メニュー
Guidance on shrinking or fully resetting conversation context between auto-engineer iterations — when to use /compact (built-in summarise), /clear (built-in wipe), or a full process restart via scripts/restart-loop.sh.
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
Seed the auto-engineer SDLC toolkit into a project. Asks for the target project path, detects the tech stack and GitHub configuration, collects any missing values interactively, then writes customized skill files and Docker infrastructure into that project. Use when the user says "seed a project", "set up auto-engineer", "install the loop", or invokes `/seed`.
Two-way sync between a seeded project and this auto-engineer repo's templates — covers skills, Dockerfile, and scripts. Detects drift in either direction, presents per-file diffs, and lets you pull from template, push to template, or skip. Harness-internal tool — never seeded into target projects. Invoked as /sync.
Autonomously drive project work end-to-end — pick an unblocked issue, plan it, implement, push a PR, resolve CI and review-bot feedback, merge when green, then repeat. Use when the user says "auto-engineer", "auto-pilot", "go run the loop", or invokes `/auto-engineer`.
Drive a large body of work end-to-end — either from a single scoping issue or a fuzzy topic ("ship GA", "harden the review app") — by discovering scope in the backlog + codebase, filing sub-issues, iterating with the user, planning parallel workstreams, then spawning `/auto-engineer` subagents to ship the whole epic autonomously. Use when the user says "auto-manager", "manage this epic", "orchestrate /auto-engineer across…", or invokes `/auto-manager`.
File a new tracker issue with the correct priority and area labels applied up front. Use whenever an agent needs to capture a new unit of work — follow-ups from a merged PR, TODOs discovered while coding, or manually-requested issues. Invoked as `/file-issue`.
Autonomously drive project work end-to-end — pick an unblocked issue, plan it, implement, push a PR, resolve CI and review-bot feedback, merge when green, then repeat. Use when the user says "auto-engineer", "auto-pilot", "go run the loop", or invokes `/auto-engineer`.
| name | context-reset |
| description | Guidance on shrinking or fully resetting conversation context between auto-engineer iterations — when to use /compact (built-in summarise), /clear (built-in wipe), or a full process restart via scripts/restart-loop.sh. |
This skill is a reference document, not an executable action.
/compactand/clearare built-in Claude Code CLI commands — they cannot be invoked by theSkilltool from within a running session. Reading this skill brings the guidance into context so you can apply the right approach for the situation.
Each auto-engineer iteration reads playbooks, fetches issues, reads source files, runs sub-agents, and produces diffs. That context stays in the conversation window across iterations unless explicitly trimmed. After several cycles the window grows large, cache misses become expensive, and prior-cycle residue (old branch names, resolved bugs, superseded plans) can subtly corrupt decisions.
| Situation | Action |
|---|---|
| Normal end-of-cycle, session still healthy, just want to trim cost | /compact (built-in) |
| Prior cycles are actively corrupting decisions — Claude is replaying resolved issues or hallucinating old state | /clear (built-in), then restart auto-engineer with explicit --iteration N |
| Container run requiring a truly clean process (token ceiling hit, env/credential refresh needed) | scripts/restart-loop.sh --iteration N (run on the host) |
/compact — the default (auto-engineer step 10)Summarises the conversation in place. The iteration counter, stop-reason history, and current PR state survive in the summary. This is why auto-engineer step 10 calls /compact rather than /clear — the ScheduleWakeup prompt carries --iteration N forward, but the surrounding rationale lives in the summary.
How to invoke: type /compact directly in the Claude Code CLI. It cannot be called from a skill or tool.
/clear — full wipeErases the entire conversation. Use only when context corruption is confirmed — you will lose the iteration counter and any uncarried state. After /clear, restart auto-engineer explicitly:
/auto-engineer --iteration <N>
How to invoke: type /clear directly in the Claude Code CLI.
scripts/restart-loop.sh — process restartKills the current container and launches a fresh one. The new Claude process starts with zero context. Use when /compact is insufficient (e.g. the context window is full even after compaction) or when you need a clean environment (fresh credentials, updated Docker image, memory hygiene).
# Start from iteration 1
scripts/restart-loop.sh
# Resume at a specific iteration
scripts/restart-loop.sh --iteration <N>
The --iteration N flag is passed through to /auto-engineer so the loop continues from where it left off with a clean slate.
/compact (auto-engineer does this automatically in step 10)./clear + /auto-engineer --iteration N./compact, or you need a fresh container → scripts/restart-loop.sh --iteration N.