deer-flow/config.example.yaml

# Configuration for the DeerFlow application
#
# Guidelines:
# - Copy this file to `config.yaml` and customize it for your environment
# - The default path of this configuration file is `config.yaml` in the project root.
#   You can set `DEER_FLOW_PROJECT_ROOT` to define that root explicitly, or use
#   `DEER_FLOW_CONFIG_PATH` to point at a specific config file.
# - Runtime state defaults to `.deer-flow` under the project root. Override it
#   with `DEER_FLOW_HOME` when you need a different writable data directory.
# - Environment variables are available for all field values. Example: `api_key: $OPENAI_API_KEY`
# - The `use` path is a string that looks like "package_name.sub_package_name.module_name:class_name/variable_name".

# ============================================================================
# Config Version (used to detect outdated config files)
# ============================================================================
# Bump this number when the config schema changes.
# Run `make config-upgrade` to merge new fields into your local config.yaml.
config_version: 26

# ============================================================================
# Logging
# ============================================================================
# Log level for deerflow modules (debug/info/warning/error)
log_level: info

# Request trace correlation for Gateway logs, HTTP response headers, and
# Langfuse metadata. Disabled by default to preserve existing HTTP/log output.
logging:
  enhance:
    enabled: false
    format: text

# ============================================================================
# Tracing / Observability (Monocle)
# ============================================================================
# Optional agent tracing via Monocle. Configured through environment variables
# (MONOCLE_TRACING, MONOCLE_EXPORTERS, OKAHU_API_KEY — like LangSmith/Langfuse),
# not config.yaml keys, and OFF by default. See README.md → "Monocle Tracing"
# for setup, what each exporter captures, and where the trace data goes.

# ============================================================================
# Token Usage
# ============================================================================
# Enable token usage collection and display.
# When enabled, DeerFlow records input/output/total tokens per model call
# and shows usage metadata in the workspace UI when providers return it.
token_usage:
  enabled: true

# ============================================================================
# Token Budget — Per-run token limits
# ============================================================================
# Prevents runaway API costs by enforcing hard token limits per run.
# When warn_threshold is crossed, the agent receives an in-context warning.
# When hard_stop_threshold is crossed, tool_calls are stripped and the agent
# is forced to produce a final answer immediately.
token_budget:
  enabled: false                # Set to true to activate budget enforcement
  max_tokens: 200000            # Total token limit (input + output) per run
  max_input_tokens: null        # Optional separate input-only limit
  max_output_tokens: null       # Optional separate output-only limit
  warn_threshold: 0.8           # Warn at 80% of the budget
  hard_stop_threshold: 1.0      # Force stop at 100% of the budget

# ============================================================================
# Recursion Limit — Hard ceiling for a client-supplied run recursion_limit
# ============================================================================
# A run's recursion_limit caps the number of LangGraph super-steps (each is at
# least one LLM call). The Gateway never trusts a client-supplied value
# verbatim: any value above this ceiling is clamped down to it, preventing
# runaway API cost / DoS. Invalid or non-positive client values fall back to
# the server default of 100. Raise this only if you legitimately run very
# deeply nested subagent graphs.
max_recursion_limit: 1000


# ============================================================================
# Models Configuration
# ============================================================================
# Configure available LLM models for the agent to use
#
# Optional per-model pricing (powers the real-cost display on the workspace
# console). Add a `pricing` block to any model entry; use ONE currency across
# all models. Prices are per one million tokens.
#
#   pricing:
#     currency: CNY                    # ISO code shown in the console (CNY, USD, ...)
#     input_per_million: 8.0           # price per 1M input tokens (cache miss)
#     output_per_million: 32.0         # price per 1M output tokens
#     input_cache_hit_per_million: 0.8 # price per 1M cache-hit input tokens
#                                      # (optional; omit → hits billed at miss price)

models:
  # Example: Volcengine (Doubao) model
  # - name: doubao-seed-1.8
  #   display_name: Doubao-Seed-1.8
  #   use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
  #   model: doubao-seed-1-8-251228
  #   api_base: https://ark.cn-beijing.volces.com/api/v3
  #   api_key: $VOLCENGINE_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   supports_thinking: true
  #   supports_vision: true
  #   supports_reasoning_effort: true
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: Volcengine Coding Plan (one key, multi-vendor gateway)
  # The Coding Plan endpoint (/api/coding/v3) lets you access models from
  # Doubao, GLM, DeepSeek, Kimi, and MiniMax with a single API key.
  # Each model may differ in thinking/vision support - configure per-model.
  #
  # - name: glm-5.2-cp
  #   display_name: GLM-5.2 (Coding Plan)
  #   use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
  #   model: glm-5.2
  #   api_base: https://ark.cn-beijing.volces.com/api/coding/v3
  #   api_key: $VOLCENGINE_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   supports_thinking: true
  #   supports_vision: false
  #   supports_reasoning_effort: true
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled
  #
  # - name: deepseek-v4-pro-cp
  #   display_name: DeepSeek-V4-Pro (Coding Plan)
  #   use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
  #   model: deepseek-v4-pro
  #   api_base: https://ark.cn-beijing.volces.com/api/coding/v3
  #   api_key: $VOLCENGINE_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   supports_thinking: true
  #   supports_vision: false
  #   supports_reasoning_effort: true
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled
  # Example: OpenAI model
  # - name: gpt-4
  #   display_name: GPT-4
  #   use: langchain_openai:ChatOpenAI
  #   model: gpt-4
  #   api_key: $OPENAI_API_KEY # Use environment variable
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 0.7
  #   supports_vision: true # Enable vision support for view_image tool

  # Example: OpenAI Responses API model
  # - name: gpt-5-responses
  #   display_name: GPT-5 (Responses API)
  #   use: langchain_openai:ChatOpenAI
  #   model: gpt-5
  #   api_key: $OPENAI_API_KEY
  #   request_timeout: 600.0
  #   max_retries: 2
  #   use_responses_api: true
  #   output_version: responses/v1
  #   supports_vision: true

  # Example: Ollama (native provider — preserves thinking/reasoning content)
  #
  # IMPORTANT: Use langchain_ollama:ChatOllama instead of langchain_openai:ChatOpenAI
  # for Ollama models. The OpenAI-compatible endpoint (/v1/chat/completions) does NOT
  # return reasoning_content as a separate field — thinking content is either flattened
  # into <think> tags or dropped entirely (ollama/ollama#15293). The native Ollama API
  # (/api/chat) correctly separates thinking from response content.
  #
  # Install: cd backend && uv pip install 'deerflow-harness[ollama]'
  #
  # - name: qwen3-local
  #   display_name: Qwen3 32B (Ollama)
  #   use: langchain_ollama:ChatOllama
  #   model: qwen3:32b
  #   base_url: http://localhost:11434   # No /v1 suffix — uses native /api/chat
  #   num_predict: 8192
  #   temperature: 0.7
  #   reasoning: true                    # Passes think:true to Ollama native API
  #   supports_thinking: true
  #   supports_vision: false
  #
  # - name: gemma4-local
  #   display_name: Gemma 4 27B (Ollama)
  #   use: langchain_ollama:ChatOllama
  #   model: gemma4:27b
  #   base_url: http://localhost:11434
  #   num_predict: 8192
  #   temperature: 0.7
  #   reasoning: true
  #   supports_thinking: true
  #   supports_vision: true
  #
  # For Docker deployments, use host.docker.internal instead of localhost:
  #   base_url: http://host.docker.internal:11434

  # Example: Anthropic Claude model (with extended thinking)
  # supports_thinking: true is required — without it, DeerFlow silently falls
  # back to non-thinking mode even when the UI thinking toggle is on.
  # budget_tokens is required by the Anthropic API when thinking.type=enabled
  # (no server default; min 1024; must be less than max_tokens).
  # - name: claude-sonnet-4
  #   display_name: Claude Sonnet 4
  #   use: langchain_anthropic:ChatAnthropic
  #   model: claude-sonnet-4-20250514
  #   api_key: $ANTHROPIC_API_KEY
  #   default_request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 16000
  #   supports_vision: true
  #   supports_thinking: true
  #   when_thinking_enabled:
  #     thinking:
  #       type: enabled
  #       budget_tokens: 4096   # required; min 1024; must be < max_tokens
  #   when_thinking_disabled:
  #     thinking:
  #       type: disabled

  # Example: Google Gemini model (native SDK, no thinking support)
  # - name: gemini-2.5-pro
  #   display_name: Gemini 2.5 Pro
  #   use: langchain_google_genai:ChatGoogleGenerativeAI
  #   model: gemini-2.5-pro
  #   gemini_api_key: $GEMINI_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   supports_vision: true

