Repository navigation
ci: Type Check runs 476 s median on every PR and merge group although its turbo cache is restored — measure the cache hit rate, fix the miss, and type-check affected packages only on PR and merge-queue runs (two-speed CI, objectstack#22998) #12222
Description
Protects: this repository's required Type Check context, the published check-run name every PR and merge group of @object-ui/* passes through; this card keeps the name and cuts its wall.
Filed by the director seat (summon #37, session_01LqNgdPZB2Jnr5HpGbaqLeB) on the maintainer's word 「Type Check 的两张后续卡也立吧」 (2026-10-12), the follow-up objectui#12214 names. Filing gate ①: a measured defect with a named landing point. ⛔ Not a claim. Tier H: .github/workflows/**; the PR stays a draft until a GOVERNED_APPROVERS member approves it.
What is measured (29 completed CI runs, 16 pull_request + 8 merge_group + push, the jobs API, 2026-10-12)
| reading | value |
|---|---|
Type Check job wall min / median / p90 / max |
2 / 10.3 / 12.4 / 13.4 min |
step Run type-check (pnpm type-check) |
476 s median, 620 p90, 680 max (82 % of the job) |
step Restore Turbo Cache |
2 s median (a two-second restore is a small or empty cache) |
the other 30 steps (guards, Type-check scripts/, setup) |
about 90 s together |
The cache keys look right on paper: key: turbo-${{ runner.os }}-type-check-${{ github.sha }}, restore-keys: turbo-${{ runner.os }}-type-check- then turbo-${{ runner.os }}-, saved whenever the exact key missed. turbo.json declares type-check with cache: true, dependsOn: ["^build", "//#type-check:e2e"] and inputs that include $TURBO_ROOT$/scripts/vite-*.ts. Yet the step re-executes for about 8 minutes on every run, so something defeats the replay: a root task (//#type-check:e2e) or a root-level input whose hash changes on every commit, a cache directory that is not .turbo/cache, or pnpm type-check not going through turbo for every package. Which one is this card's first measurement, not its premise.
What to build (one PR, Tier H, draft until approved; lands after objectui#12214's PR, which edits the same file)
- Measure the miss: one PR run with
turbo run type-check --summarize(or the run summary) read for cache hit / miss per package and the hash inputs of a missed package; name the input or task that invalidates the hash in the PR body. - Fix it so unchanged packages replay from the restored cache (narrow the invalidating input, split the root task out of
dependsOn, or point the cache at the directory turbo writes). - Type-check affected-only on
pull_request/merge_group: the packages the diff affects and their dependents (--affectedagainst the merge base, or--filter=...[<base>]), the whole workspace onpushand the hourlyschedule(the doctrine's full battery). Root-level changes (lockfile,tsconfig*.json,turbo.json,package.json) mean everything affected; state the rule in the workflow. - The job keeps its name
Type Check;scripts/check-type-check-coverage.mjsandscripts/check-required-check-set.mjsstay green.
Expected: Run type-check from 476 s to about 60–120 s on a typical PR; after objectui#12214, Type Check is the PR's critical path (10.2 min), so the PR wall follows this card down to the Lint job (6 min).
Acceptance
Type Checkmedian under 4 minutes on three consecutive merge groups and three PRs, read from the jobs API and written in the PR body with the before figures above.- The turbo summary of one PR run shows cache hits for packages the diff did not touch.
Related: objectstack#22998 · objectui#12214 · objectui#12208 · .github/workflows/ci.yml (type-check job) · turbo.json.
