hns-workflow-ci-loop

CI Loop (hns-workflow-ci-loop)

Unified CI watch + auto-fix loop. The orchestrator invokes this skill after /moai sync Phase 4 (gh pr create) returns a PR number. The skill polls required checks, classifies failures into mechanical vs semantic, attempts safe patches up to 3 iterations, and escalates semantic failures via AskUserQuestion.

Quick Reference

Trigger: /moai sync Phase 4 PR-create returns a PR number, or an existing PR needs CI monitoring.

Two phases, one skill:

1. Watch — Poll gh pr checks every 30s, classify required vs auxiliary via .github/required-checks.yml SSoT; exit on green/fail/timeout.
2. Auto-fix — On required-fail (exit 2), receive JSON handoff, run up to 3 patch iterations, escalate semantic failures immediately.

One-liner:

MOAI_CIWATCH_GH=gh sh scripts/ci-watch/run.sh <PR_NUMBER> <BRANCH>

Exit codes:

• 0 — all required passed → ready-to-merge AskUserQuestion
• 1 — fatal error → surface remediation
• 2 — required failure → JSON handoff → auto-fix phase
• 3 — 30-min hard timeout → blocker message, return control

HARD invariants:

• AskUserQuestion is orchestrator-only — CLI, shell scripts, and the manager-develop (cycle_type=autofix) subagent MUST NOT call it.
• Force-push is absolutely prohibited (--force, -f, --force-with-lease all banned).
• Max 3 auto-fix iterations; iteration 4+ triggers mandatory blocking AskUserQuestion.
• Semantic failures (race, deadlock, panic, assertion) are never auto-patched.
• Protected files: .env*, credentials, .claude/settings*.json, .github/required-checks.yml, scripts/ci-watch/run.sh.

Implementation Guide

Phase 1 — Watch Loop

Polling cadence: 30 seconds minimum (GitHub API rate-limit). Override via CIWATCH_POLL_INTERVAL env, never below 30 in production. Test mode uses MOAI_CIWATCH_NO_SLEEP=1 (single-tick exit).

30-minute hard timeout: CIWATCH_TIMEOUT_SECONDS=1800 default. On timeout, exit 3. Do not auto-restart.

Required vs auxiliary: Required checks live in .github/required-checks.yml branches.<pattern>.contexts. Auxiliary checks listed under auxiliary: MUST NOT block ready-to-merge. Hardcoding check names in scripts is prohibited.

State file: .moai/state/ci-watch-active.flag (YAML). Tracks pr_number, started_at, heartbeat_at, required_checks, abort_requested. Heartbeat staleness > 90s allows takeover. Abort: moai pr watch --abort.

Background watch standardization: For long-running PRs (5+ min), use gh pr checks <PR> --watch invoked via run_in_background: true. Sleep + poll loops are prohibited — they block the orchestrator's main session.

Status report format (stderr, state-change ticks only, no ANSI):

[ci-watch] PR #<N>: required 4/6 pass, 2 pending; advisory 0 fail

Handoff schema on exit 2 — JSON with stable fields: prNumber, branch, failedChecks[] (each entry {name, runId, logUrl}), auxiliaryFailCount, totalRequired. Field stability: name, runId, logUrl are stable contract — do not rename. Schema source: the CI-watch handoff struct.

Phase 2 — Auto-Fix Loop

Entry condition: ci-watch exit 2 + valid JSON handoff. State file: .moai/state/ci-autofix-<PR>.json (PR-scoped, 24-hour staleness threshold).

OQ2 cadence matrix (single source of truth for iteration behavior):

• iter 1, any mechanical sub_class → confirm + apply via AskUserQuestion (1st option = "Apply patch (Recommended)").
• iter 1, semantic/unknown → escalate (no patch attempt) via AskUserQuestion with diagnosis report.
• iter 2-3, mechanical + sub_class=trivial → silent apply + log (no AskUserQuestion).
• iter 2-3, mechanical + sub_class=non-trivial → confirm + apply via AskUserQuestion.
• iter 2-3, semantic/unknown → escalate (no patch) via AskUserQuestion.
• iter 4+ → mandatory blocking AskUserQuestion (no timer, options: manual fix / revise SPEC / abandon PR).

"trivial" = whitespace, gofmt/goimports, import-order (matches classify.sh RX_TRIVIAL_*).

Patch commit rule: Every patch = new commit. Format: fix(ci): auto-fix <classification> failure (iter <N>). After push, re-invoke scripts/ci-watch/run.sh to restart the watch loop.