  # Example: Gemini model via OpenAI-compatible gateway (with thinking support)
  # Use PatchedChatOpenAI so that tool-call thought_signature values on tool_calls
  # are preserved across multi-turn tool-call conversations — required by the
  # Gemini API when thinking is enabled. See:
  # https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures
  # - name: gemini-2.5-pro-thinking
  #   display_name: Gemini 2.5 Pro (Thinking)
  #   use: deerflow.models.patched_openai:PatchedChatOpenAI
  #   model: google/gemini-2.5-pro-preview   # model name as expected by your gateway
  #   api_key: $GEMINI_API_KEY
  #   base_url: https://<your-openai-compat-gateway>/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 16384
  #   supports_thinking: true
  #   supports_vision: true
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: Xiaomi MiMo model (with thinking support)
  # MiMo thinking mode returns reasoning_content and requires that field to be
  # replayed on historical assistant messages in multi-turn agent/tool-call
  # conversations. Use PatchedChatMiMo instead of plain ChatOpenAI.
  # Use https://api.xiaomimimo.com/v1 with pay-as-you-go `sk-...` keys.
  # Use your Token Plan regional URL (for example
  # https://token-plan-cn.xiaomimimo.com/v1) with Token Plan `tp-...` keys.
  # PatchedChatMiMo is model-id agnostic; use it for every MiMo thinking model
  # entry you configure (for example mimo-v2.5-pro, mimo-v2.5, mimo-v2-pro,
  # mimo-v2-omni, or mimo-v2-flash), including models referenced by subagent
  # model overrides.
  # See: https://platform.xiaomimimo.com/docs/en-US/usage-guide/passing-back-reasoning_content
  # - name: mimo-v2.5-pro
  #   display_name: MiMo V2.5 Pro
  #   use: deerflow.models.patched_mimo:PatchedChatMiMo
  #   model: mimo-v2.5-pro
  #   api_key: $MIMO_API_KEY
  #   base_url: https://api.xiaomimimo.com/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   supports_thinking: true
  #   supports_vision: false
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: DeepSeek V4 model (with thinking support)
  # - name: deepseek-v4
  #   display_name: DeepSeek V4 (Thinking)
  #   use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
  #   model: deepseek-v4-pro
  #   api_key: $DEEPSEEK_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   supports_thinking: true
  #   supports_vision: false  # DeepSeek V4 does not support vision
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: Kimi K2.5 model
  # - name: kimi-k2.5
  #   display_name: Kimi K2.5
  #   use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
  #   model: kimi-k2.5
  #   api_base: https://api.moonshot.cn/v1
  #   api_key: $MOONSHOT_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 32768
  #   supports_thinking: true
  #   supports_vision: true  # Check your specific model's capabilities
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: Novita AI (OpenAI-compatible)
  # Novita provides an OpenAI-compatible API with competitive pricing
  # See: https://novita.ai
  # - name: novita-deepseek-v3.2
  #   display_name: Novita DeepSeek V3.2
  #   use: langchain_openai:ChatOpenAI
  #   model: deepseek/deepseek-v3.2
  #   api_key: $NOVITA_API_KEY
  #   base_url: https://api.novita.ai/openai
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 0.7
  #   supports_thinking: true
  #   supports_vision: true
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: StepFun (阶跃星辰) reasoning models
  # StepFun provides OpenAI-compatible API with reasoning models.
  # With reasoning_format: deepseek-style, the API returns reasoning_content
  # (same field as DeepSeek), which must be replayed on historical assistant
  # messages in multi-turn tool-call conversations.
  # Use PatchedChatStepFun instead of plain ChatOpenAI.
  # Docs: https://platform.stepfun.com/docs/api-reference/chat-completions
  # - name: step-3.7-flash
  #   display_name: Step 3.7 Flash
  #   use: deerflow.models.patched_stepfun:PatchedChatStepFun
  #   model: step-3.7-flash
  #   api_key: $STEPFUN_API_KEY
  #   base_url: https://api.stepfun.com/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   supports_thinking: true
  #   supports_reasoning_effort: true
  #   supports_vision: true
  #   when_thinking_enabled:
  #     extra_body:
  #       reasoning_format: deepseek-style
  #   when_thinking_disabled:
  #     extra_body:
  #       reasoning_format: deepseek-style

  # Example: MiniMax (OpenAI-compatible) - International Edition
  # MiniMax provides high-performance models with 512K context window and 128K max output
  # Docs: https://platform.minimax.io/docs/api-reference/text-openai-api
  # - name: minimax-m3
  #   display_name: MiniMax M3
  #   use: deerflow.models.patched_minimax:PatchedChatMiniMax
  #   model: MiniMax-M3
  #   api_key: $MINIMAX_API_KEY
  #   base_url: https://api.minimax.io/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
  #   supports_vision: true
  #   supports_thinking: true
  #   # PatchedChatMiniMax is the MiniMax adapter: it enables reasoning_split and
  #   # maps MiniMax's structured reasoning into reasoning_content (the field
  #   # DeerFlow understands), and it strips the per-message `name` field that
  #   # DeerFlow middlewares attach — MiniMax rejects requests whose user-message
  #   # names differ with "user name must be consistent (2013)". Declare the
  #   # thinking toggle so non-thinking paths (flash mode, follow-up suggestions,
  #   # title/memory generation) truly disable reasoning instead of spending
  #   # tokens on it.
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: adaptive
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # NOTE: M2.x models always think — passing thinking:{type:disabled} has no
  # effect (per MiniMax docs), so the toggle above is omitted for M2.7. The
  # follow-up-suggestions endpoint strips inline <think> defensively regardless.
  # Still use the PatchedChatMiniMax adapter: it strips the per-message `name`
  # field DeerFlow middlewares attach, which MiniMax otherwise rejects with
  # "user name must be consistent (2013)".
  # - name: minimax-m2.7
  #   display_name: MiniMax M2.7
  #   use: deerflow.models.patched_minimax:PatchedChatMiniMax
  #   model: MiniMax-M2.7
  #   api_key: $MINIMAX_API_KEY
  #   base_url: https://api.minimax.io/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
  #   supports_vision: false  # M2.7 is text-only; M3 supports vision
  #   supports_thinking: true

  # - name: minimax-m2.7-highspeed
  #   display_name: MiniMax M2.7 Highspeed
  #   use: deerflow.models.patched_minimax:PatchedChatMiniMax
  #   model: MiniMax-M2.7-highspeed
  #   api_key: $MINIMAX_API_KEY
  #   base_url: https://api.minimax.io/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
  #   supports_vision: false  # M2.7 is text-only; M3 supports vision
  #   supports_thinking: true

  # Example: MiniMax (OpenAI-compatible) - CN 中国区用户
  # MiniMax provides high-performance models with 512K context window and 128K max output
  # Docs: https://platform.minimaxi.com/docs/api-reference/text-openai-api
  # - name: minimax-m3
  #   display_name: MiniMax M3
  #   use: deerflow.models.patched_minimax:PatchedChatMiniMax
  #   model: MiniMax-M3
  #   api_key: $MINIMAX_API_KEY
  #   base_url: https://api.minimaxi.com/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
  #   supports_vision: true
  #   supports_thinking: true
  #   # PatchedChatMiniMax is the MiniMax adapter: it enables reasoning_split and
  #   # maps MiniMax's structured reasoning into reasoning_content (the field
  #   # DeerFlow understands), and it strips the per-message `name` field that
  #   # DeerFlow middlewares attach — MiniMax rejects requests whose user-message
  #   # names differ with "user name must be consistent (2013)". Declare the
  #   # thinking toggle so non-thinking paths (flash mode, follow-up suggestions,
  #   # title/memory generation) truly disable reasoning instead of spending
  #   # tokens on it.
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: adaptive
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # NOTE: M2.x models always think — passing thinking:{type:disabled} has no
  # effect (per MiniMax docs), so the toggle above is omitted for M2.7. The
  # follow-up-suggestions endpoint strips inline <think> defensively regardless.
  # Still use the PatchedChatMiniMax adapter: it strips the per-message `name`
  # field DeerFlow middlewares attach, which MiniMax otherwise rejects with
  # "user name must be consistent (2013)".
  # - name: minimax-m2.7
  #   display_name: MiniMax M2.7
  #   use: deerflow.models.patched_minimax:PatchedChatMiniMax
  #   model: MiniMax-M2.7
  #   api_key: $MINIMAX_API_KEY
  #   base_url: https://api.minimaxi.com/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
  #   supports_vision: false  # M2.7 is text-only; M3 supports vision
  #   supports_thinking: true

  # - name: minimax-m2.7-highspeed
  #   display_name: MiniMax M2.7 Highspeed
  #   use: deerflow.models.patched_minimax:PatchedChatMiniMax
  #   model: MiniMax-M2.7-highspeed
  #   api_key: $MINIMAX_API_KEY
  #   base_url: https://api.minimaxi.com/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 4096
  #   temperature: 1.0  # MiniMax requires temperature in (0.0, 1.0]
  #   supports_vision: false  # M2.7 is text-only; M3 supports vision
  #   supports_thinking: true

  # Example: OpenRouter (OpenAI-compatible)
  # OpenRouter models use the same ChatOpenAI + base_url pattern as other OpenAI-compatible gateways.
  # - 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: $OPENAI_API_KEY
  #   base_url: https://openrouter.ai/api/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   temperature: 0.7

  # Example: Atlas Cloud (OpenAI-compatible)
  # Atlas Cloud exposes a single OpenAI-compatible endpoint in front of many open
  # models (DeepSeek, Qwen, Kimi, GLM, MiniMax, Llama, ...), so it uses the same
  # ChatOpenAI + base_url pattern as other OpenAI-compatible gateways.
  # Browse model ids at https://api.atlascloud.ai/v1/models — see https://atlascloud.ai
  # - name: atlascloud-deepseek-v3.2
  #   display_name: DeepSeek V3.2 (Atlas Cloud)
  #   use: langchain_openai:ChatOpenAI
  #   model: deepseek-ai/DeepSeek-V3.2-Exp
  #   api_key: $ATLASCLOUD_API_KEY
  #   base_url: https://api.atlascloud.ai/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   temperature: 0.7
  #   supports_vision: false
  #
  # For reasoning models on Atlas Cloud (e.g. a Qwen3 *-thinking id), use the
  # patched OpenAI-compatible adapter so reasoning_content is replayed across
  # multi-turn tool-call conversations:
  # - name: atlascloud-qwen3-thinking
  #   display_name: Qwen3 235B Thinking (Atlas Cloud)
  #   use: deerflow.models.patched_openai:PatchedChatOpenAI
  #   model: qwen/qwen3-235b-a22b-thinking-2507
  #   api_key: $ATLASCLOUD_API_KEY
  #   base_url: https://api.atlascloud.ai/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   supports_thinking: true
  #   supports_vision: false
  #   when_thinking_enabled:
  #     extra_body:
  #       thinking:
  #         type: enabled
  #   when_thinking_disabled:
  #     extra_body:
  #       thinking:
  #         type: disabled

  # Example: vLLM 0.19.0 (OpenAI-compatible, with reasoning toggle)
  # DeerFlow's vLLM provider preserves vLLM reasoning across tool-call turns and
  # toggles Qwen-style reasoning by writing
  # extra_body.chat_template_kwargs.enable_thinking=true/false.
  # Some reasoning models also require the server to be started with
  # `vllm serve ... --reasoning-parser <parser>`.
  # - 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
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 8192
  #   supports_thinking: true
  #   supports_vision: false
  #   when_thinking_enabled:
  #     extra_body:
  #       chat_template_kwargs:
  #         enable_thinking: true


