CI Sweep

CI Sweep

Reduce the time main stays red. Monitors CI, classifies each failure, and takes the smallest safe action — or escalates when human judgment is needed.

Usage

/ci-sweep                         # check CI status on current branch
/ci-sweep --branch main           # sweep a specific branch
/ci-sweep --run <run-id>          # analyze a specific workflow run
/ci-sweep --report-only           # diagnose but do not apply any fix

Run on a loop for continuous reaction:

/loop 15m /ci-sweep --branch main
/loop 5m /ci-sweep              # during active development

Behavior

Step 1 — Check CI status

# Get recent runs on the target branch
BRANCH="${ARGUMENTS#*--branch }"
BRANCH="${BRANCH:-$(git branch --show-current)}"

gh run list --branch "$BRANCH" --limit 5 \
  --json databaseId,name,status,conclusion,createdAt,headSha \
  --jq '.[] | select(.status == "completed" and .conclusion != "success")'

If all runs are passing → report "CI green" and exit immediately.

Step 2 — Get failure details

# Fetch logs for the failing job
gh run view <run-id> --log-failed 2>/dev/null | tail -100

# Identify the failing step
gh run view <run-id> --json jobs --jq '.jobs[] | select(.conclusion == "failure") | {name, steps: [.steps[] | select(.conclusion == "failure")]}'

Step 3 — Classify the failure

# Check if failure existed before recent commits
git log --oneline -10
gh run list --branch main --limit 20 --json conclusion,createdAt | \
  jq '[.[] | select(.conclusion != null)] | group_by(.conclusion) | .[] | {conclusion: .[0].conclusion, count: length}'

Step 4 — Regression: propose minimal fix

Identify the exact failing line from the error output:

# Extract file and line from test failure
echo "$FAILURE_LOG" | grep -E "FAIL|Error|assert|expected" | head -20

# Find which recent commit touched that file
git log --oneline --since="2 days ago" -- <implicated-file> | head -5

Propose a fix using the minimal-fix pattern (one problem, smallest diff, denylist respected). Cap at 3 attempts per failing job:

# Track attempt count
ATTEMPTS_FILE=".claude/cache/ci-sweep/attempts.json"
CURRENT=$(jq -r --arg job "$JOB_NAME" '.[$job] // 0' "$ATTEMPTS_FILE" 2>/dev/null || echo 0)

if [ "$CURRENT" -ge 3 ]; then
  echo "ESCALATE: 3 fix attempts exhausted for $JOB_NAME"
  exit 0
fi

After applying fix → invoke /loop-verify before staging. Do not mark the fix done yourself.

Step 5 — Flake: re-run

gh run rerun <run-id> --failed

Record the re-run in state. If the same job flakes 3 times in 24h → flag for quarantine:

FLAKE DETECTED: <test-name> has failed intermittently 3× in 24h
Suggested action: quarantine this test and open a tracking issue
Do not code-fix a flake — address the non-determinism.

Step 6 — Config fix

For YAML syntax errors or missing fields, read only the workflow file and apply the minimal correction:

# Validate workflow YAML
python3 -c "import yaml; yaml.safe_load(open('.github/workflows/<file>.yml'))" 2>&1

Config files are not on the denylist; apply minimal fix and note it.

Step 7 — Report

CI SWEEP — main @ abc1234 (2026-06-30 10:45 UTC)

Failing jobs (2)
  test-auth     REGRESSION  → minimal-fix proposed (attempt 1/3)
                            Error: AssertionError in test_refresh_token_expiry:88
                            Implicated commit: a1b2c3d "fix(auth): update token expiry logic"

  build-docker  FLAKE       → re-run triggered (2nd flake this run)

Escalated (1)
  deploy-staging ENV_FAIL   → GITHUB_TOKEN missing in workflow env; human action needed

Previously green: lint, type-check, unit-tests

Edge Cases

• All jobs passing: "CI green on <branch>" — exit immediately.
• No CI configured: Report "No CI runs found for this branch" and exit.
• Private repo without gh auth: Surface the auth error; do not guess.
• Main is perpetually red: After 3 sweep attempts with no improvement, escalate with full timeline.

Token Optimization

Expected range: 400–1,200 tokens per sweep; 50–100 tokens (CI green, early exit)

Patterns used: Bash (gh CLI), early exit (all-green), progressive disclosure (summary then per-job detail), attempt cap (3 fix attempts max)

Caching: Stores per-job attempt counts and flake history in .claude/cache/ci-sweep/ (reset on successful run).

Early exit: All runs passing → single-line report and exit (saves ~1,000 tokens).