メニュー
Patterns for headless Claude execution via claude -p --output-format json. Tool allowlist tiers, env inheritance, model override for Sonnet routing.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SOC 職業分類に基づく
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。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ansonphong/meta-dev --skill headless-workerコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Structured, report-only review across correctness, safety, patterns, coverage, and scope with host-native reviewer routing.
Consult Fable 5 before escalating a judgment call to the human. Use when a long-horizon or autonomous run hits a hard architecture / design / implementation-taste decision it would otherwise stop and ask the human approver about. Fable answers with a recommendation, a confidence score, and a falsifier; a verdict that clears 0.90 AND carries evidence is adopted and logged, anything weaker escalates to the human with Fable's recommendation as the lead option. Safety-class decisions (destructive, deploy, security, money, schema, cross-repo) never route here — they always reach the human.
The shared host-neutral execute→review→fix loop: fresh scoped workers, focused causal verification, durable commits, and one native phase review.
| name | headless-worker |
| description | Patterns for headless Claude execution via claude -p --output-format json. Tool allowlist tiers, env inheritance, model override for Sonnet routing. |
The conceptual claude -p patterns documented here are the foundation. The concrete multi-backend implementation — DeepSeek / GLM / Codex routing tables, the distilled-result contract, --repo topology resolution — lives in ${CLAUDE_PLUGIN_ROOT}/scripts/claude-headless-exec (and codex-headless-exec), fronted by /deep-execute, /glm-execute, /codex-execute, and orchestrated by /auto-execute. This file covers the general headless-model principles only; do not duplicate the backend-specific tables.
Execute structured tasks via claude -p in headless mode. Useful for batch processing, CI/CD, and cron-driven automation.
# Single prompt, JSON output
claude -p "analyze this diff: $(git diff)" --output-format json
# With tool access (read-only by default)
claude -p "check test results in /tmp/test-output.log" \
--allowedTools Bash --output-format json
# Model override for Sonnet
claude -p "complex architectural review" --model claude-sonnet-5
Stage prompts to UNIQUE per-run paths and pass them by file, not by reusing a
bare scratchpad name. All three runners (codex/grok/claude-headless-exec)
accept --prompt-file <path> (AUTHORITATIVE over any positional prompt) and now
hard-fail on a missing/empty file rather than silently skipping the work. Full
conductor-side rules — unique dir, atomic .tmp→mv, [ -s ] verify, absolute
paths — are in agentic-exec-loop/references/loop-protocol.md → "Scratchpad
staging".
See references/tool-allowlists.md for tier definitions.
| Tier | Bash | Read, Write, Edit | Git |
|---|---|---|---|
| 1 — read-only | yes (safe commands) | Read only | no |
| 2 — read+write | yes | Read + Write + Edit | no |
| 3 — full | yes | all | yes |
Default: Tier 1.
See references/backend-env.md for routing matrix.
claude -p inherits ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN from parentenv -i — it strips inherited vars