deer-flow/CHANGELOG.md

Changelog

All notable changes to DeerFlow are documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[Unreleased]

This section accumulates work toward the 2.1.0 milestone (milestone 2).

⚠ Breaking changes

• gateway: Request trace ids are now issued unconditionally, and every Gateway HTTP response carries an X-Trace-Id header. Previously both were gated behind logging.enhance.enabled, which now controls log output only — whether records carry a trace_id field, and in which format. The header cannot be turned off; installations running the default enabled: false will start seeing it after upgrading. Scheduled tasks, MCP task notification runs, IM channel messages, and the embedded DeerFlowClient bind an id per unit of work, so the id also reaches the run record, the checkpoint metadata, and Langfuse traces that previously had none. A deerflow_trace_id supplied in a run request's metadata or config.context is now ignored and overwritten so the response header, the logs, and the persisted run cannot disagree — send the X-Trace-Id request header to pin a correlation id across services. logging remains restart-required. No config keys were added or removed. (#5119)
• skills: Sandboxes now reserve /mnt/skills for managed enabled-only projections. DEER_FLOW_HOST_SKILLS_PATH and SKILLS_HOST_PATH are no longer used; Docker/AIO and hostPath deployments derive projection paths from DEER_FLOW_HOST_BASE_DIR. E2B operator mounts targeting /mnt/skills or any child path are skipped with a warning so they cannot shadow the managed projection; move extra E2B content to a different container path. User projections re-read global enable state from disk so toggles propagate across Gateway workers on the next sandbox acquire. Existing E2B sandboxes retain their creation-time snapshot until they are recreated. PVC-backed provisioner deployments still mount the operator-supplied PVC snapshot directly, so disabled-skill filesystem isolation does not apply in PVC mode until dynamic PVC materialization is implemented. (#4178)
• sandbox: E2B now enforces sandbox.replicas as a process-local capacity limit. The default wait policy waits for acquire_timeout, then fails the agent turn. DeerFlow does not retry the turn automatically. Use burst with burst_limit to permit bounded extra VMs. The reject policy can remove one warm VM before it returns a capacity error. (#4391)
• skills: A directory containing SKILL.md is now a runtime package boundary. Nested SKILL.md files inside that package are supporting data and are no longer registered as independent skills; unusual custom layouts must move independently loadable skills under a namespace directory without its own SKILL.md. (#4098)
• memory: The memory system is now pluggable (memory.manager_class selects a backend; default deermem is self-contained). DeerMem-private settings moved from the top level of memory: into memory.backend_config, and the /memory/config response (and client.get_memory_config()) changed shape. (#4122)
• memory: /memory/config and client.get_memory_config() no longer return flat DeerMem fields (storage_path, max_facts, debounce_seconds, token_counting, guaranteed_*, staleness_*, ...). They return {enabled, mode, injection_enabled, manager_class, backend_config} where backend_config is an opaque dict the active backend self-interprets. Memory data responses (/memory, /memory/status data) are unchanged. External API/SDK clients reading the old flat fields must read backend_config instead. (#4122)
• memory: Custom memory.storage_class moved: the old default path deerflow.agents.memory.storage.FileMemoryStorage no longer exists (now deerflow.agents.memory.backends.deermem.deermem.core.storage.FileMemoryStorage). Custom MemoryStorage subclasses must accept config in __init__ (was no-arg). A broken/old storage_class logs an error and falls back to FileMemoryStorage (won't crash) -- update the path + signature to restore it. (#4122)
• memory: storage_path semantics changed from a FILE path to a root DIRECTORY. Pre-abstraction, an absolute storage_path was the shared memory file (opting out of per-user isolation) and a relative value was the global file under the data base_dir. Now storage_path (absolute or relative) is the root directory; per-user memory lives at {storage_path}/users/{uid}/memory.json. An upgrade keeping the old default storage_path: memory.json (a relative file name) would orphan per-user memory or hit NotADirectoryError on save, so the legacy migration drops file-style storage_path values (ending in .json) with a warning and the factory raises if storage_path resolves to an existing file. Set memory.backend_config.storage_path to a directory for a custom root. (#4122)
• memory: memory.mode: tool with a backend that does not implement search() now fails fast at Gateway startup with a ValueError from the MemoryManager invariant, instead of starting successfully and silently returning empty results on every memory_search call. Both shipping backends implement search() (DeerMem retrieves; noop returns []), so this only affects a custom backend that onboards without overriding search(). It is intentional -- silent empties are worse than a loud startup error. Fix: switch to mode: middleware or override search() (and set supports_search=True). (#4324)
• config: database.checkpoint_delta_snapshot_frequency moved to database.checkpoint_delta.snapshot_frequency and its default changed from 1000 to 10. A legacy top-level value is still honored with a deprecation warning and mapped onto the nested key (an explicitly set nested key wins). Deployments that relied on the old default now snapshot 100x more often in delta mode -- set database.checkpoint_delta.snapshot_frequency: 1000 explicitly to keep the previous cadence. (#4516)
• docker: The published entry port now binds to loopback (127.0.0.1) by default in both compose files, matching the documented local-trust deployment model. Deployments that relied on the old 0.0.0.0 binding must set BIND_HOST to expose the stack on other interfaces. (#4618)

Added

Scheduler

• scheduler: Scheduled tasks accept interval (schedule_spec.every_seconds) in addition to once and cron. Cadence is UTC now + N with no missed-beat catch-up. N is at least scheduler.min_once_delay_seconds (default 60s) and at most 30 days.

Authentication

• auth: Personal access tokens (PAT) for programmatic API access: POST/GET/DELETE /api/v1/auth/pats manage tokens (shown once, stored as SHA-256 digests); a default-deny route policy admits only the thread/run lifecycle routes, narrowed further by the token's threads/runs scopes, and any request dimension that carries cancel capability (?action=, multitask_strategy) additionally requires runs:cancel. (#5041)
• auth: Login throttling parameters are configurable via auth.local.max_login_attempts (default 5, min 2) and auth.local.lockout_seconds (default 300), resolved live so a config reload applies on the next login without a Gateway restart — the unblock path for deployments behind corporate proxies/NAT where many users share one egress IP. Defaults are unchanged. (#5110)

Agents & runtime

• scheduler: Scheduled tasks can pin assistant_id to lead_agent (the default) or a custom agent the owner already has. Unknown or malformed names return 422. The workspace create/edit form exposes the same choice. ([#5286])
• gateway: GET /api/threads/{thread_id}/runs/page walks thread run history with a (created_at, run_id) keyset cursor ({data, has_more, next_before_created_at, next_before_run_id}). GET /api/threads/{thread_id}/runs still returns a bare array of the newest 100 runs so LangGraph SDK clients keep working. (#5282)
• middleware: New TokenBudgetMiddleware enforces a per-run token budget, shared additively across the lead agent and subagents. (#3412)
• middleware: Structured tool-result metadata and a tool-progress state machine give the runtime first-class visibility into multi-step tool flows. (#3601)
• context: Record the effective memory identity per run and persist durable context (system messages, memory, and tool state) across summarization, emitting it as structured runtime metadata so compaction no longer drops it. (#3556, #3887, #3906)
• runtime: Goal continuations let a run resume toward a goal across multiple agent turns, with continuation_count tracked and capped. (#3858)
• subagents: A system-maintained delegation ledger prevents redundant re-delegation of an in-flight task, and a total delegation cap bounds fan-out per run. (#3877, #4115)
• subagents: Persist and display subagent step history in the thread. (#3845)
• tools: Structured synopses replace raw oversized tool output in previews. (#3377)
• files: Deterministic read-before-write version gate for file tools prevents clobbering concurrent edits. (#3912)
• gateway: Cache-aware cost accounting attributes token costs to cached vs. uncached paths; a Redis stream bridge enables distributed event streaming; and manual context compaction is exposed to the user. (#3920, #3191, #3969)
• gateway: The stream-bridge heartbeat interval is configurable via stream_bridge.heartbeat_interval_seconds (default 15s), so deployments behind aggressive proxy idle timeouts can tune SSE, /wait, and internal subscribers together. (#5017)
• runtime: Dual-mode checkpoint storage with LangGraph DeltaChannel cuts thread storage from O(N²) to near-linear for long research/coding runs. (#4292)
• runtime: Delta-mode checkpoint history cache (memory/redis) with O(1) incremental composition, configured via database.checkpoint_cache. (#4638)
• agent: Config-declared lead-agent middlewares let deployments add custom AgentMiddleware classes without patching the runtime chain. (#3964)
• agents: Per-agent model and generation settings (temperature, max_tokens, thinking_enabled, reasoning_effort) override the shared model profile. (#4347)
• runtime: Record terminal artifact-delivery receipts so runs expected to present_files no longer report success when delivery fails. (#4365)
• uploads: Lazy-load historical files via a list_uploaded_files tool instead of injecting the full manifest. (#4174)
• scheduler: scheduler.recursion_limit in config.yaml sets the LangGraph super-step cap for scheduled runs (default 1000, matching the web UI's interactive budget, clamped by max_recursion_limit). (#4848)
• runtime: Every tool call now carries a runtime-stamped, tamper-evident tool receipt, and a bounded receipt ledger is injected into the model context so agents can cite execution evidence in their reports. Enabled by default via the new verification config section. (#4659)
• subagents: Subagent delegations are now verifiable, layering RFC #4651: every subagent's report contract requires citing tool receipts (e.g. [r3 write_file]) and attaching a verifiable handle to each deliverable, the lead agent cross-checks those citations against the subagent's actual execution record, and acceptance_criteria on a task delegation are checked deterministically parent-side (file existence/non-emptiness, recorded test-command exit status) with anything undecidable reported UNVERIFIED instead of silently passed. (#5076, #5090, #5109)
• clarification: Human-input (clarification) cards support structured form fields, so an agent can request exactly the input it needs instead of free text only. (#4406)
• subagents: Built-in subagents now receive the current-date context anchor, so delegated tasks involving relative dates behave like tasks the lead agent handles directly. (#4797)
• subagents: A Settings page manages a deployment-level Subagent catalog (admin-managed worker definitions alongside built-in and config.yaml ones), and Custom Agents can restrict delegation to an explicit worker allowlist enforced at both prompt and execution time. (#4887)
• subagents: Subagent concurrency is now governed by one process-wide capacity controller, and an opt-in batch_task tool runs large collections of independent items as durable, resumable SQL-backed batches with leases, bounded retries, pause/resume/cancel, and a chat panel for tracking progress. (#4998)
• agents: The current-date context injected into lead- and subagent prompts honors the optional DEER_FLOW_DATE_TIMEZONE env var (IANA name, e.g. Asia/Shanghai), so users in non-UTC deployments are no longer told the wrong "today" around midnight; unset keeps server-local behavior. (#5154)
• subagents: Delegated subagents can discover files uploaded in earlier turns: the parent run's validated uploaded_files boundary seeds the subagent's graph state, making list_uploaded_files eligible for normal tool-policy filtering (durable batch_task workers keep it disabled). (#5170)
• agents: The read-before-write gate now elides the dead payload of a blocked write_file / str_replace call (content, old_str, new_str) from model-bound requests. A blocked call never ran and must be re-issued after a re-read, so the original arguments only cost context; stored history, receipts, and the run journal keep them. Blocked results are paired with call occurrences (tool-call ids may repeat across turns), and a request whose history was rewritten drops OpenAI resp_ response ids so use_previous_response_id chaining cannot resume the original server-side history. Controlled by read_before_write.elide_blocked_payloads (default on) and read_before_write.elide_min_chars (default 2000).
• agents: ToolOutputBudgetMiddleware now also elides the content of a successful write_file call from model-bound requests once the same path was read or modified again later in the conversation. After a successful write the file on disk is the source of truth, and the read-before-write gate forces a read_file before the next modification, so the historical copy was redundant with that read and long report-writing runs carried every section twice. The newest tool_output.keep_recent_writes successful writes (default 1) always stay visible, str_replace payloads are never touched, and stored history, receipts, and the run journal keep the original arguments. Controlled by tool_output.elide_superseded_writes (default on) and tool_output.superseded_write_min_chars (default 2000).