  # Example: Qwen3-Coder deployed on MindIE Engine
  # - name: Qwen3_Coder_480B_MindIE
  #   display_name: Qwen3-Coder-480B (MindIE)
  #   use: deerflow.models.mindie_provider:MindIEChatModel
  #   model: Qwen3-Coder-480B-A35B-Instruct-Client
  #   base_url: http://localhost:8989/v1
  #   api_key: $OPENAI_API_KEY
  #   temperature: 0
  #   max_retries: 1
  #   supports_thinking: false
  #   supports_vision: false
  #   supports_reasoning_effort: false
  #   # --- Advanced Network Settings ---
  #   # Due to MindIE's streaming limitations with tool calling, the provider
  #   # uses mock-streaming (awaiting full generation). Extended timeouts are required.
  #   read_timeout: 900.0     # 15 minutes to prevent drops during long document generation
  #   connect_timeout: 30.0
  #   write_timeout: 60.0
  #   pool_timeout: 30.0

# ============================================================================
# Tool Groups Configuration
# ============================================================================
# Define groups of tools for organization and access control

tool_groups:
  - name: web
  - name: file:read
  - name: file:write
  - name: bash

# ============================================================================
# Tools Configuration
# ============================================================================
# Configure available tools for the agent to use

tools:
  # Web search tool (uses DuckDuckGo, no API key required)
  - name: web_search
    group: web
    use: deerflow.community.ddg_search.tools:web_search_tool
    max_results: 5
    # backend: auto          # DDGS backend(s): auto, duckduckgo, brave, wikipedia, etc.
    # region: wt-wt          # wt-wt is normalized for Wikipedia when backend includes auto/all/wikipedia.
    # safesearch: moderate   # on, moderate, off

  # Web search tool (uses SearXNG, self-hosted, no API key required)
  # SearXNG is a free internet metasearch engine which aggregates results from
  # various search services. Deploy your own instance: https://github.com/searxng/searxng
  # For Docker deployments, use the Docker service name instead of localhost.
  # - name: web_search
  #   group: web
  #   use: deerflow.community.searxng.tools:web_search_tool
  #   base_url: http://localhost:8088   # SearXNG instance URL (default: :8088; Docker: http://searxng:8080)
  #   max_results: 5                     # Maximum number of search results

  # Web search tool (uses Serper - Google Search API, requires SERPER_API_KEY)
  # Serper provides real-time Google Search results. Sign up at https://serper.dev
  # Note: set SERPER_API_KEY in your environment before starting the app.
  # Avoid putting literal API keys in config.yaml; use the $VAR form instead.
  # - name: web_search
  #   group: web
  #   use: deerflow.community.serper.tools:web_search_tool
  #   max_results: 5   # capped at 10 by the Serper provider
  #   # api_key: $SERPER_API_KEY  # Optional explicit env-var reference

  # Web search tool (uses Brave Search API, requires BRAVE_SEARCH_API_KEY)
  # Brave Search returns results from an independent index. Sign up at
  # https://brave.com/search/api/ to get a key. Unlike the DuckDuckGo
  # `backend: brave` option above, this calls the official Brave API directly.
  # - name: web_search
  #   group: web
  #   use: deerflow.community.brave.tools:web_search_tool
  #   max_results: 5                  # Capped at 20 by the Brave Search API
  #   # api_key: $BRAVE_SEARCH_API_KEY  # Optional if the env var is set

  # Web search tool (requires Tavily API key)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.tavily.tools:web_search_tool
  #   max_results: 5
  #   # api_key: $TAVILY_API_KEY  # Set if needed

  # Web search tool (uses InfoQuest, requires InfoQuest API key)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.infoquest.tools:web_search_tool
  #   # Used to limit the scope of search results, only returns content within the specified time range. Set to -1 to disable time filtering
  #   search_time_range: 10

  # Web search tool (uses Exa, requires EXA_API_KEY)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.exa.tools:web_search_tool
  #   max_results: 5
  #   search_type: auto  # Options: auto, neural, keyword
  #   contents_max_characters: 1000
  #   # api_key: $EXA_API_KEY

  # Web search tool (uses Firecrawl, requires FIRECRAWL_API_KEY)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.firecrawl.tools:web_search_tool
  #   max_results: 5
  #   # api_key: $FIRECRAWL_API_KEY

  # Web search tool (uses GroundRoute, requires GROUNDROUTE_API_KEY)
  # GroundRoute is a meta search layer: one API in front of six engines (Serper,
  # Brave, Exa, Tavily, Firecrawl, Perplexity). It routes each query to the cheapest
  # engine that clears a quality bar and fails over if one is down. Pricing is
  # gain-share (you keep about half of any cache savings). Get a key at
  # https://groundroute.ai/keys
  # - name: web_search
  #   group: web
  #   use: deerflow.community.groundroute.tools:web_search_tool
  #   max_results: 5                  # Clamped to 1-50 by GroundRoute
  #   # api_key: $GROUNDROUTE_API_KEY  # Optional if the env var is set
  
  # Web search tool (uses fastCRW - Firecrawl-compatible web scraper, single binary,
  # self-host or cloud. Cloud requires CRW_API_KEY; self-host may need no key.)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.fastcrw.tools:web_search_tool
  #   max_results: 5
  #   # api_key: $CRW_API_KEY
  #   # base_url: https://fastcrw.com/api   # default cloud; set to e.g. http://localhost:3000 for self-host

  # Web fetch tool (uses Browserless - headless Chrome, self-hosted or cloud)
  # Browserless renders pages with a real headless Chrome, ideal for JavaScript-heavy
  # sites and SPAs. Deploy your own: https://github.com/browserless/browserless
  # For Docker deployments, use the Docker service name instead of localhost.
  # NOTE: Only one web_fetch provider can be active at a time.
  # Comment out the Jina AI web_fetch entry below before enabling this one.
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.browserless.tools:web_fetch_tool
  #   base_url: http://localhost:3032    # Browserless instance URL (default: :3032; Docker: http://browserless:3000)
  #   # token: $BROWSERLESS_TOKEN        # API token (required for Browserless Cloud; optional for self-hosted)
  #   timeout_s: 30                       # Request timeout in seconds
  #   # allow_private_addresses: false    # SSRF guard: keep false in production. Set true ONLY for intentional internal targets.
  #   # wait_for_event: "networkidle"    # Wait for a page event before returning (e.g. "load", "networkidle")
  #   # wait_for_timeout_ms: 2000        # Extra wait after page load in milliseconds
  #   # wait_for_selector: "article"     # CSS selector to wait for before returning

  # Web fetch tool (uses Crawl4AI - self-hosted headless Chromium, no API key)
  # Crawl4AI returns server-cleaned "fit" markdown directly (no readability step needed),
  # ideal for JavaScript-heavy sites. Self-host (JWT auth is off by default, no key):
  #   docker run -d -p 11235:11235 --shm-size=1g unclecode/crawl4ai:0.8.6
  # For Docker deployments, use the Docker service name instead of localhost.
  # NOTE: Only one web_fetch provider can be active at a time.
  # Comment out the Jina AI web_fetch entry below before enabling this one.
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.crawl4ai.tools:web_fetch_tool
  #   base_url: http://localhost:11235   # Crawl4AI server URL (Docker: http://crawl4ai:11235)
  #   timeout: 30                         # Request timeout in seconds
  #   # allow_private_addresses: false     # SSRF guard: keep false in production. Set true ONLY for intentional internal targets.
  #   # filter: fit                       # Markdown filter: fit (default) | raw | bm25 | llm
  #   # token: $CRAWL4AI_TOKEN            # Bearer token (only if the server has JWT auth enabled)

  # Web capture tool (uses Browserless /screenshot to render a page as an artifact)
  # Browserless captures JavaScript-heavy pages with a real headless Chrome and
  # writes the screenshot into the current thread's outputs. It can run against
  # a self-hosted Browserless instance without a token, or Browserless Cloud with
  # BROWSERLESS_TOKEN. For Docker deployments, use the Docker service name instead
  # of localhost.
  # - name: web_capture
  #   group: web
  #   use: deerflow.community.browserless.tools:web_capture_tool
  #   base_url: http://localhost:3032       # Browserless instance URL (Docker: http://browserless:3000)
  #   # token: $BROWSERLESS_TOKEN           # Required for Browserless Cloud; optional for self-hosted
  #   timeout_s: 30                         # Request timeout in seconds
  #   output_format: png                    # png, jpeg, or webp
  #   full_page: true                       # Capture entire page instead of viewport only
  #   viewport_width: 1280
  #   viewport_height: 720
  #   # wait_for_selector: "main"           # CSS selector to wait for before capturing
  #   # wait_for_selector_timeout_ms: 5000
  #   # wait_for_timeout_ms: 1000           # Extra wait after navigation in milliseconds
  #   # best_attempt: true                  # Continue with current page state if waits time out
  #   # allow_private_addresses: false      # SSRF guard: keep false in production. Set true ONLY to
  #   #                                     # capture internal/private targets (loopback, RFC1918, etc.)

  # Web fetch tool (uses Exa)
  # NOTE: Only one web_fetch provider can be active at a time.
  # Comment out the Jina AI web_fetch entry below before enabling this one.
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.exa.tools:web_fetch_tool
  #   # api_key: $EXA_API_KEY

  # Web fetch tool (uses Jina AI reader)
  - name: web_fetch
    group: web
    use: deerflow.community.jina_ai.tools:web_fetch_tool
    timeout: 10
    # Optional proxy for restricted networks / Docker / WSL.
    # Use host.docker.internal instead of 127.0.0.1 when the proxy runs on the host.
    # proxy: $HTTPS_PROXY
    # trust_env: true

  # Web fetch tool (uses InfoQuest)
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.infoquest.tools:web_fetch_tool
  #   # Overall timeout for the entire crawling process (in seconds). Set to positive value to enable, -1 to disable
  #   timeout: 10
  #   # Waiting time after page loading (in seconds). Set to positive value to enable, -1 to disable
  #   fetch_time: 10
  #   # Timeout for navigating to the page (in seconds). Set to positive value to enable, -1 to disable
  #   navigation_timeout: 30

  # Web fetch tool (uses Firecrawl, requires FIRECRAWL_API_KEY)
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.firecrawl.tools:web_fetch_tool
  #   # api_key: $FIRECRAWL_API_KEY

  # Web fetch tool (uses GroundRoute, requires GROUNDROUTE_API_KEY)
  # Fetches a page's extracted text via GroundRoute mode=page.
  # NOTE: Only one web_fetch provider can be active at a time.
  # Comment out the Jina AI web_fetch entry above before enabling this one.
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.groundroute.tools:web_fetch_tool
  #   # api_key: $GROUNDROUTE_API_KEY
  
  # Web fetch tool (uses fastCRW - Firecrawl-compatible web scraper, single binary,
  # self-host or cloud. Cloud requires CRW_API_KEY; self-host may need no key.)
  # NOTE: Only one web_fetch provider can be active at a time.
  # Comment out the Jina AI web_fetch entry above before enabling this one.
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.fastcrw.tools:web_fetch_tool
  #   # api_key: $CRW_API_KEY
  #   # base_url: https://fastcrw.com/api   # default cloud; set to e.g. http://localhost:3000 for self-host
  #   # allow_private_addresses: false      # SSRF guard: keep false in production. Set true ONLY for intentional internal targets.

