bytedance/deer-flow

English | 中文 | 日本語 | Français | Русский | Português | 한국어

On February 28th, 2026, DeerFlow claimed the 🏆 #1 spot on GitHub Trending following the launch of version 2. Thanks a million to our incredible community — you made this happen! 💪🔥

DeerFlow (Deep Exploration and Efficient Research Flow) is an open-source super agent harness that orchestrates sub-agents, memory, and sandboxes to do almost anything — powered by extensible skills.

Learn more and see real demos on our official website. The landing-page case studies open as allowlisted, read-only showcases without requiring a sign-in.

• LLM Space - Meet our secret weapon behind DeerFlow — one desktop tool to prototype agent ideas, inspect each harness step, replay failures, and benchmark performance.

• We strongly recommend using Doubao-Seed-2.0-Code, DeepSeek v3.2 and Kimi 2.5 to run DeerFlow
• Learn more
• 中国大陆地区的开发者请点击这里

InfoQuest reader, web search, and image search use a 30-second HTTP connect/read inactivity timeout. The crawl timeout and navigation_timeout settings remain separate server-side options; they do not control the local HTTP timeout.

DeerFlow has newly integrated the intelligent search and crawling toolset independently developed by BytePlus — InfoQuest (supports free online experience)

• 🦌 DeerFlow - 2.0
  • Official Website
  • Sister Projects
  • Coding Plan from ByteDance Volcengine
  • InfoQuest
  • Table of Contents
  • One-Line Agent Setup
  • Quick Start
    • Configuration
    • Running the Application
      • Deployment Sizing
      • Option 1: Docker (Recommended)
      • Upgrading an existing checkout
      • Option 2: Local Development
      • Startup Modes
      • LangGraph Studio (Optional)
      • Docker Production Deployment
    • Advanced
      • Sandbox Mode
      • MCP Server
      • IM Channels
      • Request Trace Correlation
      • LangSmith Tracing
      • Langfuse Tracing
      • Monocle Tracing
      • Using Multiple Providers
      • Existing-Run Stream Actions
      • Personal Access Tokens
  • From Deep Research to Super Agent Harness
  • Core Features
    • Skills & Tools
      • Exporting Custom Skills
      • Claude Code Integration
    • Private Knowledge Retrieval (RAGFlow)
    • Chat Archive
    • Session Goals
    • Manual Context Compaction
    • Sub-Agents
    • Sandbox & File System
    • Agentic Browser Control
    • Context Engineering
    • Reading a Referenced Conversation
    • Current Task Notes
    • Long-Term Memory
  • Recommended Models
  • Embedded Python Client
  • Projects
    • Project instructions
    • Document shelf
    • Archive read semantics
    • Trash
  • Scheduled Tasks
    • Preview cron occurrences through the API
    • Upgrade Notes
  • Terminal Workbench (TUI)
  • Documentation
  • ⚠️ Security Notice
    • Improper Deployment May Introduce Security Risks
    • Gateway Admin Is Equivalent to Code Execution
    • External Chat Message Roles
    • Deployment Defaults
    • Security Recommendations
  • Contributing
  • License
  • Acknowledgments
    • Key Contributors
  • Star History

If you use Claude Code, Codex, Cursor, Windsurf, or another coding agent, you can hand it the setup instructions in one sentence:

Help me clone DeerFlow if needed, then bootstrap it for local development by following https://raw.githubusercontent.com/bytedance/deer-flow/main/Install.md

That prompt is intended for coding agents. It tells the agent to clone the repo if needed, choose Docker when available, and stop with the exact next command plus any missing config the user still needs to provide.

With run_events.backend: db, caller-supplied ISO 8601 event timestamps are normalized to UTC before storage, preserving their instant when SQLite drops timezone information. Timestamps without an offset are treated as UTC. Previously stored timestamps are not rewritten.

Operators can extend lead-agent, subagent, and DeerMem extraction prompts with literal prepend/append configuration without editing source templates. See prompt overlays.

Optional per-model request_admission paces requests to help stay within provider request-per-minute limits. It is disabled by default; see the linked guide to enable it. Enabling it disables exposed SDK retries and the Claude and Codex adapters' internal retry loops so middleware retries pass through request admission again. Middleware retries also cover HTTP 529 overload responses. A warning identifies retry_max_attempts values overridden by admission.

For Google's official Gemini OpenAI-compatible endpoint, use the Gemini reasoning profile.

For MindIE XML tool calls, see the argument parsing and newline compatibility guide. Both synchronous and asynchronous streams retain this compatibility: tool-enabled streams simulate chunks from a non-streaming response, while no-tool streams stay native.

1. Clone the DeerFlow repository

   git clone https://github.com/bytedance/deer-flow.git
   cd deer-flow

2. Run the setup wizard

   From the project root directory (deer-flow/), run:

   make setup

   This launches an interactive wizard that guides you through choosing an LLM provider, optional web search, and execution/safety preferences such as sandbox mode, bash access, and file-write tools. It generates a minimal config.yaml and writes your keys to .env. Takes about 2 minutes.

   The wizard also lets you configure an optional web search provider, or skip it for now.

   Brave web search preserves valid entries in mixed result lists. Malformed response containers or lists containing no usable entries return a structured format error; missing, null, or empty results keep the existing "No results found" response. Format errors also log the malformed container's path and type, or the absence of usable result objects, without including search queries, credentials, or payload values.

   GroundRoute web search and fetch contain malformed HTTP-success payloads the same way: a non-object payload, a results container that is not a list, or a non-empty list with no object entries returns the provider's format error, while non-object entries in an otherwise valid list are skipped in order. Missing, null, or empty results keep the existing "No results found" response.

   DuckDuckGo web search distinguishes execution failures from empty results: rate limits, timeouts, blocked requests, client-setup failures, SDK errors, and a missing ddgs dependency return a structured DuckDuckGo search failed: ... error instead of "No results found", while a genuinely empty result set keeps the existing "No results found" response. The structured error text lets the harness classify an outage as transient or a configuration problem instead of a query that matched nothing.

   Tavily web search and fetch contain malformed HTTP-success payloads the same way: a non-object payload, or a results/failed_results container that is not a list, returns the provider's format error, and non-object entries in an otherwise valid list are skipped in order. Missing, null, or empty results keep the existing response ([] for search, Error: No results found for fetch); missing or null search fields (title, url, content) normalize to empty text, and a non-string fetch raw_content is coerced to text before the truncation limit.

   Sofya web search and fetch contain malformed HTTP-success payloads the same way: a non-object payload, a results container that is not a list, or a non-empty list with no object entries returns the provider's format error, while non-object entries in an otherwise valid list are skipped in order. Missing, null, or empty results keep the existing "No results found" response.

   Serper web and image search preserve valid entries in mixed result lists. A malformed organic/images container, or a non-empty list holding no object entries, returns Serper's format error; missing, null, or empty containers keep the existing "No results found" / "No images found" response. Format errors log the container field name only, without queries, credentials, or payload values.

   Tencent Cloud WSA web search contains malformed HTTP-success payloads the same way: a Pages container that is not a list, or a non-empty container holding no usable page, returns the provider's format error, while malformed page entries in an otherwise valid container are skipped so the valid rows survive. Missing, null, or empty Pages keep the existing "No results found" response.

   Serply web search contains malformed HTTP-success payloads the same way in every vertical: a results/entries/articles container that is not a list, or a non-empty list with no object entries, returns the provider's format error, while non-object entries in an otherwise valid list are skipped in order. Missing, null, or empty containers keep the existing "No results found" response.

   InfoQuest web and image search contain malformed HTTP-success payloads the same way: a search_result envelope that is not an object, or an organic/top_stories.items/images_results container that is not a list or that holds no usable entry, returns the provider's format error, while malformed entries in an otherwise valid payload are skipped so the usable rows survive. Missing, null, or empty containers keep the existing empty result. Format errors log the container or entry position only, without queries or payload values.

   Jina, Browserless, and InfoQuest web fetches resolve relative links and image sources using the requested page URL (or a usable HTML base URL), so returned Markdown includes complete destinations. Link resolution preserves the surrounding HTML source, including malformed-page formatting.

   Self-hosted Browserless, Crawl4AI, Firecrawl, and fastCRW fetch backends require outbound isolation from private, loopback, link-local, and cloud-metadata networks before setting network_isolation_confirmed: true on each tool. DeerFlow rejects private or unverifiable backends by default because those services handle redirects, DNS resolution, and subresources themselves. allow_private_addresses permits intentional internal target URLs and does not bypass the backend check. See delegated fetch deployment; make doctor reports backend configurations that would be refused.

   Jina fetches support opt-in bounded retries via max_retries (default 0) and retry_budget_seconds (default 30) in the tool configuration. Valid Retry-After hints set a minimum wait for HTTP 429/503; 429 without a valid hint stays terminal. Hints that cannot fit the remaining budget stop retries. Local backoff remains randomized. Retries may increase upstream requests and cost; see Jina fetch retries.

   Jina also accepts an opt-in max_response_bytes tool setting (positive integer; omitted/null disables it). It stops oversized decoded responses before extraction, with no partial success or retry. This leaves the 4096-character output cap unchanged and does not bound HTTPX decompressor allocations or wire bandwidth; see response budget.

   Jina can also bound concurrent HTTP attempts and FIFO waiting with the opt-in request_admission tool extra. One immutable budget is shared across clients and agent threads in each process; replicas have independent budgets. See Jina request admission.

   Run make doctor at any time to verify your setup and get actionable fix hints. If you are opening a GitHub issue about a local setup or runtime problem, run make support-bundle. The command prints reporter next steps, writes a *-issue-summary.md file to paste into the issue, a *-issue-draft.md file for AI-assisted issue filing, and an optional evidence zip under .deer-flow/support-bundles/. If an AI assistant files the issue, start from the draft and replace every REQUIRED placeholder instead of inventing missing facts. Attach the zip only if a maintainer asks for it, or if the summary alone is not enough. Maintainers and AI triage tools can start with triage.json; the bundle includes redacted diagnostics and file manifests only, and does not include .env, raw conversation messages, or user file contents. Subprocess diagnostics are captured as UTF-8, with Python helpers emitting UTF-8 even on non-UTF-8 hosts and escaping unencodable characters. Thread manifests follow the local launcher's runtime paths: checkout .env values override shell exports, and a project-root override alone still uses backend/.deer-flow first. Simple variable references such as DEER_FLOW_HOME="$PWD/backend/.deer-flow" use the checkout as PWD; single-quoted references remain literal. If the checkout .env cannot be read or decoded as UTF-8, thread diagnostics retain shell exports and default storage paths. For standalone Gateway launches using backend/.env or DEER_FLOW_ENV_FILE, export the effective DEER_FLOW_HOME when collecting the bundle and ensure the checkout .env does not override it. Doctor's internal tool probes also decode UTF-8 with replacement for invalid bytes so the remaining diagnostic output stays available.

   When a thread manifest is requested, a nonempty DEER_FLOW_HOME selects the runtime data directory instead of stale checkout data. External data paths appear as {DEER_FLOW_HOME} in the report; file contents remain excluded. Runtime path settings also come from the checkout .env, with shell exports taking precedence. Relative paths resolve from the checkout as in make dev. Without an explicit home, a configured DEER_FLOW_PROJECT_ROOT selects its .deer-flow directory; external paths use that variable name in the report.

   Advanced / manual configuration: If you prefer to edit config.yaml directly, run make config instead to copy the full template. Optional dependency auto-detection accepts UTF-8 configuration files with or without a byte-order mark (BOM). See config.example.yaml for the complete reference including CLI-backed providers (Codex CLI, Claude Code OAuth), OpenRouter, Responses API, subagent runtime caps such as subagents.max_total_per_run, and more.

   Optional per-model pricing must use one currency across all priced models. DeerFlow disables Console cost estimates when currencies are mixed rather than presenting an invalid aggregate.

   Administrators can open Settings → Models to select a provider and enter only its API Key and model ID. Endpoints, authentication and adapters are backend presets for OpenAI, Anthropic, DeepSeek, Alibaba Cloud Bailian/Qwen, Volcengine/Doubao, Moonshot/Kimi, Zhipu/Z.AI and MiniMax. Region and plan-specific entries are explicitly distinguished. Arbitrary model IDs are accepted; internal names are generated automatically and remain stable when a model ID changes. Newly saved models are enabled and immediately refresh existing chat and Agent model selectors, without editing YAML or restarting the Gateway.

   Custom OpenAI-compatible still supports existing/custom base URLs and unauthenticated endpoints. Advanced settings retains display name, token limits and vision overrides; unknown IDs do not receive invented capabilities. Leave the key blank on edit to retain it for the same connection. Switching provider or custom endpoint requires a new key, and preset models must be disabled before clearing their key. YAML models remain read-only.

   Connection testing uses the same model factory and protocol as chat, sends a bounded streaming tool-call request and never saves the draft. It may incur provider charges and does not verify every agent, vision or thinking workflow. Official DeepSeek custom endpoints retain their existing adapter; a proxy does not become DeepSeek merely because of its model ID. See the Web model registration reference for presets, compatibility, extension instructions and backup requirements.

   DeepSeek regression tests run offline with the normal backend suite. To verify the real provider explicitly, set DEEPSEEK_TEST_API_KEY in your environment and run from backend/:

   DEER_FLOW_RUN_LIVE_TESTS=1 uv run --no-sync pytest tests/test_managed_deepseek_live.py -q

   These opt-in tests send short requests to DeepSeek and may incur charges; they use temporary state, never save credentials to the deployment catalog, and are skipped in CI. DEEPSEEK_TEST_MODEL optionally selects a different DeepSeek model ID (default: deepseek-flash). The same tests can be run on unfixed and fixed revisions; success is always the expected result.

   YAML models remain read-only in this page and take precedence on name conflicts. Managed models are appended after YAML models; edits apply to new configuration snapshots, while active runs retain their existing snapshot. Disabling a model removes it from future selection/resolution, so update any custom-agent or scheduled task definitions that explicitly reference it before disabling it. Managed models are shared by the deployment, not personal API-key profiles, and remain subject to the existing model authorization policy.

   The encrypted catalog and a generated local encryption key are stored in $DEER_FLOW_HOME/managed-models/ (default .deer-flow/managed-models/). Persist and back up the whole directory, restrict filesystem access, and share it across Gateway workers/replicas that should use the same catalog. The local key is protected by filesystem permissions; encryption does not protect against someone who can read both files. Losing the key requires restoring the backup. Reads and writes fail if the catalog cannot be decrypted, rather than replacing it. This storage is independent of the SQL backend and works with read-only YAML mounts.

   When several models are configured, open either model picker and use the star beside a model to favorite it. Favorites appear first in both the main chat and Side Chat pickers without changing either chat's selected or default model. They are stored for the signed-in user in the current browser, so they do not sync to another browser or device and do not require a startup setting. The compact favorites picker intentionally omits search and only adds favorite ordering to the two-line model list.

   Manual model configuration examples

   models:
     - name: gpt-4o
       display_name: GPT-4o
       use: langchain_openai:ChatOpenAI
       model: gpt-4o
       api_key: $OPENAI_API_KEY
   
     - name: openrouter-gemini-2.5-flash
       display_name: Gemini 2.5 Flash (OpenRouter)
       use: langchain_openai:ChatOpenAI
       model: google/gemini-2.5-flash-preview
       api_key: $OPENROUTER_API_KEY
       base_url: https://openrouter.ai/api/v1
   
     - name: opper-claude-sonnet-4-6
       display_name: Claude Sonnet 4.6 (Opper)
       use: langchain_openai:ChatOpenAI
       model: claude-sonnet-4-6
       api_key: $OPPER_API_KEY
       base_url: https://api.opper.ai/v3/compat
   
     - 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
   
     - name: qwen3-32b-vllm
       display_name: Qwen3 32B (vLLM)
       use: deerflow.models.vllm_provider:VllmChatModel
       model: Qwen/Qwen3-32B
       api_key: $VLLM_API_KEY
       base_url: http://localhost:8000/v1
       supports_thinking: true
       when_thinking_enabled:
         extra_body:
           chat_template_kwargs:
             enable_thinking: true

   OpenRouter and similar OpenAI-compatible gateways should be configured with langchain_openai:ChatOpenAI plus base_url. If you prefer a provider-specific environment variable name, point api_key at that variable explicitly (for example api_key: $OPENROUTER_API_KEY).

   The write_file tool's output-budget hint uses the active model's effective max_tokens. Missing or unusable limits, including YAML .inf, -.inf, .nan, and values too large for the character estimate, omit this hint without failing tool assembly; provider-specific model validation still applies.

   To route OpenAI models through /v1/responses, keep using langchain_openai:ChatOpenAI and set use_responses_api: true with output_version: responses/v1.

   The lead agent also recognizes Responses output-token caps (status: incomplete with incomplete_details.reason: max_output_tokens), including streaming responses. It suppresses tool execution, removes the suppressed calls from later model requests, and preserves text and reasoning. A capped tool-call response includes a visible notice; other incomplete reasons retain their existing handling.

   When thinking is disabled on a legacy profile, blank extra_body, thinking, and nested extra_body.thinking keys in a when_thinking_enabled template are treated as absent during dialect inference. When thinking is enabled, template values are forwarded as configured.

   Models whose provider contract differs from DeerFlow's generic thinking/effort assumptions can declare a per-model mapping-valued reasoning: block (thinking unsupported/optional/required, the accepted effort values with aliases and a default, the payload dialect, and the reasoning-history requirement). The setup wizard's Z.AI GLM-5.3-Flash profile uses it: thinking stays on for every foreground and background call, and the effort selector offers the model's own low/high/max levels. Ollama's existing boolean reasoning: true remains a native provider setting and is forwarded to ChatOllama. When migrating a profile to a custom effort path, remove any old reasoning_effort setting from the profile and thinking templates; configuration validation rejects the leftover key. The chat UI drops a remembered provider-specific effort when switching to a legacy model that does not advertise it. Profiles without the block keep their existing provider behavior. See config.example.yaml for the shape and the equivalent manual configuration.

   For vLLM 0.19.0, use deerflow.models.vllm_provider:VllmChatModel. For Qwen-style reasoning models, DeerFlow toggles reasoning with extra_body.chat_template_kwargs.enable_thinking and preserves vLLM's non-standard reasoning field across multi-turn tool-call conversations. Legacy thinking configs are normalized automatically for backward compatibility, without modifying model defaults or caller-owned extra_body dictionaries; a reused dictionary can change the legacy switch between requests. If the endpoint reports a cumulative usage snapshot on every streaming chunk, set cumulative_stream_usage: true so DeerFlow converts those snapshots into per-chunk deltas; the option is disabled by default and leaves usage unchanged when a stable completion id is unavailable. Reasoning models may also require the server to be started with --reasoning-parser .... If your local vLLM deployment accepts any non-empty API key, you can still set VLLM_API_KEY to a placeholder value.

   When prompt caching is enabled for a model configured with deerflow.models.claude_provider:ClaudeChatModel, DeerFlow preserves thinking and redacted-thinking history without placing cache breakpoints directly on those blocks. Extended-thinking tool follow-ups can keep using prompt caching.

   CLI-backed provider examples:

   models:
     - name: gpt-5.4
       display_name: GPT-5.4 (Codex CLI)
       use: deerflow.models.openai_codex_provider:CodexChatModel
       model: gpt-5.4
       supports_thinking: true
       supports_reasoning_effort: true
   
     - name: claude-sonnet-4.6
       display_name: Claude Sonnet 4.6 (Claude Code OAuth)
       use: deerflow.models.claude_provider:ClaudeChatModel
       model: claude-sonnet-4-6
       max_tokens: 4096
       supports_thinking: true

   ClaudeChatModel normalizes manual extended-thinking budgets when auto_thinking_budget: true (the default). An omitted or null budget gets 80% of max_tokens, with a minimum of 1024 tokens. Explicit budgets must be integers of at least 1024 and, for ordinary thinking, strictly below max_tokens; output limits of 1024 or less fail locally. For supported manual interleaved thinking, configure tools and betas: ["interleaved-thinking-2025-05-14"]: the thinking budget may equal or exceed the positive output limit. See Anthropic's interleaved-thinking rules for supported models. auto_thinking_budget: false bypasses this normalization and validation; disabled and adaptive thinking are unchanged.

   • Codex CLI reads ~/.codex/auth.json
   • Codex function tools preserve explicit strict: true or strict: false in wrapped or flat dictionary definitions. Missing or null settings keep the provider default. bind_tools applies the same conversion to dictionaries and BaseTool schemas.
   • Codex honors tool_choice in bind_tools and direct calls: auto, none, required, or a function name/selection. any and True mean required; None and False keep the provider default. Per-call choices override the bound policy. Bound choices are validated when binding; direct calls and overrides are validated before HTTP. required needs at least one function tool, and named selections must match a supplied function. auto and none also work without tools.
   • Completed Codex responses still return their text and tool calls when token usage is null, omitted, or empty; usage metadata remains unavailable.
   • The Codex model provider returns completed responses without waiting for the SSE connection to close. Failed or incomplete responses report the provider's error or reason; partial output is not returned as a successful answer. Non-object error details are reported as text.
   • Claude Code accepts CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_AUTH_TOKEN, CLAUDE_CODE_CREDENTIALS_PATH, or ~/.claude/.credentials.json
   • CLAUDE_CODE_OAUTH_TOKEN_FILE_DESCRIPTOR accepts a UTF-8 token handoff and reuses it for later model instances in the same process. Undecodable handoffs are skipped so Claude Code can still try its override or default credentials file.
   • CLI credential JSON files accept UTF-8 with or without a BOM, independently of the host locale. make doctor accepts the same files when checking CLI authentication. Invalid text encoding is treated as an unreadable source; Claude Code can still try its default file after an invalid override.
   • ACP agent entries are separate from model providers — if you configure acp_agents.codex, point it at a Codex ACP adapter such as npx -y @zed-industries/codex-acp
   • A bare or absolute ACP agent command is resolved before the agent starts, so npm-installed launchers work on Windows too: npx or mcode is spawned as its npx.cmd/mcode.cmd shim instead of failing with a command-not-found error, and an absolute path such as C:\Users\me\AppData\Roaming\npm\mcode resolves to the mcode.cmd shim beside npm's extensionless shell script instead of failing with WinError 193. The lookup uses the PATH the agent subprocess will actually see (acp_agents.<name>.env.PATH when set, otherwise the Gateway's) and hands the spawn an absolute path, so a relative PATH entry cannot be re-interpreted inside the agent's own workspace. A relative command (for example bin/mcode) is used as configured, and a command that cannot be resolved is still reported with the install guidance below.
   • Each ACP agent's timeout_seconds (default: 1800) is one shared budget for initialization, session creation, and the prompt, starting after the subprocess launches. On timeout, DeerFlow aborts the invocation and closes the subprocess before returning an error. Workspace/MCP preparation and subprocess cleanup are outside this budget. A TimeoutError raised by the SDK before this deadline expires retains its own error message.
   • MiniMax Code speaks ACP directly. Install and authenticate it, then add it as an ACP agent:

   npm install --global @minimax-ai/code
   mcode login

   acp_agents:
     mcode:
       command: mcode
       args: ["acp"]
       description: MiniMax Code for implementation, refactoring, debugging, and repository tasks
       auto_approve_permissions: false

   mcode must be on the Gateway process's PATH; installing it only on the Docker host does not make it available inside the Gateway container. DeerFlow invokes it through invoke_acp_agent in a per-thread ACP workspace and forwards enabled MCP servers. Keep auto_approve_permissions: false for untrusted tasks; enable it only when mcode must edit files or run commands and you trust the task.

   • On macOS, export Claude Code auth explicitly if needed:

   eval "$(python3 scripts/export_claude_code_oauth.py --print-export)"

   The exporter rejects malformed credential containers and non-string or blank access tokens before printing a token, emitting a shell export, or writing a credentials file. Valid tokens are exported unchanged, and file exports preserve the full credential container.

   API keys can also be set manually in .env (recommended) or exported in your shell:

   OPENAI_API_KEY=your-openai-api-key
   TAVILY_API_KEY=your-tavily-api-key

For an explicit backend dotenv file, export DEER_FLOW_ENV_FILE before startup, alongside DEER_FLOW_CONFIG_PATH if needed. For example, from backend/:

DEER_FLOW_ENV_FILE=/srv/deer-flow/stage.env DEER_FLOW_CONFIG_PATH=/srv/deer-flow/stage.yaml make gateway

Relative dotenv paths use the backend process working directory; absolute paths work regardless of that directory. Existing process variables win. An unset selector preserves default dotenv discovery; a specified empty, missing, non-file or unreadable path fails startup. Explicit selection also fails when PYTHON_DOTENV_DISABLED disables dotenv loading. Restart after changing the file. This selects backend dotenv input only: shell launchers, Docker Compose and the frontend retain their own environment loading. Values they already export win. It does not select ENV profiles or isolate databases, storage or tenants. See backend dotenv selection.

Use the table below as a practical starting point when choosing how to run DeerFlow:

• These numbers cover DeerFlow itself. If you also host a local LLM, size that service separately.
• Linux plus Docker is the recommended deployment target for a persistent server. macOS and Windows are best treated as development or evaluation environments.
• If CPU or memory usage stays pinned, reduce concurrent runs first, then move to the next sizing tier.

Requires Docker Desktop / Docker Engine and Docker Compose v2.24+ (docker compose version). Older Compose clients cannot parse the optional env_file syntax in docker/docker-compose.yaml and docker/docker-compose-dev.yaml.

Development (hot-reload, source mounts):

make docker-init    # Pull sandbox image (only once or when image updates)
make docker-start   # Start services (auto-detects sandbox mode from config.yaml)
make docker-logs    # View logs

make docker-start starts provisioner only when config.yaml uses provisioner mode (sandbox.use: deerflow.community.aio_sandbox:AioSandboxProvider with provisioner_url).

Docker builds use the upstream uv registry by default. If you need faster mirrors in restricted networks, export UV_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple and NPM_REGISTRY=https://registry.npmmirror.com before running make docker-init or make docker-start.

Local AIO sandbox control traffic is always direct: loopback/private addresses, single-label cluster hosts, and Docker/Podman internal hostnames do not inherit HTTP_PROXY or HTTPS_PROXY. External sandbox FQDNs and public IPs still honor environment proxy settings.

Backend processes automatically pick up config.yaml changes on the next config access, so model metadata updates do not require a manual restart during development.

Gateway runs use the top-level recursion_limit in config.yaml when an API request does not provide one. The default is 100; valid per-request values take precedence, and max_recursion_limit (default 1000) caps both. Changes apply to the next run without restarting the Gateway. This top-level setting applies to Gateway API runs; IM channel and embedded DeerFlowClient runs retain their own defaults and per-call override paths. The checkpoint storage settings database.checkpoint_channel_mode and database.checkpoint_delta.snapshot_frequency (default 10) are exceptions: both are frozen when the process first builds an agent (including through DeerFlowClient) and require a process restart to change safely.

The optional database.checkpoint_cache section (delta channel mode only) caches materialized checkpoint histories: type is memory (default) or redis, and max_entries: 0 disables the cache. The redis backend is Gateway/async-only; the sync TUI/embedded path supports memory only. The cache is performance-only — results are identical with it disabled — so it is never frozen and workers sharing one checkpoint database may safely run different cache settings. Redis thread cleanup treats a configured key_prefix as literal text, including glob metacharacters, so deletion stays within that namespace and thread.

Production (builds images locally, mounts runtime config and data):

make up     # Build images and start all production services
make down   # Stop and remove containers

Access: http://localhost:2026

make up waits for the Gateway /health endpoint before reporting success. If the Gateway does not become healthy within the startup window, deployment exits non-zero and prints the container status plus recent Gateway logs. The production image starts from its already-built environment and never resolves or installs Python dependencies at container startup.

If make up reports an unwritable runtime home or an unreadable persisted secret after make docker-start, run the printed sudo chown -R <uid>:<gid> '<home>' recovery command and retry. The check honors secret overrides from the shell or .env, accepts readable read-only secret files, and leaves make down available without reading or generating secrets.

For persistent deployments, configure database.backend as sqlite or postgres. The selected backend is shared by the LangGraph checkpointer, LangGraph Store, and DeerFlow application data. The deprecated checkpointer section, when present, overrides the first two for backward compatibility.