Memory

• memory: Memory consolidation synthesizes fragmented facts, and a staleness review prunes silently-outdated facts using LLM-assigned per-fact expected_valid_days / staleFactsToExtend. (#3996, #3860, #4143)
• memory: Guaranteed injection of correction facts (with graceful fallback) so user corrections always reach the model. (#3592)
• memory: Slim the pluggable MemoryManager interface for backend onboarding - new backends no longer implement unused abstract methods, and DeerMem-specific hook injection moves out of the shared factory. (#4326)
• memory: Incremental agent-scoped Markdown fact storage isolates per-agent facts and updates a single fact without rewriting or reindexing the whole collection. (#4279)
• memory: Memory message processing adds a conversation watermark, trivial-turn filtering, and a durable queue so extraction no longer re-feeds the full conversation every turn. (#4447)
• memory: A built-in FTS5/BM25 retrieval adapter provides full-text search over stored memories without an external retrieval service. (#4360)
• memory: New pluggable memory backends: OpenViking and mem0 over HTTP, plus Honcho as a user-model memory provider. (#4509, #4528, #4730)
• memory: A hybrid fact eviction policy blends multiple signals when deciding which stored facts to drop as memory fills. (#4789)

Skills

• skills: The built-in image-generation skill can use OpenAI-compatible Images APIs for generation and reference-image editing, with configurable endpoint, model, size, and output format.
• skills: Native SkillScan (phase 1) statically analyzes skill packages at load, and describe_skill enables deferred discovery so the model fetches a skill's schema on demand instead of loading all skills up front. (#3033, #3775)
• skills: Per-user custom skill isolation with sandbox mounting. (#3889)
• skills: The skill list reopens after a skill is selected, so several skills can be attached in a row. (#4639)
• skills: Install local .skill archives directly from the Skills settings page, reusing the existing per-user installer and security scan. (#5039)
• skills: Podcast-generation Volcengine voices are configurable per speaker gender, with trimmed blank-safe defaults. (#5156)

Models & integrations

• community: New web search/fetch engines - GroundRoute, Crawl4AI (web_fetch), and a fastCRW provider - plus a Browserless web_capture screenshot tool and Brave image_search. (#3675, #3821, #3585, #3881, #3866)
• mcp: Per-server tool_call_timeout for MCP tool calls, and routing hints that guide the model to the right server. (#3843, #4004)
• mcp: Add an official OpenViking /mcp example that exposes the native tool set through DeerFlow's generic MCP client. (#4745)
• community: Agentic browser control as a first-class thread capability - Playwright-backed browser sessions the agent operates while the user observes or takes over from the workspace. (#4187)
• community: Lark/Feishu CLI integration bundles the runtime install, the official lark-* skill pack, and an interactive auth flow so the integration is no longer environment-dependent. (#3971)
• integrations: Lark/Feishu app credentials can be switched per user from Settings > Integrations: new App ID/Secret values are validated before anything is committed, and the previous OAuth token is revoked after a successful switch. (#4703)
• acp: MiniMax Code (mcode acp) is supported and documented as a native external coding agent, and ACP thought chunks are no longer concatenated into tool results. (#4846)
• models: A Z.AI GLM-5.3-Flash profile keeps thinking permanently enabled and stops generic reasoning-effort forwarding, since the model rejects disabled thinking and only accepts its own effort values. (#5074)
• community: New web search providers - Serply (with news and scholar verticals) and Tencent Cloud WSA - plus native recency filters (day/week/month/year) shared across DDGS, Brave, Tavily, and SearXNG. (#5023, #5057, #5099)
• community: New Sofya web_search and web_fetch provider - search results carry the content of each page, capped per result so a default search stays inline. (#5239)
• knowledge: Opt-in read-only RAGFlow retrieval exposes a knowledge_search(query) agent tool over configured RAGFlow datasets, with a dataset-ID allowlist and credential/dataset-id redaction on error paths. (#4955)
• knowledge: Opt-in read-only LightRAG retrieval is an alternative knowledge_search(query) provider for the same knowledge group — operators pick RAGFlow or LightRAG by which entry is configured. It queries LightRAG's structured /query/data endpoint (no LLM generation), formats the ranked chunks as citation-numbered text, redacts the optional X-API-Key on every model-visible path, and maps server error messages to actionable tool errors. (#5209)

MCP

• mcp: A durable task runtime for MCP: long-running tool tasks survive Gateway restarts through a durable driver, and their progress and completion notifications surface in the chat UI. (#4665, #4690, #4833)
• mcp: Shared MCP servers can inject per-user credentials: a single server entry authenticates each DeerFlow user with their own header value, unmapped users are denied by default, and stored credentials are masked in Gateway API responses. (#4868)
• mcp: Per-server tool_name_prefix option lets servers that already namespace their own tools keep their original tool names; the default behavior is unchanged. (#4624)
• mcp: Settings > Tools can add, edit, and delete MCP servers through targeted Gateway endpoints, with a copy-paste JSON workflow that preserves advanced fields and masked secret placeholders. (#5022)
• mcp: Shared HTTP/SSE servers can map request-scoped secrets to headers via headers_from_context: callers supply per-request values in config.context.secrets, the config stores only key names, and missing values deny by default. (#5010)
• mcp: An optional parallel-search server entry (https://search.parallel.ai/mcp, HTTP, no auth by default) ships in extensions_config.example.json, disabled by default; enabling it exposes parallel-search_web_search and parallel-search_web_fetch, with optional Bearer authentication documented. (#5028)

Channels

• channels: Expose the IM channel_user_id to sandbox commands as DEERFLOW_CHANNEL_USER_ID. (#3926)
• channels: Queue rapid same-thread messages and preserve topic-card previews across batches. (#3988)
• channels: Inbound webhook deduplication moves to Postgres, so several Gateway pods can serve the same IM channel without double-processing events. (#4210)
• channels: DingTalk inbound messages support file and image attachments. (#4423)
• channels: New Buzz (Nostr) channel connector, including the frontend experience for the channel. (#4649, #4727)
• channels: /agent list and /agent use <name> IM commands let a conversation switch to the owner's Custom Agents: the selection is persisted in thread metadata (restored after restart) and wins over stale channel defaults, IM-created threads route through the same agent when opened in the web UI, and /agent is reserved across the slash-skill parser, frontend, and TUI so no skill can shadow the command. (#5168)

Auth & guardrails

• auth: Generic OIDC/SSO authentication with Keycloak support. (#3506)
• guardrails: Authenticated runtime context is exposed in GuardrailRequest, and security interventions are persisted as run events. (#3665, #3837)
• auth: "Keep me signed in" login option with a centralized session-cookie policy (persistent Secure cookies on HTTPS, session cookies on public HTTP). (#4255)
• auth: Deployments can close local self-registration to restrict new accounts to SSO/OIDC provisioning. (#4311)
• authz: Built-in RBAC authorization provider with a unified factory, plus tool-authorization enforcement at both assembly (tools removed before the model sees them) and runtime (denied calls blocked). (#4260, #4370)
• authz: Gateway route permissions are derived from the configured AuthorizationProvider rather than a fixed table. (#4439)
• authz: Model authorization is enforced at Gateway routes and again in the agent runtime, and sandbox:execute is checked when a sandbox is acquired - users can no longer reach models or sandboxes they are not authorized for. (#4540, #4911)
• authz: GET /auth/me now surfaces the caller's effective route permissions (RFC #4063 Phase 4), read from the AuthContext the auth middleware already stamps on every authenticated request — no extra provider evaluations — so the frontend can hide actions the caller's role cannot perform. (#5228)

Sandbox & provisioner

• sandbox: New E2B and BoxLite (micro-VM) sandbox providers; BoxLite ships with a warm pool. (#3883, #3940, #3951)
• provisioner: ClusterIP Services and scoped per-skill PVC mounts, plus a configurable sandbox container port. (#4016, #3928)
• sandbox: New cloud sandbox providers: Tenki and OpenSandbox. (#4382, #4877)
• sandbox: An optional lark-cli credential broker sidecar (K8s provisioner mode) keeps Lark app secrets and OAuth tokens out of the sandbox filesystem entirely - the sandbox sees only a shim that forwards commands to a loopback broker in the pod. Off by default. (#4501)
• sandbox: The E2B mount-upload wall-clock deadline is configurable via mount_upload_deadline_seconds (default 120s). (#4876)
• sandbox: An E2B sandbox carries a structured MountUploadResult (truncated, reason, upload totals) after creation — preserved across warm-pool reclaim within the process — so mount truncation by a resource limit is observable in code instead of only in Gateway logs. (#4884)
• sandbox: Opt-in controlled egress for local Docker AIO sandboxes: sandbox.network.mode supports isolated (per-sandbox internal bridge with no outbound route) and allowlist (the same bridge plus a domain-allowlist HTTP(S) policy sidecar that resolves destinations itself and rejects IP literals and ECH); denied public domains can be approved through the Human Input card (temporary grant or allow-for-this-sandbox), non-interactive runs fail closed, and the sandbox API is no longer published directly in restricted modes. Requires Docker Engine 28+; open remains the default. (#5152)

Extensions & plugins

• extensions: An out-of-tree Python extension system: extensions can contribute middleware, task-lifecycle and system-model observers, Gateway services, and HTTP routers, and are managed with deerflow extensions install/enable/disable/remove. (#4636, #4684, #4780)
• extensions: Extensions can observe what the agent did - message provenance, middleware policy declarations, agent-assembly fingerprints, context-compaction records, guardrail decisions, and the MCP origin of a tool. deerflow-extension-api moves to 0.2.0; extensions written against 0.1 are refused at startup with an install hint. (#4863)

Persistence

• persistence: A custom PostgreSQL schema can be selected via postgres_schema; ORM, LangGraph checkpointer, and store tables are all created there, and the schema is created automatically at startup. (#3442)

Frontend

• frontend: Branching support for assistant turns and side conversations for quoted follow-ups. (#3950, #3934)
• frontend: Regenerate the latest answer. (#3637)
• frontend: Citation-sources evidence panel, workspace change review for agent runs, and a visualized ask_clarification card. (#3907, #3945, #3956)
• frontend: Voice dictation, prompt-history recall with arrow keys, composer input polishing, and a "(thought for N seconds)" thinking-duration chip. (#4036, #3718, #3986, #3627)
• frontend: Feature-gate the agents UI behind the agents_api flag, and persist AI turn duration in backend and UI. (#3769, #3663)
• frontend: Render slash-skill activations as inline chips. (#3981)
• frontend: Localized AI-assistance disclaimer. (#4374)
• frontend: Pin recent chats. (#4442)
• frontend: Validate /goal objective length in the composer. (#4337)
• frontend: Real-time context window usage is shown as a conversation grows. (#3183)
• frontend: The latest user turn can be edited and rerun in place. (#4377)
• frontend: Replies can be typed and sent while a clarification card is pending. (#4530)
• suggestions: The number of follow-up suggestions is configurable via suggestions.max_suggestions (default 3). (#4533)
• artifacts: Text artifacts can be edited inline in the artifact panel. (#4596)
• artifacts: Markdown artifacts open rendered in a new-window reader (with "View source" and "Download" fallbacks), and all files presented in a run can be downloaded as one zip archive derived from the run's delivery receipt. (#5056, #5117)
• frontend: Browser Live is available in Custom Agent chats. (#4719)
• frontend: A conversation outline navigates long chats: past 5 user turns, a compact side menu lists the conversation's questions and jumps between them, tracking the current section. (#5025)
• frontend: Scheduled tasks can be duplicated into an editable draft that carries over the configuration but not the run history. (#5064)
• threads: Branched conversations get distinguishing titles (automatic Title (2), Title (3) sibling numbering) and the recent-chats list shows parent-child lineage with tree connectors. (#4983)
• frontend: Chats can be archived and restored: an Archive sidebar action with an Undo toast, Recent chats / Archived tabs above search, and per-chat restore controls; SQL and Memory stores apply the archive filter before pagination while messages, files, links, pin state, and the current URL are preserved. (#5236)
• projects: Project workspaces organize chats (Projects MVP Phase 1): a sidebar Projects section with flat/grouped list modes, a project detail page with a paginated thread list, project-scoped new chats whose threads are pre-created with their project so a run can never land outside it, branch membership inheritance, move-between-projects, and project create/rename/archive/restore/delete. (#5265)
• artifacts: Completed CSV/TSV artifacts preview as bounded tables (up to 200 rows × 50 columns, 50 rows per page, sticky headers, optional first-row header) parsed off the main thread over the existing 1 MiB range loader — literal strings, leading zeros, and multiline quoted cells survive, long cells open in a copyable dialog, and source view remains one toggle away. (#5284)

Observability & tooling

• observability: Trace-id correlation with enhanced logging and agent observability via Monocle. (#3902, #4024)
• tooling: A Hermes-like terminal workbench (deerflow CLI) backed by DeerFlowClient, plus a redacted community support-bundle generator. (#3760, #3886)
• setup: The setup wizard now asks whether OpenAI-compatible gateway models support thinking, and a Volcengine Coding Plan quick-setup path was added. (#3428, #4141)
• tui: clear command. (#4306)
• tui: The TUI supports a transparent terminal background. (#4631)
• gateway: New GET /health/ready readiness probe runs a bounded database SELECT 1 and returns 503 while the database is unreachable (200 not_configured for the memory backend); /health stays pure liveness, and the production compose healthcheck now gates on readiness. (#5166)
• observability: Deferred tool promotions — both routing-hint auto-promotion and explicit tool_search — persist as privacy-minimal middleware:tool_promotion run events (tool names, source, count, agent attribution; no queries, schemas, or results), observed only after skill-policy filtering so denied schemas are never reported as effective promotions. (#5183)