  # Image search tool (uses DuckDuckGo)
  # Use this to find reference images before image generation
  - name: image_search
    group: web
    use: deerflow.community.image_search.tools:image_search_tool
    max_results: 5

  # Image search tool (uses InfoQuest)
  # - name: image_search
  #   group: web
  #   use: deerflow.community.infoquest.tools:image_search_tool
  #   # Used to limit the scope of image search results, only returns content within the specified time range. Set to -1 to disable time filtering
  #   image_search_time_range: 10
  #   # Image size filter. Options: "l" (large), "m" (medium), "i" (icon).
  #   image_size: "i"

  # Image search tool (uses Serper - Google Images API, requires SERPER_API_KEY)
  # Serper provides real-time Google Images results. Sign up at https://serper.dev
  # Note: set SERPER_API_KEY in your environment before starting the app.
  # Avoid putting literal API keys in config.yaml; use the $VAR form instead.
  # - name: image_search
  #   group: web
  #   use: deerflow.community.serper.tools:image_search_tool
  #   max_results: 5   # capped at 10 by the Serper provider
  #   # api_key: $SERPER_API_KEY  # Optional explicit env-var reference

  # Image search tool (uses Brave Image Search API, requires BRAVE_SEARCH_API_KEY)
  # Brave provides independent image results and works alongside the Brave web search tool.
  # Note: set BRAVE_SEARCH_API_KEY in your environment before starting the app.
  # Avoid putting literal API keys in config.yaml; use the $VAR form instead.
  # - name: image_search
  #   group: web
  #   use: deerflow.community.brave.tools:image_search_tool
  #   max_results: 5   # capped at 200 by Brave Image Search
  #   # country: US
  #   # search_lang: en
  #   # safesearch: strict
  #   # spellcheck: true
  #   # api_key: $BRAVE_SEARCH_API_KEY  # Optional explicit env-var reference

  # File operations tools
  - name: ls
    group: file:read
    use: deerflow.sandbox.tools:ls_tool

  - name: read_file
    group: file:read
    use: deerflow.sandbox.tools:read_file_tool

  - name: glob
    group: file:read
    use: deerflow.sandbox.tools:glob_tool
    max_results: 200

  - name: grep
    group: file:read
    use: deerflow.sandbox.tools:grep_tool
    max_results: 100

  - name: write_file
    group: file:write
    use: deerflow.sandbox.tools:write_file_tool

  - name: str_replace
    group: file:write
    use: deerflow.sandbox.tools:str_replace_tool

  # Bash execution tool
  # Active only when using an isolated shell sandbox or when
  # sandbox.allow_host_bash: true explicitly opts into host bash.
  - name: bash
    group: bash
    use: deerflow.sandbox.tools:bash_tool

# ============================================================================
# Tool Search Configuration (Deferred Tool Loading)
# ============================================================================
# When enabled, MCP tools are not loaded into the agent's context directly.
# Instead, they are listed by name in the system prompt and discoverable
# via the tool_search tool at runtime.
# This reduces context usage and improves tool selection accuracy when
# multiple MCP servers expose a large number of tools.

tool_search:
  enabled: false
  # When tool_search is enabled, PR1 MCP routing metadata can auto-promote
  # matching deferred MCP tool schemas before a model call. This is the maximum
  # number of matched schemas promoted per model call. Valid range: 1..5.
  auto_promote_top_k: 3

# ============================================================================
# Tool Output Budget Protection
# ============================================================================
# Prevents oversized tool results from blowing the model context window.
# Outputs exceeding `externalize_min_chars` are persisted to disk and replaced
# with a compact typed synopsis + file reference. The model can read the full output
# via read_file. When disk persistence is unavailable, outputs exceeding
# `fallback_max_chars` are head+tail truncated instead.
#
# `exempt_tools` prevents persist→read→persist infinite loops for read tools.
# `tool_overrides` allows per-tool threshold customization.

tool_output:
  enabled: true
  externalize_min_chars: 12000
  # Sampling budget for the inline raw head/tail sample appended to every
  # typed synopsis; ignored for binary-like output, which carries its own sample.
  preview_head_chars: 2000
  preview_tail_chars: 1000
  fallback_max_chars: 30000
  fallback_head_chars: 8000
  fallback_tail_chars: 3000
  storage_subdir: ".tool-results"
  exempt_tools:
    - read_file
    - read_file_tool
  # tool_overrides:
  #   web_search: 8000
  #   bash: 20000

# ============================================================================
# Suggestions Configuration
# ============================================================================
# Configure whether the agent automatically generates follow-up question
# suggestions at the end of each response.

suggestions:
  enabled: true


# ============================================================================
# Input Polish Configuration
# ============================================================================
# Configure whether the composer can rewrite draft input before sending.

input_polish:
  enabled: true
  # Maximum draft length accepted by /api/input-polish.
  max_chars: 4000
  # Optional fast model for draft polishing. Leave null to use the default chat model.
  # For best UX, set this to your lowest-latency inexpensive model.
  model_name: null


# ============================================================================
# Loop Detection Configuration
# ============================================================================
# Detect and interrupt repeated identical tool-call loops.
# Frequency thresholds are safety limits for repeated use of the same tool type.

loop_detection:
  enabled: true
  warn_threshold: 3
  hard_limit: 5
  window_size: 20
  max_tracked_threads: 100
  tool_freq_warn: 30
  tool_freq_hard_limit: 50
  # Per-tool overrides for tool_freq_warn / tool_freq_hard_limit. Values can be
  # higher or lower than the global defaults. Commonly used to raise thresholds
  # for high-frequency tools like bash in batch workflows (e.g. RNA-seq pipelines)
  # without weakening protection on every other tool.
  # tool_freq_overrides:
  #   bash:
  #     warn: 150
  #     hard_limit: 300

# ============================================================================
# Tool Progress State Machine Configuration (RFC #3177)
# ============================================================================
# Detects tool stagnation and repetition at the (thread, tool) level.
# Tracks consecutive "no-new-info" calls (error, partial_success, near-duplicate success).
# Three transition paths (determined by deerflow_tool_meta.recoverable_by_model):
#   recoverable=true  (no_results, not_found, permission): ACTIVE → WARNED (terminal; hint re-injected each call)
#   recoverable=false (rate_limited, transient):           ACTIVE → WARNED → BLOCKED after warn_escalation_count more
#   recoverable=false + action=stop (auth, config):        ACTIVE → BLOCKED immediately
# Requires ToolErrorHandlingMiddleware to be active (always on).

# tool_progress:
#   enabled: false
#   stagnation_threshold: 3        # Consecutive problems before WARNED
#   warn_escalation_count: 2       # More problems after WARNED before BLOCKED
#   inject_assessment: true
#   jaccard_similarity_threshold: 0.8  # Word-set similarity threshold for near-duplicate detection
#   min_word_count_for_similarity: 10  # Min unique words to apply Jaccard check
#   max_tracked_threads: 100
#   exempt_tools:
#     - ask_clarification
#     - write_todos
#     - present_files
#     - task

# ============================================================================
# Read-Before-Write File Gate (issue #3857)
# ============================================================================
# Blocks write_file (append / overwrite of an existing file) and str_replace
# unless the agent has read the file's current version first; any write
# invalidates earlier reads, forcing a re-read between consecutive edits.
# Deterministic guardrail against blind duplicate appends in long tasks.

read_before_write:
  enabled: true

# ============================================================================
# Provider Safety Termination Configuration
# ============================================================================
# Intercept AIMessages where the provider stopped generation for safety reasons
# (e.g. OpenAI finish_reason='content_filter', Anthropic stop_reason='refusal',
# Gemini finish_reason='SAFETY') while still returning tool_calls. The
# tool_calls in such responses are typically truncated/unreliable and must
# not be executed. See issue #3028 for the full failure mode.
#
# Detectors are loaded by class path via reflection (same pattern as
# guardrails / models / tools). The built-in set covers OpenAI-compatible
# content_filter, Anthropic refusal, and Gemini SAFETY/BLOCKLIST/
# PROHIBITED_CONTENT/SPII/RECITATION.

safety_finish_reason:
  enabled: true
  # Leave `detectors` unset to use the built-in detector set. Set to a
  # non-empty list to fully override (use `enabled: false` to disable instead
  # of providing an empty list).
  #
  # Example — extend the OpenAI-compatible detector for a Chinese provider
  # whose gateway uses a non-standard finish_reason token:
  # detectors:
  #   - use: deerflow.agents.middlewares.safety_termination_detectors:OpenAICompatibleContentFilterDetector
  #     config:
  #       finish_reasons: ["content_filter", "sensitive", "risk_control"]
  #   - use: deerflow.agents.middlewares.safety_termination_detectors:AnthropicRefusalDetector
  #   - use: deerflow.agents.middlewares.safety_termination_detectors:GeminiSafetyDetector
  #
  # Example — add a custom detector for an in-house provider:
  # detectors:
  #   - use: my_company.deerflow_ext:WenxinSafetyDetector
  #     config:
  #       error_codes: [336003, 17, 18]

# ============================================================================
# Sandbox Configuration
# ============================================================================
# Choose between local sandbox (direct execution) or Docker-based AIO sandbox

