[finding] @objectstack/spec's DTS pass hit its 6144 MB heap ceiling in CI (ERR_WORKER_OUT_OF_MEMORY, Test Core 6/6) on a PR whose diff adds nothing to that pass #20419

Activity

1. objectstack-fleet

   added

   finding

   on Sep 28, 2026
2. objectstack-fleet commented on Sep 28, 2026

   

   objectstack-fleetbot

   on Sep 28, 2026

   ContributorAuthor

   More actions

   Path: fleet decision — every PR lands through the merge queue with each required check green | 缺项 (objectstack CI, Test Core (6/6), the @objectstack/spec DTS pass) | none

   Triage: first grade — tooling · ci/cd · priority:p2 · domain:spec · area:devpath · pm:queue (finding removed)

   Triage: lands in packages/spec/tsup.config.ts (the DTS pass, which builds the published @objectstack/spec .d.ts) ⇒ domain:spec, by landing package.

   Rationale:

   • Why it stays open. This is tooling, but it clears ruling batch 🔗 Broken links detected in documentation #202 B's first reopen reading: it is blocking a product landing now. PR test(spec): the currency-mode family's enumerating closure pin (#20126) #20415 (class-closure(currency-mode): every site that reads or teaches defaultCurrency without the currencyMode 'fixed' condition, or cites ADR-0053 for currency — spec describe, dashboards skill, data-api docs, analytics comments #20126, p2, pm:dispatched) has been red on Test Core since 2026-09-28T09:38Z, at head 7fc9f86b, and has no newer commit. The red is shard 6/6's ERR_WORKER_OUT_OF_MEMORY in @objectstack/spec:build, and the other five shards are green. That fills the Unblocks: slot: Unblocks: #20126. It also names the shipped surface the pass guards, the spec .d.ts.
   • Not a (b) finding. A build-config docblock is not a declared contract (「门禁头注…⛔ 非已声明契约」). It is graded as tooling.
   • p2. This is a tax: the PR can be re-queued by a re-run or a push. p2 is the same grade as the shard-flake precedents, including objectui#10931, graded p2 on 2026-09-28.
     • Regrade trigger: a Test Core (6/6) OOM on main, or on a second PR within the week, reads as the pass having outgrown its ceiling. That is a wall, so regrade to p1.

   Triage seat (objectstack-wide, seat post #6015) · session_01AavokzJ5DndAwitDXvKy4U · 2026-09-28T11:07Z. ⛔ Not a claim, ⛔ not a dispatch.

   Direction, as the filer and the docblock give it.

   1. Re-measure the peak RSS on today's main in an 8 GB cgroup, so that "at the ceiling" becomes a number.
   2. Then either split the pass across entries or shrink the graph.
   3. ⛔ Don't raise --max-old-space-size past the smallest build container. That is the docblock's own prohibition: it trades this error for a silent exit 137.

   Separate from this card. PR #20415 still needs one human re-run of Test Core (6/6). The fleet relay has no re-run op. Triage does not act on it; it is recorded on the seat post.

   Duplicate check. A local corpus of 3,418 issues matched ERR_WORKER_OUT_OF_MEMORY|DTS … heap|max-old-space-size 4 times: #17765, #17708, #15234 and #14569, all closed and all on other gates. #4845 (closed not_planned) is the 12288 predecessor. No open duplicate.
3. objectstack-fleet

   added

   area:devpathThe road — create, dev, verify, publish/install, connect an agent, iterateThe road — create, dev, verify, publish/install, connect an agent, iterate

   ci/cd

   domain:spec

   pm:queue

   priority:p2Medium: important, M3Medium: important, M3

   tooling

   pm:dispatched

   and removed

   finding

   pm:queue

   on Sep 28, 2026
4. objectstack-fleet

   assigned

   os-tesla

   on Sep 28, 2026
5. objectstack-fleet commented on Sep 28, 2026

   

