Commentaire d'échec de CI sur PR

Câbler les échecs de workflows Playwright dans les commentaires de pull request via une action composite, avec nettoyage des secrets et déduplication par marqueur.

Spar Skills Guide Bot
DevOpsAvancé
2006/08/2026
Claude Code
#ci#playwright#github-actions#pr-comment#gitleaks

Recommandé pour


name: post-failure-comment description: Wire a workflow's Playwright (or other captured-log) failures into the PR conversation as a marker-tagged, secret-scrubbed comment via the .github/actions/post-failure-comment composite action. Use when adding a new long-running/Playwright workflow, when a workflow's failures are opaque to agents iterating on a PR, when the user asks to "make CI failures self-report", or when refactoring inline gitleaks-scrub-and-comment blocks into the shared action.

Failure-comment composite action

Why this exists

Workflow logs in this repo are not directly readable by the Claude agent (no gh CLI, no logs surface in the GitHub MCP server, actions/runs/.../logs returns 403 unauthenticated). When a long-running workflow fails on a PR, the agent has to either ask the user to paste the log or open the run in a browser.

The composite action at .github/actions/post-failure-comment/action.yml solves this: it scrubs the captured log via gitleaks and posts (or updates, via marker-based dedup) a PR comment with the failure block. PR comments arrive as <github-webhook-activity> events the agent can read directly, and are also fetchable via mcp__github__pull_request_read.

The action is self-contained: its extractor and scrubber scripts are co-located in the action's own directory and resolved via $ACTION_PATH (${{ github.action_path }}), so the action travels with whatever repo checks the platform out — there is no dependence on a scripts/ dir at the consuming site's root.

Built once; every Playwright-running workflow should call it.

Caller convention

The action is mode-driven with caller-side gating. Two call sites per workflow: one for the failure post, one for the green-run resolve. The action itself just does what its mode input tells it to — it does not try to detect job state.

Single-job workflow (most common)

The comment step lives in the SAME job as the Playwright run.

- name: Post failure summary
  if: ${{ failure() && github.event_name == 'pull_request' }}
  uses: ./.github/actions/post-failure-comment
  with:
    mode: post
    log-file: /tmp/<your-log>.log
    marker: <unique-marker-slug>     # NO `<!-- -->` — the action wraps it
    title: <short label>             # shown in the comment heading

- name: Resolve failure summary on success
  if: ${{ success() && github.event_name == 'pull_request' }}
  uses: ./.github/actions/post-failure-comment
  with:
    mode: resolve
    marker: <unique-marker-slug>
    title: <short label>

Multi-job workflow

The comment step lives in a DOWNSTREAM job (e.g. finalize after an e2e matrix). failure() / success() reflect the FINALIZE job's state — not the matrix's — so gate on needs.<job>.result instead:

- if: ${{ needs.e2e.result == 'failure' && github.event_name == 'pull_request' }}
  uses: ./.github/actions/post-failure-comment
  with: { mode: post, log-file: /tmp/playwright-output.log, marker: e2e-failure-summary, title: E2E tests }

- if: ${{ needs.e2e.result == 'success' && github.event_name == 'pull_request' }}
  uses: ./.github/actions/post-failure-comment
  with: { mode: resolve, marker: e2e-failure-summary, title: E2E tests }

Non-pull_request triggers

For workflow_dispatch (or any event without a github.event.pull_request context), pass the parent PR explicitly and drop the github.event_name part of the if::

- if: failure()
  uses: ./.github/actions/post-failure-comment
  with:
    mode: post
    log-file: /tmp/<your-log>.log
    marker: <unique-marker-slug>
    title: <short label>
    pr-number: ${{ inputs.pr_number }}

- if: success()
  uses: ./.github/actions/post-failure-comment
  with:
    mode: resolve
    marker: <unique-marker-slug>
    title: <short label>
    pr-number: ${{ inputs.pr_number }}

Gating the call further

For a workflow whose runtime gate matches the upload-artifact step (e.g. vars.PROD_PLAYGROUND_MODE == 'true'), match the if: on each call to the upload step's gate so the comment behaviour is consistent with the rest of the failure-handling chain. For example:

if: ${{ failure() && vars.PROD_PLAYGROUND_MODE == 'true' }}
if: ${{ success() && vars.PROD_PLAYGROUND_MODE == 'true' }}

Patterns that DON'T work — don't repeat

  • ${{ job.status }} in with: — silently expands to empty string inside a composite action's with: block. Two iterations confirmed: no comment ever lands.
  • failure() / success() inside the action's own step if: — unreliable in composite contexts; whatever GitHub Actions evaluates failure() against inside a composite step doesn't reflect the calling job's state in our case.