# Option 1: Local Sandbox (Default)
# Executes commands directly on the host machine
uploads:
  # Application-level upload limits enforced by the gateway and exposed to the
  # frontend before file selection.
  max_files: 10
  max_file_size: 52428800    # 50 MiB
  max_total_size: 104857600  # 100 MiB
  # Automatic Office/PDF conversion runs on the backend host before sandbox
  # isolation applies. Keep this disabled unless uploads come from a fully
  # trusted source and you intentionally accept host-side parser risk.
  auto_convert_documents: false
  # Controls which PDF-to-Markdown converter is used whenever PDF conversion
  # runs. Automatic upload conversion is gated separately by
  # auto_convert_documents.
  # auto        — prefer pymupdf4llm when installed; fall back to MarkItDown for
  #               image-based or encrypted PDFs (recommended default).
  # pymupdf4llm — always use pymupdf4llm (must be installed: uv add pymupdf4llm).
  #               Better heading/table extraction; faster on most files.
  # markitdown  — always use MarkItDown (original behaviour, no extra dependency).
  pdf_converter: auto

sandbox:
  use: deerflow.sandbox.local:LocalSandboxProvider
  # Host bash execution is disabled by default because LocalSandboxProvider is
  # not a secure isolation boundary for shell access. Enable only for fully
  # trusted, single-user local workflows.
  allow_host_bash: false
  # Optional: Mount additional host directories into the sandbox.
  # Each mount maps a host path to a virtual container path accessible by the agent.
  # Note: with LocalSandboxProvider under `make up` (docker-compose), host_path is
  # checked from inside the deer-flow-gateway container — you must also bind-mount
  # the same directory into services.gateway.volumes in docker/docker-compose.yaml
  # for this mount to take effect (see issue #3244).
  # mounts:
  #   - host_path: /home/user/my-project    # Absolute path; see note above for Docker mode
  #     container_path: /mnt/my-project     # Virtual path inside the sandbox
  #     read_only: true                     # Whether the mount is read-only (default: false)

  # Tool output truncation limits (characters).
  # bash uses middle-truncation (head + tail) since errors can appear anywhere in the output.
  # read_file and ls use head-truncation since their content is front-loaded.
  # Set to 0 to disable truncation.
  bash_output_max_chars: 20000
  read_file_output_max_chars: 50000
  ls_output_max_chars: 20000

  # Maximum wall-clock seconds a single host bash command may run before it is
  # terminated (process group and all). A blocking foreground command — e.g. a
  # server started without backgrounding — is killed after this long so the
  # agent's turn cannot hang. Start long-lived processes in the background with
  # output redirected (e.g. `your-command > /tmp/server.log 2>&1 &`) when you
  # need logs; unredirected background output is drained with bounded capture
  # and excess output is discarded.
  bash_command_timeout: 600

# Option 2: Container-based AIO Sandbox
# Executes commands in isolated containers (Docker or Apple Container)
# On macOS: Automatically prefers Apple Container if available, falls back to Docker
# On other platforms: Uses Docker
# Uncomment to use:
# sandbox:
#   use: deerflow.community.aio_sandbox:AioSandboxProvider
#
#   # Optional: Container image to use (works with both Docker and Apple Container)
#   # Default: enterprise-public-cn-beijing.cr.volces.com/vefaas-public/all-in-one-sandbox:latest
#   # The mirror's `:latest` tag is frozen on an old pre-1.9.3 digest that lacks
#   # the /v1/bash/* routes required-secrets skills need (see #3921/#3922), so
#   # pin an explicit version instead — recommended: 1.11.0 (multi-arch, works
#   # on both x86_64 and arm64). Custom images should extend the default image
#   # or implement the same AIO sandbox HTTP API used by agent-sandbox. See
#   # backend/docs/CONFIGURATION.md.
#   # image: enterprise-public-cn-beijing.cr.volces.com/vefaas-public/all-in-one-sandbox:1.11.0
#
#   # Optional: Base port for sandbox containers (default: 8080)
#   # port: 8080

#   # Optional: Maximum number of concurrent sandbox containers (default: 3)
#   # When the limit is reached the least-recently-used sandbox is evicted to
#   # make room for new ones. Use a positive integer here; omit this field to use the default.
#   # replicas: 3
#
#   # Optional: Prefix for container names (default: deer-flow-sandbox)
#   # container_prefix: deer-flow-sandbox
#
#   # Optional: Additional mount directories from host to container
#   # NOTE: Skills directory is automatically mounted from skills.path to skills.container_path
#   # mounts:
#   #   # Other custom mounts
#   #   - host_path: /path/on/host
#   #     container_path: /home/user/shared
#   #     read_only: false
#   #
#   #   # DeerFlow will surface configured container_path values to the agent,
#   #   # so it can directly read/write mounted directories such as /home/user/shared
#
#   # Optional: Environment variables to inject into the sandbox container
#   # Values starting with $ will be resolved from host environment variables
#   # environment:
#   #   NODE_ENV: production
#   #   DEBUG: "false"
#   #   API_KEY: $MY_API_KEY        # Reads from host's MY_API_KEY env var
#   #   DATABASE_URL: $DATABASE_URL  # Reads from host's DATABASE_URL env var

# Option 3: BoxLite micro-VM Sandbox
# Runs each sandbox as a BoxLite micro-VM. Released boxes stay in an in-process
# warm pool and can be reclaimed by the same user/thread without a cold start.
# Requires the boxlite runtime and host virtualization support (KVM on Linux,
# Hypervisor.framework on macOS).
# sandbox:
#   use: deerflow.community.boxlite:BoxliteProvider
#   image: python:3.12-slim
#
#   # Optional: Per-box memory and CPU limits.
#   # memory_mib: 1024
#   # cpus: 2
#
#   # Optional: Maximum active + warm BoxLite VMs per gateway process (default: 3).
#   # Active boxes are never evicted; only warm-pool boxes are stopped to make room.
#   # replicas: 3
#
#   # Optional: Seconds before an idle warm-pool VM is stopped (default: 600).
#   # Set to 0 to keep warm VMs until shutdown or replica eviction.
#   # idle_timeout: 600
#
#   # Optional: Skip the reclaim health check for very recently released VMs.
#   # Default 0.0 keeps reliability-first validation before warm reuse.
#   # health_check_skip_seconds: 0.0
#
#   # Optional: Environment variables to inject into every command.
#   # environment:
#   #   PYTHONUNBUFFERED: "1"
#
# Option 4: Provisioner-managed AIO Sandbox (docker-compose-dev)
# Each sandbox_id gets a dedicated Pod in k3s, managed by the provisioner.
# Recommended for production or advanced users who want better isolation and scalability.:
# sandbox:
#   use: deerflow.community.aio_sandbox:AioSandboxProvider
#   provisioner_url: http://provisioner:8002
#   # API key for provisioner authentication. Must match PROVISIONER_API_KEY
#   # set on the provisioner container. Both sides must have the same value set;
#   # the provisioner rejects all /api/* requests when PROVISIONER_API_KEY is unset.
#   # Generate a strong key: openssl rand -hex 32
#   # provisioner_api_key: $PROVISIONER_API_KEY
#   # Note: provisioner-created Pods use the provisioner's SANDBOX_IMAGE
#   # environment variable, not sandbox.image from this config file.

# ============================================================================
# Subagents Configuration
# ============================================================================
# Configure timeouts for subagent execution
# Subagents are background workers delegated tasks by the lead agent

# subagents:
#   # Default timeout (seconds) for built-in subagents (default: 1800 = 30 min).
#   # Custom agents use their own timeout_seconds (default 900) unless overridden.
#   timeout_seconds: 1800
#   # Optional global max-turn override for all subagents.
#   # Built-in defaults: general-purpose=150, bash=60. Leave unset to keep them.
#   # max_turns: 120
#
#   # Total number of subagent delegations allowed in one lead-agent run.
#   # This is a deterministic backstop against repeated planning checkpoints
#   # launching legal-sized batches forever. The default 6 allows two full
#   # batches at the default concurrency of 3. Valid config range: 1-50.
#   # Per-request runtime context can temporarily override this with
#   # `max_total_subagents`, clamped to the same 1-50 range.
#   max_total_per_run: 6
#
#   # Per-run token ceiling for subagents (#3875 Phase 2). A backstop against a
#   # subagent that burns tokens on trivial work. At the hard-stop threshold the
#   # in-flight turn is capped (tool calls stripped, finish_reason forced to
#   # "stop") so the run completes naturally with a final answer; the result is
#   # stamped `completed` + `subagent_stop_reason=token_capped` so the lead and
#   # UI can tell a budget-capped completion from a clean one. The 2,000,000
#   # default is a generous ceiling — lower it to tighten cost controls. A
#   # per-agent `token_budget` override (see `agents:` below) wins over this.
#   # token_budget:
#   #   enabled: true
#   #   max_tokens: 2000000
#   #   warn_threshold: 0.7     # log a warning once this fraction of the budget is spent
#
#   # Optional per-agent overrides (applies to both built-in and custom agents)
#   agents:
#     general-purpose:
#       timeout_seconds: 2700  # 45 minutes for very long deep-research tasks
#       max_turns: 250         # raise above the 150 default for very deep tasks
#       # token_budget:        # per-agent override of the global token_budget above
#       #   max_tokens: 3000000  # raise the ceiling for deep-research tasks
#       # model: qwen3:32b     # Use a specific model (default: inherit from lead agent)
#       # skills:               # Skill whitelist (default: inherit all enabled skills)
#       #   - web-search
#       #   - data-analysis
#     bash:
#       timeout_seconds: 300   # 5 minutes for quick command execution
#       max_turns: 80
#       # skills: []            # No skills for bash agent
#
#   # Custom subagent types: define specialized agents with their own prompts,
#   # tools, skills, and model configuration. Custom agents are available via
#   # the `task` tool alongside built-in types (general-purpose, bash).
#   # custom_agents:
#   #   analysis:
#   #     description: "Data analysis specialist for processing datasets and generating insights"
#   #     system_prompt: |
#   #       You are a data analysis subagent. Focus on:
#   #       - Processing and analyzing datasets
#   #       - Generating visualizations
#   #       - Providing statistical insights
#   #     tools:                  # Tool whitelist (null = inherit all)
#   #       - bash
#   #       - read_file
#   #       - write_file
#   #     skills:                 # Skill whitelist (null = inherit all, [] = none)
#   #       - data-analysis
#   #       - visualization
#   #     model: inherit          # 'inherit' uses parent's model
#   #     max_turns: 80
#   #     timeout_seconds: 600
#
#   # Model override: by default, subagents inherit the lead agent's model.
#   # Set `model` to use a different model (e.g., a local Ollama model for cost savings).
#   # The model name must match a name defined in the `models:` section above.

