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.
# - Set `DEER_FLOW_DATE_TIMEZONE` to an IANA timezone name (e.g. `Asia/Shanghai`)
#   when the conversation date injected into agents should follow a zone other
#   than the server's local timezone. It is read at runtime by the date-context
#   middlewares (not a config-schema field), so it can be set on a container
#   without mounting a `config.yaml`.
# - 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# Config Version (used to detect outdated config files)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: 55

# Optional operator-owned literal prompt extensions (see CONFIGURATION.md):
# lead_prompt_overlay:
#   prepend: "Apply our organization-wide writing conventions."
#   append: "End reports with a concise decision summary."
# Subagents: subagents.agents.<name>.prompt_overlay.{prepend,append}
# DeerMem: memory.backend_config.{prompt_prepend,prompt_append}

# Optional per-model request pacing (inside a models[] entry):
# request_admission:
#   requests_per_minute: 60
#   group: shared-provider-account  # optional; defaults to the model config name
#   max_wait_seconds: 300
#   max_queue_size: 256
# Off unless configured. Shared groups require identical settings; restart after
# changes. Process-local RPM only, not TPM or a cluster-wide quota. See the
# "Model request admission" section in backend/docs/CONFIGURATION.md.

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

# Trace ids are always issued and always returned in the `X-Trace-Id` response
# header; this block controls log output only — whether records carry a
# `trace_id` field, and in which format. Off by default because enabling it
# changes the log format. Restart required (see reload_boundary.py).
logging:
  enhance:
    enabled: false
    format: text

# ============================================================================
# Agent Extensions
# ============================================================================
# Optional AgentMiddleware classes loaded into the lead and subagent runtime
# middleware chains after built-in runtime middlewares, but before the
# safety/clarification tail. Missing packages, invalid classes, and broken
# modules fail at agent creation with an actionable import error; constructor
# errors fail at agent creation with the underlying exception.
# The same list applies to both lead and subagent runtimes; lead-only vs
# subagent-only configuration is not expressible yet. Each entry is a class
# path or `{class, kwargs}` for constructor arguments. kwargs values must be
# JSON types (object, array, string, number, boolean, or null); YAML dates
# and timestamps are coerced to ISO strings so they match JSON. Treat these
# files as trusted operator config because middleware classes execute code.
# Uncomment this block to define middlewares in config.yaml. Leaving it commented
# lets extensions_config.json remain the source of truth for this fixed-slot
# (not deprecated) middleware list. Packaged plugins use the `plugins:` block below.
# extensions:
#   middlewares:
#     - my_company.deerflow_middlewares:DomainGuardMiddleware
#     - class: my_company.deerflow_middlewares:LatencyStampingMiddleware
#       kwargs:
#         header: X-DeerFlow-Latency

# ============================================================================
# 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 Limits
# ============================================================================
# A run's recursion_limit caps the number of LangGraph super-steps (each is at
# least one LLM call). Clients may override the default per request; invalid or
# non-positive client values fall back to this configured default.
recursion_limit: 100

# The Gateway never trusts a configured or client-supplied value above this
# ceiling, preventing runaway API cost / DoS. Raise it 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.
#
# Two token fields look similar but mean different things:
# - `max_tokens` is the per-call OUTPUT cap passed to the provider.
# - `context_window` is a positive integer for the total context capacity
#   (prompt + completion). It drives the real-time "% context used" indicator
#   in the chat UI and feeds the model's langchain profile, which fraction-based
#   summarization triggers (`summarization.trigger: [{type: fraction, ...}]`)
#   resolve their thresholds from. Third-party OpenAI-compatible models carry no
#   built-in profile, so without `context_window` a fraction trigger degrades
#   (dropped with a warning) instead of crashing the agent build.
# Leave `context_window` unset if the provider limit is unknown; the percentage
# will not render. Verify configured values against the provider's model docs.
#
# 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. Mixed currencies disable cost reporting to prevent invalid sums.
# 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
  #   context_window: 262144     # Total prompt + completion capacity
  #   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

  # Reasoning capability contract (optional, per model)
  #
  # A mapping-valued `reasoning:` block opts into this contract. The existing
  # Ollama `reasoning: true` / `false` boolean, and the `low` / `medium` / `high`
  # level string gpt-oss style models take, remain native provider options and
  # are still forwarded to ChatOllama.
  #
  # `supports_thinking` / `supports_reasoning_effort` only say whether a knob
  # exists. Models whose provider contract differs from DeerFlow's generic
  # assumptions can declare a `reasoning:` block instead; the two booleans are
  # then derived from it and every model-creation path (chat, summarization,
  # title generation, subagents) enforces the same normalized policy:
  #
  #   reasoning:
  #     thinking: optional             # unsupported | optional | required
  #     on_disable_request: keep_enabled   # required only: keep_enabled | reject
  #     dialect: auto                  # auto | openai_extra_body | anthropic |
  #                                    # vllm_chat_template | ollama | none
  #     history: clear                 # preserve | clear (omit for provider default)
  #     effort:
  #       values: [low, medium, high]  # provider vocabulary, in display order
  #       default: medium              # applied when the caller does not choose
  #       aliases: {minimal: low}      # DeerFlow generic value -> provider value
  #       path: reasoning_effort       # dotted identifier path (default)
  #
  # `required` thinking never receives a disable payload, unknown effort values
  # never reach the provider (they map through `aliases` or fall back to
  # `default`), and `dialect` lets the factory synthesize the on/off payload
  # when `when_thinking_enabled` / `when_thinking_disabled` are absent
  # (`auto` infers it from `when_thinking_enabled`). `default` also applies to
  # callers that never choose an effort — summarization, title generation and
  # subagents — so pick a level you are happy to pay for on every background
  # call. Contradictory profiles fail at startup: `required` plus
  # `when_thinking_disabled`, a `default` outside `values`, an effort value in
  # the profile or in `when_thinking_*` templates that the contract rejects, a
  # stale `reasoning_effort` beside a custom `effort.path`, or a `path` that
  # would overwrite a whole mapping such as `extra_body`.
  # Profiles without `reasoning:` behave exactly as before.

  # Example: Z.AI GLM-5.3-Flash (required thinking, restricted effort)
  #
  # GLM-5.3-Flash cannot disable thinking and only accepts low/high/max effort.
  # The contract keeps thinking on for every foreground and background call,
  # maps DeerFlow's generic minimal/medium presets onto low/high, and
  # serializes clear_thinking=true so the model does not require exact
  # reasoning replay after summarization. `default: high` keeps background
  # calls off the provider's deepest budget; `max` stays selectable in the UI.
  #
  # - name: glm-5.3-flash
  #   display_name: GLM-5.3-Flash
  #   use: deerflow.models.patched_deepseek:PatchedChatDeepSeek
  #   model: glm-5.3-flash
  #   api_base: https://api.z.ai/api/paas/v4
  #   api_key: $ZAI_API_KEY
  #   timeout: 600.0
  #   max_retries: 2
  #   temperature: 1.0
  #   top_p: 0.95
  #   max_tokens: 131072
  #   context_window: 1000000
  #   supports_vision: true
  #   stream_usage: false
  #   reasoning:
  #     thinking: required
  #     dialect: openai_extra_body
  #     history: clear
  #     effort:
  #       values: [low, high, max]
  #       default: high
  #       aliases:
  #         minimal: low
  #         medium: high
  #   extra_body:
  #     tool_stream: true

  # 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          # Per-call output cap
  #   context_window: 128000    # Total prompt + completion capacity
  #   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
  #   # Optional, forwarded to ChatOpenAI as-is (default false): send only the new
  #   # turn plus previous_response_id instead of replaying the full history. The
  #   # chained context is still billed as input tokens, and any client-side
  #   # history rewrite (e.g. blocked-write payload elision) is only honored when
  #   # the history is replayed.
  #   use_previous_response_id: false
  #   context_window: 400000
  #   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
  #   context_window: 32768              # Match the context length served by Ollama
  #   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
  #   context_window: 32768
  #   supports_thinking: true
  #   supports_vision: true
  #
  # For Docker deployments, use host.docker.internal instead of localhost:
  #   base_url: http://host.docker.internal:11434

  # Example: llama.cpp (OpenAI-compatible local server)
  #
  # llama-server exposes an OpenAI-compatible /v1/chat/completions endpoint,
  # so use ChatOpenAI and point base_url at the server's /v1 root. --alias
  # gives the served model a stable API name independent of the GGUF filename.
  #
  # Start llama.cpp, for example:
  #   llama-server -m /path/to/model.gguf --host 0.0.0.0 --port 8080 --alias qwen-local
  #
  # Tool calling depends on the model's chat template supporting tools. Current
  # llama.cpp enables its Jinja template engine by default; if you override the
  # template, use one that supports function/tool calls.
  #
  # - name: qwen-llamacpp
  #   display_name: Qwen (llama.cpp)
  #   use: langchain_openai:ChatOpenAI
  #   model: qwen-local
  #   api_key: no-key              # Non-empty placeholder when llama-server auth is disabled
  #   base_url: http://localhost:8080/v1
  #   request_timeout: 600.0
  #   max_retries: 2
  #   supports_vision: false
  #
  # If llama-server was started with --api-key, replace the placeholder with
  # that key (prefer an environment variable). In Docker, localhost is the
  # DeerFlow container itself; use an address reachable from the container,
  # for example on Docker Desktop:
  #   base_url: http://host.docker.internal:8080/v1

  # 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         # Per-call output cap
  #   context_window: 200000    # Total prompt + completion capacity
  #   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
  #   context_window: 1048576
  #   supports_vision: true

  # Example: Gemini model via Google's official OpenAI-compatible endpoint
  # 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
  # Google accepts reasoning_effort, not extra_body.thinking. Gemini 3.1 Pro
  # cannot disable thinking; dialect: none avoids emitting a thinking toggle.
  # https://ai.google.dev/gemini-api/docs/openai#thinking
  # - name: gemini-3.1-pro-preview
  #   display_name: Gemini 3.1 Pro (Thinking)
  #   use: deerflow.models.patched_openai:PatchedChatOpenAI
  #   model: gemini-3.1-pro-preview
  #   api_key: $GEMINI_API_KEY
  #   base_url: https://generativelanguage.googleapis.com/v1beta/openai/
  #   request_timeout: 600.0
  #   max_retries: 2
  #   max_tokens: 16384
  #   context_window: 1048576
  #   supports_vision: true
  #   reasoning:
  #     thinking: required
  #     dialect: none
  #     effort:
  #       values: [minimal, low, medium, high]

