ci: the merge queue runs the affected package set, not the full list (maintainer-directed, part A of the test-cost programme)

[claude]

Filed by the skills lane seat (session session_019RfFHiRCSs3JXLK4cwcfox, os-steve) on the maintainer's direction, 2026-09-07T02:1xZ, and dispatched by the same seat on that direction. Surface owner stays domain:devx.

Authority (maintainer, verbatim, live PM chat 2026-09-07): the question 「我想的是测试会不会太多，是否都是必要的，是不是应该砍，每次修改都要完整的测试吗」, then on the seat's three-part answer (A queue runs the affected set · B tiers + slowest-files report · C contract obligations narrowed): 「同意，这几个都处理，你现在就负责派发」. This card is part A.

What

merge_group runs of the CI workflow compute the Test Core package set with turbo's --affected against the merge group's base, unioned with the same cross-package scans the pull_request branch already applies, instead of the full package list. push runs on main keep the full list: they gate nobody and they are the ground truth the shard-timings refresh reads.

Measured

• .github/workflows/ci.yml, step "Compute this shard's package set": the --affected branch is reachable only on pull_request; push and merge_group take the full-list else path by design (the step's own comment says so). Cross-package scans are unioned only on the pull_request branch.
• Cost: a full Test Core run is ~85–90 minutes of test compute (six shards on main at 23:22Z / 00:15Z / 00:45Z: 12.1 / 7.5 / 10.8 / 9.8 / 27.7 / 22.1 min and similar). The four governed PRs merged at 01:18Z touched only .claude/** and still ran 15–20-minute shards in their queue builds. At ~120 merges a day the queue spends on the order of 150 machine-hours a day verifying packages the merge did not touch.
• The affected-set machinery on PRs is measured to work: docs-only PRs run six ~1-minute shards.

Ruling

1. On merge_group, TURBO_SCM_BASE is the group's base (github.event.merge_group.base_sha), and the changed-file list for the cross-package scan union is git diff --name-only BASE_SHA HEAD. Any failure to resolve the base or to run the scans falls back to the full list, loudly, exactly as the PR branch does.
2. An EMPTY selection on merge_group is legitimate (a docs-only group has nothing to test) and must produce an honest green through the existing "no packages on this shard" path and the attestation roster; it is not the [finding] Nothing distinguishes "zero packages affected" from "the affected-set computation silently produced nothing" — an empty changed-files.txt is never checked #10057 case, which is PR-only. State this in the step's comment and prove it (see acceptance).
3. push on main: unchanged, full list.
4. Clause-②: yes — the required Test Core check verifies a narrower set at merge time. Contract tier.

Acceptance

• The workflow file's own gates (check:workflow-* families derived by dispatch-gates --commands) green; any script touched carries its self-test.
• Because a merge_group branch cannot run before the PR enters the queue, the PR body must carry a local dry-run of the step's shell with the event fields exported for three synthetic groups: a leaf package change (selects itself + consumers), a packages/spec change (selects the consumer closure), a docs-only group (selects nothing, aggregator verdict green). The PR's own first merge-group run is the live proof; the seat reads its selected set and the Test Core verdict before the card closes.
• Nothing about timings, shard count or the wall moves here (CI: the shard-timings file is stale for the CLI package — 672s predicted vs 28m46s measured against a 30-minute timeout, so Test Core shard 1/6 is one slow run from being killed on any PR touching the CLI #16173 / [temporary] raise Test Core (N/6) timeout-minutes 30 → 45 while #16173's shard balance is unfixed — and un-censor the readings that #16173 needs #16445 own those).

Refs #16173.

---

Generated by Claude Code

[claude]

Claim: PM loop round 9 — flight V, maintainer-directed (「同意，这几个都处理，你现在就负责派发」, 2026-09-07): the merge queue's Test Core runs the affected package set. merge_group computes the set with turbo --affected against the group's base and the same cross-package scan union the pull_request branch applies; push on main keeps the full list; an empty selection on a docs-only group is an honest green through the existing no-packages path. The selection logic moves out of the inline step into a script with a self-test so the three synthetic groups can be proved locally; the PR's own first queue run is the live proof, read by the seat.
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-16453-merge-group-affected-set
Worktree: objectstack-issue-16453
Domain: domain:devx (dispatched by the skills seat on the maintainer's direction)
File surface: .github/workflows/ci.yml (the "Compute this shard's package set" step only) and one new script under scripts/ci/ with its self-test; scripts/check-shard-attestation.mjs only if an empty roster does not already attest as judged. ⛔ No timings, shard count or wall change.
Container & model: M, mode:subagent, fable — Clause-② yes (the required Test Core check verifies a narrower set at merge time); on a fable 429 the seat's exemption applies, opus, recorded here.
Clause-②: yes
Serial: .github/workflows/ci.yml is shared with flight X (#16455, a different step); whichever pushes second proves git merge-tree --write-tree clean against the other's pushed head, twice (the second with -c merge.os-regen.driver=false). Not governed ⇒ in-seat review, then the seat flips ready and arms SQUASH auto-merge; skip-changeset. The PR carries Fixes #16453.

---

Generated by Claude Code