# ============================================================================
# ACP Agents Configuration
# ============================================================================
# Configure external ACP-compatible agents for the built-in `invoke_acp_agent` tool.

# acp_agents:
#   claude_code:
#     # DeerFlow expects an ACP adapter here. The standard `claude` CLI does not
#     # speak ACP directly. Install `claude-agent-acp` separately or use:
#     command: npx
#     args: ["-y", "@zed-industries/claude-agent-acp"]
#     description: Claude Code for implementation, refactoring, and debugging
#     model: null
#     # auto_approve_permissions: false  # Set to true to auto-approve ACP permission requests
#     # env:                             # Optional: inject environment variables into the agent subprocess
#     #   ANTHROPIC_API_KEY: $ANTHROPIC_API_KEY  # $VAR resolves from host environment
#
#   codex:
#     # DeerFlow expects an ACP adapter here. The standard `codex` CLI does not
#     # speak ACP directly. Install `codex-acp` separately or use:
#     command: npx
#     args: ["-y", "@zed-industries/codex-acp"]
#     description: Codex CLI for repository tasks and code generation
#     model: null
#     # auto_approve_permissions: false  # Set to true to auto-approve ACP permission requests
#     # env:                             # Optional: inject environment variables into the agent subprocess
#     #   OPENAI_API_KEY: $OPENAI_API_KEY  # $VAR resolves from host environment

# ============================================================================
# Skills Configuration
# ============================================================================
# Configure skills directory for specialized agent workflows

skills:
  # Path to skills directory on the host (relative to project root or absolute)
  # Default: skills under the project root
  # Override with DEER_FLOW_SKILLS_PATH when this field is omitted.
  # Uncomment to customize:
  # path: /absolute/path/to/custom/skills

  # Path where skills are mounted in the sandbox container
  # This is used by the agent to access skills in both local and Docker sandbox
  # Default: /mnt/skills
  container_path: /mnt/skills

  # Deferred skill discovery (default: false)
  # When enabled, only skill names appear in the system prompt (<skill_index>).
  # The LLM discovers skill details on demand via the describe_skill tool.
  # This keeps the system prompt compact and prefix-cache friendly when many
  # skills are installed.
  # deferred_discovery: true

# ============================================================================
# SkillScan Configuration
# ============================================================================
# Native deterministic skill safety scanning. This runs before the LLM skill
# scanner on skill install/update and agent-managed skill writes.
skill_scan:
  # Set false to disable the new deterministic analyzers. Safe archive
  # extraction and the LLM skill scanner still run.
  enabled: true

# Note: To restrict which skills are loaded for a specific custom agent,
# define a `skills` list in that agent's `config.yaml` (e.g. `agents/my-agent/config.yaml`):
# - Omitted or null: load all globally enabled skills (default)
# - []: disable all skills for this agent
# - ["skill-name"]: load only specific skills

# ============================================================================
# Title Generation Configuration
# ============================================================================
# Automatic conversation title generation settings

title:
  enabled: true
  max_words: 6
  max_chars: 60
  model_name: null # null = fast local fallback; set a model name to use LLM title generation

# ============================================================================
# Summarization Configuration
# ============================================================================
# Automatically summarize conversation history when token limits are approached
# This helps maintain context in long conversations without exceeding model limits

summarization:
  enabled: true

  # Model to use for summarization (null = use default model)
  # Recommended: Use a lightweight, cost-effective model like "gpt-4o-mini" or similar
  model_name: null

  # Trigger conditions - at least one required
  # Summarization runs when ANY threshold is met (OR logic)
  # You can specify a single trigger or a list of triggers
  trigger:
    # Trigger when token count reaches 32000
    - type: tokens
      value: 32000
    # Uncomment to also trigger when message count reaches 50
    # - type: messages
    #   value: 50
    # Uncomment to trigger when 80% of model's max input tokens is reached
    # - type: fraction
    #   value: 0.8

  # Context retention policy after summarization
  # Specifies how much recent history to preserve
  keep:
    # Keep the most recent 10 messages (recommended)
    type: messages
    value: 10
    # Alternative: Keep specific token count
    # type: tokens
    # value: 3000
    # Alternative: Keep percentage of model's max input tokens
    # type: fraction
    # value: 0.3

  # Maximum tokens to keep when preparing messages for summarization
  # Set to null to skip trimming (not recommended for very long conversations)
  trim_tokens_to_summarize: 15564

  # Custom summary prompt template (null = use default LangChain prompt)
  # The prompt should guide the model to extract important context
  summary_prompt: null

  # Loaded SKILL.md references (read_file calls under skills.container_path) are
  # captured into the durable skill_context channel and re-injected after
  # compaction as name/path/description reminders. The full skill body is not
  # persisted; the agent should re-read the file before applying instructions.
  # Tool names counted as skill reads:
  # Legacy preserve_recent_skill_* summarization settings are no longer used;
  # skill retention is handled by this durable reference channel instead. Set
  # this list to [] to disable durable skill-reference capture.
  skill_file_read_tool_names:
    - read_file
    - read
    - view
    - cat

# ============================================================================
# Memory Configuration
# ============================================================================
# Global memory mechanism (pluggable + self-contained).
#
# Shared fields (host level, backend-agnostic):
#   enabled             - Master switch for the memory mechanism (call-site gate)
#   injection_enabled   - Whether to inject memory into the system prompt (call-site gate)
#   shutdown_flush_timeout_seconds - Hard budget (s) to drain pending updates on Gateway graceful shutdown (default: 30)
#   manager_class       - Backend selector: registered name (deermem/noop) or dotted path
#   backend_config      - Backend-private config dict (passthrough; each backend self-interprets)
#
# DeerMem-private fields live under ``backend_config`` (NOT at the memory: top level):
#   storage_path        - Data root. Empty = deer-flow base_dir (factory injects absolute
#                         runtime_home). Absolute path = that root. Relative = CWD-relative.
#   model               - LLM config for memory extraction: {provider, model, api_key, base_url,
#                         temperature}. Omit all fields = no extraction (non-LLM ops still work).
#   debounce_seconds    - Debounce wait before processing queued updates (default: 30)
#   max_facts           - Maximum facts to store (default: 100)
#   fact_confidence_threshold - Minimum confidence for storing facts (default: 0.7)
#   max_injection_tokens     - Token budget for memory injection (default: 2000)
#   token_counting          - tiktoken (accurate, network on first use) or char (network-free)
#   guaranteed_categories   - Fact categories always injected (default: ["correction"])
#   guaranteed_token_budget - Token ceiling for guaranteed categories (default: 500)
#   staleness_review_enabled       - Enable staleness pruning (default: true)
#   staleness_age_days             - Age threshold for staleness candidates (default: 90)
#   staleness_min_candidates       - Minimum stale facts to trigger review (default: 3)
#   staleness_max_removals_per_cycle - Safety cap on removals per cycle (default: 10)
#   staleness_protected_categories - Categories exempt from staleness review (default: ["correction"])
#   staleness_max_lifetime_multiplier - Creation-time cap multiplier for a fact LLM-assigned
#                                       expected_valid_days; new facts clamped to
#                                       staleness_age_days x multiplier. Default 20.0
#                                       (90 x 20 = 1800d ~ 5 years) supports the very-stable
#                                       prompt tier. (default: 20.0, range: 1.0-100.0)
#   staleness_max_extension_days   - Absolute ceiling (in days) on expected_valid_days after a
#                                    lifetime extension (staleFactsToExtend). Prevents timedelta
#                                    overflow and LLM misfire from permanently deferring a fact.
#                                    (default: 3650, range: 90-36500)
memory:
  enabled: true
  injection_enabled: true # Whether to inject memory into system prompt
  # Hard budget (seconds) to drain pending memory updates on Gateway graceful
  # shutdown. Each pending item does one LLM call, so large IM batches may need
  # more. Must fit inside the pod's K8s terminationGracePeriodSeconds (channel
  # stop + this drain + buffer) or K8s SIGKILLs the drain -- set that on the
  # gateway Helm deployment (see deploy/helm/deer-flow). Default 30s.
  shutdown_flush_timeout_seconds: 30.0
  # Memory backend selector. Either a registered backend name (matching a
  # backends/<name>/ folder that exposes MANAGER_CLASS, e.g. deermem / noop)
  # or a dotted import path to a MemoryManager subclass.
  manager_class: deermem
  # Memory operation mode:
  #   middleware (default) - passive background extraction after each turn.
  #   tool - experimental opt-in; the model calls memory_search/memory_add/
  #     memory_update/memory_delete directly. This gives the model agency over
  #     memory writes, but effectiveness depends on model tool-use behavior.
  # Only one mode runs at a time. (tool mode calls the MemoryManager ABC --
  # memory_search/add/update/delete go through the active backend; backends
  # without fact-CRUD return a JSON error instead of crashing.)
  mode: middleware
  # Backend-private config (a dict), passed verbatim to the backend __init__.
  # Each backend self-interprets it (DeerMem parses it into DeerMemConfig).
  backend_config:
    storage_path: ""            # empty = deer-flow base_dir (factory injects absolute runtime_home); a non-empty path is the root DIRECTORY (per-user memory under {storage_path}/users/{uid}/memory.json)
    debounce_seconds: 30 # Wait time before processing queued updates
    model:                      # LLM for memory extraction; omit all fields = no extraction (non-LLM ops still work; an update raises)
      # provider: openai
      # model: gpt-4o-mini
      # api_key: $OPENAI_API_KEY
      # base_url:               # optional, for OpenAI-compatible gateways (e.g. DeepSeek)
      # temperature:            # optional
    max_facts: 100 # Maximum number of facts to store
    fact_confidence_threshold: 0.7 # Minimum confidence for storing facts
    max_injection_tokens: 2000 # Maximum tokens for memory injection
    # Token counting strategy for memory-injection budgeting:
    #   tiktoken (default) - accurate, but the encoding BPE data may be
    #     downloaded from a public network endpoint on first use. In
    #     network-restricted environments this download can block for a long
    #     time (see issues #3402 / #3429). Pre-cache the encoding or set this
    #     to "char" to avoid it.
    #   char - network-free CJK-aware character-based estimate; never touches
    #     tiktoken. Slightly less precise budgeting, zero network I/O.
    token_counting: tiktoken
    # Guaranteed injection: fact categories that bypass the regular token budget
    # and draw from a reserved allowance, so high-signal corrections (e.g.
    # "don use pip, use uv") survive even when the budget is tight.
    guaranteed_categories:
      - correction
    guaranteed_token_budget: 500
    # Staleness review: periodically prune aged facts that may no longer reflect
    # the user current situation. When triggered, the LLM reviews facts older
    # than their individual review window (expected_valid_days, or
    # staleness_age_days as fallback) during the normal memory-update call (same
    # LLM invocation - no extra API call) and decides KEEP, REMOVE, or EXTEND
    # for each. The LLM assigns expected_valid_days when creating a fact; EXTEND
    # (staleFactsToExtend) recalibrates that window at review time.
    staleness_review_enabled: true
    staleness_age_days: 90
    staleness_min_candidates: 3
    staleness_max_removals_per_cycle: 10
    staleness_protected_categories:
      - correction
    staleness_max_lifetime_multiplier: 20.0   # creation cap = staleness_age_days x multiplier (90 x 20 = 1800d)
    staleness_max_extension_days: 3650        # absolute ceiling on extended expected_valid_days (~10 years)
    # Memory consolidation (opt-in, lossy: source facts are replaced by a
    # synthesized one; only consolidatedFrom IDs are kept). Runs in the same
    # memory-update LLM call as extraction/staleness - no extra API cost.
    # consolidation_enabled defaults to false because consolidation is lossy.
    consolidation_enabled: false       # set true to opt into fact consolidation
    consolidation_min_facts: 8         # min facts in one category to trigger review (3-30)
    consolidation_max_groups_per_cycle: 3  # max groups merged per update cycle (1-10)
    consolidation_max_sources: 8       # max source facts per consolidation group (2-20)