  # 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
  #   # Enable only when the endpoint reports cumulative usage on every stream chunk.
  #   cumulative_stream_usage: true
  #   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

# ============================================================================
# Knowledge Capability
# ============================================================================
# Optional knowledge capability. The feature is disabled by default. Provider
# connection and retrieval settings belong to the matching entry in `tools:`
# below (for example, the RAGFlow `knowledge_search` tool), not here.
knowledge_base:
  enabled: false
  # Custom-agent chat only. Keeps the selector hidden unless the effective
  # knowledge_search provider is the bundled RAGFlow tool.
  scope_selection_enabled: false

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

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

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

tools:
  # Read explicitly referenced conversations (Gateway API only, opt-in).
  # Each run must submit conversation_references; see backend/docs/API.md.
  # - name: read_conversation
  #   group: conversation
  #   use: deerflow.tools.conversation:read_conversation

  # RAGFlow knowledge retrieval (read-only). Uncomment this single entry.
  # `datasets` is optional. Omit it to list every tenant-visible dataset at
  # search time; an explicit `datasets: []` is invalid. Empty datasets are
  # skipped; searchable datasets are grouped by
  # embedding model and retrieved with at most four groups in parallel. Group
  # ranks are interleaved under the global `page_size` limit; scores are omitted
  # when multiple groups are searched because they are not comparable.
  # Configure stable IDs only to restrict scope. IDs and catalog listing never
  # reach the model.
  # - name: knowledge_search
  #   group: knowledge
  #   use: deerflow.community.ragflow.tools:knowledge_search_tool
  #   base_url: http://localhost:9380  # Docker: use a backend-reachable URL
  #   api_key: $RAGFLOW_API_KEY
  #   datasets:  # Optional operator-controlled allowlist
  #     - 0123456789abcdef0123456789abcdef  # Replace with a RAGFlow dataset ID
  #   timeout: 30
  #   page_size: 8
  #   similarity_threshold: 0.2
  #   vector_similarity_weight: 0.3
  #   top_k: 256
  #   max_chars_per_chunk: 800
  #   max_total_chars: 8000
  # - name: list_knowledge_bases
  #   group: knowledge
  #   use: deerflow.community.ragflow.tools:list_knowledge_bases_tool

  # LightRAG knowledge retrieval (read-only). Alternative provider for the same
  # knowledge_search tool; uncomment this entry instead of the RAGFlow one —
  # duplicate names keep the first entry, so configure exactly one. Requires
  # LightRAG v1.4.9+ (response envelope and citation fields). Searches the deployment's
  # single indexed workspace through the data retrieval endpoint (no LLM
  # generation inside LightRAG). `api_key` is optional; only omit it for
  # loopback/trusted-network servers — when auth is enabled the key is sent as
  # the X-API-Key header. `mode` picks the retrieval strategy: naive, local,
  # global, hybrid, or mix (default mix, matching LightRAG's own default).
  # Internal chunk and reference identifiers never reach the model.
  # - name: knowledge_search
  #   group: knowledge
  #   use: deerflow.community.lightrag.tools:knowledge_search_tool
  #   base_url: http://localhost:9621  # Docker: use a backend-reachable URL
  #   api_key: $LIGHTRAG_API_KEY  # Omit only for unauthenticated trusted servers
  #   mode: mix
  #   timeout: 30
  #   top_k: 60
  #   chunk_top_k: 8  # Optional; omit to use the server default
  #   max_chars_per_chunk: 800
  #   max_total_chars: 8000

  # 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
  #   include_domains: []  # Optional domain-only lists, at most 10 entries each
  #   exclude_domains: []  # Exclusion wins; see backend/docs/CONFIGURATION.md
  #   # api_key: $SERPER_API_KEY  # Optional explicit env-var reference

  # Web search tool (uses Serply - Google Search API, requires SERPLY_API_KEY)
  # Serply returns live Google results and covers Google News and Google Scholar
  # with the same key. Sign up at https://serply.io (docs: https://serply.io/docs)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.serply.tools:web_search_tool
  #   max_results: 5                 # Serply accepts 1-100 per request
  #   # vertical: search             # search (default), news, or scholar
  #   # gl: us                       # Optional country code for the results
  #   # hl: en                       # Optional interface language
  #   # api_key: $SERPLY_API_KEY     # Optional if the env var is set

  # 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
  #   # Optional search-source filters; omit or use [] for no restriction of that kind.
  #   # include_domains:
  #   #   - docs.python.org
  #   #   - developer.mozilla.org
  #   # exclude_domains: []
  #   # 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 Tencent Cloud Web Search API, requires a service API key)
  # Create a service API key in the Tencent Cloud WSA console, then set
  # TENCENTCLOUD_WSA_APIKEY in the Gateway environment. Do not use Tencent Cloud
  # SecretId/SecretKey for this provider.
  # - name: web_search
  #   group: web
  #   use: deerflow.community.tencent_wsa.tools:web_search_tool
  #   max_results: 5                 # 1-50; values above 10 request Tencent's Cnt option
  #                                    # (requires a Tencent plan that supports Cnt)
  #   # mode: 0                      # Optional: 0=web, 1=VR, 2=mixed
  #   # api_key: $TENCENTCLOUD_WSA_APIKEY  # Optional explicit env-var reference

  # 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. Cloud requires FIRECRAWL_API_KEY; self-host may need no key.)
  # - name: web_search
  #   group: web
  #   use: deerflow.community.firecrawl.tools:web_search_tool
  #   max_results: 5
  #   # api_key: $FIRECRAWL_API_KEY
  #   # base_url: http://<host>:3002  # optional: self-hosted Firecrawl; no cloud API key needed

  # 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 search tool (uses Sofya, requires SOFYA_API_KEY)
  # Sofya returns the content of the result pages, not only their snippets.
  # Sign up at https://sofya.co to get a key.
  # Note: set SOFYA_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.sofya.tools:web_search_tool
  #   max_results: 5                # capped at 20 by the Sofya provider
  #   # search_depth: basic         # basic reads the result pages, snippets returns snippets only
  #   contents_max_characters: 2000 # per-result content cap; 0 disables it
  #   # api_key: $SOFYA_API_KEY     # Optional explicit env-var reference

  # 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: $BROWSERLESS_TOKEN        # API token (required for Browserless Cloud; optional for self-hosted)
  #   timeout_s: 30                       # Request timeout in seconds
  #   # allow_private_addresses: false    # SSRF guard: keep false in production. Set true ONLY for intentional internal targets.
  #   # wait_for_event: "networkidle"    # Wait for a page event before returning (e.g. "load", "networkidle")
  #   # wait_for_timeout_ms: 2000        # Extra wait after page load in milliseconds
  #   # wait_for_selector: "article"     # CSS selector to wait for before returning
  #   # wait_for_selector_timeout_ms: 5000  # How long to wait for that selector, in milliseconds
  #   # reject_resource_types: ["image", "media", "font"]  # Skip these resource types
  #   # reject_request_pattern: ["\\.css$", "\\.png$"]     # Skip URLs matching these regexes
  #   #                                    # Both accept a YAML list or a
  #   #                                    # comma-separated string. Regexes containing
  #   #                                    # commas must use the YAML list form.

  # Web fetch tool (uses Crawl4AI - self-hosted headless Chromium, no third-party API key)
  # Crawl4AI returns server-cleaned "fit" markdown directly (no readability step needed),
  # ideal for JavaScript-heavy sites. Self-host it:
  #   docker run -d -p 11235:11235 --shm-size=1g \
  #     -e CRAWL4AI_API_TOKEN=$CRAWL4AI_TOKEN unclecode/crawl4ai:0.9.2
  # Crawl4AI >= 0.9 is secure-by-default: a bearer token is REQUIRED on every request
  # except GET /health, and a server started without CRAWL4AI_API_TOKEN binds 127.0.0.1
  # only (so a container/remote DeerFlow cannot reach it at all). Set the same value in
  # the server env and in `token:` below, or requests fail with HTTP 401.
  # Use >= 0.8.7: earlier images, including 0.8.6, carry known pre-auth RCEs.
  # 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)
  #   token: $CRAWL4AI_TOKEN              # Bearer token; required by Crawl4AI >= 0.9
  #   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

  # 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 (`timeout` also accepted)
  #   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.)

  # Agentic browser control (stateful navigate → observe → click/type loop).
  # Unlike web_fetch/web_capture (read-only, stateless), these tools keep a live
  # per-thread Playwright browser so the agent can interact with JavaScript-heavy
  # pages, fill forms, and follow multi-step flows. Elements are addressed by the
  # numeric [ref] index returned in each page snapshot.
  #
  # Requires the optional dependency, installed once per environment:
  #   cd backend && uv sync --extra browser && uv run playwright install chromium
  # make dev / Docker startup auto-detect an active browser_navigate entry and
  # preserve the browser extra during dependency sync.
  #
  # Uncomment the whole block to enable. Launch and URL-policy settings are read
  # canonically from browser_navigate and shared by every browser tool + Live.
  # - name: browser_navigate
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_navigate_tool
  #   headless: true                   # Set false only for local, trusted debugging
  #   timeout_ms: 30000                # Per-action navigation/interaction timeout
  #   viewport_width: 1280
  #   viewport_height: 720
  #   # allow_private_addresses: false # SSRF guard: keep false in production. Also applies to the
  #   #                                # pinned egress proxy every launched browser connects through.
  #   # cdp_url: http://127.0.0.1:9222 # Attach to YOUR running Chrome (Codex-style)
  #   # allow_unguarded_cdp: false      # REQUIRED true with cdp_url: attaching to an
  #   #                                # existing Chrome prevents DeerFlow from
  #   #                                # enforcing its subresource/redirect SSRF guard.
  #   #                                # instead of launching a private headless one.
  #   #                                # Start Chrome with --remote-debugging-port=9222
  #   #                                # so you watch the agent drive your real browser
  #   #                                # with your real login sessions. Local/trusted only;
  #   #                                # never enable for an untrusted CDP endpoint.
  #   #                                # Chrome must be version-matched to the bundled
  #   #                                # Playwright; a much newer Chrome can reject the
  #   #                                # CDP handshake. If so, use headless: false to launch
  #   #                                # Playwright's own visible browser instead.
  #   # Multi-worker note: browser sessions live in one Gateway worker's memory.
  #   # Keep the Gateway to one worker process while this tool group is enabled:
  #   # GATEWAY_WORKERS=1, and WEB_CONCURRENCY unset or 1 on the launches that pass
  #   # uvicorn no --workers (backend/Dockerfile, scripts/serve.sh).
  # - name: browser_snapshot
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_snapshot_tool
  # - name: browser_click
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_click_tool
  # - name: browser_type
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_type_tool
  # - name: browser_get_text
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_get_text_tool
  #   max_chars: 8000                  # Truncation cap for browser_get_text output
  # - name: browser_back
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_back_tool
  # - name: browser_screenshot
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_screenshot_tool
  # - name: browser_close
  #   group: browser
  #   use: deerflow.community.browser_automation.tools:browser_close_tool

