deer-flow

🦌 DeerFlow - 2.0

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

[Python] [Node.js] [License: MIT]

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.

https://github.com/user-attachments/assets/a8bcadc4-e040-4cf2-8fda-dd768b999c18

[!NOTE] DeerFlow 2.0 is a ground-up rewrite. It shares no code with v1. If you're looking for the original Deep Research framework, it's maintained on the 1.x branch — contributions there are still welcome. Active development has moved to 2.0.

Official Website

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.

Sister Projects

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

Coding Plan from ByteDance Volcengine

• 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

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)

---

Table of Contents

• 🦌 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

One-Line Agent Setup

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.

Quick Start

Configuration

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.

Run event streams include run.terminal for the owning worker's confirmed final Run status, while run.end continues to describe each root graph completion. Terminal publication is best effort; event persistence failure does not undo the Run outcome or prevent stream completion. Recovery with built-in SQL Run and Event stores sharing an engine uses one pooled connection for each guarded event write, so it can publish with only one pool slot available. See Run Event Stream.

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. Newline compatibility preserves literal escapes inside backtick and tilde code fences, including unclosed fences, in complete and streamed replies.

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.

   Unbrowse web fetch preserves the extracted text but prefixes a warning when the target page's HTTP status is 400–599, even when the scraping service itself returns a successful response. Empty-page errors retain that status too. Both messages include up to 500 characters of the reported final URL after removing any username and password; final URLs that cannot be parsed as HTTP(S) URLs with a host are omitted.

   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.

   SearXNG web search contains malformed HTTP-success payloads the same way: a payload that is not an object, a body that is not JSON, a results container that is not a list, or a non-empty list holding no object entries returns the provider's format error, while non-object entries in an otherwise valid page are skipped in order, each skip logged with its entry position, so the valid rows survive and provider-side drift leaves a trace. Missing, null, or empty results keep the existing empty result ([]). This matters most for a self-hosted instance: one behind a reverse proxy or captive portal answers an HTML or plain-text error page with HTTP 200, and reading that as an empty search ends a research loop as "no hits".

   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.

   InfoQuest web and image search screen result URLs through the shared public-URL policy and omit rejected rows. Responses without the required search_result envelope return a format error instead of exposing the raw provider payload.

   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.

   At DEBUG level, Jina emits one bounded crawl completion summary with attempts, elapsed/backoff seconds, outcome and stop reason. See diagnostics for configuration and correlation.

   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. The setup wizard and health check support legacy terminal encodings, including Windows cp936, replacing unsupported characters so both commands can finish. 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). It checks an existing DEER_FLOW_CONFIG_PATH first, then DEER_FLOW_PROJECT_ROOT/config.yaml, before the checkout's default locations. Local startup also accepts the project-root configuration when the checkout has no config.yaml. 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. For older runs without a per-model usage breakdown, Console attributes the recorded total and input token counts to the run's model.

   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.

   The Moonshot (China) preset recognizes kimi-k3.1-pro and kimi-k3.1-pro-highspeed with required thinking, image input and native low / high / max reasoning effort (default max). Thinking stays on even when a caller requests it off. Requests to the official endpoint omit the unsupported thinking switch and fixed sampling parameters. See the model registration reference for the image-input and parameter constraints.

   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.

   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: api-route-gpt-6.1-sol
       display_name: GPT 6.1 Sol (API Route)
       use: langchain_openai:ChatOpenAI
       model: gpt-6.1-sol
       api_key: $API_ROUTE_API_KEY
       base_url: https://global.api-route.com/v1
   
     - 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).

   API Route is also available in the setup wizard. It saves your key as API_ROUTE_API_KEY and uses the Chat Completions endpoint above. Model availability depends on your key's group and permissions; check the authenticated GET https://global.api-route.com/v1/models catalog before replacing the model in config.yaml, and use its returned ID unchanged without an upstream provider prefix.

   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.

   • ClaudeChatModel honors integer-second and HTTP-date Retry-After hints in its provider retry loop. Dates use UTC; past dates retry immediately, while invalid hints or delays beyond 24 hours retain local backoff.

