remote-dev-bot.yaml

# Remote Dev Bot Configuration
# Usage: /agent-<mode>[-<model>]
#
#   /agent-resolve   — implement issue, open PR
#   /agent-design    — design analysis, post comment
#   /agent-review    — code review on a PR, post comment
#   /agent-workshop  — design analysis + multi-model council critique
#   /agent-build     — implement issue + multi-model council code review
#
# To use a specific model, append it: /agent-resolve-claude-large
# This works for any mode. See models: below for available aliases.

default_model: claude-small

# Modes define what the agent does. Each mode specifies optional defaults;
# omit a key to inherit from the agent: section below.
#
# extra_files lists differ by mode deliberately:
#   - design/workshop/review: read static context files for analysis
#     (contributing guidelines, agent conventions, etc.)
#   - resolve/build: explore the codebase dynamically via tools and only
#     need light orientation files; extra context would add noise
modes:
  resolve:
    default_model: claude-small
    max_iterations: 50       # Iteration limit for the agent loop
    extra_files:
      - AGENTS.md
      - README.md
    # Optional: extra_instructions appended to the resolve prompt.
    # extra_instructions: |
    #   Always add type annotations to any Python functions you write.

  design:
    default_model: claude-small
    max_iterations: 10       # Iteration limit for the agentic loop
    extra_files:
      - README.md
      - CONTRIBUTING.md
      - AGENTS.md
      - CLAUDE.md
      - .gemini/GEMINI.md
    # Optional: extra_instructions appended to the canonical design prompt.
    # Use this to add project-specific guidance (e.g., preferred patterns,
    # areas to focus on, or terminology). Do not duplicate the base prompt.
    # extra_instructions: |
    #   When analyzing issues, always consider the impact on our deploy pipeline.

  review:
    max_iterations: 10       # Iteration limit for the review loop
    extra_files:
      - AGENTS.md
      - README.md

  workshop:
    default_model: claude-small  # Model for design stage (Stage 1)
    max_iterations: 15       # Iteration limit for the design loop
    extra_files:
      - README.md
      - CONTRIBUTING.md
      - AGENTS.md
      - CLAUDE.md
      - .gemini/GEMINI.md
    # Council: models that review the design. Defaults to all configured models
    # (including the design model) if not specified. Override per-repo to
    # select a different set. The value of the council is diverse perspectives
    # across model families; depth within one family adds less value.
    council:
      - claude-small
      - gpt-small
      - gemini-small

  build:
    default_model: claude-small
    max_iterations: 50       # Iteration limit for the implementation loop
    extra_files:
      - AGENTS.md
      - README.md
    # Council: models that review the PR. Defaults to all configured models if not specified.
    council:
      - claude-small
      - gpt-small
      - gemini-small


models:
  claude-small:
    id: anthropic/claude-sonnet-4-6
    description: "Balanced cost/capability — good default for most tasks"

  claude-large:
    id: anthropic/claude-opus-4-6
    description: "Most capable — for complex multi-file features"

  gpt-small:
    id: openai/gpt-5.1-codex-mini
    description: "OpenAI GPT-5.1 Codex Mini — good default"

  gpt-large:
    id: openai/gpt-5.3-codex
    description: "OpenAI GPT-5.3 Codex — best coding model"

  gemini-small:
    id: gemini/gemini-2.5-flash
    description: "Google Gemini 2.5 Flash — good default"

  gemini-large:
    id: gemini/gemini-2.5-pro
    description: "Google Gemini 2.5 Pro"

# Agent settings — apply to all modes unless overridden per-mode above.
# Note: 'openhands:' is accepted as a backward-compatible alias for 'agent:'
agent:
  # Max agent iterations per run (controls cost ceiling).
  # Override per-invocation with: max_iterations = 25 (inline arg in comment body)
  max_iterations: 50
  # PR type: "draft" or "ready"
  pr_type: ready
  # What to do when the agent can't fully resolve the issue (success=False):
  #   draft   — open a draft PR with partial changes (default)
  #   comment — post a comment only, no PR
  on_failure: draft
  # Assign the triggering user to the issue when the agent starts
  assign_issue: true
  # Assign the triggering user to the resulting PR
  assign_pr: true
  # Watchdog timeout in minutes. The runner kills the agent process after this
  # many minutes, then cleanup steps (cost report) still run.
  # Override per-invocation with: timeout_minutes = 60 (inline arg in comment body)
  # Note: GitHub Actions has a hard 6-hour cap. If you need longer runs, set
  # timeout-minutes in your calling workflow's job definition.
  timeout_minutes: 60
  # Bash output truncation limit in characters. Long outputs are trimmed to
  # first half + last half of this length to prevent context bloat.
  # Set to 0 to disable truncation.
  bash_output_limit: 8000
  # Rolling status log: post a one-sentence status update every N iterations
  # as an issue comment. Set to 0 to disable (default — enable per-repo).
  status_log_interval: 0
  # Number of recent tool call/result pairs to keep in context. Older pairs
  # are replaced with a placeholder to prevent O(N²) token growth on long runs.
  # Set to 0 to keep all tool results (not recommended for runs > ~20 iterations).
  context_keep_tool_results: 10
  # Context window compaction: when the context grows too large, the oldest
  # messages are summarized by the LLM and replaced with a compact summary.
  # Set max_context_tokens to 0 to disable compaction entirely.
  # Compaction triggers at 85% of max_context_tokens (hardcoded).
  # Typical model limits: Claude Sonnet 200k, Gemini 2.5 Flash 1M, GPT-4o 128k.
  max_context_tokens: 0
  # Fraction of conversation history (from the oldest end) to consider for
  # compaction when triggered. 0.5 = compact the oldest 50% of messages.
  compaction_coverage: 0.5
  # Fraction of the selected messages to actually summarize and replace.
  # 0.5 = replace 50% of the selected range, keeping the rest verbatim.
  compaction_factor: 0.5
  # Graceful wrap-up: inject instructions when approaching max iterations.
  # This helps the agent commit partial work before hitting the limit.
  graceful_wrapup:
    # Enable/disable graceful wrap-up instructions
    enabled: true
    # Threshold as fraction of max_iterations (0.8 = 80%)
    # At this point, the agent receives instructions to wrap up
    threshold: 0.8