The caller-side gating in the patterns above is the only approach that has been empirically confirmed to fire (PR #517 finalize job at run 25561744480 — first successful auto-comment after switching to v3 mode-driven design).

Required upstream conditions

  1. Capture the log. The Playwright invocation must 2>&1 | tee /tmp/<your-log>.log. The action reads from this path verbatim. If the log is missing or empty, the action emits a "(no log captured)" comment instead of failing.

  2. Pin a unique marker. The action wraps <your-marker> as <!-- your-marker -->. Two workflows MUST NOT share a marker — they will clobber each other's comments. Existing markers in use:

    • e2e-failure-summarye2e-tests.ymlfinalize (aggregates the e2e matrix)
    • unit-failure-summarye2e-tests.ymlunit
    • e2e-real-failure-summarye2e-tests.ymle2e-real
    • parity-failure-summarye2e-tests.ymlparity
    • select-failure-summarye2e-tests.ymlselect
    • host-loop-failure-summarycms-publish-loop-host.yml
    • prod-mutate-failure-summarycms-publish-loop-prod.yml
    • preview-loop-failure-summarycms-publish-loop-preview.yml
  3. Check out the platform so the action is on disk. The action shells out to its co-located $ACTION_PATH/extract-playwright-failures.sh and $ACTION_PATH/scrub-secrets.js ($ACTION_PATH = ${{ github.action_path }}). The scripts ship inside the action's own directory, so you do NOT need a scripts/ dir at the consuming repo's root — but the action's directory must be present, which means the workflow must check out the repo/path that contains .github/actions/post-failure-comment/ (e.g. the reusable e2e-tests.yml checks the platform out into .cms-platform/ and references the action as ./.cms-platform/.github/actions/post-failure-comment). You still want actions/checkout for the site so the captured log path exists, but the extractor + scrubber no longer depend on a site-root scripts/ dir.

  4. Grant pull-requests: write to the workflow. The default GITHUB_TOKEN works for posting comments on the same repo, BUT only if the workflow's permissions: block explicitly grants it. Without this, the embedded actions/github-script call 403s silently — the workflow log shows the error but no comment is posted, which can fool a casual review. The composite action does NOT require CMS_E2E_PAT. Minimum block:

    permissions:
      contents: read
      pull-requests: write
    

What the action does internally

| Step | Gate | Notes | |---|---|---| | Install gitleaks | outcome == 'failure' | Installs to $HOME/.local/bin (no sudo). Adds to $GITHUB_PATH so subsequent steps see it. | | Extract & scrub | outcome == 'failure' | Runs $ACTION_PATH/extract-playwright-failures.sh; falls back to tail -c 80000 if the extractor finds nothing. Scrubs via $ACTION_PATH/scrub-secrets.js (gitleaks). Truncates to 60 KB. | | Post comment | outcome == 'failure' AND (pull_request event OR pr-number set) | Marker-based update-or-create. | | Resolve on success | outcome == 'success' AND (pull_request event OR pr-number set) | Replaces an existing failure comment with a "passing on <sha>" stub. |

The action no-ops on cancelled, skipped, or null outcomes — those rarely indicate a real failure that warrants a comment.

Security: env vars, not interpolation

Inputs to the embedded actions/github-script calls are passed via env: and read as process.env.X, not inlined as ${{ inputs.x }} into the script body. This is the pattern the github-script README explicitly requires; inlining input strings into a JS body is a classic script-injection vector. If you extend the action with new fields, follow the same pattern.

