agentleFS
Sign inStart free
willink-osswillink-oss/willink-claude-kit/skills/self-heal-ci/SKILL.md

self-heal-ci

Skill0 starsChanged 4 months agoMITView on GitHub

Detect a red CI and relay \"fix → re-verify\" until CI is green, BOUNDED by an attempt cap. A goal-loop wrapper whose stop condition is a deterministic gate (gh conclusion=success), not the model's self-report. Triggers: self-heal-ci, self heal ci, CI self-heal, red CI heal, fix failing CI, keep fixing CI until green.

The file

Create a free account to mount
Files
SKILL.md

self-heal-ci

A wrapper over scripts/goal-loop.sh (the stop primitive) that self-heals a red CI: when the latest run is red it relays "fix → re-verify" until CI is green, or until an attempt cap is reached. The stop / counting discipline is delegated to goal-loop.sh — this wrapper only drives detection and the fix cycle, it never loops forever on its own.

The goal test is a single deterministic point: gh run list -L1 → latest run's conclusion == "success". Everything else (failure / cancelled / timed_out / null=in-progress / ""=fetch failed) is treated as red (fail-safe — an empty gh result is "red (unknown)", never "0 runs / green"). "Done" is decided by that gate's exit code, never by the model reporting "I fixed it".

How to run

  1. From the target repository's working tree (or with --repo O/R), start the loop:
    bash "${CLAUDE_PLUGIN_ROOT:-.}/scripts/self-heal-ci.sh" [--repo <owner/repo>] [--max <N>] [--escalate-file <path>]
  2. Each round the wrapper calls goal-loop.sh and measures whether the latest run is green:
    • green (exit 0) → ✅ done, stop.
    • red & under cap (exit 1) → run ONE fix cycle. If claude is on PATH it runs headless (one fix = feature branch + PR, following CYCLE-PROMPT.md). With DRY_RUN=1 or no claude on PATH it is gate-only: it returns exit 1 so an outer caller fixes and re-runs.
    • cap reached (exit 2) → escalate (see below) and stop, so it never burns tokens forever.
  3. Fixes are PR-only. No self-merge / tag push / push to main — that boundary lives in CYCLE-PROMPT.md (the prompt the fix agent reads).
  4. The Verifier of "is it really green?" is a deterministic gate, not the fixer. Independent review of the actual diff can be done by the kit's dev-reviewer agent (read-only), a /review session, or a human.

Modes / options

Flag / envMeaning
(none)run the self-heal loop (default)
--repo O/Rtarget repository (defaults to the cwd repo)
--max Nattempt cap (default 5; SELF_HEAL_MAX env also works)
--escalate-file <path>on CAP, append the escalation line here instead of stdout (SELF_HEAL_ESCALATE_FILE env also works)
--state <path>where the attempt counter persists (SELF_HEAL_STATE env; default .self-heal-ci.state)
--ci-checkexit 0 if latest run is green, exit 1 if red (the body of goal-loop's --check)
--self-testgh-independent deterministic self-test (exit 0 = PASS)
DRY_RUN=1gate-only: detect once and exit; do not launch claude (smoke test)

Escalation sink

On CAP the wrapper prints a single escalation line to stdout by default. Pass --escalate-file <path> (or set SELF_HEAL_ESCALATE_FILE) to append it to a file of your choice instead — e.g. an action list, a log, or an issue body draft. No path is hardcoded.

Deterministic --check (self-test)

The skill / script's own health is machine-verified without any dependency on gh:

bash "${CLAUDE_PLUGIN_ROOT:-.}/scripts/self-heal-ci.sh" --self-test
  • exit 0 = PASS. It verifies (no hardcoded success):
    1. bash -n (its own syntax).
    2. that it references goal-loop.sh and that the stop primitive exists.
    3. the ci_conclusion_ok predicate on fixtures (only success is green; every other value, including empty and SUCCESS, is red).
    4. the --ci-check path end-to-end with a fake gh (hermetic fixture): success→exit0 / failure→exit1 / empty→exit1 (fail-safe).
  • non-zero: a gap in the conclusion truth table, the wiring, or syntax.

When to use / not

  • ✅ A repo has a red CI you want driven to green with a bounded, deterministic stop.
  • ❌ No gh-reported CI, or a one-off manual fix that needs no re-verify loop (use /build).

Related

  • primitive: scripts/goal-loop.sh (exit 0=GOAL MET / 1=CONTINUE / 2=CAP REACHED)
  • wrapper: scripts/self-heal-ci.sh (--ci-check / --self-test / --repo / --max / --escalate-file)
  • fix cycle prompt: skills/self-heal-ci/CYCLE-PROMPT.md
  • Verifier role: the dev-reviewer agent (read-only), a /review session, or a human
  • command: /goal-loop; the green-CI pattern lives in examples/ci/all-checks-pass-pattern.md

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free