  # 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
    # Opt-in retries may increase upstream requests and cost.
    # max_retries: 0
    # retry_budget_seconds: 30
    # 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. Cloud requires FIRECRAWL_API_KEY; self-host may need no key.)
  # - name: web_fetch
  #   group: web
  #   use: deerflow.community.firecrawl.tools:web_fetch_tool
  #   # api_key: $FIRECRAWL_API_KEY
  #   # base_url: http://<host>:3002  # optional: self-hosted Firecrawl; no cloud API key needed

  # 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.

  # Web fetch tool (uses Sofya, requires SOFYA_API_KEY)
  # Returns the page as markdown. Also handles PDF, DOCX and other documents.
  # 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.sofya.tools:web_fetch_tool
  #   # api_key: $SOFYA_API_KEY

  # Web fetch tool (uses Unbrowse, requires UNBROWSE_API_KEY)
  # Returns the page as markdown. Fetches over plain HTTP when that is enough and
  # falls back to a hosted cloud browser for pages that need JavaScript.
  # Get a key at https://unbrowse.ai/app.
  # 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.unbrowse.tools:web_fetch_tool
  #   # render: auto                  # auto | never (plain HTTP only) | always (cloud browser)
  #   # api_key: $UNBROWSE_API_KEY

  # 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.
#
# The same middleware also elides the `content` argument of successful
# write_file calls from model-bound requests once the same path was read or
# modified again later (issue #5328): the file on disk is then the source of
# truth, and the read-before-write gate forces a read_file before the next
# modification anyway, so the historical copy only costs context. Stored
# history, receipts, and the run journal keep the original arguments.

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
  # Replace a successful write_file call's `content` with a short placeholder
  # pointing at read_file once a later successful read_file / write_file /
  # str_replace of the same path exists. Only the model-bound request changes.
  elide_superseded_writes: true
  # Only elide content at least this many characters long (a character count,
  # not tokens: CJK text costs ~3-4x more per character than ASCII). 0 elides
  # every non-empty content.
  superseded_write_min_chars: 2000
  # The newest N successful write_file calls (across all paths) always stay
  # visible, even when superseded, so the model can still say what it just
  # wrote without a read. 0 keeps none.
  keep_recent_writes: 1

# ============================================================================
# Tool Artifact Handle Registry Configuration (issue #4676)
# ============================================================================
# Captures lightweight artifact references (paths, URLs, task ids) from tool
# results into durable per-thread state (`ThreadState.tool_artifacts`) so they
# survive context compaction. The model references artifacts by short handles
# (`art_xxxxxxxx`) that are resolved to real references before tool execution.
# Handles are listed in the model context and rendered as badges in the UI.

tool_artifacts:
  enabled: true
  # Maximum artifact entries retained per thread (a sliding window: the
  # oldest entries are evicted when the cap is exceeded).
  max_entries: 100
  # Conservatively scan free-text tool results for sandbox paths / remote file URLs.
  detect_refs_in_text: true
  # Project available artifact handles into the model request as durable context.
  inject_model_context: true
  # Resolve artifact handles found in tool arguments to real references before execution.
  resolve_handles_in_args: true

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

suggestions:
  enabled: true
  max_suggestions: 3


# ============================================================================
# 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
  # Compatibility name: bounds retained (thread_id, run_id) histories, not whole threads.
  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

# ============================================================================
# PII Redaction Middleware Configuration (issue #3190)
# ============================================================================
# Deterministic PII redaction in model-bound context. When enabled, emails,
# phone numbers, API keys, credit-card numbers (Luhn-validated) and national
# IDs found in genuine user messages and remote-content tool results
# (web_fetch / web_search / image_search / web_capture and MCP tools) are
# rewritten to keyed, value-derived placeholders like
# [EMAIL_bdqkxwcfmqvrqzblcchjkqbcdqw] before they reach the model. Tokens are
# HMAC digests keyed by token_secret, so they cannot be re-derived offline.
# Raw text stays in thread state for the UI; queued memory-extraction
# payloads are redacted too. Compaction and configured LLM title inputs are
# redacted as well; placeholders are value-derived, so an identity redacted
# before compaction renders the same token afterwards.
pii_redaction:
  enabled: false
  # redact_email: true
  # redact_api_key: true
  # redact_credit_card: true
  # redact_phone: true
  # redact_national_id: true
  # Required when enabled: true (non-empty, at least 16 characters) — the
  # deployment-scoped HMAC secret for the placeholder digests. Tokens are
  # linkable only within this deployment and cannot be re-derived offline
  # without it. Enabling redaction without a valid secret fails config
  # validation: an unkeyed digest would be a publicly computable, globally
  # linkable fingerprint of the raw values (guessable for phone numbers).
  # token_secret: change-me-to-a-long-random-string

# ============================================================================
# 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
  # A blocked call never ran and must be re-issued after a re-read, so its
  # payload (write_file content, str_replace old_str/new_str) is dead weight in
  # every later model call. Replace it with a short placeholder in model-bound
  # requests; stored history, receipts, and the run journal keep the original.
  elide_blocked_payloads: true
  # Only elide payload fields at least this many characters long (a character
  # count, not tokens: CJK text costs ~3-4x more per character than ASCII).
  # 0 elides every non-empty payload.
  elide_min_chars: 2000

# ============================================================================
# Projects
# ============================================================================
# User projects group threads under a name plus free-form instructions. Member
# threads receive the project's current instructions as a bounded, request-
# scoped <project> block rendered from the snapshot pinned at run admission;
# the block never enters the system prompt or persisted history. The shelf
# index bounds below cap the request-scoped <documents> listing rendered from
# the same pinned snapshot. Oversized instructions are rejected at write time
# (422) and never truncated.

projects:
  # Hard UTF-8 byte cap for project instructions (range: 256-262144). Multi-
  # byte characters count as their UTF-8 byte length, not as characters.
  instructions_max_bytes: 8192
  # Maximum shelf entries rendered into the <documents> index per run
  # (range: 1-500).
  shelf_index_max_entries: 50
  # UTF-8 byte cap for the rendered <documents> index, counting IDs, escaped
  # names, wrappers, separators and the overflow note (range: 512-65536).
  # Usually binds before the entry cap for CJK document names.
  shelf_index_max_bytes: 4096
  # Days a trashed project document stays recoverable before the retention
  # sweep may purge it permanently (range: 1-3650).
  trash_retention_days: 30

# ============================================================================
# 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]

# ============================================================================
# Uploads
# ============================================================================
# Limits and conversion settings for files attached to a conversation.
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 Configuration
# ============================================================================
# Choose between local sandbox (direct execution), Docker-based AIO sandbox,
# or BoxLite micro-VM sandbox
# Option 1: Local Sandbox (Default)
# Executes commands directly on the host machine
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.
  # A read_file cut lands on a line boundary when the budget allows, and the marker names
  # the next start_line to continue with, so the agent can page through a long file exactly.
  # Set to 0 to disable truncation.
  bash_output_max_chars: 20000
  read_file_output_max_chars: 50000
  ls_output_max_chars: 20000

  # Provider command deadline (default: 600). AIO images on the supported
  # semver line (1.9.3+, recommended 1.11.0) enforce it server-side through
  # `hard_timeout`; the frozen legacy `all-in-one-sandbox:latest` image only
  # gets the bounded host-side request. `bash_command_timeout` is used by
  # providers that explicitly wire this setting (currently LocalSandbox,
  # AioSandbox, and OpenSandbox). Other providers retain their provider-specific
  # command defaults unless a caller supplies an explicit timeout. For providers
  # that enforce this deadline server- or process-side, blocking foreground
  # commands are terminated at the configured deadline. Long-lived processes
  # should be started in the background with output redirected; unredirected
  # background output is drained with bounded capture.
  bash_command_timeout: 600

