AI Agent & Automation Workflows

Use when modifying assets (rules, skills, agents, hooks) in ai-dev-setting and deciding whether changes need reinstallation across projects, or when running setup/update-all commands.

실행 모드(once/scheduled/realtime) 간 경계 위반과 LLM SDK 직접 호출을 감지한다. Use when writing or reviewing code that touches execution modules, LLM calls, or import statements across mode directories.

반복되는 결함·리뷰 지적·경계 위반을 docs/design-docs/core-beliefs.md 의 R 룰로 승격하고, 대응하는 강제 장치(테스트·린터 규칙)를 스캐폴드한다. Use when the same defect/review comment/boundary violation appears 2+ times, when the user says "이거 룰로 박아"/"rule로 올려", or when weekly-doc-gardening issues a rule-candidate. Invoke before continuing implementation if the pattern is load-bearing.

비자명한 작업에 계획 → 구현 → 검증 흐름을 적용하고, 큰 변경일 때만 깊은 리뷰로 승격한다. Use when starting a new feature, multi-file change, or any non-trivial implementation task.

Crystallize repeated conversation patterns into reusable skill files. Triggered automatically by the Hermes learning loop (Stop hook) when the same pattern is detected 3+ times (patterncount count >= 3, crystallized = 0). Saves the new skill under [project]/.hermes/skills/, registers it in skillindex, and marks the pattern as crystallized.

Run Hermes dreaming — consolidate accumulated rolling summaries into crystallized/evolved skills and propose junk-skill deletions. Use when the user types /hermes-dream [apply]. Without 'apply' it crystallizes/evolves automatically and only PROPOSES deletions (dry-run); with 'apply' it also EXECUTES the proposed deletions.

Goal-based autonomous loop for the CURRENT project. Trigger when the user types /hermes-loop (with or without a goal), or asks in natural language to run/start a Hermes loop (e.g. "헤르메스 루프 하자", "헤르메스 루프 돌려줘", "목표 루프 작업하자", "run a hermes loop"). On trigger, FIRST present the options form and ask the user to fill the required goal, letting them skip any optional field, THEN run. The target is always the current project where this skill is installed — never ask which project. Destructive actions go through the standard permission prompts.

Recall past conversation context from Hermes rolling summaries. Use when the user types /hermes-recall [keyword]. Reads the project's .hermes/state.db sessionsummary table and prints matching 5-slot summaries (decisions/open tasks). Without a keyword, shows the most recent session summary.

Show the overall Hermes engineering status. Use when the user types /hermes-status. Reads the project's .hermes/state.db and prints skill/rule/session counts, pending crystallization patterns, and the latest/most-used crystallized skills, then offers to run crystallization if patterns are pending.

Node.js runtime patterns — ESM, package manager discipline, async/await, streams, error handling, environment config.

Use when a plan or spec contains an acceptance-criteria checklist ("수용 기준", "acceptance criteria", "manual verification", "수동 검증", browser walk-through, feature sign-off) that the agent needs to confirm before marking the work done. Split every item into programmatically-verifiable (auto) and human-only (human) parts, run the auto parts immediately, record evidence in the audit doc, and ask the user only for the residual human items. Invoke even when the plan labels the step "수동 검증" or "manual test" — most such checklists have items an agent can prove without user eyes via a DB query, API call, grep, or log scan. The point is to stop dumping the entire checklist on the user when half of it is mechanically verifiable.

Use when creating new files, planning features, or writing exec-plans — before any code is written, to ensure each file has a single responsibility and folders are organized by…

Testing FastAPI apps with pytest-asyncio and httpx.AsyncClient — fixtures, dependency overrides, DB isolation, auth.