Changed

• frontend performance: Keep the public root and localized docs static; lazy-load closed workspace panels and editor/highlighter dependencies; incrementally derive streamed message state; bound streaming Markdown work; virtualize long message and chat lists; pause offscreen decorative effects; and enforce representative route JS/CSS budgets.
• browser: Negotiate binary Browser Live JPEG frames, retain the legacy JSON/base64 protocol for older clients, coalesce presentation to the latest frame per refresh, and revoke replaced object URLs.
• artifacts: Stream regular text artifacts with HTTP byte-range support and limit the initial Web UI preview to 1 MiB until the user explicitly loads the complete file.
• sandbox: The Helm chart now defaults per-sandbox Services to ClusterIP instead of NodePort, so the code-execution sandbox is reachable only inside the cluster via Service DNS (http://sandbox-<id>-svc.<ns>.svc.cluster.local) and is no longer bound on every node's interfaces - including the externally-reachable ones on GKE/EKS/AKS. Existing chart installs flip NodePort -> ClusterIP on upgrade. To preserve the old reachability (an external probe hitting the 30xxx port, or the Docker-Compose/hybrid path where the gateway is not in K8s), set provisioner.sandboxServiceType: NodePort (with provisioner.nodeHost if needed). The provisioner itself is unchanged (mode-aware since #4016). (#4190)
• skills: An active restrictive skill must explicitly list task in allowed-tools to delegate to a subagent. Read-only discovery infrastructure (tool_search and describe_skill) remains available, but cannot grant schema visibility or execution for a denied business tool. (#4098)
• memory: Pre-abstraction top-level memory.* DeerMem fields (storage_path, max_facts, debounce_seconds, model_name, token_counting, staleness_*, consolidation_*, ...) are auto-migrated into backend_config on load with a warning, so an upgrade does NOT silently revert customized settings to defaults (model_name -> backend_config.model.model). Move them under memory.backend_config in config.yaml to silence the warning. (#4122)
• memory: Added memory.mode (middleware | tool); tool mode registers memory tools (memory_search/add/update/delete) the model calls directly instead of passive per-turn summarization. manager_class resolution is now fail-fast (raises ValueError on an unknown backend instead of silently falling back). (#4023)
• middleware: Declarative layered middleware builder; ThreadData now runs before Uploads. (#3809)
• sandbox: The host->virtual output-masking regex now has a single owner, eliminating duplicated pattern compilation. (#4108)
• docs: AGENTS.md is now the source of truth for agent guidance, imported by CLAUDE.md via @AGENTS.md; module guides refreshed. (#3770)
• memory: The OpenViking memory backend now uses the official OpenViking adapter; the old trusted-mode auth_mode/account fields are rejected in favor of a credential-bound USER API key. (#4707)
• gateway: Threads created before the run-event journal have their checkpoint history backfilled as seed events before the first new run, so legacy conversations stay visible and correctly ordered after an upgrade. (#4590)
• agents: Subagent delegation is now routed by net benefit: the lead agent defaults to direct execution unless parallel latency, specialist capability, or context isolation clearly pays off. (#4384)