Gateway startup automatically repairs the missing run-change schema affecting some existing databases (#5516). The repair preserves run history and existing change positions; downgrading the repair to its predecessor also retains the schema and positions required by that version.

For lightweight single-process event persistence, run_events.backend: jsonl keeps Unicode message content intact, including line and paragraph separators. Existing valid JSONL records remain readable without rewriting the files.

For per-call usage audits, an immediate LLM response replay with populated usage updates both nested usage and top-level token counters, even if initial usage was absent or zero. The original request, model, and status metadata stay intact; flushed events are not rewritten. See the run event contract.

The unified nginx endpoint is same-origin by default and does not emit browser CORS headers. If you run a split-origin or port-forwarded browser client, set GATEWAY_CORS_ORIGINS to comma-separated exact origins such as http://localhost:3000; the Gateway then applies the CORS allowlist and matching CSRF origin checks.

When fine-grained authorization is enabled, Live Browser connections require threads:write as well as ownership of the thread, even when only viewing frames: the same connection can control the browser. Permission checks run when connecting. Restart Gateway after upgrading to disconnect sessions admitted by older code.

Browser login uses HttpOnly session cookies. The login page offers a "keep me signed in" option that extends the browser session when the request is HTTPS (including trusted X-Forwarded-Proto: https) or localhost HTTP. The localhost exception uses the direct request Host and ignores forwarded host headers. Public HTTP deployments, including many temporary sandbox URLs, fall back to session cookies by default. DeerFlow never stores the password in browser storage; the UI may remember only the email address.

OIDC/SSO authorization endpoints may include provider query parameters such as ?prompt=login. DeerFlow preserves those parameters when adding the login request, including for configured authorization_endpoint overrides.

Administrators can suspend and restore accounts from Settings → Users (admin-only; the section is hidden for ordinary users). The page lists each account's email, role, and status and exposes the same enable/disable operation as PATCH /api/v1/admin/users/{id}; the last remaining active admin cannot be disabled, and the Gateway's rejection detail is shown verbatim. A disabled account is rejected by every authentication surface, and its live sessions end on the next request with a login page that states the account has been disabled. Password login for a disabled account deliberately still answers "incorrect email or password" so the state is not disclosed to whoever holds the credentials.

DeerFlow still uses Forwarded / X-Forwarded-* headers to recover the browser-facing scheme and origin behind a proxy. The bundled nginx sets X-Forwarded-Proto, but preserves an upstream HTTPS value and does not overwrite every forwarded header. Configure the outer trusted proxy to replace or strip client-supplied forwarding headers before traffic reaches DeerFlow.

Cancelling a model recovery probe, including while it is queued or waiting to retry, lets the next call check whether the provider has recovered. Cancellation does not count as a provider failure or release another call's active recovery probe.

Model retries honor valid provider Retry-After hints up to 24 hours, including waits longer than the local backoff cap. Malformed, non-finite, or over-limit hints use normal backoff instead. This applies to numeric/date hints in LLM error handling and integer-second hints in Claude's retry path.

With lease heartbeat enabled, a transient RunStore renewal error is retried only until the last confirmed lease expires; the stale worker then cancels local execution and suppresses checkpoint, completion-hook, delivery-receipt, and thread-status finalization. A remote tool side effect already in flight may still be outside local cancellation.

Reconciliation uses an atomic takeover claim that re-checks the lease after candidate selection, so a successful owner renewal wins over orphan recovery and only one reconciler can report a run as recovered. When multiple Gateway workers share the Docker/AIO or E2B sandbox backend, also configure sandbox.ownership.type: redis; E2B uses the leases during background startup and periodic reconciliation so duplicate/orphan cleanup cannot terminate a live peer's sandbox.

See CONTRIBUTING.md for detailed Docker development guide.

Keep config.yaml, .env, and extensions_config.json. Stop the services you currently use, run git pull --ff-only, then start the same mode again. Do not run make config or make docker-init again for a routine source upgrade. If the new version requires configuration changes, run make config-upgrade before restarting. See Operations and Troubleshooting for the commands for each mode.

When rolling out the resume-command idempotency fix across multiple Gateway workers or Pods, route all keyed resume submissions and retries (command.resume with Idempotency-Key, on thread-scoped /runs, /runs/stream, or /runs/wait) only to upgraded workers until every worker serving these endpoints is upgraded. Older workers ignore the private resume identity and compare only input: with input: null, retrying deny can incorrectly reuse an earlier approve run. If the load balancer cannot isolate upgraded workers, pause keyed resume traffic until the rollout finishes, or stop all old workers before restarting on the new version. The additive database migration keeps old run-history readers compatible; it does not make old workers safe for resume admission. An upgraded worker returns 409 when retrying an identity-less legacy resume run, even for the same decision; inspect that run and the current thread state before deciding to submit a new action. See the run API contract.

Custom run stores must explicitly support every supplied idempotency field. Keyed input and resume requests return HTTP 503 before run admission when the store cannot guarantee that support. The built-in Memory and SQL stores support both fields; older custom stores with an explicit idempotency_key parameter continue to accept keyed input requests.

If you prefer running services locally:

Prerequisite: complete the "Configuration" steps above first (make setup). make dev requires a valid config.yaml in the project root. Set DEER_FLOW_PROJECT_ROOT to define that root explicitly, or DEER_FLOW_CONFIG_PATH to point at a specific config file. Runtime state defaults to .deer-flow under the project root and can be moved with DEER_FLOW_HOME; skills default to skills/ under the project root and can be moved with DEER_FLOW_SKILLS_PATH. Run make doctor to verify your setup before starting.

Configuration checks and agent-storage migrations use the installed backend environment through uv run --no-sync --project backend from the repository root. This preserves relative runtime paths and installed optional dependencies; see the setup checks and agent-storage migration.

On Windows, run the local development flow from Git Bash. Native cmd.exe and PowerShell shells are not supported for the bash-based service scripts, and WSL is not guaranteed because some scripts rely on Git for Windows utilities such as cygpath.

The documented root make commands invoke repository .sh files through Bash explicitly. They therefore continue to work from source archives or filesystems that do not preserve POSIX executable bits. When calling a script directly from such a checkout, use bash ./scripts/<name>.sh ....

1. Check prerequisites:

   make check  # Verifies Node.js 22+, pnpm, uv, nginx

   The local make check, make install, make dev, and make start entry points use a direct pnpm executable when available and otherwise fall back to corepack pnpm. With native Windows Python, the shared runner checks pnpm.cmd before the generic pnpm lookup, which follows PATH/PATHEXT and may select an .exe or .bat in the same or an earlier PATH directory. The Corepack fallback likewise checks corepack.cmd before corepack. POSIX Python keeps the generic names first, including when running under MSYS/Cygwin. The runner and diagnostics resolve repository paths absolutely, so these checks work regardless of the caller's current directory. Corepack runs from frontend/, so it honors the packageManager version pinned in frontend/package.json; enabling a global pnpm shim is not required.

   make check decodes tool diagnostics as UTF-8 and preserves Unicode pnpm errors on non-UTF-8 hosts. Invalid output bytes are replaced so the checker can report dependency failures instead of aborting while decoding output.

2. Install dependencies:

   make install  # Install backend + frontend dependencies + pre-commit hooks

   Hook setup calls pre-commit through uv, so uv's tool directory need not be on PATH.

3. (Optional) Pre-pull sandbox image:

   # Recommended if using Docker/Container-based sandbox
   make setup-sandbox

   Reads the configured sandbox image from UTF-8 config.yaml, with or without a leading BOM, using LF or CRLF line endings. On macOS, a successful Apple Container pull completes this step even when Docker is not installed. If Docker is available, its image is also pulled.

4. Start services:

   make dev

5. Access: http://localhost:2026

6. (Optional) Load sample memory data for local review: open Settings > Memory, click Import memory, and select backend/docs/memory-settings-sample.json. The browser imports into the signed-in user's memory.

   To replace memory for every registered user in a disposable review environment:

   cd backend
   uv run python ../scripts/load_memory_sample.py --all-users

   Bulk mode supports SQLite/PostgreSQL user registries, creates timestamped backups under .deer-flow/memory-sample-backups/, and rejects the non-persistent database.backend: memory mode. See backend/docs/MEMORY_SETTINGS_REVIEW.md for the complete review flow.

Local services always use their internal ports (8001, 3000, and 2026). The root .env variable PORT configures only the published Docker ingress; it does not change the Next.js port used by make dev.

DeerFlow runs the agent runtime inside the Gateway API. Development mode enables hot-reload; production mode uses a pre-built frontend.

make start and make start-daemon rebuild the frontend with next build on every run. To reuse the last build instead, pass SKIP_FRONTEND_BUILD=1 (or add --skip-frontend-build when calling ./scripts/serve.sh --prod directly). This is opt-in: it fails fast when frontend/.next has no completed build.

Gateway owns /api/langgraph/* and translates those public LangGraph-compatible paths to its native /api/* routers behind nginx.

Cold agent imports during run creation use a dedicated worker pool, keeping unrelated Gateway requests responsive while the agent stack loads. A failed factory import prevents the run from being admitted.

For a read-only demo without the Gateway, run make build-static from frontend/, then HOSTNAME=127.0.0.1 PORT=3000 node --env-file=.env .next/standalone/server.js from the same directory. The build includes public demo assets and resolves supported demo API reads locally; writes are unavailable. To display the homepage GitHub star count, set GITHUB_OAUTH_TOKEN in frontend/.env before starting Node. The token stays on the server; missing credentials or GitHub failures hide the count. Restart Node after changing the token; no rebuild is needed.

The default make dev topology uses DeerFlow's Gateway-embedded runtime and does not require LangGraph Studio. To inspect and test the registered lead-agent graph with the standalone development server, run the command from backend/ so the CLI discovers langgraph.json:

cd backend
uv run langgraph dev --allow-blocking

The command prints the local API and Studio UI URLs. This in-memory server is for development and testing only. The flag permits DeerFlow's synchronous configuration and graph-factory setup during local Studio requests; it must not be treated as a production-server setting. Local Studio authentication is handled automatically, so the connection does not require custom headers. Use DeerFlow's documented production startup modes or a supported LangSmith deployment for production workloads. Assistant ownership and provenance in this standalone mode are server-owned: Studio can discover registered graphs and the assistants it creates, and normal assistant-version selection remains available. Before the locked local runtime loads its persisted development store, DeerFlow repairs legacy assistant rows and version history so historical client metadata cannot restore server privileges or be discarded by the runtime's startup cleanup. Keep the backend dependencies synchronized with uv sync; this compatibility path requires the declared LangGraph runtime versions and logs a warning if the persisted-store contract no longer matches its expectations. The documented command uses LangGraph's file-based custom-app loader, which is also covered directly by DeerFlow's regression tests.

Standalone runs using if_not_exists="create" retain config and run metadata on the newly created thread, including searchable tags; run metadata takes precedence for duplicate keys. Thread ownership and MCP incarnation remain server-owned, and later runs do not replace the thread's creation metadata.

For workflows that invoke backend/langgraph.json through LangGraph Studio or a direct LangGraph Server, DeerFlow consumes the authenticated identity published by that runtime and uses it for custom-agent configuration/SOUL, user skills and skill policy, uploads, thread data, and memory reads/writes. This keeps authenticated runs out of the shared default filesystem bucket, and the server-owned identity takes precedence over ordinary client-supplied user_id values. External identities such as email addresses are mapped to stable, collision-resistant directory-safe user IDs before accessing DeerFlow storage. The default DeerFlow service topology remains the Gateway-embedded runtime described above.

Gateway runs automatically enforce native delivery for artifacts created or modified under /mnt/user-data/outputs: present_files must present at least one output produced by the current run, and the terminal run.delivery receipt must be durably recorded. Virtual artifact paths are resolved within the same authenticated user and thread scope that produced the output before the output-directory boundary is validated. Runs that do not produce output artifacts keep ordinary conversational behavior.

Thread-scoped runs.wait() calls that finish with status: error report the current run error instead of an earlier answer. The asynchronous Python LangGraph SDK raises by default; pass raise_error=False to inspect the returned status and error.

DeerFlow's built-in custom events are available through both LangGraph streaming interfaces: native clients can continue subscribing to stream_mode="custom", while callback-based integrations can consume the same payloads as on_custom_event records from astream_events(version="v2"). The callback event name matches the payload's type field.

./scripts/deploy.sh supports building and starting separately:

# One-step (build + start)
./scripts/deploy.sh

# Two-step (build once, start later)
./scripts/deploy.sh build       # build all images
./scripts/deploy.sh start       # start pre-built images

# Stop
./scripts/deploy.sh down

DeerFlow supports multiple sandbox execution modes:

• Local Execution (runs sandbox code directly on the host machine)
• Docker Execution (runs sandbox code in isolated Docker containers)
• Docker Execution with Kubernetes (runs sandbox code in Kubernetes pods via provisioner service)

Sandbox references in conversation state are server-owned. External run and thread-state APIs reject caller-supplied sandbox values; when restoring a checkpoint, the runtime resolves the reference against the authenticated user and thread before a tool can reuse it. A missing runtime thread ID raises an error even when the referenced sandbox is cached.

When host Bash is enabled for Local Execution, DeerFlow starts OS detection with uname -s, then uses sw_vers on Darwin. On Linux, it reads host system files such as /etc/os-release only when the active sandbox policy permits it. Host filesystem path checks still apply; after a blocked path, the agent is directed to use a permitted command-only probe or virtual path instead of repeating the rejected command.

For Docker development and production, service startup follows the configured sandbox mode. With AioSandboxProvider, an omitted provisioner_url, an empty string ("" or ''), or YAML null (null, ~, or a blank value), including trailing comments, selects aio: provisioner is not started and the Docker socket overlay is enabled. This grants the gateway host root-equivalent Docker access; see the sandbox security notes. A nonempty URL selects provisioner mode without that overlay. Quoted "null" and "~" are nonempty strings, not YAML nulls; mode detection does not validate URLs. The default local provider never enables the socket overlay.

Local AIO sandbox port allocation stops at TCP port 65535. If the remaining ports from the configured starting port are occupied, allocation reports that no port is available instead of attempting an out-of-range bind.

See the Sandbox Configuration Guide to configure your preferred mode.

Remote directory listings report traversal failures (for example, unreadable directories) as incomplete results, even when no entries were returned. A missing start path is reported separately as “Directory not found.”

The optional Tenki cloud sandbox provider uses Tenki SDK 1.4.0 or newer. Timed-out commands preserve partial output and report Exit Code: 124; unsuccessful health checks cannot reclaim a warm sandbox. Health probes tolerate login-shell output around the ok line, and failures log the sandbox ID and probe output before replacing the sandbox.

The optional OpenSandbox provider mirrors generated output files to local thread storage at turn completion for Gateway previews and downloads; see its synchronization limits and failure behavior. An invalid UTF-8 filename is skipped without blocking other outputs, and same-size content edits are still copied on release. Outputs with case- or Unicode-equivalent file or directory names are skipped before copying, and aliases of existing local paths are rejected, preserving previous complete copies.

BoxLite shutdown rejects late VM registration and keeps its SDK loop open while in-flight acquisitions finish. If they cannot drain within five seconds, shutdown fails with resources still owned and can be retried.

In the chat UI, enable Token Usage → Debug to inspect generic/MCP tool calls. Each Tool details panel starts collapsed and shows the tool name, call ID, input, and received result or explicit error. Large previews are truncated; fields whose names exceed the remaining preview budget are omitted rather than renamed. Structured previews retain complete JSON syntax, including escaped strings and closing delimiters. Array previews stop when the text budget cannot display another element; literal ellipsis values are preserved. Consecutive generated markers at an array's end share one ellipsis indicating an omitted suffix; markers before later values retain their positions. Text results retain their original representation, including large numeric IDs and duplicate JSON keys, without reparsing. Text exceeding the limit is shown as a prefix with a truncation notice; structured objects and arrays are formatted separately. Copy actions copy only the displayed preview. This is a frontend view of data already received by the browser, without an additional secret-redaction layer.

In plan mode, malformed TODO statuses return normal tool-validation errors without aborting token attribution, so the agent can correct the call.

Tool-produced paths and URLs can be retained as short artifact handles across context compaction (tool_artifacts in config.yaml). Handles distinguish separate tool-result occurrences, even when a provider reuses call IDs. Standard MCP image URLs are captured alongside file references, including results from stdio and HTTP servers. Detected file URLs preserve their query strings and fragments. When PII redaction is enabled, model-visible artifact labels follow that policy; internal references stay intact for tool argument resolution. The configured registry limit retains the newest artifacts, while checkpointed processing identities prevent evicted results from being recaptured after restart. Resolution runs before authorization and write-safety checks; unknown or expired handles return an error without executing the tool. Small unknown structured results may be retained as complete JSON up to 4096 UTF-8 bytes; empty or oversized payloads are skipped. Handles are agent-local: task arguments resolve parent handles to concrete references, and delegated reports must return concrete references rather than child-local handles. A truncated model projection reports how many handles are omitted.

DeerFlow supports configurable MCP servers and skills to extend its capabilities. When durable MCP background tasks are enabled, agents can use list_background_tasks(status="failed") or status="input_required" to find failed tasks or tasks awaiting input in the current chat. Other supported statuses are submitted, working, completed, and cancelled; omitting the status preserves existing behavior. The database applies the filter before limiting results to the 20 most recent matching tasks. When combined with active_only=true, both filters apply: active statuses are submitted, working, and input_required, so terminal statuses return an empty list. GET /api/threads/{thread_id}/mcp-tasks supports the same status and active_only filters before its limit (default 50, range 1–100); for example, ?status=failed&limit=20. Unknown statuses return 422, and omitting the filters preserves the existing response. For HTTP/SSE MCP servers, OAuth token flows are supported (client_credentials, refresh_token). Missing, malformed, or out-of-range token response expires_in values use a one-hour default lifetime. This includes lifetimes that cannot be added to the current time without overflowing the expiry timestamp. Durable HTTP/SSE task status and cancellation calls select configured user_auth credentials using the persisted task owner, including after restart; per-request secrets are not retained for background calls. If a request-scoped credential overrides submit authentication, both credentials must authorize access to the same remote task. For stdio MCP servers, per-tool call timeouts can be configured with tool_call_timeout; durable background-task calls honor the same setting for HTTP/SSE servers as well. The persistent stdio session pool evicts only idle sessions at capacity. If every pooled session is executing a tool, a new session request fails with a capacity error and can be retried after a call finishes. For stdio file outputs, a bare filename is linked to a uniquely matching file created or changed by that call. Filenames embedded in unrelated paths, including Windows backslash paths, are left intact. Explicit remote URLs (scheme://...) retain their original query parameters and fragments even when these contain a generated file's name or path, including URLs with backslash-escaped quotes or markup delimiters; local references outside the URL still translate normally. For HTTP/SSE background-task calls, session_init_timeout separately bounds connection setup (including the SSE endpoint event) and MCP initialization together; it stops applying once the tool call begins. Initialization deadline errors identify the server and configured time limit. Ordinary task subagents retain the parent run's captured thread incarnation for MCP calls, including legacy threads, so delegation preserves the same lifecycle scope. MCP tool names are prefixed with <server_name>_ by default to prevent collisions across servers. If a server already namespaces its own tools, set tool_name_prefix: false on that server in extensions_config.json to keep the original names. Disable the prefix only when the resulting names remain unique across all enabled servers. Signed-in users' notification toggle, default model, conversation mode, and reasoning effort are saved to their account and restored on other browsers or after clearing browser storage. Browser notification permission still needs to be granted on each device. Changes retry after network failures; unsent changes survive a reload in the same tab. Concurrent edits to different fields are preserved; for the same field, the last server write wins. Existing unscoped browser preferences are not uploaded automatically because they have no account owner; reselect those settings once after upgrading. Static demos and auth-disabled development keep browser-local settings. Thread-specific model overrides and other display preferences remain local.

In a new chat, the submitted question stays above its streamed reasoning and tool steps while the server creates the conversation and confirms the message.

Markdown and JSON conversation exports from the chat header or sidebar read all persisted history pages, including earlier turns outside the loaded view or compacted model context. A failed history read stops the export instead of downloading a partial transcript. Public demos export their bundled messages.

Capability Center groups plugins by office collaboration, documents and knowledge, search and research, business and data, and development and operations. The directory includes setup references alongside existing MCP configurations and Lark. Recommended integrations and built-in support do not imply an installed or verified connection; the Installed filter shows configured MCP entries and installed Lark only.

Personal MCP connections configured in the web interface are persisted per user. Deployment tools remain shared. Administrators can add, edit, enable, disable and delete shared MCP servers under Platform provided; ordinary users see their status without controls. Personal plugin switches affect only the signed-in user's connections. See connection ownership.

For plugin manifests, adapter registration, and Agent capability selection, see Capability Center integration contract.

DingTalk and WeCom group notifications and HubSpot CRM are bundled configurable plugins. Administrators supply robot credentials or a HubSpot private app token; Agents can then send requested group notifications, list companies, or create contacts. Saving configuration performs no external write. These plugins reuse the existing MCP lifecycle and require no separate plugin service. See the integration contract above for required fields, scopes, and feature boundaries.

Plugin brand icons are bundled locally. When adding or editing one personal MCP plugin, users can upload a PNG, JPG, or WebP image (up to 2 MB), preview it, or restore the default icon. Changes take effect only after Save; custom icons persist across browsers as a normalized 128px PNG in the server entry's display-only presentation.icon metadata. They are not sent to the MCP transport.

Capability Center > Plugins adds, replaces, and deletes one MCP server at a time through targeted mutations that preserve concurrent sibling changes; deletes use a bodyless URL-addressed request. An invalid stdio command on one server no longer blocks toggling another, while enabling that invalid server remains protected by the command allowlist and surfaces the backend validation message in the UI. Targeted updates accept both DeerFlow's type field and the MCP-spec transport field for SSE/HTTP servers. Runtime MCP and skill updates replace extensions_config.json atomically, so an interrupted write cannot leave the shared configuration truncated or partially written. Every Gateway worker or instance that reads the same file picks up a new revision on its next read (the cache checks the file's content signature), so MCP and skill changes made through one replica apply to the others without a restart; a missing, truncated, or invalid revision keeps the previous configuration until a complete one lands, including when the file disappears during a reload. The admin MCP cache reset advances a durable generation marker in the writable config directory. Every Gateway worker mounting that same directory retires its own cached tools and pooled sessions before the next lookup; replicas with independent filesystems are not implicitly covered. If no config path is available, the API reports a process-local reset instead. The parsed extensions configuration and its recorded content digest come from the same read, so a racing edit followed by a timestamp-preserving backup restore cannot leave a different revision cached indefinitely. Changes to the global mcpInterceptors chain rebuild every server's tool wrappers, including when a server configuration is edited at the same time.

extensions_config.json accepts UTF-8 with or without a leading byte-order mark (BOM), including files saved as UTF-8 with BOM by an editor.

When deferred MCP tool discovery is enabled, tool_search supports opt-in literal keyword queries such as keywords:notebook jupyter. Keywords match names and descriptions in any order, ranked by distinct term coverage and then name hits; ties retain catalog order. This mode uses the first 256 characters after the prefix and at most 16 unique whitespace-separated terms, returning up to five tools. Punctuation stays literal. Bare queries and +required ranking keep their existing regular-expression semantics; select:tool_a,tool_b still fetches all exact, case-sensitive name matches.

MCP routing hints can also prefer a specific MCP tool for matching requests without forbidding other tools. When tool_search defers MCP schemas, matching routing metadata can auto-promote up to tool_search.auto_promote_top_k deferred schemas before the model call.

Deferred tool_search regex queries share a 100 ms execution budget across the catalog, including +required ranking. Patterns are limited to 256 characters and searchable names/descriptions to 65,536 characters each. Queries that exceed a limit return a retry hint without promoting partial results; use a simpler pattern, keywords:, or select: with exact names. Oversized catalog fields are reported with the tool name, field and length, and logged with the MCP server name when available; use keywords: or select: because simplifying the pattern cannot shrink a catalog field. Exact selection remains uncapped and bypasses regex limits; the literal keywords: mode also bypasses regex limits. Invalid regex syntax still falls back to a literal substring match.

OpenViking users can register the official Streamable HTTP endpoint at /mcp with an owner-bound USER API key. The native forget tool is exposed for capability parity; deletion is irreversible, so it should be called only after explicit user confirmation. DeerFlow does not enforce that confirmation. This explicit, model-selected MCP tool path can run alongside the separate automatic OpenViking memory backend; it does not replace automatic turn capture or recall. See the OpenViking MCP tools configuration.

The Gateway can adapt an MCP server's ordinary submit / status / cancel tools into durable background tasks. The Agent sees only the configured submit tool and a DeerFlow-local task ID; remote IDs are persisted before the submit call returns, while status and cancel stay internal to the runtime. Polling uses cross-worker leases, exponential retry backoff, scoped MCP sessions, bounded result storage, and restart recovery. A status-tool isError is retained as a bounded diagnostic and retried; servers report a permanent remote-task outcome through a normal structured result with status: "failed". Remote poll hints are finite positive numbers capped at 24 hours, artifact-reference JSON is limited to 64 KiB, and task/server identifiers are validated against their durable SQL column limits before persistence. Input-required and terminal updates wake the current chat through idempotent Agent runs, while list_background_tasks and cancel_background_task let the Agent manage tasks without asking users for remote handles. Current-thread tasks are available through GET /api/threads/{thread_id}/mcp-tasks, its detail endpoint, and POST /api/threads/{thread_id}/mcp-tasks/{task_id}/cancel; when the task runtime actually starts, the Web UI exposes the same safe local view from the chat header with live status refresh, cancellation, and on-demand result, artifact, input-request, status-error, and cancellation-retry details. Default-disabled and memory-backend deployments hide that UI and do not poll the task endpoints. A failed remote cancellation remains queued with backoff, and its latest bounded error and attempt count stay visible in the expanded task card. Enable mcp_tasks in config.yaml, configure task_toolsets with exact raw tool names in extensions_config.json, and use a SQL database backend (sqlite or postgres). Task-enabled server connection, authentication, interceptor, timeout, or binding changes require a Gateway restart so Agent tool discovery and background calls cannot use different configuration versions. input_required is notification-only for now: DeerFlow can display the request but cannot yet submit the user's answer back to the remote task.

For deployment-level HTTP/SSE servers with task_toolsets, discovery, ordinary tools, and background task calls share OAuth token state within one Gateway process, including across tool-cache resets. Rotated refresh tokens stay in memory; they are not written back to configuration or shared across processes. After a restart, the configured refresh token must still be valid.

Notification launch and failed Agent-run deliveries use capped exponential backoff with a visible attempt count and stop after five failed attempts. When a bounded ordinary release exceeds its drain deadline, the service retains ownership until it settles. A permanently rejected target such as a deleted chat is dead-lettered immediately instead of retried forever or recreated. Cancellation endpoints return after durably recording the request; the background service owns the potentially slow remote MCP call and its retry schedule.

Notification runs keep their trusted delivery instruction separate from the framed, untrusted remote event payload. The process-started task runtime—not a hot config read—controls whether the task-management tools are exposed, so changing mcp_tasks requires a Gateway restart. When a skill's allowed-tools policy is active, list_background_tasks and cancel_background_task must be declared explicitly like other business tools. See the MCP Server Guide for detailed instructions.

Security: pass per-request MCP credentials only through config.context.secrets; credentials must never be placed in either run metadata surface (metadata.auth_token or config.metadata.auth_token). See MCP credential migration and cleanup for the supported interceptor flow and the required rotation and retained-copy cleanup when migrating from legacy metadata credentials.

DeerFlow supports receiving tasks from messaging apps. Channels auto-start when configured — no public IP required for any of them.

Cancelling a channel restart discards its pending configuration reload, preserving newer runtime settings applied afterward.

DeerFlow can also expose user-owned IM channel connections in the workspace UI. When channel_connections is enabled, logged-in users can bind Telegram, Slack, Discord, Feishu/Lark, DingTalk, WeChat, WeCom, QQ, or Buzz from the sidebar / Settings > Channels. It reuses the existing outbound channels.* transports, so no public IP or provider callback URL is required. Incoming IM messages then run under the connected DeerFlow user account. See IM Channel Connections for setup and security notes. Buzz backlog paging stops with a warning if timestamp-based queries cannot advance, and channel removal closes both live and backlog subscriptions.

