BACKEND
+
+
+def _normalize_setting(value: str | None, default: str) -> str:
+    return str(value).strip() if value else default
+
+
+def _backend_includes_wikipedia(backend: str | list[str] | tuple[str, ...] | None) -> bool:
+    backend = _normalize_backend(backend)
+    return any(part.strip().lower() in WIKIPEDIA_BACKENDS for part in backend.split(","))
+
+
+def _contains_codepoint(query: str, ranges: tuple[tuple[int, int], ...]) -> bool:
+    return any(start <= ord(char) <= end for char in query for start, end in ranges)
+
+
+def _infer_wikipedia_region(query: str) -> str:
+    """Pick a valid Wikipedia language region when DDGS' worldwide region is used."""
+    if _contains_codepoint(query, ((0x3040, 0x30FF), (0x31F0, 0x31FF))):
+        return "jp-ja"
+    if _contains_codepoint(query, ((0xAC00, 0xD7AF), (0x1100, 0x11FF), (0x3130, 0x318F))):
+        return "kr-ko"
+    if _contains_codepoint(query, ((0x3400, 0x9FFF),)):
+        return "cn-zh"
+    if _contains_codepoint(query, ((0x0400, 0x04FF),)):
+        return "ru-ru"
+    if _contains_codepoint(query, ((0x0370, 0x03FF),)):
+        return "gr-el"
+    if _contains_codepoint(query, ((0x0590, 0x05FF),)):
+        return "il-he"
+    if _contains_codepoint(query, ((0x0600, 0x06FF),)):
+        return "xa-ar"
+    return DEFAULT_WIKIPEDIA_REGION
+
+
+def _resolve_ddgs_region(query: str, region: str | None, backend: str | list[str] | tuple[str, ...] | None) -> str:
+    """
+    DDGS' wikipedia engine treats the second part of region as a Wikipedia
+    subdomain. Its default worldwide region, wt-wt, becomes wt.wikipedia.org.
+    """
+    normalized_region = _normalize_setting(region, DEFAULT_REGION).lower()
+    if not _backend_includes_wikipedia(backend):
+        return normalized_region
+
+    if normalized_region == DEFAULT_REGION:
+        return _infer_wikipedia_region(query)
+
+    if "-" not in normalized_region:
+        return DEFAULT_WIKIPEDIA_REGION
+
+    country, language = normalized_region.split("-", 1)
+    return f"{country}-{WIKIPEDIA_LANGUAGE_ALIASES.get(language, language)}"
+
 
 def _search_text(
     query: str,
     max_results: int = 5,
-    region: str = "wt-wt",
-    safesearch: str = "moderate",
+    region: str | None = DEFAULT_REGION,
+    safesearch: str | None = DEFAULT_SAFESEARCH,
+    backend: str | list[str] | tuple[str, ...] | None = DEFAULT_BACKEND,
 ) -> list[dict]:
     """
     Execute text search using DuckDuckGo.
@@ -26,6 +99,7 @@ def _search_text(
         max_results: Maximum number of results
         region: Search region
         safesearch: Safe search level
+        backend: DDGS backend(s), e.g. "auto", "duckduckgo", or "duckduckgo,brave"
 
     Returns:
         List of search results
@@ -39,11 +113,15 @@ def _search_text(
     ddgs = DDGS(timeout=30)
 
     try:
+        backend = _normalize_backend(backend)
+        safesearch = _normalize_setting(safesearch, DEFAULT_SAFESEARCH)
+        effective_region = _resolve_ddgs_region(query, region, backend)
         results = ddgs.text(
             query,
-            region=region,
+            region=effective_region,
             safesearch=safesearch,
             max_results=max_results,
+            backend=backend,
         )
         return list(results) if results else []
 
@@ -64,14 +142,23 @@ def web_search_tool(
         max_results: Maximum number of results to return. Default is 5.
     """
     config = get_app_config().get_tool_config("web_search")
+    region = DEFAULT_REGION
+    safesearch = DEFAULT_SAFESEARCH
+    backend = DEFAULT_BACKEND
 