Security: gitleaks pass-through is non-optional

Every comment that lands on a PR via this action runs through $ACTION_PATH/scrub-secrets.js (which shells out to gitleaks detect --no-git --source <log>) inside the action's Extract and scrub failure summary step. There is no caller-side switch to disable it.

If you extend the action with new modes or new emit-paths, keep the scrubber call on every code path that emits log content into a comment body. A leaked PAT, AWS key, or token in failure output that bypasses gitleaks would be visible to anyone with read access to the PR (which on a public repo means the open internet, including search-engine indexers). The platform's secrets-scan.yml workflow prevents secrets from reaching git history; the gitleaks pass-through here is the equivalent guard for everything that reaches a PR comment.

Triple-check this when:

  • Adding a new mode: value (e.g. digest, summary, anything that emits log content).
  • Refactoring the extract / post sequence (don't accidentally hoist the github-script call above the scrubber).
  • Adding a new fallback path inside the action's Extract and scrub failure summary step.

Common refactor pitfalls

  • Missing pull-requests: write permission — the most subtle and most common. The default GITHUB_TOKEN is read-only unless explicitly elevated; without elevation the comment-post 403s silently and the workflow looks like it ran the comment step but produced no comment. ALWAYS add pull-requests: write to a workflow's permissions: block when wiring this action in. Verified by listing PR comments after a deliberately-broken push (see "Testing your wiring" below).
  • The action's directory isn't on disk — happens when slotting the action into a one-step workflow that never checks out the repo/path holding .github/actions/post-failure-comment/. The extractor + scrubber are co-located in that directory and resolved via $ACTION_PATH, so the action dir itself must be present (a uses: ./... path can't resolve otherwise). In the reusable platform model the harness checks the platform out into .cms-platform/ and references ./.cms-platform/.github/actions/post-failure-comment. A separate site actions/checkout is still wanted so the captured log path exists.
  • Overlapping markers — copy-pasting from another workflow without changing the marker. Always pick a globally-unique slug.
  • Wrong outcome source${{ job.status }} is the canonical source for single-job workflows. For workflows with multiple dependent jobs (like e2e-tests.yml's post-job), pass ${{ needs.<job>.result }} instead — and remember needs.<job>.result only reflects ONE upstream job. If the workflow has multiple matrix jobs that should all gate the comment, AND-combine them in an intermediate step.
  • Skipping the gate on pull_request — for workflows that ALSO fire on schedule or workflow_dispatch, the action's if: will skip posting (no PR context) but the gitleaks install still runs. To save runtime, add && github.event_name == 'pull_request' to the caller's if: or use pr-number for explicit-target workflows.

When NOT to use this

  • Non-Playwright workflows with short logs. The extractor matches Playwright's --reporter=list blocks; for other tooling it falls back to tail -c, which is usually less useful than the workflow's own native error output. If your workflow already prints a clear single-line error, a comment with 60 KB of tail noise is worse than no comment.
  • Workflows with no PR context AND no parent PR. schedule-only workflows (e.g. canary-prod.yml) have no PR to comment on. Open an issue instead, which is what canary-prod.yml already does in its if: failure() branch.
  • One-off / hotfix workflows. The marker dedup convention requires a stable name; ad-hoc workflows would pollute the PR with stale "passing" stubs.

Testing your wiring

  1. Push a deliberately-broken commit (e.g. add await page.waitForTimeout(99999999) to a spec).
  2. Watch for the workflow to fail.
  3. Check the PR for a new <!-- your-marker --> comment — it should arrive ~30s after the workflow ends.
  4. Push a fix that turns the workflow green.
  5. The comment should refresh to ## <title> passing on <sha> within ~30s of the green run.

Files

All three live together in the action's directory and travel with it (resolved at runtime via $ACTION_PATH = ${{ github.action_path }}):

  • .github/actions/post-failure-comment/action.yml — the composite action.
  • .github/actions/post-failure-comment/extract-playwright-failures.sh — awk-based extractor for Playwright --reporter=list failure blocks.
  • .github/actions/post-failure-comment/scrub-secrets.js — gitleaks-backed scrubber. Replaces detected secrets in-place with <REDACTED:RuleID>.
  • AGENTS.md "Failure-comment composite action" — short reference; this skill is the long-form.
Skills similaires