Attachments saved by the shared IM ingestion pipeline or Feishu/DingTalk's embedded downloads keep distinct filenames, including when concurrent uploads choose the same name. The final filename is passed to the agent and used for sandbox sync; an existing conversation file is not overwritten.

Configuration in config.yaml:

Discord's channels.discord.allowed_guilds accepts one positive numeric guild ID (quoted or unquoted) or a YAML list. Unset, null, [], or a blank string allows all guilds. Invalid entries are ignored with a warning; any other configured value yielding no valid ID denies every guild and logs an error. allowed_channels accepts one channel ID (quoted or unquoted) or a YAML list of IDs exempt from mention_only, within allowed guilds. An empty value gives no exemptions, so mention_only applies everywhere when enabled.

Discord inbound attachments have a 25 MiB download limit. Oversized-file notices show the enforced limit so senders can resize or compress the attachment.

channels:
  # LangGraph-compatible Gateway API base URL (default: http://localhost:8001/api)
  langgraph_url: http://localhost:8001/api
  # Gateway API URL (default: http://localhost:8001)
  gateway_url: http://localhost:8001

  # Maximum queued or provider-reserved inbound messages (default: 1000)
  inbound_queue_maxsize: 1000
  # Fixed number of long-lived inbound handler workers (default: 5)
  max_concurrency: 5
  # Seconds to drain accepted work before cancelling active handlers (default: 3)
  shutdown_grace_period_seconds: 3

  # Optional: global session defaults for all mobile channels
  session:
    assistant_id: lead_agent  # or a custom agent name; custom agents are routed via lead_agent + agent_name
    config:
      recursion_limit: 100
    context:
      thinking_enabled: true
      is_plan_mode: false
      subagent_enabled: false

  feishu:
    enabled: true
    app_id: $FEISHU_APP_ID
    app_secret: $FEISHU_APP_SECRET
    # domain: https://open.feishu.cn       # China (default)
    # domain: https://open.larksuite.com   # International

  qq:
    enabled: true
    app_id: $QQ_APP_ID
    client_secret: $QQ_CLIENT_SECRET
    allowed_users: []  # QQ OpenIDs, not QQ account numbers; empty allows all; one ID is one entry

  wecom:
    enabled: true
    bot_id: $WECOM_BOT_ID
    bot_secret: $WECOM_BOT_SECRET
    # Optional: extra host suffixes inbound media downloads may come from, in
    # addition to the built-in qq.com family and WeCom's official COS media
    # host (ww-aibot-img-1258476243.<region>.myqcloud.com); add one here if
    # WeCom rotates to a new COS account or media goes through a proxy
    allowed_media_hosts: []

  slack:
    enabled: true
    bot_token: $SLACK_BOT_TOKEN     # xoxb-...
    app_token: $SLACK_APP_TOKEN     # xapp-... (Socket Mode)
    allowed_users: []               # empty = allow all

  telegram:
    enabled: true
    bot_token: $TELEGRAM_BOT_TOKEN
    # Optional: render final Markdown replies as Telegram Rich Messages.
    rich_messages: false
    allowed_users: []               # numeric user IDs, not @usernames; empty = allow all

  wechat:
    enabled: false
    bot_token: $WECHAT_BOT_TOKEN
    ilink_bot_id: $WECHAT_ILINK_BOT_ID
    qrcode_login_enabled: true      # optional: allow first-time QR bootstrap when bot_token is absent
    allowed_users: []               # iLink user IDs; empty allows all; one ID is one entry
    polling_timeout: 35             # timing values must be positive finite seconds
    polling_retry_delay: 5
    qrcode_poll_interval: 2
    qrcode_poll_timeout: 180
    state_dir: ./.deer-flow/wechat/state
    max_inbound_image_bytes: 20971520
    max_outbound_image_bytes: 20971520
    max_inbound_file_bytes: 52428800
    max_outbound_file_bytes: 52428800
    # Inbound media downloads stream with the caps above and are restricted to
    # these host suffixes (plus *.qq.com and the cdn_base_url host by default)
    allowed_media_hosts: []

    # Optional: per-channel / per-user session settings
    session:
      assistant_id: mobile-agent  # custom agent names are also supported here
      context:
        thinking_enabled: false
      users:
        "123456789":
          assistant_id: vip-agent
          config:
            recursion_limit: 150
          context:
            thinking_enabled: true
            subagent_enabled: true

  dingtalk:
    enabled: true
    client_id: $DINGTALK_CLIENT_ID             # Client ID of your DingTalk application
    client_secret: $DINGTALK_CLIENT_SECRET     # Client Secret of your DingTalk application
    allowed_users: []                          # empty = allow all
    card_template_id: ""                       # Optional: AI Card template ID for streaming typewriter effect

Notes:

• assistant_id: lead_agent calls the default LangGraph assistant directly.
• If assistant_id is set to a custom agent name, DeerFlow still routes through lead_agent and injects that value as agent_name, so the custom agent's SOUL/config takes effect for IM channels.
• IM channel workers call Gateway's LangGraph-compatible API internally and automatically attach process-local internal auth plus the CSRF cookie/header pair required for thread and run creation.
• Inbound work is bounded to inbound_queue_maxsize pending messages plus max_concurrency active workers. When capacity is exhausted, socket/polling providers drop new messages before sending DeerFlow's working acknowledgment and emit a rate-limited warning. Buzz leaves its replay cursor unchanged and reconnects for relay replay; GitHub webhooks return 503, marking the delivery failed for manual/API redelivery. Shutdown closes admission immediately, keeps channel transports available while accepted messages drain for up to shutdown_grace_period_seconds, then cancels and awaits active handlers before closing provider resources; the Gateway's outer timeout can cancel an incomplete shutdown without detaching those resources.
• Telegram shutdown retains ongoing cleanup across a timeout or cancellation so a retry can finish it. Stopping during startup also cancels initialization and closes the partially initialized bot's HTTP clients before retiring its worker loop. If the Telegram loop has already stopped, shutdown warns that download clients may remain open and releases the abandoned cleanup state; later stops do not repeat that warning.
• Slack stop waits up to five seconds for an in-flight Socket Mode connection and its cleanup. If cleanup is still pending, the previous instance stays tracked and a restart can be retried after it finishes. SDK cleanup failures report their original error and remain retryable.
• In Telegram private chats, text, attachment and streaming replies (including "Working on it...") can still be delivered if their referenced user or bot message was deleted, such as after clearing chat history during a run. Group reply-target requirements remain unchanged to preserve forum routing.
• Feishu/Lark now queues rapid follow-up messages per mapped DeerFlow thread_id instead of immediately surfacing the generic busy reply, and topic replies keep a per-message card with a compact source-message preview across queued/running/final patches.
• Streaming IM channels treat backend error events like transport failures and follow the same reply and retry path. The channel logs the error type and message for diagnosis.

Set the corresponding API keys in your .env file:

# Telegram
TELEGRAM_BOT_TOKEN=123456789:ABCdefGHIjklMNOpqrSTUvwxYZ

# Slack
SLACK_BOT_TOKEN=xoxb-...
SLACK_APP_TOKEN=xapp-...

# Feishu / Lark
FEISHU_APP_ID=cli_xxxx
FEISHU_APP_SECRET=your_app_secret

# WeChat iLink
WECHAT_BOT_TOKEN=your_ilink_bot_token
WECHAT_ILINK_BOT_ID=your_ilink_bot_id

# WeCom
WECOM_BOT_ID=your_bot_id
WECOM_BOT_SECRET=your_bot_secret

# DingTalk
DINGTALK_CLIENT_ID=your_client_id
DINGTALK_CLIENT_SECRET=your_client_secret

Telegram Setup

1. Chat with @BotFather, send /newbot, and copy the HTTP API token.
2. Set TELEGRAM_BOT_TOKEN in .env and enable the channel in config.yaml.
3. The bot accepts inbound text, photos, and documents (with or without captions). Hosted Bot API downloads are limited to 20 MB per attachment.

Slack Setup

1. Create a Slack App at api.slack.com/apps → Create New App → From scratch.
2. Under OAuth & Permissions, add Bot Token Scopes: app_mentions:read, chat:write, im:history, im:read, im:write, files:write.
3. Enable Socket Mode → generate an App-Level Token (xapp-…) with connections:write scope.
4. Under Event Subscriptions, subscribe to bot events: app_mention, message.im.
5. Set SLACK_BOT_TOKEN and SLACK_APP_TOKEN in .env and enable the channel in config.yaml.

Feishu / Lark Setup

1. Create an app on Feishu Open Platform → enable Bot capability.
2. Add permissions: im:message, im:message.p2p_msg:readonly, im:resource.
3. Under Events, subscribe to im.message.receive_v1 and select Long Connection mode.
4. Copy the App ID and App Secret. Set FEISHU_APP_ID and FEISHU_APP_SECRET in .env and enable the channel in config.yaml.
5. The bot supports inbound text, image, and file messages. Inbound attachment downloads are limited to 20 MB per attachment.
6. (Optional) Interactive clarification cards: set card_interaction_enabled: true under channels.feishu to let the agent's option-based clarifications render as clickable button cards, preserving any background context before the question. Then, in the developer console, open Callback Configuration (回调配置 — a separate tab from Event Subscriptions), subscribe to card.action.trigger in Long Connection mode, and publish a new app version. A click resolves the clarification exactly like typing the choice; only the user who triggered the clarification can operate its card. Button answers are saved in the conversation history, and subsequent clarifications continue in the same thread.

Scheduled-task result and lifecycle notices can also be pushed to a bound Feishu/Lark user, without a new inbound message. They use the same notification outbox and localized summaries as WeCom. The app must be able to send to that user; see Feishu scheduled-task notifications.

WeChat Setup

1. Enable the wechat channel in config.yaml.
2. Either set WECHAT_BOT_TOKEN in .env, or set qrcode_login_enabled: true for first-time QR bootstrap.
3. When bot_token is absent and QR bootstrap is enabled, watch backend logs for the QR content returned by iLink and complete the binding flow.
4. After the QR flow succeeds, DeerFlow persists the acquired token under state_dir for later restarts.
5. For Docker Compose deployments, keep state_dir on a persistent volume so the get_updates_buf cursor and saved auth state survive restarts.
6. Outbound images/files enforce max_outbound_image_bytes / max_outbound_file_bytes (20 MiB / 50 MiB defaults) while reading, including files that grow after resolution. Oversize reads are rejected before encryption/upload instead of sending a truncated prefix. Non-positive limits disable the corresponding cap.
7. allowed_users takes iLink user IDs. Unset, null, [], or a blank string allows everyone. A single ID is one entry, not a sequence of characters, and an unquoted integer-valued number is stored as that integer's text. A scalar string containing commas or interior whitespace logs a warning but remains one literal ID; use a YAML list for multiple IDs. Invalid entries are ignored with a warning; any other configured value that yields no valid ID denies every user and logs an error. /connect is still accepted before that check, and a denied sender is dropped before inbound media is downloaded.
8. Shutdown waits for in-flight cursor writes. On token expiry, DeerFlow persists the cursor reset and removes the saved token before completing poller cancellation.

WeCom Setup

1. Create a bot on the WeCom AI Bot platform and obtain the bot_id and bot_secret.
2. Enable channels.wecom in config.yaml and fill in bot_id / bot_secret.
3. Set WECOM_BOT_ID and WECOM_BOT_SECRET in .env.
4. Make sure backend dependencies include wecom-aibot-python-sdk. The channel uses a WebSocket long connection and does not require a public callback URL.
5. The current integration supports inbound text, image, and file messages. Final images/files generated by the agent are also sent back to the WeCom conversation.

DingTalk Setup

1. Create a DingTalk application in the DingTalk Developer Console and enable Robot capability.
2. Set the message receiving mode to Stream Mode in the robot configuration page.
3. Copy the Client ID and Client Secret, set DINGTALK_CLIENT_ID and DINGTALK_CLIENT_SECRET in .env, and enable the channel in config.yaml.
4. (Optional) To enable streaming AI Card replies (typewriter effect), create an AI Card template on the DingTalk Card Platform, then set card_template_id in config.yaml to the template ID. You also need to apply for the Card.Streaming.Write and Card.Instance.Write permissions.

DingTalk's Markdown fallback displays tables as labeled blockquotes, preserving escaped pipes (\|) within cells and empty leading or trailing cells so values stay under the correct column labels.

When DeerFlow runs in Docker Compose, IM channels execute inside the gateway container. In that case, do not point channels.langgraph_url or channels.gateway_url at localhost; use container service names such as http://gateway:8001/api and http://gateway:8001, or set DEER_FLOW_CHANNELS_LANGGRAPH_URL and DEER_FLOW_CHANNELS_GATEWAY_URL.

Commands

Once a channel is connected, you can interact with DeerFlow directly from the chat:

Messages without a command prefix are treated as regular chat — DeerFlow creates a thread and responds conversationally.

Agent selection is conversation-scoped: /agent use <name> starts a fresh conversation and pins that Custom Agent in the thread metadata. Existing conversations never switch agents midway, the selection survives a Gateway restart, and opening the IM-created thread in the Web UI continues through the same Custom Agent. Use /agent use lead_agent to return to the default agent in a new conversation.

Model selection is conversation-scoped too: /model <name> pins a model to the current conversation — validated against the caller-visible model list, persisted in the thread metadata so it survives a Gateway restart, and applied from the next message without starting a new conversation. /model shows the effective model and its source, /model default clears the pin, and /models reports the pinned model.

Every Gateway HTTP response carries an X-Trace-Id header. The id is inherited from an inbound X-Trace-Id when the caller sends one and generated otherwise, so a proxy or an upstream service can pin one id across services. It needs no configuration and cannot be turned off.

The same id stays attached to work that outlives the HTTP response: the detached run task, any subagents it delegates to, and the background memory-update threads. It is recorded as deerflow_trace_id on the run record (visible in the runs API), in the thread's checkpoint metadata, and in Langfuse traces. Scheduled tasks, MCP task notification runs, and IM channel messages start outside HTTP and mint their own id per occurrence.

Log records carry that id only when enhanced logging is on:

logging:
  enhance:
    enabled: true   # print trace_id into log records
    format: text    # or json

This is off by default because turning it on changes the log format. logging is restart-required, so edit config.yaml and restart the Gateway. The setting affects log output only — the id, the response header, and the run metadata are unaffected.

deerflow_trace_id is a DeerFlow correlation id: it is not a run id, and it is not a provider's native trace id. It is not a lookup key either — nothing resolves a thread or a run from it; use it to correlate log lines. A deerflow_trace_id sent in a run request's metadata or config.context is ignored and overwritten, so the response header, the logs, and the persisted run can never disagree. To pin a correlation id, send the X-Trace-Id header.

Gateway run history also records one terminal run.delivery receipt per run, including zero-output and crash-recovered runs. The receipt is persisted before the durable terminal run status during normal execution. Orphan recovery first atomically claims an expired lease and then idempotently backfills the receipt, so a stale recovery scan cannot overwrite a live run's detailed delivery facts. Receipt persistence remains best-effort during an event-store outage. Runs that fail checkpoint preflight (or are cancelled while waiting for prior finalization) keep the existing completion-data behavior: they receive the zero-delivery receipt but do not overwrite RunStore completion fields with an empty snapshot.

When tool_progress.enabled is true, the same run event history also records result-quality guard phase changes. It records loop-detection decisions and deferred MCP tool promotions for both the lead agent and ordinary task subagents. Promotion events identify newly promoted deferred-tool names and whether routing metadata or tool_search selected them, without copying the search query, routing keywords, schemas, arguments, results, or catalog hash into the promotion event itself.

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

Add the following to your .env file:

LANGSMITH_TRACING=true
LANGSMITH_ENDPOINT=https://api.smith.langchain.com
LANGSMITH_API_KEY=lsv2_pt_xxxxxxxxxxxxxxxx
LANGSMITH_PROJECT=xxx

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 instance, set LANGFUSE_BASE_URL to your deployment URL.

Trace correlation fields. Every agent run is annotated with Langfuse's reserved trace attributes so the Sessions and Users pages light up automatically:

• session_id = LangGraph thread_id — groups every trace of the same conversation
• user_id = effective user from get_effective_user_id() (falls back to default in no-auth mode)
• trace_name = assistant id (defaults to lead-agent)
• tags = [env:<DEER_FLOW_ENV>, model:<model_name>] (built-in Gateway lead-agent and embedded client runs tag the selected model after default or fallback resolution; custom graph factories may only provide the requested model; the environment tag is omitted when unset)
• metadata.deerflow_trace_id = DeerFlow request correlation id, matching X-Trace-Id when request trace correlation is enabled

These are injected into RunnableConfig.metadata at the graph invocation root for both the gateway path (runtime/runs/worker.py::run_agent) and the embedded path (client.py::DeerFlowClient.stream), so any LangChain-compatible callback can read them. Set DEER_FLOW_ENV (or ENVIRONMENT) to tag traces by deployment environment.

DeerFlow also supports Monocle, an OpenTelemetry-based tracer for agentic applications. It records each run end-to-end: LLM calls, agent steps, and tool and MCP invocations, with their inputs, outputs, timings, and token counts.

Add the following to your .env file:

MONOCLE_TRACING=true
MONOCLE_EXPORTERS=file          # file, console, okahu, s3, blob, gcs (default: file)
OKAHU_API_KEY=okh_xxxxxxxx      # required only for the `okahu` exporter

Each run writes one trace file to .monocle/; open it in the Monocle VS Code extension to inspect the span timeline and token counts. Connect to Okahu, an agent-observability platform, to analyze traces across runs and run trace-based and agentic evaluations (via the okahu exporter).

Traces capture span inputs and outputs verbatim — prompts, tool arguments, and model responses — plus token usage and timings. The file exporter keeps them on local disk and never rotates or cleans them up, so prune .monocle/ periodically; the remote exporters (okahu, s3, blob, gcs) send that same data off-box, so enable only destinations you trust. Monocle is initialized once at Gateway startup: a configuration error (unknown exporter, missing OKAHU_API_KEY) is logged there and tracing stays off until the Gateway restarts.

LangSmith and Langfuse attach as LangChain callbacks, so you can enable both and DeerFlow reports each run to both. If an enabled provider is missing required credentials or fails to initialize, DeerFlow fails fast and names it. Monocle uses a global OpenTelemetry provider rather than a callback; Langfuse shares that provider, so all three can run together. Because both span processors sit on the same shared provider, Monocle's exporters also see Langfuse's spans when both are enabled.

For Docker deployments, tracing is disabled by default. Set LANGSMITH_TRACING=true and LANGSMITH_API_KEY in your .env to enable it.

Existing-run SSE joins are observation-only on GET: supplying action=interrupt|rollback returns 405. Cancellation on this stream route is POST-only and requires the runs:cancel permission. Accordingly, the OpenAPI contract exposes action and wait only on POST; the GET operation exposes only its path parameters.

Non-interactive clients (CI pipelines, scripts, server-to-server integrations) can call the Gateway API with a personal access token (PAT) instead of a browser session. Create one while logged in via POST /api/v1/auth/pats — the raw dfp_... value is shown exactly once; only its SHA-256 digest is stored — then send it as a Bearer credential:

POST /api/threads/search
Authorization: Bearer dfp_...
Content-Type: application/json

{}

Each token runs with its owning user's identity (owner filtering and per-user memory keep working), carries a scope set that can only narrow that user's permissions, and is admitted only to the thread/run lifecycle routes — every other route answers 403 to PAT callers, and a PAT never carries admin capability. Tokens can be listed and revoked at any time; revocation is immediate. PATs require a database backend (SQLite/PostgreSQL). Full reference: API Reference — Personal Access Tokens.

DeerFlow started as a Deep Research framework — and the community ran with it. Since launch, developers have pushed it far beyond research: building data pipelines, generating slide decks, spinning up dashboards, automating content workflows. Things we never anticipated.

That told us something important: DeerFlow wasn't just a research tool. It was a harness — a runtime that gives agents the infrastructure to actually get work done.

So we rebuilt it from scratch.

DeerFlow 2.0 is no longer a framework you wire together. It's a super agent harness — batteries included, fully extensible. Built on LangGraph and LangChain, it ships with everything an agent needs out of the box: a filesystem, memory, skills, sandbox-aware execution, and the ability to plan and spawn sub-agents for complex, multi-step tasks.

Use it as-is. Or tear it apart and make it yours.

Open Capability Center from the workspace sidebar to manage Plugins (MCP servers and Lark/Feishu integration) and Skills. Both catalogs support search; skill cards show concise descriptions with full descriptions in a detail view. Built-in and user-created/imported skills are listed separately. The Community tab supports importing .skill archives into My skills. General preferences remain in Settings.

Administrators can see Could not load warnings in Capability Center > Skills when their own custom skill's SKILL.md contains invalid YAML. Warnings show the package-relative file and, when available, its line and column, with a quoting hint for unquoted colon values. Fix the file on disk and select Reload skills to refresh the list and invalidate skill prompt caches for subsequent runs; active runs retain their existing snapshot. This covers caller-owned custom skills only, excluding legacy/global, public, integration, and external linked packages. These warnings cover YAML syntax errors only: missing frontmatter, non-mapping metadata, missing/invalid name or description, invalid allowed-tools or required-secrets declarations, non-UTF-8 files, and file-read failures (including permission denied) are not reported. A missing warning does not establish that a package is valid or readable. The scope note remains visible even when there are no YAML warnings; check Gateway logs for other load failures. Parsing remains strict and invalid skills stay unavailable to agents.

Configure knowledge_search with either deerflow.community.ragflow.tools:knowledge_search_tool or deerflow.community.lightrag.tools:knowledge_search_tool in config.yaml. Both support attributable retrieval evidence. LightRAG and RAGFlow responses without complete provider locators use the shared DeerFlow source contract; complete RAGFlow responses retain their existing citation snapshots and scope controls. Knowledge citations show the provider, document title, and retrieved excerpt in chat, including after reload and when a subagent performed the retrieval. Missing page numbers and public URLs remain absent; sources without a URL are readable as local evidence. Inline citations remain readable after Markdown sanitization, and supplied page numbers use the selected interface language. New providers use the shared envelope; existing RAGFlow citation histories remain supported. Expand a completed subtask card to inspect its report's forwarded knowledge sources, including when reopening a saved conversation. Shortened mixed-provider task reports remove dangling citation fragments without flattening retained shared-source links. Existing web citations and the text-only Python knowledge_search(query) APIs remain supported. No knowledge management service or page is required. See the knowledge contract guide.

Knowledge scope selection also works with operator-installed read-only Provider extensions. The shared selector shows only the filters a Provider supports: knowledge-base selection requires a catalog and dataset filtering; file selection additionally requires document listing and document filtering. Unsupported or inaccessible filters fail without expanding the search scope. Choosing no knowledge search stops retrieval before loading the Provider. Provider resource IDs are preserved up to 256 Unicode code points; longer IDs are rejected without truncation. RAGFlow keeps its existing configuration and citations, and search-only Providers remain usable without a catalog or management page. See the offline Provider example for extension registration and configuration. WeKnora's adapter is planned for the next phase.

Skills are what make DeerFlow do almost anything.

A standard Agent Skill is a structured capability module — a Markdown file that defines a workflow, best practices, and references to supporting resources. DeerFlow ships with built-in skills for research, report generation, slide creation, web pages, image and video generation, and more. But the real power is extensibility: add your own skills, replace the built-in ones, or combine them into compound workflows.

Skills are loaded progressively — only when the task needs them, not all at once. This keeps the context window lean and makes DeerFlow work well even with token-sensitive models.

When deferred skill discovery is enabled, describe_skill ranks installed skills by bounded, Unicode-normalized intent-term coverage across names and descriptions. Natural multi-term requests can therefore find a relevant skill without requiring one exact phrase, while exact select: and required-name +prefix lookups remain available. Ranked searches use up to 256 characters and return up to five results; exact select: lists are not truncated and return all requested catalog matches.

A skill directory is a package boundary: once DeerFlow finds its SKILL.md, nested SKILL.md files under that package (for example evaluation fixtures) remain supporting data and are not registered as runtime skills. This applies to managed integration packs as well as public and custom skills. Namespace directories without their own SKILL.md can still group nested skills.

Discovery follows operator-managed directory symlinks, but skips links back to an ancestor directory so a cyclic namespace does not repeatedly rescan the same tree. Independent links to the same external skill tree remain supported.

Skill Markdown and bundled text resources use UTF-8. Skill-creator CLI and review utilities read and write text explicitly as UTF-8 so localized skills behave consistently across operating systems.

SKILL.md YAML frontmatter must open and close with --- on its own line (trailing spaces or tabs are allowed). A line such as ---suffix is not a closing delimiter.

Custom skill history preserves Unicode line separators inside saved content and metadata, keeping those revisions readable for history inspection and rollback. Malformed JSON history records are still rejected.

The bundled github-deep-research API client also works without requests, using Python's standard-library HTTP client. Its fallback preserves spaces, Unicode, and reserved characters in search queries and label filters.

Users can explicitly activate an enabled skill for a single turn by starting the request with /skill-name, for example /data-analysis analyze uploads/foo.csv. DeerFlow loads that skill's SKILL.md as hidden current-turn context while leaving the base prompt limited to skill metadata. Slash activation respects disabled skills, custom-agent skill whitelists, and existing channel commands such as /new and /help.

After an answer loads skills, its toolbar includes Skills used. Hover over or click the icon to see the skills and their sources; hovering a skill name underlines it. Select a skill to inspect its SKILL.md in the resizable side panel (a drawer on mobile). The view uses snapshots captured during successful configured read-tool loads or explicit slash activation, so later edits or removal of a skill do not rewrite its history. Repeated loads appear once per run, in first-load order. Range reads and bounded snapshots are labeled as partial; older conversations without captured evidence do not show this menu. Copy returns the captured Markdown, including YAML frontmatter. Package-relative links and images remain readable references rather than navigating away from the conversation. Skill loads performed through other tools, such as shell commands, are not inferred from answer text.

An enabled skill's allowed-tools policy applies only after that skill is explicitly slash-activated or captured in the agent's active skill context after a read_file load. Merely enabling, advertising, or listing a skill in a custom agent or subagent skills allowlist does not reduce that agent's normal toolset; subagents use the same progressive discovery and activation policy as the lead agent. During a slash-activated run, that explicit skill's policy is authoritative: reading another SKILL.md may provide instructions but cannot widen the slash skill's tools. Without slash activation, policies from skills actually loaded into active context retain their union semantics. Once active, the policy filters both model-visible tool schemas and tool execution. Framework discovery tools (tool_search and describe_skill) remain available so an allowed deferred tool or installed skill can still be discovered, but discovery and promotion never grant permission to execute a business tool omitted from allowed-tools. task is not framework-exempt; a restrictive skill must list it explicitly to delegate to a subagent. Per-step policy decisions are internal runtime context and are removed from observable or persisted context copies. Registry failures and an active set with no remaining valid skill fail closed to framework-safe tools; individual stale paths are ignored only when another valid active skill remains. This is best-effort behavioral scoping, not a hard security boundary: loading skill instructions through another tool is not captured, and active-skill entries can be evicted from bounded context.

When you install .skill archives through the Gateway, DeerFlow accepts standard space-separated allowed-tools, optional frontmatter metadata, and the Claude-compatible argument-hint field instead of rejecting otherwise valid external skills. YAML lists remain supported for allowed-tools and preserve exact runtime names. Exact portable spellings such as WebFetch, WebSearch, Glob, Grep, and Read map to DeerFlow's web_fetch, web_search, glob, grep, and read_file tools; lowercase or otherwise unknown scalar names remain unchanged so custom and MCP tools keep their exact runtime spelling. Parenthesized entries such as Bash(tvly *) are tokenized as one literal entry, including spaces, quoted text, and escaped parentheses, but remain inactive because DeerFlow does not inspect tool arguments; declare bash only when the skill may use the full Bash tool.