Fixed

• sandbox: Stop remote glob and grep from reporting "no matches" when their output was cut off. BoxLite, Tenki, E2B, and OpenSandbox cap the search's raw output and then filter it in Python (ignored directories such as node_modules, the pattern or glob scope), but they reported truncated only when max_results was reached. When the capped lines were all filtered out, a search with real matches past the cap came back empty and complete. The search now passes one line beyond its cap so a cut-off result is reported as truncated, and the glob and grep tools say an empty truncated result is incomplete instead of "No matches found". (#5427)
• sandbox: Stop host paths reaching the model when output joins them with :, as $PATH and $PYTHONPATH do. The matched path ran on through the rest of the list, so every later entry under the same root was left unmasked; extra masking passes recovered one entry each, which hid the leak for short lists. Masking now ends a matched path at :. A symlink inside a mount whose target lies outside every mount is now shown by its mount path instead of the target's host path in command output and glob results. (#5418)
• sandbox: Stop BoxLite grep from ignoring the directory part of glob. It compared only file names, so src/*.js matched every .js file in the tree. The glob now applies to the path relative to the search root, the same scope as glob() and the other providers. (#5419)
• models: Stop every Claude model after the first from losing its credential when the Claude Code OAuth token is handed off through CLAUDE_CODE_OAUTH_TOKEN_FILE_DESCRIPTOR. Every ClaudeChatModel instance loaded credentials again, but a descriptor can be drained only once, so the title, summarization, and subagent models — and every later run — had no credential and failed with TypeError: Could not resolve authentication method. The token is now read once per process and reused. (#5411)
• models: Stop the lead agent from failing to build whenever a model with supports_reasoning_effort: true also gets a reasoning_effort from its profile — at the top level, in when_thinking_enabled or when_thinking_disabled, or from the extra_body.thinking disable path. The regular lead-agent build forwards the requested effort even when unset, so the key reached the provider constructor twice and raised TypeError: got multiple values for keyword argument 'reasoning_effort'. The requested value now layers like per-agent model_settings: it replaces a top-level profile value, an unset request keeps that value, and the thinking-mode settings still decide the final one. Codex keeps its own level check. (#5403)
• runtime: Stop a keyed run retry from failing with 500 on the SQL run store. HTTP admissions do not pass a user_id; the SQL store stamps the request user on the row, but the process-local run record kept None. A retry with the same Idempotency-Key that reached another Gateway worker, or the same worker after the finished run was cleaned up, compared the two owners, took its own run for another user's, and raised. The same mismatch dropped HTTP runs from owner-scoped history reads and skipped the MCP background_tasks projection for them, so values events for these runs now include background_tasks. RunManager now resolves an omitted owner from the request user the way the SQL store does, so every store records the same owner. (#5401)
• runtime: Stop a cross-worker idempotent run reuse from permanently blocking the thread on the reusing worker. The reuse registered the hydrated store row as a local run record, but only the owning worker finalizes and cleans up its records, so the copy kept its admission-time pending/running status forever: every later reject admission for that thread on the worker returned 409 until a restart, its run reads kept reporting the stale status, and orphan reconciliation skipped the run if the owner crashed. A cancel sent to that worker also took the local-owner path and marked the owner's still-running row interrupted. The reusing worker now returns a detached store-only handle instead, so cancel follows the non-owner contract. (#5393)
• skills: Stop writing resolved secrets into extensions_config.json when a skill is toggled. The Gateway skill toggle and DeerFlowClient.update_skill loaded the file through ExtensionsConfig.from_file(), which replaces every $VAR value with the environment value, and wrote that model back — so a "$GITHUB_TOKEN" reference was persisted as the plaintext token and an unset variable was permanently replaced with "". DeerFlowClient.update_mcp_config did the same for every key other than mcpServers. These writers now edit the raw on-disk JSON and validate the candidate the way the runtime loads it, so placeholders and hand-written structure survive; the MCP router shares the same raw loader. Files rewritten by an earlier toggle keep their plaintext values: restore the $VAR references and rotate the exposed credentials. (#5357)
• gateway: Honor disable_clarification and github_token only for internally-authenticated callers, the way non_interactive already was. Both keys were forwarded from body.context regardless of the caller and were not scrubbed from the free-form body.config that the run config copies verbatim, so any session or PAT caller could set them. disable_clarification is the stronger of the two: ClarificationMiddleware answers every clarification — risk_confirmation included — with "proceed without asking", and SandboxMiddleware reads it as the same non-interactive signal as non_interactive. github_token reached runtime.context, where the bash tool exports it as GH_TOKEN/GITHUB_TOKEN, and a copy smuggled through body.config['configurable'] was persisted in the checkpoint store. The scheduler, IM channels, and the GitHub webhook channel authenticate over the internal request channel and are unaffected. (#5338)
• artifacts: Keep PUT /api/threads/{id}/artifacts/{path} confined to /mnt/user-data/outputs. The outputs-only guard was a string-prefix check on the raw path, so a percent-encoded .. (outputs/%2e%2e/uploads/x.txt) — which nginx forwards untouched and Starlette decodes — passed it, and the resolver only confines to user-data/, letting a caller overwrite a sibling upload or workspace file in their own thread. Dot segments are now collapsed before the prefix check, and the resolved host path is re-checked against the resolved outputs root so a symlink planted inside outputs/ cannot redirect the write either. The rule now lives in one shared helper that IM-channel attachment delivery uses as well, so the two copies cannot drift. (#5321)
• gateway: Stop persisting a caller-supplied deerflow_trace_id on the run record. body.metadata reaches both the live run config, which the run worker restamps, and the run record echoed verbatim by the runs API; only the first was covered, so a client could make the most durable surface of a run disagree with the X-Trace-Id and the log lines from the same request. The id is now stamped once at the trust boundary, config.context is closed off the same way, and a thread's own metadata is no longer seeded with the run-scoped id of whichever run created it. (#5119)
• gateway: Expose X-Trace-Id in Access-Control-Expose-Headers. It is not CORS-safelisted, so split-origin browser clients — the ones that cannot read the Gateway's logs either — could not read the correlation id they are meant to quote in a bug report. (#5119)
• gateway: Keep X-Trace-Id on unhandled-exception 500s. Starlette's ServerErrorMiddleware emits those through the raw send outside every user middleware, so the 500 for a server bug — the response most in need of correlation — was the only one shipped without the id. TraceMiddleware now sends its own 500 carrying the header before re-raising; the server's exception logging is untouched and mid-stream failures propagate unchanged. This fallback is emitted outside CORSMiddleware and stays CORS-opaque, so split-origin browser clients cannot read the id on this one response — same as the ServerErrorMiddleware 500 it replaces. (#5119)
• gateway: Strip a forged deerflow_trace_id from the persisted request echo. body.config is stored verbatim as runs.kwargs_json and served back by the runs API, so a forged id in config.metadata or config.context survived on that one surface while every other carried the real id. redact_config_secrets now drops the key from both containers, and build_run_config merges run metadata onto a copy so the server-stamped id can no longer be written through into the caller's request body. (#5119)
• artifacts: Keep explicit full-file loading scoped to the source thread, so a same-path artifact in another conversation keeps its 1 MiB preview. (#4634)
• sandbox: SandboxAuditMiddleware no longer blocks ordinary command substitution that only captures output. The rule now judges position instead of matching any $(: x=$(curl url), echo $(curl url), an argument, and a for word list all run normally, while a substitution in command position ($(curl url), after a |/&&/;, behind leading assignments or an env/nohup/time style wrapper, or as an eval/source argument) still blocks because it executes fetched content. An interpreter's code-string flag (bash -c, python -c, perl -e, node -p, php -r, and the <<< here-string) is treated as an execution context wherever it appears, so bash -c "$(curl url)" blocks; source <(curl url) and the backtick spelling of eval/source now block too, neither of which was detected before. An unquoted newline separates statements like ;, so echo hi followed by a new line starting $(curl url) blocks as well, while heredoc bodies are consumed as data — writing a file whose content happens to start a line with $(curl url) is not a command. Variable expansions whose name merely starts with a risky executable ($shell, $bashrc, $python_version) and lookalike binaries (shellcheck, shasum) are no longer false positives. (#4611, #4623)
• mcp: Isolate Settings > Tools enable/disable updates to one MCP server, so an unrelated disallowed stdio command no longer blocks every switch; allow disabling a disallowed target while still rejecting its re-enable, preserve the raw extensions config, honor the MCP-spec transport alias when enabling SSE/HTTP servers, surface backend validation details in the UI, and atomically replace the shared config for MCP, skill, and embedded-client updates so interrupted writes cannot leave it truncated. (#4574, #4577)
• runtime: Thread metadata now switches to running only after the run passes the startup barrier, so pending-cancelled runs no longer briefly project running; clients may observe the prior thread status during worker startup. (#4450)
• runtime: Re-check orphan candidates through an atomic, lease-aware takeover claim so a successful heartbeat after the scan keeps the run active and only one reconciler reports recovery. (#4424, #4434)
• skills: Apply allowed-tools only to slash-activated or actually loaded lead-agent skills, preventing passive enabled skills and evaluation fixtures from removing MCP, web, file, and delegation tools from every run. (#4095, #4098, #4192)
• models: Honor api_base on every BaseChatOpenAI subclass (VllmChatModel, MindIEChatModel, PatchedChatMiMo, PatchedChatStepFun, PatchedChatMiniMax), not just ChatOpenAI / PatchedChatOpenAI. Those five previously dropped the configured endpoint silently and then failed every request with an opaque unexpected keyword argument 'api_base'; the unknown-config-key warning was disabled for them as well. Both now gate on issubclass(BaseChatOpenAI). (#4146)
• agents: Coalesce SystemMessages before the LLM request; ensure a visible response after tool runs; avoid a default LLM title call before stream end; reserve ellipsis room so the local title respects max_chars; and snap the tool-output tail forward so fallback truncation respects max_chars. (#3711, #4033, #3885, #4052, #4017)
• agents: Skip dateless reminders in the dynamic-context date scan; load SOUL.md from agent dirs without config.yaml; require config.yaml in update_agent's legacy-agent guard; and refuse empty SOUL.md updates. (#3685, #4136, #4166, #4219)
• middleware: Window the loop-detection tool-frequency counter so long runs no longer false-trip; prevent the title middleware from streaming tokens; fix positional fallback consuming an unrelated todo when the same-content list is exhausted; acquire the token-budget lock across _apply, before_agent, _clear_run_state, and _drain_pending_warnings; drop orphan ToolMessages so strict providers don't 400; sanitize invalid tool-call arguments; and recover from empty tool-call names and malformed tool-call ids in dangling repair. (#4072, #3566, #3709, #3714, #4080, #4193, #4008, #4246)
• subagents: Inherit LoopDetectionMiddleware and summarization middleware so tool loops break and steps are captured; surface the turn-budget cap as MAX_TURNS_REACHED with a partial result; unify guardrail caps on the additive stop_reason + token_budget; inject durable context before compaction; preserve the parent checkpoint namespace; prohibit the task tool in the general-purpose system prompt; re-buffer subagent events on flush failure to avoid losing steps; and fix the lost loop_capped stop reason when a subagent's run_id is None. (#3931, #4009, #3949, #3980, #4040, #4215, #4161, #4082, #4059)
• memory: Harden against null/empty edge cases - skip whitespace-only facts; coerce null confidence / source.confidence in updates, searches, and the three remaining raw reads; treat explicit null backend_config values as omitted; fix KeyError / UnboundLocalError when a fact has no id or the facts list is empty; stop the busy-spin in the debounced update queue; and flush the memory queue on graceful shutdown to prevent loss. (#3719, #4074, #4076, #4034, #4217, #3993, #3992, #4073, #4181)
• runs: Close multi-worker ownership gaps in run atomicity; fail-stop local execution when lease renewal cannot be confirmed before its deadline and fence late completion writes after peer takeover; degrade cancel to lease takeover for multi-worker; keep create_thread idempotent when the insert loses a race; read stop_reason from runtime context; and persist run duration in checkpoints for history reads. (#4003, #4064, #4414, #3800, #4188, #4118, #4431)
• runtime: Serialize SQLite event-store writes to prevent per-thread sequence collisions; skip hidden human messages in the journal; and drop the silent delta-discard in _merge_stream_text. (#4077, #3698, #4085)
• gateway: Attach thread-message feedback by real event_type; offload blocking filesystem IO in artifact serving, gateway uploads, and the Discord channel; limit the uploaded-file context manifest; and live-tail malformed Redis reconnect ids. (#3651, #3551, #3935, #3927, #3917, #4012)
• uploads: Claim the converted-Markdown companion filename before writing it, so two convertible uploads sharing a stem (or a convertible plus a same-stem .md upload) no longer silently clobber each other within one request. When uploads.auto_convert_documents is on, the companion .md now gets a unique name (e.g. a_1.md); POST /threads/{id}/uploads and DeerFlowClient.upload_files both report the actual name in markdown_file. (#4288)
• config: Coerce null object config sections to their defaults; honor the unified database configuration in the store and sync checkpointer; and have legacy DB backfill create missing Index objects on existing tables. (#3573, #3904, #3994, #4090)
• models: Apply the stream_chunk_timeout default to all BaseChatOpenAI subclasses; and normalize api_base -> base_url for ChatOpenAI with a warning on unknown config keys. (#4102, #3790)
• mcp: Isolate tool-discovery failures per server; synchronize the session-pool singleton lifecycle; invalidate the tools cache on config content
  • path (not just newer mtime); validate MCP tool names at load so deferred prompts stay inert; and route tools by source server, not name prefix. (#3772, #3797, #4124, #4154, #3812)
• skills: Activate a slash skill once per run, not per model call; close the skill-install security-scan coverage gap; recognize fully deleted skill packages in review CI and remaining requests / httpx methods as network sinks in SkillScan; reuse the resolved app config in the no-arg skills prompt section; and reload mounted skills without restarting the Gateway. (#4103, #3924, #4169, #4130, #4160, #4264)
• sandbox: Guard the reverse path-translation and output-masking regexes with segment boundaries; handle one-sided line ranges and empty files in read_file / str_replace; align the AIO bash working directory; use os.sep in the reverse-resolve containment check on Windows; normalize Windows backslash paths in bash commands; stop glob / grep / ls from surfacing disabled skills' files; and allow valid heredoc commands in the sandbox audit. (#4035, #4053, #4078, #4079, #4051, #4058, #3869, #4096, #3786)
• sandbox: Synchronize the sandbox provider singleton lifecycle (with concurrency regression tests) and keep k8s calls off the event loop in the provisioner. (#3730, #3941)
• sandbox: Align sandbox artifact mounts with the channel user; fix local-dev (make dev) on non-root / NFS hosts; reap macOS nginx processes on stop; and fix production Postgres UV-extras detection in Docker. (#3729, #3590, #3828, #3897)
• channels: Validate the channel provider before resolving its config; dedupe GitHub webhook redeliveries and drop redundant GitHub review-comment webhook fan-out; scope the slash-skill whitelist check to the run's owner; batch Feishu file messages into one thread and dispatch Feishu group commands prefixed with a bot @mention; accept leading @mentions before /connect bind codes and don't treat a bare "connect" as a bind command; stop Feishu from creating thread topics and throttle card updates; let the UI runtime channel config win over config.yaml; fix require_mention gating on whitespace-only bot_login / mention_login; guard null quote fields in WeCom; and key inbound dedupe on chat-scoped workspaces so Telegram, Feishu, WeChat and DingTalk redeliveries stop re-running the agent on a default (unbound) configuration, releasing the dedupe key on transient failures so a redelivery can still recover. (#4100, #4104, #4131, #4129, #3753, #4229, #4222, #4251, #3810, #3674, #4055, #4069, #4287)
• frontend: Preserve messages and durable context across summarization; preserve artifacts and stabilize artifact paths during streaming; resolve relative artifact image paths; retain presented artifacts in the header dropdown; keep orphan tool messages visible; show assistant text during tool steps; reset new chat on client-side navigation; prevent stream cancellation on concurrent submit; fix stale-run reconnect and cancel handling; fix chat math rendering, single-tilde markdown, double reasoning rendering, UTF-16 markdown binary classification, and <memory> tags in Streamdown; make recent-chat rows fully clickable; validate attachment limits before upload and fix uploaded-file metadata in message copy; fix mobile workspace and accessibility blockers, the card tool-message bug, and side-chat toolbar / panel-button behavior; block unresolved suggestion-template placeholders; refresh notification permissions; show the branch action only for completed turns; enable regenerate in custom agent chats; and generate a fallback title for interrupted first-turn runs. (#3826, #3791, #4094, #4038, #3854, #3880, #4114, #3673, #3878, #3908, #3557, #4245, #3870, #3966, #4209, #3733, #3900, #3944, #3740, #3976, #3959, #3961, #3764, #3768, #4147, #3967, #3874, #3644)
• tui: Interrupt an active run before /quit exits. (#4235)
• harness: Don't flag the outline as truncated at exactly MAX_OUTLINE_ENTRIES headings. (#3856)
• tracing: Attach Langfuse trace metadata to the goal evaluator. (#4202)
• context: Resolve the context-compress bug. (#4065)
• threaddata: Fix AttributeError when runtime.context is None. (#3989)
• goal: Stop continuation_count double-bump during stand-down. (#4199)
• circuit-breaker: Stop wedging after a non-retriable half-open probe. (#3991)
• github: Match allow_authors logins case-insensitively. (#4218)
• community: image_search now returns the full-resolution image URL. (#3990)
• skills: Offload blocking filesystem IO in the skill-history endpoint. (#3563)
• skills: Don't treat a lazily evaluated PEP 695 type alias as a network sink in SkillScan. (#4315)
• tracing: Resolve the Langfuse trace user from runtime context. (#3794)
• guardrails: Propagate internal owner attribution into the guardrail context. (#3839)
• subagents: Clamp the subagent limit consistently with MIN_SUBAGENT_LIMIT. (#4081)
• subagents: Load user-scoped skills. (#4356)
• mcp: Per-server fail-soft OAuth priming, and persist rotated refresh tokens. (#4084)
• mcp: Ignore malformed path-like text. (#4456)
• auth: Resolve email accounts case-insensitively. (#4101)
• auth: Recover from setup-status timeouts. (#4371)
• scheduler: Close a dispatch race that could launch two runs for one scheduled task. (#4105)
• channels: Buffer and drain GitHub comments queued during a busy run. (#4133)
• channels: Escape Slack reserved characters before mrkdwn conversion. (#4197)
• channels: Check response.success() on Feishu card/reaction SDK calls. (#4234)
• channels: Drop inbound DingTalk messages that carry no conversation identity. (#4316)
• channels: Receive inbound Telegram attachments. (#4392)
• memory: Consolidated facts inherit expected_valid_days from their sources. (#4225)
• config: Sync _memory_config with AppConfig auto-reload. (#4208)
• postgres: Harden the async engine with pool_recycle and command_timeout to stop stale-connection 504s. (#4230)
• harness: Add a timeout to invoke_acp_agent to prevent indefinite hangs. (#4238)
• community: Surface the target-page error status in web_fetch (Browserless). (#4239)
• sandbox: Widen the BoxLite/AIO tenant hash and verify identity on reclaim. (#4171)
• sandbox: Make an empty old_str a no-op in str_replace on any file. (#4256)
• sandbox: Serialize E2B release transitions. (#4355)
• sandbox: Bound E2B output-synchronization resources. (#4364)
• sandbox: Unwrap Overwrite-wrapped sandbox state in after_agent. (#4381)
• sandbox: Bypass proxies for local AIO traffic. (#4444)
• models: Surface length-capped model responses instead of dropping them. (#4309)
• streaming: Keep large file generation responsive. (#4354)
• streaming: Expose custom events to astream_events. (#4403)
• streaming: Signal replay history gaps. (#4426)
• summarization: Summarize with the run model and fall back on summary-provider failure. (#4361)
• runtime: Remove transient image context after model calls. (#4267)
• runtime: Stop subgraph stream frames from impersonating root frames. (#4407)
• runtime: Reject unsupported run options and stream modes. (#4430)
• runtime: Serialize checkpoint writes with active runs, linearize delta-mode checkpoint resume, and accept the SDK's default stream_resumable=false to avoid resume races. (#4437, #4460, #4468)
• checkpoint: Unwrap Overwrite first writes into empty channels. (#4383)
• nginx: Allow long chat prompts through /api/langgraph/ without a raw 500. (#4277)
• gateway: Prefer X-Trace-Id over metadata.deerflow_trace_id when the header is set. (#4283)
• gateway: Seed branch run-events so inherited history survives forking. (#4385)
• gateway: Scope branch-history seed run ids per inherited turn. (#4459)
• frontend: Harden artifact and markdown rendering. (#4117)
• frontend: Classify a symlink replacing a file distinctly from deleted in workspace-change review. (#4170)
• frontend: Offload blocking filesystem IO in the workspace-change text-cache lifecycle. (#4268)
• frontend: Encode artifact URL path segments. (#4278)
• frontend: Clarify run-duration display. (#4348)
• frontend: Preserve regenerate state in branched threads. (#4358)
• frontend: Default the reasoning-effort label to Medium when unset. (#4373)
• frontend: Strip and parse the <current_uploads> upload-context tag. (#4402)
• frontend: Keep leading orphan tool messages visible. (#4408)
• frontend: Keep completed subtask cards stable after reload. (#4432)
• frontend: Apply message-image maxWidth via inline style. (#4446)
• frontend: Restore resizing for the artifacts and sidecar panels. (#4469)
• frontend: Allow dev-server access from non-localhost hosts. (#4471)
• safety: Backfill empty content-filter responses so they don't poison the thread. (#4394)
• tools: Exclude injected runtime from the list_uploaded_files schema. (#4376)
• mcp: Bound MCP server bring-up — tool discovery (subprocess spawn + initialize + tools/list) and persistent stdio session initialization — with a new per-server session_init_timeout (default 60s, null disables), so a hung stdio server can no longer block agent construction, or the whole Gateway event loop, indefinitely. tool_call_timeout still bounds individual stdio tool calls. (#4657)
• runtime: Tool-output budget externalization no longer trips run delivery verification. The default .tool-results storage dir (and any custom tool_output.storage_subdir) is excluded from workspace-change snapshots and produced-artifact detection, so a run that only externalized oversized tool outputs succeeds instead of failing as an error. (#4657)
• frontend: Hide stale follow-up suggestion chips while a turn is still streaming. (#3396)
• frontend: Fix streaming render glitches: stop the word animation from replaying, keep step text stable, preserve message order during long runs, and keep reasoning above the answer. (#4266, #4510, #4513, #4578)
• frontend: Encode thread IDs in chat routes so IDs with special characters no longer break navigation. (#4302)
• frontend: Render citation links from React children. (#4486)
• frontend: Localize conversation export failure messages. (#4493)
• frontend: Sync side panel state when a drag collapses the panel. (#4556)
• frontend: Render one workspace-change card per run instead of duplicates. (#4559)
• frontend: Refresh the active artifact's content when it changes. (#4584)
• gateway: Reject non-positive read limits in API requests. (#4284)
• gateway: Handle a null config.configurable when resolving the thread id instead of failing. (#4301)
• gateway: Unify thread id validation across API routes. (#4589)
• gateway: Merge concurrent thread metadata updates instead of letting them silently overwrite each other's changes. (#4489)
• gateway: Expose the run metadata response header to cross-origin clients, so a split-origin frontend learns new run ids instead of staying stuck on the new-thread placeholder route until reload. (#4535)
• gateway: Replay edit and rerun from a settled checkpoint so the edited prompt actually runs (previously a first turn's edit replayed the original prompt and vanished after reload), and keep a manual rename through the rerun. (#4534, #4539)
• runtime: Cancel a run from any live gateway worker, not only the one that owns it, so the stop button no longer depends on request routing. (#4500)
• runtime: Close a replacement run when interrupt or rollback admission is cancelled mid-flight, instead of stranding an unseen active run on the thread. (#4472)