-    # Override max_results from config if set
-    if config is not None and "max_results" in config.model_extra:
+    if config is not None:
+        # Override tool call defaults from config if set.
         max_results = config.model_extra.get("max_results", max_results)
+        region = config.model_extra.get("region", region)
+        safesearch = config.model_extra.get("safesearch", safesearch)
+        backend = config.model_extra.get("backend", backend)
 
     results = _search_text(
         query=query,
         max_results=max_results,
+        region=region,
+        safesearch=safesearch,
+        backend=backend,
     )
 
     if not results:

backend/packages/harness/deerflow/community/jina_ai/jina_client.py

@@ -9,7 +9,7 @@ _api_key_warned = False
 
 
 class JinaClient:
-    async def crawl(self, url: str, return_format: str = "html", timeout: int = 10) -> str:
+    async def crawl(self, url: str, return_format: str = "html", timeout: int = 10, proxy: str | None = None, trust_env: bool = True) -> str:
         global _api_key_warned
         headers = {
             "Content-Type": "application/json",
@@ -23,7 +23,10 @@ class JinaClient:
             logger.warning("Jina API key is not set. Provide your own key to access a higher rate limit. See https://jina.ai/reader for more information.")
         data = {"url": url}
         try:
-            async with httpx.AsyncClient() as client:
+            client_kwargs: dict[str, object] = {"trust_env": trust_env}
+            if proxy:
+                client_kwargs["proxy"] = proxy
+            async with httpx.AsyncClient(**client_kwargs) as client:
                 response = await client.post("https://r.jina.ai/", headers=headers, json=data, timeout=timeout)
 
             if response.status_code != 200:

backend/packages/harness/deerflow/community/jina_ai/tools.py

@@ -9,6 +9,38 @@ from deerflow.utils.readability import ReadabilityExtractor
 readability_extractor = ReadabilityExtractor()
 
 
+def _coerce_bool(value: object, default: bool) -> bool:
+    if isinstance(value, bool):
+        return value
+    if isinstance(value, str):
+        normalized = value.strip().lower()
+        if normalized in {"1", "true", "yes", "on"}:
+            return True
+        if normalized in {"0", "false", "no", "off"}:
+            return False
+    return default
+
+
+def _coerce_timeout(value: object, default: int) -> int:
+    if isinstance(value, bool):
+        return default
+    if isinstance(value, int):
+        return value
+    if isinstance(value, str):
+        try:
+            return int(value)
+        except ValueError:
+            return default
+    return default
+
+
+def _coerce_proxy(value: object) -> str | None:
+    if not isinstance(value, str):
+        return None
+    proxy = value.strip()
+    return proxy or None
+
+
 @tool("web_fetch", parse_docstring=True)
 async def web_fetch_tool(url: str) -> str:
     """Fetch the contents of a web page at a given URL.
@@ -22,10 +54,14 @@ async def web_fetch_tool(url: str) -> str:
     """
     jina_client = JinaClient()
     timeout = 10
+    proxy = None
+    trust_env = True
     config = get_app_config().get_tool_config("web_fetch")
-    if config is not None and "timeout" in config.model_extra:
-        timeout = config.model_extra.get("timeout")
-    html_content = await jina_client.crawl(url, return_format="html", timeout=timeout)
+    if config is not None:
+        timeout = _coerce_timeout(config.model_extra.get("timeout"), timeout)
+        proxy = _coerce_proxy(config.model_extra.get("proxy"))
+        trust_env = _coerce_bool(config.model_extra.get("trust_env"), trust_env)
+    html_content = await jina_client.crawl(url, return_format="html", timeout=timeout, proxy=proxy, trust_env=trust_env)
     if isinstance(html_content, str) and html_content.startswith("Error:"):
         return html_content
     article = await asyncio.to_thread(readability_extractor.extract_article, html_content)