# Option 2: Container-based AIO Sandbox
# Executes commands in isolated containers (Docker or Apple Container)
# On macOS: Prefers Apple Container when available, but keeps Docker while this
# sandbox prefix still has managed Docker containers to reconcile after a mode change
# 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: outbound network policy for locally managed Docker sandboxes.
#   # open keeps the current behavior. isolated denies all outbound traffic.
#   # allowlist permits the listed HTTP/HTTPS domains; denied public domains can
#   # optionally be approved from the Human Input card. Restricted modes require
#   # Docker Engine 28+ and are not supported by Apple Container or provisioner mode.
#   # Private, loopback, link-local, multicast, and cloud metadata addresses are always denied.
#   network:
#     mode: allowlist                    # open | isolated | allowlist (default: open)
#     allow_domains:
#       - pypi.org
#       - files.pythonhosted.org
#       - registry.npmjs.org
#       - github.com
#     approval: prompt                  # deny | prompt
#     temporary_grant_ttl: 300          # 30-3600 seconds
#     proxy_image: ghcr.io/bytedance/deer-flow-sandbox-network-proxy:latest
#
#   # Optional: Override whether the sandbox already sees the gateway's
#   # thread workspace/uploads/outputs through shared mounts.
#   # Omit this field to auto-detect from the backend (local containers: true;
#   # remote/provisioner backends: false). Set true only when the deployment
#   # guarantees both sides use the same thread user-data directories, such as
#   # a correctly aligned shared PVC, NFS volume, hostPath, or bind mount.
#   # true skips per-upload sandbox acquire/sync; false forces explicit sync.
#   # thread_data_mounts: true
#
#   # 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
#   # AIO semver images default MAX_SHELL_SESSIONS to 10. DeerFlow leaves that
#   # image default unchanged while subagent_runtime.max_running + 1 fits, and
#   # otherwise injects the required value for local and provisioner sandboxes.
#   # An explicit value is allowed but must be at least max_running + 1.
#   # environment:
#   #   MAX_SHELL_SESSIONS: "13"  # Example for subagent_runtime.max_running: 12
#   #   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
#   #   # Required when using the OpenAI-compatible image-generation provider:
#   #   IMAGE_GENERATION_PROVIDER: $IMAGE_GENERATION_PROVIDER
#   #   IMAGE_GENERATION_API_KEY: $IMAGE_GENERATION_API_KEY
#   #   IMAGE_GENERATION_BASE_URL: $IMAGE_GENERATION_BASE_URL
#   #   IMAGE_GENERATION_MODEL: $IMAGE_GENERATION_MODEL
#   #   IMAGE_GENERATION_SIZE: $IMAGE_GENERATION_SIZE  # Optional fixed-size override
#
#   # Optional: Cross-instance container ownership (issue #4206).
#   #
#   # Gateway instances share sandbox containers but each keeps its own in-memory
#   # warm pool. Without shared ownership state, one instance's startup
#   # reconciliation adopts a container another instance is actively using and
#   # later idle-destroys it — tool calls then fail with 502 / connection refused.
#   #
#   # Single gateway instance? Leave this out; `memory` is the default and the
#   # cross-instance kill cannot happen.
#   #
#   # MULTIPLE gateway instances / workers sharing one container backend
#   # (load-balanced deployments, Docker Compose) MUST set type: redis. Docker
#   # Compose already sets DEER_FLOW_STREAM_BRIDGE_REDIS_URL, which is taken as
#   # proof the deployment is multi-instance, so redis ownership is inferred even
#   # if this section is omitted.
#   #
#   # The redis ownership store 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
#   #
#   # ownership:
#   #   type: memory                 # single gateway instance only
#   #
#   # ownership:
#   #   type: redis                  # required for multi-instance / load-balanced
#   #   redis_url: redis://redis:6379/0
#   #   renewal_interval_seconds: 30 # how often an owner refreshes its leases
#   #   ttl_multiplier: 4            # lease TTL = interval x this (min 2, so a
#   #                                # single missed renewal cannot expire a live
#   #                                # owner). Liveness is deliberately independent
#   #                                # of idle_timeout: renewal keeps running even
#   #                                # at idle_timeout: 0.
#   #   key_prefix: deerflow:sandbox:owner
#   #
#   # NOTE: the redis ownership store is fail-closed, matching the stream bridge's
#   # fail-hard policy. Redis.from_url is lazy so a down Redis does not block
#   # startup, but a sandbox whose ownership cannot be published is not handed out
#   # — acquiring raises instead. The alternative (proceed unowned) is exactly the
#   # #4206 cross-instance kill. Run Redis with HA / a restart policy.
#   #
#   # NOTE: the other boundary is the lease TTL (renewal_interval_seconds x
#   # ttl_multiplier). A Redis outage longer than the TTL can let a reconciling
#   # instance adopt a live owner's container: a lapsed lease is indistinguishable
#   # from a dead owner, so one TTL of adoption grace is the whole safety margin.
#   # Size the TTL against your Redis availability target (HA Redis keeps this
#   # window out of reach).

# 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.

# Option 5: Tenki cloud microVM Sandbox
# Runs each sandbox as an isolated Tenki cloud microVM. Released sandboxes stay
# in an in-process warm pool and are reclaimed by the same user/thread without a
# cold start. Requires the optional SDK: pip install "deerflow-harness[tenki]".
# sandbox:
#   use: deerflow.community.tenki:TenkiSandboxProvider
#   # api_key: $TENKI_API_KEY        # falls back to TENKI_API_KEY / TENKI_AUTH_TOKEN env var
#   # base_url: https://tenki.cloud  # optional; SDK default when omitted
#   # image: my-base-image           # optional; Tenki account default base image when omitted
#   # workspace_id: ws_...           # optional; auto-selected if the account has exactly one
#   # Migration: sandbox.project_id is gone — Tenki 1.x removed projects, so scope
#   # is the workspace alone. Set workspace_id if the account has more than one.
#   # A leftover project_id is ignored and logs a warning at startup.
#   # cpu_cores: 2                   # optional per-sandbox vCPUs
#   # memory_mb: 2048                # optional per-sandbox memory
#   # replicas: 3                    # active + warm microVM cap per gateway process
#   # idle_timeout: 600              # warm microVM idle seconds before terminate; 0 disables
#   # max_duration: 14400            # Tenki sandbox lifetime in seconds; 0 uses the account default
#   # sticky: false                  # pin the microVM to its host (only matters with pause/resume)
#   # home_dir: /home/tenki          # writable dir backing /mnt/user-data
#   # environment:                   # injected into every command (and as create-time env)
#   #   PYTHONUNBUFFERED: "1"

# Option 6: OpenSandbox remote Sandbox
# Runs each sandbox through an OpenSandbox server. Released sandboxes stay in an
# in-process warm pool and are reclaimed only by the same user/thread. Requires
# the optional SDK: pip install "deerflow-harness[opensandbox]".
# sandbox:
#   use: deerflow.community.opensandbox:OpenSandboxProvider
#   image: python:3.11
#   # api_key: $OPEN_SANDBOX_API_KEY  # optional when the SDK env var is set
#   # domain: localhost:8080          # optional; OPEN_SANDBOX_DOMAIN fallback
#   # protocol: http                  # localhost only; use https for remote domains
#   # request_timeout: 30             # finite positive management API request timeout seconds
#   # ready_timeout: 30               # finite positive create/readiness timeout seconds
#   # use_server_proxy: false         # proxy execd/file traffic through server
#   # sandbox_timeout: 14400          # finite non-negative remote lifetime seconds; 0 = explicit cleanup
#   # bash_command_timeout: 600       # default remote command timeout seconds
#   # replicas: 3                     # active + warm cap per gateway process
#   # idle_timeout: 600               # warm seconds before destroy; 0 disables
#   # environment:                    # create-time and per-command defaults
#   #   PYTHONUNBUFFERED: "1"

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

# Process-wide execution capacity. These fields are restart-required and are
# shared by ordinary `task` calls and durable batches. Waiting work is async;
# queued subagents do not occupy a scheduler thread.
subagent_runtime:
  max_running: 3
  max_queued: 64
  admission_policy: queue       # queue or reject when all slots are occupied
  queue_timeout_seconds: 300

# 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. One turn is one
#   # model call plus the tools it runs; the turn count is converted into
#   # LangGraph's step budget, so this is the number of turns you actually get.
#   # 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 discovery/activation allowlist (default: all enabled)
#       #   - web-search
#       #   - data-analysis
#     bash:
#       timeout_seconds: 300   # 5 minutes for quick command execution
#       max_turns: 80
#       # skills: []            # No discoverable/activatable 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 discovery/activation allowlist (null = 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.

# Durable native-subagent batches. Disabled by default because enabling it can
# materially increase model usage and requires database.backend sqlite/postgres.
# The three limits are intentionally separate:
# - total: all persisted items in one batch
# - live: pending work admitted as queued/running at one time
# - running: items from one batch allowed to hold real execution slots
subagent_batches:
  enabled: false
  poll_interval_seconds: 1
  lease_seconds: 120
  max_items_per_batch: 5000
  default_max_live_items: 100
  max_live_items_per_batch: 1000
  default_max_running_items: 3
  max_running_items_per_batch: 64
  max_attempts: 3
  max_result_chars: 100000
  result_preview_max_chars: 2000

# ============================================================================
# Tool Result Verification
# ============================================================================
# Deterministic receipts are stamped onto tool results and injected into the
# model context so final reports can cite executed actions. Disable them only
# if the extra provenance context is not wanted. The selective judge settings
# are reserved for acceptance-criteria review and remain off by default.
verification:
  receipts_enabled: true
  # Lead-chain ledger rendering: 'delegation_only' renders only while
  # processing subagent results (subagent chains always render).
  receipts_render_mode: "delegation_only"
  judge_enabled: false
  judge_model_name: null

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

# acp_agents:
#   mcode:
#     # MiniMax Code speaks ACP directly; no adapter package is required.
#     # Install with `npm install --global @minimax-ai/code`, then run `mcode login`.
#     command: mcode
#     args: ["acp"]
#     description: MiniMax Code for implementation, refactoring, debugging, and repository tasks
#     # auto_approve_permissions: false  # Enable only for trusted tasks that need MCode to edit files or run commands
#     # timeout_seconds: 1800            # Abort + kill the subprocess if it doesn't respond in time (default: 1800 = 30 min)
#
#   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
#     # timeout_seconds: 1800            # Abort + kill the subprocess if it doesn't respond in time (default: 1800 = 30 min)
#     # 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
  # AIO/provisioner and E2B modes require a canonical absolute non-root path
  # that does not overlap reserved platform mounts. Providers snapshot this
  # path for identity, mounts, metadata, and synchronization, so restart the
  # Gateway after changing it.
  # 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 (nested-archive,
  # secret-pattern, and other content-level checks). Safe archive extraction
  # (path traversal, symlinks, executable-binary, total-size, and entry-count
  # limits) and the LLM skill scanner still run unconditionally.
  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

