.coderabbit.yaml

# CodeRabbit configuration.
# CodeRabbit review comments are advisory only — they do not block merges
# and do not replace human review.
# Schema reference: https://docs.coderabbit.ai/reference/yaml-template

language: "en-US"

reviews:
  # assertive: surface nitpicks, edge cases, and style issues, not just
  # high-confidence blockers. chill was too quiet — clean-looking PRs got a
  # bare "no actionable comments" even when there was worth flagging.
  profile: "assertive"
  # Config half of a merge gate: CodeRabbit submits a formal "Request changes"
  # review on actionable findings and only approves once the PR is clean. This
  # becomes actual enforcement only when main gets a branch-protection rule
  # requiring CodeRabbit's approving review (GitHub Pro on a private repo, or a
  # public repo).
  request_changes_workflow: true
  high_level_summary: true
  poem: false
  review_status: true
  collapse_walkthrough: true
  auto_review:
    enabled: true
    drafts: false
    # Review every push to an open PR, not just the first one.
    auto_incremental_review: true
    base_branches:
      - "main"
  # Docs are load-bearing here (DESIGN.md is the spec, ADRs record decisions),
  # so markdown IS reviewed. Only exclude generated/vendored noise.
  path_filters:
    - "!**/testdata/**"
    - "!**/*.lock"
    - "!bin/**"
  path_instructions:
    - path: "**/*.go"
      instructions: >
        Review for idiomatic Go, error handling, and test coverage. Flag
        missing tests for exported functions and any unhandled or swallowed
        errors. This project's core invariant is the NodeRunner seam:
        internal/noderunner (ClaudeCLIRunner) must be the ONLY os/exec touch
        point — flag any new os/exec, exec.Command, or direct process spawning
        elsewhere. Child processes must run on the user's subscription auth:
        flag anything that could leak ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN
        into a child environment, and flag any use of the Agent SDK or a
        `--bare` claude invocation. For auto-planned graphs, the tool-safety
        allowlist gates untrusted LLM output — flag permission_mode:
        bypassPermissions or unbounded allowed_tools in generated nodes.
    - path: "**/*_test.go"
      instructions: >
        Tests must use FakeRunner, never a real claude subprocess, so CI stays
        hermetic and free. Flag tests that shell out to `claude` or depend on
        network/auth. Prefer failure-first cases (invalid input, nonzero exit,
        early return) alongside the happy path.
    - path: "**/*.md"
      instructions: >
        Flag claims that the code contradicts (stale invariants, wrong flags,
        renamed symbols, deferred features described as shipped). Keep the docs
        English-only — except README.ko.md, the deliberate Korean translation
        (see its own instruction below). DESIGN.md is the spec — flag drift
        between it and the implementation.
    - path: "README.ko.md"
      instructions: >
        This file is the repo's one deliberate Korean artifact: the Korean
        translation of README.md (PR #77; CONTRIBUTING's "Sync the Korean
        README" release-checklist bullet). Korean prose, alt text, and
        captions are correct here — never flag them for not being English.
        Do flag content drift against README.md, the English source of truth.
    - path: "**/*.{sh,yaml,yml}"
      instructions: >
        Flag unsafe shell (unquoted expansions, injection via interpolated
        inputs) and, in graph YAML, over-broad allowed_tools or
        permission_mode values that weaken the tool-safety gate.

  # Static analysis raises review quality well beyond the LLM pass alone.
  tools:
    golangci-lint:
      enabled: true
    gitleaks:
      enabled: true
    shellcheck:
      enabled: true
    markdownlint:
      enabled: true

chat:
  auto_reply: true