backend/packages/harness/deerflow/config/app_config.py

@@ -7,7 +7,7 @@ from typing import Any, Self
 
 import yaml
 from dotenv import load_dotenv
-from pydantic import BaseModel, ConfigDict, Field
+from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from deerflow.config.acp_config import ACPAgentConfig, load_acp_config_from_dict
 from deerflow.config.agents_api_config import AgentsApiConfig, load_agents_api_config_from_dict
@@ -18,8 +18,10 @@ from deerflow.config.guardrails_config import GuardrailsConfig, load_guardrails_
 from deerflow.config.loop_detection_config import LoopDetectionConfig
 from deerflow.config.memory_config import MemoryConfig, load_memory_config_from_dict
 from deerflow.config.model_config import ModelConfig
+from deerflow.config.reload_boundary import format_field_description
 from deerflow.config.run_events_config import RunEventsConfig
 from deerflow.config.runtime_paths import existing_project_file
+from deerflow.config.safety_finish_reason_config import SafetyFinishReasonConfig
 from deerflow.config.sandbox_config import SandboxConfig
 from deerflow.config.skill_evolution_config import SkillEvolutionConfig
 from deerflow.config.skills_config import SkillsConfig
@@ -29,6 +31,7 @@ from deerflow.config.summarization_config import SummarizationConfig, load_summa
 from deerflow.config.title_config import TitleConfig, load_title_config_from_dict
 from deerflow.config.token_usage_config import TokenUsageConfig
 from deerflow.config.tool_config import ToolConfig, ToolGroupConfig
+from deerflow.config.tool_output_config import ToolOutputConfig
 from deerflow.config.tool_search_config import ToolSearchConfig, load_tool_search_config_from_dict
 
 load_dotenv()
@@ -83,15 +86,27 @@ def apply_logging_level(name: str | None) -> None:
 class AppConfig(BaseModel):
     """Config for the DeerFlow application"""
 
