Nan Gao 4fcb4bc366
feat(subagents): persist and display subagent step history (#3779) (#3845)
* feat(subagents): persist and display subagent step history (#3779)

Capture both assistant turns and tool outputs during subagent execution,
stream them in task_running events, and persist them as subagent.* run
events so the subtask card's step timeline survives a reload.

Backend:
- step_events.py: pure layer (capture_step_message, build_subagent_step,
  subagent_run_event) shared by streaming and persistence
- executor.py: capture ToolMessage outputs, not just AIMessage turns
- worker.py: persist task_* custom events to RunEventStore (category
  "subagent" keeps them out of the thread feed; list_events backfills)

Frontend:
- core/tasks/steps.ts + api.ts: SubtaskStep model, messageToStep,
  eventsToSteps, mergeSteps, fetchSubtaskSteps
- subtask card accumulates live steps and backfills on expand
- carry run_id onto history content messages for the events endpoint

* fix(subagents): show AI turns in subtask card + paginate step backfill (#3779)

Two follow-ups to the subagent step-history feature:

Problem 1 — reload backfill could silently truncate the step timeline because
list_events capped at 500 events (seq-ASC) across the whole run. Add task_id
filtering + an after_seq forward cursor to list_events (all three stores +
abstract base + the /events route), and make fetchSubtaskSteps page through one
task's subagent.step events until a short page. No schema migration: the DB
filter rides the existing run-scoped index via event_metadata["task_id"].

Problem 2 — the card only rendered tool steps, so persisted AI turns were never
shown. Replace toolStepsForDisplay with stepsForDisplay: interleave AI reasoning
turns (with text) and tool steps by message_index, drop blank-text AI turns, and
drop the trailing final-answer AI turn when completed (already shown as result).
Card renders AI steps as muted clamped markdown with a sparkles icon.

Tests: store task_id/after_seq filtering + pagination across memory/db/jsonl,
the /events route forwarding, stepsForDisplay rules, and fetchSubtaskSteps
pagination. Docs updated in both AGENTS.md.

* make format

* fix(subagents): capture full multi-tool step tail, batch step persistence, cap tool-call args (#3779)

Address PR review findings on the subagent step-history feature:

1. executor.py streamed on stream_mode="values" and captured only
   messages[-1] per chunk, so a multi-tool-call turn (ToolNode appends
   one ToolMessage per call in a single super-step) lost all but the last
   tool output in both the live task_running stream and the persisted
   history. Replace with capture_new_step_messages, which walks the
   newly-appended tail (and still re-checks the trailing message on
   no-growth chunks so id-less in-place replacements survive).

2. worker.py persisted each step with the store's low-frequency put()
   (a per-thread advisory lock per call); a deep subagent (max_turns=150)
   emits hundreds of steps on the hot stream loop. Replace with
   _SubagentEventBuffer, which batches via put_batch (flush on terminal
   subagent.end, at FLUSH_THRESHOLD, and in the worker finally).

3. build_subagent_step capped only text; tool_calls[].args were copied
   verbatim, so a large write_file/bash payload produced an unbounded
   subagent.step row. Cap each call's serialized args at
   SUBAGENT_STEP_MAX_CHARS, flagged args_truncated.

Tests updated/added for all three; AGENTS.md refreshed.

* fix(subagents): merge backfill into latest subtask state; reuse message_content_to_text (#3779)

Address the remaining two PR review findings:

4. subtask-card's fetchSubtaskSteps().then(updateSubtask) closed over a
   stale tasks snapshot: a late-resolving backfill wrote setTasks({...stale}),
   clobbering SSE steps/status and sibling subtasks that arrived during the
   fetch. useUpdateSubtask now reads/writes through a tasksRef mirroring the
   latest state (ref-to-latest), and the pure per-subtask transition is
   extracted to core/tasks/subtask-update.ts::computeNextSubtask (unit-tested).

5. step_events._content_to_text duplicated deerflow.utils.messages.
   message_content_to_text; call the shared helper instead (guarding None
   content with 'or ""' so a tool-call-only turn still renders as "").

