pipeline-eng
Skill8 starsChanged 5 months agoMITView on GitHub
Reference and checklist for building and maintaining CI/CD workflows, inter-workflow communication, dispatch scripts, and cloud agent infrastructure. Use when modifying any workflow YAML, dispatch script, or sentinel-based communication.
- Reads credentials
In this repository31 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- accelerateSkill
- implement-issueSkill
- pipeline-orchestrateSkill
- pipeline-peSkill
- pipeline-scorerSkill
- qa-acceleratedSkill
- qa-prSkill
- qa-webSkill
- releaseSkill
- review-feedbackSkill
- roadmap-prioritizationSkill
- roadmap-refinementSkill
- security-auditSkill
- accelerateSkill
- implement-issueSkill
- pipeline-engSkill
- pipeline-orchestrateSkill
- pipeline-peSkill
- pipeline-scorerSkill
- qa-acceleratedSkill
- qa-hitlSkill
- qa-prSkill
- qa-webSkill
- releaseSkill
- review-feedbackSkill
- roadmap-prioritizationSkill
- roadmap-refinementSkill
- screenshotsSkill
- security-auditSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. About 0 are common, 4 are distinctive, and 9 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
9
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 17
When invoked with a specific file (e.g., /pipeline-eng pipeline-triage.yml), focus guidance on that file's triggers, sentinels, tokens, and …
- Distinctive: line 52
Before modifying any workflow file or dispatch script:
- Very specific: line 76
When a job posts a comment that triggers another workflow, the token identity determines the actor filter — github-actions[bot] vs …
- Very specific: line 89
Testing strategy: Every issue_comment-triggered workflow should also accept workflow_dispatch with a pr_number input.
- Very specific: line 108
Always use the
's'flag: when parsing sentinel JSON from comment bodies. - Very specific: line 143
Use AbortSignal.timeout() for per-request deadlines
- Very specific: line 151
When Anthropic releases new model versions, update all three files.
- Very specific: line 161
When dispatching cloud agents (via /accelerate or manual gh workflow run):
- Distinctive: line 164
Before dispatching multiple agents in parallel, check for file overlap.
- Distinctive: line 171
Verify the target branch exists and is up to date
- Very specific: line 172
Check that no other agent is already working on the same PR (look for recent <!-- agent-fix-attempt: sentinels)
- Very specific: line 173
Use secrets.PROJECT_TOKEN (not github.token) for cross-workflow dispatch
- Distinctive: line 197
After modifying any workflow or script, verify:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Pipeline Engineering
Institutional knowledge for building and maintaining the Prose CI/CD pipeline. This is not about running the pipeline (see pipeline-scorer, pipeline-pe, pipeline-orchestrate) — it's about modifying it safely.
Usage
/pipeline-eng — full reference
/pipeline-eng <file> — targeted guidance for a specific workflow or scriptWhen invoked with a specific file (e.g., /pipeline-eng pipeline-triage.yml), focus guidance on that file's triggers, sentinels, tokens, and downstream consumers.