Iteration 4+ escalation (mandatory blocking, no silent timeout):

1. (Recommended) Manual fix — investigate and fix manually
2. Revise SPEC — revise the SPEC and restart implementation
3. Abandon PR — close the PR and abandon this approach

manager-develop (cycle_type=autofix) spawn prompt injects: handoff JSON, classification + sub_class, failed CI log + PR diff, mode directive (mechanical → propose unified-diff patch; semantic/unknown → return diagnosis only, no patch). HARD: no AskUserQuestion call from the subagent — return Markdown only.

Audit log: .moai/logs/ci-autofix/<PR-NNN>-<YYYY-MM-DD>.md. Append-only. Each iteration records classification, sub_class, action, patch_sha, escalation_reason.

Protected Files (never auto-modified)

• **/.env, **/.env.*
• **/credentials*, **/*_key.json, **/*secret*
• .claude/settings.json, .claude/settings.local.json
• .github/required-checks.yml (Wave 1 SSoT, read-only for Wave 2/3)
• scripts/ci-watch/run.sh (Wave 2 invariant)

If the manager-develop (cycle_type=autofix) subagent proposes a patch touching any of these, reject and escalate.

Go Helpers and Shell Scripts

Go helpers: the CI-watch classifier (required-vs-auxiliary), the handoff JSON-schema struct, the watch state file, and the PR-watch report emitters (EmitReadyToMergeReport, EmitFailureHandoff). Shell: scripts/ci-watch/run.sh (main loop, mock via MOAI_CIWATCH_GH); scripts/ci-watch/lib/classify.sh (yq + grep fallback); scripts/ci-autofix/log-fetch.sh (failure log + PR diff); scripts/ci-autofix/classify.sh (mechanical vs semantic).

gh CLI compat: requires gh >= 2.50 for the workflow JSON field. On older gh, classify.sh falls back to name-based heuristics from the required: list.

Works Well With

• manager-develop (cycle_type=autofix) — failure diagnosis + patch proposal subagent

• manager-git — commit/push of auto-fix patches

• .claude/rules/local/ci-watch-protocol.md — HARD watch invocation contract (dev-only; governs scripts/ci-watch/)

• .claude/rules/local/ci-autofix-protocol.md — HARD autofix invocation contract (dev-only; governs scripts/ci-autofix/)

  Both are the dev-repo twins, not the distributed rules of the same basename. The distributed .claude/rules/moai/workflow/ci-autofix-protocol.md is deliberately script-free and governs the manager-develop autofix cycle instead; it does not describe the scripts this skill runs. ci-watch-protocol.md is not distributed at all.

• .github/required-checks.yml — Wave 1 SSoT

Common Rationalizations

• "Skip watch loop for small PRs" — small PRs fail CI too. Loop costs nothing, saves manual polling.
• "Auxiliary failures should block merge" — advisory by SSoT definition. Edit required-checks.yml to change classification.
• "Try auto-patching even semantic failures" — semantic failures (race, assertion) cannot be auto-patched without context; wrong patch is worse.
• "Clean up history with force-push" — force-push destroys reviewer diff visibility. Always a new commit.
• "Time out after iter 3" — silent timeout prohibited; user must decide explicitly.
• "Apply trivial fixes without confirming" — iter 1 always confirms; iter 2+ trivial may silent apply.

Red Flags

• manager-develop (cycle_type=autofix) subagent calls AskUserQuestion (HARD: orchestrator-only)
• Iteration 4 auto-continues without blocking AskUser
• git push --force / -f / --force-with-lease anywhere in scripts
• Semantic classification produces a patch attempt
• State file missing → iteration counter lost → infinite loop risk
• Watch polling interval < 30s (rate-limit risk)
• required-checks.yml modified without moai github init re-run

Verification

• bash scripts/ci-watch/test/run_test.sh passes all shell tests
• go test ./internal/ciwatch/... ./internal/cli/pr/... -race passes
• internal/ciwatch/ coverage >= 85%
• No ANSI codes in FormatStatusUpdate() output
• EmitReadyToMergeReport first option carries (Recommended)
• CLI does NOT call AskUserQuestion
• grep -r 'push -f\|push --force' scripts/ci-autofix/ scripts/ci-watch/ returns no matches
• Audit log .moai/logs/ci-autofix/<PR>-<DATE>.md contains every iteration