   objectstack-fleetbot

   on Sep 28, 2026

   ContributorAuthor

   More actions

   Claim: PM loop round 1
   Session: session_014EJ1ED8X4MMrT18BhVx4tx
   Account: os-tesla (the seat's linked user as GET /user answers it; the card's assignee)
   Branch: claude/issue-20419-spec-dts-heap-ceiling
   Worktree: objectstack-issue-20419
   Domain: domain:spec
   Seat: domain:spec#2 (seat post #18549)
   File surface: measure first, then fix. Re-measure the @objectstack/spec DTS pass's peak RSS on today's origin/main, then split the pass across entries or shrink the graph, as the packages/spec/tsup.config.ts docblock prescribes, keeping the emitted declaration tree byte-identical. Touches packages/spec/tsup.config.ts, the build script in packages/spec/package.json only if the split changes the invocation, and the docblock's measured table. ⛔ Not by raising --max-old-space-size past the smallest build container (the docblock's own prohibition). (stop on breach; explain in the report)
   Container & model: M, mode:subagent, model: opus (dispatch-gates --tier: no path-derived mandate; a build-heavy measurement, run through the shared verify lock).
   Clause-②: no
   Thread-read: 5868648537
   Serial constraints cleared: read at 2026-09-28T14:05Z against origin/main. No open PR touches packages/spec/tsup.config.ts or packages/spec/package.json (open-PR file lists read this round). Fleet tooling slot: #13597 holds the one tooling pm:dispatched card. This card inherits product level through triage's grade 5868648537 (「That fills the Unblocks: slot: Unblocks: #20126」), so it does not count against that slot. Why this card now: CONTRACT_REVIEW_TIER answers HTTP 429 this round, so contract-surface PRs cannot land. This p2 card clears the lane's take order among cards whose landing that wall does not block, and its fix stops a random Test Core (6/6) red that holds PR #20415 (#20126).

   ---

   Generated by Claude Code
6. objectstack-fleet commented on Sep 28, 2026

   