Tests added for computeNextSubtask and the None-content case; AGENTS.md docs updated.
2026-07-02 07:43:09 +08:00
..
2026-01-14 09:57:52 +08:00

DeerFlow Backend

DeerFlow is a LangGraph-based AI super agent with sandbox execution, persistent memory, and extensible tool integration. The backend enables AI agents to execute code, browse the web, manage files, delegate tasks to subagents, and retain context across conversations - all in isolated, per-thread environments.


Architecture

                        ┌──────────────────────────────────────┐
                        │          Nginx (Port 2026)           │
                        │      Unified reverse proxy           │
                        └───────┬──────────────────┬───────────┘
                                │
            /api/langgraph/*    │    /api/* (other)
            rewritten to /api/* │
                                ▼
               ┌────────────────────────────────────────┐
               │        Gateway API (8001)              │
               │        FastAPI REST + agent runtime    │
               │                                        │
               │ Models, MCP, Skills, Memory, Uploads,  │
               │ Artifacts, Threads, Runs, Streaming    │
               │                                        │
               │ ┌────────────────────────────────────┐ │
               │ │ Lead Agent                         │ │
               │ │ Middleware Chain, Tools, Subagents │ │
               │ └────────────────────────────────────┘ │
               └────────────────────────────────────────┘

Request Routing (via Nginx):

  • /api/langgraph/* → Gateway LangGraph-compatible API - agent interactions, threads, streaming
  • /api/* (other) → Gateway API - models, MCP, skills, memory, artifacts, uploads, thread-local cleanup
  • / (non-API) → Frontend - Next.js web interface

Core Components

Lead Agent

The single LangGraph agent (lead_agent) is the runtime entry point, created via make_lead_agent(config). It combines:

  • Dynamic model selection with thinking and vision support
  • Middleware chain for cross-cutting concerns (9 middlewares)
  • Tool system with sandbox, MCP, community, and built-in tools
  • Subagent delegation for parallel task execution
  • System prompt with skills injection, memory context, and working directory guidance

Middleware Chain

Middlewares execute in strict order, each handling a specific concern:

# Middleware Purpose
1 ThreadDataMiddleware Creates per-thread isolated directories (workspace, uploads, outputs)
2 UploadsMiddleware Injects newly uploaded files into conversation context
3 SandboxMiddleware Acquires sandbox environment for code execution
4 SummarizationMiddleware Reduces context when approaching token limits (optional)
5 TodoListMiddleware Tracks multi-step tasks in plan mode (optional)
6 TitleMiddleware Auto-generates conversation titles after first exchange
7 MemoryMiddleware Queues conversations for async memory extraction
8 ViewImageMiddleware Injects image data for vision-capable models (conditional)
9 ClarificationMiddleware Intercepts clarification requests and interrupts execution (must be last)

Sandbox System

Per-thread isolated execution with virtual path translation:

  • Abstract interface: execute_command, read_file, write_file, list_dir
  • Providers: LocalSandboxProvider (filesystem) and AioSandboxProvider (Docker, in community/). Async runtime paths use async sandbox lifecycle hooks so startup, readiness polling, and release do not block the event loop. AioSandboxProvider validates active-cache and warm-pool containers during acquire/reuse, dropping definitively dead entries so a thread can provision a fresh sandbox after an unexpected container exit while keeping get() as an in-memory lookup. Backend health-check failures are treated as unknown, not dead, and a container that cannot be verified during discovery is simply not adopted (acquire falls through to create instead of failing).
  • Virtual paths: /mnt/user-data/{workspace,uploads,outputs} → thread-specific physical directories
  • Skills path: /mnt/skillsdeer-flow/skills/ directory
  • Skills loading: Recursively discovers nested SKILL.md files under skills/{public,custom} and preserves nested container paths
  • File-write safety: str_replace serializes read-modify-write per (sandbox.id, path) so isolated sandboxes keep concurrency even when virtual paths match
  • Tools: bash, ls, read_file, write_file, str_replace (write_file overwrites by default and exposes append for end-of-file writes; bash is disabled by default when using LocalSandboxProvider; use AioSandboxProvider for isolated shell access)

Subagent System

Async task delegation with concurrent execution:

  • Built-in agents: general-purpose (full toolset) and bash (command specialist, exposed only when shell access is available)
  • Concurrency: Max 3 subagents per turn, 15-minute timeout
  • Execution: Background thread pools with status tracking and SSE events
  • Flow: Agent calls task() tool → executor runs subagent in background → polls for completion → returns result

Memory System

LLM-powered persistent context retention across conversations:

  • Automatic extraction: Analyzes conversations for user context, facts, and preferences
  • Structured storage: User context (work, personal, top-of-mind), history, and confidence-scored facts
  • Debounced updates: Batches updates to minimize LLM calls (configurable wait time)
  • System prompt injection: Top facts + context injected into agent prompts
  • Storage: JSON file with mtime-based cache invalidation

Tool Ecosystem

Category Tools
Sandbox bash, ls, read_file, write_file, str_replace
Built-in present_files, ask_clarification, view_image, task (subagent)
Community Tavily (web search), Jina AI (web fetch), Firecrawl (scraping), fastCRW (scraping), DuckDuckGo (image search)
MCP Any Model Context Protocol server (stdio, SSE, HTTP transports)
Skills Domain-specific workflows injected via system prompt

Gateway API

FastAPI application providing REST endpoints for frontend integration:

Route Purpose
GET /api/models List available LLM models
GET/PUT /api/mcp/config Manage MCP server configurations
POST /api/mcp/cache/reset Reset cached MCP tools so they reload on next use
GET/PUT /api/skills List and manage skills
POST /api/skills/install Install skill from .skill archive
GET /api/memory Retrieve memory data
POST /api/memory/reload Force memory reload
GET /api/memory/config Memory configuration
GET /api/memory/status Combined config + data
POST /api/threads/{id}/uploads Upload files (auto-converts PDF/PPT/Excel/Word to Markdown, rejects directory paths, auto-renames duplicate filenames in one request)
GET /api/threads/{id}/uploads/list List uploaded files
DELETE /api/threads/{id} Delete DeerFlow-managed local thread data after LangGraph thread deletion; unexpected failures are logged server-side and return a generic 500 detail
GET /api/threads/{id}/artifacts/{path} Serve generated artifacts

IM Channels

The IM bridge supports Feishu, Slack, and Telegram. Slack and Telegram still use the final runs.wait() response path, while Feishu now streams through runs.stream(["messages-tuple", "values"]) and updates a single in-thread card in place.

For Feishu card updates, DeerFlow stores the running card's message_id per inbound message and patches that same card until the run finishes, preserving the existing OK / DONE reaction flow.


Quick Start

Prerequisites

  • Python 3.12+
  • uv package manager
  • API keys for your chosen LLM provider

Installation

cd deer-flow

# Copy configuration files
cp config.example.yaml config.yaml

# Install backend dependencies
cd backend
make install

Configuration

Edit config.yaml in the project root:

models:
  - name: gpt-4o
    display_name: GPT-4o
    use: langchain_openai:ChatOpenAI
    model: gpt-4o
    api_key: $OPENAI_API_KEY
    supports_thinking: false
    supports_vision: true

  - name: gpt-5-responses
    display_name: GPT-5 (Responses API)
    use: langchain_openai:ChatOpenAI
    model: gpt-5
    api_key: $OPENAI_API_KEY
    use_responses_api: true
    output_version: responses/v1
    supports_vision: true

Set your API keys:

export OPENAI_API_KEY="your-api-key-here"

Running

Full Application (from project root):

make dev  # Starts Gateway + Frontend + Nginx

Access at: http://localhost:2026

Backend Only (from backend directory):

# Gateway API + embedded agent runtime
make dev

Direct access: Gateway at http://localhost:8001

Terminal Workbench (TUI) — a terminal-native UI over the embedded harness, no services required:

uv pip install 'deerflow-harness[tui]'   # optional 'textual' dependency
deerflow                                 # launch the TUI
deerflow --print "summarize this repo"   # headless one-shot

Sessions opened in the TUI appear in the Web UI sidebar (it writes the shared threads_meta store under the local default user). See docs/TUI.md.