   • 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.

   • If an ACP agent exits before answering (for example because it is not authenticated or a configured argument is wrong), DeerFlow reports its exit code within a few seconds instead of waiting for timeout_seconds, and records the end of the agent's stderr in the Gateway log. Agent stderr is read continuously, so an agent that logs heavily cannot stall on a full pipe.

   • 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
   

   make dev and make start read .env as data, never as a shell script, so a value means what the Gateway reads: $$, backticks and $(...) stay as written, and an unquoted value may contain spaces. $NAME, ${NAME}, ${NAME:-default} and an unquoted leading ~ still expand, and single-quoted values stay literal.

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.

Custom blob storage backends accept blob_storage.backend: module.Class or module:Class, including a single .py module importable by the Gateway. The selected class must implement BlobStore; backend options go under blob_storage.backend_config. See blob storage.

Viewed-image artifact URLs can use shared blobs when the serving Gateway lacks the local file. This fallback requires an existing thread record accessible to the caller; deleting that record revokes access even if checkpoint cleanup fails. When blob storage is disabled, a missing local file returns 404 without loading checkpoint state. Shared-store failures also return 404 and emit a warning with the exception type, without backend error details.

Running the Application

Deployment Sizing

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

Deployment target	Starting point	Recommended	Notes
Local evaluation / make dev	4 vCPU, 8 GB RAM, 20 GB free SSD	8 vCPU, 16 GB RAM	Good for one developer or one light session with hosted model APIs. 2 vCPU / 4 GB is usually not enough.
Docker development / make docker-start	4 vCPU, 8 GB RAM, 25 GB free SSD	8 vCPU, 16 GB RAM	Image builds, bind mounts, and sandbox containers need more headroom than pure local dev.
Long-running server / make up	8 vCPU, 16 GB RAM, 40 GB free SSD	16 vCPU, 32 GB RAM	Preferred for shared use, multi-agent runs, report generation, or heavier sandbox workloads.

• 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.

Option 1: Docker (Recommended)

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). Direct Compose users can add --profile provisioner to the dev command for the same mode; the default stack does not require Kubernetes. With USERDATA_PVC_NAME, sandbox data, extra mounts and Lark broker credentials share one user-data volume with separate subPath mounts, avoiding duplicate references to the same claim on NFS CSI drivers.

Docker builds use the upstream uv registry by default. The gateway image installs with uv sync --locked, and uv.lock pins https://pypi.org/simple, so UV_INDEX_URL has to keep that value: exporting a mirror such as https://pypi.tuna.tsinghua.edu.cn/simple fails the build with The lockfile at `uv.lock` needs to be updated, but `--locked` was provided (reproduce in an already-synced environment with cd backend && UV_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple uv sync --locked --check; if the environment itself is stale, --check reports that instead for either index, so use uv lock --check to see it without syncing). Installing from a mirror requires regenerating the lock or dropping --locked, and both change what the image installs, so keep the default index for a checkout you intend to contribute back. NPM_REGISTRY=https://registry.npmmirror.com is independent of this: it applies to the frontend image's pnpm install and to the gateway image's Lark CLI install (@larksuite/cli, which backend/Dockerfile installs with npm).

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.

After native runtime shutdown and MCP teardown, Gateway closes an initialized sandbox provider, including AIO sandboxes first created by a worker-thread tool call. It does not start an unused sandbox provider solely for cleanup.

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 and IM channel runs. A recursion_limit in an IM channel's session.config takes precedence, and GitHub channel runs keep their own floor. Embedded DeerFlowClient runs retain their own defaults and per-call override paths. A run that reaches its limit stops with a plain message that names up to five files it created in /mnt/user-data/outputs without presenting them; send a message to continue. IM channels reply with the same message.

A custom frontend can add an optional context.client to a Gateway run request: a client name, capabilities such as artifacts or charts, and display preferences. DeerFlow sanitizes it and passes the values to the agent as untrusted, request-scoped hints, so skills can pick an output shape the client can show. It is never an authorization or identity signal. See Create Run. Client hints are omitted from stored run configuration and must be sent again for each run; later runs, including server-started tasks, do not inherit them. 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.