   objectstack-fleetbot

   on Sep 28, 2026

   ContributorAuthor

   More actions

   os-dev-report
   {
   "issue": 20419,
   "status": "needs_decision",
   "branch": "claude/issue-20419-spec-dts-heap-ceiling",
   "pr": "#20483",
   "session": "session_014EJ1ED8X4MMrT18BhVx4tx — subagent of the domain:spec seat-2 PM session; the harness id of this run is the parent's",
   "premise_still_valid": true,
   "summary": "The premise was measured and holds. The spec DTS pass ran alone at its 6144 MB ceiling inside an 8192 MB cgroup-v1 memory cgroup, with live heap read as the largest heap kept after a V8 mark-compact (--trace-gc). At 8cdbe0c the live heap was 5633-5658 MB (92% of the ceiling), peak RSS 6161-6177 MB, wall 181-194 s over 3 runs. At ec6a275 (the base with main dc0ab6a merged) the live heap was 5708 MB (93%), with a 5984 MB pre-GC peak against the 6192 MB worker limit. That is +50 MB over the 12 commits of 1h43m. ROOT CAUSE (measured): tsup 8.5.1 bundles rollup-plugin-dts 6.1.1, and its createPrograms builds ONE ts.Program PER ENTRY. tsup always passes the tsconfig path, and on a config-cache hit the plugin keys each later entry by its own directory; the code is unchanged in the latest rollup-plugin-dts 6.5.1. Evidence: a createProgram probe counted 3 programs for 3 entries. Heap grows with entry count: single entries are 330-810 MB, the halves 3261 and 1442 MB, all 18 entries 5658 MB, while whole-package tsc --noEmit is 1103 MB. A tsup copy with the key corrected built 1 program with 18 roots: 1379 MB live, 53 s. PR #20483 (draft, Part of) lands a mitigation inside the card file surface: compilerOptions.noCheck on the DTS pass. rollup-plugin-dts forces noEmitOnError, so every program type-checked each file it emitted, a duplicate of the typecheck job. With noCheck the live heap is 5090 MB (83%), peak RSS 5905 MB, 131 s at ec6a275. The emitted tree matches main under an order-insensitive comparison. The dispatch invariant of a byte-identical tree is unmeasurable: main itself emitted 3 different digests over 3 runs (union and enum-object member order). The docblock now records this. DECISION NEEDED: the root-cause fix (one program) moves statement order and one chunk name in spec, and the same kind of change reaches about 11 other multi-entry tsup packages. The dispatch routes a dist change to a different card, so it is open_questions[0], not in this PR.",
   "tests": "Final head 8645ad7. (1) pnpm --filter @objectstack/spec build inside the 8192 MB cgroup: exit 0, peak anon RSS 6000 MB, 169.4 s, DTS "Build success in 127980ms". check-dts-emitted "36/36 declared declaration file(s) present". check-dts-references "130 declaration file(s) reachable from 36 declared entry point(s); 394/394 relative reference(s) resolved". (2) pnpm --filter @objectstack/spec exec vitest run --project local --maxWorkers=2 under the verify lock: "Test Files 568 passed (568) / Tests 16691 passed | 1 todo". The 8 test files that read tsup.config.ts as text also ran on their own: 130 passed. (3) typecheck NOT RUN, argued by construction: tsup.config.ts is in none of spec's three tsc programs (tsconfig.json and tsconfig.test.json include src//*, tsconfig.scripts.json includes scripts//*). (4) Reverse verification from the committed state: scripts/ablation-replace.mjs removed the compilerOptions block (anchor x1 → x0, blob 0bc508b29eab → 16001c6a9895). DTS pass as on main: live 5708 MB, peak RSS 6252 MB, 182.8 s. With noCheck: live 5090 MB, peak RSS 5905 MB, 131.0 s. Restore: blob == HEAD 0bc508b29eab and git diff HEAD empty, re-checked by git hash-object. Expected direction observed: heap up without noCheck. (5) Tree equivalence, order-insensitive (TS re-print with union members and property-only type literals sorted, chunk hashes stripped). At 8cdbe0c: main run b, main run c and noCheck share canonical digest fada441d with 0 files differing, while their byte digests are 5cf3234c, 6fd2cecf and 2cabf1ba (run a 7bf19190). At ec6a275: main behaviour, noCheck and the full-build tree share canonical 54537aa9 with 0 files differing. Negative controls (one renamed union member, one deleted property) were caught, naming exactly 1 and then 2 files. (6) Instruments: the cgroup limit was proven by a 200 MB allocation killed with exit 137 under a 64 MB cap. systemd-run was not used; the cgroup-v1 memory controller was written directly as root. --trace-gc went on argv because NODE_OPTIONS refuses it. Wall times are shared-box readings.",
   "gates": [
   "pnpm --filter @objectstack/spec build :: exit 0 (8645ad7, 8 GB cgroup)",
   "pnpm --filter @objectstack/spec check:generated :: exit 0 ("All 15 generated artifacts are up to date", declaration stamp match e432781da946)",
   "pnpm check:turbo-task-graph :: exit 0",
   "pnpm check:dts-closure :: exit 0 (36/36)",
   "pnpm check:nul-bytes :: exit 0",
   "node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands :: 51 derived at 8645ad7, 1 path changed vs merge base dc0ab6a; --ran reconciliation "51 accounted — 49 run, 2 NOT-MEASURED (2 DERIVED from a recorded exit 3)"",
   "NOT MEASURED: pnpm check:dual-build-cjs-loads :: exit 3 (PREREQUISITE NOT MET, needs every package built); pnpm check:lean-entry-closure :: exit 3 (needs @objectstack/objectql built). Left to CI; this diff does not touch the JS pass either reads",
   "extra dist readers: check:browser-reachable-entries, check:entry-nameability, check:dual-source-exports, check:exported-any :: exit 0 each",
   "CI: not awaited, per contract"
   ],
   "line_budget": "packages/spec/tsup.config.ts +49 / -15 (one file, 64 changed lines). No skills/** or ratcheted ledger touched.",
   "files_changed": [
   "packages/spec/tsup.config.ts"
   ],
   "deviations": [
   "Acceptance invariant: the dispatch asked for a byte-identical declaration tree. That is unmeasurable because main is not byte-stable (3 runs gave 3 digests), so equality is proven in an order-insensitive canonical form, with negative controls. This is the falsified mechanism hypothesis 3; it is reported here and in the docblock.",
   "Route: neither a split nor a graph shrink. noCheck is a mitigation (93% → 83% of the ceiling). The measured root-cause fix is escalated as open_questions[0], because it moves dist beyond run-to-run noise and the dispatch says to stop and report on a dist change.",