Project Structure

backend/
├── src/
│   ├── agents/                  # Agent system
│   │   ├── lead_agent/         # Main agent (factory, prompts)
│   │   ├── middlewares/        # 9 middleware components
│   │   ├── memory/             # Memory extraction & storage
│   │   └── thread_state.py    # ThreadState schema
│   ├── gateway/                # FastAPI Gateway API
│   │   ├── app.py             # Application setup
│   │   └── routers/           # 6 route modules
│   ├── sandbox/                # Sandbox execution
│   │   ├── local/             # Local filesystem provider
│   │   ├── sandbox.py         # Abstract interface
│   │   ├── tools.py           # bash, ls, read/write/str_replace
│   │   └── middleware.py      # Sandbox lifecycle
│   ├── subagents/              # Subagent delegation
│   │   ├── builtins/          # general-purpose, bash agents
│   │   ├── executor.py        # Background execution engine
│   │   └── registry.py        # Agent registry
│   ├── tools/builtins/         # Built-in tools
│   ├── mcp/                    # MCP protocol integration
│   ├── models/                 # Model factory
│   ├── skills/                 # Skill discovery & loading
│   ├── config/                 # Configuration system
│   ├── community/              # Community tools & providers
│   ├── reflection/             # Dynamic module loading
│   └── utils/                  # Utilities
├── docs/                       # Documentation
├── tests/                      # Test suite
├── langgraph.json              # LangGraph graph registry for tooling/Studio compatibility
├── pyproject.toml              # Python dependencies
├── Makefile                    # Development commands
└── Dockerfile                  # Container build

langgraph.json is not the default service entrypoint. The scripts and Docker deployments run the Gateway embedded runtime; the file is kept for LangGraph tooling, Studio, or direct LangGraph Server compatibility.


Configuration

Main Configuration (config.yaml)

Place in project root. Config values starting with $ resolve as environment variables.

Key sections:

  • models - LLM configurations with class paths, API keys, thinking/vision flags
  • tools - Tool definitions with module paths and groups
  • tool_groups - Logical tool groupings
  • sandbox - Execution environment provider
  • skills - Skills directory paths
  • title - Auto-title generation settings
  • summarization - Context summarization settings
  • subagents - Subagent system (enabled/disabled)
  • memory - Memory system settings (enabled, storage, debounce, facts limits)

Provider note:

  • models[*].use references provider classes by module path (for example langchain_openai:ChatOpenAI).
  • If a provider module is missing, DeerFlow now returns an actionable error with install guidance (for example uv add langchain-google-genai).

Extensions Configuration (extensions_config.json)

MCP servers and skill states in a single file:

{
  "mcpServers": {
    "github": {
      "enabled": true,
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {"GITHUB_TOKEN": "$GITHUB_TOKEN"}
    },
    "secure-http": {
      "enabled": true,
      "type": "http",
      "url": "https://api.example.com/mcp",
      "oauth": {
        "enabled": true,
        "token_url": "https://auth.example.com/oauth/token",
        "grant_type": "client_credentials",
        "client_id": "$MCP_OAUTH_CLIENT_ID",
        "client_secret": "$MCP_OAUTH_CLIENT_SECRET"
      }
    }
  },
  "skills": {
    "pdf-processing": {"enabled": true}
  }
}

Environment Variables

  • DEER_FLOW_CONFIG_PATH - Override config.yaml location
  • DEER_FLOW_EXTENSIONS_CONFIG_PATH - Override extensions_config.json location
  • Model API keys: OPENAI_API_KEY, ANTHROPIC_API_KEY, DEEPSEEK_API_KEY, etc.
  • Tool API keys: TAVILY_API_KEY, GITHUB_TOKEN, etc.

LangSmith Tracing

DeerFlow has built-in LangSmith integration for observability. When enabled, all LLM calls, agent runs, tool executions, and middleware processing are traced and visible in the LangSmith dashboard.

Setup:

  1. Sign up at smith.langchain.com and create a project.
  2. Add the following to your .env file in the project root:
LANGSMITH_TRACING=true
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=lsv2_pt_xxxxxxxxxxxxxxxx
LANGSMITH_PROJECT=xxx

Legacy variables: The LANGCHAIN_TRACING_V2, LANGCHAIN_API_KEY, LANGCHAIN_PROJECT, and LANGCHAIN_ENDPOINT variables are also supported for backward compatibility. LANGSMITH_* variables take precedence when both are set.

