メニュー
Use when: changing or auditing output channel configuration, no-problems behavior, output templates, target links, author mentions, or dispatch contracts; do not use for agent runtime or VCS fetch internals.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | output-channel-contracts |
| description | Use when: changing or auditing output channel configuration, no-problems behavior, output templates, target links, author mentions, or dispatch contracts; do not use for agent runtime or VCS fetch internals. |
| user-invocable | false |
Use this skill when work touches how AICR decides whether, where, and how review results are rendered to PR/MR comments, issues, bots, or other notification channels.
no_problems / LGTM / empty-summary publishing policy.Do not use this skill for LLM prompt layering, agent runtime materialization, or VCS scoped-fetch implementation unless those changes directly affect output rendering.
Read the authoritative surfaces
../../../docs/ai/architecture.md §3.9 and §3.10.../../../Plan.md when roadmap status or remaining milestone scope matters.../../../docs/output-channels.md.../../../packages/core/src/config.ts and ../../../packages/core/test/config.test.ts for schema coverage.../../../packages/server/src/bootstrap.ts and ../../../packages/server/src/review-orchestrator.ts for publisher routing and empty-summary behavior.../../../packages/outputs/src/template-engine.ts and related tests for template variables.Preserve per-channel policy semantics
no_problems.action in this order: built-in defaults → outputs.no_problems → outputs.channels[].no_problems → workspaces.defaults.outputs.* → workspaces.instances.<id>.outputs.*.publish or suppress); avoid negative booleans such as skip_no_problems.review.problem_issue.max_recent_issues (default 20, global → workspace override, range 1..100) before closing or deleting stale issues.Render target links accurately
event.url as View PR.Keep docs, examples, and tests aligned
Plan.md roadmap summary, ../../../docs/ai/architecture.md, docs/output-channels.md, example/config.yaml, and config schema tests together.update_existing marker parsing (aicr:managed, aicr:scope, aicr:problems), same-scope PATCH behavior, cross-scope non-overwrite behavior, and whitespace-tolerant fingerprint parsing.publishProblem buffers, publishSummary flushes), one-body Markdown aggregation, line-comments-only route flushing, fallback to one issue comment on 403/422, and non-inline problem inclusion in the same body.<!-- aicr:problem-meta=BASE64_JSON --> marker roundtrip, readable title rendering ([SEVERITY] category — file:line), legacy no-metadata fallback without raw fingerprints, and anchor ID generation for both open and resolved items.review.labels.ignore causes the webhook handler to return ignored_by_label before scheduling a review. This is checked per-event, not per-channel.review.labels.auto_tag and review.labels.reviewed_tag are applied by PR/MR and issue dispatchers during publishSummary or publishAggregatedProblems. They are resolved/created via the platform Labels API and attached alongside any severity_label_prefix labels.gitea_problem_issue applies both tags at issue creation time via body.labels.workspaces.defaults.review.labels and workspaces.instances.<id>.review.labels.Load these sibling files only when the task touches the named surface:
references/im-bot-message-contracts.md: IM bot Markdown/card rendering, mentions, truncation, and structured-output repair rules.references/managed-problem-issues.md: gitea_problem_issue / github_problem_issue lifecycle, issue modes, scope fingerprints, file-scope resolution guard, deduplication, and cleanup behavior.review.skip_lgtm with output routing; the former guides review behavior, the latter decides dispatch per channel.publishEmptySummary boolean when different channels have different policies.per_problem and consolidated modes, stale issues from the previous mode will be cleaned up during the next reconciliation cycle.AicrOutputCollector.reportProblem deduplicates by fingerprint or sha256(file:line:category:message)[:16] at collection time; collectCompletionOutputs in the orchestrator skips text re-parsing when MCP state or stream tool-call events already produced review output; reconcileProblems in github_problem_issue / gitea_problem_issue dispatchers calls dedupProblemsByFingerprint before building issue bodies; and the bootstrap publishSummary wrappers for those channels guard reconcileProblems with a per-publisher reconciled flag so it runs at most once per review.Use when: deploying AICR to a remote server, repairing managed issue state after deployment, updating deployed config.yaml, syncing source files, or troubleshooting podman build/run issues; do not use for local development or CI pipeline changes.
Use when: comparing roadmap/docs with implementation, auditing or retiring completed task plans/specs, identifying gaps, fixing issues, or adding missing tests; do not use for unrelated feature work.
Use when: validating repository-wide build, lint, test, markdown, Docker, docs, CI, root tooling, or workspace topology changes; do not use for isolated package-only feature edits.
Use when: auditing, creating, or optimizing AI agent prompts, bridge files, skills, SKILL.md metadata, and cross-tool compatibility; do not use for ordinary feature edits that leave AI-facing assets unchanged.
Use when: implementing or auditing agent CLI runtime materialization, LLM config translation, MCP tool mapping, prompt/instruction layering, and skills merging; do not use for ordinary review logic or output rendering changes.
Use when: planning, coding, refactoring, reviewing, or editing prompts for non-trivial tasks where assumptions, scope control, simplicity, or verification matter; do not use for obvious one-line fixes.