ci-integration（CI 集成）

CI Integration

Value: Feedback -- CI pipelines produce signals. Deterministic interaction patterns ensure those signals are received, classified, and acted on correctly. Undisciplined CI interaction (pushing over failing runs, ignoring flaky tests) degrades the signal until the pipeline is noise.

Purpose

Teaches disciplined CI/CD interaction: one pending run at a time, structured failure triage, automated self-healing for mechanical failures, and structured evidence output for pipeline consumption. Prevents the most common CI failure mode: pushing again before understanding why the last run failed.

Practices

Push-and-Wait Discipline

One pending CI run at a time. Never push while a run is in progress.

1. Push the commit
2. Poll or wait for the CI run to complete
3. Read the full result before taking any action
4. Only after the run completes (pass or fail) may you push again

Do:

• Wait for CI completion before starting new work that would require a push
• Read the complete CI output, not just the status badge

Do not:

• Push a "fix" while the previous run is still pending
• Assume a run will pass and push follow-up commits
• Cancel a run to push a new one (unless the run is clearly stale)

Failure Triage

Classify every CI failure before attempting a fix. The classification determines the fix strategy.

Read the full CI log to classify. Do not guess from the status alone.

Self-Healing

Mechanical failures that have deterministic fixes should be resolved automatically:

1. Lint/format failures: Run the project's formatter or linter with auto-fix, commit the result, re-push. This is a single retry -- if the formatter does not resolve the issue, classify as build-failure.
2. Infra failures: Retry the CI run (max 2 retries). If all retries fail, escalate to the user with the infra error details.
3. Test failures: Never auto-retry. Route to debugging-protocol for investigation. A test failure is a signal, not noise.
4. Build failures: Check dependency lockfiles, build configuration, and environment differences. Do not blindly retry.

Flaky Test Detection

A test is flaky if it passes on retry without any code changes.

When detected:

1. Flag the test with its name, file, and the failure output from the flaky run
2. Record in project memory (.factory/audit-trail/flaky-tests.json if factory mode is active, or project notes otherwise)
3. Report to the user -- flaky tests erode CI trust and must be addressed
4. Do not treat a flaky pass as a real pass for quality gate purposes until the flakiness is resolved

Structured Output

Every CI interaction produces a CI_RESULT evidence packet:

{
  "run_id": "string (CI run identifier)",
  "status": "string (passed | failed | cancelled)",
  "duration": "number (seconds)",
  "failure_type": "string (test-failure | lint-failure | build-failure | flaky-test | infra-failure) -- omit if passed",
  "failure_details": "string (summary of failure) -- omit if passed",
  "fixes_applied": ["string (description of each auto-fix applied)"],
  "retry_count": "number (0 if no retries)"
}

This packet is consumed by pipeline orchestrators and audit trails. Always produce it, even for passing runs.

Enforcement Note

• Pipeline mode: Gating. CI pass/fail mechanically gates merge.
• Standalone mode: Advisory. The agent self-enforces push-and-wait discipline.

Hard constraints:

• Push-and-wait: [H] in pipeline mode, [RP] in standalone
• Flaky pass not counted as real pass: [H]

Constraints

• Push-and-wait: One pending run means one. Not "one per branch" or "one per type of change." If a CI run is in progress, do not push regardless of how unrelated the change seems. The discipline exists because CI resource contention and merge ordering matter.
• Failure classification: Classify BEFORE attempting any fix. Not "I tried a fix and it worked, so it must have been a lint failure." The classification determines the fix strategy, not the other way around. If you can't confidently classify, investigate further -- don't guess and fix.
• Flaky test: A flaky test is one that produces different results on the same code. "Passes on retry without code changes" is the definition. If you changed the environment, the retry is not evidence of flakiness -- it's evidence of an environment-dependent test. Those are different problems with different fixes.

Verification

After each CI interaction, verify:

• Waited for the previous CI run to complete before pushing
• Read the complete CI output (not just status)
• Classified the failure type before attempting a fix
• Applied the correct fix strategy for the classification
• Did not retry a test failure (routed to debugging-protocol instead)
• Lint/format auto-fixes were limited to one attempt
• Infra retries did not exceed 2
• Flaky tests were flagged and recorded
• Produced a CI_RESULT evidence packet

If any criterion is not met, revisit the relevant practice before proceeding.

Dependencies

This skill works standalone for any project with a CI/CD pipeline. It integrates with:

• debugging-protocol: Test failures route to the debugging protocol for systematic investigation rather than blind fix attempts
• tdd: CI failures during TDD cycles feed back into the RED-GREEN loop; a CI test failure means the cycle is not complete
• pipeline: When used within factory mode, the pipeline orchestrator consumes CI_RESULT packets to evaluate quality gates

Missing a dependency? Install with:

npx skills add jwilger/agent-skills --skill debugging-protocol