# ============================================================================
# Custom Agent Management API
# ============================================================================
# Controls whether the HTTP gateway exposes custom-agent SOUL/USER.md management.
# Keep this disabled unless the gateway is behind a trusted authenticated admin boundary.
agents_api:
  enabled: false

# ============================================================================
# Skill Self-Evolution Configuration
# ============================================================================
# Allow the agent to autonomously create and improve skills in skills/custom/.
skill_evolution:
  enabled: false # Set to true to allow agent-managed writes under skills/custom
  moderation_model_name: null # Model for LLM-based security scanning (null = use default model)

# ============================================================================
# Checkpointer Configuration (DEPRECATED — use `database` instead)
# ============================================================================
# Legacy standalone checkpointer config. Kept for backward compatibility.
# Prefer the unified `database` section below, which drives the LangGraph
# checkpointer, LangGraph Store, and DeerFlow application data (runs,
# feedback, events) from a single backend setting.
#
# If both `checkpointer` and `database` are present, `checkpointer`
# takes precedence for the LangGraph checkpointer and Store only.
#
# checkpointer:
#   type: sqlite
#   connection_string: checkpoints.db
#
# checkpointer:
#   type: postgres
#   connection_string: postgresql://user:password@localhost:5432/deerflow

# ============================================================================
# Database
# ============================================================================
# Unified storage backend for the LangGraph checkpointer, LangGraph Store,
# and DeerFlow application data (runs, threads metadata, feedback, etc.).
#
# backend: memory  -- No persistence, data lost on restart
# backend: sqlite  -- Single-node deployment, files in sqlite_dir
# backend: postgres -- Production multi-node deployment
#
# If this section is omitted or empty in config.yaml, DeerFlow uses:
#   backend: sqlite
#   sqlite_dir: .deer-flow/data
#
# SQLite mode uses a single deerflow.db file with WAL journal mode
# for the checkpointer, Store, and application data.
#
# Postgres mode: put your connection URL in .env as DATABASE_URL,
# then reference it here with $DATABASE_URL.
#
# Install the driver — Issue #2754 fix lands `UV_EXTRAS` in every code path:
#   Local      `make dev` auto-detects from `database.backend: postgres` below
#              and passes `--extra postgres` to `uv sync` on every restart, so
#              the extra is no longer wiped. To opt in explicitly (or layer
#              extras like `postgres,ollama`), set in project-root .env:
#                  UV_EXTRAS=postgres
#   Docker dev `make docker-start` reads `UV_EXTRAS` from project-root .env via
#              `env_file`. Set:
#                  UV_EXTRAS=postgres
#              Multiple extras (`postgres,ollama`) supported here too — see
#              docker/dev-entrypoint.sh.
#   Docker img build-arg `UV_EXTRAS=postgres,discord docker compose build`
#              supports comma- or whitespace-separated extras at build time
#              (backend/Dockerfile expands them to repeated `--extra` flags).
#
# First-time bootstrap (before `make dev`):
#   cd backend && uv sync --all-packages --extra postgres
#   (--all-packages propagates the extra into workspace members — see PR #2584)
#
# NOTE: When both `checkpointer` and `database` are configured,
# `checkpointer` takes precedence for the LangGraph checkpointer and Store;
# `database` still controls DeerFlow application data.
# If you use `database`, you can remove the `checkpointer` section.
# database:
#   backend: sqlite
#   sqlite_dir: .deer-flow/data
#
# database:
#   backend: postgres
#   postgres_url: $DATABASE_URL
database:
  backend: sqlite
  sqlite_dir: .deer-flow/data

# ============================================================================
# Run Events Configuration
# ============================================================================
# Storage backend for run events (messages + execution traces).
#
# backend: memory  -- No persistence, data lost on restart (default)
# backend: db      -- SQL database via ORM, full query capability (production)
# backend: jsonl   -- Append-only JSONL files (lightweight single-node persistence)
#
# run_events:
#   backend: memory
#   max_trace_content: 10240    # Truncation threshold for trace content (db backend, bytes)
#   track_token_usage: true     # Accumulate token counts to RunRow
run_events:
  backend: memory
  max_trace_content: 10240
  track_token_usage: true

# ============================================================================
# Scheduled Tasks Configuration
# ============================================================================
# Background scheduler for one-time and recurring (cron) agent runs.
# All fields are restart-required (captured at Gateway lifespan startup).
#
# Multi-worker note: the scheduler runs once per uvicorn worker. SQLite silently
# ignores row-level locks, so multiple workers can double-fire the same task.
# For multi-worker deployments (GATEWAY_WORKERS > 1), use the Postgres database
# backend, where FOR UPDATE SKIP LOCKED serializes claims correctly.
#
# scheduler:
#   enabled: false               # Master switch for the background poller
#   poll_interval_seconds: 5     # How often to scan for due tasks
#   lease_seconds: 120           # Claim lease; a crashed process's task becomes reclaimable after this
#   max_concurrent_runs: 3       # Global cap on active scheduled runs; each poll claims only into the remaining budget
#   min_once_delay_seconds: 60   # Minimum future offset for one-time tasks at creation time
scheduler:
  enabled: false
  poll_interval_seconds: 5
  lease_seconds: 120
  max_concurrent_runs: 3
  min_once_delay_seconds: 60

# ============================================================================
# Run Ownership Configuration
# ============================================================================
# Controls cross-process run ownership for multi-worker deployments.
# When GATEWAY_WORKERS > 1, each worker claims runs with a lease; the heartbeat
# renews leases, and reconciliation recovers orphaned runs from crashed workers.
#
# CLOCK-SYNC REQUIREMENT (multi-worker only): reconciliation compares another
# worker's UTC lease timestamp against this worker's datetime.now(UTC). Worker
# clocks MUST be synced (NTP / chrony / systemd-timesyncd — default on K8s and
# cloud VMs) within a few seconds. grace_seconds is the skew budget; worst case
# (owning worker's heartbeat just about to fire), a peer whose clock is more
# than grace_seconds ahead can mis-reclaim a still-live run as an orphan. Raise
# grace_seconds if your environment cannot keep clocks within a few seconds;
# the trade-off is longer recovery latency for genuinely dead workers
# (lease_seconds + grace_seconds from last heartbeat to reclaim).

run_ownership:
  lease_seconds: 30           # Seconds before a run lease expires if not renewed.
                              # Heartbeat renews every lease_seconds / 3.
  grace_seconds: 10           # Extra seconds past expiry before reclaiming an orphaned run.
                              # Also the cross-worker clock-skew budget — see note above.
  heartbeat_enabled: false    # Set to true for GATEWAY_WORKERS > 1

# ============================================================================
# Stream Bridge Configuration
# ============================================================================
# The stream bridge carries live agent events from gateway workers to SSE
# clients. Docker Compose sets DEER_FLOW_STREAM_BRIDGE_REDIS_URL automatically,
# so Docker deployments use Redis Streams even if this section is omitted.
#
# The redis bridge requires the optional `redis` extra. It is auto-detected from
# this section on `make dev`, and always installed in the Docker image. To install
# it manually: cd backend && uv sync --all-packages --extra redis
#
# stream_bridge:
#   type: memory                 # single-process only
#   queue_maxsize: 256
#
# stream_bridge:
#   type: redis                  # recommended for Docker / multi-worker gateway
#   redis_url: redis://redis:6379/0
#   queue_maxsize: 256           # events retained per run (redis stream MAXLEN)
#   stream_ttl_seconds: 86400    # rolling TTL for retained stream buffers.
#                                # Refreshed on each publish/publish_end; set 0
#                                # to disable. This is not a run timeout.
#   recovered_stream_cleanup_delay_seconds: 60  # seconds to wait after
#                                # publishing END for a recovered orphan run
#                                # before deleting the stream key.
#   max_connections: 100         # optional pool ceiling. Each live SSE client
#                                # holds one connection blocked in XREAD ... BLOCK
#                                # for up to heartbeat_interval (15s), so hundreds
#                                # of concurrent clients open hundreds of
#                                # connections. Unset = redis-py default (unbounded).
#
# NOTE: the redis bridge is fail-hard in v1. Redis.from_url is lazy, so a down
# Redis does not block gateway startup, but the first publish/xread raises. A
# mid-run Redis outage fails the active run — there is no automatic retry/backoff
# or fallback to the in-memory bridge. Run Redis with HA / a restart policy.