Activity
objectstack-fleet commentedon Oct 12, 2026
Claim: PM loop round 1 · the director seat dispatching on the maintainer's standing order, verbatim 「三仓的 ci 优化卡你负责派发」 (this seat's session, 2026-10-12T12:31Z)
Session:
session_01LqNgdPZB2Jnr5HpGbaqLeBAccount:
os-zhuangBranch:
claude/issue-12222-type-check-cache-affectedWorktree:
objectui-issue-12222Domain:
domain:devxSeat:
domain:devx#9File surface:
.github/workflows/ci.yml(thetype-checkjob only: its turbo cache steps, the affected filter and the root-change rule),turbo.json(thetype-checktask'sinputs/dependsOnif the miss is there),scripts/check-type-check-coverage.mjsandscripts/check-required-check-set.mjswith their tests only if a pin must move honestly; ⛔ not thetestjob ofci.yml(objectui#12208 PR-2), ⛔ notspec-main-shape-gate.yml(objectui#12240) (stop on breach; explain in the report)Container & model:
M,mode:subagent,model: opus(the default judgment tier; objectstack'sdispatch-gates --tier --repo objectstack-ai/objectuion the paths answers no path-derived mandate; finding the cache miss is a measurement-and-judgment task). Tier H:.github/workflows/**is a governed surface in this repository; the PR stays a draft until aGOVERNED_APPROVERSmember approves it.Clause-②: no
Responsibility:
n/a — not a defect cardThread-read: none
Ruling-ref: the card (filed on the maintainer's word verbatim 「Type Check 的两张后续卡也立吧」), objectstack-ai/objectstack#22998 (the doctrine), objectstack-ai/objectstack#23017 (the objectstack sibling: a turbo cache that never restores, the same defect shape)
Serial constraints cleared: PR #12232 (objectui#12214, the last PR on
ci.yml) is MERGED (21fdeb34); objectui#12240 (claimed in this act) editsspec-main-shape-gate.ymland the required-set pin only; objectui#12208 PR-2 (thetestjob ofci.yml) is dispatched after this PR lands. Mergeorigin/mainbefore opening the PR.objectstack-fleet commentedon Oct 12, 2026
os-dev-report
{
"issue": 12222,
"status": "done",
"branch": "claude/issue-12222-type-check-cache-affected",
"pr": "#12251",
"session": "session_01LqNgdPZB2Jnr5HpGbaqLeB — the dispatching seat's session (this run is its subagent)",
"premise_still_valid": true,
"summary": "Measured and fixed. The invalidating task is the root task
//#type-check:e2e: every package'stype-checkdependsOn it, it declared noinputs, and a root task hashes every tracked file (8591 on 923504f; 8593 in CI run 1), so every commit moved all 45-46 type-check hashes (local probe: one untracked docs/ file moved 46/46 type-check and 0/36 build hashes; after fcccda1, 0/46). CI run 1 also showed the restore side: the bareturbo-Linux-restore key matched thedocsjob's cache saved minutes earlier on the same PR ref (actions/cache tries every key on the run's own ref beforemain), about 2 GB with 321/401 entries unused; by run 4 nothing was restorable at all (run 3's cache, saved 4 min earlier, and anymaintype-check cache were gone). The real working set is 80 entries / 9.0 MB. PR #12251 (draft, 4 commits): e33ccd3 turbo run summary +Report what the turbo cache replayednotice; fcccda1 root task inputs = its e2e program, plus 8 import-closure files atsc --listFilesOnlysweep found outside package+dependency closure added to type-check inputs; 95368ae two speeds (scripts/select-type-check-scope.mjs: turbo--filter=...[BASE]on pull_request/merge_group, derived root rule widens to the workspace, fail-open; push/schedule run every package with TURBO_FORCE); bcfabaf restore only theturbo-Linux-type-check-prefix, save only on push/schedule after--pruneto what that run used. No CI run on this PR could replay a type-check (run 2 new hashes by construction; run 3 shadowed by another job's cache; run 4 nothing to restore and PR runs no longer save) — CI replay is observable only after landing, once the first push run seeds main's pruned cache; replay is proven locally (@object-ui/types: miss 42.6 s before vs hit 4.1 s after, same docs/ probe).","measured": {
