[rwgk]

Description

Restacking a chain of PRs currently starts CI for every rewritten PR, even when the author only wants to update the stack. Introduce an opt-in ci:intentional label that suppresses the 29 automatic GitHub Actions test jobs on participating PRs, independently of draft/readiness state. Existing conditions still apply to other PRs and to scheduled, branch-push, release, and manual runs.

Add a small experimental entry point for /ci full, /ci quick, and /ci help. During the pilot, any PR author can request runs on their own open PRs carrying ci:intentional. The listener reads the live PR state, captures the head/base SHAs, and dispatches a fixed workflow on intentional-ci-experiment. It never checks out PR code.

This is a bootstrap, not the finished CI implementation: the dispatched worker is an explicit placeholder that runs no tests and publishes no PR CI result. After this merges, I will create the shared intentional-ci-experiment development branch from master and iterate on the worker there without requiring review of each experimental step. The listener and ordinary workflows on master can remain stable. I will submit the completed design and any proposed project-wide switch together in a final PR.

The pilot documentation explains setup, separate experimental result/concurrency/cache names, the initial-label race, and existing-stack rebasing. The opt-out covers Actions test jobs; AppVeyor, Read the Docs, and the preview-link/label-maintenance workflows are separate.

Validation

• Full pre-commit run --all-files passed, including GitHub workflow schema validation.
• actionlint passed for all seven changed/new workflows.
• GitHub's @actions/expressions evaluator passed 6,496 condition cases and eight listener event/authorization cases, covering labeled/unlabeled PRs, draft/ready states, ready_for_review, python dev, and non-PR events.
• An independent baseline audit confirmed that all existing conditions and workflow definitions are preserved apart from the opt-out.

The final head has no CI skip instruction, so this update starts normal PR CI. Live comment-to-dispatch behavior can only be exercised after the entry point exists on the default branch; it has not been tested on GitHub yet.

[henryiii]

Fable 5.1 review:

🤖 AI text below 🤖

.github/workflows/ci.yml
● 9 [correctness] Removing the ci:intentional label never re-triggers CI (types lack labeled/unlabeled), so a PR can go from labeled to merged with zero test runs; the doc tells authors when to add the label but not that removal requires a new push or close/reopen.
● 36 [simplification] The identical two-line label guard is copy-pasted into 26 jobs across 6 workflows; any job added later without it (or any typo in one copy) silently runs for labeled PRs and nothing catches it.

.github/workflows/intentional-ci.yml
● 53 [security] The listener lets any PR author trigger a workflow_dispatch of the intentional-ci-experiment branch; workflow_dispatch runs on an in-repo branch receive repository secrets and whatever token permissions that branch's YAML declares, so the 'read-only, no secrets' rule in intentional-ci.md is only a convention the listener cannot enforce.
● 21 [usability] Recognized-but-rejected commands exit silently with a green job: startsWith() in the expression is case-insensitive and prefix-only, while the JS regex is exact, and the author/label/state check just returns, so the user gets no feedback at all.
● 4 [efficiency] issue_comment: created with a job-level if creates a workflow run record for every comment on every issue and PR in pybind/pybind11, not only /ci comments.
● 70 [usability] A 404/422 from createWorkflowDispatch is reported only in the listener's Actions job (pull-requests: read means no PR comment/reaction), so the PR author who posted /ci full has no signal that the request failed or succeeded.

.github/workflows/format.yml
● 25 [design] Skipping the pre-commit Format job on labeled PRs removes the only GitHub-side lint gate while giving nothing back: pre-commit.ci already runs on every PR regardless of label, so the opt-out saves about a minute and the clang-tidy job is the only expensive part.

.github/intentional-ci.md
● 22 [style] 'Any PR author is welcome to join the pilot.' is mannered prose and 'retaining'/'retain' (lines 4-5, 17) are non-Simplified-Technical-English verbs, per the global CLAUDE.md convention.

The review of PR #6218 (master...HEAD, 6 commits, 8 files) found 8 issues, reported above. The two that matter most:

• Label removal never re-triggers CI. A PR can be labeled, pushed with all jobs skipped, unlabeled, and merged with zero test runs. Either document "push or close/reopen after removing the label" or add unlabeled to the trigger types.
• Any PR author can dispatch the experiment branch, and those runs have secrets. The "read-only, no secrets" rule is only a convention. Pin the worker job to permissions: {} and consider a ruleset on the branch.

The rest are feedback gaps in the /ci listener (silent exits, no PR-side signal on dispatch failure), the 26-copy label guard, the skipped-run noise from issue_comment, the pre-commit opt-out being pointless since pre-commit.ci still runs, and some wording in intentional-ci.md.

[henryiii]

Also, I'd be fine to replace our draft/non-draft mechanism with this, even as an experiment.