• runtime: Regenerating a response now preserves the thread's current title and supports the latest interrupted response whose partial message never reached a checkpoint. (#4480, #4524)
• agents: Classify web_fetch error pages such as 404s as errors rather than successful evidence, so retries and stagnation guards can react. (#4314)
• agents: Handle XML-to-dict option shapes when normalizing clarification choices. (#4527)
• subagents: Run delegated subagents with isolated callbacks and lazy skill activation, fixing cross-event-loop failures and passive skills stripping baseline tools like write_file. (#4497)
• sandbox: Handle overwrite-wrapped state when ensuring the sandbox is initialized. (#4429)
• sandbox: Reconcile E2B sandboxes safely: pick the first healthy candidate, adopt the canonical instance per user and thread, defer a peer's live duplicates, and reap orphans after a grace window. (#4443)
• sandbox: Claim ownership before destroying a sandbox that failed its readiness check, so a peer gateway can no longer adopt the not-yet-ready sandbox and kill a live turn. (#4505)
• sandbox: Allow grep to search a single file. (#4512)
• sandbox: Enforce the E2B capacity limit deployment-wide when sandbox ownership uses Redis, so multiple gateways cannot create past it. (#4575)
• skills: Activate managed integration skills from the managed integrations root on slash invocation. (#4570)
• skills: Offload blocking filesystem IO when updating a skill and serialize concurrent writes. (#3565)
• mcp: Ignore oversized path-like text. (#4582)
• memory: Harden long-term memory: reject duplicate facts inside the create critical section, truncate injected mem0 context on entry boundaries, and keep task-scoped instructions such as "inspect only" out of long-term memory. (#4599, #4600, #4604)
• scheduler: Keep a successfully launched scheduled run's slot and run id when post-launch bookkeeping fails, preventing a later dispatch from launching a duplicate run. (#4504)
• config: Treat a deleted extensions config file as absent instead of raising, so tool and skill config resolution keeps working. (#4275)
• config: Normalize the postgres:// short scheme for the async ORM engine. (#4293)
• console: Disable cost reporting when model pricing mixes currencies instead of reporting a meaningless cross-currency total. (#4564)
• browserless: Accept the timeout config key and harden its coercion. (#4519)
• docker: Send Connection: upgrade only when the browser requests it, fixing login-page refresh loops when the Docker dev stack is accessed via a remote host. (#4250)
• runtime: Group JSONL batch event writes by run, so a batch covering several runs no longer lands all events in the first run's file and makes later runs unreadable through per-run APIs. (#4938)
• runtime: Restore standalone LangGraph Studio compatibility: the graph entrypoint and file-based app load again, the Studio identity can discover system assistants, and the documented langgraph dev workflow works. (#4760, #4838)
• gateway: Stamp turn_duration on a run's last AI message only in /messages/page, so multi-step turns no longer repeat the same run lifetime as thinking latency on every intermediate message. (#4755)
• gateway: Preserve exact history attribution beyond the event page limit, so older AI messages on long-lived threads are no longer credited to a later turn's run and duration. (#4953)
• gateway: Reject MCP task cancellation with HTTP 503 when the task worker is stopped, instead of acknowledging a cancellation that would never run. (#4963)
• middleware: Correct four context-handling defects: fallback dynamic-context injection targets the latest user message instead of resurrecting an old prompt as the current turn; bare string blocks in list-form user content are sanitized like all other user text; duplicate placeholders are no longer emitted for the same invalid tool call; and summarization no longer compresses away the current request's user message while leaving the previous turn's behind. (#4667, #4668, #4693, #4882)
• middleware: Restore the system-prompt injection that teaches the model about the write_todos tool, which the todo middleware's model-call override had silently dropped. (#4735)
• agents: Make SQL agent-store signatures content-sensitive, so an agent update that reuses its previous timestamp no longer leaves the GitHub agent registry serving stale webhook routing. (#4709)
• tools: Resolve presented files with the runtime user, so present_files no longer rejects valid artifacts as outside the outputs directory when the request user context is unavailable. (#4677)
• tools: Retain a strong reference to deferred subagent cleanup tasks, so garbage collection can no longer destroy a pending cleanup and leak cancelled subagent records, locks, and memory. (#4928)
• subagents: Give every background subagent run a server-side execution ID, so concurrent runs that reuse a provider tool-call ID can no longer overwrite, poll, or cancel each other's state. (#4758)
• harness: Offload ACP workspace creation and MCP config loading from the event loop, so invoking an ACP agent no longer raises blocking-IO errors or stalls other async work. (#4965)
• mcp: Reject non-finite poll_after_seconds values on task snapshots when they arrive, so a bad polling interval no longer crashes scheduling and persistence after a successful poll. (#4750)
• mcp: Keep the configured grant_type authoritative over extra_token_params during OAuth token exchange, so extra parameters can no longer silently switch the configured flow and be rejected by the token endpoint. (#4860)
• mcp: Exclude the internal stdio MCP temp directory (.mcp/tmp) from workspace changes, so MCP temporary and debug files no longer appear alongside user deliverables or crowd real changes out of the file budget. (#4898)
• mcp: Cancel the remote task when a durable task submission is cancelled mid-flight, so an interrupted submission no longer leaves a remote task running with no record to poll or stop. (#4933)
• sandbox: Accept the documented E2B reconciliation config fields, so valid E2B configuration no longer produces misleading startup warnings. (#4772)
• sandbox: Bound E2B mount upload resource use per file, per mount, and across the whole upload pass (shared size and file budgets plus a wall-clock deadline), so large mounts can no longer spike Gateway memory or hold sandbox capacity indefinitely. (#4812, #4842)
• sandbox: Preserve trailing whitespace in E2B-synced filenames and tolerate out-of-range remote mtimes, so output sync no longer re-downloads files repeatedly or aborts mid-sync. (#4861)
• sandbox: Reject non-finite Redis lease-timing values in sandbox ownership config at parse time instead of crashing with an OverflowError during startup. (#4960)
• sandbox: Resolve structured skill reads through the sandbox provider's path mappings, so read_file opens legacy and per-user custom skills under the same enabled-state projection as ls and shell execution. (#4792)
• skills: Parse Responses API content blocks in the moderation scanner, so valid skill-management decisions returned as content blocks are no longer rejected as unparseable. (#4936)
• memory: Reject non-positive and non-finite timeout and character-limit settings in the Honcho and Mem0 backends at config parse time, so a bad value fails fast instead of silently truncating stored text or crashing on the first HTTP call. (#4783, #4823)
• memory: Scope custom-agent bootstrap facts to the selected agent's bucket, so facts learned during setup no longer leak into the default bucket and influence ordinary lead-agent conversations. (#4804)
• artifacts: Support atomic saves on Windows, and serve a SHA-256 ETag on artifact reads so inline preview and editing work on non-secure contexts such as plain-HTTP LAN origins where crypto.subtle is unavailable. (#4629, #4865)
• frontend: Keep conversation order stable around long runs: the submitted user message no longer renders twice or sinks below its own processing steps, and after a mid-run page reload a turn's steps can no longer appear above the user message that started the run. (#4620, #4660, #4834)
• frontend: Stop matching <header> as <head> when injecting the base href into HTML artifact previews, so relative assets in report fragments that begin with <header> now load in the sandboxed preview iframe. (#4625)
• frontend: Open landing-page case studies on a public read-only /showcase/ route so anonymous visitors are no longer redirected to login. (#4635)
• frontend: Sort the chats page by pinned state, so pinned threads no longer render below unpinned ones. (#4643)
• frontend: Keep <think> pairs written inside markdown inline code in the rendered content instead of hollowing them out into the Reasoning panel, and restore the copy button for turns that contain only reasoning. (#4647)
• frontend: Surface model-loading failures with a workspace error banner and retry action instead of a silently empty model list. (#4840, #5021)
• frontend: Preserve copy and other actions on completed assistant messages while a later turn is still streaming. (#4844)
• frontend: Keep the browser live stream connected after a successful reconnect instead of tearing down the new socket and immediately creating another. (#4951)
• frontend: Reuse the shared clipboard fallback when copying the Lark authorization link, so the copy action works in browsers without the Clipboard API. (#4767)
• frontend: Use consistent "DeerFlow" casing in the composer disclaimer and fix the "What's New" heading on the landing page. (#4970)
• channels: Bound inbound intake with a fixed worker pool and bounded admission queues, and await real cross-thread tasks on shutdown, so message floods are rejected promptly instead of accumulating and channel shutdown no longer tears down transports with work still in flight. (#4800, #4816)
• channels: Offload outbound attachment file IO for Feishu, Telegram, and WeCom to worker threads, so sending a large artifact no longer stalls the Gateway event loop. (#4633)
• channels: Run Telegram connection-identity lookups on the Gateway event loop, so inbound messages and commands no longer crash with a cross-loop error when channel connections are enabled. (#4815)
• feishu: Keep file receiving off the event loop and preserve every inbound attachment: duplicate provider filenames no longer overwrite each other, writes can no longer be redirected outside the thread bucket, and a failed attachment no longer blocks the rest of the message. (#4627, #4903)
• dingtalk: Strip leading @bot mentions before command classification, so slash commands like /new sent in group chats are recognized instead of treated as plain chat. (#4724)
• discord: Refuse to start typing-indicator loops after the channel stops, so shutdown no longer leaves an infinite typing task sending events in the background. (#4752)
• wecom: Serialize WebSocket start/stop transitions and await the SDK's real receive-task shutdown, so stopping the WeCom channel can no longer return before the socket closes or clear a newer connection's state. (#4762)
• buzz: Drop replayed events across reconnects using a persistent seen-id store, so the agent no longer re-answers the last message in a channel after a relay or Gateway restart. (#4888)
• lark: Keep the CLI lock directory writable inside sandboxes while the credential-bearing config root stays read-only, restoring Lark API commands that previously failed with a read-only filesystem error. (#4701)
• scheduler: Enforce the global max_concurrent_runs budget for manual triggers too, returning HTTP 409 when the cap is reached instead of letting manual launches exceed it. (#4769)
• scheduler: Coerce serialized task timestamps on read, so scheduled-task operations no longer fail when string-form timestamp values reach the database layer. (#4785)
• scheduler: Support safe multi-instance scheduler recovery: startup no longer treats live runs owned by peer Gateway instances as local leftovers, so a restarting instance cannot interrupt a live run or trigger a duplicate execution; multi-instance mode is opt-in via scheduler.multi_instance. (#4713)
• scheduler: Enqueue busy scheduled task runs instead of skipping them: occurrences that hit a busy reused thread now wait in a durable queue (bounded by scheduler.queue_timeout_seconds) and survive Gateway restarts, and the UI explains the queueing behavior. (#4918)
• cli: Add --recursion-limit to headless --print, --json, and --cli runs, so long-running agent loops are no longer stuck at the default recursion limit of 100. (#4615)
• dev: Exclude backend runtime state from the Uvicorn reload watcher in the backend make dev launcher, so an agent task writing files under the runtime tree can no longer restart the Gateway and reset concurrent users' requests. (#4759)
• dev: Resolve diagnostic script paths from the script's own location, so root diagnostic commands work when invoked from any working directory. (#4736)
• docker: Harden local and container startup: make up waits for the Gateway health probe before declaring the stack ready, Docker startup no longer aborts when .env is missing, the Gateway can write extensions_config.json in production, runtime data stays out of the image build context, log commands resolve the checkout root correctly, and the default loopback origins are allowed so the dev setup page can hydrate. (#4658, #4806, #4852, #4853, #4956, #4959)
• gateway: Stamp the server-authoritative feed position onto persisted messages, so an early user message no longer vanishes or jumps into the middle of the step stream once history exceeds one page and context compaction has fired. (#4696)
• lark: Preserve the new app secret during managed credential switches by clearing the previous app's OAuth data before the replacement is written, so the subsequent browser authorization no longer resolves an empty client_secret. (#4820)
• messages: Drop legacy <uploaded_files> tag handling: the backend treats the pre-#4174 spelling as ordinary content and strips only <current_uploads>, while the frontend keeps stripping the legacy tag so old threads still render cleanly. (#4826)
• skills: Reject a blank SKILL.md description at the write gate, matching what the loader already requires, so editing a custom skill with an empty description no longer writes a file the loader then rejects - which destroyed the skill on disk. (#4867)
• sandbox: Make the model-facing description argument optional (empty by default) across bash, ls, glob, grep, read_file, write_file, str_replace, and task, so providers that omit it are no longer rejected before execution. (#4878)
• sandbox: Bound Windows command execution: host commands run in a new process group killed via taskkill /T /F on timeout so a descendant cannot hold the call open, and output flows through the existing bounded 10 MiB capture. (#4946)
• sandbox: Scope the Windows MSYS path-conversion exclusion to safe virtual path prefixes instead of disabling conversion globally, so host-native CLI launchers that need normal conversion work again. (#5003)
• skills: Rebuild per-user skill storage after an app-config hot reload, so it no longer stays bound to paths from the previous config instance. (#4972)
• skills: Tokenize portable allowed-tools scalars with parenthesis awareness, so Bash(tvly *)-style entries stay intact, unmatched parentheses are rejected instead of silently fragmenting, and argument-scoped entries remain literal rather than broadening access. (#4984)
• agents: Normalize ToolMessages returned inside Command results, so error payloads no longer earn a default success receipt and tool-progress tracking sees them. (#4977)
• mcp: Tear down the in-flight session owner when get_session is cancelled during eviction, so a cancelled caller no longer leaks the owner task or parks past its timeout. (#5008)
• mcp: Reconnect ordinary stdio tools after a transport disconnect: the failed pooled session is evicted (only if still registered), the original error surfaces without automatic replay, and a later retry starts a fresh subprocess. (#5018)
• mcp: Preserve pooled stdio sessions after protocol timeouts during durable MCP task polling - a 408 is not a disconnect - so task state survives and the next poll no longer reports task_not_found. (#5027)
• mcp: Reject credentials that cannot travel as HTTP header values (trailing newline or whitespace, non-ASCII) at the config boundary, so the transport's exception - which echoes the full value - can no longer leak a secret into model context, checkpoints, and traces. (#5066)
• subagents: Clean up the background-task entry when the poller exits unexpectedly and drop a PENDING registry entry when submission fails, so a failed or crashed poll no longer leaks the entry or leaves the subagent running unattended. (#5069)
• subagents: Stop the zombie PENDING registry entry on the submit-failure path, and derive the capacity snapshot's queued count from the waiters' length instead of iterating a deque other threads mutate. (#5086)
• channels: Synchronize ChannelStore reads with mutations, so get_thread_id()/list_entries() can no longer raise dictionary changed size during iteration. (#5083)
• discord: Retain strong references to ack-reaction tasks and drain them on shutdown, so a GC pass can no longer silently drop a reaction or pin the channel across restart cycles. (#5049)
• buzz: Move seen-event persistence off the event loop with coalesced atomic writes, preserving dirty generations when events arrive mid-write and awaiting the final flush on shutdown. (#5103)
• streaming: Stop an IndexError in MemoryStreamBridge._make_gap when a subscriber reconnects to an empty or drained stream with an expired cursor. (#5047)
• uploads: Keep deduplicated filenames within the 255-byte limit by truncating the stem on a UTF-8 code-point boundary, so two max-length files that differ only by a dedupe suffix upload successfully instead of failing the whole batch. (#5059)
• frontend: Format structured upload error details (FastAPI validation issues, objects, arrays) instead of showing [object Object]. (#5071)
• frontend: Keep a renamed thread's title in sync across the active chat header, document title, search results, and metadata caches without a reload. (#5045)
• frontend: Truncate selected model names to the selector button width, so long model names ellipsize in the composer and sidecar instead of overflowing. (#5050)
• frontend: Truncate long subtask card titles to one line with a tooltip, so a delegation whose model omitted description (falling back to the full prompt) no longer overflows the chat layout. (#5136)
• dev: Default the frontend dev server to Webpack on all platforms (DEER_FLOW_DEV_BUNDLER=turbo opts back into Turbopack), avoiding Turbopack's macOS PostCSS worker leak and its Windows runtime panics. (#5036, #5133)
• scripts: Run repo shell scripts through an explicit interpreter (bash scripts/...), so a lost executable bit - zip/tarball downloads, core.fileMode=false, non-POSIX filesystems - no longer breaks make docker-start and friends with Permission denied. (#5031)
• deps: Depend on the renamed tenki package instead of the PyPI-removed tenki-sandbox (same tenki_sandbox import), so clean checkouts can resolve dependencies again on make dev/uv sync. (#5087)
• memory: Memory reads configured to stop the turn now raise a backend-neutral MemoryReadError that prompt assembly preserves instead of swallowing: strict OpenViking (read: raise), Mem0, and Honcho reads propagate, OpenViking scope-resolution failures follow the configured read policy, and the 5-second injection deadline honors the same policy (fail-open continues without the context; strict raises with the timeout as its cause). (#4726)
• memory: Buffered memory extraction is cancelled when a custom agent is deleted or cleared, so pending debounce timers can no longer resurrect the deleted per-agent memory scope or overwrite a fresh clear with a stale pending update. (#5123)
• agents: Conversation titles are generated from the user's original message content when upload (or other) context wrappers have been injected into the text, so titles no longer quote server-injected <current_uploads> context; attachment-only messages keep the New Conversation fallback. (#4729)
• agents: Fraction summarization triggers resolve against the model's declared context_window (now translated into the LangChain profile), and an unresolvable fraction clause degrades to a never-firing trigger with a warning instead of crashing the whole agent build; percent-style and non-finite trigger values are rejected at config load. (#4901)
• agents: Custom-agent storage falls back to file-backed storage only when search mode cannot resolve the main application config — invalid configuration and missing config paths surface instead of silently switching storage backends — and store IO moves off the event loop in async agent routes. (#4952)
• middleware: Loop-detection hard stops win across a whole tool-call batch: selecting a soft warning no longer ends inspection, so a later call in the same response that crosses an operator-configured hard limit is rejected instead of riding along with the earlier warning. (#5245)
• sandbox: list_dir and glob in the five remote sandbox providers (E2B, OpenSandbox, AIO, Tenki, BoxLite) return filenames verbatim instead of stripping whitespace, so files whose names begin or end with spaces are no longer listed under paths that then do not exist. (#4980)
• sandbox: Concurrent subagents sharing one thread sandbox run under process-local execution leases with task-scoped AIO shell sessions, so one sibling finishing can no longer release the shared sandbox underneath the others or corrupt the implicit persistent session; a healthy replacement session is promoted after corruption instead of returning to it. (#5134)
• sandbox: The Bash tool guides the agent to detect its execution environment with evidence (uname -s, sw_vers, uname -a) instead of model assumptions, and rejected host paths direct it to command-only probes or allowed virtual paths rather than repeating the blocked command. (#5111)
• sandbox: The Docker AIO compatibility capability allowlist gains FOWNER, so AIO images whose startup chmods /run/user/1000 (e.g. 1.11.0) start again under the hardened default capabilities while no-new-privileges stays on. (#5163)
• sandbox: File appends no longer destroy existing content when the pre-read fails: E2B append treats only missing-file errors as an empty file and re-raises anything else instead of overwriting the file with just the appended tail, and AIO appends use the server's native append mode without a pre-read at all. (#5261, #5278)
• skills: Skill markdown is read explicitly as UTF-8, so localized skills validate on Windows hosts whose default code page is not UTF-8 instead of failing with UnicodeDecodeError. (#4995)
• mcp: The MCP tools cache re-initializes after a runtime config change: a module-level asyncio.Lock bound to a closed event loop and an unsynchronized initialized flag had left every post-update call failing with Lock is bound to a different event loop or racing across worker threads. (#5062)
• mcp: Sync-wrapped MCP tools keep LangGraph ToolRuntime injection (the annotation-less sync wrapper is now functools.wraps-transparent), so per-user scope resolution and durable task submission no longer run with runtime=None — which had routed completion-notification runs under the default lead agent instead of the thread's custom agent. (#5164)
• auth: A duplicate OAuth identity no longer reports "Email already registered" — the two integrity violations are distinguished — and the partial OAuth-identity index now declares postgresql_where so Postgres builds the intended partial index instead of a full one. (#5026)
• frontend: The mobile sidebar trigger stays clickable on regular and custom-agent welcome pages, where multi-line (longer localized) welcome text in a same-z-index overlay could cover the header's tappable area. (#5149)
• subagents: SubagentResult lifecycle timestamps are UTC-aware on every writer, matching the repo-wide convention instead of stamping local wall-clock time on non-UTC hosts. (#5153)
• browser: Background live-frame scheduler tasks retain strong references, so garbage collection can no longer silently stop the Browser Live view from refreshing by stranding a pending guard that was cleared only in a collected task's finally. (#5155)
• runtime: Duplicate on_llm_end callbacks for the same LangChain run id persist one durable llm.ai.response event (replayed usage merged by generation position, first callback canonical), so append-only message APIs no longer return duplicate responses from providers that re-fire the callback with usage populated. (#5187)
• runtime: Terminal finalization completes after a cancellation raised inside the completion hook or task-stop fan-out, so extension observers run and the stream END marker is published before the interruption is re-raised; the finalization tail stays interruptible. (#5191)
• models: A cancelled LLM call releases its owned circuit-breaker recovery probe across provider execution, concurrency admission, and backoff, so subsequent calls no longer see CircuitBreakerOpen after a cancellation; probe ownership is fenced by a per-call token so cancelling an older call cannot release another call's probe. (#5197)
• runtime: The embedded DeerFlowClient keys its graph cache by effective user in every authorization mode and materializes the same user in runtime context, so sequential reuse for different users can no longer serve a graph assembled with another user's prompt and workspace state. (#5206)
• runtime: Cancelled workspace-change snapshot captures drain their already-running scan and clean up the per-run text cache instead of leaking deerflow-workspace-changes-* directories, while metadata-only captures propagate cancellation promptly without waiting on the scan. (#5232, #5234)
• persistence: Gateway startup tolerates a database already migrated to the explicitly-reviewed newer revision (0019_thread_incarnations), so rolling back to this image after a newer deployment stays possible; other unknown revisions, an empty version table, and multiple version rows still fail closed. (#5219)
• community: Tavily Extract results without a title fall back to the result or requested URL as the display heading instead of raising KeyError and discarding usable page content. (#5280)
• uploads: Fenced code blocks are excluded from uploaded-document outlines, so code comments and bold examples inside fences no longer crowd real sections out of the 50-entry heading budget. (#5281)
• subagents: After compaction, the delegation ledger distinguishes execution completion from task acceptance and preserves bounded examples of a completed subagent's unmet and unverified acceptance criteria, so the lead agent repairs the remaining gaps instead of treating the completed result as fully done. (#5287)
• dev: _pick_python() validates interpreter candidates through /usr/bin/env, mirroring how the frontend is launched, so make dev starts the frontend on Windows where Microsoft Store Python stubs satisfy Bash-side probes but not env. (#5181)