"before_card": "Run type-check 476 s median / 620 p90 / 680 max; Type Check job 10.3 min median / 12.4 p90 (29 runs, jobs API, per the card)",
"run1_e33ccd3c": "CI run 38254636135 job 114829419183: Run type-check 356 s, job wall 470 s (7.8 min), Restore 20 s. Notice: build 30/35 replayed, type-check 0/45, uncached //#type-check:e2e (8593 input files), cache 401 entries 1997.8 MB (321 unused), restored turbo-Linux-docs-339344d7 (another job, same PR ref)",
"run2_fcccda14": "CI run 38255726542 job 114833669552: Run type-check 642 s, job wall 798 s (13.3 min), Restore 18 s. Notice: build 8/35, type-check 0/45 (new hashes, cold by construction), root task 44 input files, cache 473 entries 2004.7 MB (393 unused), restored turbo-Linux-type-check-339344d7 (run 1's)",
"run3_95368ae8": "CI run 38257816387 job 114840337613: scope=workspace (root rule: turbo.json, tsconfig.e2e.json in the diff). Run type-check 311 s, job wall 444 s (7.4 min), Restore 25 s. Notice: build 34/35, type-check 0/45, cache 457 entries 2498.7 MB (377 unused), restored turbo-Linux-6fb281d6 (bare-key match on another job's cache; run 2's type-check entry not restored)",
"run4_bcfabaf2": "CI run 38258825685 job 114844174815: Run type-check 676 s, job wall 797 s (13.3 min), Restore 1 s. Notice: build 0/35, type-check 0/45, cache 80 entries 9.0 MB (0 unused), restored nothing (cold); Prune and Save skipped (pull_request), as designed",
"invalidating_input": "
//#type-check:e2e(root task, cache:false, no inputs => every tracked file) folded into every package type-check hash via dependsOn","hypotheses": "H1 path/save fine, restore returns another job's cache (bare key, own ref searched first) or nothing — remedied in bcfabaf; H2 confirmed (root task); H3 refuted as cause (build hashes stable; builds replay 30/35 and 34/35 when a cache holds them); H4 implemented, root-only change selects 0 packages (measured) so guarded by the derived root rule; H5 refuted (coverage guard reads package scripts, unchanged and green)",
"head_checks": "All check runs on bcfabaf completed success (Type Check, Test aggregate + 12 shards, Lint, Repo Gates, Build & E2E, Build Docs, Changeset Declaration, Governed Surface Queue Guard, Line Citation Gate, Spec Main Shape Gate) as of 14:08Z"
},
"tests": "At HEAD bcfabaf:
pnpm exec vitest runover every test file reading ci.yml / turbo.json / tsconfig.e2e.json plus the two new ones — Test Files 65 passed (65), Tests 2285 passed (2285);pnpm type-check:scriptsexit 0; eslint --no-inline-config over the 8 changed script/test files exit 0. Local turbo proof (shared 4-core box): @object-ui/types type-check before fix: same tree hit 3.1 s, + untracked docs/ file miss 42.6 s (root hash fbb2d654 -> 956fc4f3); after fix: seed miss 53.2 s, + docs/ file hit 4.1 s (root hash 4804aa2f unchanged); TURBO_FORCE=true writes an entry the next TURBO_FORCE=false run replays (389474f8). Ablation (after committing fcccda1, via objectstack scripts/ablation-replace.mjs with trap restore): deleting the root task inputs -> e2e-type-check.test.ts 1 failed | 14 passed (the new pin), restored blob == HEAD 02add335, git diff HEAD empty; first attempt was a no-op (helper refused an empty replacement before running the test), not counted. Single-package local turbo only; fullpnpm type-checkmeasured by CI (4 runs above). NOT MEASURED: actionlint (absent), repo-wide pnpm lint locally (CI Lint green).","gates": "objectstack dispatch-gates.mjs refuses a sister repo by design (exit 2, REFUSING objectstack-ai/objectui), so hand-derived from objectui package.json + workflows, all exit 0 at bcfabaf: check-control-bytes, check-action-ref-convention, check-shell-escape-residue, check-doc-links, check-doc-example-ids, check-doc-component-types, check:new-line-citations, check-type-check-coverage, check-required-check-set --self-test, check:pre-install-import-graph, check:bash32-floor, check:test-path-roots, lint:coverage, check-changeset-presence (no changeset owed), check:entry-guard; ci.yml YAML parse ok; --ran reconciliation NOT MEASURED (script refuses this repo).