Pipeline Map
The full workflow chain, trigger to output:
| Workflow | Trigger | Reads sentinel | Writes sentinel | Dispatches to |
e2e.yml | pull_request, /test comment | — | — (artifact: e2e-results) | — |
web-e2e.yml | pull_request (with accelerated label), /test comment | — | — | — |
ci-gate.yml | workflow_run on "E2E Tests" | <!-- e2e-fix-attempt:, <!-- agent-fix-attempt:, <!-- ci-gate-sha: | Posts /review <!-- ci-gate-sha: ... --> comment; <!-- e2e-fix-escalation --> | Codex.yml (via /review comment) |
Codex.yml (auto-review) | /review comment on PR | — | <!-- review-verdict: clean --> or <!-- review-verdict: issues-found --> | — |
review-feedback.yml (clean-gate) | issue_comment with <!-- review-verdict: clean --> | <!-- review-verdict: clean -->, absence of <!-- review-feedback-analysis --> | <!-- review-feedback-analysis --> | — |
review-feedback.yml (analyze) | issue_comment with ## Code Review (no verdict trailer) | absence of <!-- review-feedback-analysis -->, <!-- review-verdict: clean -->, <!-- review-verdict: issues-found --> | <!-- review-feedback-analysis -->, <!-- pipeline-bypass-warning --> | — |
pipeline-triage.yml (security-gate) | issue_comment with <!-- review-verdict: issues-found --> | <!-- security-gate: true --> in review body | <!-- orchestrator-verdict: hitl-full --> (short-circuit) | — |
pipeline-triage.yml (scorer + PE) | issue_comment with <!-- review-verdict: issues-found --> (security gate NOT detected) | <!-- review-verdict: issues-found -->, absence of <!-- scorer-output: / <!-- pe-output: | <!-- scorer-output: {...} -->, <!-- pe-output: {...} --> | — |
pipeline-triage.yml (orchestrate) | needs: [run-scorer, run-pe-analysis] | <!-- scorer-output: {...} -->, <!-- pe-output: {...} --> | <!-- orchestrator-verdict: auto-fix|auto-fix-verify|hitl-light|hitl-full --> | — |
pipeline-fix.yml | issue_comment with <!-- orchestrator-verdict: auto-fix --> or auto-fix-verify | <!-- orchestrator-verdict: ... -->, absence of <!-- agent-fix-attempt: AND <!-- e2e-fix-attempt: | <!-- agent-fix-attempt: 1 -->, <!-- agent-fix-escalation --> | — |
dispatch.yml | /triage, /fix, /pipeline comment | — | — | pipeline-triage.yml, pipeline-fix.yml, or posts /review |
notify.yml | PR labeled needs-review or complex | — | — | Slack webhook |
Sync hazard: run-pe-analysis.mjs and Codex.yml auto-review both hardcode the same privilege-boundary path list (src/main/**, src/preload/**, electron-builder.*, electron.vite.config.*). If you update one, update the other. validate-pipeline.sh invariant #20 enforces this.
Sync hazard: Codex.yml emits and pipeline-triage.yml greps the <!-- security-gate: true --> sentinel. If either side drifts, the short-circuit silently breaks and privilege-boundary PRs waste API calls on scorer+PE. validate-pipeline.sh invariant #19 enforces this.
Dual /test trigger: Posting /test on a PR fires BOTH e2e.yml (Electron Playwright) and web-e2e.yml (browser Playwright) in parallel. This is intentional — it runs the full test matrix. For pull_request events, only e2e.yml runs automatically; web-e2e.yml only runs on accelerated-labeled PRs.
Unified circuit breaker: Both ci-gate.yml and pipeline-fix.yml count both <!-- e2e-fix-attempt: and <!-- agent-fix-attempt: sentinels in their fix-attempt guards. validate-pipeline.sh invariant #18 enforces this.
Pre-Flight Checklist
Before modifying any workflow file or dispatch script:
- Token choice — Which token does each
ghcall / API call use? Does it need to cross a workflow boundary? (See Tokens section) - Trigger model — Will this workflow fire on the correct branch? Can it be tested on a feature branch? (See Triggers section)
- Loop risk — Does the comment this workflow posts contain text that could re-trigger itself or an upstream workflow?
- Sentinel consumer — Who reads the sentinel this workflow writes? Will your change break their regex?
- Permissions — Does the
permissions:block grant exactly what's needed, nothing more? - Parser correctness — If parsing JSON from a sentinel, is the regex using the
's'(dotAll) flag? - Circuit breaker — Is there a guard preventing infinite re-execution (attempt counter, sentinel-absence check)?