In delta mode, a run selecting an older checkpoint reports an error before rewriting state if either the selected checkpoint or the current head has pending tool execution. For Command(resume=...), other pending graph nodes also prevent the rewrite, except for the synthetic input entry node. Resume pending work from the latest checkpoint; to regenerate completed work, submit new input there. Full-mode resumes and ordinary replay without pending execution keep their existing behavior.

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. Returned state snapshots have independent nested values, so editing one cannot change later checkpoint reads. Redis thread cleanup treats a configured key_prefix as literal text, including glob metacharacters, so deletion stays within that namespace and thread. Redis rejects malformed write records on ingestion; malformed or wrong-channel cached histories are rebuilt in full from authoritative checkpoints. Seeds and write values remain opaque to the cache: these checks do not validate schema-specific state contents. Checkpoint reads, valid neighboring writes and optional-cache cleanup continue even if a warning handler fails.

Gateway checkpoint cleanup runs when the next run wins durable same-thread admission, before its first checkpoint read. It checks only the root head and its parent, removes at most one metadata-only duration leaf without pending writes, and preserves explicit checkpoint selections and delta replay state. See checkpoint operations for the retention constraints.

[!TIP] On Linux, if Docker-based commands fail with permission denied while trying to connect to the Docker daemon socket at unix:///var/run/docker.sock, add your user to the docker group and re-login before retrying. See CONTRIBUTING.md for the full fix.

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.

Production Docker commands can select a root dotenv profile with DEER_FLOW_COMPOSE_ENV_FILE=/srv/deer-flow/stage.env make up. The same file feeds Compose interpolation, Gateway/provisioner variables, and deployment secret checks. YAML and runtime-state locations remain separate selectors; see production env files.

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.

The bundled Docker, local and Helm nginx configs proxy Live Browser WebSocket upgrades with a 600-second read timeout for quiet streams. Helm uses nginx's default connection and send timeouts for the in-cluster Gateway. Any additional Ingress or reverse proxy must also allow WebSocket upgrades.

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.

If an authentication refresh fails during a temporary network outage, a visible tab rechecks the unknown session when connectivity returns or on a later tab switch. Reconnects received in a background tab are remembered until it becomes visible, without waiting for the tab-switch throttle. Identity is restored only after a successful auth response; a confirmed expired session or explicit logout still requires signing in again.

Password forms accept non-ASCII email addresses and trim surrounding email whitespace. Sign-in requires at least six password characters; creating an account or choosing a new password requires eight. These checks run before sending a request. Passwords, including surrounding spaces, are submitted unchanged.

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.

After an SSO sign-in, DeerFlow returns you to the page that asked you to sign in instead of the default workspace. Opening the login page while already signed in does the same for its next target.

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.