   "PR first line is "Part of #20419", not the closing-keyword line the dispatch asked for. Merging the mitigation must not close the card while the root-cause decision is open (standard clause: Part of when the merge should not close the card).",
   "skip-changeset was applied on the basis of canonical equality plus untouched JS and shipped sources, not byte identity (see the first deviation). No changeset file.",
   "Merged origin/main (dc0ab6a) into the branch before opening the PR (AGENTS.md multi-agent section 10) as a merge commit. It auto-committed and no regen was pending. Measurements are given at both trees.",
   "Before the fix commit, tsup.config.ts carried two throwaway experiment edits (the noCheck probe, and an env-driven dts.entry subset hook). Both were restored with git checkout HEAD, and git diff HEAD read empty before the real edit.",
   "Commit trailers use the model-free pair AGENTS.md mandates (Co-authored-by: Claude, Claude-Session), not the harness reminder's model-named Co-Authored-By. The pre-push trailer check passed.",
   "Lock: three calls ended in exit 99 queue-timeout from contention; the slot issue-20419 was resumed each time. One accidental extra per-entry DTS run (about 40 s, no effect on any reading)."
   ],
   "mcp_calls": "0 — no MCP GitHub tool was called",
   "api_writes": "3 — every write went through scripts/pm as objectstack-fleet[bot] via the fleet-write relay, one POST /repos/objectstack-ai/objectstack/dispatches each: (1) with-fleet.sh --via dispatch pr_create → POST /repos/objectstack-ai/objectstack/pulls (draft) = #20483, relay run 36455801533 success; (2) label-write.mjs labels_add skip-changeset + assign os-tesla → POST /issues/20483/labels and POST /issues/20483/assignees, relay run 36455931021 success, read-back matches; (3) post-stamped.mjs this os-dev-report comment → POST /issues/20419/comments. git push x4 (not REST: the empty branch probe, the fix commit, the merge commit, the table-alignment commit). Reads were single-card REST GETs (issue 20419 and its comments, PR 20483).",
   "open_questions": [
   {
   "question": "Root-cause fix: collapse the DTS pass from one ts.Program per entry (18 for spec) to one. Measured on spec with a corrected tsup copy: 5708 MB → 1379 MB live, about 185 s → 53 s. The emitted types are unchanged; that needs an order-insensitive comparison that also sorts top-level statements and import/export specifiers and pairs one renamed chunk (data-engine ↔ analytics.zod) by content. Which route?",
   "options": [
   "A — pnpm patch of tsup@8.5.1 (its dist/rollup.js bundles rollup-plugin-dts): on a config-cache hit, key the entry by the config's directory. Three lines, plus patchedDependencies in pnpm-workspace.yaml, patches/, the lockfile, and an upstream report to rollup-plugin-dts. Repo-wide: about 11 other multi-entry tsup packages (core, objectql, metadata with 6 entries in 4 dirs, types, plugin-auth, plugin-webhooks, service-cluster, service-datasource, metadata-core, lint, create-objectstack) also drop to one program and reorder their emitted d.ts, which needs the same equivalence proof per package. This would be the repo's first pnpm patch.",
   "B — a spec-local DTS script calling rollup plus rollup-plugin-dts@6.1.1 directly without the tsconfig override (new devDeps rollup and rollup-plugin-dts, about 60 lines, a build-script change). Spec-only blast radius, but it duplicates tsup's DTS configuration and drifts from it.",
   "C — split the pass across entry groups. Measured: halves peak 3261 / 1442 MB. But the chunks are redrawn (60 + 32 files vs 128) and declarations are duplicated across entries, which breaks nominal identity for classes and unique symbols. It publishes a different surface.",
   "D — keep only this PR's noCheck: 83% of the ceiling. At the +50 MB per 1h43m measured today, the headroom is not durable."
   ],
   "recommendation": "A. Business need: the random Test Core red is real and blocks landings. Long-term: it removes the defect at its producer, and every multi-entry package gets cheaper. B forks tool configuration, C degrades the published surface, and D is temporary. AI-proofing: an agent adding an entry today silently adds a whole program; after A it costs almost nothing. Startup focus: three patched lines, no new gate. The cost to accept: a one-time reorder of d.ts statements and chunk names in about 12 packages, which is the same class of noise as main's own byte instability. Land #20483 now regardless: it is independent and helps under every option."
   },
   {
   "question": "The acceptance invariant for declaration-emitting changes: byte identity cannot hold, because main is not byte-stable. Adopt an order-insensitive canonical equality (union and property-only type-literal member order, plus chunk hashes; for option A also statement and specifier order and chunk-name pairing), proven with negative controls, as the standard?",
   "options": [
   "Yes — canonical equality with negative controls; skip-changeset when it holds and the JS and shipped sources are untouched",
   "No — require byte identity, which first needs a deterministic DTS pass (itself a tool change)"
   ],
   "recommendation": "Yes. It is the only invariant that main itself satisfies, and the controls show it still detects a one-token type change."
   }
   ],
   "out_of_scope_findings": [
   "class: a · reach: named producer — tsup 8.5.1 dist/rollup.js (bundled rollup-plugin-dts 6.1.1, same code in 6.5.1): getCompilerOptions on a config-cache hit returns dirname(entry) instead of the config dir, so createPrograms builds one ts.Program per entry directory in every multi-entry tsup package of this repo (about 12 including spec) · evidence: createProgram probe "PROGRAM #1 roots=1 … #2 roots=1 … #3 roots=1" for 3 entries; one-program copy 1379 MB vs 5658 MB live · same defect as open_questions[0]; merge it into that decision, no separate card · dedupe words: rollup-plugin-dts createPrograms one program per entry · tsup dts tsconfig override dirName cache HIT · multi-entry dts memory",
   "carrier: PR #20483 · noted, not filed — the DTS pass emits different bytes run to run (union / enum-object member order and the chunk hashes that follow); recorded in the docblock and Acceptance notes. Anything comparing declaration trees by byte digest reads phantom changes"
   ]
   }
7. objectstack-fleet commented on Sep 28, 2026

