ci-cd
Use when PR merge waits are long or CI/CD is being designed, audited, or sped up: targets, PR/main/nightly tiers, opt-in labels, roll-forward, merge queue, release and privileged workflows.
- License 许可
- MIT
- 最近更新
- 2026-10-20
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- ci-cd
- license
- MIT
- description
- Use when PR merge waits are long or CI/CD is being designed, audited, or sped up: targets, PR/main/nightly tiers, opt-in labels, roll-forward, merge queue, release and privileged workflows.
03 SKILL.md 正文
CI/CD tiers and PR merge time
Learned in a large monorepo: the median open→merge time was about 30 minutes, and the gate took 6–11 minutes. Fix the loop first, then the slowest leg.
Targets (evidence)
- Merge-blocking gate ≤ 10 min p90: Fowler's 10-minute build.
- Open→merge median ≤ 15 min for auto-merge PRs. DORA elite lead time is < 1 h, and the LinearB elite merge time is < 1 h.
- Small changes: elite PR size is < ~90 changed lines. Google's median review is < 4 h with one reviewer.
Measure first
- Gate wall-clock:
gh run list --workflow <gate> --event pull_request --status success -L 40 --json createdAt,updatedAt. Report p50, p90 and max. - Open→merge:
gh pr list --state merged -L 30 --json createdAt,mergedAt. Report the median. - Runs per PR and the failure share. Re-push loops usually cost more than slow legs do.
- Per-job durations plus queue time. The critical path is the slowest leg, not the sum.
Three tiers
| Tier | Runs | Rule |
|---|---|---|
| PR (blocking) | path-filtered affected-module build + unit tests, lint/guards, required governance | only what MUST block; light |
| Push to main | the same path filters (tests for unrelated modules never run), never cancelled, plus slow OS/E2E legs | a red leg opens or updates one tracker issue, which auto-closes on green |
| Nightly/scheduled | full cross-OS acceptance, live/external suites, security scans, EAP verifiers | tracker issue, no Closes keyword |
| Release | full verify, always | never weakened |
Move a leg off PR only when all three hold:
- it is > 5 min or is the critical path;
- it caught no PR-only regression in its last ~20–30 runs;
- the main tier runs it.
Keep a live-label opt-in such as ci:<leg>, plus ci:full, and re-run the gate when the label is added. Read labels from the API, because re-runs replay the old event payload.
Quality layers are never removed
Tests, static analysis, coverage and review may be rearranged or added, but never removed. The one exception is a layer that repeated, evidenced analysis proves an utter waste. Moving a layer, for example from PR to main or from blocking to reporter, must keep equal or better coverage:
- a blocking in-repo equivalent;
- findings resolved before merge;
- a tracker issue on main.
A new or changed gate is done only after it ran on one real pass case and one real fail case; a coverage gate no PR of its kind had exercised blocked every such PR.
Never auto-revert
Main rolls forward. A failing post-merge leg opens or updates one rolling issue (P0, assigned to the owner), and the next green run closes it. Do not open revert PRs automatically.
First push green
- Run the cheap CI tests locally before every push: workflow README/inventory, timeout, path-filter and PR-body validators.
- Check the PR body against the repo's required sections, and name a targeted test, before creating the PR.
- A body edit does not re-run a PR-body guard that reads the event payload. Push a commit instead.
- Contract tests that CI runs with plain
unittestmust import only the standard library. A module installed on the dev box but not on the runner, such as PyYAML, passes locally and fails in CI. - Normalize paths with
removeprefix("./"), neverlstrip("./").lstripalso strips the dot from.github/, so path-triggered pre-push checks silently stop running. - Before babysitting,
git ls-remote origin "$BRANCH"must equalgit rev-parse HEAD. An unpushed local commit leaves CI testing the old head. - Run
git statusbefore each push and check for unexpected deletions. A box restore can drop tracked files under cache-like directory names (build/,coverage/) in every worktree; restore them withgit ls-files -d -z | xargs -0 git checkout --. - Never delete a worktree or branch before checking that the cherry-pick or merge landed (
git status,git log). If you already did, recover the commit withgit fsck --lost-found. - Validate every third-party CI config (coverage, analysis) against the vendor's validator in pre-push and CI. Some vendors silently ignore a whole file over one unknown key, and fall back to defaults.
- Coverage carryforward reads the vendor's record of the git parent. An old branch whose parent chain lacks the full flag set never goes green. Move its commits to a fresh branch from main, and make every main push upload every flag.
- A gate that reads an external status, such as Codecov, must wait for a settled verdict (an unchanged title and completion time across polls). The vendor recomputes per upload, so a mid-upload red is not final.
Release and privileged workflows
- New release safety is additive: a hard gate before publishing, and
continue-on-errorfor anything after publishing, such as attestation or SBOM upload, so it can never strand a published release. - An API-compatibility baseline must be the published artifact downloaded as a file. Comparing against the same version number resolves to the reactor jar, so the build compares the jar with itself. Give the comparator any missing optional dependencies precisely; never set a global ignore-missing-classes, which hides real breaks.
pull_request_targetandworkflow_runjobs never check out or run PR code. Use two workflows instead: the unprivilegedpull_requestjob computes a result and uploads an artifact. The privilegedworkflow_runjob re-reads the PR from the API, validates the artifact against an allow-list, applies it as data and pushes.
Delivery order, queue, stacks
- One PR at a time: babysit it to merged, then open the next. With a merge queue,
gh pr merge --auto --mergeprints "merge strategy is set by the merge queue" but may still enable auto-merge. ConfirmautoMergeRequestis set. If the PR is green and not queued, call the GraphQLenqueuePullRequestmutation: its error names the exact blocker, such as a cancelled or expected required check. - A code-scanning alert on changed lines opens a review conversation. Under a resolve-conversations rule, it blocks the enqueue even when the scanner check is not required. Fix the finding, or, if it is accepted by design, reply with the reason and resolve the thread. Resolving does not dismiss the alert.
- A required check must never share a concurrency group across runs. GitHub keeps only one pending run per group, so a burst of PR events cancels the queued one, and a cancelled required check blocks the queue. Use a per-run group (
github.run_id) and no cancellation. - Merge queue: add
merge_group:to every workflow that reports a required check, and run the PR tier there. External apps that do not report onmerge_groupmust leave the required list first. - Stacked PRs: run cheap checks on every layer, and expensive legs only on the lowest unmerged layer or the top of the stack (
github.event.pull_request.stack).
Report
Before/after p50/p90 gate time, the open→merge median, the PRs merged, and the admin changes left for the owner (rulesets, merge queue).
04 同仓库更多 Skills
ponytail
Lazy senior dev mode: the smallest change that fully solves the task, and a reply a busy human understands in one read. Use on any coding task (writing, fixing, refactoring, reviewing, choosing dependencies) and when the user says "ponytail", "be lazy", "simplest solution", "yagni", or complains about over-engineering or bloat. Levels: lite, full (default), ultra.
work-item
Use when opening or rewriting a work item on any git-based SCM. Source-control agnostic; GitHub, GitLab, and Azure Boards are adapters only.
self-improve
Use when running ChaosEngine Learning Session self-improve. Harness lessons are GitHub issues only. Product lessons may queue after delivery or on request.
omniroute
Use when local-runtimes selected the OmniRoute process for bounded work: runner, delegate continuity, capability enforcement and proof of dispatch.
local-runtimes
Use when choosing an optional local inference runtime for a narrow mechanical or offline job.
local-openai-compat
Loopback OpenAI-compatible peers (Ollama, LM Studio, llamacpp on 127.0.0.1). Use when a narrow task should go to an already READY loopback server.