[!IMPORTANT] The Gateway still owns active run tasks in process, so production defaults to a single Gateway worker (GATEWAY_WORKERS=1). Multi-worker deployments require Postgres, the Redis stream bridge (stream_bridge.type: redis), run_ownership.heartbeat_enabled: true, and run_events.backend: db; process-local memory/JSONL event stores cannot enforce singleton delivery receipts across workers. Kubernetes replicas run one worker per Pod, which the worker count cannot see: declare them with deployment.multi_instance: true (or DEER_FLOW_MULTI_INSTANCE=1, which deploy tooling such as a Helm chart can set from its replica count) so the startup gate enforces the same prerequisites instead of staying inert. Declared instances that store credentials (channel_connections, or MCP servers using the OAuth refresh_token grant, whose rotating tokens they share through the database) must also share one at-rest DEER_FLOW_CREDENTIALS_KEY (the Helm chart and make up generate it; back it up, see backend/docs/CONFIGURATION.md). The bridge shares SSE delivery and bounded Last-Event-ID replay across workers. When a trimmed reconnect cursor cannot be verified for complete replay, or a subscriber that already established an empty-stream wait falls behind before its first delivery, Memory and Redis emit a machine-readable SSE gap event instead of silently returning a partial replay; the Web UI reloads durable thread/event state and resumes from the retained tail. The Memory bridge can replay without recovery when the cursor is its exact last evicted event: every successor is still retained. Older or unverifiable evicted cursors still trigger recovery. An already-observed Redis stream removed during replay also signals a gap, so slow subscribers reload durable state instead of heartbeating forever after cleanup. Redis also tracks terminal stream producers with a short lease: if a worker dies after committing a terminal run status but before publishing END, SSE and /wait finish after the lease expires instead of waiting for the retained stream's TTL. Lease reconciliation marks runs from dead workers as errors, persists their delivery receipts, publishes the terminal stream marker, schedules retained-stream cleanup, and updates the affected thread status. Redis END publication is atomic and idempotent: owner and recovery retries confirm an existing marker without appending another, events cannot be appended once END is the stream tail, and an unowned recovery retry after cleanup does not recreate the stream. Legacy streams with an earlier END followed by late data can still accept data. The Redis compatibility scan (stream_bridge.legacy_end_scan_enabled: true) also runs on first END publication; disable it only after old writers stop and all legacy replay buffers expire or are removed, then restart every Gateway to use tail-only END detection. SSE, /wait, and internal stream consumers use stream_bridge.heartbeat_interval_seconds (default 15) for idle liveness checks; changing it requires a Gateway restart. Malformed Redis reconnect IDs live-tail new events instead of replaying the retained buffer, and the rolling retained-buffer TTL (stream_ttl_seconds) remains a cleanup safety net rather than a run timeout. Failed-login counters and lockouts for POST /api/v1/auth/login/local (auth.local.max_login_attempts / lockout_seconds) are kept in the shared login_throttle table whenever the application database is SQLite or Postgres (auth.local.throttle_storage: auto, the default), so every replica enforces one lockout per client IP; memory keeps the historical per-process counter, which under N replicas hands an attacker N × max_login_attempts guesses and logs a startup warning. IM chat-to-thread bindings live in the shared channel_thread_bindings table whenever the application database is SQLite or Postgres (an existing channels/store.json is imported once at startup and renamed store.json.migrated); IM bot credentials entered in the web UI live in the shared channel_runtime_configs table, encrypted with DEER_FLOW_CREDENTIALS_KEY (an existing channels/runtime-config.json is imported once and then deleted), so a save on any instance reaches the provider's connection within one lease renewal; each provider's platform connection is held by one instance through the shared channel_leases table — channels.leader_lease (default auto, on wherever the multi-process signal is) hands it to another instance when the holder dies, and the startup gate refuses enabled: false while an IM channel other than GitHub is enabled (see IM channel leader lease); follow-up buffers still need their own multi-worker coordination. To try this topology on one machine, make dev-multi starts two Gateways on shared Postgres, Redis and DEER_FLOW_HOME, and make dev-multi-check runs the cross-instance checks; see the local two-Gateway harness in backend/docs/CONFIGURATION.md.

The producer-lease recovery behavior above is gated by stream_bridge.producer_lease_recovery_enabled. Keep it disabled while old Gateway binaries or their runs remain; enable it on every instance only after the writer rollout and drain complete.

With run ownership heartbeats enabled, terminal success requires acknowledgment from that worker's own durable write. Run creation and replacement serialize per conversation: a slow admission does not hold the manager-wide lock needed by another conversation's cached run reads or heartbeat snapshots. Their own Store I/O can still encounter database or pool contention. New runs remain hidden until their initial Store write succeeds. A matching database row alone cannot confirm success, and completion metadata cannot retry a rejected terminal transition. Each run renews independently, so a stalled store operation does not delay other runs. If acknowledgment remains unavailable when the last confirmed lease expires, the worker stops conservatively even if its commit actually reached the database. Expired leases cannot be renewed. Interrupt-and-replace admission acknowledges only the old runs actually claimed by its atomic transaction. Rollback keeps renewing its acknowledged interrupted row until restoration and the separate final error commit finish. An active-only renewal that raced that admission gets one bounded retry after the interrupted acknowledgment, without weakening the owner or live-lease checks. A failed renewal leaves a local expiry guard armed. Shutdown bounds its wait for orphan recovery, but cancellation-resistant storage may still be using its resources.

