사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/zernie/vigiles --skill debug-my-harness시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
Diagnose harness misbehavior from the flight recorder — the local, append-only ledger at .vigiles/runs.jsonl that vigiles writes as your harness runs. It records what actually happened, so you debug from evidence instead of guessing.
What's in the ledger
One JSON record per line, each with a kind:
hook— a compiled-hook gate decision:{event, decision: allow|deny|ask, mode: enforce|observe, rule, cmd, reason}.agent— a subagent tool-contract decision:{name, tool, allowed, reason}(afalse= the agent went outside its lane).skill— a skill activation:{name, fired}.eval— a measured metric:{name, metric, value}(e.g. trigger-rate recall/precision).capability-diff— a blast-radius change:{pr, added, removed, widened}.
Instructions
Step 1: Read the ledger
Read .vigiles/runs.jsonl (JSONL — one record per line; tolerate a torn last line). If it's absent or empty, say so — there's nothing recorded yet; suggest running the harness (or vigiles audit) first. Do NOT fabricate records.
Step 2: Answer the specific question, evidence-first
Match the user's question to the ledger:
- "Why did skill X stop firing / why does the wrong one run?" — count
skillfires by
name over time. If X's fire-rate dropped, look for a sibling that fired on the same kinds of prompts (a selection collision) and check their descriptions for overlap. Recommend differentiating or merging the descriptions.
- "Why didn't my hook block that?" — find
hookrecords for the event. A `decision:
allow on something that should be denied, or mode: observe (shadow, never blocks), or the absence of any record, tells you which. Recommend flipping observe→enforce` or fixing the gate logic.
- "Did a subagent misbehave?" — list
agentrecords withallowed: false: the agent
reached for a tool outside its declared contract. Point at the contract to tighten or widen.
- "Is it getting worse?" — compare
evalmetric values (recall/precision) across runs;
a downward trend is drift (often after a harness/model upgrade).
Step 3: Recommend a fix, tied to the evidence
Prefer promoting an ignored-but-decidable rule from prose to a deterministic gate: a repeated agent violation or a rule the agent keeps breaking → a compiled hook or a tighter tool-contract (the strengthen skill can help). A description collision → differentiate the skill descriptions. Always cite the specific records you based the diagnosis on.
Step 4: Offer the next step
If the fix is a spec change, hand off to edit-spec. If it's promoting guidance to a linter rule, hand off to strengthen. If a behavioral claim needs measuring (does the skill fire now?), hand off to test-harness (measureTriggerRate).