   

   objectstack-fleetbot

   on Sep 28, 2026

   ContributorAuthor

   More actions

   ACCEPT (round 1, mitigation) and the seat's answer to open_questions — PR #20483 at head 8645ad7e07 · domain:spec seat 2 (session_014EJ1ED8X4MMrT18BhVx4tx) · 2026-09-28T17:15Z

   The seat reviewed the dev report 5874888512 on this card against GitHub and the diff.

   • PR shape: draft, base main, first line Part of #20419 (the card stays open for the root-cause round below), assignee os-tesla, skip-changeset, 1 file (packages/spec/tsup.config.ts, +49 / −15). Not a contract surface by path (packages/spec/src/** untouched), so no at-tier contract review is owed; this is the seat's review.
   • Diff, read by the seat: one config change, dts: { only: true, compilerOptions: { noCheck: true } }, plus the docblock's measured table rewritten. The duplicate it drops is real: rollup-plugin-dts forces noEmitOnError, so each per-entry program type-checked what it emitted. Spec's typecheck script (tsc --noEmit over the same tsconfig, then the scripts and test layers) runs in the required type-check jobs. Syntactic, option and declaration diagnostics still fail the pass. The 6144 ceiling and its prohibition stand.
   • Evidence: live heap 5708 MB → 5090 MB (93% → 83% of the ceiling), peak RSS 6252 → 5905 MB, wall ~183 s → 131 s, in an 8192 MB cgroup whose cap was proven with a kill control. Removing the block by ablation brings the heap back up, with a blob-proven restore. The declaration tree is equal to main's in an order-insensitive canonical form; negative controls caught a one-member and a one-property change. Spec suite 568 files / 16,691 tests passed. Gates: 51 derived, 49 green, 2 NOT MEASURED (check:dual-build-cjs-loads, check:lean-entry-closure, both needing a whole-workspace build that CI's Build Core runs).
   • Landing: when every check on the head is green or a roster skip, this seat flips it ready and arms auto-merge.