The local two-Gateway harness disables IM providers under both channels and channel_connections by default. DEERFLOW_MI_KEEP_CHANNELS=1 preserves them; channels.leader_lease stays as configured.

Cleaning up an in-memory run stream, or closing its bridge, releases retained payloads and signals a replay gap to subscribers with unfinished replay. They reload durable state instead of waiting on a detached buffer or consuming stale history. Established readers that fully consumed a terminal stream retain their normal end signal. A new reader whose initial attachment loses the race with cleanup receives a gap so it can recover the saved result.

During SSE replay-gap recovery, the Web UI preserves a reconnect pointer claimed by another run while durable state is loading. An older recovery cannot overwrite that pointer when it rejoins its own stream. Even a stale foreign pointer is preserved, so the next reload may first check that run before active-run recovery rejoins, adding a reconnect cycle.

Store-only SSE and /wait consumers can finish after a heartbeat when run-manager or scheduled-task reconciliation recovers a dead worker's run, even if no END marker was published. Ordinary completions still wait for END to preserve tail events.

In single-process JSONL deployments, cancelling an admitted event-store mutation waits for its background file I/O, rollback, and bookkeeping to settle before releasing the thread write lock. This prevents an older cancelled write from recreating deleted records or rolling back a later successful write. Cancellation can therefore wait on slow storage; it does not stop an in-flight filesystem operation. Callers still waiting to acquire the lock can cancel without starting a mutation. A batch spanning multiple threads drains its current thread group before propagating cancellation; subsequent thread groups do not start.

After a run publishes its terminal stream marker, its process-local RunRecord remains available for the existing five-minute grace period before cleanup; durable run history remains available through RunStore, while the stream bridge retains its delivery tail on its separate cleanup schedule.

With in-memory run storage, replacing a run snapshot under another thread removes its previous thread membership. Old-thread history and usage queries exclude it, and deleting the old thread preserves the reassigned run.

Run cancellation may land on any Gateway worker. A non-owning worker now persists the interrupt or rollback request for the live owner, which observes it during lease renewal and performs the normal cancellation flow; load-balancer routing alone no longer produces a 409. The first accepted action wins even if a retry lands on the owner, and accepted cancellation competes atomically with owner completion. Dead owners still follow lease takeover and orphan recovery. Cancellation latency is therefore bounded by the lease heartbeat interval.

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 and HTTP-date hints in LLM error handling and Codex's retry path, and integer-second hints in Claude's retry path. Codex prioritizes numeric retry-after-ms hints; past HTTP-date hints retry immediately.

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.

Upgrading an existing checkout

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.

To continue a thread paused at a static interrupt_before or interrupt_after breakpoint, submit input: null (or omit input) to its /runs, /runs/wait, or /runs/stream endpoint. This continues pending graph work with the source turn's knowledge scope from the selected checkpoint (the head by default). An explicit empty object is new input, not a continuation; dynamic interrupts still require command.resume. In delta mode, historical continuation rejects rewrites that would discard pending execution; continue at the head instead. Continuations retain the checkpoint's MCP background_tasks projection; only dictionary input refreshes it from the task repository. Old null-input runs stored as {} cannot safely match a keyed null-input retry after upgrading and return 409. Check the original run before submitting a new key.

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.

Gateway runs also assign IDs to input messages that omit one before storing the run record and invoking the graph. This keeps the message visible as one turn when the Web UI reconnects to an active run started by an API client. During a rolling upgrade, route keyed runs with ID-less input and their retries to upgraded workers until all Gateway workers are upgraded; older workers may reject those retries with 409. See the run API contract.

Option 2: Local Development

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 a distinct timestamped backup directory for every import 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.

Local startup waits up to 30 seconds for the Gateway port. On a slow cold start, set a longer wait without editing the launcher:

DEER_FLOW_GATEWAY_STARTUP_TIMEOUT=240 make dev

The value must be a positive integer in seconds. You can also set it in the root .env, whose value takes precedence over the shell. It applies to local make dev, make start, and their daemon variants, not Docker health checks or LLM request timeouts. Check logs/gateway.log for startup errors before increasing the wait; an expired deadline still stops the services.