Disabling a skill also removes it from the sandbox filesystem view, so shell commands and structured file tools follow the same enabled state. Local, Docker/AIO, hostPath provisioner, and newly created E2B sandboxes source /mnt/skills from enabled-only projections that update when public, custom, legacy, or managed integration skills are toggled, edited, created, deleted, or installed. Structured read_file calls (including line ranges and read-before-write checks) use the sandbox provider's mount mapping, so the user identity captured when the sandbox was acquired remains authoritative. Managed integration packages remain shared, while their projected filesystem visibility follows each user's enabled state. Multi-worker Gateways re-read on-disk enable state while rebuilding user projections, so a toggle handled by one worker is honored by another worker's next sandbox acquire. Existing E2B sandboxes retain their creation-time snapshot until they are recreated. PVC-backed provisioner skills keep their configured PVC snapshot/layout for now; dynamic PVC materialization is tracked separately.

# Paths inside the sandbox container
/mnt/skills/public
├── research/SKILL.md
├── report-generation/SKILL.md
├── slide-creation/SKILL.md
├── web-page/SKILL.md
└── image-generation/SKILL.md

/mnt/skills/custom
└── your-custom-skill/SKILL.md      ← yours

/mnt/skills/integrations
└── lark-cli/lark-doc/SKILL.md      ← managed, read-only

The built-in image-generation skill supports Gemini, MiniMax, and OpenAI-compatible Images APIs. Select the latter with IMAGE_GENERATION_PROVIDER=openai, then configure IMAGE_GENERATION_API_KEY, IMAGE_GENERATION_BASE_URL, and IMAGE_GENERATION_MODEL. For a containerized sandbox or an explicitly enabled local host shell, expose these variables through sandbox.environment; sandbox commands intentionally do not inherit API keys from the Gateway process. The local provider uses the resolved configuration values unchanged, and request-scoped secrets override operator values. Local bash output and command errors redact credential values while leaving benign settings such as model names and base URLs readable.

For LocalSandboxProvider, this is a managed tool-path boundary rather than host filesystem isolation. Explicit per-Agent skill policies are accepted only while host bash is disabled (the default), because a host subprocess can address canonical paths without using the provider's virtual-path mappings. Use Docker/AIO, the Kubernetes provisioner, or E2B when the filesystem boundary must remain enforceable alongside shell access.

On Windows, the local sandbox's cmd.exe fallback decodes builtin output using the console output code page, including UTF-8 after chcp 65001, independently of Python's UTF-8 mode. If no console page is available, it uses the system OEM page, with ANSI as a last resort. Native programs can choose a different output encoding; configure those programs to match the console page when needed.

Managed integrations install shared read-only skill packs without mixing them into custom skills. The Lark/Feishu CLI integration is available under Capability Center → Plugins → Lark / Feishu; an administrator installs or upgrades the official lark-* pack once under {DEER_FLOW_HOME}/integrations/skills/lark-cli, and every user discovers that same pack with an independent enabled state. Each user's app configuration and OAuth data remain isolated under {DEER_FLOW_HOME}/users/{user_id}/integrations/lark-cli/{config,data}. These secret directories are restricted to 0700, regular credential files to 0600, and symlinks are rejected. On Windows, waiting for another process's integration file lock retries for up to 15 minutes; other file-lock errors are reported immediately.

After installation, users can click Connect Lark to open a browser authorization link; no terminal authorization is required. The same UI can request additional permission domains such as Calendar, Docs, or Drive, or a specific OAuth scope reported by lark-cli. A cheap status refresh only inspects the local credential tree, so the UI reports Credentials configured (not live-verified) until an explicit browser completion performs live token verification. The action then remains Reconnect Lark so users can replace or extend authorization. If an agent hits missing Lark authorization during a conversation, the managed lark-shared guidance points the user back to the same plugin configuration with /workspace/capabilities?tab=plugins&plugin=lark.

Once configured, Change Lark app lets a user point their DeerFlow account at a different Lark/Feishu app without a reinstall — either by pasting an existing app's App ID / App Secret or by re-registering an app in the browser. Switching is per-user (it never touches another user's credentials), validates the new credentials through the official CLI's live tenant-token probe before replacing the active app, and revokes/removes the previous app's OAuth tokens. A rejected credential change does not supersede an in-progress setup or authorization flow. The previous OAuth data is cleared before the CLI stores the replacement app, so the new file-backed keychain secret remains available during reconnection. DeerFlow then immediately opens browser authorization for the newly bound app so the switch ends in a usable connection.

Installing the Lark skill pack resolves the latest official larksuite/cli release from GitHub and downloads that version's skills at install time, so the Gateway needs outbound internet access for that step (it falls back to a bottom-line pinned version if the release lookup fails). The settings page shows the installed version and, when available, the newest published version so an admin can reinstall to upgrade. Air-gapped deployments can pre-stage the archive and point DEER_FLOW_LARK_CLI_SKILLS_ARCHIVE at the local file. Integrity does not depend on a pinned archive byte hash (GitHub does not guarantee stable source-archive bytes); instead the download is restricted to the official GitHub host, every archive member passes structural safety guards, and a content hash of the effective installed skill tree (including DeerFlow's injected shared guidance) is recorded so content changes are auditable across reinstalls.

When sandbox.use selects the AIO provider, the same install also downloads the official Linux amd64 and arm64 CLI release archives, verifies their published SHA-256 checksums, safely extracts one executable per architecture, and mounts the resulting runtime read-only at /mnt/integrations/lark-cli/runtime. An architecture-selecting launcher in that mount makes lark-cli available in the sandbox PATH. Air-gapped AIO deployments can pre-stage a symlink-free runtime tree containing bin/lark-cli plus both linux-{amd64,arm64}/lark-cli files and set DEER_FLOW_LARK_CLI_SANDBOX_RUNTIME_DIR to that directory.

Sandbox trust boundary: the browser never receives the Lark app secret, but agent conversations run lark-cli inside the sandbox, so the per-user credential directories are mounted into it: config (holding the long-lived appSecret) is mounted read-only, its otherwise empty config/locks subdirectory is over-mounted writable for lark-cli coordination files, and data (refreshable OAuth tokens) is writable. The credential-bearing config and data mounts remain readable by any process the agent runs there, so code reached via prompt injection in a tool result could read them. Treat the sandbox as inside the Lark credential trust boundary until the sidecar credential-broker follow-up removes these mounts from sandbox execution.

For remote/Kubernetes deployments (the provisioner backend), the sandbox lark-cli runtime can instead be supplied by an optional init container that copies the binaries into a shared emptyDir — no install-time GitHub download and no hostPath/PVC runtime mount. Publish the image under docker/lark-cli-init and set LARK_CLI_INIT_IMAGE on the provisioner (with the Helm chart, provisioner.larkCliInitImage / provisioner.larkCliBrokerImage); it stays off (legacy behavior) when unset. The Lark integration status (GET /api/integrations/lark/status) reports sandbox_runtime_mode, sandbox_runtime_probed, and sandbox_runtime_ready. sandbox_runtime_probed marks whether runtime readiness was actually evaluated; responses from older backends may omit the flag, in which case the Settings mutation cache keeps the last probed runtime fields instead of overwriting them with an unevaluated fallback — so the Settings UI shows whether lark-cli will actually be present in the sandbox at chat time, rather than a green status hiding a later command not found.

In Lark broker mode, AIO's fresh Bash runs close inherited pipe stdin. Persistent terminal commands keep their input; the shim ignores terminal stdin. Explicit pipelines, heredocs, and file redirections still supply input normally, and the shim forwards that input only after EOF. A stdin idle timeout aborts without executing the command, and broker execution logs omit argument values. See the broker image guide for timeout settings and image rebuild requirements.

If a trusted operator manages the configured skills directory through an external mount such as MinIO, NFS, or CSI, an administrator can call POST /api/skills/reload after changing files. This invalidates skill prompt caches in the handling Gateway process, waits up to the bounded refresh timeout, and then advances a durable reset marker (.extensions_config.json.skills-cache-reset.json) in the writable config directory. Every Gateway worker or Pod mounting that directory notices the marker within about a second of its next prompt build and rescans the latest files, so one call reaches every replica sharing the volume; the response reports scope: shared_config, or scope: process when no extensions config path is available. Running tasks are unchanged. Skill installs, edits, deletions, rollbacks and enable/disable toggles made through the API publish the same marker, so a change made on one replica takes effect on the others without a restart. A loader-level filesystem failure returns a generic server error and preserves the last successfully loaded process cache rather than publishing an empty catalog. Replicas with independent filesystems are not covered. Direct mount writes bypass the validation, SkillScan, and history applied by DeerFlow's install/edit APIs, so only operator-controlled systems should have write access.

Skill installs and agent-managed skill edits run through SkillScan, a native deterministic safety scanner before the LLM-based skill scanner. Phase 1 runs offline with no Semgrep/OpenGrep dependency, blocks high-confidence CRITICAL findings such as private keys or shell execution, and passes warning findings to the LLM scanner for contextual review. Code files (anything under scripts/, a script suffix such as .py, .sh, or .js, or an extensionless file starting with #!) that are not NUL-free UTF-8 text raise a warning and are still analyzed over a lossy decode, so a single stray byte cannot hide them from CRITICAL checks. The moderation adapter normalizes both plain-text model responses and LangChain Responses API text blocks before parsing the required JSON decision. Python instance-client exfiltration checks follow a minimal same-scope evidence chain: a simple name bound to a known client constructor, optional name-to-name aliases, and an actual outbound method or context-manager use supported by that constructor. Constructor roots must be proven imports; bare canonical-looking names are not inferred as modules. Nested scopes do not inherit client handles and inherit only constructor import aliases that are never rebound in the enclosing scope. Comprehensions, walrus-bearing statements, annotations, complex binding targets, unsupported operations, and ambiguous branch flows produce no finding from this signal; skipped constructs conservatively invalidate every name they may bind so stale client state cannot create a finding. A deterministic work budget or recursion limit reached by this best-effort analysis does not discard findings already collected for the file. Set skill_scan.enabled: false in config.yaml to disable only the deterministic analyzers; safe archive extraction and the LLM scanner still run.

Windows scripts (.bat, .cmd, .ps1, .psm1, .js, .jse, .vbs, .vbe, .wsf), HTML applications (.hta), and scriptlets (.sct) count as code even outside scripts/, regardless of filename case. They receive both SkillScan analysis and the installer's executable-code policy. SkillScan warns about remote downloads piped into common shells, including env launchers, sudo, interpreter paths, and shell line continuations. It distinguishes environment dumps from env commands, including assignments and GNU split-string escapes. Shell parsing and split expansion are bounded; reaching a limit preserves critical findings already detected. Pipes to non-shell tools such as jq and tee do not trigger this warning.

SkillScan treats HTTP/HTTPS scheme spellings and HTTP hostnames case-insensitively and recognizes bracketed IPv6 loopback ([::1]) URLs as local. External IPv6 endpoints still trigger network findings, and cloud-metadata hostname detection is case-insensitive.

For Python credential mappings, SkillScan checks literal values while treating ordinary dictionary keys as labels. A mapping such as tokens = {"access_token": os.getenv("ACCESS_TOKEN")} does not report a hardcoded credential. Keys matching a recognized cloud or API token format are still checked as embedded credentials.

SkillScan blocks embedded legacy and modern sk- API tokens, including OpenAI project, service-account, and admin keys and Anthropic keys, even in Bearer headers without a secret assignment. Token bodies may contain - and _; finding evidence is fully redacted. Placeholder examples do not suppress detection of later credentials in the same file. Scheduled-task notification summaries also redact these key formats before sending them to external IM platforms.

DeerFlow also ships with skill-reviewer, a public skill for read-only skill quality review. It uses the built-in review_skill_package tool to inspect installed skills, local packages, archives, or pasted SKILL.md content without activating the target skill, binding its secrets, executing its scripts, or installing it. The tool returns a compact, tag-neutralized JSON payload to the model context and keeps the full raw review payload in the tool artifact for programmatic consumers. The deterministic review core reuses DeerFlow parsing and SkillScan facts, emits versioned JSON contracts under contracts/skill_review/, and can be run from the backend CLI:

cd backend
uv run python -m deerflow.skills.review.cli ../skills/public/data-analysis --format text --fail-on error --fail-on-incomplete

Public-skill CI waivers are exact, expiring exceptions in .github/skill-review-waivers.v1.json. Because only the trusted base manifest can suppress a finding, a file-changing pull request can be preauthorized safely by first merging a manifest-only change that lists the reviewed future full-file SHA-256 in preapproved_file_sha256s; the file change can then land in a later pull request.

Tools follow the same philosophy. DeerFlow comes with a core toolset — web search, web fetch, rendered web capture, file operations, bash execution — and supports custom tools via MCP servers and Python functions. The bundled DDG, Brave, Tavily, SearXNG, and Serper search providers accept an optional time_range of day, week, month, or year; omitting it preserves existing search behavior. For DDG recency searches, DeerFlow excludes DDGS backends that ignore time limits. Swap anything. Add anything.

For DDG web search and DDG image search, max_results in config.yaml can be a positive integer or an environment-variable reference such as max_results: $DDG_MAX_RESULTS with DDG_MAX_RESULTS=5. The configured value takes precedence over the tool call's max_results argument. Invalid values (for example, abc, an empty string, or 3.5), zero, and negative counts produce a warning and fall back to the default of 5 results. DDG, Brave, and Serper image search share one result-URL screen that omits non-HTTP(S), localhost, cloud-metadata hostnames, and non-public address literals, including multicast and IPv6 spellings embedding private IPv4 addresses. The screen does not resolve DNS; downloaders must validate the resolved address when connecting. A result with neither a safe full image URL nor a safe thumbnail is omitted.

Image result screening also recognizes Unicode hostname separators (。, ．, ｡) as dots, matching IDNA client normalization before classifying local/private hosts. Accepted result URLs retain their original spelling.

Stdio MCP servers can set cwd in extensions_config.json when their entrypoint or data files depend on a specific working directory. The setting applies to both discovery and tool calls; see MCP configuration. Omitted, null, or empty values keep the default working directories.

Tavily web_search also accepts optional include_domains and exclude_domains lists in its config.yaml tool entry to control search sources. Non-empty include_domains uses Tavily's filter mode to restrict results to those domains. These are deployment settings; the model still supplies only query and optional time_range. Omitted filters preserve the existing SDK request; an explicit empty list is forwarded and imposes no restriction of that kind. See the tool configuration example.

Serper web and image search can use another Serper-compatible provider through SERPER_BASE_URL and that provider's SERPER_API_KEY. Each tool can instead set its own base_url and api_key in config.yaml; each request reads them from one configuration snapshot. Use an absolute HTTP(S) base URL with a host, without /search, /images, a query string or a fragment. Invalid overrides return a configuration error before HTTP; blank environment values keep Serper's default unless a tool overrides them. See Serper endpoints for precedence and how the provider receives the API key.

Serper web_search supports deployment-level include_domains and exclude_domains too. It checks returned URL hosts (including subdomains), with exclusion taking precedence. Filters can return fewer results, including zero; there are no refill requests. This selects sources, not factual accuracy or a global URL-access policy. The model arguments and image search are unchanged. See Serper configuration for validation and query-length limits.

Serper web and image search support opt-in transient recovery with max_retries: 1 (default 0, maximum 3) in each tool's configuration. retry_budget_seconds bounds retry scheduling, not active synchronous requests. Budget exhaustion is logged separately while the tool returns the last error. Extra requests may consume quota or incur cost. See Serper retries.

When using Tavily for web_fetch, extracted pages without a title use their URL as the heading; their content remains available to the agent. Chat tool-step titles accept leading blank lines and up to three spaces before a page's first H1 heading. Indented code, including mixed spaces and tabs, is not used as a title; the tool step falls back to the URL. Tavily search and fetch each read api_key from their own tool entry in config.yaml, falling back to TAVILY_API_KEY when omitted. Fetch does not reuse the search entry's key, so search can use a different provider. If you previously configured a shared Tavily key only under web_search, also set it under web_fetch or use TAVILY_API_KEY for both.

For news-focused research, select Webz.io News Search in make setup, or replace the web_search tool's use with deerflow.community.webz.tools:web_search_tool and set WEBZ_API_KEY. An explicit tool api_key takes precedence over the environment variable. An explicit call's max_results overrides the configured default; omission uses configuration or 5 (clamped to 1–100). Invalid configured counts, including booleans and fractional numbers, fall back to 5. Results include the matching passage, title, URL, publication date, and source metadata. The chat search step displays source titles and links, including after reloading a conversation, and skips malformed source entries without losing valid links. returned_results counts the returned page, not all matching news articles. Malformed entries are skipped with a warning while valid entries are retained; invalid responses or pages with no valid entries return an error. Empty result lists remain valid. The tool accepts language, country, source-domain, sentiment, category, and publication-date filters. Its time_range maps to a UTC lower date bound of 1, 7, 30, or 365 days; an explicit published_from overrides that bound. Webz searches recent news rather than the general web: date filters cannot extend the provider's documented 30-day coverage.

Administrators can export their own custom skills from Capability Center → Skills → My skills → View details → Export. Review the file list and declared environment requirements, then choose Download .skill. The archive contains the currently saved skill, including supporting files and empty directories; disabled skills can also be exported. If the skill changes after preview, refresh the file list before downloading. Import the archive on another DeerFlow instance with Install .skill; existing-name conflicts and normal installation security checks still apply.

Account settings, conversations and history outside the skill folder are excluded. Files inside the folder are preserved unchanged, including any credentials an author placed there; filename notices are advisory. Configure dependencies and credentials on the destination. Linked folders/files, hard links, unsupported executable binaries, nested SKILL.md files and nonportable paths cannot be exported. Export supports hosts with descriptor-relative no-follow filesystem APIs (Linux/macOS); unsupported hosts fail explicitly. Limits: 4096 ZIP entries, 64 MiB per file, 100 MiB total content/archive and 1 MiB frontmatter. YAML aliases and excessively complex declarations are not supported. Ordinary script executable semantics are preserved on POSIX import, without restoring special permissions. See the export API contract.

The claude-to-deerflow skill lets you interact with a running DeerFlow instance directly from Claude Code. Send research tasks, check status, manage threads — all without leaving the terminal.

Install the skill:

npx skills add https://github.com/bytedance/deer-flow --skill claude-to-deerflow

Then make sure DeerFlow is running (default at http://localhost:2026) and use the /claude-to-deerflow command in Claude Code.

What you can do:

• Send messages to DeerFlow and get streaming responses
• Choose execution modes: flash (fast), standard, pro (planning), ultra (sub-agents)
• Check DeerFlow health, list models/skills/agents
• Manage threads and conversation history
• Upload files for analysis

Environment variables (optional, for custom endpoints):

DEERFLOW_URL=http://localhost:2026            # Unified proxy base URL
DEERFLOW_GATEWAY_URL=http://localhost:2026    # Gateway API
DEERFLOW_LANGGRAPH_URL=http://localhost:2026/api/langgraph  # LangGraph API

See skills/public/claude-to-deerflow/SKILL.md for the full API reference.

Answers can cite retrieved RAGFlow evidence with clickable knowledge citations. Click a citation, or an entry in the answer's knowledge sources list, to see the original retrieved excerpt, dataset and document names, and page numbers when RAGFlow supplies them. These are retrieval-time snapshots retained with the conversation, including sources forwarded by ordinary task subagents; they remain inspectable after reloading the conversation. An excerpt is not a live copy of the full document: changes in RAGFlow do not rewrite past evidence. Missing source records are shown as unavailable rather than turned into guessed links. Source snapshots do not add a knowledge-management page or expose the RAGFlow API key. Durable batch exports and standalone Markdown files do not carry these interactive conversation source records. Ordinary document-title links in a Sources section open the same evidence as inline citations. When a tool-output budget applies, only complete evidence entries that fit remain citable; omitted sources are reported rather than retaining a source record for a cut-off excerpt.

DeerFlow can optionally connect to a tenant-scoped RAGFlow deployment. The knowledge_search Agent tool resolves the configured dataset scope, groups datasets by embedding model, and retrieves those groups in parallel so mixed embedding models do not cause a provider error. Dataset IDs and API keys are never exposed to the model. The optional list_knowledge_bases tool returns names only.

Main and custom-agent chats can optionally expose a page-local, icon-only Knowledge selector beside the mode control. Its persistent highlight indicates that knowledge retrieval is active; the neutral state means retrieval is off. Set knowledge_base.scope_selection_enabled: true in config.yaml while using the built-in RAGFlow knowledge_search provider to allow all permitted datasets, selected datasets/files, or no retrieval for a turn. The same config flag controls both chat types; when disabled, neither composer shows the selector or submits a scope. The choice resets to the custom agent’s saved default (or all when unbound) when the page is refreshed or another conversation is opened; each sent human message keeps an immutable scope snapshot for replay and history. The Gateway validates every snapshot, intersects it with the operator's dataset allowlist, propagates the execution-only scope to native and durable subagents, and removes it from model inputs and external traces. Client-supplied internal runtime controls and credentials are also stripped from run context before execution or checkpoint persistence. Idempotent retries accept both canonical snapshots and legacy raw run inputs, preserving retry compatibility across upgrades. Custom agents can save a Default knowledge selection from Agents → Agent settings, including optional file filters or retrieval off. Selecting all knowledge bases clears the binding. The same knowledge_scope field is available on agent create/update APIs and in the agent's stored configuration; omitted updates preserve it and null clears it. It is a default, not an authorization boundary: an explicit per-message selection overrides it, and the operator's allowlist still applies at retrieval time. Gateway runs without a message scope (including scheduled and channel turns) use and snapshot the saved default even when the composer selector is hidden. Regenerate/resume retain the original turn's scope, including legacy unscoped turns, rather than picking up later configuration changes. Unknown or unavailable selections never broaden retrieval. Idempotent retries keep the original run and scope when a default is added, changed, or cleared, including unscoped runs accepted before this feature. This default applies to Gateway-hosted custom-agent turns; direct harness/client integrations continue to supply their own execution scope.

In Agent settings, Tool access can inherit every configured tool group, allow selected groups, or disable all groups. Saved groups stay visible and removable while the catalog loads or fails to load. A saved group is marked unavailable only when a successful catalog load confirms it is missing. Saving a selection that removes knowledge also clears the saved default knowledge selection. When enabling knowledge access for an agent that previously lacked it, save the tool access first, then reopen settings to choose its default knowledge bases.

The knowledge_base block is provider-neutral and only controls whether the knowledge capability and selector are enabled. RAGFlow connection, dataset allowlist, and retrieval parameters (base_url, api_key, datasets, page_size, thresholds, and output limits) must be configured on the tools[].name: knowledge_search entry; they are never read from knowledge_base. Custom-agent chat requests carry the selected agent name as both assistant_id and context.agent_name, so Gateway scope admission and runtime agent loading use the same identity. Main chat requests use lead_agent; both identities are admitted only when the shared configuration enables the RAGFlow provider. When answering a pending clarification, an explicitly submitted current selector snapshot wins; clients that omit it inherit the prior turn's accepted scope. Edit-and-regenerate follows the same fallback, and the file catalog is loaded only after a dataset is switched from all files to selected files. This release does not add an independent Knowledge item to the workspace sidebar or a DeerFlow knowledge-management page; create, upload, parse, and delete datasets and documents directly in RAGFlow.

Each message can still select up to 1000 documents. When more than 100 documents are selected from a single dataset, DeerFlow validates them in batches of at most 100 while preserving the complete selection. If any batch contains an inaccessible or non-searchable document, retrieval is rejected.

Advanced deployments can enable pluggable authorization with authorization.enabled in config.yaml. A configured AuthorizationProvider filters denied tools before they reach the model or deferred-tool catalog, then the same provider is checked again before every business-tool execution through the existing guardrail middleware. Gateway threads:* and runs:* route permissions are derived from the same provider, while existing owner checks and admin-only management gates remain in force. Every HTTP route that starts or enables a future Agent run requires runs:create: this includes the stateless POST /api/runs/stream and POST /api/runs/wait endpoints plus scheduled-task create, update, resume, and manual-trigger mutations. Scheduled-task mutations retain their existing threads:write requirement, and the stateless routes separately enforce ownership when the optional thread ID is supplied in the request body. A generated tool_search may bypass the second tool check only when it fronts the current build's already-filtered deferred catalog. Model access follows the same provider: the Gateway models list is filtered per principal, model:use is enforced on model detail requests and again when the runtime resolves the agent's model, and a denied default model falls back to the first remaining candidate that also passes model:use. The built-in RBAC provider supports per-role tools, routes, models, skills, and sandbox allow/deny policies and validates that default_role names a configured role; authorization is disabled by default. Custom roles declared in the provider's roles mapping become reachable by assigning them to a user through the admin API (GET /api/v1/admin/users, PATCH /api/v1/admin/users/{id}; admin-only, last-admin demotion protected) — a declared role nobody holds grants nothing. See config.example.yaml and the authorization RFC.

Follow-up suggestions also check model:use before calling the selected model, including the default model when no name is supplied. A denied model returns HTTP 403 without an LLM call; authorization-provider failures follow the configured fail_closed policy.

For vision-capable agents, view_image and the subsequent model-context image read also require sandbox:execute. Allowing the tool name alone does not grant image-file access; a role denied sandbox execution cannot reread previously recorded image metadata after its permissions change. External run input and thread-state updates cannot set viewed_images or thread_data; when a saved image is read from a host copy, its path must resolve to the current user's thread and the recorded virtual image path. Each model request includes at most four recently viewed images and 28 MiB of base64 image data. Images excluded by these limits are reported as omitted; the agent can call view_image again to bring one back into context.

Advanced deployments can also extend the agent runtime itself by declaring AgentMiddleware classes under extensions.middlewares in config.yaml or extensions_config.json. Each entry is a module.path:ClassName string (zero-argument constructor) or an object {class, kwargs} whose kwargs are passed to the constructor. kwargs values must be JSON types (object, array, string, number, boolean, or null); YAML dates and timestamps are coerced to ISO strings so they match JSON. DeerFlow loads the same configured list into the lead-agent and subagent pipelines after their built-in runtime middlewares and loop/token guards, but before the terminal-response/safety/clarification tail, so enterprise forks can add domain guardrails, tool-call governance, or observability hooks without patching the built-in middleware builders. Missing packages, invalid classes, broken modules, and constructor errors fail loudly at agent creation. Treat config.yaml and extensions_config.json as trusted operator-controlled files: middleware paths are code execution, just like custom tool, model, sandbox, guardrail, MCP server, and MCP interceptor declarations. Gateway skill/MCP toggle endpoints preserve this field but do not expose an API write path for extensions.middlewares. Separate lead-only/subagent-only middleware lists are not supported yet.

For packaged and configurable runtime integrations, use DeerFlow's extension manager. It accepts a Python package requirement, a public HTTPS Git URL, or a local directory, installs the package into the backend's dedicated extensions dependency group, updates backend/uv.lock, and adds an enabled entry to the startup-only top-level plugins: list in config.yaml:

# PyPI — pin a version for a reproducible deployment
make extension-install SOURCE="deerflow-extension-acme==1.2.3"

# Public HTTPS Git — pin an immutable commit
make extension-install \
  SOURCE="git+https://github.com/acme/deerflow-extension-acme.git@0123456789abcdef0123456789abcdef01234567"

# Local package — an absolute path avoids Make's backend-relative working directory
make extension-install SOURCE="$PWD/examples/deerflow-extension-example"

make extension-list
make extension-upgrade SOURCE="$PWD/examples/deerflow-extension-example"
make extension-disable NAME=acme
make extension-enable NAME=acme
make extension-remove NAME=acme

Installation is interactive because package installation can execute Python build hooks, and the loaded extension later runs with Gateway privileges. For an already-reviewed source, automation can acknowledge that boundary explicitly with cd backend && uv run --frozen --no-group extensions deerflow extensions install <source> --yes. The manager requires uv 0.8.0 or newer; the provided Docker images pin uv 0.11.1. The other direct commands are deerflow extensions upgrade SOURCE, list, enable NAME, disable NAME, and remove NAME; NAME may be the extension name, Python distribution, or module:install value. Do not put credentials in a source URL — a URL carrying embedded userinfo or a credential-looking query parameter is rejected before uv runs. Remote Git sources must use public HTTPS; SSH Git URLs are rejected because the stock Docker builder does not forward host SSH credentials. Installing from a loopback URL is allowed for local tooling but warns, because 127.0.0.1 recorded in the lock is a different machine inside the Docker builder.

