triage-task
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-08-02
- Stars
- ★ 1
- Forks
- 0
03 SKILL.md 正文
Triage Task
Use the configured small model tier. The adapter resolves this to a concrete model at runtime (for Claude Code, the small tier maps to the haiku alias by default).
Purpose
Classify the request from LeanRigor's deterministic evidence packet and recommend the next workflow. Do not inspect the repository, implement, edit files, run tests, create a detailed plan, launch agents, or commit.
Required output
Return only JSON matching the ModelTriageRecommendation schema in
src/core/triage-schema.ts.
The output must contain:
- one-sentence task summary;
- task type;
- separate complexity, ambiguity, blast-radius, architecture, security, data-integrity, and operational assessments;
- model recommendation only;
- confidence from 0 to 1;
- sequential or candidate parallelism;
- review and test levels;
- at most one blocking clarification question;
- concrete additional-inspection questions only when materially justified;
- at most three escalation reasons;
- at most three assumptions;
- explicit constraints.
Decision rules
- Select the lowest workflow mode that safely handles the identified risks.
- Never equate implementation difficulty with workflow rigor.
- Fast requires positive evidence: low ambiguity, low blast radius, low architectural impact, and no security, data-integrity, operational, migration, or public-contract risk.
- Rigorous requires an explicit trigger such as authentication, authorisation, credentials, payments, migrations, destructive data work, public API compatibility, production infrastructure, concurrency, privacy, compliance, or broad architectural boundary changes.
- Bug fixes normally require Standard mode with targeted regression validation.
- A difficult read-only investigation may remain Standard.
- Recommend only whether parallelism is a candidate. The execution-graph skill makes the final task split.
- Request targeted inspection only through structured questions with exact allowed paths. Do not ask for broad repository analysis.
- Ask only one blocking question, and only when its answer could change scope, architecture, compatibility, safety, mode, or acceptance criteria.
- Do not produce implementation advice or an execution plan.
Confidence behaviour
0.80–1.00: proceed to policy overrides.0.55–0.79: permit one narrow context-enrichment pass.- Below
0.55: ask one blocking question or use the safer adjacent mode. Low confidence alone does not justify Rigorous.
Validation and fallback
The orchestrator validates the response and then applies deterministic repository policy. Invalid output may receive one structured repair request. Provider failure falls back to deterministic local triage unless a known high-risk trigger requires Rigorous.
04 同仓库更多 Skills
leanrigor-sequential-workflow
Use when running LeanRigor's persisted sequential workflow in Claude Code.
validate-integration
暂无简介
prepare-commits
暂无简介
inspect-repository
暂无简介
evaluate-options
暂无简介
dispatch-execution
暂无简介