-    log_level: str = Field(default="info", description="Logging level for deerflow and app modules (debug/info/warning/error); third-party libraries are not affected")
+    log_level: str = Field(
+        default="info",
+        description=format_field_description(
+            "log_level",
+            field_doc="Logging level for deerflow and app modules (debug/info/warning/error); third-party libraries are not affected.",
+        ),
+    )
     token_usage: TokenUsageConfig = Field(default_factory=TokenUsageConfig, description="Token usage tracking configuration")
     models: list[ModelConfig] = Field(default_factory=list, description="Available models")
-    sandbox: SandboxConfig = Field(description="Sandbox configuration")
+    sandbox: SandboxConfig = Field(
+        description=format_field_description(
+            "sandbox",
+            field_doc="Sandbox provider configuration (local filesystem or Docker-based aio sandbox).",
+        ),
+    )
     tools: list[ToolConfig] = Field(default_factory=list, description="Available tools")
     tool_groups: list[ToolGroupConfig] = Field(default_factory=list, description="Available tool groups")
     skills: SkillsConfig = Field(default_factory=SkillsConfig, description="Skills configuration")
     skill_evolution: SkillEvolutionConfig = Field(default_factory=SkillEvolutionConfig, description="Agent-managed skill evolution configuration")
     extensions: ExtensionsConfig = Field(default_factory=ExtensionsConfig, description="Extensions configuration (MCP servers and skills state)")
+    tool_output: ToolOutputConfig = Field(default_factory=ToolOutputConfig, description="Tool output budget protection configuration")
     tool_search: ToolSearchConfig = Field(default_factory=ToolSearchConfig, description="Tool search / deferred loading configuration")
     title: TitleConfig = Field(default_factory=TitleConfig, description="Automatic title generation configuration")
     summarization: SummarizationConfig = Field(default_factory=SummarizationConfig, description="Conversation summarization configuration")
@@ -102,11 +117,51 @@ class AppConfig(BaseModel):
     guardrails: GuardrailsConfig = Field(default_factory=GuardrailsConfig, description="Guardrail middleware configuration")
     circuit_breaker: CircuitBreakerConfig = Field(default_factory=CircuitBreakerConfig, description="LLM circuit breaker configuration")
     loop_detection: LoopDetectionConfig = Field(default_factory=LoopDetectionConfig, description="Loop detection middleware configuration")
+    safety_finish_reason: SafetyFinishReasonConfig = Field(default_factory=SafetyFinishReasonConfig, description="Provider safety-filter finish_reason interception middleware configuration")
     model_config = ConfigDict(extra="allow")
-    database: DatabaseConfig = Field(default_factory=DatabaseConfig, description="Unified database backend configuration")
-    run_events: RunEventsConfig = Field(default_factory=RunEventsConfig, description="Run event storage configuration")
-    checkpointer: CheckpointerConfig | None = Field(default=None, description="Checkpointer configuration")
-    stream_bridge: StreamBridgeConfig | None = Field(default=None, description="Stream bridge configuration")
+    database: DatabaseConfig = Field(
+        default_factory=DatabaseConfig,
+        description=format_field_description(
+            "database",
+            field_doc="Unified database backend for run/feedback metadata (memory, sqlite, or postgres).",
+        ),
+    )
+    run_events: RunEventsConfig = Field(
+        default_factory=RunEventsConfig,
+        description=format_field_description(
+            "run_events",
+            field_doc="Run-event store backend (memory for dev, db for production queries, jsonl for lightweight single-node persistence).",
+        ),
+    )
+    checkpointer: CheckpointerConfig | None = Field(
+        default=None,
+        description=format_field_description(
+            "checkpointer",
+            field_doc="LangGraph state-persistence checkpointer configuration.",
+        ),
+    )
+    stream_bridge: StreamBridgeConfig | None = Field(
+        default=None,
+        description=format_field_description(
+            "stream_bridge",
+            field_doc="Stream bridge connecting agent workers to SSE endpoints.",
+        ),
+    )
+
+    @field_validator("models", "tools", "tool_groups", mode="before")
+    @classmethod
+    def _coerce_null_list_sections(cls, value: Any) -> Any:
+        """Treat a present-but-empty config section as an empty list.
+
+        Commenting out every entry under a top-level YAML key — e.g. ``models:``
+        with only comments beneath it, exactly as shipped in
+        ``config.example.yaml`` — makes PyYAML parse the value as ``None``.
+        Without this, the documented ``cp config.example.yaml config.yaml``
+        first-run flow crashes with an opaque ``Input should be a valid list``
+        pydantic error. Coercing ``None`` to ``[]`` keeps that flow working and
+        matches the field's own ``default_factory=list``.
+        """
+        return [] if value is None else value
 
     @classmethod
     def resolve_config_path(cls, config_path: str | None = None) -> Path:
@@ -169,6 +224,11 @@ class AppConfig(BaseModel):
         config_data["extensions"] = extensions_config.model_dump()
 
         result = cls.model_validate(config_data)
+        if not result.models:
+            logger.warning(
+                "No models are configured in %s. Add at least one entry under `models:` (see the commented examples in config.example.yaml) or run `make setup`.",
+                resolved_path,
+            )
         acp_agents = cls._validate_acp_agents(config_data.get("acp_agents", {}))
         cls._apply_singleton_configs(result, acp_agents)
         return result

backend/packages/harness/deerflow/config/checkpointer_config.py

@@ -41,6 +41,20 @@ def set_checkpointer_config(config: CheckpointerConfig | None) -> None:
     _checkpointer_config = config
 
 
+def ensure_config_loaded() -> None:
+    """Lazily load app config when checkpointer config has not been initialized."""
+    from deerflow.config.app_config import _app_config, get_app_config
+
+    config = get_checkpointer_config()
+    if config is not None or _app_config is not None:
+        return
+
+    try:
+        get_app_config()
+    except FileNotFoundError:
+        pass
+
+
 def load_checkpointer_config_from_dict(config_dict: dict | None) -> None:
     """Load checkpointer configuration from a dictionary."""
     global _checkpointer_config

backend/packages/harness/deerflow/config/extensions_config.py

@@ -5,7 +5,7 @@ import os
 from pathlib import Path
 from typing import Any, Literal
 