A managed package declares exactly one standard PEP 621 entry point:

[project.entry-points."deerflow.extensions"]
acme = "acme_deerflow_extension:install"

That callable uses the standalone deerflow-extension-api contract and can register several contribution kinds: isolated middleware at semantic lead/subagent model or tool positions, lead and subagent task-lifecycle hooks, observers for DeerFlow-owned model calls that are not wrapped by middleware model-call hooks (goal, memory, title, and summarization), Gateway-lifetime services, and eager FastAPI HTTP routers. The contract package has no framework dependencies; extensions must declare FastAPI, LangChain, LangGraph, or other libraries they import. The fetched-content screening example contributes one such middleware at the visible tool position. After operator opt-in it classifies the sanitized text a remote tool result shows the model, PII-redacted when pii_redaction is enabled, and adds an advisory warning through a lifecycle state update; it imports only the extension contract.

Full-stack contributions can additionally provide browser pages, conversation actions, authenticated backend operations and model tools through the plugin APIs. The independent bookmarks example demonstrates one package with persistent user data, its own sidebar page and a read-only search tool. Model tools can opt into separate structured artifacts with response_format="content_and_artifact" and PluginToolResult (extension-api 0.2.6). Artifact object keys must be strings at every depth. Existing tools keep their JSON text output. The local retrieval example demonstrates bounded, read-only results and checkpoint replay without an external knowledge service; artifacts do not automatically enable citations. It also appears in the extension catalog with localized installation-guide metadata. Reopening a bookmark resolves the conversation's current agent through the host, so custom-agent conversations retain their original chat entry point, including older bookmarks. The independent Agent teams example lets full Custom Agents collaborate through native @ mentions, shared messages and asynchronous peer requests, with a separate team page and persistent member conversations. Capability Center lists the example with localized installation information even before it is installed. Installation and activation remain deployment-controlled; Capability Center shows plugin information and status. Browser code runs as trusted same-origin code. The browser API and inline BrowserModule.code transport are experimental. The plugin guide describes an additive path to manifests and packaged static resources.

DeerFlow allocates a task-scoped extension store only for middleware, lifecycle, or system-model observation. Services receive app-scoped runtime dependencies after Gateway persistence is ready and stop in reverse order after active runs drain. The optional ExtensionRuntimeDeps.run_evidence_reader is a stable, read-only interface for audit, evaluation, synchronization, and observability services: it discovers changed runs with an opaque resumable cursor, pages a known run's persisted events with after_seq, and reads the authoritative run status separately from event evidence. A database-backed run store keeps the discovery cursor valid across Gateway restarts; the memory run store provides the same ordering only for the current process lifetime. Event metadata is secret-redacted at this boundary, but event content is returned unchanged. Both payloads are detached snapshots, so modifying nested values cannot change the host's stored evidence. The production Gateway supplies an app-scoped, cross-user reader to trusted operator extensions; an embedded host may bind the same adapter to one user. Changed-run pages contain creations and changes to retained rows, not deletion tombstones; consumers that reconcile deletions must poll status for known runs and treat a missing result as absent. Extensions still execute with Gateway privileges and retain the legacy session_factory, so the reader is an API-stability and least-accidental-privilege boundary, not a sandbox for untrusted Python packages. Extension HTTP routers are mounted after every host route; definite shadows and routes entering the host's authentication- or CSRF-exempt paths are rejected with attributed diagnostics, while unrelated routers continue to load. Because the host's public paths are a reserved prefix list that extensions cannot enter, every contributed endpoint requires an authenticated session — there is currently no way for an extension to expose an unauthenticated route, so inbound provider webhooks and public status endpoints are out of scope for this release. Within that, an extension distinguishes an ordinary user from an administrator through deerflow_extension_api.auth: resolve_principal(request) returns the caller, require_admin(request) raises PermissionError for anyone else and fails closed when identity cannot be determined. Extensions receive a projection — user id, admin flag, internal flag, roles — never the host's auth context. Router startup/shutdown hooks, custom lifespans, Mounts, and WebSocket routes are not accepted; lifetime resources belong in ExtensionService, and WebSocket contributions require a future host-owned authentication/Origin wrapper. Lifecycle and system-model callbacks use the Gateway's canonical notification loop, including subagents on isolated loops.

User-facing extension routes can call deerflow_extension_api.require_run_evidence_reader(request) (extension API 0.2.2+). The Gateway requires authenticated runs:read permission and fixes the reader's scope to that user, including administrators and internal callers. Caller-supplied IDs cannot change that scope. Invisible and missing runs return the same result; changed-run cursors cannot be reused directly across users. The optional resolve_run_evidence_reader(request) returns None when the host does not support the capability; the required helper raises NotImplementedError instead. Denied access raises PermissionError. Routes should map these exceptions to HTTP 503 and 403 respectively; resolver failures never fall back to the global service reader.

Extension services can optionally receive a host model invoker (extension API 0.2.4+). Operators explicitly map allowed logical roles under plugins[].host_access.model_invocation; the host handles credentials, concurrency, timeouts, and schema-validated output. Admission is bounded; timed-out provider work keeps its concurrency slot until completion, and schema validation runs in terminable child processes. A provider's own cancellation becomes ModelInvocationFailed, while cancellation of the calling extension task still propagates normally. Without a grant the capability is None.

Plugin order is deterministic, per-plugin configuration is passed to install(), and required: true makes load failure abort startup; otherwise failures are reported and skipped. enabled: false skips resolution and import. The manager preserves the extension's private config when toggling it and writes name, package, use, enabled, and required metadata for managed installs. Installs are recorded required: false so a later broken extension is reported rather than blocking Gateway startup; pass extensions install <source> --required when the package's absence should abort startup instead. Plugins load once when the Gateway app is constructed, so install, enable, disable, remove, and manual plugins: edits all require a Gateway restart. Because this imports Python code, plugins: is intentionally unavailable through the API-writable extensions_config.json.

Management commands bootstrap the checkout environment without the extension group via uv run --frozen --no-group extensions. Frozen mode lets disable and remove start even when an installed extension's remote source or managed snapshot has become unavailable, while a fresh checkout can still create the non-extension environment from the existing lock. The manager itself owns the subsequent locked dependency transaction. Mutations for one checkout are serialized through a process lock. Failed extension operations restore dependency files and configuration atomically with their original file permissions, preserving access for a Gateway that reads them through group permissions. The initial manager surface is create/remove rather than in-place upgrade: to change an installed source, save its private plugins[].config, remove it, reinstall the new pin, and restore that config.

Local-directory installs are copied into backend/extensions/sources/<normalized-distribution>/; this deployable snapshot, rather than the original directory, is recorded in the lock. Git metadata, virtual environments, bytecode caches, symbolic links, and likely credential files are not accepted as snapshot content. Review what you install anyway: filtering accidental files does not sandbox an extension, its build backend, or its runtime code.

Local make dev/make start, Docker development, and the production Gateway image all consume the same backend/pyproject.toml and backend/uv.lock. Local and Docker-dev launchers perform a locked sync before starting; the production image performs that sync during its build and includes managed local snapshots in the build context. Gateway runtime commands then use the already-created environment without resolving or installing packages. Local and Docker-development pre-start syncs may download missing locked artifacts. A production deployment instead downloads them only during the explicit install or image build; starting the resulting production Gateway container never resolves or installs extensions from the network. A local wheel or file:// Git URL is rejected because it would not exist in the Docker build context; pass a source directory to create a managed snapshot instead. Because environment configuration (such as a UV_FIND_LINKS wheelhouse) can still resolve a plain package name to a local wheel, the manager audits every new lock before enabling the extension: any local reference the stock image build cannot reproduce rolls back the entire install or removal. Rebuild with make up after changing the managed extension set. See config.example.yaml and the reference extension for a complete example.

Gateway-generated follow-up suggestions now normalize both plain-string model output and block/list-style rich content before parsing the JSON array response, so provider-specific content wrappers do not silently drop suggestions. Follow-up suggestions and input polishing skip content blocks with non-string types, preserving the surrounding text instead of failing the request

Visible-content detection (has_visible_content, shared by the empty-response, terminal-response and model-length middlewares) and the subagent dispatch snapshot (ParentContextSnapshot.from_state) probe content-block types by equality rather than set membership, so a provider-emitted block with an unhashable type such as type: [] degrades to a skipped block — keeping the surrounding history — instead of raising TypeError out of the middleware chain

Backend response cleanup preserves unrelated tag names such as <think-tank> and <think:note> in follow-up suggestions and polished drafts; only the exact <think> name (optionally followed by attributes) starts a reasoning block. Self-closing <think/> and <think /> tags are empty reasoning blocks and leave the following answer intact. Delimiters inside quoted attributes, such as <think note=">"/>, also leave the following answer intact.

The Web UI composer can polish draft input before sending. The rewrite runs as a short Gateway LLM request using the input_polish model configuration, keeps slash skill prefixes such as /data-analysis, and only replaces the local draft after the user clicks the polish button; it does not create a thread run or persist a message.

When the agent asks for clarification, the Web UI shows the structured response card but keeps the normal composer available. Users can complete the card or send a free-form chat message to bypass it; that message closes the latest pending clarification and becomes the agent's next input. Accompanying answer text stays outside the execution steps panel, and answering the request keeps previously completed text in the conversation while the agent continues.

Unsent Web UI composer drafts survive page reloads and switching between conversations within the same browser tab. Drafts are isolated by user, agent, and conversation, include the selected skill and explicit conversation references, and are cleared once a send is accepted. Project documents already attached to the thread also survive reloads; browser-local uploads and quoted message excerpts are not persisted.

The Web UI composer supports @ references at the cursor or through the @ button. Search enabled skills, documents in the current project, and other conversations in a grouped picker; selected references appear inline alongside the draft text and can be removed with normal editing. Multiple skills can be selected in one message (up to 16 unique skills); selecting a checked skill again removes it, with each selection checked against the current user registry and agent allowlist. Project files are attached only after ingestion succeeds; conversation references require the read_conversation capability and retain its per-run limit. If capability discovery fails, saved conversation references remain intact and sending waits for a successful retry. Polishing preserves complete reference labels when instructions are reordered. The file group also offers local uploads. /goal, /compact, and legacy /skill-name input remain supported.

After picking an inline reference, typing continues immediately after it. Moving the cursor keeps the position you chose, including when editing text beside a reference.

The Web UI composer also supports browser-based voice dictation when the browser exposes the Web Speech API. The microphone button transcribes speech into the local draft only; DeerFlow receives only the transcribed text, while audio handling is delegated to the browser or operating system speech-recognition service according to that environment's policy. Users can review or edit the text before sending. Dictation keeps words separated when browsers trim transcript segments, preserves Chinese and Japanese joins, and attaches leading punctuation and percent signs without adding spaces.

The Web UI displays a localized AI-generated-content disclaimer below the composer in both standard and custom-agent conversations, reminding users to verify important information.

Interrupted first-turn runs still persist a fallback conversation title, so stopping a streaming response does not leave the thread as "Untitled" after refresh.

Streaming Markdown responses animate only newly arrived words; text that is already visible is not faded out and replayed when the next chunk extends the same block.

In the Web UI, completed assistant turns can be branched into a new main conversation. Automatically inherited branch titles use the next free language-neutral numeric suffix (Title (2), then Title (3) for another sibling or a branch of the numbered conversation) so generated sibling titles remain distinct without persisting a locale-specific label; matching explicit or renamed sibling titles also reserve their displayed suffix even though they carry no generated sequence metadata. An explicit API title is preserved. Renaming a branch clears its generated sequence, so its next automatic branch starts from the renamed title at (2). Recent chats also groups loaded branches directly beneath a loaded parent with subtle tree connectors. Missing parents, malformed or cyclic lineage, and branches in a different pinned state stay visible at the top level instead of being hidden or moved across the pinned boundary. The new thread starts from that turn's checkpoint and keeps the preceding replay checkpoint, so the branched response can be regenerated immediately. The latest response can also be regenerated after an interruption, even when its streamed partial text never reached a checkpoint. Regenerating the latest response preserves the thread's current title, including a title you renamed manually after the original response. Legacy or imported histories without checkpoint parent links use a bounded chronological fallback; if no earlier replay checkpoint exists, branching still succeeds with the legacy single-checkpoint shape, while regeneration remains unavailable for that inherited response. Existing single-checkpoint branches are left unchanged rather than attempting an unsafe checkpoint copy. Because workspace files are not checkpointed, the branch only receives a best-effort copy of the current workspace when you branch from the latest turn; branching from an older turn keeps just the restored message history so the branch never inherits files that were created in a later part of the conversation.

The Web UI reports completed task time once per run. This is total wall-clock time—including model reasoning, tool calls, and waiting—not a per-step or model-only thinking duration. Reasoning content remains available through its own separate disclosure.

While a response streams, reasoning-only messages stay in the processing panel, including Anthropic thinking blocks. Once answer content arrives alongside reasoning, it appears in an assistant bubble.

Literal <think> tags in fenced, indented, or inline code remain part of the answer and its copied text, rather than being moved into the reasoning disclosure. This includes fences opened on list-item lines: real reasoning after the code is still extracted. Unfinished inline code spans are preserved while streaming within a paragraph, but end at a blank line or an interrupting heading, list, thematic break or fence. Indented paragraph continuations do not start a code block.

In the Web UI, the latest completed user turn can also be edited and rerun from the message toolbar. DeerFlow restores the conversation checkpoint before that user message, submits the edited text as a new user message, and hides the superseded turn once the replay is in progress or succeeds. This is a conversation-state replay only: files, memory updates, and external tool side effects are not undone.

Web UI chat links percent-encode custom thread identifiers before placing them in route segments, so reserved URL characters such as # and ? do not change which conversation is opened.

Deleting a chat from the sidebar requires confirmation showing its title. Deletion removes the conversation and its files and cannot be undone.

Use Archive chat in a recent chat's sidebar menu to hide completed work while keeping its messages, files, and original link. The success message offers Undo. Open Chats → Archived to find archived conversations and restore them individually; an open archived conversation also shows a restore button in its header. Search filters the titles of loaded conversations, with Load more for older entries.

Archive and restore preserve the chat's activity time and pinned state. Archiving does not stop a running task or pause its schedules, and new activity does not automatically restore it. Use the existing Delete action when you intend to remove a conversation and its files.

Use /goal <completion condition> to attach one active completion condition to the current thread. The goal is thread-scoped state, not a skill activation, so it stays active across turns until DeerFlow determines it has been satisfied or you clear it. The agent sees the active goal on every model call, so it keeps working toward it after context compaction removes the message that stated it.

Supported commands:

/goal finish the implementation and make all tests pass
/goal              # show the active goal
/goal clear        # clear it

After each Gateway-backed run, DeerFlow evaluates the visible conversation, including the assistant's tool calls and shortened tool results, against the active goal with a non-thinking evaluator model. A successful tool result alone does not satisfy a goal, and when the assistant had to guess missing or ambiguous information the evaluator reports needs_user_input. The evaluator must return a typed blocker (missing_evidence, needs_user_input, run_failed, external_wait, or goal_not_met_yet) plus visible evidence. DeerFlow only injects a hidden continuation when the latest assistant turn is durably checkpointed, the blocker is goal_not_met_yet, the thread did not change during evaluation, and the no-progress breaker has not fired. The safety cap defaults to 8 hidden continuations, and repeated identical non-progress evaluations stop after 2 attempts. /goal clear and any user-authored new input win over queued continuations. When the evaluator considers the goal satisfied, DeerFlow clears it after the run successfully finalizes and publishes the updated thread state. The same checkpoint records the completion as goal_outcome, which stays until the next goal change. If required artifact delivery or its receipt persistence fails, the run reports an error and the goal remains active for a later retry; this does not start another hidden continuation.

Regenerating an earlier response keeps the goal currently set for the chat and its used continuation budget. It does not restore a goal that was cleared or achieved. The earlier answer's completion record is removed when its historical checkpoint is replayed.

The Web UI shows the active goal above the composer, says why auto-continue stopped, and shows a "Goal met" notice after a goal is met. Editing the last message is off while a goal is set. The same command is available from the TUI and supported IM channels. In the Web UI and supported IM channels, setting /goal <completion condition> also starts a run with the condition as the task; status and clear commands only manage goal state. Setting or clearing a goal is rejected while that thread has a run in flight, including a run owned by another Gateway worker, so the goal checkpoint cannot branch away from an active run's checkpoint lineage.

When your role lacks runs:create, the Web UI rejects a new task or /goal <completion condition> before preparing the thread or saving the goal, and keeps your draft for retry. Goal status, goal clearing, and /compact remain governed by their own endpoint permissions.

Automatic and manual compaction process all selected history in chronological chunks before replacing it. summarization.trim_tokens_to_summarize bounds raw input per step; generated summaries use at most half to leave room for source. Oversized summaries retain marked head/tail text; raw history is never clipped. A pass is limited to 64 chunks and leaves history unchanged on failure or limit exhaustion. A limit breach starts a five-minute per-process cooldown for the same thread/history window and policy, including rebuilt middleware. A larger input budget or smaller source window permits an immediate retry. Model fallback and SDK retries can add calls. See the summarization guide.

With task_continuity.enabled, history_search searches the current task's active and compacted history. Its optional role accepts user, assistant, or tool and filters before the eight-result limit; omitting it or passing null preserves search across all roles. Search excerpts contain at most 600 characters around the earliest locatable matching occurrence that fits and include excerpt_start / excerpt_end, a zero-based, half-open character range in the original source. Oversized occurrences are skipped in favor of later fitting matches. If no complete matching occurrence can be located within the excerpt, the excerpt falls back to the source opening with excerpt_match=false. Pass the start as the offset to history_read to continue reading. Use history_read to verify the original source; historical user messages do not grant current authorization. See task continuity.

Automatic and manual compaction exclude old todo reminder messages from both the summary input and retained context. The current todo list stays in thread state. In planning mode, if the original write_todos call is no longer visible, DeerFlow adds a reminder using the latest task statuses before the next model call. Skipped or failed compaction leaves the existing messages unchanged.

Optional pii_redaction.enabled redacts detected identifiers in user messages, remote tool results, compaction input, reinjected summaries, configured LLM title input, and the /goal evaluator's input. Memory admission, including pre-compaction flushes, also redacts detected identifiers in supported text/JSON content, parsed and invalid call arguments/error text, provider-raw/legacy function calls and supported user-content provenance. It is off by default; enabled deployments supply a secret for stable, keyed value-derived placeholders. No token-to-raw mapping is persisted. Raw thread text and local fallback titles remain available for display; memory redaction copies messages without changing the caller's history or tool execution. The hidden /goal continuation is redacted before it is stored.

The Web UI preserves persisted message order when merging history with live updates. Streaming steps around a persisted result inside the loaded history stay together, including steps that arrive after the result. Steps captured during compaction also remain visible before their persisted result when history has not refreshed and the UI has not rendered them yet.

Compaction keeps the current user request and summarizes older assistant/tool activity. When rescuing that request leaves an assistant/tool-only summary window, input trimming favors its most recent content. For mixed histories whose user-message anchor falls outside the trimming budget, compaction retains the existing final-message fallback. summarization.trim_tokens_to_summarize (4000 by default) controls trimming of the raw summary input; escaping and prompt formatting add overhead beyond that budget. Setting this option to null disables input trimming for the summary model; choose that only when the model can accept the full history being compacted.

If the same rendered summary prompt and model candidates recur, DeerFlow reuses the existing summary when the first candidate previously returned it unchanged. An unchanged fallback response is not cached, so reuse cannot prevent the primary model from being retried after an invocation failure. Cache failures fall back to normal generation. See summarization telemetry for the summary result and reuse counters.

Use /compact in the Web UI composer to summarize older context for the current thread. DeerFlow keeps the full chat visible, but future model calls use the compacted summary plus recent messages. The command is ignored when there is not enough history to compact, and it is blocked while the thread has a run in flight, including when that run is owned by another Gateway worker. If a multi-worker reservation loses its lease, DeerFlow cancels the checkpoint writer before the replacing run proceeds and returns a retryable conflict after cleanup. Thread-title edits are serialized through the same state-write boundary and show a conflict without closing the rename dialog when a run is active.

Manual compaction also returns HTTP 409 when execution is pending or the checkpoint is unavailable, so it cannot confirm the thread is idle.

The chat header also shows a context-window gauge when the selected model has a positive context_window configured. It estimates the latest materialized checkpoint's message tokens and keeps the previous same-thread percentage visible while data refetches, independently of the cumulative token-usage setting.

When token_budget.enabled and token_usage.enabled are both enabled, each lead-agent run budget includes completed subagent usage, including the final batch. Hidden goal continuations share the budget; a new user run starts fresh.

When a sub-agent ends with return_direct=True tools, including tools contributed by extension middleware, their outputs are returned in tool-call order. A failed tool marks the task as failed while preserving the batch outputs.

Ordinary task calls accept context_mode="isolated" (default) or context_mode="snapshot". Isolated tasks receive their delegated prompt as before. Snapshot tasks also receive the parent's retained conversation and compaction summary, captured at dispatch as historical background. This helps handoffs that depend on earlier requirements or failed approaches, at the cost of additional input tokens. Retained text, tool-call descriptions/results, and JSON-serializable media input blocks are carried over. Binary or otherwise unserializable media blocks become an explicit omission notice; surrounding conversation remains available. Parent system prompts, hidden framework messages (such as injected memory and todo reminders), reasoning blocks, tool execution metadata, and pending tool calls are excluded. Tool-call descriptions require a retained matching result, including calls alongside the current task. Valid hidden user clarification responses remain part of the conversation. The child keeps its own role, model, tools, and skill restrictions. Parent tool records cannot satisfy child execution checks. Parent and child histories evolve independently afterward; shared sandbox/filesystem behavior is unchanged. Snapshot mode does not restore already-compacted messages or promise prompt cache reuse. Durable batch_task items still require self-contained prompts.

Native document blocks are retained in snapshots. Their textual content, titles, and context use the same reserved-tag and input-boundary neutralization as other historical text; encoded media and URL sources are preserved without decoding.

For a manual, synthetic comparison of complete handoffs and snapshots, see the context snapshot evaluation.

Custom Agents support an optional Unicode display name, including Chinese and emoji. Open an agent's Agent settings → Display name to set it (up to 100 Unicode code points), or leave it blank to show the existing identifier. Control characters and bidirectional formatting controls are rejected; ordinary multilingual text and emoji are supported. Invisible-only names and invisible formatting characters such as zero-width spaces are rejected. Invalid display names in older or hand-edited storage fall back to the agent identifier when read; a warning identifies the affected agent. Re-bootstrapping preserves valid display names. The gallery, chat header, and welcome page use this label; URLs and API calls continue to use the stable English name. API callers can pass display_name to agent creation or update requests; an omitted update preserves it and null clears it. The same optional field is supported in the agent's config.yaml.

Saving an unrelated Agent setting preserves its existing thinking and reasoning defaults, including while the model catalog is loading, unavailable, or missing the configured model. Explicit reasoning edits are saved independently; choosing a different known model clears defaults that the new model cannot support. Saving a model switch waits until the selected model is available; unrelated edits remain saveable without the catalog. Incompatible saved reasoning values are shown explicitly so they can be replaced with Inherit or a supported value.

Custom Agents migrated to agent_storage.backend: db remain editable through Agent settings and update_agent while the importer retains their old files for rollback. See Agent Storage for the migration procedure.

Custom Agents can also be moved between DeerFlow users or deployments from the Agents gallery. Export Agent downloads a versioned *.deerflow-agent.json package containing the portable definition and SOUL.md; Import Agent validates that package and lets the receiving user choose a local name. Invalid packages show validation messages so you can correct the package and retry. Imports never overwrite an existing agent. Packages carry model, tool/MCP, skill, knowledge, sub-agent, reasoning, and memory-policy settings, but deliberately exclude memory contents, conversations, credentials, and operator-owned GitHub bindings.

Sub-agents are an optimization, not the default response to a complex request.

After Stop interrupts a delegated task before it returns a reply, the next user turn marks that earlier task as cancelled in the agent's durable context so it can retry. Existing replies are preserved. Older replies without status metadata may still appear in progress; their outcome is not inferred from their text. The durable delegation ledger distinguishes calls by run and provider tool-call ID, so a later user turn can reuse an ID without replacing earlier work or losing that turn's per-run delegation count.

The lead agent can spawn sub-agents on the fly — each with its own scoped context, tools, and termination conditions — when delegation has clear net benefit from real parallel latency, specialist capability, or context isolation. It keeps interdependent scopes and overlapping side effects out of parallel dispatch; a bounded sequential chain can still run in one sub-agent when specialist or context-isolation benefit clearly wins. The lead uses the fewest useful sub-agents and re-evaluates later batches instead of fanning out solely because a task is large or multi-step. Sub-agents report back structured results, and the lead agent verifies and synthesizes them into a coherent output. Deterministic tool receipts cover both direct tool messages and state-updating Command results such as delegated task responses; when the receipt ledger reaches its context budget, it retains the newest actions and their original receipt IDs. Operators can disable this provenance layer with verification.receipts_enabled: false. Their configured skills are resolved from the same user-scoped catalog as the lead agent, so user-owned custom skills remain available without exposing another user's version. Their internal AI and tool messages stay scoped to the delegated graph instead of entering the parent chat stream. Reloaded thread history enforces the same boundary: callback-captured sub-agent AI responses remain available in run-event diagnostics but are excluded from the parent transcript, while the parent task result remains attached to its subtask card. Long-running sub-agents compact older history when summarization is enabled and re-inject the summary as guarded, hidden durable context before continuing, so recent assistant/tool activity remains grounded in the task. Their system instructions, including the role and report contract, survive compaction; if only those instructions and the current request would be summarized, compaction is skipped. Provider/model request failures are reported as failed sub-agent tasks rather than successful results, so the lead agent and Web UI can react to them correctly. Concurrent parent runs also receive independent server-side sub-agent execution IDs, so a provider that reuses a tool-call ID cannot make one run poll, cancel, or clean up another run's background task. Collapsed sub-agent cards show the effective model and, when the provider returns usage metadata, a cumulative token total that updates after each completed sub-agent LLM call and persists after a reload. When token usage tracking is enabled, completed sub-agent usage is attributed back to the dispatching step from that run's terminal tool-message metadata rather than a process-global provider-ID cache.

Sub-agent cards keep unrecognized structured task statuses in progress. Accompanying result text does not turn an unknown status into success or failure.

Cancelled or timed-out background sub-agent executions retain provider-reported token usage from completed model calls, including responses received before their next progress update. Final usage delivery to the parent run does not count earlier progress snapshots twice.

For tests_passed:go test ./..., packages marked [no test files] or [no tests to run] do not veto a passing summary from another package. Runs with only zero-test package summaries remain UNVERIFIED; failures still take precedence. Under -v, a record whose per-case lines are all --- SKIP also remains UNVERIFIED: Go exits 0 and prints the same ok banner, so nothing ran to pass. Non-verbose output cannot separate an all-skipped package from a passing one and stays trusted as recorded. This checks recorded execution evidence, not claim correctness.

For tests_passed:cargo test, a complete passing target summary can coexist with empty unit, integration or documentation targets. Complete zero-pass successful summaries do not veto that sibling evidence. All-empty or all-ignored runs and bare test result: ok banners remain UNVERIFIED; incomplete zero-test summaries are kept as vetoes, and failures still take precedence.

For tests_passed:python -m unittest, a complete Ran N tests in ...s / OK summary must include at least one ordinary passing test. Skips and expected failures do not count as passes. Successful runs with all-skipped, all-expected-failure, empty or incomplete summaries remain UNVERIFIED; nonzero exits, failures and unexpected successes still take precedence.

For tests_passed:node --test, complete TAP or spec counter summaries must contain an actual pass and no failed or cancelled tests. Individual TAP ok points do not prove the run passed. Empty, all-skipped, all-todo, incomplete or inconsistent summaries remain UNVERIFIED; execution and shell-provenance checks still apply.