Performance

• runtime: Index MemoryRunStore by thread_id and MemoryRunEventStore events by run_id to avoid O(n) scans. (#3562, #3686)
• subagents: Deduplicate streamed AI messages via a seen-id set (O(n²) -> O(n)). (#3687)
• sandbox: Cache LocalSandbox path-rewrite regexes and local-path masking patterns per instance instead of recompiling per search match. (#3648, #3713)
• messages: Index tool-call results per group. (#4411)
• frontend: Coalesce streaming renders to a frame budget instead of per chunk. (#4425)
• frontend: Stop re-deriving message content on every stream chunk. (#4441)
• sandbox: read_file reads only the requested line range from the sandbox instead of fetching the whole file first. (#3824)
• browser: Encode Browser Live progress frames as JPEG to cut progress payload size. (#4836)
• middleware: Inject view_image content via wrap_model_call instead of a checkpointed hidden message, so up to 20 MB of base64 no longer sits in two checkpoints per viewed image and an interrupted run can no longer leave the payload behind. (#5014)
• frontend: Cache settled copy-data derivation across streaming chunks, so each chunk no longer re-derives toolbar/copy text for every settled message. (#5095)
• runtime: Bound gateway memory after terminal runs, stopping the post-GC low-water mark from creeping upward across completed sessions. (#5112)
• frontend: Chat streams request messages-tuple + updates + custom instead of full values state snapshots — the retransmitted historical values messages were ~75% of SSE payload — folding reducer events into rendered state locally while keeping full snapshots only for replay-gap recovery. (#5159)

