AGENT SKILL

health-check

Read-only health report on the development harness — always-on budget headroom, eval cases, pending proposal age, specs stuck in flight, stale gate markers, vendor opt-in, required checks, main's CI, open pull requests with unresolved threads, and local branches whose pull request merged

License 许可
Apache-2.0
所属仓库
turbokast/mythhelm
最近更新
2026-11-10
Stars
★ 0
Forks
1

01 前言 Front Matter

name
health-check
description
Read-only health report on the development harness — always-on budget headroom, eval cases, pending proposal age, specs stuck in flight, stale gate markers, vendor opt-in, required checks, main's CI, open pull requests with unresolved threads, and local branches whose pull request merged
argument-hint
[--offline] [--json] [--stale-days N]

03 SKILL.md 正文

Health Check

One report on whether the harness is in a state to trust. It changes nothing: every finding comes with the command that would act on it, for a maintainer or another skill to run. The checks and their statuses are defined in scripts/harness/health.py; knowledge/learning-loop.md explains where each fits in the loop.

Input

$ARGUMENTS, all optional: --offline (skip the GitHub checks), --json (the machine-readable report), --stale-days N (the age at which a proposal or an in-flight spec is reported, default 7).

Invocation contexts

  • Slash command: runs both steps and prints the report.
  • Model-invoked: the same.
  • Non-interactive: the same; the report is the final message. The skill writes nothing, so it has no gates.

Steps

  1. Report. Unless --offline, refresh the remote refs first so origin/main is current (a fetch changes no branch or file):

    git fetch -q origin
    python3 scripts/harness/health.py $ARGUMENTS
    

    The script prints every check with one status: ok, warn, fail, unknown or skipped. A check that could not run is unknown with its reason, never omitted, so the report shows checks=10 on every run. Quote its output; do not re-derive or summarise away a status.

  2. Next steps. For each warn, fail or unknown, name the action, without taking it:

    Check Action
    budget move always-on text to path-conditional rules or knowledge/ (.claude/rules/agent-config-conventions.md)
    evals the named case lost its behaviour: fix the change that broke it, or update the case in the same pull request that deliberately changes the behaviour
    proposals /apply-proposals
    specs /run-spec <spec> or /finalize-spec <spec> to resume, or /evaluate-spec <spec>
    gate-markers scripts/harness/gate.sh <scope> in that worktree before any completion claim
    required-checks a maintainer restores the ruleset, or updates .claude/data/required-checks.json and docs/automation.md together
    main-ci real: a fix pull request; infra or unknown: the maintainer decides on a re-run
    open-prs answer and resolve the threads on each pull request
    merged-branches scripts/harness/prune-merged.sh lists what it would remove; with --apply it deletes each branch and its clean worktree after verifying the merge (a squash merge leaves git branch -d unable to tell)

    Never delete a branch or worktree, re-run a job, merge or edit anything from this skill.

Output

The script's report verbatim, followed by the next-step lines for every check that is not ok or skipped:

Harness health
  ok       budget           always-on 13291 of 16384 bytes, headroom 3093
  warn     proposals        2 pending, oldest 21 days; decide them with /apply-proposals
  ...
checks=10 ok=7 warn=2 fail=0 unknown=1 skipped=0
Next: proposals -> /apply-proposals; ...

04 同仓库更多 Skills