For Maven and Gradle test acceptance, a BUILD SUCCESS / BUILD SUCCESSFUL banner alone remains UNVERIFIED: builds can succeed without running tests. Complete Tests run: N, Failures: F, Errors: E, Skipped: S or N tests completed summaries establish a hold only when at least one test ran without a reported failure. Zero-test or all-skipped modules do not supply positive evidence; failures in any module take precedence over passing counts.

For file acceptance criteria, an empty regular file in the shared workspace can satisfy file:<path> exists and file_written:<path>, including on remote sandboxes. It fails file:<path> non-empty with a deterministic empty-file result.

To require CSV header columns, set a task or batch_task item's criterion to file:../outputs/report.csv csv-columns:["id","name"]. The JSON array must contain unique, nonempty, UTF-8-encodable column names. Names match exactly (case and whitespace included), in any order; extra columns are allowed. Checks read only a bounded prefix and parse the first comma-separated logical record, supporting double-quoted fields, doubled quotes, embedded newlines, and an optional UTF-8 BOM. Data rows are not validated. Empty files, missing required columns, malformed or invalid UTF-8 headers, and empty or duplicate header names return does not hold. Headers over 50,000 bytes, incomplete reads, invalid criteria, unavailable workspace access, or exhausted read budgets return UNVERIFIED. A large file can satisfy the check when its complete header fits the budget. A definite syntax error within that header remains a checked negative regardless of later data rows; truncated quoting remains UNVERIFIED. Reads use the same authorized, scoped transport as JSON checks below: 50,001 bytes maximum; remote reads have a 5-second command timeout (10 seconds for the provider call). Local reads and parsing share a cooperative 5-second deadline; remote parsing has its own 5-second deadline starting after the read returns. These checks do not interrupt an in-progress filesystem or parser call. Concurrent writes are not an atomic snapshot. This does not validate data types, row counts, or business meaning.

To request JSON syntax validation, explicitly set a task or batch_task item's acceptance_criteria to ["file:../outputs/report.json json-valid"]. Checks cover only complete UTF-8 JSON files within the shared workspace, up to 50,000 bytes. Valid syntax returns holds; empty files, syntax errors, non-UTF-8 content, or NaN/Infinity return does not hold. Oversize files, incomplete reads, out-of-scope paths, or parser resource limits return UNVERIFIED. Reads are capped at 50,001 bytes to detect oversize content. Remote results are checked for a completion marker, the read exit code, and the probed size. Remote providers lacking the required probe tools leave the result unverified rather than falling back to full-content reads. Concurrent writes of the same size do not yield an atomic snapshot. Local checks authorize sandbox access before resolving paths or probing metadata, and recheck before reading content; revoked permissions return UNVERIFIED without revealing existence. Remote probes report missing files only with a searchable ancestor and in-scope canonical path; inaccessible directories remain UNVERIFIED. UTF-8 BOMs are rejected. Top-level scalars, duplicate keys, and syntactically valid large numbers can pass, without validating schemas, fields, or business semantics. Other file criteria are unchanged, and .json files are not checked automatically. This criterion standardizes verdicts, path boundaries, and read limits instead of requiring ad hoc bash/Python checks, and makes insufficient evidence explicit. Execution completion status and automatic retry policy are unchanged.

Content-less sub-agent final messages report No response generated instead of the literal text None. A content-less provider-error fallback reports its structured error detail when available.

An ordinary task also receives a defensive snapshot of the dispatching run's current uploads. This lets eligible sub-agents use list_uploaded_files to find earlier-turn files without returning same-turn attachments as historical. Delayed or recovered batch_task workers leave this tool disabled because they have no valid turn-local upload boundary.

Embedded Python uploads use the same captured content for the saved document and its Markdown conversion, even if the caller's source changes during upload.

历史上传发现支持稳定续页：list_uploaded_files 默认每页 20 项、最多 100 项， 将返回的 next_cursor 作为下一次调用的 cursor，保留相同过滤条件，直到末页不再 返回游标。目录元数据或调用上下文变化时会明确要求重新枚举；详情见 文件上传文档。

Durable batch_task workers use one app-owned plugin snapshot for tool assembly and execution. Recovered tasks adopt the new worker's plugin snapshot after a Gateway restart; plugin objects are never stored in durable task records.

Ordinary task delegation and explicit durable batch_task execution share the startup-scoped subagent_runtime process capacity. Batch mode keeps large independent item sets in SQL with separate total, live, and running limits, restart recovery, bounded results, and a thread-scoped Web UI panel. The panel pages through bounded previews on demand; bulk stored results are available through the owner-scoped JSONL export, while internal execution and authorization context never enters owner-facing responses. If the batch worker is later stopped or disabled, threads with persisted batches retain read-only item inspection and JSONL export; execution controls remain disabled until the worker is running again. See config.example.yaml and the implementation contract for limits and recovery semantics.

While durable batch tools are available, a later owner turn can explicitly inspect selected results with read_batch_result(batch_id, position=0, offset=0, max_chars=4000). The reader is restricted to the current user and thread, including in explicitly bound SDK graphs. It returns one window of a JSON document containing the stored report, execution/error state and separate acceptance criteria/verdict, without starting or waiting for work. Concatenate content windows before parsing: pass next_offset and the returned revision as expected_revision until next_offset is null. A changed item returns restart_required; discard prior windows and start at zero. Use next_position for a selected next item within batch_status.total_items. The maximum window is 8,192 characters; metadata shares the window budget, and JSON response framing adds overhead. result_truncated means storage already capped the report; continuation cannot recover discarded content. A succeeded item is not an accepted task, and unchecked criteria remain UNVERIFIED. No automatic result ingestion, waiting loop or UI change is added.

The complete escaped response also fits within 10,000 characters and the bound host's positive tool_output externalization/fallback limits, so normal budgeting preserves continuation JSON. Smaller limits can shorten a window; budget_too_small means the operator must raise the budget instead of retrying a zero-length page. Limits too small even for that error envelope cannot provide a usable structured response. Explicit SDK graphs with a custom output-budget middleware should supply the same AppConfig through SubagentRuntime; default graphs use default reader limits.

Offline acceptance: cd backend && PYTHONPATH=. uv run --locked pytest tests/test_batch_result_reader.py -k 'real_submission_worker or sdk_factory_graph' -q. This exercises real submission, deterministic worker completion, database reopen and registered graph consumption; external model execution is replaced by a fake model/executor.

Durable knowledge-research results also export nullable result_artifact.knowledge_sources snapshots. Each retained source pairs a citation ID in the stored report with the exact excerpt captured by knowledge_search or a delegated task, plus its RAGFlow dataset/document/chunk locator and pages. The existing authenticated GET /api/threads/{thread_id}/subagent-batches/{batch_id}/results.jsonl remains the entry point: find a report's #knowledge-<id> link in that row's sources by id to inspect the evidence after a Gateway restart or provider change. Compact item queries and the batch panel do not include these snapshots; this does not add clickable citations to batch previews.

Snapshots preserve whole excerpts, not freshly fetched provider content. They inherit the 100-source/1,000,000-text-character forwarding limits; the serialized snapshot has a separate character budget equal to subagent_batches.max_result_chars, in addition to the report's existing budget. knowledge_sources.omitted_count reports referenced records omitted by forwarding, snapshot projection checks or snapshot budgeting; malformed or unsupported producer artifacts filtered before candidate selection are not counted. Report truncation limits selection to complete citations in the stored text. Old/plain results have no snapshot (null); no historical evidence is reconstructed. Gateway startup adds a nullable JSON column through migration 0033_batch_result_artifact. Back up the database before migrating; downgrading drops evidence snapshots while preserving report text. Cancelled, failed and stale attempts do not publish evidence, and explicit retries clear previous snapshots.

Explicitly enabled nested task delegation propagates omitted opaque source IDs in its captured artifact. Each outer report filters and deduplicates those IDs, and independently captured evidence can restore an omitted source. The durable snapshot publishes only the final count. Legacy task artifacts without omission IDs cannot reconstruct earlier forwarding losses; unknown prose citations are not treated as captured evidence.

If an item's final lease expires after exhausting the retry budget, the batch reaches a terminal state once all items are terminal: failed when none succeeded, or completed when at least one succeeded. The item remains failed in both cases so partial results stay visible.

Direct create_deerflow_agent(...) integrations can own the same boundary explicitly instead of relying on Gateway startup. Construct one SubagentRuntime and share it across every graph in that application; its max_running, ordinary per-run total, bound task tool, and optional durable-batch tools then use the same caller-owned snapshot and execution controller. A runtime with a batch repository owns a worker and must be started before graph construction and stopped during application shutdown:

from deerflow.agents import RuntimeFeatures, create_deerflow_agent
from deerflow.subagents import SubagentRuntime

runtime = SubagentRuntime.from_app_config(app_config, batch_repository=batch_repository)
async with runtime:
    graph = create_deerflow_agent(
        model,
        features=RuntimeFeatures(subagent=True),
        subagent_runtime=runtime,
    )
    # Serve or invoke graph while the durable worker is running.

SubagentRuntime.stop() waits for its owned batch service to finish before propagating caller cancellation, including repeated cancellation. This drain has no timeout; repository operations and child cleanup must terminate. If service shutdown also fails or is cancelled, the first caller cancellation is preserved with the service failure as its cause.

Gateway run finalization also retains ownership through caller cancellation: it finishes terminal persistence, stream completion, and resource cleanup before propagating the interruption. Backend I/O must terminate or enforce its own timeout. Gateway shutdown includes workers still finalizing after their terminal status commits; an expired shutdown drain budget is logged and does not cancel their outstanding writes.

The factory still does not load YAML or create SQL infrastructure: the caller supplies the config snapshot, repository, and lifecycle. Because it accepts a caller-owned system_prompt, direct integrations also own any model-visible wording about those limits; the default middleware enforces the runtime limits regardless. The factory does not mount the Gateway owner-scoped HTTP routes or Web UI, so direct applications must expose their own result API/UI if they need those surfaces. For ordinary delegation only, SubagentRuntime(...) needs no asynchronous startup.

Administrators can add, edit, disable, and delete reusable worker definitions from Settings → Subagents. Built-in and config.yaml definitions remain visible there as read-only entries. The default Lead Agent can use every enabled runtime sub-agent; each page-created Custom Agent can instead allow all, none, or a selected set. That selection is enforced both in the model-visible directory and by the server-side task tool. Managed definitions are deployment-wide in this version and follow agent_storage.backend: atomic files for a local deployment or the shared application database for multiple instances.

For example, independent read-only research can run concurrently when the wall-clock savings outweigh duplicated discovery and synthesis cost, while a repository refactor with shared files and sequential test feedback remains with the lead agent. When max_concurrent_subagents is 1, parallel and multi-batch routing guidance is disabled; delegation remains available only for material specialist or context-isolation benefit.

AIO sandboxes recycle after an uncertain implicit-shell outcome or session creation; a confirmed hard_timeout remains eligible for warm reuse. The Gateway waits for all execution and upload holders to finish, including command-session cleanup. An in-flight session creation alone does not trigger recycling or interrupt a concurrent run. Before recycling, the Gateway records the container ID or Pod UID under {DEER_FLOW_HOME}/sandbox-quarantine. These records survive failed stops and Gateway restarts. Gateways sharing AIO containers must share this home in addition to their ownership store. A new runtime instance can reuse the thread's sandbox ID; an unverifiable generation fails closed against existing records. Before recreating a quarantined ID, the Gateway retires its old records only after the backend confirms that no container or Pod remains, while holding the normal teardown fences. This also permits recovery when a runtime cannot report a generation. If a partial deletion leaves a container or a Pod behind, the Gateway inspects it independently of health checks and Service availability, then retries cleanup only for a quarantined generation. A replacement generation keeps its resources. A delete acknowledgement or failed health check is insufficient to clear the records. Recovery also clears leftover local proxy and network resources before recreation.

Lark broker provisioning requires confirmed capabilities. Probe failures cannot select credential mounts or overwrite a known broker requirement. Update the Gateway and provisioner together: the provisioner reports each Pod's actual broker mode, and incompatible Pods are replaced through the normal ownership fences before reuse. Lark commands use that admitted mode. If an older provisioner omits a Pod's broker mode, the Gateway refuses reuse with an explicit upgrade error and preserves the Pod. An omitted mode is not evidence that the existing runtime needs replacement. Capability probes share one request per provisioner and briefly back off after failure. A Pod mode that contradicts the cache triggers a fresh observation before replacement. Failed admission after creation uses ownership-fenced cleanup. A create request whose lark provisioning mode conflicts with the provisioner's current configuration is marked as a capability refresh; the Gateway drops its cached observation so the next acquire re-probes instead of repeating the failure until the cache entry expires. On sandboxes without an attested broker mode, lark-cli commands fail with an explicit unverified-mode error instead of running against an unauthenticated profile.

Host-externalized tool outputs use the Gateway's normal file-creation umask. Host and sandbox outputs use deterministic filenames hashed from the raw tool call ID and output content. Missing, colliding, or oversized IDs cannot overwrite different output or make the filename too long; identical output can reuse its path. Host and remote sandbox persistence failures, including invalid Unicode during filename hashing, use the configured inline fallback. tool_output character/count budgets, including per-tool overrides, require non-negative integers; YAML booleans are rejected rather than treated as 0 or 1. An explicit zero per-tool override disables externalization while preserving any positive global fallback limit. Structured JSON tool results follow these budgets, including results mixed with media; externalization keeps the full JSON available through a file reference. MindIE renders JSON tool results as escaped text, with configured tool exemptions preserved. Enabled PII redaction scans nested JSON keys and values before rendering. Mounted sandboxes running under another UID need read access through the shared storage permissions. An unclean shutdown can leave .tool-output-*.tmp files in tool_output.storage_subdir (default .tool-results) under thread outputs. Remove leftovers during thread-data maintenance with all Gateway writers stopped, or when deleting the corresponding inactive thread's data.

E2BSandboxProvider uses wait as its default overflow policy. It waits for acquire_timeout, then fails the agent turn. DeerFlow does not retry the turn automatically. Clients can use the structured error to schedule a retry.

Use burst with burst_limit to permit bounded extra VMs. The wait and reject policies use only replicas. The reject policy can remove one warm VM before it returns an error.

With in-memory ownership, replicas limits one Gateway process. With Redis ownership, E2B shares one capacity Hash between workers using the same sandbox.ownership.key_prefix; replicas (plus a configured burst) is then a deployment-wide hard limit. Use one unique prefix and the same effective limit per deployment. To change the limit, stop its Gateways, delete the capacity Hash, and restart; mismatched workers fail closed.

The Hash counts remote VMs and in-flight creates, repairs interrupted creates from E2B metadata, grace-protects stale inventory omissions, and blocks new creates while Redis or initial inventory is unavailable. Run Redis with persistence, non-evicting memory, and HA.

E2B reconciliation renews active VMs with a positive timeout covering its configured cadence, even when idle_timeout is zero or shorter than that cadence. Warm VMs retain the configured idle-timeout behavior. Expired local warm entries do not directly free shared capacity: remote inventory must confirm disappearance through the existing grace period first. Active renewal completes before release sets the warm timeout, so a concurrent maintenance pass cannot extend an idle VM's lifetime. Warm-entry cleanup also preserves ownership acquired by a new request during the sweep. Ownership heartbeats remain independent of slow E2B timeout requests, preventing control-plane delays from making active sandbox leases expire.

E2B snapshots skills.container_path when the provider starts and includes the canonical root in its thread identity, warm-pool seed, and remote metadata. A VM created for a different root is never adopted; reconciliation reaps it after the configured grace period once no live peer owns it. Restart the Gateway after changing the root.

E2B acquisition uses a bounded executor. Waiting acquisitions do not use the default asyncio executor.

When an E2B sandbox expires between conversation turns, its replacement restores the same user's saved thread files from workspace, uploads, outputs, and the ACP workspace before tools can use it. Live sandbox reuse does not overwrite remote files with older saved copies. Restoration failure or an exceeded upload limit rejects the new sandbox. Transient failures may succeed on a later attempt. Capacity breaches report the resource tree and exceeded cap: 100 MiB per file, 2,000 files / 512 MiB per tree, and 2,000 files / 512 MiB across the creation pass. An operator must archive or remove unneeded files from the Gateway's persisted thread directories to fit those limits before retrying; retrying unchanged data cannot resolve a capacity breach. Deleting only the remote copies is insufficient. Other restoration errors identify the resource tree and cause category; Gateway logs retain the underlying exception for diagnosis. Recovery requires the Gateway's persisted thread directories to remain available; files that never synchronized before the VM disappeared cannot be recovered. The existing release sync saves workspace/output files but does not mirror file deletions, so saved copies of remotely deleted files can reappear on recreation.

Each E2B mount upload pass accepts at most 512 MiB and 2,000 files. The pass also has a cooperative 120-second deadline. Thread restoration runs first and shares these limits with skill projections and configured mounts. The provider checks the deadline before each mount and during directory preflight. The deadline stops new file uploads after it expires. It does not interrupt active filesystem or E2B SDK calls.

An E2B VM keeps its slot until E2B confirms destruction. This rule covers create and reclaim operations. Discovery can find a VM from another Gateway. Shutdown closes an unowned discovery client without destroying its VM. Release stops counting a transition when the VM enters the warm pool. Shutdown races retry remote cleanup after a transient kill failure. Reset destroys tracked active and warm E2B VMs. The old provider instance cannot accept new acquisitions.

DeerFlow doesn't just talk about doing things. It has its own computer.

Each task gets its own execution environment with a full filesystem view — skills, workspace, uploads, outputs. The agent reads, writes, and edits files. It can view images and, when configured safely, execute shell commands.

The read-before-write gate ties each read mark to that read_file call's result, including custom tools returning multi-message Command updates. An unrelated result cannot authorize a write after a failed read or hide a successful read.

Loop warnings let agents save collected results before their final answer. If the save is blocked by the read-before-write gate, the agent may read the target file and retry the save once. A looping or stopped tool must be replaced by an available alternative; if no permitted save path exists or the retry fails, the final answer includes the collected results and reports that the file was not saved. Recovery calls still count toward the loop limits, and a hard stop prevents further tool execution.

Ranged read_file calls count lines the same way on every sandbox provider: a line ends only at a newline. On E2B, BoxLite, Tenki, and OpenSandbox, a file with bare carriage returns (such as a saved progress log), form feeds, or Unicode line separators therefore returns the same lines as on the local sandbox, and the start_line a truncated read suggests points at the next unread line.

Concurrent reads and writes to the same file share a gate across synchronous and asynchronous tool calls. Async callers waiting for that gate do not occupy worker threads needed to finish the current read or write. Cancelling a waiting call leaves the current operation running; a call that already started file inspection still waits for that work to finish before releasing its gate.

The built-in grep tool searches either one text file or all matching text files below a directory, so an agent can search an uploaded document directly without first broadening the request to the entire uploads directory. E2B's glob filter preserves spaces, quotes, and dollar signs in filename patterns, while wildcard matching and root-relative directory scoping remain unchanged. E2B grep also preserves colons in file and directory names when reporting matching paths and line numbers. glob and grep preserve non-LF separator characters, such as vertical tabs and Unicode line separators, inside returned paths on E2B, BoxLite, Tenki, and OpenSandbox; grep also preserves them inside matched text.

Remote ls excludes ignored descendants before applying its 500-entry listing limit, so dependency and build trees do not crowd out visible files. Explicitly listing an ignored directory still lists its contents; normal depth and output limits remain in effect.

AIO directory listings discard missing shell sessions so the next request can recover. After a dropped connection, directory listings and persistent shell commands report an unknown outcome without replaying the operation; later calls use a fresh session.

Current-upload context tolerates malformed attachment metadata: invalid filename references and files that cannot be checked are skipped individually, while valid attachments in the same message remain available. Unusable or display-overflowing file sizes use the existing zero-size display fallback without dropping the file. Probe errors are logged at debug level. Sizes of 1 TB or more that can be rendered safely display as >= 1 TB to keep the prompt compact; stored byte counts remain unchanged.

Uploaded Markdown outlines recognize ATX heading syntax, clean closing markers with a linear suffix scan, and skip HTML comment blocks and fenced and indented code examples, so hashtags and code comments do not crowd out real document sections from the agent's heading preview. Indented bold examples are also excluded; PDF-style bold headings with up to three leading spaces remain supported. Split-bold numeric table rows, including parenthesized years, signed values, and currency-prefixed amounts, are excluded when any block after the section number is a numeric column, so they do not consume the outline's heading budget. UTF-8 Markdown files with or without a byte-order mark (BOM) produce the same outlines and fallback previews, with original line numbers preserved. Outline titles are limited to 200 characters and fallback previews to 2,000 characters per file, with truncation markers. Full uploaded files remain available for targeted reads.

Converted upload outlines and previews require a matching source version, including modification timestamps. A detected source-version change invalidates its previous conversion, including equal-length edits. Older ownership records without source timestamps are also rejected; re-upload the source with uploads.auto_convert_documents: true to restore conversion-backed outlines. Files remain available, and unvalidated Markdown conversions appear as standalone files in the agent's historical listing. On Windows, st_ctime_ns may represent creation time; a same-size rewrite that restores the original mtime can evade validation. Timestamp checks are conservative metadata validation, not a guarantee of content equality.

Sandbox image downloads enforce the 20 MiB vision limit while reading, including model-context rereads. Oversized streams stop at the first over-limit chunk. Custom sandbox providers must accept and enforce the download_file(..., max_bytes=...) keyword; older implementations fail the image read instead of falling back to an unrestricted download.

Image bytes loaded for a vision-model call are transient: DeerFlow removes the hidden base64 message after the model consumes it so later checkpoints do not keep duplicating that payload.

After each run, DeerFlow records a workspace change summary for the run-owned workspace and outputs directories. The Web UI shows a compact "files changed" badge on the assistant turn; opening it reveals created, modified, and deleted files with text diffs when safe to display. Uploads are excluded because they are user inputs, not agent-generated changes, and stdio MCP temporary/debug files under the DeerFlow-owned .mcp/ namespace are excluded because they are process-internal state (like .git/ and node_modules/, any directory named .mcp is excluded at any depth). Large, binary, or sensitive-looking files are shown as metadata only.

Files presented through present_files remain part of the thread's artifact state, and the Web UI restores the artifact panel and selected document after a page refresh. When a completed response successfully presents between 2 and 50 files, its final file card also offers one ZIP download. Archive membership comes from the terminal delivery receipt rather than browser-supplied paths, and the ZIP contains the current file versions, which may have changed since the response. The currently selected formal artifact is refreshed once when the run finishes so edits become visible without a manual reload. Existing UTF-8 text artifacts under /mnt/user-data/outputs can also be edited and explicitly saved from the panel on Unix and Windows while the thread is idle; saves use content revisions to prevent overwriting agent changes. Source previews also recognize extensionless Dockerfile and Makefile artifacts by their file names. Unknown file types, including names such as constructor and __proto__, retain the download fallback.

Unsaved artifact edits survive switching chats, workspace pages, or visiting the home page and returning in the same browser tab. Drafts stay in memory for the current account until saved or discarded; changing accounts or signing out clears them. Public showcase visits and unavailable Gateway auth probes retain hidden drafts until authentication is known. Recovering as another account or confirming logout clears the retained content. Refreshing or closing the tab warns while any chat has unsaved edits. Draft file contents are not stored in browser storage and do not survive a reload.

Artifact content revisions refresh when an output file is atomically replaced, even if its size and modification time are preserved. Refreshing the preview then provides the new revision for saving; an older preview still requires a reload. Regular files over the 2 MiB editing limit use file identity and change metadata for range validators without hashing the whole file. Conditional byte ranges for regular files require a matching ETag; date-form If-Range requests receive the full current file. Saving also bounds the existing-file read to 2 MiB plus one detection byte, so a file that grows or is replaced after the size check is rejected without loading the entire oversized file into memory.

CSV and TSV artifacts open as tables in the artifact panel and in a separate window. The preview preserves text values (including leading zeros), supports an optional header row, and pages through up to 200 rows and 50 columns from the initial sample. Long or multiline cells can be opened and copied in full. Switch to source to inspect or edit the file; downloads and separate windows use the saved version.

HTML artifact previews resolve relative assets from the artifact directory unless the document contains a real <base> element. Base-tag examples in comments or script text do not change that resolution.

With authentication enabled, static resources in the same thread's artifact route (images, responsive images, classic scripts and CSS, including nested imports) are fetched by the parent and inlined without weakening the opaque iframe sandbox. Preparation has resource, byte, CSS-depth and time limits and reports partial previews. Dynamic JavaScript requests, ES module imports, nested iframes and navigation remain subject to sandbox restrictions; see the frontend preview notes for the supported scope and limits. Optional resource failures, including 401 responses, keep the conversation open with a partial-preview notice. HTML doctypes and doctype-less rendering modes are preserved, and CSS references inside inline SVG styles are prepared too.

Markdown report previews in the artifact panel and separate window resolve local images from the report's own directory and conversation. Both relative paths such as ./charts/summary.png and absolute /mnt/... paths are supported; external image URLs keep their existing behavior.

If the sample cuts a CRLF line ending in half, the preview keeps the earlier complete rows and omits the incomplete final record, including when its last field is quoted.

Text artifacts are streamed with HTTP byte-range support. The Web UI initially loads at most 1 MiB, shows the preview size when a file is larger, and waits for an explicit Load full file action before fetching the remainder or mounting the full code editor. Editing becomes available once the complete file has loaded; unsaved-change detection uses the full content, including when reverting edits or deleting the portion beyond the initial preview. If a reload fails or returns only a preview, users can still exit editing while retaining their unsaved draft. Active HTML, XHTML, and SVG artifacts remain forced downloads at the Gateway boundary.

Artifact previews and downloads preserve literal percent sequences in file names: report%20final.md and report final.md remain distinct files. Markdown links continue to accept URL-encoded paths.

The write_file guidance reflects the active model's output token limit, including custom-agent and thinking-mode overrides. For longer documents, the agent is guided to write sections with append=True; models without a known limit receive no numeric budget hint.

With AioSandboxProvider, shell execution runs inside isolated containers. With LocalSandboxProvider, file tools still map to per-thread directories on the host, but host bash is disabled by default because it is not a secure isolation boundary. Re-enable host bash only for fully trusted local workflows. Host bash commands have a wall-clock timeout, and long-lived processes should be started in the background with output redirected to a workspace log. On Windows, Git Bash/MSYS argument-conversion exclusions are limited to safe non-root virtual path prefixes, so host-native CLI launchers retain their normal MSYS compatibility. When the local sandbox falls back to PowerShell, it captures output as UTF-8 so CJK text does not depend on the Gateway host locale.

Docker AIO sandboxes default to their existing open egress behavior for compatibility. Operators can set sandbox.network.mode to isolated or allowlist; allowlist mode supports operator-defined domains and an interactive Human Input card for temporary or sandbox-lifetime HTTP(S) approval. Private, loopback, link-local, multicast, and cloud metadata addresses remain unapprovable. Denied hostnames are rejected before DNS resolution. Runs in scheduled, webhook, or autonomous interaction mode auto-deny without opening a card. These unattended runs proceed with minimal assumptions only for low-risk, reversible work; high-risk or irreversible work without sufficient authorization returns a structured BLOCKED result naming the missing decision, even if the model attempts to ask for clarification. The trusted sidecar uses a dedicated per-sandbox egress bridge rather than Docker's shared default bridge, and rejects ambiguous HTTP field names before forwarding. See Sandbox configuration for runtime requirements and the complete policy model.

AioSandboxProvider normally detects thread-data mounts from its backend: local containers use the mounted gateway directories, while remote/provisioner sandboxes receive uploaded files through explicit synchronization. Deployments where both sides are guaranteed to share the same thread user-data directories can set sandbox.thread_data_mounts: true to skip that per-upload sandbox acquire and sync. Leave the field unset for automatic detection; setting it incorrectly can make uploaded files unavailable inside the sandbox.

Failed upload commits report the original error even if temporary-file cleanup also fails, for example because of a Windows sharing violation. Once the file is published, a temporary-file cleanup failure is logged without failing the upload; hidden staging files are left for the startup sweep. If that sweep cannot scan an uploads directory, it logs a warning and preserves its files while continuing cleanup elsewhere. A directory that disappears during cleanup is skipped quietly.

