config.yaml.example

models:
  # Optional: choose the active main/sub profile explicitly.
  # If omitted, the first profile is used for main and sub defaults to main.
  # active_main: "gpt-4.1"
  # active_sub: "gpt-4.1-mini"
  profiles:
    gpt-4.1:
      model: "gpt-4.1"
      support_modal: [text, image] # Accepted inputs; defaults to [text].
      api_key: "your-api-key-here"
      provider: "openai-compatible"
      base_url: "https://api.openai.com/v1"
      max_tokens: 4096
      temperature: 0.0
      max_context_tokens: 128000
      # Optional automatic context-policy overrides; omitted fields inherit
      # the global context section below.
      # context:
      #   auto_snip: false
      # Keep provider-returned reasoning_content in history for later rounds.
      preserve_reasoning_content: true
      # For stricter gateways, inject an empty reasoning_content on assistant tool-call
      # messages if a previous round/history entry is missing it.
      backfill_reasoning_content_for_tool_calls: false

    gpt-4.1-mini:
      model: "gpt-4.1-mini"
      support_modal: [text, image]
      api_key: "your-api-key-here"
      provider: "openai-compatible"
      base_url: "https://api.openai.com/v1"
      max_tokens: 4096
      temperature: 0.0
      max_context_tokens: 128000
      preserve_reasoning_content: true
      backfill_reasoning_content_for_tool_calls: false

    # Responses API keeps the canonical conversation locally and sends a
    # stateless request on every round. Omitting request_mode keeps the legacy
    # chat-completions default used by the profiles above.
    gpt-5.6-luna-responses:
      model: "gpt-5.6-luna"
      api_key: "your-api-key-here"
      provider: "openai-compatible"
      request_mode: "responses"
      base_url: "https://api.openai.com/v1"
      max_tokens: 8192
      temperature: 0.0
      max_context_tokens: 128000
      responses:
        state: "local"
        cache:
          # Implicit is the default; use explicit only on supporting models.
          mode: "implicit"

    # Example DeepSeek reasoning profile for tool-calling workloads.
    deepseek-v4-pro:
      model: "deepseek-v4-pro"
      api_key: "your-deepseek-api-key-here"
      provider: "openai-compatible"
      base_url: "https://api.deepseek.com"
      max_tokens: 8192
      temperature: 0.0
      max_context_tokens: 128000
      preserve_reasoning_content: true
      backfill_reasoning_content_for_tool_calls: false
      thinking_enabled: true
      reasoning_effort: "high"
      reasoning_replay_mode: "tool_calls"

modes:
  active: "coder"
  profiles:
    planner:
      description: "Planning-first mode; keep implementation changes minimal unless asked."
      tools: ["read_file", "view_image", "glob", "grep", "history_search", "history_read", "artifact_read", "update_plan", "report_progress"]
      prompt_append: "Focus on architecture, trade-offs, and actionable implementation plans."
      allowed_subagent_modes: ["explore"]

approval:
  default_mode: "require_approval"
  # Route approval prompts to the user (default) or a fail-closed reviewer.
  reviewer: "user" # user | auto_review
  # Required for auto_review; never silently falls back to the work/sub model.
  auto_review_model_profile: null
  auto_review_policy: ""
  auto_review_timeout_seconds: 15
  rules:
    - tool_name: "read_file"
      action: "allow"
    - tool_name: "glob"
      action: "allow"
    - tool_name: "grep"
      action: "allow"
    - tool_name: "list_file"
      action: "allow"
    - tool_name: "history_search"
      action: "allow"
    - tool_name: "history_read"
      action: "allow"
    - tool_name: "artifact_read"
      action: "allow"
    - tool_name: "write_file"
      action: "require_approval"
    - tool_name: "edit_file"
      action: "require_approval"
    - tool_name: "shell"
      action: "require_approval"
    - tool_name: "agent"
      action: "require_approval"
    - tool_source: "mcp"
      action: "require_approval"

# Optional user/workspace-level system prompt injection.
# prompt:
#   system_append: |
#     Always answer in Chinese unless explicitly requested otherwise.
#     Keep responses concise and actionable.