-from pydantic import BaseModel, ConfigDict, Field
+from pydantic import BaseModel, ConfigDict, Field, model_validator
 
 from deerflow.config.runtime_paths import existing_project_file
 
@@ -47,6 +47,24 @@ class McpServerConfig(BaseModel):
     description: str = Field(default="", description="Human-readable description of what this MCP server provides")
     model_config = ConfigDict(extra="allow")
 
+    @model_validator(mode="before")
+    @classmethod
+    def _accept_transport_alias(cls, data: Any) -> Any:
+        """Accept the MCP-spec ``transport`` field as an alias for ``type``.
+
+        The official MCP configuration schema uses ``transport`` to indicate
+        the transport mechanism (``stdio``/``sse``/``http``). Earlier versions
+        of this project only honored ``type``, which caused remote SSE/HTTP
+        servers configured with just ``transport`` to be incorrectly treated as
+        ``stdio`` (the default). This validator normalizes the two so either
+        spelling works, with ``type`` taking precedence when both are provided.
+        """
+        if isinstance(data, dict):
+            transport = data.get("transport")
+            if transport and not data.get("type"):
+                data = {**data, "type": transport}
+        return data
+
 
 class SkillStateConfig(BaseModel):
     """Configuration for a single skill's state."""
@@ -141,7 +159,7 @@ class ExtensionsConfig(BaseModel):
         try:
             with open(resolved_path, encoding="utf-8") as f:
                 config_data = json.load(f)
-            cls.resolve_env_variables(config_data)
+            config_data = cls.resolve_env_variables(config_data)
             return cls.model_validate(config_data)
         except json.JSONDecodeError as e:
             raise ValueError(f"Extensions config file at {resolved_path} is not valid JSON: {e}") from e
@@ -149,7 +167,7 @@ class ExtensionsConfig(BaseModel):
             raise RuntimeError(f"Failed to load extensions config from {resolved_path}: {e}") from e
 
     @classmethod
-    def resolve_env_variables(cls, config: dict[str, Any]) -> dict[str, Any]:
+    def resolve_env_variables(cls, config: Any) -> Any:
         """Recursively resolve environment variables in the config.
 
         Environment variables are resolved using the `os.getenv` function. Example: $OPENAI_API_KEY
@@ -160,23 +178,26 @@ class ExtensionsConfig(BaseModel):
         Returns:
             The config with environment variables resolved.
         """
-        for key, value in config.items():
-            if isinstance(value, str):
-                if value.startswith("$"):
-                    env_value = os.getenv(value[1:])
-                    if env_value is None:
-                        # Unresolved placeholder — store empty string so downstream
-                        # consumers (e.g. MCP servers) don't receive the literal "$VAR"
-                        # token as an actual environment value.
-                        config[key] = ""
-                    else:
-                        config[key] = env_value
-                else:
-                    config[key] = value
-            elif isinstance(value, dict):
-                config[key] = cls.resolve_env_variables(value)
-            elif isinstance(value, list):
-                config[key] = [cls.resolve_env_variables(item) if isinstance(item, dict) else item for item in value]
+        if isinstance(config, str):
+            if not config.startswith("$"):
+                return config
+            env_value = os.getenv(config[1:])
+            if env_value is None:
+                # Unresolved placeholder — store empty string so downstream
+                # consumers (e.g. MCP servers) don't receive the literal "$VAR"
+                # token as an actual environment value.
+                return ""
+            return env_value
+
+        if isinstance(config, dict):
+            return {key: cls.resolve_env_variables(value) for key, value in config.items()}
+
+        if isinstance(config, list):
+            return [cls.resolve_env_variables(item) for item in config]
+
+        if isinstance(config, tuple):
+            return tuple(cls.resolve_env_variables(item) for item in config)
+
         return config
 