--tier --repo objectstack-ai/objectui: no path-derived mandate.","files_changed": [
".github/workflows/ci.yml (type-check job only: Restore keys, Decide which packages to type-check, Run type-check, Report, Prune, Save if)",
"turbo.json (type-check inputs +8 import-closure files; //#type-check:e2e inputs)",
"tsconfig.e2e.json (comment)",
"content/docs/guide/ci-cd-pipeline.md (type-check row + two-speeds paragraph)",
"scripts/report-turbo-cache.mjs (new)",
"scripts/select-type-check-scope.mjs (new)",
"scripts/tests/report-turbo-cache.test.ts (new)",
"scripts/tests/select-type-check-scope.test.ts (new)",
"scripts/tests/e2e-type-check.test.ts (new pin)",
"scripts/tests/turbo-type-check-inputs.test.ts (INPUTS_NOT_DERIVABLE entries)",
"scripts/tests/vitest-setup-type-check.test.ts and vitest-config-type-check.test.ts (step locator moved honestly)"
],
"deviations": [
"Per-commit figures are NOT in the PR body: the dispatch asked for them there, os-dev.md says the dev writes the PR body once and never PATCHes; os-dev.md wins. Seat amendment for #12251 body: (1) "Three commits" -> "Four commits" and add row
bcfabaf2= restore onlyturbo-Linux-type-check-, Save on push/schedule only, new stepPrune the turbo cache to what this run used; (2) fill the Figures table frommeasuredrun1..run4; (3) replace the H1 sentence with the run-1 finding (bare key matched the docs job cache on the PR ref, own ref searched before main; run 4 restored nothing); (4) Two speeds: add "pull-request and queue runs restore and never save".","Four commits, not the suggested three: bcfabaf was added after run 1 read the restore side; it was amended twice locally before its first push (no force-push ever).",
"File surface wider than the claim's literal list (ci.yml type-check job, turbo.json, the two check scripts): two new scripts + tests, tsconfig.e2e.json comment, the CI/CD guide row (AGENTS #2 + the doc command-parity pin), e2e-type-check / turbo-type-check-inputs pins, and two older pins (vitest-setup/vitest-config type-check) that located the step by the literal
run: pnpm type-check. The ci.ymltestjob and spec-main-shape-gate.yml untouched.","First
git pushof the empty branch answeredremote: fatal error in commit_refs(not 403); succeeded on the first retry.","Pushing each commit cancelled the previous CI run's still-running Test shards (workflow concurrency); the Type Check job had completed in each before the next push."
],
"mcp_calls": "0",
"api_writes": "2 relay dispatches (POST /repos/objectstack-ai/objectstack/dispatches, executed as objectstack-fleet[bot]): pr_create = POST /repos/objectstack-ai/objectui/pulls (draft) + POST /repos//issues/12251/assignees; comment = POST /repos//issues/12222/comments (this report). label-write: 0 (objectui labels are path-driven; labeler applied documentation, configuration, ci/cd, tests). git push: 5 (empty branch + 4 commits; not REST).",
"open_questions": [
{
"question": "Every package type-check still inherits the e2e root task hash, and turbo force-adds the root package.json and turbo.json to a root task's inputs (a
!package.jsonnegation is ignored, measured), so a PR touching the root package.json (23 of the last 300 main commits), e2e/ (16) or turbo.json (4) still re-checks every package cold. Decouple?","options": [
"A keep as landed: root task inputs = its e2e program; cost above, no contract change",
"B take //#type-check:e2e out of type-check dependsOn and run
pnpm type-check:e2eas its own ci.yml step (like type-check:scripts), roottype-checkscript runs both; removes the coupling, but changes whatpnpm type-checkruns and moves objectui#4471 pins (e2e-type-check.test.ts)"],
"recommendation": "A now; decide B after landing if the after-landing medians show root-package.json PRs dominating the slow tail"
},
{
"question": "The card files
.github/workflows/**as Tier H here, butnode scripts/check-governed-queue-guard.mjs --test .github/workflows/ci.yml turbo.jsonanswers NOT GOVERNED in objectui. Which is authoritative for #12251?","options": [
"A the guard (ordinary PR; seat lands after CI)",
"B the card (wait for a GOVERNED_APPROVERS approval)"
],
"recommendation": "The PR stays a draft either way (dispatched PR); the seat decides the landing route — flagged, not chosen here"