Langfuse Tracing

DeerFlow also supports Langfuse observability for LangChain-compatible runs.

Add the following to your .env file:

LANGFUSE_TRACING=true
LANGFUSE_PUBLIC_KEY=pk-lf-xxxxxxxxxxxxxxxx
LANGFUSE_SECRET_KEY=sk-lf-xxxxxxxxxxxxxxxx
LANGFUSE_BASE_URL=https://cloud.langfuse.com

If you are using a self-hosted Langfuse deployment, set LANGFUSE_BASE_URL to your Langfuse host.

Dual Provider Behavior

If both LangSmith and Langfuse are enabled, DeerFlow initializes and attaches both callbacks so the same run data is reported to both systems.

If a provider is explicitly enabled but required credentials are missing, or the provider callback cannot be initialized, DeerFlow raises an error when tracing is initialized during model creation instead of silently disabling tracing.

Docker: In docker-compose.yaml, tracing is disabled by default (LANGSMITH_TRACING=false). Set LANGSMITH_TRACING=true and/or LANGFUSE_TRACING=true in your .env, together with the required credentials, to enable tracing in containerized deployments.


Development

Commands

make install    # Install dependencies
make dev        # Run Gateway API + embedded agent runtime (port 8001)
make gateway    # Run Gateway API without reload (port 8001)
make lint       # Run linter (ruff)
make format     # Format code (ruff)
make detect-blocking-io  # Inventory blocking IO that may block the backend event loop
make migrate-rev MSG="..."  # Autogenerate a new alembic revision against the live ORM models

Schema Migrations

DeerFlow's application tables (runs, threads_meta, feedback, users, run_events, and the channel_* tables) are owned by alembic. The Gateway runs alembic upgrade head automatically on startup via bootstrap_schema(engine, backend=...), so operators do not run alembic manually in production. Bootstrap is concurrency-safe (Postgres advisory lock across processes; per-engine asyncio.Lock inside one SQLite process) and idempotent against pre-existing schemas (empty / legacy / versioned).

When you add or change an ORM model, ship the change as a new revision under packages/harness/deerflow/persistence/migrations/versions/:

make migrate-rev MSG="add foo column to runs"

The target invokes scripts/_autogen_revision.py, which builds a fresh temp SQLite at head and diffs the live models against it — so a clean checkout does not need a pre-existing ./data/deerflow.db. Review the generated file and switch raw op.add_column / op.drop_column calls to the idempotent helpers in migrations/_helpers.py before committing. There is no make migrate / make migrate-stamp target on purpose — Gateway startup is the only execution path, which keeps operational mistakes off the table. See backend/CLAUDE.md (Schema Migrations) for the full design.

Code Style

  • Linter/Formatter: ruff
  • Line length: 240 characters
  • Python: 3.12+ with type hints
  • Quotes: Double quotes
  • Indentation: 4 spaces

Testing

uv run pytest

make detect-blocking-io statically scans backend business code for blocking IO that may run on the backend event loop and is not test-coverage-bound. It prints a concise summary for human review and writes complete JSON findings to .deer-flow/blocking-io-findings.json at the repository root (regardless of whether the target is invoked from the repo root or from backend/). JSON findings include both broad IO category and review-oriented fields such as priority, location, blocking_call, event_loop_exposure, reason, and code. priority is a deterministic review ordering from the operation type, not proof of a bug. Bare-name same-file calls are resolved by function name, so duplicate helper names in one file can conservatively over-report async reachability.


Technology Stack

  • LangGraph (1.0.6+) - Agent framework and multi-agent orchestration
  • LangChain (1.2.3+) - LLM abstractions and tool system
  • FastAPI (0.115.0+) - Gateway REST API
  • langchain-mcp-adapters - Model Context Protocol support
  • agent-sandbox - Sandboxed code execution
  • markitdown - Multi-format document conversion
  • tavily-python / firecrawl-py - Web search and scraping

Documentation


License

See the LICENSE file in the project root.

Contributing

See CONTRIBUTING.md for contribution guidelines.