# ============================================================================
# User-Owned IM Channel Connections
# ============================================================================
# Lets logged-in users connect their own IM accounts from the DeerFlow frontend
# while reusing the existing `channels` runtime configuration below.
#
# Security notes:
# - No public IP, OAuth callback URL, or provider webhook is required.
# - Provider bot/app credentials stay under `channels.*`.
# - `channel_connections` stores per-user bindings and one-time connect codes.
# - Telegram uses a deep link when `bot_username` is configured.
# - Slack, Discord, Feishu, DingTalk, WeChat, and WeCom use `/connect <code>`
#   through the already-running bot/app.
#
# channel_connections:
#   enabled: false
#   # Security: keep this enabled unless you intentionally want legacy open-bot behavior.
#   # Disabling it lets unbound external IM users create DeerFlow threads/runs.
#   require_bound_identity: true
#
#   telegram:
#     enabled: false
#     bot_username: $TELEGRAM_BOT_USERNAME
#
#   slack:
#     enabled: false
#
#   discord:
#     enabled: false
#
#   feishu:
#     enabled: false
#
#   dingtalk:
#     enabled: false
#
#   wechat:
#     enabled: false
#
#   wecom:
#     enabled: false

# ============================================================================
# IM Channels Configuration
# ============================================================================
# Connect DeerFlow to external messaging platforms.
# All channels use outbound connections (WebSocket or polling) — no public IP required.

# channels:
#   # LangGraph-compatible Gateway API base URL for thread/message management (default: http://localhost:8001/api)
#   # For Docker deployments, use the Docker service name instead of localhost:
#   #   langgraph_url: http://gateway:8001/api
#   #   gateway_url: http://gateway:8001
#   langgraph_url: http://localhost:8001/api
#   # Gateway API URL for auxiliary queries like /models, /memory (default: http://localhost:8001)
#   gateway_url: http://localhost:8001
#   #
#   # Docker Compose note:
#   # If channels run inside the gateway container, use container DNS names instead
#   # of localhost, for example:
#   # langgraph_url: http://gateway:8001/api
#   # gateway_url: http://gateway:8001
#   # You can also set DEER_FLOW_CHANNELS_LANGGRAPH_URL / DEER_FLOW_CHANNELS_GATEWAY_URL.
#
#   # Optional: default mobile/session settings for all IM channels
#   session:
#     assistant_id: lead_agent  # or a custom agent name; custom agents route via lead_agent + agent_name
#     config:
#       recursion_limit: 100
#     context:
#       thinking_enabled: true
#       is_plan_mode: false
#       subagent_enabled: false
#
#   feishu:
#     enabled: false
#     app_id: $FEISHU_APP_ID
#     app_secret: $FEISHU_APP_SECRET
#     # domain: https://open.feishu.cn       # China (default)
#     # domain: https://open.larksuite.com   # International
#
#   slack:
#     enabled: false
#     bot_token: $SLACK_BOT_TOKEN     # xoxb-...
#     app_token: $SLACK_APP_TOKEN     # xapp-... (Socket Mode)
#     allowed_users: []               # empty = allow all; can also be a single Slack user ID string, e.g. U123456, but list form is recommended
#
#   telegram:
#     enabled: false
#     bot_token: $TELEGRAM_BOT_TOKEN
#     allowed_users: []               # empty = allow all
#
#   wechat:
#     enabled: false
#     bot_token: $WECHAT_BOT_TOKEN
#     ilink_bot_id: $WECHAT_ILINK_BOT_ID
#     # Optional: allow first-time QR bootstrap when bot_token is absent
#     qrcode_login_enabled: true
#     # Optional: sent as iLink-App-Id header when provided
#     ilink_app_id: ""
#     # Optional: sent as SKRouteTag header when provided
#     route_tag: ""
#     allowed_users: []               # empty = allow all
#     # Optional: long-polling timeout in seconds
#     polling_timeout: 35
#     # Optional: QR poll interval in seconds when qrcode_login_enabled is true
#     qrcode_poll_interval: 2
#     # Optional: QR bootstrap timeout in seconds
#     qrcode_poll_timeout: 180
#     # Optional: persist getupdates cursor under the gateway container volume
#     state_dir: ./.deer-flow/wechat/state
#     # Optional: max inbound image size in bytes before skipping download
#     max_inbound_image_bytes: 20971520
#     # Optional: max outbound image size in bytes before skipping upload
#     max_outbound_image_bytes: 20971520
#     # Optional: max inbound file size in bytes before skipping download
#     max_inbound_file_bytes: 52428800
#     # Optional: max outbound file size in bytes before skipping upload
#     max_outbound_file_bytes: 52428800
#     # Optional: allowed file extensions for regular file receive/send
#     allowed_file_extensions: [".txt", ".md", ".pdf", ".csv", ".json", ".yaml", ".yml", ".xml", ".html", ".log", ".zip", ".doc", ".docx", ".xls", ".xlsx", ".ppt", ".pptx", ".rtf"]
#
#     # Optional: channel-level session overrides
#     session:
#       assistant_id: mobile-agent  # custom agent names are supported here too
#       context:
#         thinking_enabled: false
#
#       # Optional: per-user overrides by user_id
#       users:
#         "123456789":
#           assistant_id: vip-agent
#           config:
#             recursion_limit: 150
#           context:
#             thinking_enabled: true
#             subagent_enabled: true
#   wecom:
#     enabled: false
#     bot_id: $WECOM_BOT_ID
#     bot_secret: $WECOM_BOT_SECRET
#
#   dingtalk:
#     enabled: false
#     client_id: $DINGTALK_CLIENT_ID
#     client_secret: $DINGTALK_CLIENT_SECRET
#     allowed_users: []               # empty = allow all
#     card_template_id: ""            # Optional: AI Card template ID for streaming updates
#
#   discord:
#     enabled: false
#     bot_token: $DISCORD_BOT_TOKEN
#     allowed_guilds: []              # empty = allow all guilds; can also be a single guild ID
#     mention_only: false             # If true, only respond when the bot is mentioned
#     allowed_channels: []              # Optional: channel IDs exempt from mention_only (bot responds without mention)
#     thread_mode: false               # If true, group a channel conversation into a thread

# ============================================================================
# Guardrails Configuration
# ============================================================================
# Optional pre-execution authorization for tool calls.
# When enabled, every tool call passes through the configured provider
# before execution. Three options: built-in allowlist, OAP policy provider,
# or custom provider. See backend/docs/GUARDRAILS.md for full documentation.
#
# Providers are loaded by class path via resolve_variable (same as models/tools).

# --- Option 1: Built-in AllowlistProvider (zero external deps) ---
# guardrails:
#   enabled: true
#   provider:
#     use: deerflow.guardrails.builtin:AllowlistProvider
#     config:
#       denied_tools: ["bash", "write_file"]

# --- Option 2: OAP passport provider (open standard, any implementation) ---
#   The Open Agent Passport (OAP) spec defines passport format and decision codes.
#   Any OAP-compliant provider works. Example using APort (reference implementation):
#     pip install aport-agent-guardrails && aport setup --framework deerflow
# guardrails:
#   enabled: true
#   provider:
#     use: aport_guardrails.providers.generic:OAPGuardrailProvider

# --- Option 3: Custom provider (any class with evaluate/aevaluate methods) ---
# guardrails:
#   enabled: true
#   provider:
#     use: my_package:MyGuardrailProvider
#     config:
#       key: value

# ============================================================================
# Authorization Configuration
# ============================================================================
# Fine-grained resource authorization (RBAC and beyond). Disabled by default;
# every authenticated user has access to all resources. Schema is present but
# inert; the built-in RBAC provider and runtime wiring arrive in subsequent phases.
# See RFC: https://github.com/bytedance/deer-flow/issues/4063
authorization:
  enabled: false

# ============================================================================
# Circuit Breaker Configuration
# ============================================================================
# Circuit breaker for LLM calls prevents repeated requests to a failing provider.
# When the failure threshold is reached, subsequent calls fast-fail until recovery.
#
# This is useful for:
# - Avoiding rate-limit bans during provider outages
# - Reducing resource exhaustion from retry loops
# - Gracefully degrading when LLM services are unavailable

# circuit_breaker:
#   # Number of consecutive failures before opening the circuit (default: 5)
#   failure_threshold: 5
#   # Time in seconds before attempting to recover (default: 60)
#   recovery_timeout_sec: 60

# ============================================================================
# SSO / OIDC Authentication (optional)
# ============================================================================
# Enable SSO login via any OIDC-compatible provider (Keycloak, Google, Azure AD, Okta, etc.).
# When enabled, the login page will show SSO buttons alongside the standard email/password form.
#
# Provider configuration:
#   - issuer:      The OIDC issuer URL (e.g. https://keycloak.example.com/realms/deerflow)
#   - client_id:   OAuth2 client ID from the provider
#   - client_secret: OAuth2 client secret ($ENV_VAR references supported)
#   - redirect_uri: Callback URL. In production, this must match what you configure in
#                   the provider. Defaults to a self-derived URL in development.
#
# Keycloak setup:
#   1. Create a client with type "confidential" and Standard Flow enabled
#   2. Add Valid Redirect URI: http://localhost:8001/api/v1/auth/callback/keycloak
#   3. Add Web Origin: http://localhost:8001 (or your frontend origin)

# auth:
#   oidc:
#     enabled: true
#     # Base URL of the frontend, used for redirects after SSO callback.
#     # In production behind a reverse proxy, set this to the public frontend URL
#     # (e.g. https://deerflow.example.com). In development, leave unset.
#     # frontend_base_url: http://localhost:3000
#     providers:
#       keycloak:
#         display_name: Keycloak
#         issuer: https://keycloak.example.com/realms/deerflow
#         client_id: deerflow
#         client_secret: $KEYCLOAK_CLIENT_SECRET
#         # Optional: explicitly set the callback URL.
#         # redirect_uri: https://deerflow.example.com/api/v1/auth/callback/keycloak
#         scopes:
#           - openid
#           - email
#           - profile
#         token_endpoint_auth_method: client_secret_post
#
#         # User provisioning settings (safe defaults shown below):
#         auto_create_users: true          # Auto-create DeerFlow account on first SSO login
#         require_verified_email: true     # Reject SSO logins without verified email
#         # allowed_email_domains:          # Restrict to specific email domains
#         #   - example.com
#         admin_emails: []                 # Auto-grant admin role to these emails
#
#         # Security features (enabled by default):
#         pkce_enabled: true               # PKCE (S256) for authorization code flow
#         nonce_enabled: true              # Nonce validation in ID tokens