Uploads, new skill support files, and new local sandbox paths reject Windows reserved device names on every platform, including COM¹, LPT², the console aliases CONIN$ and CONOUT$, and names with extensions such as com³.txt. ASCII spaces before the extension do not make a device name portable: NUL .txt and COM1 .log are rejected too. Ordinary names such as report .txt are preserved unchanged. Rename these files before creating or uploading them so the same file tree remains usable on Windows.

Uploaded filenames matching .upload-*.part are rejected because that pattern is reserved for temporary staging files. Rename such a file before uploading it. The restriction includes Windows aliases with trailing dots or spaces and case variants, such as .upload-notes.part., .upload-notes.part , and .UPLOAD-NOTES.PART, on every host. The HTTP check recognizes both / and \ as path separators, including on Linux, when extracting the basename. The endpoint returns 400 with a rename hint before publishing any file in a batch containing a reserved name, so the chat reports the upload error instead of continuing without the attachment. The SDK validates filenames for the whole batch before copying any files. New project document names follow the same restriction, whether defaulted from the source or explicitly supplied on upload or promotion to the shelf; an older shelf document with a reserved name remains downloadable but cannot be attached directly. Download it, rename it, and upload it to the thread. This change does not recover or migrate older thread uploads that already match the staging pattern.

Upload listings remain available during concurrent cleanup: removed files are omitted, and if the upload directory disappears or is replaced by a file, the listing returns any entries already collected. Permission and other operational errors still surface.

This is the difference between a chatbot with tool access and an agent with an actual execution environment.

# Paths inside the sandbox container
/mnt/user-data/
├── uploads/          ← your files
├── workspace/        ← agents' working directory
└── outputs/          ← final deliverables

Browser screenshots and Browserless captures publish complete content atomically. New files use the host's normal creation permissions (typically 0644), and replacements preserve existing permissions, so sandbox code can read captures when those permissions allow it.

Browserless reserves capture names with empty placeholders to prevent concurrent captures from overwriting one another. Failed publication releases the placeholder; if cleanup fails, Gateway logs a warning with its path and preserves the original publication error or interruption.

Browser dependency auto-detection accepts name, group, and use in any order within a tool entry, with either indented or indentless YAML lists.

Reading a page is not the same as using one. Alongside the read-only web_fetch and web_capture tools, DeerFlow ships an optional agentic browser tool group that keeps a live, per-conversation browser session so the agent can actually operate a page — navigate, read the interactive elements, click, type, submit forms, and follow multi-step flows on JavaScript-heavy sites.

Each action returns a fresh snapshot of the page's interactive elements, each addressed by a stable [ref] number, so the agent acts on what it just observed instead of guessing selectors. Outbound URLs are SSRF-screened by default, and the browser's TCP connections go through a local proxy that pins each one to the screened addresses, so a DNS answer that changes after the check cannot redirect them to a private host (WebRTC UDP is not covered). It is powered by Playwright and shipped as an optional extra so the core install stays lean:

cd backend
uv sync --extra browser
uv run playwright install chromium

Then uncomment the group: browser tool entries in config.yaml (browser_navigate, browser_snapshot, browser_click, browser_type, browser_get_text, browser_back, browser_screenshot, browser_close). make dev / Docker startup detects an enabled browser_navigate tool and preserves the browser extra on dependency syncs. The Gateway fails startup if browser control is configured but Playwright is missing, and /api/features hides the Browser UI unless the backend can actually serve it. Keep headless: true and allow_private_addresses: false for anything but local, trusted debugging. Attaching to an existing Chrome with cdp_url cannot enforce DeerFlow's subresource/redirect SSRF guard and therefore fails closed unless allow_unguarded_cdp: true explicitly acknowledges that risk; use it only with a trusted local browser. Browser sessions are process-local; keep the Gateway to a single worker process while this tool group is enabled because ordinary uvicorn worker dispatch does not provide thread affinity. That means GATEWAY_WORKERS=1, and on the launches that pass uvicorn no worker count at all (backend/Dockerfile, scripts/serve.sh) also WEB_CONCURRENCY unset or 1, since uvicorn takes the process count from it.

Existing, non-mock Custom Agent chats expose the same Browser Live controls when browser control is available and the agent either leaves tool_groups unrestricted or includes the browser group. An explicit tool-group allowlist without browser keeps those controls hidden.

The workspace Browser Live client negotiates binary JPEG WebSocket frames, keeps only the newest pending frame per display refresh, and revokes replaced object URLs. Gateway control messages remain JSON, and clients that do not request the binary capability retain the legacy JSON/base64 frame protocol.

Table Tool-Output Synopses: CSV/TSV row counts reflect logical records, so newlines inside quoted fields do not add rows. If parsing fails or exceeds the field limit, the synopsis reports an undetermined count. Outputs beyond the input-size limit still skip structured parsing.

Isolated Sub-Agent Context: Each sub-agent runs in its own isolated context. This means that the sub-agent will not be able to see the context of the main agent or other sub-agents. This is important to ensure that the sub-agent is able to focus on the task at hand and not be distracted by the context of the main agent or other sub-agents.

Summarization: Within a session, DeerFlow manages context aggressively — summarizing completed sub-tasks, offloading intermediate results to the filesystem, compressing what's no longer immediately relevant. This lets it stay sharp across long, multi-step tasks without blowing the context window.

Tool Output Preview Budget: Structured synopses of externalized results have a character budget. Long JSON keys and paths are shortened with ..., and oversized synopsis bodies include a truncation notice. Raw head/tail samples, the full file reference, and read_file guidance are preserved; read the saved file for exact keys or complete output.

Strict Tool-Call Recovery: When a provider or middleware interrupts a tool-call loop, DeerFlow now strips provider-level raw tool-call metadata on forced-stop assistant messages and injects placeholder tool results for dangling calls before the next model invocation. This keeps OpenAI-compatible reasoning models that strictly validate tool_call_id sequences from failing with malformed history errors.

Visible Tool-Run Completion: For interactive turns, DeerFlow retries an empty post-tool final response once, then surfaces a visible error instead of reporting a silent successful run.

Run History: Image-only user input is recorded once per run, even when the agent makes multiple model calls. Its media content remains in history without requiring a text summary.

Gateway API callers can opt into read_conversation and submit a conversation_references list with a run. The lead agent can then read bounded pages of the current visible text of those owned conversations. Read permission expires with the run, and text in old messages does not grant access. Text the agent has already read stays in the destination conversation after access expires or the source is deleted. A message too long for one read carries a continuation, so the agent can read the rest; it asks for the missing part only if that read is unavailable. SDK clients that cannot add top-level request fields may send the same list as context.conversation_references, and GET /api/features reports whether the tool is enabled. When it is, the web composer shows a "Reference a conversation" button next to the attachment button: pick up to three of your recent conversations, and they are attached to the next message only, shown as chips in the composer and in the transcript. There is no automatic history search. See configuration and the request contract.

Enable task notes and history recall with task_continuity.enabled: true. A task can retain up to eight notes. Parallel additions beyond the remaining slots return note_capacity, preserving existing notes. When artifact-handle resolution is enabled, capacity counts resolved keys; aliases for the same note share one slot. Malformed non-dict sibling arguments do not consume slots or disrupt valid note calls. Notes with invalid keys, content over 750 characters, more than four sources, or malformed source IDs also reserve no slot. Existing keys can still be replaced or deleted. Slots freed by sibling deletions or runtime failures (such as unavailable sources or policy denial) become available in the next batch, when rejected additions can be retried.

DeerMem debounces pending updates independently per conversation, user, and agent. memory.backend_config.debounce_seconds sets the quiet period (default 30 seconds), and max_wait_seconds caps debounce deferral from the first pending update (default 300 seconds). Activity in another conversation cannot extend that wait. Extraction remains serial, so backlog and LLM latency can still delay completion.

The opt-in DeerMem scope-isolation benchmark checks semantic safety across facts and summaries, and fact routing across users and agents. Failed extraction attempts are retryable execution errors, not safety passes.

For DeerMem, memory.backend_config.storage_class: markdown opts into tolerant summary reads while keeping JSON writes and the existing UI. A hand-edited memory.json can contain its JSON object inside a fenced memory-json block; embedded backticks and later fenced notes are supported. Unparseable summary text is moved to memory.json.corrupt-<timestamp> for recovery before rebuilding. storage_path must point to the data root directory, not an existing JSON file.

Gateway shutdown drains memory updates before closing the backend, even when shutdown is cancelled. Config reload failures are logged without aborting runtime teardown. For Kubernetes, budget terminationGracePeriodSeconds for all shutdown hooks, config/backend resolution, memory.shutdown_flush_timeout_seconds, and backend close plus a safety margin. The flush timeout does not bound close(): custom backends must make close quick or internally bounded, or shutdown can wait until the process is forcibly terminated.

Most agents forget everything the moment a conversation ends. DeerFlow remembers.

DeerMem can optionally suppress near-duplicate extracted facts with memory.backend_config.fact_dedup_enabled: true and fact_dedup_similarity_threshold (default 0.7, range 0.5–1.0). This local, deterministic word/CJK-bigram heuristic compares facts only within the same user, agent, and category; it is not semantic equivalence detection. It keeps the existing ID, text, and creation time, raises confidence to the maximum, and refreshes the source only when confidence increases. Explicit correction replacements and facts proposed for removal are protected from near-duplicate merging. A merge does not count as user confirmation. The gate is off by default and does not affect targeted fact updates.

DeerFlow also includes an optional openviking memory backend. It uses the official langchain-openviking package to capture completed turns into stable OpenViking Sessions and recall memory for prompt injection while leaving DeerMem as the default. The initial integration supports one DeerFlow user with one credential-bound OpenViking USER API key in memory.mode: middleware and does not inherit arbitrary HTTP headers from ovcli.conf. See OpenViking memory backend for its configuration, behavior, and current boundaries.

Across sessions, DeerFlow builds a persistent memory of your profile, preferences, and accumulated knowledge. The more you use it, the better it knows you — your writing style, your technical stack, your recurring workflows. Memory is stored locally and stays under your control.

DeerMem and mem0 capture user inputs and final assistant replies. Assistant turns that request tools, including malformed or provider-raw calls, are excluded from memory extraction even when they contain visible text.

DeerMem remains the default local backend. An opt-in mem0 backend is also available for the hosted mem0 Platform API or API-compatible self-hosted servers. Its token-bearing base_url must use HTTPS by default; plaintext HTTP requires an explicit local-development opt-in. See the mem0 backend guide.

An opt-in honcho backend is available for self-hosted or hosted Honcho (v3 API). It builds user-model memory — long-term preferences and a cross-session working representation — on Honcho's server side, so the backend makes no LLM calls locally. Each user gets an isolated workspace derived from user_id; a missing user id fails closed instead of falling back to a shared workspace. Fact CRUD and Settings-page fact editing are not available for this backend. See the Honcho backend guide.

Memory updates now skip duplicate fact entries at apply time, so repeated preferences and context do not accumulate endlessly across sessions. Legacy memory files are normalized when read or imported, including recoverable fact metadata, so older local data remains usable as the schema evolves. Frontend and backend normalization use confidence 0.5 when missing or invalid, default blank sources to unknown, and trim fact content.

In the default DeerMem middleware mode, automatic extraction now classifies every proposed fact by scope, durability, and authority before a deterministic write gate accepts it. Only durable, descriptive user-level facts are stored; current-thread or project constraints and one-time action permissions stay in conversation state. User-global summaries require both user scope and descriptive authority, contradiction removals are scope-gated, and a replacement-dependent removal is applied only when its replacement actually survives validation and storage. These classification labels are extraction-only metadata, add no extra LLM call, and are not written into the fact files. The explicit CRUD tools in memory.mode: tool remain a separate, model-directed path. Deployments that override the bundled DeerMem prompts via memory.backend_config.prompts_dir must add the new classification fields to their custom templates (the memory_update fact/summary/removal formats and the consolidation consolidated-fact schema): the write gate fails closed, so an un-migrated template stops every extraction-driven fact, summary, and removal write, surfacing only through the rejected_by_scope_gate metrics and the high-rejection-rate warning.

When a fact scope reaches max_facts, DeerMem still uses the historical confidence-only eviction order by default. Operators can opt in to memory.backend_config.fact_eviction_policy: hybrid-v1, which combines bounded confidence (65%), explicit-confirmation freshness (25%), and query-driven access heat (10%). Hybrid signal metadata is collected only while hybrid-v1 or shadow mode is enabled. Explicit confirmation is returned as factsToReinforce by the existing memory-update LLM call and is accepted only when deterministic message processing also detects a user reinforcement signal; it also resets the fact's staleness-review clock. This deterministic gate is batch-level: it establishes only that a human message among the last six filtered messages in the current extraction batch matched a reinforcement pattern. The LLM-selected factsToReinforce ID supplies the fact binding; DeerMem does not independently verify a signal-to-fact correspondence. Repeated extraction or automatic injection never confirms a fact. Custom memory_update prompts should add the optional factsToReinforce array to participate in confirmation freshness. Access heat is stored in a separate decaying sidecar and increases only when memory_search actually returns the fact, so reads do not rewrite canonical Markdown or its updatedAt. Hybrid mode also reserves a bounded minimum of correction slots (10% of the cap, at most 10; unused slots return to normal competition). Capacity deletion remains physical, but a bounded metadata-only audit records fact IDs, categories, policy scores, and reasons without copying fact content. fact_eviction_shadow_enabled: true evaluates hybrid-v1 alongside the default policy without changing actual retention. This feature adds no LLM invocation and can be rolled back by selecting confidence.

File-backed memory now separates global user context from agent facts. Each user has one memory.json containing only the project-independent user and history summaries; every fact is a canonical Markdown file below agents/{agent_name}/facts/. Existing lead-agent middleware, API, Settings, import/export, and embedded-client calls that omit agent_name resolve inside DeerMem to the reserved __default__ bucket. That bucket is outside the valid custom-agent name grammar, so a real custom agent named lead-agent has a separate fact repository and deleting a custom agent cannot delete a memory-only directory without config.yaml. Public agent identifiers are case-insensitive and canonicalized to lowercase. Runtime/API readers still receive a compatibility facts array for the selected/default agent, so the frontend does not read agent facts from memory.json; structured Markdown source metadata is projected to the historical string field at the MemoryManager boundary. An unscoped Clear All first migrates facts from unread legacy per-agent JSON without adopting its soon-to-be-cleared summaries, then removes shared summaries and facts from every agent bucket while preserving agent configuration files, so a later read cannot resurrect skipped legacy facts; an explicitly agent-scoped clear removes only that agent's facts. On first normal read, old facts embedded in the user JSON are migrated automatically to __default__; facts written to the earlier implicit lead-agent bucket are also moved when that directory is not a real custom agent. Migration and normal writes notify the configured retrieval adapter only after durable storage locks are released. DeerMem uses a scope-aware SQLite FTS5/BM25 adapter by default, stores only rebuildable derived index data under .retrieval/ (or under memory.backend_config.retrieval_index_path when set; a relative value resolves against storage_path), and rebuilds it in the background during Gateway startup or lazily on the first scoped search. A corrupt derived index is recreated automatically. When several Gateway instances share one storage_path, point retrieval_index_path at instance-local disk: SQLite's WAL mode is unsupported on network filesystems, each instance re-syncs a user's index on its next search after a peer writes, and a declared multi-instance deployment (deployment.multi_instance: true / DEER_FLOW_MULTI_INSTANCE=1) that leaves the index under the memory root logs a warning at startup. The Helm chart mounts a Pod-local emptyDir for it. Set memory.backend_config.retrieval_adapter to an empty string to disable it and use the local substring fallback. Chinese tokenization is optional; install the backend memory-zh extra (uv sync --extra memory-zh) for jieba-assisted sub-phrase search. Journaled writes, a shared user lock, and optimistic user-memory revisions prevent silent lost updates.

Set memory.backend_config.retrieval_relevance_enabled: true to opt into deterministic relevance/confidence ranking and query-aware injection. This bypasses retrieval_adapter for search, including the default FTS5/BM25 and custom adapters; indexing remains configured. retrieval_relevance_weight controls the blend and retrieval_diversity_weight enables near-duplicate penalties (default 0). Scoring uses at most the first 4096 characters and 128 tokens per query/fact. Search diversifies only up to top_k; injection diversifies guaranteed and regular facts independently until their token budgets are reached. Leave the feature disabled to retain the existing retrieval and injection behavior.

Lexical relevance measures IDF-weighted coverage of distinct query terms, so repeated partial matches cannot tie a complete match merely by saturating the score. Prefix matches require one complete token to prefix the other, not just four shared characters. For uploads, query-aware injection uses the preserved user request rather than the prepended file descriptions; attachment-only messages retain query-less injection. Older custom memory backends can keep their existing get_context signature: prompt injection and the inherited async wrapper pass query only when that callable supports the keyword and the hint is not None. Search and automatic injection use the same IDF weighting for the same user/agent candidate scope. Category-filtered search and injection's separate guaranteed/regular token budgets can still yield different final selections.

Memory injection follows the configured operation mode. In middleware mode, DeerMem injects the user-global summaries and the selected agent's facts. Custom-agent bootstrap conversations use that agent's fact bucket as well, so setup details do not leak into the default agent's memory. In tool mode, the automatic <memory> block contains only the global user and history summaries; agent facts are retrieved explicitly through memory_search, avoiding duplicate automatic and tool-returned fact context. Setting memory.injection_enabled: false still disables the entire block in either mode. When a fact ID is already known, memory_get(fact_id) returns that exact fact from the current user and agent scope without searching by content. Missing or out-of-scope IDs return a not-found error; backends without the required read or agent-scope capability return an unsupported error.

An individual Custom Agent can opt out of memory without changing the global setting. Add memory_enabled: false to that agent's users/{user_id}/agents/{name}/config.yaml. The agent still receives the current-date reminder, but DeerFlow does not inject recalled memory, queue passive or summarization-driven memory updates (including manual /compact), expose memory tools, or add memory-tool instructions for that agent. If an existing agent is switched off, its previously injected memory block is removed from checkpoint state before the next model call while its date reminder and conversation remain. Omitting the field (or setting it to true) preserves the existing global memory behavior.

Single-fact repository operations are genuinely incremental: an upsert/delete reads, journals, writes, and re-indexes only the addressed fact files, and returns an explicit incomplete delta rather than a cache-dependent fake full document. Summary change sets merge the supplied user/history child keys over the persisted sections so a partial update cannot erase omitted siblings; full imports normalize both sections to the complete compatibility schema before applying replacement values. Imports reject malformed fact lists or blank/non-text content with HTTP 400 before changing stored memory; recoverable legacy metadata still receives defaults. An explicit empty fact list remains an intentional clear. Manager/API compatibility methods materialize a fresh full document only when their public response contract requires one. Fact-level point operations use separate expected user-memory and fact revisions and may explicitly rebase when every addressed fact precondition still holds. Snapshot-derived operations such as scoped clear, capped create, consolidation, and trimming never replay stale delete/trim sets: a manifest conflict reloads the complete document and recomputes the operation, with a bounded retry. Fact paths use the first two hexadecimal characters of SHA-256(fact_id) so generated fact_* IDs distribute across shards. The cache token combines the shared JSON's nanosecond mtime, size, and persisted revision; this prevents coarse-mtime same-size writes from returning stale data without scanning fact files. Direct out-of-band Markdown edits require an explicit reload. Storage-specific conflicts and corruption are translated at the MemoryManager boundary; the Gateway returns conflict as HTTP 409 and a stable, non-sensitive corruption error as HTTP 500. Full-document save() remains a compatibility API and computes a diff before writing; malformed or missing facts can no longer silently erase an agent's Markdown files. Legacy migration preserves non-empty user/history before deleting an agent memory.json; conflicting summaries keep the legacy file and fail loudly instead of choosing a winner.

Legacy facts in memory.json migrate automatically into the reserved __default__ Markdown bucket on the user's first normal memory read. Operators who prefer to audit or complete the migration before serving traffic can run the optional idempotent CLI from backend/:

PYTHONPATH=. python scripts/migrate_memory_markdown.py --all-users --dry-run
PYTHONPATH=. python scripts/migrate_memory_markdown.py --all-users
# A custom DeerMem root or original non-directory-safe identity can be explicit:
PYTHONPATH=. python scripts/migrate_memory_markdown.py --storage-path /path/to/deerflow-home --user-id 'test@example.com'

The v1-to-v2 storage migration is one-way for a running application: pre-PR code does not read Markdown facts. Before upgrading a persistent deployment, stop DeerFlow and take a filesystem snapshot or full backup of the configured memory storage root. The migration also durably retains each destructive JSON source beside the original path as {manifest_filename}.v1.bak before writing v2 data; an existing mismatched backup or a backup-write failure stops migration without modifying the v1 source. This local backup preserves pre-migration data but is not a substitute for a full snapshot and does not contain facts created after the upgrade.

--user-id may be repeated. --all-users discovers the existing directory-safe buckets below the selected storage root; standalone integrations that passed raw IDs containing characters such as @ should use the original value with --user-id. A failed user's migration is reported without hiding the rest of the audit, and the command exits non-zero when any user fails. The automatic first-read path remains enabled, so running this CLI is not required for startup.

DeerFlow is model-agnostic — it works with any LLM that implements the OpenAI-compatible API. That said, it performs best with models that support:

• Long context windows (100k+ tokens) for deep research and multi-step tasks
• Reasoning capabilities for adaptive planning and complex decomposition
• Multimodal inputs for image understanding and video comprehension
• Strong tool-use for reliable function calling and structured outputs

MindIE XML tool-call arguments support signed, fractional, and exponent JSON numbers. If a number overflows, underflows to zero, or exceeds Python's integer digit limit, the complete argument stays a string, including nested JSON lists and objects, so parsing does not silently change its value. Python-literal sets and tuples containing non-finite floats also retain the complete argument as text, including when nested in other containers. Dictionary keys and values are checked recursively as well.

chat() and stream() preserve text appended by model middleware before the first state snapshot, emitting only the suffix after the streamed text.

For DeerFlowClient(agent_name="researcher"), the named agent's mcp_plugins selection applies to both the lead agent and its task / batch_task delegations: null inherits all enabled MCP plugins, [] selects none, and installation IDs select only those plugins. Call client.reset_agent() after editing the saved agent configuration to refresh the selection.

DeerFlowClient.stream() includes summary_text in each values event. This is the current compacted context summary, or None when absent. Consumers can record changes without reading checkpoint internals; repeated snapshots may carry the same summary, and an initial snapshot may already contain one from an earlier turn.

DeerFlow can be used as an embedded Python library without running the full HTTP services. The DeerFlowClient provides direct in-process access to all agent and Gateway capabilities, returning the same response schemas as the HTTP Gateway API. The HTTP Gateway also exposes DELETE /api/threads/{thread_id} to remove DeerFlow-managed local thread data after the LangGraph thread itself has been deleted:

For database-backed run events, deleting a run preserves its thread's sequence watermark. Thread deletion removes that watermark once no events remain, allowing a recreated thread to restart at sequence 1. Owner-scoped deletion preserves the watermark when another owner's events remain.

Single-process JSONL event storage also retains the thread sequence watermark across run deletion and restarts, so clients using after_seq do not miss later messages. Keep runs/.seq-watermark with the run files when backing up this backend. New or replaced watermarks inherit the deleted run file's permission bits. Deleting the complete thread removes the watermark and resets allocation.

Thread IDs may be supplied by callers and do not have to be UUIDs. Explicit IDs must contain 1–64 ASCII letters, digits, hyphens, or underscores (^[A-Za-z0-9_-]{1,64}$). DeerFlow generates a UUID only when thread_id is omitted or None; an explicitly supplied empty string is invalid. Existing route-addressable threads created under older, looser rules remain readable and deletable, but cannot start new runs or create new filesystem or sandbox state. Legacy deletion skips local path cleanup when the ID is not safe under the canonical contract. For canonical legacy threads whose conversation exists only in LangGraph checkpoints, DeerFlow seeds an empty run-event feed from the checkpoint before the first new run so /messages/page keeps both the old and new turns.

from deerflow.client import DeerFlowClient

client = DeerFlowClient()

# Chat
response = client.chat("Analyze this paper for me", thread_id="my-thread")

# Streaming (LangGraph SSE protocol: values, messages-tuple, end)
for event in client.stream("hello"):
    if event.type == "messages-tuple" and event.data.get("type") == "ai":
        print(event.data["content"])
    elif event.type == "messages-tuple" and event.data.get("type") == "tool" and "artifact" in event.data:
        # Structured tool artifacts (for example, ask_clarification cards)
        # are preserved when the ToolMessage provides one.
        print(event.data["artifact"])

# Configuration & management — returns Gateway-aligned dicts
models = client.list_models()        # {"models": [...]}
skills = client.list_skills()        # {"skills": [...]}
client.update_skill("web-search", enabled=True)
client.upload_files("thread-1", ["./report.pdf"])  # {"success": True, "files": [...]}
client.set_goal("thread-1", "finish the implementation and make all tests pass")
client.get_goal("thread-1")       # {"goal": {...}} or {"goal": None}
client.clear_goal("thread-1")

When continuing a thread, values still contains the full conversation state, while messages-tuple and end.usage cover only the current turn.

The embedded client's end.usage sums usage deltas across streamed chunks, including providers such as native Gemini that report usage more than once per reply. The completed message's usage is reconciled without counting those chunks again; a lower completed-message count does not subtract tokens already reported by chunks. Signed provider corrections in chunks remain supported, as do providers that report usage only at completion.

Concurrent upload_files() calls preserve filenames absent at the start of each call by assigning numeric suffixes and returning the actual saved names. The same rule applies to new Markdown companions. Exact names already present retain replacement behavior. Convertible files use one private source capture through all filename retries and conversion, preserving the original permissions and timestamps. Detected staging-file substitution is rejected; staging cleanup failures are logged without masking the upload outcome.

The HTTP Gateway accepts values, messages-tuple, updates, debug, tasks, checkpoints, and custom stream modes. Unsupported modes such as messages and events, unsupported non-default run options such as webhooks, delayed execution, or multitask_strategy="enqueue", and undeclared SDK options such as checkpoint durability overrides return 422 before execution instead of being silently ignored or downgraded.

All dict-returning methods are validated against Gateway Pydantic response models in CI (TestGatewayConformance), ensuring the embedded client stays in sync with the HTTP API schemas. See backend/packages/harness/deerflow/client.py for full API documentation.

Projects group related conversations under a shared name, instructions, and document shelf.

A conversation joins a project at creation time (when a project is selected) or later through the move menu. Runs never modify membership: submitting a message cannot assign or reassign a conversation. Moving a conversation out of a project keeps it unassigned until it is explicitly moved again.

Moving a conversation refreshes its header affiliation as well as the project lists, including when an older metadata request is still in flight.

Projects require the current database tables and columns. A database stamped 0019_thread_incarnations from the older 0018-based rollout is rejected at startup if the project schema is missing. Follow the offline database recovery procedure before starting this build against that database.

Each project stores free-form instructions — background, conventions, and constraints that apply to every conversation in the project — editable on the project page's Instructions tab with a live byte counter. When a run starts on a member thread, the Gateway pins the project's current state once and renders the instructions as a bounded, request-scoped <project> block for that run only: the block never enters the system prompt or persisted history, and every new run sees the latest saved instructions. Instructions are capped at projects.instructions_max_bytes UTF-8 bytes (default 8192, range 256–262144); multi-byte characters count as their UTF-8 byte length. Oversized instructions are rejected with a 422 at write time and are never silently truncated.

Each project has a document shelf for files the whole project shares, managed from the project page's Documents section:

• Upload a file (button or drag-drop, one file per request). Shelf size limits reuse uploads.max_file_size (default 50 MiB); re-uploading identical content returns the existing entry instead of creating a duplicate.
• List entries with name, size, modified time, and provenance (uploaded vs. saved from a conversation), and preview or download any entry.
• Save to project from a thread file: the read-only conversation-files browser below the shelf lists member threads' uploads and outputs, each with a Save to project action.
• Attach to thread: copy a shelf file into a thread's uploads through the normal ingestion pipeline, so the conversation can work with it directly.