Startup Modes

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

	Local Foreground	Local Daemon	Docker Dev	Docker Prod
Dev	./scripts/serve.sh --devmake dev	./scripts/serve.sh --dev --daemonmake dev-daemon	./scripts/docker.sh startmake docker-start	—
Prod	./scripts/serve.sh --prodmake start	./scripts/serve.sh --prod --daemonmake start-daemon	—	./scripts/deploy.shmake up

Action	Local	Docker Dev	Docker Prod
Stop	./scripts/serve.sh --stopmake stop	./scripts/docker.sh stopmake docker-stop	./scripts/deploy.sh downmake down
Restart	./scripts/serve.sh --restart [flags]	./scripts/docker.sh restart	—

Local and Docker development stop commands also stop running sandbox containers whose names start with deer-flow-sandbox-. Containers with that text elsewhere in their names are left alone.

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.

LangGraph Studio (Optional)

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.

Docker Production Deployment

./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

Advanced

Sandbox Mode

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)

The optional Tenki provider retains failed termination handles for retry and rejects late acquisitions after reset or shutdown. If shutdown reports outstanding work, retry after its SDK workers finish or the remote service recovers. At process exit, outstanding cleanup is logged without an exception traceback.

In Local Execution, glob and grep can search the skills root (default /mnt/skills) across its mounted categories. Patterns stay relative to that root, and one result limit covers all categories. User isolation, disabled-skill filtering and read-only mounts still apply. Configured category mounts such as build or logs remain searchable; ignore rules apply inside each mount. A broken category mount stops the root search, and the Gateway warning log identifies its virtual path.

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.

AIO teardown failures retain a generation-bound cleanup record. Partial shutdown reports failure and can be retried; retries verify the runtime identity and absence before completing cleanup. Healthy replacements with the same logical ID remain reusable, including after background recovery into the warm pool.

See the Sandbox Configuration Guide to configure your preferred mode. For absolute skills paths and custom directory mappings on Linux, macOS, Windows, WSL, or a Docker Gateway, see Skills and sandbox paths.

Managed async AIO runs can opt into bounded sandbox prewarming to overlap startup with the first model call. Direct SDK graphs keep lazy acquisition. It defaults off; see the guide and sandbox benchmarks before enabling it on a shared host.

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.

The sandbox benchmark also supports OpenSandbox cold acquisition, warm reuse and concurrent scopes, using the configured service connection and reporting unconfirmed cleanup as a failed run. Live runs require a disposable OpenSandbox service and runtime home.

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.

Successful read_file calls treat a JSON document's error field as file data, not a failure of the read tool. This keeps receipts accurate when inspecting saved API error responses and avoids blocking later reads when the optional tool_progress guard is enabled. Actual read failures and output truncation still retain their error or partial-result status.

MCP Server

Tool steps mark explicit execution failures. Scheduled-task steps show completed operations only after a valid response, rather than when the request is issued.

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. Free-text sandbox paths retain spaces inside matching ASCII quotes, backticks, or curly quote pairs, including backslash-escaped wrappers. Unterminated spans and delimited values containing a line break or tab are skipped. Free-text artifact references retain balanced ASCII and CJK brackets in literal paths and URLs while trimming surrounding prose punctuation. 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. Inline MCP audio results are not supported. Stdio tools report an explicit error, matching HTTP/SSE tools, rather than treating base64 audio as a successful text result. 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. Durable background-task status and cancellation calls renew their claims while they run, so a slow response can still be saved after the original lease duration. If ownership is lost or renewal cannot be confirmed, the worker stops its local call and leaves recovery to the durable state. See control-call leases for recovery and shutdown behavior. 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. Unexpected failures when changing deployment or personal MCP configuration return fixed error messages; exception details stay in server logs. 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 MCP schema promotions last for the current user turn. They remain active across that turn's model/tool loop and checkpoint resume, then become deferred again when a new user messag