Tokens & Identity
| Token | Identity when posting | Can trigger other workflows? | Use case |
github.token | github-actions[bot] | Yes for issue_comment triggers | Default for most in-workflow API calls |
secrets.PROJECT_TOKEN | The PAT owner (human) | Yes | Cross-workflow dispatch, label ops in orchestrate, ci-gate /review posting |
secrets.ANTHROPIC_API_KEY | N/A (API auth) | N/A | Codex API calls in scorer, PE, auto-fix, auto-review |
Rules:
github.tokencannot triggerworkflow_dispatchin another workflow — usesecrets.PROJECT_TOKENorgh workflow runwith a PATid-token: writeis for OIDC federation only (used byCodex-action) — do not add it unless the job uses OIDCactions: writepermission is required forgh workflow rundispatch calls- When a job posts a comment that triggers another workflow, the token identity determines the
actorfilter —github-actions[bot]vsCodex[bot]vs a human username. Mismatched identity breaksif:conditions downstream.
Trigger Model
| Trigger | Runs from branch | Testable on feature branch? | Notes |
pull_request | PR head (merge ref) | Yes | Standard |
issue_comment | Default branch (main) | No — always runs main's version | Must merge workflow changes before testing |
workflow_run | Default branch | No — same as issue_comment | ci-gate.yml always runs from main |
workflow_dispatch | Specified branch | Yes — can target any branch | Escape hatch for testing issue_comment workflows |
Testing strategy: Every issue_comment-triggered workflow should also accept workflow_dispatch with a pr_number input. This lets you test workflow changes on a feature branch via gh workflow run <workflow> --ref <branch> -f pr_number=<N>. All current pipeline workflows follow this pattern.
Sentinel & Loop Guards
Three failure modes
1. Reading the trigger comment instead of the response
A workflow triggers on issue_comment containing sentinel A, then reads comments looking for sentinel B. If the search isn't scoped, it may match the triggering comment itself instead of a downstream response. Always filter by author and/or createdAt > trigger time.
2. .match() returning the first instead of the last sentinel
Multiple bot comments may contain the same sentinel prefix (e.g., after a re-run). String.match() returns the first match. When you need the latest, iterate comments in reverse chronological order, or use matchAll and take the last result.
3. dotAll flag mismatch with embedded JSON
Sentinel JSON may span multiple lines if pretty-printed or if the comment body contains newlines before the sentinel. The run-orchestrator.mjs parser uses the 's' (dotAll) flag on its regex for this reason:
const regex = new RegExp(`<!-- ${prefix}: (\\{.*?\\}) -->`, 's')Always use the 's' flag when parsing sentinel JSON from comment bodies.
Existing guard patterns
Absence-based loop prevention (preferred):
# pipeline-triage.yml — don't re-run scorer if output already exists
if: >-
contains(github.event.comment.body, '<!-- review-verdict: issues-found -->')
&& !contains(github.event.comment.body, '<!-- scorer-output:')Attempt counter (for fix workflows):
# pipeline-fix.yml — check if fix already attempted
# Shell step counts <!-- agent-fix-attempt: occurrences across all commentsSentinel-in-own-output (always prepend a unique loop-prevention sentinel):
run-scorer.mjsprepends<!-- scorer-output-comment -->to its outputrun-pe-analysis.mjsprepends<!-- pe-output-comment -->analyze-review-feedback.mjsprepends<!-- review-feedback-analysis -->
contains() through backticks
In GitHub Actions if: expressions, sentinel strings inside contains() must be wrapped in single quotes. If the sentinel itself contains single quotes, you'll need to use a different guard strategy (e.g., shell step with grep).