     def get_enabled_mcp_servers(self) -> dict[str, McpServerConfig]:

backend/packages/harness/deerflow/config/model_config.py

@@ -32,6 +32,16 @@ class ModelConfig(BaseModel):
         description="Extra settings to be passed to the model when thinking is disabled",
     )
     supports_vision: bool = Field(default_factory=lambda: False, description="Whether the model supports vision/image inputs")
+    stream_chunk_timeout: float | None = Field(
+        default=None,
+        description=(
+            "Maximum seconds to wait between successive streaming chunks before "
+            "langchain-openai raises StreamChunkTimeoutError. None means use the "
+            "factory default (240s for OpenAI-compatible clients). Tune higher for "
+            "reasoning models with long thinking pauses; lower for latency-sensitive "
+            "interactive endpoints. Has no effect on non-OpenAI-compatible providers."
+        ),
+    )
     thinking: dict | None = Field(
         default_factory=lambda: None,
         description=(

backend/packages/harness/deerflow/config/paths.py

@@ -1,3 +1,4 @@
+import hashlib
 import os
 import re
 import shutil
@@ -10,6 +11,8 @@ VIRTUAL_PATH_PREFIX = "/mnt/user-data"
 
 _SAFE_THREAD_ID_RE = re.compile(r"^[A-Za-z0-9_\-]+$")
 _SAFE_USER_ID_RE = re.compile(r"^[A-Za-z0-9_\-]+$")
+_UNSAFE_USER_ID_CHAR_RE = re.compile(r"[^A-Za-z0-9_\-]")
+_SAFE_USER_ID_DIGEST_HEX_LEN = 16
 
 
 def _default_local_base_dir() -> Path:
@@ -31,6 +34,23 @@ def _validate_user_id(user_id: str) -> str:
     return user_id
 
 
+def make_safe_user_id(raw: str) -> str:
+    """Normalize an external identity into the user-id charset (``[A-Za-z0-9_-]``).
+
+    IM channel ids (Feishu/Slack/Telegram) may contain characters that
+    :func:`_validate_user_id` rejects. Already-safe ids pass through unchanged;
+    lossy ones get a short digest suffix so two distinct inputs never share a
+    storage bucket.
+    """
+    if not raw:
+        raise ValueError("user_id must be a non-empty string.")
+    sanitized = _UNSAFE_USER_ID_CHAR_RE.sub("-", raw)
+    if sanitized == raw:
+        return raw
+    digest = hashlib.sha1(raw.encode("utf-8")).hexdigest()[:_SAFE_USER_ID_DIGEST_HEX_LEN]
+    return f"{sanitized}-{digest}"
+
+
 def _join_host_path(base: str, *parts: str) -> str:
     """Join host filesystem path segments while preserving native style.
 