Security

• prompt-injection: New input-sanitization middleware defends against prompt-injection, forged framework tags in the input guardrail are blocked, and system context is injected as a SystemMessage for role isolation. (#3662, #4155, #3661)
• prompt-injection: HTML-escape untrusted content rendered into model prompts
  • memory facts and summaries, SOUL.md, subagent descriptions, skill metadata, and the conversation block in the memory-update prompt - and neutralize prompt-injection tags in web_capture tool results. (#4028, #4119, #4137, #4157, #4162, #4099, #4060, #4097, #4128)
• prompt-injection: Close two input-sanitization bypasses. hide_from_ui and a human name="summary" tell is_genuine_user_message that the framework authored a message, which skips sanitization entirely. Untrusted run input and thread-state writes carrying either marker are now marked server-side and sanitized regardless, so a caller can no longer land a raw <system-reminder> outside the user-input boundary markers that the lead-agent prompt declares trusted framework data. The markers themselves are preserved, so messages that use hide_from_ui only to stay out of the transcript — quoted conversation context, sidecar context, the agent save command, HumanInputCard replies — keep doing that, and trusted internal launchers are unaffected. Sanitization also covers every genuine user message instead of only the newest: the transformation is request-scoped, so a last-turn-only scan neutralized a payload for exactly one model call and then replayed it verbatim from the next turn on. (#5375)
• secrets: Scrub inherited secret environment variables (MYSQL_PWD, REDISCLI_AUTH, abbreviated *_PASS, and Postgres PGPASSFILE) from the skill environment; request-scoped secrets are bound for both slash-activated and autonomously-invoked skills. (#4018, #4026, #3871, #3938)
• web_fetch: SSRF guard for self-hosted providers. (#3942)
• guardrails: An empty allowlist now denies all tools instead of failing open. (#4067)
• authz: Global skills-management endpoints now require admin; the legacy skills mount is gated by user visibility; artifacts honor a trusted owner-user-id header; and the trusted authorization principal is propagated through the runtime. (#3855, #3985, #3982, #4203)
• auth: Persist the csrf_token cookie for the access-token lifetime. (#3872)
• storage: Stop persisting base64 image data in checkpoint state. (#4140)
• mcp: Reject legacy MCP credentials in run metadata. (#4448)
• mcp: Constrain stdio launcher arguments and environment variables at the config API, rejecting launcher flags and env names that could turn an allowlisted npx/uvx server registration into arbitrary code execution. (#4617)
• auth: Harden validation of the post-login next path. (#4587)
• runtime: Honor the LangGraph Server's authenticated user identity across agents, uploads, thread data, memory, and skills, and reject client-supplied auth identity fields. (#4538)
• frontend: Send the session cookie on model, workspace-change, and ranged artifact reads in split-origin deployments. (#4827)
• frontend: Restore sanitization in custom streamdown rehype chains, so artifact markdown previews and the memory settings summary can no longer render hostile HTML such as javascript: links or on* event handlers. (#4987)
• skills: Copy projected skill files instead of hardlinking them, so a sandboxed write can no longer mutate the canonical skill source, and fail closed on a drifted projection namespace on every platform, including Windows. (#4825, #4830)
• scripts: Redact secret-shaped keys (db_pass, signing_key, ...) wherever they appear in bundled config, not only under well-known key names. (#4242)
• sandbox: Sanitize MCP-sourced tool results through the same trust boundary as the built-in web tools, so a hostile or compromised MCP server can no longer hand the model forged <system-reminder> or user-input boundary tags. (#4839)
• sandbox: Harden local Docker sandbox containers: published ports bind the Docker bridge gateway instead of 0.0.0.0 when the sandbox host is non-loopback (DEER_FLOW_SANDBOX_BIND_HOST=0.0.0.0 restores the broad bind), Docker's default seccomp profile replaces unconditional seccomp=unconfined (opt back in with DEER_FLOW_SANDBOX_SECCOMP_UNCONFINED=1), and containers drop all capabilities, get no-new-privileges, and run with bounded resources. (#4986)
• authz: Enforce run-create authorization on stateless stream/wait endpoints (runs:create), and require both threads:write and runs:create for scheduled-task create, update, resume, and manual-trigger mutations. (#5030)
• authz: Re-check the authorization policy before reusing a persisted sandbox, so a revoked sandbox:execute grant takes effect on the next sandbox-backed turn instead of outliving the policy in the cached sandbox. (#5006)
• skills: Enforce custom-Agent skill allowlists at the sandbox filesystem level: an explicit skills policy materializes a signed per-user/thread skills view, so a custom agent with shell or file tools can no longer read skills its policy excludes. (#5077)
• runs: Reject cancel/rollback actions on GET stream joins with 405 Method Not Allowed - cancel-then-stream is a POST operation - closing a state change that CSRF middleware deliberately exempted on safe methods; action-less GET joins are unchanged. (#5092)
• lark: Lark CLI credential trees on Windows enforce private ACLs (gateway-SID-only protected DACL, reparse-point rejection, handle-relative traversal), extending the POSIX 0700/0600 confidentiality contract against inherited grants, junction redirection, hard-link aliasing, and TOCTOU replacement. (#5141)
• sandbox: SSH_AUTH_SOCK is scrubbed from the sandbox subprocess environment — inheriting the host ssh-agent socket lets sandboxed code sign and authenticate with every key the agent holds — unless a skill explicitly declares it via required-secrets. (#5145)
• artifacts: Serve XML artifacts as download attachments like HTML and SVG. GET /api/threads/{id}/artifacts/{path} rendered .xml, .xsl, and .rdf files — and +xml types such as .rss wherever the host MIME database maps them — inline in the application origin, so an XML document with an XHTML-namespaced <script>, written by a prompt-injected agent and opened from a chat link, could call the API with the viewer's session. Every XML MIME type (text/xml, application/xml, text/xsl, any +xml subtype) is now treated as active content, including .skill archive members; the artifacts panel keeps previewing XML through its ranged fetch. (#5353)

Documentation

• docs: Clarify how LocalSandboxProvider resolves sandbox.mounts[].host_path under production Docker, with gateway bind-mount and config examples. (#3833)
• docs: Document that Crawl4AI >= 0.9 requires a bearer token. (#4518)
• docs: Document the GitHub inbound-dedupe TTL semantics, including what redeliveries are not deduped, and tighten the redelivery tests. (#4274)
• docs: Update the agent AGENTS.md and ARCHITECTURE.md guides. (#4817)
• docs: Document the Honcho memory backend with a dedicated guide and a long-term memory section entry in the README. (#4822)
• docs: Align custom-agent documentation with the API across the English and Chinese agents/threads/lead-agent pages: the required ASCII name request field, lowercase storage, /api/agents/check name-availability behavior, and no auto-derived slug from display_name. (#4944)

Internal

• tests: Migrate frontend unit tests to rstest and run hook-level tests in a DOM environment. (#3703, #4453)
• tests: Require explicit opt-in for live client tests. (#4482)
• tests: Rename the LLM-error test stand-in instead of the shared FakeError. (#4744)
• tests: Replace the magic unwritable absolute path in tool-output tests with a self-constructed failure condition. (#4722)
• tests: Add multi-turn message-stream invariants as graph integration tests. (#3708)
• tests: Add trace-based behavioral tests with Monocle Test Tools, asserting agent routing, tool calls, and token/duration cost. (#4025)
• tests: Cover passive skill tool visibility in the MCP layer. (#4247)
• tests: Add SQL and concurrent-reconciler coverage for lease-aware orphan recovery. (#4427)
• tests: Restore memory updater regression coverage. (#4490)
• tests: Lock in POST logout from the gateway-offline banner. (#4506)
• tests: Document known instance-client false negatives in the SkillScan tests. (#4644)
• refactor: Extract frontend placeholder detection into a tested utility. (#3783)
• refactor: Consolidate E2B client lifecycle helpers and reuse the kill helper during warm-pool eviction. (#4262, #4298)
• refactor: Name the E2B capacity-ledger meta-field count so the admission offset is explicit. (#4764)
• dev: Trace self/cls attribute chains and local aliases in the blocking-IO detector's call graph, closing false negatives. (#4200)
• ci: Publish the lark-cli-init and lark-broker images. (#4558)
• dev: Route host-side pnpm consumers through a shared runner with a Corepack fallback so local workflows work without a pnpm shim. (#4405)
• bench: Add an isolated checkpoint channel-mode benchmark comparing full and delta across latency, storage, and replay metrics. (#4395)
• deps: Bump cryptography 49.0.0 -> 50.0.0, postcss 8.4.31 -> 8.5.25, h2 4.3.0 -> 4.4.1, langgraph-checkpoint-sqlite and langgraph-checkpoint-postgres 3.1.0 -> 3.1.1, and nanoid 5.1.6 -> 5.1.16. (#4681, #4683, #4737, #4738, #4747, #4748)
• bench: Add a reproducible hybrid memory-eviction evaluation under backend/scripts/benchmark/deermem_eviction/ with a deterministic, blind-by-construction grader for the #4789 policy. (#4810)
• bench: Measure Postgres checkpoint/blob/write storage growth in the checkpoint benchmark alongside memory and SQLite. (#5051)
• tests: Exclude tests/blocking_io/ from make test; the dedicated make test-blocking-io suite (and its CI workflow) remains the owner. (#5105)
• refactor: Share sandbox identity derivation and acquire serialization across the five remote sandbox providers (RFC #4741), replacing five per-provider lock tables that grew unboundedly with process lifetime; derived ids are pinned byte-identical by per-provider golden vectors. (#5089)
• ci: Split the backend unit-test workflow into four parallel shards — each on its own runner with isolated Postgres/Redis — using pytest-split with a committed duration baseline and failing closed when the baseline is missing; make test remains the canonical full offline suite. (#5137)
• dev: Launch the Playwright webServer's Next.js through pnpm exec, so Windows E2E runs resolve the platform package binary instead of failing on the extensionless POSIX shim. (#5185)
• tests: Skip the POSIX mode-bit skill-permission assertions on Windows, where the chmod contract is unobservable, so Windows contributors can reach a green backend-suite baseline. (#5244)

2.0.0 — 2026-06-15

DeerFlow 2.0 is a ground-up rewrite around a "super agent" harness with sub-agents, persistent memory, sandbox execution, and an extensible skills/tools system. It shares no code with the 1.x line, which now lives on the main-1.x branch.

This release closes milestone 2.0.0 with 180 merged pull requests since the first 2.0 milestone tag.