# Optional task notes and keyword recall of compacted messages. See docs/task-continuity.md.
task_continuity:
  enabled: false
  max_batches: 32
  max_records_per_batch: 256
  max_record_chars: 16000

summarization:
  enabled: true

  # Model to use for summarization.
  #   null = summarize with the model the run actually uses (the lead run's model, a
  #          subagent's own model, or a thread's custom-agent model), NOT models[0].
  #   set  = that model generates; if its provider fails, compaction falls back to the
  #          run's own model so a broken summary provider cannot disable compaction.
  # 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.
    # The percentage resolves from the SUMMARY model's declared `context_window`
    # (summarization.model_name when set, else the run's own model): declare it
    # on that models entry — third-party OpenAI-compatible models carry no
    # built-in profile. Without it the fraction clause is dropped with a warning
    # and any remaining absolute clauses keep working. If a separate summary
    # model with a larger window is configured, prefer absolute `tokens`
    # thresholds sized for the run model instead.
    # - 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/mem0/noop/openviking) 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 / mem0 / 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.
  # Normally only one mode runs at a time. A backend that needs conversation-
  # level extraction may retain passive writes in tool mode while still
  # exposing query-aware search (mem0 does this). Tool calls go through the
  # active MemoryManager; unsupported fact CRUD returns a JSON error.
  mode: middleware
  # Pre-extraction cost gate (optional, OFF by default). Decides whether a batch
  # of new conversation is worth paying for the extraction LLM call, using one
  # `noul` question ("does this contain durable, cross-task, user-level
  # information?").
  #   off (default) - no request, no behavior change.
  #   shadow - decide and record the verdict, still extract (the safe first step).
  #   enforce - a low-probability batch skips the call AND advances the watermark,
  #     so that batch is consumed as "nothing durable here". Enable only after the
  #     measurement gate that scripts/eval_memory_prescreen.py implements and
  #     backend/docs/MEMORY_IMPROVEMENTS.md documents
  #     passes; `skip_threshold` below is provisional and unmeasured.
  # The failure direction is always extraction: an error, timeout, unusable
  # response, missing verdict, or an over-limit batch extracts as usual. A skip is
  # not a safety boundary, gates no execution, and changes no post-extraction
  # gate (confidence / scope / durability / authority / dedup / eviction).
  #
  # EGRESS: the formatted conversation text (user + final assistant turns, which
  # may contain file paths or credentials) is sent to the TypeSafe API -- a second
  # destination beyond your own extraction model. Existing memory is never sent.
  # `pii_redaction_middleware` does not cover this path.
  #
  # NOTE: a batch is not judged at all while `staleness_review_enabled` (true
  # below) or `consolidation_enabled` is on -- a skip would also skip that batch's
  # maintenance review -- and never on the emergency (pre-summarization) flush or
  # the shutdown drain. With the defaults shipped here, shadow/enforce therefore
  # sees no batches until staleness_review_enabled is turned off.
  # prescreen:
  #   mode: "off"
  #   use: deerflow.agents.memory.prescreen.typesafe:TypeSafeMemoryPrescreen
  #   config:
  #     skip_threshold: 0.2          # p below this skips; provisional, unmeasured
  #     max_state_chars: 6000        # characters of the judged batch; over it extracts
  #     cache_size: 256              # digest-keyed, successful verdicts only
  #     cache_ttl_seconds: 300.0
  #     # Connection/model/timeouts come from the top-level `typesafe:` block
  #     # unless set here (api_key_env / api_key / base_url / model / timeout /
  #     # deadline_seconds / max_attempts / retry_backoff).
  #
  # Signal classification (optional, OFF by default): two `noul` questions per
  # batch ("does this explicitly endorse / reject something the user stated
  # earlier?") mapped to the deterministic `reinforcement` / `correction` hints.
  #   off (default) - no request.
  #   shadow - decide and record only.
  #   hints - merge the labels into the extraction hint text and, only while the
  #     pre-screen is in `enforce`, veto a skip the model endorsed.
  # A model verdict never decides extraction, never drives deletion, and never
  # satisfies the reinforcement evidence gate (that gate stays deterministic).
  # signal_classification:
  #   mode: "off"
  #   use: deerflow.agents.memory.signals.typesafe:TypeSafeSignalClassifier
  #   combine: auto                  # auto | always | never: request sharing with the pre-screen
  #   config:
  #     hint_threshold: 0.5          # provisional, unmeasured
  #     max_state_chars: 6000        # a mismatch with prescreen => each side keeps its own request
  #     cache_size: 256
  #     cache_ttl_seconds: 300.0
  # Backend-private config (a dict), passed verbatim to the backend __init__.
  # Each backend self-interprets it (DeerMem parses it into DeerMemConfig).
  #
  # OpenViking example (replace this DeerMem backend_config block when
  # manager_class is openviking). This first official-adapter integration uses
  # one credential-bound OpenViking USER key for one DeerFlow user and supports
  # middleware mode only. Use owner_user_id: default when DeerFlow auth is off.
  #
  # backend_config:
  #   base_url: http://openviking:1933
  #   owner_user_id: default
  #   api_key_env: OPENVIKING_API_KEY
  #   timeout_seconds: 30
  #   default_peer_id: deerflow
  #   max_seen_message_ids: 512 # bounded hash-only capture cursor
  #   startup_policy: fail_fast
  #   failure_policy:
  #     read: fail_open
  #     write: log_and_drop
  #   retrieval:
  #     top_k: 8
  #     score_threshold: 0.25
  #     max_injection_chars: 12000
  #     content_mode: overview
  #     injection_query: >-
  #       user profile preferences important entities events ongoing goals
  #       constraints and prior decisions
  #
  # For a host-installed OpenViking used by Docker DeerFlow, set base_url to
  # http://host.docker.internal:1933 and allow_insecure_http: true. The bundled
  # optional Compose overlay uses the internal http://openviking:1933 address.
  #
  # Honcho example (replace this DeerMem backend_config block when
  # manager_class is honcho). This remote HTTP adapter uses Honcho's server-side
  # deriver to build user-model memory representations; no local LLM calls.
  #
  # backend_config:
  #   base_url: http://localhost:8000
  #   # api_key: $HONCHO_API_KEY   # hosted Honcho; plain-http + api_key needs allow_insecure_http: true
  #   workspace_prefix: deerflow-u-   # one isolated workspace per user id
  #   # workspace_overrides: {}    # map specific user ids to custom workspaces
  #   # user_peer_overrides: {}    # map specific user ids to custom peer names
  #   assistant_peer: deerflow
  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)
    storage_class: file         # file (default) or a dotted MemoryStorage class path; invalid persistent backends fail fast
    strict_user_scope: false    # set true in authenticated deployments after all callers propagate user_id
    manifest_filename: memory.json # user-global JSON: version/revision/time + user/history only; no facts or fact index
    file_lock_timeout_seconds: 10 # per-scope cross-process advisory lock timeout (single-machine local filesystem)
    retrieval_adapter: fts5     # fts5 (default), empty to disable, or a dotted RetrievalPort factory(config)
    debounce_seconds: 30 # Wait time before processing queued updates
    # Backpressure cap on pending items. 0 = unlimited. At the cap, new
    # non-signal updates are rejected (QueueFull); signal updates are always
    # admitted so important memories are never shed.
    queue_max_depth: 1000
    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
    # Capacity eviction defaults to the historical confidence-only ranking.
    # Opt in to hybrid-v1 to score confidence (65%), explicit-confirmation
    # freshness (25%, 90-day half-life), and query-driven access heat (10%,
    # 30-day half-life). Actual memory_search hits count; default injection does
    # not. A bounded 10% correction reserve aligns storage with guaranteed
    # correction injection. Confirmation detection is batch-level: it checks
    # human messages among the last six filtered messages, while the LLM binds
    # that signal to a factsToReinforce ID without a second correspondence
    # check. Shadow mode audits hybrid disagreements without changing which
    # facts confidence-only keeps.
    fact_eviction_policy: confidence # confidence | hybrid-v1
    fact_eviction_shadow_enabled: false
    eviction_confidence_weight: 0.65
    eviction_confirmation_weight: 0.25
    eviction_access_weight: 0.10
    eviction_confirmation_half_life_days: 90
    eviction_access_half_life_days: 30
    eviction_correction_reserved_fraction: 0.10
    eviction_correction_reserved_max: 10
    eviction_audit_max_entries: 200 # metadata-only events per user/agent scope; 0 disables
    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)
    # extraction_callback is a host-injected post-extraction observability hook
    # (token usage, facts accepted/rejected, rejection rate). The factory
    # injects a logging default; set programmatically to emit a Langfuse span.
    # Message processing (externalized patterns / prompt templates):
    #   patterns_dir        - dir with correction.yaml / reinforcement.yaml overriding
    #                         the bundled signal-detection patterns. None (default) =
    #                         bundled core/message_patterns/. When explicitly set, both
    #                         files must exist (typos or missing mounts raise an error).
    #   prompts_dir         - dir with custom memory-extraction prompt templates
    #                         (memory_update.chat.yaml, staleness_review.yaml,
    #                         consolidation.yaml, fact_extraction.yaml). None (default)
    #                         = bundled core/prompts/. Supports per-agent subdirectories.
    # patterns_dir: ""            # empty = bundled defaults
    # prompts_dir: ""             # empty = bundled defaults

# ============================================================================
# Blob Storage Configuration
# ============================================================================
# Content-addressed blob store: content is addressed by its sha256, so writes
# are idempotent, identical bytes dedupe to one object, and reads verify the
# digest. Mirrors the memory mechanism: a `backend` name matches a
# storage/backends/<name>/ folder that exposes STORE_CLASS (e.g. local_fs), or
# takes a dotted import path to a BlobStore subclass. The factory resolves the
# backend at get_blob_store() time and fails fast on an unresolvable value --
# blobs are persistent state, so a broken backend name must not silently swap
# the storage location.
#
# Off by default and purely opt-in: no producer is migrated yet, so a config
# that never sets this key behaves exactly as before. Contract, the two named
# producers, and the migration order: backend/docs/blob-storage.md.
blob_storage:
  enabled: false               # master switch (inert until the producer follow-ups land)
  backend: local_fs            # folder name under storage/backends/, or a dotted import path
  # Backend-private config (dict), passed verbatim to the backend; each backend
  # self-interprets it. local_fs reads `root` -- empty defaults to
  # {runtime_home}/blobs; point it at a shared volume (NFS / EFS / a
  # ReadWriteMany PVC) for multi-instance deployments:
  #
  # backend_config:
  #   root: /mnt/shared/deerflow-blobs
  backend_config: {}

# ============================================================================
# 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)
  security_fail_closed: true # Moderation model unavailable: true blocks all writes; false allows non-executable content with a warning (executable is always blocked)