backend/packages/harness/deerflow/config/reload_boundary.py

@@ -0,0 +1,104 @@
+"""Single source of truth for the config hot-reload boundary.
+
+Bytedance/deer-flow issue #3144: gateway request dependencies resolve
+``AppConfig`` through ``get_app_config()`` on every request, so per-run
+fields take effect on the next message without restarting the gateway.
+The fields listed in this module are the **infrastructure** subset that
+the gateway captures once at startup — engines, singletons, IM clients,
+the logging handler — and that therefore require a process restart to
+change at runtime.
+
+The registry covers two kinds of entries:
+
+- Top-level ``AppConfig`` fields (``database``, ``checkpointer``,
+  ``run_events``, ``stream_bridge``, ``sandbox``, ``log_level``). For
+  these, :func:`format_field_description` produces the standardised
+  ``"startup-only: ..."`` prefix that the matching Pydantic
+  ``Field(description=...)`` carries, so the boundary surfaces in IDE
+  hover next to the field itself.
+- Top-level ``config.yaml`` sections that are not part of the
+  ``AppConfig`` schema (``channels``). These cannot be standardised at
+  the schema level, so the registry is their only canonical location.
+
+Any future "needs restart" scanner — operator tooling, lint hooks, doc
+generators — should drive off this registry rather than re-parsing
+prose.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Iterator
+
+#: The standardised prefix every restart-required field description starts
+#: with. ``test_reload_boundary`` enforces both directions: registered
+#: fields must use this prefix in the schema, and any schema field using
+#: this prefix must be in the registry.
+STARTUP_ONLY_PREFIX = "startup-only:"
+
+
+#: Restart-required field paths mapped to the human-readable reason.
+#:
+#: The reason text is what surfaces in ``Field(description=...)``, so it
+#: must explain *what* code captures the snapshot — not just that the
+#: field is restart-required — so an operator changing the value knows
+#: which subsystem to restart.
+STARTUP_ONLY_FIELDS: dict[str, str] = {
+    "database": ("init_engine_from_config() runs once during langgraph_runtime() startup; the SQLAlchemy engine holds the connection pool and is not rebuilt on config.yaml edits."),
+    "checkpointer": ("make_checkpointer() binds the persistent checkpointer once at startup, including SQLite WAL / busy_timeout settings."),
+    "run_events": ("make_run_event_store() picks the memory- vs SQL-backed implementation at startup and is frozen onto app.state.run_events_config to stay paired with the underlying event store."),
+    "stream_bridge": ("make_stream_bridge() constructs the stream-bridge singleton once during startup."),
+    "sandbox": ("get_sandbox_provider() caches the provider singleton (``_default_sandbox_provider``); a different ``sandbox.use`` class path only takes effect on next process start."),
+    "log_level": (
+        "apply_logging_level() runs only during app.py startup; it sets the deerflow/app logger levels and may lower root handler thresholds so configured messages can propagate. A freshly reloaded AppConfig does not retrigger it."
+    ),
+    # Not part of the AppConfig Pydantic schema — channel credentials are
+    # consumed directly by ``start_channel_service()`` once at lifespan
+    # startup and the live channel clients are not rebuilt on
+    # config.yaml edits.
+    "channels": ("start_channel_service() is invoked once during startup; the live IM channel clients (Feishu, Slack, Telegram, DingTalk) are not rebuilt when channels.* changes."),
+}
+
+
+def iter_startup_only_field_paths() -> Iterator[str]:
+    """Yield every registered restart-required field path."""
+    return iter(STARTUP_ONLY_FIELDS)
+
+
+def is_startup_only_field(field_path: str) -> bool:
+    """Return ``True`` when *field_path* is registered as restart-required.
+
+    Accepts only top-level paths (``"database"``, ``"sandbox"`` etc.);
+    nested keys like ``"database.url"`` are not modelled here because the
+    boundary is per-section, not per-leaf.
+    """
+    return field_path in STARTUP_ONLY_FIELDS
+
+
+def format_field_description(field_path: str, *, field_doc: str | None = None) -> str:
+    """Build the standardised description for a registered field.
+
+    Used inside ``AppConfig`` ``Field(description=...)`` so the hover
+    text in IDEs matches the registry and the drift tests can pin one
+    side against the other.
+
+    Args:
+        field_path: A registered top-level field path (e.g. ``"log_level"``).
+        field_doc: Optional human-facing description for the field itself
+            (allowed values, semantics, etc.). When supplied, it is
+            appended after the ``startup-only:`` marker block separated by
+            a blank line so IDE hover shows both the restart-required
+            reason *and* the field's normal documentation. Composition
+            keeps the marker as the leading token machine-readable tooling
+            pivots on while restoring the prose that ``Field(description=)``
+            used to carry before the registry took over.
+
+    Raises:
+        KeyError: when *field_path* is not registered. This is deliberate
+            — silently returning a placeholder would let a typo bypass
+            the drift coverage.
+    """
+    reason = STARTUP_ONLY_FIELDS[field_path]
+    header = f"{STARTUP_ONLY_PREFIX} {reason}"
+    if field_doc is None:
+        return header
+    return f"{header}\n\n{field_doc.strip()}"