}
],
"out_of_scope_findings": [
"carrier: 承接者:无 · noted in PR #12251 Acceptance notes, not filed —
linthas the same coupling: uncached root task//#lint:root(default inputs = every tracked file) is inlintdependsOn;turbo run lint --dry=jsonbefore/after one untracked docs/ file moves 46/46 lint hashes, so the Lint job's turbo cache can never replay · dedupe: lint:root turbo hash, lint cache never replays, root task default inputs","carrier: 承接者:无 · noted, not filed — other
.turbo/cachejobs (ci.yml docs job: restore-keysturbo-Linux-docs-then bareturbo-Linux-; lint.yml; the e2e job) keep the bareturbo-${{ runner.os }}-fallback and save multi-GB caches on every run; measured here as the source of the shadowing (run 1, run 3) and the likely eviction pressure (run 4) · dedupe: turbo cache bare restore key, actions cache eviction, multi-GB turbo cache","carrier: 承接者:无 · noted in turbo-type-check-inputs.test.ts and the PR body — that guard derives config root files only; with replay live, an import-closure edge outside a package and its dependency closure is unhashed unless listed (8 listed from a one-off
tsc --listFilesOnlysweep, run without dist/); a derived import-closure guard would keep the list honest · dedupe: type-check inputs import closure, INPUTS_NOT_DERIVABLE, listFilesOnly"]
}
Generated by Claude Code
objectstack-fleet commentedon Oct 12, 2026
Review: ACCEPT · PR #12251 (draft, head
bcfabaf2) · director seat dispatching on the maintainer's order, 2026-10-12T14:20ZChecked on GitHub, not on the report's say-so:
main(923504f4), body line 1Fixes #12222(the only closing keyword), assigneeos-zhuang; 12 files, +1,221/−13, four commits pushed one at a time so each CI run measured one (e33ccd3cmeasure,fcccda14the fix,95368ae8the two speeds,bcfabaf2restore and save); labels path-driven. The body was amended by this seat from the report's four named spots (one body edit, read back identical). No changeset owed (check-changeset-presenceexit 0).//#type-check:e2edeclared noinputs, a root task hashes every tracked file (8,591–8,593), and every package'stype-checkdependsOnit, so every commit moved all 45–46 type-check hashes. Local probe: one untrackeddocs/file moved 46/46 type-check hashes and 0/36 build hashes beforefcccda14, 0/46 after;@object-ui/typesmissed in 42.6 s before and hit in 4.1 s after on the same probe. H2 confirmed; H3 (build misses) refuted as a cause; H5 refuted (the coverage guard reads package scripts, unchanged and green). H1's restore side, found on the PR's own runs: the bareturbo-Linux-fallback matched another job's 2 GB cache on the PR's own ref (searched beforemain), 321 of 401 entries unused; by run 4 nothing could be restored at all. Fixed inbcfabaf2: restore only theturbo-Linux-type-check-prefix, save onpush/scheduleonly after pruning to what the run used (80 entries, 9.0 MB).turbo.json(+11/−2): the root task's inputs are its e2e program; eight import-closure files found by atsc --listFilesOnlysweep join thetype-checkinputs.ci.yml, thetype-checkjob only (+149/−2): the restore keys,Decide which packages to type-check(scripts/select-type-check-scope.mjs:--filter=...[BASE]onpull_request/merge_group, the derived root rule widening to the whole workspace, any selector failure → the workspace;push/scheduleforced over every package),Report what the turbo cache replayed(scripts/report-turbo-cache.mjs, a notice),Prune,Save if. Two new scripts with tests (217 + 193 lines), the e2e pin (e2e-type-check.test.ts, +52),turbo-type-check-inputs.test.ts(INPUTS_NOT_DERIVABLEentries), two older pins that located the step byrun: pnpm type-checkmoved honestly, the pipeline guide's row and paragraph, atsconfig.e2e.jsoncomment. Thetestjob andspec-main-shape-gate.ymluntouched (objectui#12208 and ci(spec-main-shape-gate): drop themerge_grouptrigger once the ruleset reads INTACT on seven — an advisory verdict on a queue build is read by nobody (objectui#12240 follow-up, letter B) #12252 own them).bcfabaf2: all 30 check runs success (Type Check, the Test aggregate and 12 shards, Lint, Repo Gates, Build & E2E, Build Docs, Changeset Declaration, Governed Surface Queue Guard, Line Citation Gate, Spec