# ============================================================================
# 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
#   # A filesystem path or ':memory:'; SQLite `file:` URIs are rejected.
#   connection_string: checkpoints.db
#
# checkpointer:
#   type: postgres
#   connection_string: postgresql://user:password@localhost:5432/deerflow
#   # Optional: place LangGraph checkpointer/store tables in this schema.
#   # Leave empty to use the server default search_path, usually public.
#   postgres_schema: 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
#   # Optional: place DeerFlow ORM tables and LangGraph checkpointer/store
#   # tables in this schema. Leave empty to use the server default
#   # search_path, usually public.
#   postgres_schema: deerflow
#
# postgres_schema only takes effect for new tables. On startup DeerFlow runs
# CREATE SCHEMA IF NOT EXISTS and pins the connection search_path; it never
# moves existing tables, so third-party tables in `public` are left untouched.
#
# Migrating an EXISTING deployment from `public` to a dedicated schema:
#   1. Stop the DeerFlow services.
#   2. Move EVERY DeerFlow-owned table AND the Alembic state into the new schema
#      (run as the DB owner). Missing any table strands its rows: after restart
#      DeerFlow recreates an empty counterpart in the target schema and the old
#      rows stay invisible in `public`. Moving the ORM tables but leaving
#      `alembic_version` behind is especially dangerous -- bootstrap then treats
#      the target schema as unversioned, re-baselines it, and replays migrations.
#        CREATE SCHEMA IF NOT EXISTS deerflow;
#        -- Application ORM tables:
#        ALTER TABLE public.runs                  SET SCHEMA deerflow;
#        ALTER TABLE public.run_events            SET SCHEMA deerflow;
#        ALTER TABLE public.threads_meta          SET SCHEMA deerflow;
#        ALTER TABLE public.feedback              SET SCHEMA deerflow;
#        ALTER TABLE public.users                 SET SCHEMA deerflow;
#        ALTER TABLE public.agents                SET SCHEMA deerflow;
#        -- IM channel tables:
#        ALTER TABLE public.channel_connections   SET SCHEMA deerflow;
#        ALTER TABLE public.channel_credentials   SET SCHEMA deerflow;
#        ALTER TABLE public.channel_oauth_states  SET SCHEMA deerflow;
#        ALTER TABLE public.channel_conversations SET SCHEMA deerflow;
#        -- Scheduled-task tables:
#        ALTER TABLE public.scheduled_tasks       SET SCHEMA deerflow;
#        ALTER TABLE public.scheduled_task_runs   SET SCHEMA deerflow;
#        -- Alembic migration state (REQUIRED -- see note above):
#        ALTER TABLE public.alembic_version       SET SCHEMA deerflow;
#      The DeerFlow-owned set grows over time; discover any tables the list
#      above misses (as the DeerFlow DB owner/role) with:
#        SELECT table_name FROM information_schema.tables WHERE table_schema='public';
#      LangGraph checkpointer/store table names also vary by version -- list them
#      the same way, then ALTER ... SET SCHEMA deerflow for each (e.g. checkpoints,
#      checkpoint_blobs, checkpoint_writes, checkpoint_migrations, store,
#      store_migrations).
#   3. Set postgres_schema: deerflow here and restart.
#   4. Verify with `SHOW search_path;` and a smoke test (create a run, read
#      its history). The `public` schema should gain no new DeerFlow tables.
database:
  backend: sqlite
  sqlite_dir: .deer-flow/data

# ============================================================================
# Inbound webhook dedupe storage (cross-pod redelivery dedup — issue #4120)
# ============================================================================
# Where the ChannelManager records inbound webhook dedupe state.
#
# backend: auto     -- (default) Postgres whenever database.backend=postgres,
#                       including multi-replica K8s where each pod runs a single
#                       worker but shares one DB. Otherwise an in-process memory
#                       store (single-DB, single-pod).
#                       NOTE: multi-replica deployments must use a Postgres DB;
#                       if database.backend is sqlite/memory, 'auto' falls back to
#                       the per-pod memory store and cross-pod dedupe is disabled.
# backend: memory   -- In-process store only. Per-pod: a webhook redelivered to
#                       a DIFFERENT replica is NOT deduped. Not recommended for
#                       multi-worker deployments (logs a startup WARNING).
# backend: postgres -- Share dedupe state across pods via the application DB.
#                       REQUIRED for any multi-replica deployment. If
#                       database.backend is not 'postgres', a WARNING is logged and
#                       the store falls back to the in-process memory store.
#
# dedupe_storage:
#   backend: auto

  # Recycle app ORM PostgreSQL connections before the environment's idle cutoff.
  pool_recycle: 300
  # App ORM PostgreSQL command timeout. Set to null to disable it or raise it
  # for intentionally long commands.
  command_timeout: 30
  # Restart required. Use one value across every process sharing this database.
  # full: current full-message checkpoints; delta: DeltaChannel for messages.
  checkpoint_channel_mode: full
  # Delta-mode tuning (only applies when checkpoint_channel_mode is delta).
  # Restart required, like the mode itself: the cadence is compiled into each
  # graph's channel table. All processes sharing this database must agree.
  # snapshot_frequency: full messages snapshot every N per-step writes
  # (higher = smaller checkpoints, slower materialization).
  checkpoint_delta:
    snapshot_frequency: 10
  # Size caps for the compiled checkpoint graph caches (per process).
  # Hot-reloadable, no restart needed: the value is re-read on each eviction
  # check and only changes when a cache evicts, never graph semantics.
  checkpoint_graph_cache:
    # Gateway thread-state accessor graphs (per assistant/mode/cadence).
    accessor_graph_max: 64
  # Delta-mode checkpoint history cache (only used when checkpoint_channel_mode: delta).
  # Pure performance policy: safe to differ across workers, never frozen.
  # checkpoint_cache:
  #   type: memory            # memory | redis (redis is Gateway/async only)
  #   max_entries: 128        # 0 disables the cache
  #   redis_url: null         # or DEER_FLOW_CHECKPOINT_CACHE_REDIS_URL / REDIS_URL
  #   ttl_seconds: 86400      # redis leak safety net; bounds residual copies if a
  #                           # thread-delete purge fails (purge itself is immediate).
  #                           # 0 explicitly disables expiry (redis maxmemory only)
  #   key_prefix: ""          # default: hash of the database identity

# ============================================================================
# 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

# ============================================================================
# Agent Storage Configuration
# ============================================================================
# Where custom agent DEFINITIONS (config.yaml + SOUL.md) and deployment-level
# managed subagent definitions are stored. This is separate from `database`
# (run/thread/event data) and from agent memory.
# Restart-required (the backend is captured at Gateway lifespan startup).
#
# backend: file  -- Custom Agents use per-user files; managed subagents use one
#                   JSON file each under {base_dir}/managed-subagents/ (default).
#                   Node-local without a shared mount.
# backend: db    -- Both definition types use the shared SQL persistence layer,
#                   so every node sees them. Requires database.backend to be
#                   'sqlite' or 'postgres' (rejected at startup on 'memory').
#
# Switching an existing install to 'db'? Import the on-disk Custom Agents and
# managed subagents once with:
#     python backend/scripts/migrate_agents_to_db.py   # --dry-run to preview
# The source files are left in place, so reverting to 'file' is a clean rollback.
#
# agent_storage:
#   backend: file
agent_storage:
  backend: file

# ============================================================================
# Scheduled Tasks Configuration
# ============================================================================
# Background scheduler for one-time, cron, and interval agent runs.
# Poller fields (enabled, multi_instance, poll_interval_seconds, lease_seconds,
# max_concurrent_runs, max_concurrent_runs_per_user, queue_timeout_seconds,
# min_once_delay_seconds) are restart-required.
# recursion_limit is read at dispatch and applies to the next scheduled run
# without a Gateway restart. The default (1000) matches the web UI's
# interactive budget so scheduled and interactive runs behave identically out
# of the box. Values above max_recursion_limit are clamped.
#
# Execution slots are shared fairly between task owners. Each owner may hold at
# most max_concurrent_runs_per_user launching/running scheduled runs (manual
# "run now" included); the default 2 leaves one of the 3 default global slots
# for everyone else. The waiting queue is drained owner by owner, so one
# owner's backlog never pushes another owner's run out of the batch, even with
# the per-owner cap off (0). A run that waits longer than queue_timeout_seconds
# is skipped; its run history reads "Skipped: it waited too long for a free
# slot" and no chat or IM notice is sent, so keep queue_timeout_seconds well
# above an owner's typical run length.
#
# The scheduler is single-instance by default. Set multi_instance: true only
# when every Gateway instance shares Postgres, run ownership heartbeats, and
# database-backed run events; startup recovery then preserves live peer runs.
#
# When channel_connections.enabled is true and IM channels are running, the
# same poll_interval_seconds also drives the notification delivery worker that
# pushes scheduled-task updates to the owner's bound IM identities (only WeCom
# supports proactive push today; other apps get no notices). Each occurrence
# sends at most one message: a finished run, a failed run, a missed goal, the
# automatic pause after three misses, a pause by the agent (stop condition met)
# or the task finishing (all max_runs done, end_at reached). Notices are queued
# in the same transaction that records the outcome, so runs finalized by crash
# or lease recovery notify once. Plain manual "run now" trials and interrupted
# runs stay silent. The text is in the owner's web UI language (else
# channel_connections.notification_locale) and has no IDs and no links.
#
# scheduler:
#   enabled: false               # Master switch for the background poller
#   multi_instance: false        # Opt into lease-aware recovery across Gateway instances
#   poll_interval_seconds: 5     # How often to scan for due tasks (+ IM outbox delivery)
#   lease_seconds: 120           # Claim lease; a crashed process's task becomes reclaimable after this
#   max_concurrent_runs: 3       # Global cap on launching/running scheduled runs across multi-instance Pods
#   max_concurrent_runs_per_user: 2  # Per-owner cap (0 = off); never more than max_concurrent_runs
#   queue_timeout_seconds: 3600  # Maximum durable queue wait before an occurrence fails
#   min_once_delay_seconds: 60   # Floor for one-time run_at offset and interval every_seconds
#   recursion_limit: 1000        # LangGraph super-step cap for scheduled runs (matches the web UI)
scheduler:
  enabled: false
  tool_enabled: false          # Offer schedule tools to authorized Gateway conversations
  multi_instance: false       # Opt into lease-aware recovery across Gateway instances
  poll_interval_seconds: 5
  lease_seconds: 120
  max_concurrent_runs: 3
  max_concurrent_runs_per_user: 2
  queue_timeout_seconds: 3600
  min_once_delay_seconds: 60
  recursion_limit: 1000

# ============================================================================
# Long-running MCP Tasks Configuration
# ============================================================================
# Durable runtime for ordinary MCP submit/status/cancel task toolsets. It is
# disabled by default; task_toolsets are configured per server in
# extensions_config.json.
# All fields are restart-required (captured at Gateway lifespan startup).
#
# mcp_tasks:
#   enabled: false               # Master switch for the background status poller
#   poll_interval_seconds: 5     # Scan interval and default task retry interval
#   lease_seconds: 120           # Expired claims become recoverable after this delay
#   max_concurrent_polls: 8      # Maximum status calls started by one worker per scan
#   max_poll_backoff_seconds: 300             # Cap for exponential retries after transient errors
#   input_required_poll_interval_seconds: 60  # Minimum poll interval while waiting for user input
#   tracking_degraded_after_errors: 3         # Consecutive errors before query API reports degraded tracking
#   max_result_bytes: 65536                    # Full JSON result storage limit
#   result_preview_max_chars: 2000             # Text preview retained when a result exceeds the limit
mcp_tasks:
  enabled: false
  poll_interval_seconds: 5
  lease_seconds: 120
  max_concurrent_polls: 8
  max_poll_backoff_seconds: 300
  input_required_poll_interval_seconds: 60
  tracking_degraded_after_errors: 3
  max_result_bytes: 65536
  result_preview_max_chars: 2000