backend/packages/harness/deerflow/config/safety_finish_reason_config.py

@@ -0,0 +1,47 @@
+"""Configuration for SafetyFinishReasonMiddleware.
+
+Mirrors the shape of GuardrailsConfig: detectors are loaded by class path
+through ``deerflow.reflection.resolve_variable`` (same loader the
+``guardrails.provider`` config uses) so users can drop in custom provider
+detectors without modifying core code.
+"""
+
+from __future__ import annotations
+
+from pydantic import BaseModel, Field
+
+
+class SafetyDetectorConfig(BaseModel):
+    """One detector entry under ``safety_finish_reason.detectors``."""
+
+    use: str = Field(
+        description=("Class path of a SafetyTerminationDetector implementation (e.g. 'deerflow.agents.middlewares.safety_termination_detectors:OpenAICompatibleContentFilterDetector')."),
+    )
+    config: dict = Field(
+        default_factory=dict,
+        description="Constructor kwargs passed to the detector class.",
+    )
+
+
+class SafetyFinishReasonConfig(BaseModel):
+    """Configuration for the SafetyFinishReasonMiddleware.
+
+    The middleware intercepts AIMessages where the provider signaled a
+    safety-related termination (e.g. OpenAI ``finish_reason='content_filter'``)
+    while still returning tool calls, and suppresses those tool calls so the
+    half-truncated arguments never execute.
+    """
+
+    enabled: bool = Field(
+        default=True,
+        description="Master switch for the SafetyFinishReasonMiddleware.",
+    )
+    detectors: list[SafetyDetectorConfig] | None = Field(
+        default=None,
+        description=(
+            "Custom detector list. Leave unset (None) to use the built-in "
+            "set covering OpenAI-compatible content_filter, Anthropic "
+            "refusal, and Gemini SAFETY/BLOCKLIST/PROHIBITED_CONTENT/SPII/"
+            "RECITATION. Provide a non-null list to fully override."
+        ),
+    )

backend/packages/harness/deerflow/config/sandbox_config.py

@@ -4,7 +4,20 @@ from pydantic import BaseModel, ConfigDict, Field
 class VolumeMountConfig(BaseModel):
     """Configuration for a volume mount."""
 
-    host_path: str = Field(..., description="Path on the host machine")
+    host_path: str = Field(
+        ...,
+        description=(
+            "Source path for the mount. Resolution depends on the active provider: "
+            "``LocalSandboxProvider`` checks this path from the gateway process — in "
+            "``make dev`` that is the host machine, but in Docker deployments "
+            "(``make up`` / docker-compose) it is the path *inside* the "
+            "``deer-flow-gateway`` container, so the host directory must also be "
+            "bind-mounted into the gateway service for the mount to take effect. "
+            "``AioSandboxProvider`` (DooD) passes this value straight to ``docker -v`` "
+            "for the sandbox container, where it is resolved by the host Docker daemon "
+            "from the host machine's perspective."
+        ),
+    )
     container_path: str = Field(..., description="Path inside the container")
     read_only: bool = Field(default=False, description="Whether the mount is read-only")
 

backend/packages/harness/deerflow/config/title_config.py

@@ -51,3 +51,16 @@ def load_title_config_from_dict(config_dict: dict) -> None:
     """Load title configuration from a dictionary."""
     global _title_config
     _title_config = TitleConfig(**config_dict)
+
+
+def reset_title_config() -> None:
+    """Restore the title configuration to its pristine ``TitleConfig()`` default.
+
+    Public API so that tests do not have to reach into the private
+    ``_title_config`` module attribute. ``AppConfig.from_file()`` calls
+    :func:`load_title_config_from_dict`, which permanently mutates the
+    singleton; tests that need a clean slate between cases should call
+    this between tests.
+    """
+    global _title_config
+    _title_config = TitleConfig()