Main Shape Gate). The dev's local union, exits quoted: vitest over every test file that readsci.yml/turbo.json/tsconfig.e2e.jsonplus the two new ones 65 files / 2,285 tests;type-check:scriptsexit 0; eslint over the 8 changed script/test files exit 0; fifteencheck:*gates exit 0 (the list hand-derived:dispatch-gatesrefuses objectui by design); an ablation of the root task's inputs reds the new pin (1 failed / 14 passed) and is restored by blob hash; NOT MEASURED locally: actionlint (CI's YAML parse and the green workflow stand in), repo-widepnpm lint(CI Lint green).Run type-check356 / 642 / 311 / 676 s against 476 s median before; none replayed a type-check (run 2 cold by construction, run 3 shadowed by another job's cache, run 4 with nothing to restore because PR runs no longer save): the CI replay is first visible after landing, once the firstpushrun seedsmain's pruned cache; the card's acceptance (three merge groups and three PRs under 4 minutes) is this seat's reading after landing.package.json,e2e/orturbo.jsonstill re-checks every package cold: 23 + 16 + 4 of the last 300maincommits; B, decoupling//#type-check:e2efromtype-check, becomes a card only if the after-landing medians show those PRs dominating the slow tail). 2 → the route is the card's: a draft until aGOVERNED_APPROVERSapproval (the round's posture for workflows; objectui#12255 is queued for the table row), and, under the two-speed clause, this lane's PRs do not arm the queue while theTest (coverage)main-red (objectui#12256) stands: this PR lands after objectui#12208 PR-2's first commit lands, or with it.testjob and the spec gate untouched); four commits instead of three; the body written once and amended by the seat; the first push'sremote: fatal error in commit_refs(one retry). Writes: 2 relay strokes (pr_create with assignee; the report 6126065113);mcp_calls0; no label writes.linthas the same coupling (//#lint:rootwith default inputs inlint'sdependsOn: 46/46 lint hashes move on an untrackeddocs/file, so the Lint job's cache never replays): a card for the lane once this lands shows the pattern; the other jobs that use.turbo/cache(docs, lint.yml, e2e) keep the bare fallback and save multi-GB caches every run (the shadowing and eviction measured here); theturbo-type-check-inputsguard derives config root files only, so an import-closure edge outside a package is unhashed unless listed (8 listed).Landing: the round's posture: the maintainer approves; the seat arms the queue, after the main-red fix per the two-speed clause. This card stays
pm:dispatcheduntil the PR lands; the landing record and the six-run reading follow.objectstack-fleet commentedon Oct 12, 2026
LANDED — PR #12251 is MERGED (2026-10-12T14:41Z, approved, readied and queued by the maintainer's own hand while the
Test (coverage)main-red objectui#12256 still stands: the maintainer's call over the two-speed clause, recorded as such; merge commit1ce8db8d) and read in place onorigin/main1ce8db8d. Director seat, summon #37,session_01LqNgdPZB2Jnr5HpGbaqLeB, 2026-10-12T14:42Z.origin/mainturbo.json://#type-check:e2edeclares its owninputsci.ymltype-checkjob: restore keysturbo-<os>-type-check-<sha>and the prefix only; stepsDecide which packages to type-check,Report what the turbo cache replayed,Prune the turbo cache to what this run usedscripts/select-type-check-scope.mjs,scripts/report-turbo-cache.mjsin the treeFixesline;pm:dispatchedstripped and the assignee released in this act.main's firstpushrun after this merge seeds the pruned cache (theReportnotice says what it saved); then three merge groups and three PRs: theRun type-checkstep (476 s median before) and theType Checkjob wall (10.3 min median), with the cache notice's replay counts; the 4-minute line decides whether question 1's B (decoupling the e2e root task) becomes a card.domain:devx@ objectui lane:linthas the same coupling (//#lint:rootwith default inputs inlint'sdependsOn: the Lint job's cache never replays) — a CI card the director proposes to the maintainer; the docs, e2e and lint.yml jobs keep the bareturbo-<os>-fallback and save multi-GB caches every run; theturbo-type-check-inputsguard lists eight import-closure edges from a one-off sweep.Fixes #12222alone.