技能 2026-10-10 20:39 UTC
skills
由提交人 chrisbanes 提交,主要更新类型为:常规更新与维护。提交说明:Keep drain alive under a read-only Project watcher (#113)。对项目做了常规维护与小幅调整。
Commit提交信息原文
Keep drain alive under a read-only Project watcher (#113) ## Summary `run-github-project` `drain` used to return `waiting-for-human` as soon as only human-gated work remained, and its "monitor all remote slots" step never said how to watch. This adds a deterministic, read-only watcher so a drain can wait cheaply and wake on relevant GitHub changes. - **`scripts/watch_project.py`** (stdlib, `gh api graphql` only): - `snapshot` prints a fingerprint of every Project item (Status, assignees, labels, state), watched PRs (state, head SHA, checks rollup, review decision, mergeability) and watched issues (latest comment). - `wait` exits with one JSON line: `changed` (keys and field names only), `deadline` or `error`. - Transient `UNKNOWN` mergeability is ignored. - Each `gh` call is bounded by the deadline, and polling backs off on rate limits. - **Drain rules** (`references/drain-scheduler.md`, a new `### Project Watcher` section): - The watcher runs whenever a PR is in remote wait or human-gated work exists. - The baseline is taken before each read, and the watcher is relaunched through host cancellation, never by process match. - A report is only a hint: it never satisfies the Refresh Gate or authorizes anything. - Each drain keeps a 24h no-progress deadline, and three consecutive failures mean monitoring is unavailable. - Hosts: Claude Code runs the script in the background. Other hosts (Codex) relay it through the cheapest subagent and `wait_agent`. - `next` is unchanged. - **Other references:** human-frontier, authority-and-pauses, project-config (an optional poll interval, default 120s), `SKILL.md` and the README are updated to match. - **Evals:** four calibration-only cases (direct, novel, `next` negative, restraint counterexample). The calibration count goes 61 → 65, and the 38-case benchmark is unchanged. ## Validation - `npm test`, `npm run lint`, `npm run build` and `npm run evals:validate` (167 cases) pass. - Watcher tests run against a fake `gh`, with no network. - An independent review returned `fix-first`; the repairs were re-reviewed and approved as `ship`. ## Not checked - No live run against a real Project or GitHub's live GraphQL schema. - Not checked whether Codex or Claude Code background commands can block for up to 24h. - No live model evals.