orchestrator: add /ci-test-review skill (in THIS repo) + drop Phase 3r from loops queue

The on-demand AI review layer is now an orchestration-repo skill built directly
by the orchestrator, NOT a loops phase in the cc-ci product repo:

- .claude/skills/ci-test-review/{SKILL.md,run-all-recipes.sh}: runs the real
  cc-ci harness across all enrolled recipes (deterministic, AI-free execution),
  then AI diagnoses each failure and classifies it as needing a recipe PR / a
  CI-server PR / a stale-test update — or reports "ALL PASSED, recipes + tests
  up to date". Proposes PRs; never decides pass/fail; never auto-merges.
- .gitignore: track .claude/skills/ (shareable) while still ignoring local
  claude session state (locks, history) under .claude/.
- launch.sh: remove Phase 3r from PHASES_SPEC; loops sequence back to
  1c 1b 1d 1e 2w 2pc 2 2b 3 4. Deleted plan-phase3r (superseded by the skill).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-05-29 16:57:26 +01:00
co-authored by Claude Opus 4.8
parent 5f84f8c028
commit 2530845e50
3 changed files with 167 additions and 83 deletions
+80
View File
@@ -0,0 +1,80 @@
---
name: ci-test-review
description: On-demand AI review of the cc-ci server — runs the deterministic recipe test suite across ALL enrolled recipes, and for any failure diagnoses the root cause and classifies it as needing a RECIPE PR (with the proposed change), a CI-SERVER PR, or a stale-test update; or reports "all passed, recipes + tests up to date". Use when the operator wants a full health check + actionable diagnosis of the recipe CI, beyond the deterministic pass/fail of normal !testme runs. Invoke as /ci-test-review.
---
# ci-test-review
An **on-demand, AI-driven review** of the cc-ci recipe CI server. cc-ci's primary use is
**deterministic** testing (normal `!testme` / nightly — no AI). This skill adds a review layer: run
the full suite across **all** recipes, then **AI-diagnose** any failures and say exactly what fixes
them — a recipe PR, a CI-server PR, or a test update — or confirm everything is green and current.
**Hard boundary:** the test *execution* is deterministic (the cc-ci harness). **AI is used ONLY for
diagnosis / classification / proposing the PR — never to decide pass/fail.**
## Access (this skill runs from the orchestration repo, drives cc-ci over the tailnet)
- cc-ci is reachable via `ssh cc-ci` (root, through the userspace-tailscaled SOCKS proxy on
127.0.0.1:1055). If `ssh cc-ci` fails, restart the proxy (`systemctl restart cc-ci-tailscaled`) per
the orchestrator's access notes, then retry.
- The cc-ci repo + harness live at `/root/cc-ci` on the cc-ci VM; the runner is
`cc-ci-run runner/run_recipe_ci.py` with `RECIPE=<name>` (+ optional `STAGES=...`).
## Procedure
### 1. Run the deterministic sweep (no AI)
Run the helper, which enumerates enrolled recipes and runs the **full suite per recipe** on cc-ci,
emitting one structured result line per recipe (per-tier pass/fail/skip + the tested vs latest
published version):
```
bash "$(dirname "$0")/run-all-recipes.sh" # or: .claude/skills/ci-test-review/run-all-recipes.sh
```
It writes a results table to stdout and a machine-readable summary to
`/srv/cc-ci/.cc-ci-logs/ci-test-review-<runid>.json`. Do NOT hand-judge pass/fail — trust the
harness's per-tier results.
### 2. If everything is green AND current → report all-clear
If every recipe's tiers are `pass`/`skip(N/A)` (no `fail`) **and** every recipe was tested at its
**latest published version** with no stale test flagged, output:
> **ALL PASSED — recipes up to date, tests up to date.** <N recipes, versions listed>
Stop here. Nothing to fix.
### 3. For each failure → diagnose root cause + CLASSIFY (this is the AI part)
For every `fail` (after confirming it's not a flake — see below), read the run log, the recipe's
compose/code (on cc-ci `~/.abra/recipes/<recipe>`), and the harness, then determine the **root cause**
and classify it as **exactly one** of:
- **RECIPE bug → recipe PR.** The recipe itself is broken/fragile (bad healthcheck, missing backup
hook, startup race, etc.). Output the **concrete proposed change** (e.g. "add a collabora WOPI
healthcheck + `start_period`", "add a `pg_dump` backup hook"). This is a PR to the recipe
(recipe-maintainer); it is "working" only once cc-ci verifies it green, then the operator merges.
- **CI-SERVER bug → cc-ci PR.** The harness/test/infra is wrong (a test asserts the wrong thing, a
readiness gate is off, an infra config bug). Output the proposed change to the cc-ci repo.
- **TEST out-of-date → cc-ci test update.** The recipe legitimately changed upstream and the cc-ci
test/overlay needs updating to match. Output the update.
- **FLAKY / infra-transient.** Distinguish from a real failure: **re-run the recipe once or twice**;
if it then passes, classify as flaky (note it; suggest a robustness fix if there's a pattern) — do
NOT report a flake as a recipe/CI bug.
### 4. Freshness (part of "up to date")
For each recipe, compare the **tested version** to the **latest published** recipe version (the
sweep records both). Flag any recipe tested at a stale pin, and any cc-ci test that lags a recipe
change — "all passed" must mean "passes against current upstream", not against an old pin.
### 5. Output a structured report
- A per-recipe **status table** (recipe | tiers | version tested / latest | verdict).
- For each failure: **root cause → classification (recipe PR / CI PR / test-update / flaky) →
proposed change** (concrete enough to act on).
- Or the all-clear summary from step 2.
## Guardrails
- **Deterministic execution stays AI-free** — only diagnose; never override the harness's pass/fail.
- **Propose, don't auto-merge.** This skill outputs recommendations + proposed PRs. It does NOT
create or merge PRs. (A recipe PR is operator-merged only after cc-ci verifies it green.) If asked,
it may hand a specific proposed recipe PR to the `recipe-create-pr` skill as an explicit follow-up.
- **Don't weaken any test** to make the review "pass."
- **Flake ≠ bug** — re-run before classifying a failure as a real recipe/CI defect.