New shelf names, including explicit upload name and promotion shelf_name, follow ordinary upload filename validation. Names containing NUL, Windows reserved device names (such as CON.txt), or trailing dots are rejected with 400 before bytes are staged.

On native Windows, the document shelf uses extended-length filesystem paths so deep workspace directories and long filenames can be uploaded, downloaded, restored, and purged without enabling the system-wide long-path setting. Filesystem limits on individual path components still apply.

Runs on member threads also receive a bounded <documents> index rendered per run from the pinned snapshot (capped by projects.shelf_index_max_entries and projects.shelf_index_max_bytes), and the agent can page the shelf and read documents with the list_project_documents and read_project_document tools.

Document character counts use a bounded, thread-safe cache; concurrent reads do not serialize full file scans behind its cache lock.

Archiving a project freezes writes but keeps reads. Threads in an archived project still run and still receive the project's instructions and shelf index, and the shelf remains fully readable: listing, preview/download, the conversation-files browser, and attach-to-thread all keep working. Uploads, save-to-project, and moving individual shelf files to trash require an active project, and a trashed document cannot be restored into an archived one. Deleting an archived project remains allowed and moves its whole shelf to trash.

Deleting a shelf document moves it to trash instead of erasing it: the entry keeps its bytes and a snapshot of its origin project for projects.trash_retention_days (default 30) before the retention sweep may purge it permanently. The /workspace/trash page — reachable from the project page's Documents section and the sidebar Projects header — lists trashed documents with their origin project and remaining retention, with per-entry Restore and Delete permanently actions plus an Empty trash action that permanently deletes every document in the trash — immediately, not after the retention window; the window only bounds how long an entry may sit there before the retention sweep reclaims it. Restore returns the document to its origin project, or to a project you pick when the origin is gone or archived; if the target already holds an identical active file, the entries merge. Deleting a project moves its entire shelf to trash in the same step.

DeerFlow now includes a first-class scheduled-task MVP in the workspace.

Editing a one-time task's title or prompt preserves its original execution time, including seconds and the selected occurrence during a daylight-saving clock rollback. Changing its date, time, or timezone recalculates the execution time. Switching tasks while editing loads the selected task's own title, prompt, and schedule.

An unchanged one-time execution time can still be saved within scheduler.min_once_delay_seconds of its run. New execution times and reactivated finished tasks must satisfy that minimum delay.

Current MVP capabilities:

• Manage tasks at /workspace/scheduled-tasks
• One-time task forms reject local times skipped by daylight-saving transitions; select another time before creating or saving the task.
• Existing one-time tasks display the scheduler's execution time even when a conversation or API saved a local time skipped by daylight saving. Local times without an explicit offset use the first occurrence of a repeated hour.
• Choose whether each scheduled task reuses a thread and its conversation history or creates a fresh thread per run
• Pin each task to lead_agent (default) or a custom agent the owner already has; unknown names are rejected
• Sending assistant_id: null in a scheduled-task PATCH resets the task to lead_agent; omitting assistant_id preserves the current agent, including when that custom agent has since been deleted
• Duplicate an existing task into the create form as an editable draft without copying its run history
• Support once, cron, and interval schedules
• Editing or duplicating an interval task preserves its saved cadence until the interval is explicitly changed, including sub-minute intervals allowed by the operator's scheduler configuration
• Run background scheduled executions as non-interactive DeerFlow runs (ask_clarification is not exposed there)
• Persist a due execution as queued when its reused thread or the global execution budget is busy, then launch it when capacity is available; queued occurrences survive Gateway restarts and fail after scheduler.queue_timeout_seconds
• Share execution slots fairly between task owners: one owner may have at most scheduler.max_concurrent_runs_per_user scheduled runs starting or running at a time (default 2, never more than max_concurrent_runs; 0 turns the per-owner cap off), and the waiting queue is drained owner by owner, so one owner's backlog never holds back another owner's run. A run that waits longer than scheduler.queue_timeout_seconds is skipped, and its history reads "Skipped: it waited too long for a free slot"
• Freeze a task's definition while an occurrence is queued, launching, or running, so a durable occurrence cannot silently pick up a different prompt, thread, or schedule; transitioning a task to paused or deleting it cancels an existing waiting occurrence, while launching/running work must finish before those mutations are retried and an explicit manual trigger may still wait and run without resuming a paused schedule
• Pause, resume, trigger, inspect history, and delete tasks
• Search task titles or prompts, combined with status/type filters and the current thread scope.
• Execute scheduled work through the normal DeerFlow run lifecycle
• When channel_connections.enabled: true, send scheduled task updates to the task owner's connected IM identities (outbox + delivery worker) on apps that support proactive push, which today is WeCom, Feishu/Lark, WeChat and QQ private chat; Settings → Notifications controls proactive updates per connected app, independently of browser notifications and ordinary chat replies. Each occurrence queues at most one message per target: a finished, failed or goal-missed run, the automatic pause after three missed goals, a pause by the agent (its stop condition was met) or the task finishing (all max_runs done, end_at reached); when several apply, the pause or finish wins and still says how the last run went. A one-time task sends only its run's outcome. The message is queued in the same database transaction that records the outcome, so runs finalized after a crash or a lost lease do not enqueue a duplicate notice for the same target. It reads on its own, in your web UI language (else channel_connections.notification_locale): the task title, what happened, a one-line result when the agent replied, and "Open DeerFlow → Scheduled tasks for details.", with no IDs and no links. Plain manual "run now" trials and interrupted runs stay silent, and so do occurrences that end without a finished run (launch error, queue timeout, interrupted by a restart). Channel/transport outages park deliveries without exhausting retries, for up to about a day; platform rejections retry for roughly 15 minutes before settling as failed. An identity you disconnect while a delivery is waiting is never pushed to: the row is dropped as failed.
• Browse execution history in pages of 50; older pages pause automatic refresh, with an explicit return to the latest runs. Counts appear only after a successful read; loading and failed reads are not reported as zero runs.

Derived run summaries omit leading inline <think> reasoning before limiting the text, so a long thought does not replace the actual answer in previews or scheduled notices. Original response history and literal tags in the answer are unchanged; previously stored run summaries are not rewritten.

Filter execution history through the API

To inspect failures without downloading every successful occurrence, authenticated clients with threads:read can request GET /api/scheduled-tasks/{task_id}/runs?status=failed&limit=50&offset=0 for an owned task. The optional status accepts queued, launching, running, success, failed, skipped, interrupted, or unmet; these are occurrence statuses, so task statuses such as completed are invalid (422).

Filtering happens before pagination. limit (1–200, default 50) and offset (nonnegative, default 0) apply to matching records, ordered by creation time then ID, both descending. Omitting status preserves the existing mixed-history array response; no matches return []. The API does not change task execution, and the workspace history UI remains unfiltered.

Current MVP limits:

• No text-only notification jobs
• No channel or GitHub dispatch targets (result push above is not a dispatch target)

Enable background polling with config.yaml -> scheduler.enabled. Manual trigger uses the same scheduled-task resource and execution path.

In the task form, switching an hourly, daily, weekly, or monthly schedule to Custom cron starts from its current timing, including selected weekdays or day of the month. Switching back to Custom cron after preset edits uses those edits rather than an older expression.

Open Settings → Notifications → IM notifications to turn scheduled-task updates on or off for each connected, push-capable app. Existing connections keep receiving updates until you opt out. Turning a channel off cancels its pending and retrying notices, including an explicitly selected group; turning it on again does not resend the old backlog. Notifications whose platform delivery already started may still arrive. Chat replies and scheduled execution are unaffected. All Gateway workers must run the upgraded version for the preference to be enforced.

WeChat needs a private message after binding to save its current conversation context securely; an old binding or changed bot token also needs a fresh private message. QQ requires a C2C/private binding and platform permission to send proactively. WeChat remains private-only; QQ and WeCom also support explicitly selected group destinations. See setup and delivery limits.

The same page offers separate Task & goal run notifications for completion, terminal failure and requests for input. These ordinary web/background reminders are off by default. A pending input request takes precedence over completion; unmet goal stand-downs stay silent, and IM-originated chats keep their original reply rather than receiving a second proactive notice. Turning an event off also cancels its waiting backlog. Messages direct you back to Conversations.

In a task's create or edit form, Also notify a group (optional) adds one chosen group to the existing personal notifications. Add the app bot to the group, choose its supported channel in the binding panel, then send the displayed /connect command while mentioning the bot in that group. QQ and WeCom require a name in the binding panel so you can identify the receiving group; the name travels with the owner-issued binding code and is saved on the connection. Refresh and select the group by name. Feishu additionally needs im:chat:read to retrieve its group names; every provider needs permission to send proactively.

While the form checks group-notification support, it shows a loading hint and retains the saved selection.

No group is selected automatically. Clearing the selection stops new group notices; changing it affects future notices, while already queued notices keep their original target. A disconnected binding never receives a waiting notice. The first version is configured through the task form or REST API; conversation schedule tools keep their existing inputs. See the group delivery guide.

• The tasks page and the REST API (POST / PATCH /api/scheduled-tasks) accept the same per-run goal (goal_objective), safety cap (max_runs, end_at) and stop condition (stop_condition) as a conversation. Sending null in a PATCH clears any of these four; an end_at without a UTC offset is wall-clock time in the task's timezone.
• A stop condition is the user's "stop when …" rule. It is stored in its own field (migration 0031), never inside the task instructions. Only when a run starts does DeerFlow append it to that run's message and ask the run to call stop_scheduled_task when the rule holds. While scheduler.tool_enabled is on, every scheduled run can stop its own schedule, whether a chat or the tasks page created the task; with it off, the run is asked to report a met rule instead of calling a tool it does not have.
• Goal tasks created on the tasks page are now evaluated like chat-created ones, and their runs also receive the saved notes and the previous-run reference.
• Resume computes the next run from now, so a long pause never causes a catch-up run. A one-time task whose time has passed returns 422 once_time_passed and needs a new time. Resuming an active task changes nothing; pausing a finished task returns 409 task_finished.
• max_runs is a lifetime total of automatic runs; trial runs never count. Reactivating a task whose cap is used up (Resume, or a PATCH that re-arms a finished task's schedule) returns 409 limits_exhausted unless the same request renews the limit that ran out: a used-up run limit needs a higher max_runs or null, a passed end time needs a later end_at or null (a later end_at alone does not renew a used-up max_runs). POST /api/scheduled-tasks/{task_id}/resume accepts an optional {"max_runs": …, "end_at": …} body for that (null clears a cap; chat-created sub-hourly tasks must keep one). A PATCH that only changes the cap of a finished task saves it and leaves the task finished.
• Goal-check failures (the evaluator failed, or the conversation changed during the check) neither count toward the three-miss automatic pause nor reset it. Changing the goal, the instructions or the stop condition, or adding a note, starts a new count; Resume keeps it.
• While this Gateway process's scheduler is not running, creating a task (including Duplicate) returns 409 scheduler_not_running, because the task would never run on schedule. GET /api/features reports scheduled_tasks.available, running, tool_enabled and min_interval_seconds.
• Errors from /api/scheduled-tasks* are {"detail": {"code", "message", "params"}}; see backend/docs/API.md and contracts/scheduled_task_errors_contract.json.

Set both scheduler.enabled: true and scheduler.tool_enabled: true, then restart Gateway. An authorized interactive turn can use schedule_task to create, update, list, pause, resume or delete tasks, start a trial run, or save a note. For example: “Every weekday at 9:00, check release-checklist.md and tell me what is still unchecked; stop when everything is checked.” In the web app the result is a live card with the schedule, the stop condition and buttons, and the agent replies in one or two sentences; IM and other non-web turns describe the schedule, the next run and the stop condition in plain text.

• Which tasks a conversation manages. The tasks created in it, and, in a run conversation (the chat a scheduled run posted into), the task that run belongs to: “pause this” or “move it to 10:00” work there too. This applies only to turns you send. A scheduled run itself can only pause its own schedule with stop_scheduled_task.
• Edits keep the task. Changing the time, instructions, goal, stop condition or safety cap is an update of the same task, so its ID and run history stay. resume restarts a paused or finished task without a catch-up run. When the cap is used up, the agent asks how to renew the limit that ran out (a higher max_runs or none; a later end_at or none) and sends that with the resume.
• Timezone. A zone you name wins. Otherwise a new task uses the browser timezone the web app sends with each message (context.client_timezone, read only for this), and the result says which zone was used. Intervals and one-time times with a UTC offset need no zone; for a cron schedule or a local one-time time with no known zone (for example from IM), the agent asks. Edits keep the saved zone; the browser zone never changes an existing task.
• Where results appear. Each run posts its result in a new chat of its own, titled “{task} · {local time}”, or in the originating chat when the task runs there. When the schedule is paused by the agent, is paused automatically or finishes, the originating chat shows one line where the conversation stood, with a link to that run or to the task; nothing else is posted back to it. The line stays after the task is deleted. A run chat shows the task instructions as one collapsed “Task instructions” block under the run's header instead of a long user message.
• Language. The agent writes the title, instructions and stop condition in your language, and scheduled runs answer in the language of the instructions.

New tasks default to a fresh conversation for each occurrence. A configured goal_objective applies only to that occurrence: success does not stop a recurring schedule. The running agent can request stop_scheduled_task for its own schedule when your stop condition has been met; the request takes effect during terminal finalization. max_runs counts automatic launches only, and end_at provides a deadline. Either end condition takes precedence over a pause request. Tool-created sub-hourly schedules require an end condition; each owner may keep at most 20 live tool-created tasks, including paused tasks.

An unmet occurrence is recorded as unmet, distinct from an execution failure. Three eligible automatic unmet occurrences pause a recurring task. Accepted success resets the streak, including a success relying on disclosed assumptions; manual trials, interruption, execution failure, external waiting and goal-check failures do not advance it. Resume retains the streak, so another eligible unmet occurrence can pause the task again. Connected IM apps with proactive push receive one notice per occurrence through the same durable outbox (goal missed, auto-paused, paused by the agent, finished); plain manual trials stay silent.

You can ask the agent in a conversation that manages the task to save an explicit note for future runs (at most 10 notes of 500 characters). Fresh recurring runs may read the previous executed occurrence through opt-in read_conversation, with the same owner and read-permission checks. This provides a source reference, not an automatic summary or a post-back into the originating chat.

For a trial, ask directly, for example “Run it now”, “OK, run it now” or “先跑一次吧”; in the web app the card's Run once now button does the same. The host accepts a bounded set of English/Chinese direct-run requests from the current user turn, optionally after a short acknowledgement such as “Sure,” or “好的，”. A bare “yes” or “好”, task mentions, and quoted or conditional requests do not start a paid run. When a run is already waiting to start, no extra trial is added and the agent says so. A trial does not count toward max_runs.

A goal occurrence can use up to nine agent turns, with an evaluator request after each. Evaluator requests and provider-reported tokens are included in run usage; missing usage makes the corresponding cost estimate unknown. token_budget limits the main graph and is checked after model calls; evaluator usage is additional, so it is not a strict whole-run or dollar limit.

Goal evaluation upgrade: existing scheduled, webhook and autonomous goal runs can accept disclosed low-risk, reversible assumptions and record relied_on_assumption in their verdict. Interactive goal evaluation remains strict. This policy also applies when conversation schedule tools are disabled; it does not authorize sensitive actions or substitute for user authorization.

Scheduled runs use scheduler.recursion_limit in config.yaml (default 1000, matching the web UI's interactive budget). Values above max_recursion_limit are clamped. This field is read at dispatch, so the next scheduled run picks it up without a Gateway restart.

The background scheduler is single-instance by default. For a multi-pod deployment, set scheduler.multi_instance: true and use shared Postgres, run_ownership.heartbeat_enabled: true, and run_events.backend: db; startup and periodic recovery then preserve live peer runs, atomically return expired launch claims to the queue, take over only expired run leases, and fence stale launch writes. max_concurrent_runs is a shared global cap across Pods for launching/running occurrences; waiting queued rows do not consume it. Without those settings, enable the scheduler on exactly one Gateway pod. These scheduler fields are startup-only; restart all Gateway Pods together when changing them.

Authenticated clients with threads:read can call POST /api/scheduled-tasks/preview-cron before creating a task:

{"cron":"0 9 * * 1-5","timezone":"Asia/Shanghai","count":3,"start_at":"2026-09-12T00:00:00Z"}

The response contains normalized cron, timezone, the effective UTC start_at, and occurrences with UTC run_at and offset-bearing local_time. In this example the first occurrence is 2026-09-14T01:00:00Z / 2026-09-14T09:00:00+08:00.

count is an integer from 1 to 10 (default 5). start_at must include a timezone; omit it to capture server time once. Cron expressions use the scheduler's five-field syntax (maximum 256 characters); timezone names are at most 128 characters. Invalid inputs or schedules without the requested future occurrences return 422. Preview shares the scheduler's DST behavior, creates no task, thread or run, and does not reserve execution. This is an API capability; the workspace form does not yet display these occurrences.

• Occurrence ordering applies to rows admitted by upgraded Gateway instances, which project only sequenced occurrences onto the parent task and defer recovery while any occurrence is still live, whichever instance admitted it; a task whose history is entirely unsequenced keeps the previous timestamp ordering until its first sequenced admission. During a rolling upgrade, rows admitted by pre-upgrade instances are projected by those instances themselves, as before the upgrade, and the ordering guarantees hold once every Gateway writer runs the upgraded version. Existing history is not backfilled; the upgrade does not reconstruct past order or repair historical counts.
• Before upgrading a deployment with GATEWAY_WORKERS > 1 and scheduler.enabled: true, either keep the scheduler on exactly one Gateway worker or configure scheduler.multi_instance: true with shared Postgres, run_ownership.heartbeat_enabled: true, and run_events.backend: db. The upgraded Gateway rejects the unsafe combination at startup instead of starting silently.
• In multi-instance mode, scheduler.max_concurrent_runs is a cluster-wide execution cap, not a per-Pod cap. It includes launching and running scheduled occurrences, so capacity does not multiply with the number of replicas; durable waiting rows remain outside the cap.
• scheduler.multi_instance and the related scheduler, ownership, and run-event settings are startup-only. Apply changes with a coordinated restart of all Gateway Pods; changing the ConfigMap alone does not activate multi-instance recovery.

deerflow is a terminal-native workbench for people who live in the shell. It runs embedded over DeerFlowClient — no Gateway, frontend, nginx, or Docker required — while honoring the same config.yaml, checkpointer, skills, memory, MCP, and sandbox settings as the rest of DeerFlow.

Parallel synchronous stdio MCP calls use independent sessions on their own event loops. They do not cancel each other's connections, but they do not share server-side state; session reuse requires the same loop. See the MCP session notes for details. Manually managed event loops must drain pending session owners before closing; the normal asyncio.run() path does this automatically.

uv pip install 'deerflow-harness[tui]'        # optional 'textual' dependency

deerflow                                      # launch the terminal UI (TTY required)
deerflow --tui-transparent                    # use the terminal's default background
deerflow --continue                           # resume the most recent thread
deerflow --resume THREAD                      # resume a thread by id
deerflow --print "summarize this repo"        # headless one-shot answer to stdout
deerflow --json  "hello"                       # headless newline-delimited StreamEvents; failure -> one {"type": "error"} record, exit 1
deerflow --recursion-limit 250 --print "task" # override the headless agent-loop limit

Headless --print and --json exit with status 1 when the run fails, including provider errors returned as fallback messages. --print still writes the fallback text to stdout; --json appends a terminal error record.

A keyboard-driven chat surface with a streaming transcript (Markdown-rendered answers), compact tool-activity cards, a / slash-command palette, display-only /clear, /goal goal management, /model and /threads pickers, input history, PageUp/PageDown transcript navigation, and Esc / Ctrl+C interrupt. The composer preserves line breaks and indentation in pasted code, stack traces, and multi-paragraph prompts; Enter sends the complete document. Transcript refreshes preserve your reading position after you scroll upward and resume following new output when you return to the bottom. /clear removes rows from the current terminal display without deleting the thread or its persisted conversation; /new and /clear ask you to wait during an active run instead of resetting in-flight display state. Sessions opened in the TUI also appear in the Web UI sidebar — it writes the shared thread store under the local default user, so terminal and web stay in sync without running the Gateway.

During an active run, /resume, /threads, and /switch ask you to wait before switching conversations. An invalid /resume reference displays an error without closing the TUI or changing the current conversation. After an interrupt, late stream actions from the previous run cannot change the next run's display or status, even when both prompts use the same conversation. A tool call already in progress still finishes, so until it does a new prompt in that conversation is held back with a notice instead of racing it; /new and /resume stay available. If a run cannot start, the TUI reports an error and returns to idle so you can retry. A message that is held back or cannot start stays in the composer, so Enter resends it.

At the last composer row, Down leaves an unsent draft untouched unless you are browsing input history; after recalling history, it moves forward to restore your saved draft.

At the first composer row, Up also leaves the draft, cursor, and undo history untouched when no input history is available. Recalling identical history text or a saved draft also preserves the cursor and undo history.

See backend/docs/TUI.md for the full guide.

• Contributing Guide - Development environment setup and workflow
• Configuration Guide - Setup and configuration instructions
• Architecture Overview - Technical architecture details
• Backend Architecture - Backend architecture and API reference

DeerFlow has key high-privilege capabilities including system command execution, resource operations, and business logic invocation, and is designed by default to be deployed in a local trusted environment (accessible only via the 127.0.0.1 loopback interface). If you deploy the agent in untrusted environments — such as LAN networks, public cloud servers, or other multi-endpoint accessible environments — without strict security measures, it may introduce security risks, including:

• Unauthorized illegal invocation: Agent functionality could be discovered by unauthorized third parties or malicious internet scanners, triggering bulk unauthorized requests that execute high-risk operations such as system commands and file read/write, potentially causing serious security consequences.
• Compliance and legal risks: If the agent is illegally invoked to conduct cyberattacks, data theft, or other illegal activities, it may result in legal liability and compliance risks.

An admin can register stdio MCP servers, which run commands inside the Gateway container. The API restricts them to an allowlist (npx, uvx by default, extended via DEER_FLOW_MCP_STDIO_COMMAND_ALLOWLIST) and rejects arguments and environment variables that would evaluate arbitrary code. That is defense in depth, not a boundary: these launchers exist to fetch and run remote packages, so treat Gateway admin as equivalent to code execution on the host and grant it accordingly.

Gateway run requests and manual thread-state updates reject client-supplied system / developer messages with HTTP 400, including equivalent serialized message forms. Ordinary chat, attachments, and assistant/tool history replay remain supported. Session or PAT authentication does not grant system-prompt authority; trusted internal run producers retain that ability.

This check prevents new role injection; it does not rewrite existing checkpoints. If an older version accepted an injected system message, use a fresh thread or have an operator review and clean the affected state. Restarting the service does not remove persisted instructions, and restoring an older checkpoint can restore them.

For local verification, run python backend/tests/poc_external_system_message_injection.py --help. The same opt-in PoC supports --expect vulnerable on an isolated old revision and --expect blocked after the fix. Its help includes PAT creation, thread-ID selection, browser follow-up, and the distinction between persistence and model obedience. Use a fresh disposable thread for each run; the test appends messages.

On an isolated unfixed checkout, --expect vulnerable demonstrates acceptance only when the request returns 200 and the exact injected message remains in the checkpoint as type=system across a normal follow-up. A marker in a web answer is model-dependent and is not evidence by itself that the role was promoted. After applying the fix, run the same script with --expect blocked: it requires the specific role-rejection 400, an unchanged checkpoint, a successful ordinary follow-up, and absence of the rejected message IDs. Other 400 responses and authentication, conflict, or server errors are inconclusive rather than passes.

The PoC does not clean up automatically. When verification is complete, delete the disposable chat with the web sidebar's delete action and revoke the short-lived PAT if one was created. Restarting the service does not remove a persisted injected instruction.

The Docker stack publishes its entry port on 127.0.0.1 only, and local make dev / make start bind nginx, the Gateway, and the frontend to loopback, matching the local-trusted-environment model described above. To reach it from another machine, set BIND_HOST in .env (e.g. BIND_HOST=0.0.0.0) — and only after putting the security measures below in place. Local runs apply BIND_HOST to nginx on port 2026 only; the Gateway and frontend stay on loopback behind it.

Complete first-run setup before the host becomes reachable. A fresh instance has no accounts yet, so create the admin account through /setup immediately after starting any deployment that is not loopback-only.

Note: We strongly recommend deploying DeerFlow in a local trusted network environment. If you need cross-device or cross-network deployment, you must implement strict security measures, such as:

• IP allowlist: Use iptables, or deploy hardware firewalls / switches with Access Control Lists (ACL), to configure IP allowlist rules and deny access from all other IP addresses.
• Authentication gateway: Configure a reverse proxy (e.g., nginx) and enable strong pre-authentication, blocking any unauthenticated access.
• Network isolation: Where possible, place the agent and trusted devices in the same dedicated VLAN, isolated from other network devices.
• Stay updated: Continue to follow DeerFlow's security feature updates.

We welcome contributions! Please see CONTRIBUTING.md for development setup, workflow, and guidelines.

Backend make test runs four duration-balanced shards concurrently and reports failure if any shard fails. Use make test TEST_JOBS=1 to run the shards sequentially. It excludes live external-API and blocking-I/O coverage. Run cd backend && make test-blocking-io for strict blocking-I/O checks. Maintainers can run the real DeerFlowClient suite with cd backend && make test-live. This command requires a valid root config.yaml and API credentials. It can incur API costs and create local sandboxes, artifacts, or files. Direct pytest runs additionally require DEER_FLOW_RUN_LIVE_TESTS=1.

Regression coverage includes Docker sandbox mode detection and provisioner kubeconfig-path handling tests in backend/tests/. Backend blocking-IO diagnostics are available from the repository root with make detect-blocking-io: it statically scans backend business code for blocking IO that may run on the backend event loop, prints a concise summary, and writes complete JSON findings to .deer-flow/blocking-io-findings.json. The JSON includes compact review records with priority, location, blocking_call, event_loop_exposure, reason, and code. Gateway artifact serving now forces active web content types (text/html and XML documents such as .xml, .xhtml, and .svg) to download as attachments instead of inline rendering, reducing XSS risk for generated artifacts.

Frontend route asset budgets can be checked with cd frontend && pnpm perf:check. The command measures /login from a normal production build, then performs a production static-demo build for the fixture-backed workspace routes. It measures the unique JavaScript and CSS referenced by representative routes and writes the detailed result to .next/performance-results.json.

The script workflows extension adds inspectable JavaScript orchestration through full AgentRuns, background launch, bounded parallel execution, progress, retained results and cancellation. It is disabled by default and requires a POSIX Gateway and operator-authorized Agent references. Versioned definitions, retained step outputs, live-process pause/resume and authenticated human checkpoints support staged inspection reports. Confirmation binds the exact retained output or captured file version (SHA-256); preview and downstream reads use the same immutable bytes. Each child retains host-generated assembly evidence for its resolved instructions, model, tools and Skills. Gateway restart ends resumability. There is no restart replay or hard monetary budget. Workflow deadlines and child-observation timeouts are reported separately from uncertain admission. Accepted launches retain their request identity across inspection refresh failures, with a separate inspection retry action. A runnable example, installation steps and QuickJS source-build prerequisites for other Python/CPU combinations are included.

This project is open source and available under the MIT License.

DeerFlow is built upon the incredible work of the open-source community. We are deeply grateful to all the projects and contributors whose efforts have made DeerFlow possible. Truly, we stand on the shoulders of giants.

We would like to extend our sincere appreciation to the following projects for their invaluable contributions:

• LangChain: Their exceptional framework powers our LLM interactions and chains, enabling seamless integration and functionality.
• LangGraph: Their innovative approach to multi-agent orchestration has been instrumental in enabling DeerFlow's sophisticated workflows.

These projects exemplify the transformative power of open-source collaboration, and we are proud to build upon their foundations.

A heartfelt thank you goes out to the core authors of DeerFlow, whose vision, passion, and dedication have brought this project to life:

• Daniel Walnut
• Henry Li

Your unwavering commitment and expertise have been the driving force behind DeerFlow's success. We are honored to have you at the helm of this journey.