# ============================================================================
# Deployment Topology
# ============================================================================
# Declare whether more than one Gateway instance shares this database.
#
# GATEWAY_WORKERS / WEB_CONCURRENCY only count the uvicorn workers of one
# process tree, so a Kubernetes Deployment with replicas > 1 (one worker per
# Pod) is invisible to the startup safety gate. Without this declaration each
# Pod passes every check, and its startup orphan reconciliation writes the
# other Pods' lease-less runs off as crashed on every rolling update.
#
# With multi_instance: true (or the environment variable
# DEER_FLOW_MULTI_INSTANCE=1, which deploy tooling such as a Helm chart can
# set from its replica count) startup requires database.backend: postgres, run_events.backend: db,
# run_ownership.heartbeat_enabled: true and a Redis stream bridge, and refuses
# an explicit sandbox.ownership.type: memory, process-local browser tools, and
# scheduler.enabled: true without scheduler.multi_instance: true.
# Restart required; apply the same value to every instance.
deployment:
  multi_instance: false

# ============================================================================
# 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).
#
# The owning worker itself does NOT use grace_seconds as extra execution time.
# If renewal cannot be confirmed before lease_seconds expires, it cancels the
# local run and suppresses durable finalization; grace_seconds only delays when
# a peer may reclaim the expired row.

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           # data events retained per run (min: 1). A reconnect
#                                # older than this window receives SSE `gap`.
#   heartbeat_interval_seconds: 15  # idle interval (max: 86400) between
#                                   # heartbeats sent to SSE, wait, and internal
#                                   # stream consumers.
#
# stream_bridge:
#   type: redis                  # recommended for Docker / multi-worker gateway
#   redis_url: redis://redis:6379/0
#   queue_maxsize: 256           # data events retained per run (Redis MAXLEN, min: 1).
#                                # Older valid cursors receive SSE `gap`.
#   heartbeat_interval_seconds: 15  # idle interval (max: 86400) between
#                                   # heartbeats sent to SSE, wait, and internal
#                                   # stream consumers.
#   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: 86400).
#   max_connections: 100         # optional pool ceiling. Each live SSE client
#                                # holds one connection blocked in XREAD ... BLOCK
#                                # for up to the configured heartbeat interval, 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, WeCom, and QQ use `/connect <code>`
#   through the already-running bot/app.
# - When enabled (and channels are running), scheduled-task updates (runs,
#   pauses, finishes) are pushed to the owner's connected identities on apps
#   with proactive push, WeCom today (see scheduler section above). Settings
#   shows per app whether updates are sent there. Plain manual trials and
#   interrupted runs do not notify. Notices carry no links.
#
# 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
#   # Language of scheduled-task IM notices (en-US or zh-CN) for owners without
#   # a web UI language preference (e.g. auth disabled). Restart-required.
#   notification_locale: en-US
#
#   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
#
#   buzz:
#     enabled: false
#
#   qq:
#     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
#   # Maximum queued or provider-reserved inbound messages. Must be a positive integer.
#   inbound_queue_maxsize: 1000
#   # Fixed number of long-lived inbound handler workers. Must be a positive integer.
#   max_concurrency: 5
#   # Seconds to drain accepted inbound work before cancelling active handlers.
#   # Must be a non-negative finite number. Cancelled handlers are awaited; the Gateway's
#   # outer shutdown timeout remains the process-level bound for incomplete cleanup.
#   shutdown_grace_period_seconds: 3
#   #
#   # 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
#
#   # QQ Open Platform bot; choose WebSocket in the developer console.
#   # Text-only C2C and group @mentions; no public callback URL required.
#   qq:
#     enabled: false
#     app_id: $QQ_APP_ID
#     client_secret: $QQ_CLIENT_SECRET
#     allowed_users: []  # QQ OpenIDs, not numeric QQ account numbers; empty allows all
#     sandbox: false    # REST sandbox host only; console sandbox membership is separate
#
#   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: []               # numeric user IDs, not @usernames; empty = allow all
#     rich_messages: false             # Bot API 10.1 Rich Messages for final Markdown responses
#
#   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: []               # iLink user IDs; empty allows all; a single ID is one entry; unparseable non-empty lists fail closed
#     # Optional: extra host suffixes inbound media downloads may come from,
#     # in addition to the platform CDN defaults (*.qq.com and the cdn_base_url host).
#     # Entries are suffixes: "example.com" and "*.example.com" are equivalent.
#     allowed_media_hosts: []
#     # Optional: timing values must be positive finite seconds
#     polling_timeout: 35
#     polling_retry_delay: 5
#     # QR poll interval when qrcode_login_enabled is true
#     qrcode_poll_interval: 2
#     # QR bootstrap timeout
#     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
#     # Optional: extra host suffixes inbound media downloads may come from, in
#     # addition to the built-in qq.com family and WeCom's official COS media
#     # host (ww-aibot-img-1258476243.<region>.myqcloud.com). Add a suffix here
#     # if WeCom rotates to a new COS account or media is routed through a proxy.
#     allowed_media_hosts: []
#
#   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 joins as a bot; requires the `discord` dependency extra
#   # (uv sync --extra discord) for discord.py.
#   discord:
#     enabled: false
#     bot_token: $DISCORD_BOT_TOKEN
#     # Allowed guild IDs: one positive numeric ID (quoted or unquoted) or a YAML list.
#     # Unset, null, [], or a blank string allows all guilds.
#     # Invalid entries are ignored with a warning. Any other configured value
#     # with no valid ID denies every guild and logs an error.
#     allowed_guilds: []
#     mention_only: false             # If true, only respond when the bot is mentioned
#     # Channel IDs exempt from mention_only: one ID (quoted or unquoted) or a YAML list.
#     # Empty = no exemptions; mention_only applies to every channel when enabled.
#     allowed_channels: []
#     thread_mode: false               # If true, group a channel conversation into a thread
#
#   # Buzz (https://github.com/block/buzz) — Nostr-relay team workspace.
#   # DeerFlow joins as a member identity; generate a keypair and register it as
#   # a relay member, then @mention or DM DeerFlow in Buzz. Requires the `buzz`
#   # dependency extra (uv sync --extra buzz).
#   buzz:
#     enabled: false
#     relay_url: wss://buzz.example.com
#     private_key: $BUZZ_PRIVATE_KEY        # hex or nsec1…
#     allowed_users: []                     # pubkeys (hex or npub) allowed to trigger runs.
#                                           # DENY-BY-DEFAULT (unlike other channels): empty
#                                           # means nobody, and DeerFlow logs a startup warning.
#     require_mention: true                 # @mention needed in channels
#     mention_free_channels: []             # channel UUIDs that respond to every message

# ============================================================================
# Guardrails Configuration
# ============================================================================
# Optional pre-execution authorization for tool calls.
# When enabled, every tool call passes through the configured provider
# before execution. Four options: built-in allowlist, OAP policy provider,
# custom provider, or the TypeSafe (Jev) risk gate. 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

# --- Option 4: TypeSafe (Jev) pre-execution risk gate (network provider) ---
#   Sends the tool name and its full argument JSON to POST /v1/systemone and
#   denies calls whose risk probability reaches `threshold`. Requires an API
#   key: export TYPESAFE_API_KEY=... (or set `api_key` below).
#
#   `tools: ["bash"]` is a SMOKE-TEST scope, not a production recommendation:
#   the operation can move to an unprobed tool, an MCP tool or a subagent, and
#   long heredocs / large write_file payloads are denied locally once the
#   argument text exceeds max_state_chars. Probe every tool that can perform the
#   risky work instead of narrowing the gate. Pointing `guardrails.provider` at
#   TypeSafe REPLACES the allowlist/OAP provider: the allow half of an old
#   allowlist moves into `allowed_tools` below, while deny rules still belong to
#   `authorization.*` (see backend/docs/GUARDRAILS.md, Option 4).
# guardrails:
#   enabled: true
#   fail_closed: true
#   provider:
#     use: deerflow.guardrails.typesafe:TypeSafeGuardrailProvider
#     config:
#       model: jev-latest           # pin an exact version (e.g. jev-1.13.0) when
#                                   # decisions must stay traceable across releases
#       threshold: 0.5              # example value; calibrate with the online evaluation
#       tools: ["bash"]
#       # allowed_tools: ["bash", "read_file", "write_file"]   # hard permission list:
#       #   unlisted tools are refused locally before any state is built, and
#       #   listed tools still have to pass the risk gate. Omit to enforce none;
#       #   [] refuses every tool.
#       # Optional: api_key / api_key_env (default TYPESAFE_API_KEY) / base_url /
#       #           timeout / deadline_seconds / max_attempts / retry_backoff /
#       #           max_state_chars / cache_size / cache_ttl_seconds /
#       #           instructions / criteria

# ============================================================================
# TypeSafe (Jev) Client Defaults (optional)
# ============================================================================
# One block for every TypeSafe consumer: the guardrail risk gate
# (guardrails.provider.config) and, once enabled, memory pre-screening and
# signal classification. Each consumer overrides what it needs in its own
# `config`; precedence is consumer config > this block > built-in defaults.
#
# Not here, on purpose: failure policy (the gate denies on error, the memory
# paths carry on), cache settings, thresholds and question content. Those are
# the consumer's own; two consumers with different thresholds can still share
# one request, and each publishes its own policy identity.
#
# Credentials: prefer api_key_env. `api_key` here puts the key in this file.
# A consumer's `config` and this block are separate layers: the first layer that
# sets `api_key` or `api_key_env` decides the credential, so a consumer with its
# own `api_key_env` keeps reading that variable. Inside one layer a literal
# `api_key` wins over that layer's `api_key_env`.
#
# These values are checked when a consumer is built, not on the first call: a
# `base_url` that is not http(s) or that carries a query / fragment / embedded
# credentials, a key with surrounding whitespace (a secret mounted from a file),
# and a non-printable key character are all configuration errors. Numbers here are
# validated strictly, so write `max_attempts: 2` — `true` is rejected rather than
# silently read as `1`.
#
# typesafe:
#   api_key_env: TYPESAFE_API_KEY        # default; the variable holding the key
#   # api_key: <explicit key>            # alternative to api_key_env
#   base_url: https://api.typesafe.ai    # default; http(s) only, no query/fragment/credentials
#   model: jev-latest                    # default; pin an exact version for traceability
#   timeout: 5.0                         # per-request timeout (seconds)
#   deadline_seconds: 10.0               # whole-evaluation budget (seconds)
#   max_attempts: 2                      # includes the first; 429/529 and transport errors only
#   retry_backoff: 0.5                   # base seconds of exponential backoff