Robustness
API error handling
The current scripts (run-scorer.mjs, run-pe-analysis.mjs, analyze-review-feedback.mjs) make a single fetch() call with no retry. They exit with code 1 on non-OK responses. If you add retry logic:
- Handle HTTP 429 (rate limit) and 529 (overloaded) with exponential backoff
- Wrap the entire operation in a
totalTimeoutto prevent zombie jobs - Use
AbortSignal.timeout()for per-request deadlines
Model ID management
Model IDs are hardcoded in each script:
run-scorer.mjs:Codex-sonnet-4-6run-pe-analysis.mjs:Codex-opus-4-6analyze-review-feedback.mjs:Codex-sonnet-4-6
When Anthropic releases new model versions, update all three files. Consider using an environment variable override (e.g., process.env.SCORER_MODEL || 'Codex-sonnet-4-6') to allow CI-level model pinning without code changes.
The validate-pipeline.sh script checks for stale model IDs — update its patterns when adding new models.
Codex-action@v1 model pinning. The action does not expose a model: input — it delegates to the bundled Codex CLI, which the action rolls forward on every release. CLI defaults can change silently (the May 8–9 default shift to Opus 4.7 sent daily CI spend from ~$2 to $36 on identical workload). Always pass --model <id> via claude_args on every Codex-action@v1 invocation. Current pin: Codex-sonnet-4-6 in Codex.yml (auto-review + @Codex handler) and pipeline-fix.yml.
Agent Dispatch
When dispatching cloud agents (via /accelerate or manual gh workflow run):
File overlap
Before dispatching multiple agents in parallel, check for file overlap. Two agents editing the same file will create merge conflicts. Use integration branches or sequential dispatch for overlapping scopes.
The "implementer never merges" rule
The agent session that implements changes should NOT merge the PR. A separate review step (human or automated) must validate before merge. This is enforced by the pipeline structure: auto-fix agents push commits, then the pipeline re-runs review.
Pre-flight for dispatch
- Confirm the target workflow accepts
workflow_dispatchwith the expected inputs - Verify the target branch exists and is up to date
- Check that no other agent is already working on the same PR (look for recent
<!-- agent-fix-attempt:sentinels) - Use
secrets.PROJECT_TOKEN(notgithub.token) for cross-workflow dispatch
NEVER Rules
These are absolute prohibitions — violations cause subtle, hard-to-debug pipeline failures:
- NEVER add
id-token: writeunless the job uses OIDC federation (currently onlyCodex-actionjobs). It's not a general-purpose permission. - NEVER use
github.tokento post comments that must trigger another workflow.github.tokenevents are suppressed from triggering further workflows by design. Usesecrets.PROJECT_TOKEN. - NEVER use
.match()for sentinel extraction without considering multiple matches. A PR with re-runs will have multiple comments containing the same sentinel. Either iterate in reverse or use the's'-flagged regex on the last matching comment. - NEVER edit privilege-boundary paths in only one location. The list appears in both
Codex.yml(auto-review prompt) andrun-pe-analysis.mjs. Update both or neither. - NEVER dispatch overlapping agents without an integration branch strategy. Two agents editing
src/main/ipc.tssimultaneously will produce irreconcilable conflicts. - NEVER trust workflow run status as proof a PR was created. Cloud agents complete "successfully" even when they fail to create branches or PRs. Always verify with
gh pr list --head <branch>.
Review Checklist
After modifying any workflow or script, verify:
- Token audit — Every
gh/curl/fetchcall uses the correct token. No accidentalgithub.tokenwherePROJECT_TOKENis needed (or vice versa). if:condition targeting — Conditions match the correct comment author (github-actions[bot],Codex[bot], or human) and sentinel.- Sentinel parsing — All regex patterns use the
's'(dotAll) flag when parsing JSON from comment bodies. Test with multi-line input. workflow_dispatchfallback — Everyissue_comment-triggered workflow also acceptsworkflow_dispatchwithpr_numberfor testability.- Path list sync — If privilege-boundary paths changed, both
Codex.ymlandrun-pe-analysis.mjsare updated. - Attempt guards — Fix workflows have attempt counters and escalation paths. A failing agent cannot retry infinitely.
- Permissions minimized — The
permissions:block grants only what the job needs. Nowrite-allor leftover permissions from copy-paste.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