⚠ Breaking changes

• harness: Hydrate runs from RunStore and persist interrupted status. Run cancellation/multitask semantics now require a working RunStore on the worker that owns the run; cross-worker cancels return 409 instead of silently appearing successful. (#2932)

Added

Agents & runtime

• agent: Custom-agent self-updates with user isolation — agents can persist edits to their own SOUL.md / config.yaml from inside a normal chat. (#2713)
• loop-detection: Make loop detection configurable with per-tool frequency overrides; keep configurable on/off switch. (#2586, #2711)
• loop-detection: Defer warning injection so detector pairs cleanly with tool-call lifecycle. (#2752)
• run: Propagate model_name from the gateway request through the runtime and persistence stack into the SQLite-backed store. (#2775)
• subagents: Stream subagent token usage to the header via terminal task events. (#2882)
• memory: Add memory.token_counting config to opt out of tiktoken for network-restricted deployments. (#3465)
• suggest: Make AI follow-up question suggestions optional. (#3591)

Models & integrations

• models: Add StepFun reasoning model adapter. (#3461)
• community: Add Brave Search web search tool. (#3528)
• channels: Enhance Discord with mention-only mode, thread routing, and typing indicators. (#2842)
• im: Add user-owned IM channel connections — users can bind their own Slack/Telegram/Discord/Feishu/DingTalk/WeChat/WeCom accounts on top of the operator-configured bots. (#3487)
• models: Add patched MiMo reasoning content support. (#3298)
• models: Add MiniMax provider for image/video/podcast skills plus a new music-generation skill. (#3437)
• community: Add SearXNG and Browserless web search/fetch tools. (#3451)
• community: Add Serper Google Images provider for image_search. (#3575)
• channels: Stream Telegram agent replies by editing the placeholder message in place. (#3534)

Observability

• trace: Set the LangGraph trace name to lead_agent (or the custom agent's agent_name) for cleaner Langfuse/LangSmith traces. (#3101)
• frontend: Refine token usage display modes. (#2329)
• defaults: Enable token usage tracking by default. (#2841)
• defaults: Raise default summarization trigger threshold. (#3174)
• trace: Attribute subagent spans to the parent thread's Langfuse trace. (#3611)

Skills

• skill: Add blocking-io-guard skill for blocking-IO triage and runtime anchors. (#3503)
• skill: Add maintainer issue and PR workflow skill. (#3554)
• skill: Strengthen the maintainer orchestrator review workflow. (#3606)

Performance

• harness: Push thread metadata filters into SQL instead of post-filtering in Python. (#2865)
• runtime: Index runs by thread_id to avoid O(n) scans in RunManager. (#3499)
• runtime: Index messages in MemoryRunEventStore to avoid O(n) scans. (#3531)
• persistence: Cache Base.to_dict column reflection per class. (#3654)
• sandbox: Speed up should_ignore_name in glob/grep walks. (#3657)

Security

• upload: Reject symlinked upload destinations. (#2623)
• uploads: Add Windows support for safe symlink-protected uploads. (#2794)
• mcp: Mask sensitive values in MCP config API responses. (#2667)
• mcp: Harden the MCP config endpoint against malformed input. (#3425)
• auth: Reject cross-site auth POSTs. (#2740)
• gateway: Cap skill artifact preview decompression to prevent zip-bomb-style abuse. (#2963)
• sandbox: Mount the host Docker socket only in aio (DooD) sandbox mode. (#3517)
• sandbox: Do not bind-mount host CLI auth dirs by default. (#3521)

Fixed

Runtime, gateway & persistence

• runtime: Rollback restore checkpoint now supersedes newer checkpoints. (#2582)
• runtime: Persist run message summaries. (#2850)
• runtime: Bound write_file execution-failure observations to keep failure traces from blowing out the context. (#3133)
• runtime: Protect the sync singleton's init and reset paths. (#3413)
• runtime: Avoid PostgreSQL aggregate FOR UPDATE on run events. (#2962)
• runs: Restore historical runs from persistent store after a gateway restart. (#2989)
• gateway: Return ISO 8601 timestamps from threads endpoints. (#2599)
• gateway: Make cancel idempotent for already-interrupted runs. (#3058)
• gateway: Split stream_existing_run into per-method routes for unique OpenAPI operationIds. (#3228)
• events: Serialize structured DB event content. (#2762)
• persistence: Emit timezone-aware timestamps from SQLite-backed stores. (#3130)
• persistence: Reuse token usage model grouping expression. (#2910)
• runs: Ignore stale run reconnect conflicts. (#3284)
• nginx: Defer CORS to the gateway allowlist instead of double-applying it. (#2861)
• persistence: Fix runtime journal run lifecycle events. (#3470)
• gateway: Enforce thread ownership on stateless run endpoints. (#3473)
• runtime: Propagate interrupt through SSE values events for the LangGraph SDK. (#3605)
• serialization: Strip base64 image data from streamed values events. (#3631)
• history: Strip base64 image data from REST endpoint responses. (#3535)
• gateway: Attribute token usage to the actual models. (#3658)

Agents, subagents & middleware

• subagents: Make subagent timeout terminal state atomic. (#2583)
• subagents: Use model override for tools and middleware. (#2641)
• subagents: Consolidate system_prompt and skills into a single SystemMessage. (#2701)
• subagent: Isolate subagents from the parent run's checkpointer. (#3559)
• agents: Make update_agent honor runtime.context user_id like setup_agent does. (#2867)
• agents: Resolve duplicate todos channel type conflict in TodoMiddleware. (#3200)
• agents: Offload blocking filesystem IO in the custom-agent router off the event loop. (#3457)
• agents: Keep new agent bootstrap in user scope. (#2784)
• loop-detection: Keep tool-call pairing on warn injection. (#2725)
• middleware: Sync raw tool-call metadata. (#2757)
• middleware: Handle invalid tool calls in dangling pairing middleware. (#2891)
• middleware: Prevent todo completion reminder IM-message leak. (#2907)
• middleware: Normalize tool result adjacency before model calls. (#2939)
• agents: Require config.yaml in resolve_agent_dir to skip memory-only directories. (#3481)
• agents: Sync agent_name across context/configurable and reject empty soul. (#3553)
• middleware: Offload the uploads scan in UploadsMiddleware off the event loop. (#3311)
• middleware: Offload memory injection off the event loop to prevent tiktoken blocking. (#3411)
• middleware: Externalize oversized tool output into the sandbox for non-mounted sandboxes. (#3417)
• middleware: Preserve the sandbox reducer in middleware state. (#3629)
• subagents: Raise general-purpose max_turns to 150 and default timeout to 30 min. (#3610)

Memory & tracing

• memory: Replace short-lived asyncio.run() with a persistent event loop. (#2627)
• memory: Isolate queued memory updates by agent. (#2941)
• memory: Parse wrapped memory-update JSON responses. (#3252)
• tracing: Propagate session_id and user_id into Langfuse traces. (#2944)
• trace: Decode unicode escape sequences in non-ASCII memory trace info. (#3104)

Tools, sandbox & MCP

• mcp: Fix env resolution in MCP config lists. (#2556)
• models: Record Codex token usage in usage_metadata. (#2585)
• sandbox: Supplement list_running in RemoteSandboxBackend. (#2716)
• sandbox: Disable MSYS path conversion for Git Bash on Windows. (#2766)
• sandbox: Avoid blocking sandbox readiness polling. (#2822)
• sandbox: Uphold the /mnt/user-data contract at the Sandbox API boundary. (#2881)
• sandbox: Scope provisioner PVC data by user. (#2973)
• sandbox: Merge idempotent sandbox state updates. (#3518)
• tools: Introduce Runtime type alias to eliminate Pydantic serialization warnings. (#2774)
• tools: Preserve tool_search promotions across re-entrant get_available_tools. (#2885)
• harness: Wrap async-only config tools for sync client execution. (#2878)
• harness: Wrap all async-only tools for sync clients. (#2935)
• tool-search: Reliably hide deferred MCP schemas by removing the ContextVar. (#3342)
• search: Fix DDGS Wikipedia region handling. (#3423)
• web_fetch: Support a proxy for the Jina reader in restricted networks. (#3430)
• sandbox: Persist lazily-acquired sandbox state via Command. (#3464)
• sandbox: Fix stale AIO sandbox cache reuse. (#3494)
• sandbox: Create a shell session before retrying on a fresh id. (#3577)
• sandbox: Stop flagging string-literal path fragments as unsafe absolute paths. (#3623)
• sandbox: Return an actionable hint when read_file hits a binary file. (#3624)
• mcp: Make stdio MCP-produced files resolvable via virtual sandbox paths. (#3600)
• mcp: Surface admin-required state on the settings tools page. (#3533)
• mcp: Add a tools cache reset endpoint. (#3602)
• uploads: Fix the upload file size contract. (#3408)

Skills & channels

• skills: Enforce allowed-tools metadata. (#2626)
• skills: Harden slash skill activation across chat channels. (#3466)
• skills: Fix custom skill install permissions. (#3241)
• channels: Authenticate gateway command requests. (#2742)
• skills: Surface the offending line and a quoting hint on SKILL.md YAML errors. (#3335)
• skills: Keep skill archive installation off the event loop. (#3505)
• channels: Ignore hidden control messages when extracting replies. (#3270)
• channels: Reload config on channel restart. (#3514)
• channels: Surface WeCom WebSocket connection failures. (#3526)
• channels: Close the Discord file handle after upload. (#3561)
• channels: Require a bound identity for user-owned IM messages. (#3578)
• channels: Scope IM files and helper commands to the owner. (#3579)
• channels: Make runtime provider state authoritative. (#3580)
• channels: Harden runtime credential management APIs. (#3581)
• channels: Make the channel connect flow deterministic. (#3582)
• channels: Centralize shared channel retry helpers. (#3583)
• channels: Add operational guardrails. (#3584)
• channels: Unsubscribe channel listeners by equality. (#3608)

Auth

• auth: Replace setup-status 429 rate limit with a cached response. (#2915)
• auth: Persist auto-generated JWT secret so it survives restarts. (#2933)
• auth: Align auth-disabled mode with mock history loading. (#3471)

Frontend

• frontend: Restore localhost fallback for getGatewayConfig in prod mode. (#2718)
• chat: Prevent the first user message from being swallowed in new conversations. (#2731)
• frontend: Use backend thread token usage for the header total. (#2800)
• frontend: Wait for async chat submit before clearing the input. (#2940)
• frontend: Resolve login page flickering and the resize-observer loop. (#2954)
• frontend: Deduplicate restored thread messages. (#2958)
• frontend: Avoid duplicate optimistic user message. (#3002)
• frontend: Hide the copy button for streaming assistant messages. (#3176)
• frontend: Show a new thread in the sidebar immediately on creation. (#3283)
• frontend: Isolate new chat thread messages. (#3508)
• frontend: Cap deeply nested list indentation to prevent render crashes. (#3393, #3570)
• token-usage: Dedupe token usage aggregation by message id. (#2770)
• frontend: Fall back to Streamdown clipboard copy. (#3397)
• frontend: Remove the Backspace shortcut for deleting prompt attachments. (#3410)
• frontend: Restructure the Memory settings toolbar into two rows. (#3433)
• suggestions: Strip inline <think> reasoning before parsing follow-up questions. (#3435)
• frontend: Stop fetching follow-up suggestions when they are disabled. (#3599)
• frontend: Paginate the workspace chat list beyond 50 threads. (#3485)
• frontend: Prevent user message bubble overflow with long unbreakable strings. (#3488)
• frontend: Keep the workspace interactive when the SSR auth probe cannot reach the gateway. (#3495)
• frontend: Render user messages as plain text and cap blockquote nesting. (#3502)
• frontend: Reset the active chat after deletion. (#3519)
• frontend: Improve the mobile workspace layout. (#3646)
• frontend: Render full content for multi-part AI messages. (#3649)

Build, deploy, scripts & config

• packaging: Add postgres extra for store/checkpointer support; clarify install guidance. (#2584)
• harness: Resolve runtime paths from the project root. (#2642)
• docker: Force nginx to resolve upstream names at request time. (#2717)
• docker: Default Gateway to a single worker to prevent multi-worker breakage. (#3475)
• scripts: Preserve uv extras across make dev restarts. (#2767, #2754)
• scripts: Clean up local nginx on stop. (#3005)
• deploy: Fall back to python / openssl when python3 is absent for secret generation. (#3074)
• config: Make the reload boundary discoverable from code. (#3144, #3153)
• replay-e2e: Key replay fixtures by caller and conversation. (#3453)
• setup: Refresh LLM provider wizard defaults. (#3421)
• config: Coerce null config.yaml list sections to an empty list. (#3434)
• scripts: Exclude runtime state from gateway reload. (#3426)
• scripts: Create the backend/sandbox dir before the uvicorn reload-exclude. (#3460)
• scripts: Stop next-server correctly after make start-daemon. (#3498)
• makefile: Fix per-commit hooks installation. (#3569)
• replay-e2e: Match replay by conversation, not the living system prompt. (#3436)

Changed

• provider (refactor): Share assistant payload replay matching across providers. (#3307)
• lead-agent (refactor): Make build_middlewares public to drop the last cross-module private import. (#3458)
• todo (refactor): Remove the unused completion reminder counter. (#3530)

Documentation

• Document blocking-IO detection usage and maintenance. (#3233)
• Clean standalone LangGraph server remnants from docs. (#3301)
• Add AI assistance disclosure to the PR template and CONTRIBUTING. (#3398)
• Document custom AIO sandbox images. (#3548)

Internal

• dev: Add async/thread boundary detector. (#2936)
• runtime: Add lifecycle end-to-end coverage. (#2946)
• windows: Add PYTHONIOENCODING and PYTHONUTF8 to backend Makefile targets. (#3069)
• blocking-io: Fail-loud repo-root resolution and shared detector CLI shim. (#3512)
• runtime: Add a Blockbuster runtime anchor for JsonlRunEventStore async IO. (#3313)
• ci: Consolidate PR/issue labeling and fix the reviewing-job crash and label thrash. (#3455)