# ============================================================================
# Authorization Configuration
# ============================================================================
# Fine-grained resource authorization (RBAC and beyond). Disabled by default;
# every authenticated user has access to all resources.
# See RFC: https://github.com/bytedance/deer-flow/issues/4063
#
# authorization:
#   enabled: true
#   fail_closed: true            # block on provider error / unresolved identity
#   default_role: user           # applied when user_role is None; built-in RBAC requires this role below
#   provider:
#     use: deerflow.authz.rbac:RbacAuthorizationProvider
#     config:
#       # A known role with no policy for a resource is unrestricted for it.
#       # Define both `tools` and `routes` wherever access should be constrained.
#       # `models` / `skills` govern the per-caller listing/detail surfaces
#       # (which models/skills the caller sees on GET /api/models and the
#       # GET /api/skills[.../custom, /{name}] skill routes); skill
#       # management endpoints (install/edit/export) stay admin-gated.
#       #
#       # Plugin resources: `plugin_actions` gates invoking a registered
#       # backend action (POST /api/plugins/{ns}/actions/{name}), and
#       # `plugin_management` gates an enterprise-contributed management route
#       # that calls require_plugin_management. Targets are
#       # "<namespace>/<action-name>" and "<namespace>/permissions.read"
#       # / "<namespace>/permissions.write". **Omitting a key leaves that
#       # resource unrestricted for the role**, so list it explicitly to
#       # constrain it. Declared plugin pages are not gated yet: their
#       # `plugin_pages` row ships with the page-authorization slice.
#       roles:
#         admin:
#           tools: {allow: "*"}
#           routes: {allow: "*"}
#           models: {allow: "*"}
#           sandbox: {allow: "*"}
#           skills: {allow: "*"}
#           plugin_actions: {allow: "*"}
#           plugin_management: {allow: "*"}
#         user:
#           tools: {allow: "*", deny: ["update_agent"]}
#           routes: {allow: "*"}
#           models: {allow: "*"}
#           sandbox: {allow: "*"}
#           skills: {allow: "*"}
#           plugin_actions: {allow: "*"}                 # invoke any installed action
#           plugin_management: {allow: []}               # no management route access
#         guest:
#           # web-only role: sandbox-dependent tools (read_file, bash, glob,
#           # grep, write_file, ...) are omitted — with sandbox:execute denied
#           # they could only ever return the deny error, so allow just the
#           # non-sandbox tools.
#           tools: {allow: ["web_search"]}
#           routes: {allow: ["threads:read", "runs:read"]}
#           models: {allow: ["gpt-4o-mini"]}
#           sandbox: {allow: false}  # deny sandbox execution
#           skills: {allow: []}      # no skills visible in the listing
#           plugin_actions: {allow: ["acme.reports/export"]}
#           plugin_management: {allow: ["acme.reports/permissions.read"]}
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

# ============================================================================
# LLM Call Concurrency Configuration
# ============================================================================
# Cap the number of concurrently in-flight LLM calls process-wide. A provider
# burst-rate (limit_burst_rate) error fires on the *slope* of the request rate,
# not on a static quota - so the morning peak (e.g. 08:30) ramping from ~0 to
# full throttle in seconds gets rejected even when total RPM is within budget.
# Capping concurrency caps that slope. Retries alone make it worse (they add
# demand inside the very burst being throttled); pair this cap with the
# decorrelated-jitter backoff already built into the LLM error-handling
# middleware, and ideally an nginx `limit_req` at the ingress.
#
# 0 disables the cap (default) - existing deployments see no behavior change.
#
# Per-process, not per-cluster: the cap bounds in-flight LLM calls within ONE
# gateway process. With GATEWAY_WORKERS > 1 the aggregate cap across the
# deployment is effectively `max_concurrent_calls * GATEWAY_WORKERS`, and a
# multi-node rollout multiplies it further - size the per-process value with
# that in mind (and pair it with an nginx `limit_req` at the ingress for a
# true cluster-wide slope cap).
#
# Startup-only: the cap is captured at the first LLM run and frozen for the
# process lifetime. Editing `max_concurrent_calls` here takes effect only after
# a gateway restart; the other `llm_call.*` knobs below remain hot-reloadable.
# Freezing the cap avoids the downscale and config-freshness races that a
# runtime-mutable, process-wide/cross-loop limiter would otherwise hit (a
# lowered cap could keep admitting queued waiters, or a stale config snapshot
# constructed after a fresher one could restore a higher cap).

# llm_call:
#   # Max concurrently in-flight LLM calls across the whole process (default: 0 = disabled)
#   max_concurrent_calls: 0
#   # Max LLM call attempts for retriable transient errors, 1 = no retry (default: 3)
#   retry_max_attempts: 3
#   # Base delay (ms) for the decorrelated-jitter retry backoff (default: 1000)
#   retry_base_delay_ms: 1000
#   # Hard cap (ms) on any single retry backoff delay (default: 8000)
#   retry_cap_delay_ms: 8000
#   # Backoff base (ms) used ONLY for burst-rate (limit_burst_rate) 429s - higher
#   # than retry_base_delay_ms so the single burst retry lands after the throttle
#   # window subsides. Ignored when the provider sends Retry-After (default: 5000)
#   burst_retry_base_delay_ms: 5000

# ============================================================================
# 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)

# ============================================================================
# Local (email/password) authentication
# ============================================================================
# Self-registration via POST /api/v1/auth/register is open by default: anyone
# who can reach the Gateway may create a regular user account.
#
# The OIDC provisioning policy below (allowed_email_domains, require_verified_email,
# auto_create_users) is enforced only in the SSO callback. It does NOT constrain local
# registration, so on an SSO-provisioned deployment leaving this open lets a visitor
# create an account outside that policy. Set allow_registration: false there.
#
# The first admin account is always created through /initialize regardless of this
# setting, so turning it off cannot lock you out of a fresh install.

# auth:
#   local:
#     allow_registration: false
#     # Login throttling per client IP (in-process, per worker). Raise
#     # max_login_attempts when many users share one egress IP (corporate
#     # proxy / NAT) so shared-IP lockouts don't block the whole office;
#     # lower it for a stricter posture (minimum 2 — one failed attempt must
#     # never lock an IP, or a single typo would block the shared egress).
#     # Defaults preserve the historical hardcoded policy (5 failures /
#     # 5 minutes). Live-read per login, so a config reload applies without
#     # a Gateway restart: lowering lockout_seconds releases an active lock
#     # early and tightening max_login_attempts keeps already-counted
#     # failures; a raised lockout_seconds extends a still-active lock but
#     # never resurrects one that already expired.
#     # max_login_attempts: 5
#     # lockout_seconds: 300
#
#   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

# --- Plugins ----------------------------------------------------------------
# Extension packages, loaded once at startup in the order listed here. Each
# entry names an install entry point as "module.path:install"; the `config`
# block is handed to that package verbatim and validated by the package itself.
#
# Prefer the extension manager over editing this list by hand:
#   make extension-install SOURCE="deerflow-extension-acme==1.2.3"
#   make extension-install SOURCE="git+https://github.com/acme/deerflow-extension-acme.git@<commit>"
#   make extension-install SOURCE="/absolute/path/to/local-extension"
#   make extension-list
#   make extension-enable NAME=acme
#   make extension-disable NAME=acme
#   make extension-remove NAME=acme
# The installer updates backend/pyproject.toml, backend/uv.lock, and this block
# together. It also asks the operator to confirm that the source is trusted;
# extensions and their build systems execute with Gateway privileges.
#
# Every install, enable, disable, remove, or manual edit requires a Gateway
# restart. Development launchers may fetch missing locked artifacts before
# handoff; a built production Gateway never installs plugins during startup.
# Local directories are copied to backend/extensions/sources as deployable
# snapshots; use a pinned version or public HTTPS Git commit for reproducible
# remote sources. SSH Git URLs are rejected because the stock Docker builder
# does not forward host SSH credentials.
# Loading remains explicit: a package installed outside the manager does nothing
# until it is listed here, and list order fixes contribution ordering.
#
# This is deliberately separate from the `extensions:` block above (MCP servers,
# skills, config-declared middlewares). That one is backed by
# extensions_config.json, which the Gateway can rewrite through an HTTP
# endpoint; a list that causes code to be imported must stay in this
# operator-controlled file only.
#
# Plugins can contribute semantically placed middleware, task-lifecycle hooks,
# system-model observers, Gateway-lifetime services, and eager FastAPI HTTP routers.
# See examples/deerflow-extension-example for a package that exercises all five.
#
# plugins:
#   - name: example                       # PEP 621 entry-point name
#     package: deerflow-extension-example # Python distribution managed by uv
#     use: deerflow_extension_example:install
#     enabled: true                       # false skips import and registration
#     required: false                     # true makes load failure abort Gateway startup
#                                         # (install --required opts in)
#     host_access:                        # optional host-owned grants (restart required)
#       model_invocation:
#         roles:                          # logical role -> configured models[].name
#           default: my-configured-model
#         max_concurrency: 2              # shared by this installation's services
#         timeout_seconds: 60             # includes waiting for capacity
#         max_input_chars: 262144         # includes the response-schema instruction
#         max_output_chars: 65536         # returned text, checked after generation
#     table_prefix: example_ext_          # optional: table-name prefix this extension
#                                         # owns under its own MetaData/migration chain.
#                                         # Keeps `alembic revision --autogenerate`, run
#                                         # directly against a live database, from seeing
#                                         # those tables and proposing to drop them.
#                                         # (`make migrate-rev` diffs a throwaway SQLite
#                                         # built from the migration chain, so it never
#                                         # sees them in the first place.)
#                                         # Rejected at startup if empty, or if it would
#                                         # also match a host-owned table name.
#     config:
#       label: example                    # extension-private values, if any