# Optional image budgets. Paste a local image path to attach it; see docs/images.md.
# attachments:
#   image:
#     normal_max_bytes: 262144             # 256 KiB binary per ordinary image
#     max_edge_px: 2000
#     detail_max_base64_bytes: 2097152     # 2 MiB Base64 for detail/ROI
#     originals_cache_max_bytes: 1073741824 # 1 GiB originals per session

# Optional context compression settings. Omit to use defaults.
# context:
#   image_retention: history # history | user_turn; steering/goal continuations keep the turn.
#   # Automatic strategies are independent; manual /compact force remains available.
#   auto_snip: true
#   auto_summarize: true
#   auto_collapse: true
#   # A models.profiles.<name>.context section may override any switch above.
#   # Number of recent tool calls to protect from snipping.
#   snip_keep_recent_tools: 5
#   # Min content length (chars) to trigger tool output snip.
#   snip_threshold_chars: 1500
#   # Min line count to trigger tool output snip.
#   snip_min_lines: 6
#   # Number of recent complete API rounds to protect during LLM summarize.
#   summarize_keep_recent_turns: 5
#   # Reserve output and safety headroom before calculating input thresholds.
#   reserved_output_tokens: 8192
#   safety_margin_tokens: 2048
#   # Optional estimates for stable prompt/tool-schema overhead.
#   fixed_prompt_tokens: 0
#   tool_schema_tokens: 0

# Both web tools use environment proxies when omitted. See docs/web-tools.md.
# web:
#   proxy: env # env | direct | "http://127.0.0.1:7890" | "socks5://127.0.0.1:1080"

# Optional output/session/history settings. Omit to use defaults.
# tool_output:
#   max_chars: 12000
#   max_lines: 120
#   store_full_output: true
#   # store_dir: ".rcoder/tool-outputs"
#
# UI output is independently bounded. The default summary mode uses a tool's
# structured summary, or a head+tail folded preview for legacy/tool text.
# ui:
#   verbosity: "compact"          # compact | standard | debug
#   tool_output: "summary"        # errors | summary | preview | full
#   max_preview_lines: 20
#   max_preview_chars: 1200
#   show_tool_args: true
#   reasoning_display: "indicator" # hidden | indicator | inline
#   notification_threshold: "info" # debug | info | warning | error
#
# session:
#   auto_save: true
#   # dir: ".rcoder/sessions"
#
# Optional remote relay settings.
# Use `rcoder --server` to start host mode; the service will listen exactly on
# the configured relay_bind address.
# remote_exec:
#   enabled: true
#   host_mode: true
#   relay_bind: "127.0.0.1:8765"
#   bootstrap_access_secret: "replace-with-a-long-random-secret"
#   bootstrap_token_ttl_sec: 120
#   peer_token_ttl_sec: 3600
#   heartbeat_interval_sec: 10
#   heartbeat_timeout_sec: 30
#   default_tool_timeout_sec: 30
#   shell_timeout_sec: 120
#
# cli:
#   # history_file: "~/.rcoder/history"

skills:
  enabled: true
  scan_project: true
  scan_user: true
  disabled: []

# Optional MCP server examples.
# Uncomment and adjust paths/tokens for your environment.
# mcp:
#   servers:
#     filesystem:
#       command: "npx"
#       args:
#         - "-y"
#         - "@modelcontextprotocol/server-filesystem"
#         - "/path/to/your/project"
#       env: {}
#       cwd: null
#       enabled: true
#
#     github:
#       command: "npx"
#       args:
#         - "-y"
#         - "@modelcontextprotocol/server-github"
#       env:
#         GITHUB_TOKEN: "your-github-token-here"
#       cwd: null
#       enabled: true
#
#     postgres:
#       command: "uvx"
#       args:
#         - "mcp-server-postgres"
#       env:
#         DATABASE_URL: "postgresql://user:pass@localhost/db"
#       cwd: null
#       enabled: false

goal:
  default_token_budget: null # Cumulative input - cached input + output; null means no limit.

lsp:
  edit_wait_timeout_ms: 1000 # Wait briefly for diagnostics in edit/write results; 0 only checks ready results.