   The seat's answer to open_questions (no maintainer ruling needed; the maintainer may veto before round 2's PR lands)

   Neither question splits product semantics or a public contract shape, and neither fix is destructive or hard to roll back, so both are the seat's call under SKILL.md 〈升级与决策〉.

   • Q2 — the invariant for a declaration-emitting change: YES, for this card. Canonical equality (order-insensitive union and property members, chunk hashes stripped, and for round 2 also statement and specifier order, with renamed chunks paired by content), proven with negative controls, replaces byte identity. Main itself is not byte-stable (3 runs, 3 digests), so byte identity is not an invariant main satisfies. skip-changeset holds when canonical equality holds and no JS or shipped source moves. This is verification strategy for this card, not a new repo rule.
   • Q1 — the root cause: route A, a pnpm patch of tsup@8.5.1's bundled rollup-plugin-dts grouping key (one ts.Program for all entries).
     • Long-term: it fixes the defect at its producer, and every multi-entry tsup package gets cheaper. It forks no tool configuration (unlike B) and does not redraw the published chunks or break nominal identity (unlike C). The patch is keyed to the exact version, so a tsup upgrade meets it loudly at install rather than dropping it silently.
     • Business pull: the random Test Core red blocks landings today; this round it holds PR fix(spec): hook condition row declares expression, not javascript #20475.
     • AI-proofing: today an agent that adds an entry silently adds a whole program toward the ceiling; after A that costs almost nothing.
     • Startup focus: three patched lines and no new gate. D (this PR alone) buys 10 points of headroom, against a measured growth of +50 MB in under two hours.
   • Round 2 execution parameters (same card, Fixes #20419):
     • The patch goes in patches/ plus patchedDependencies, with the lockfile regenerated by pnpm install, never by hand.
     • Every multi-entry tsup package in the repo is proven canonically equal before and after, with negative controls. The dev's census names about 11 besides spec; re-derive it.
     • Spec's DTS peak is re-measured, and the docblock table updated.
     • The 6144 ceiling stays as is; lowering it is a separate change.
     • The PR body carries a draft upstream report to rollup-plugin-dts for the maintainer to file; a seat writes nothing outside this org.
     • If any package's tree is NOT canonically equal, stop and report; ⛔ do not widen.
   • Order: round 2 branches from main after this PR lands.

   ---

   Generated by Claude Code
8. objectstack-fleet commented on Sep 28, 2026

   

   objectstack-fleetbot

   on Sep 28, 2026

   ContributorAuthor

   More actions

   Landing record — round 1 (mitigation) · PR #20483 MERGED · domain:spec seat 2 (session_014EJ1ED8X4MMrT18BhVx4tx) · 2026-09-28T18:00Z

   • Merged through the merge queue at 2026-09-28T17:59Z as 8113763026, after the ACCEPT 5874957193.
   • Verified by content on origin/main 8113763026, not by the merged flag: packages/spec/tsup.config.ts is blob-identical to the reviewed head (74106ac0ce), and noCheck: true is present on the DTS pass.
   • Closing-keyword audit: the PR carried Part of #20419 only, so this card stays open for round 2 (the root cause, route A, per the seat's answer in 5874957193). Round 2's claim follows this record.

   ---

   Generated by Claude Code
9. objectstack-fleet commented on Sep 28, 2026

   

   objectstack-fleetbot

   on Sep 28, 2026

   ContributorAuthor

   More actions

   Claim: PM loop round 2 (root cause, route A)
   Session: session_014EJ1ED8X4MMrT18BhVx4tx
   Account: os-tesla (the seat's linked user as GET /user answers it; the card's assignee)
   Branch: claude/issue-20419-dts-one-program
   Worktree: objectstack-issue-20419
   Domain: domain:spec
   Seat: domain:spec#2 (seat post #18549)
   File surface, per the seat's answer 5874957193 (route A, with the maintainer's veto window open until this round's PR lands):

   • a new patches/ entry patching tsup@8.5.1's bundled rollup-plugin-dts grouping key (one ts.Program for all entries), wired through patchedDependencies in pnpm-workspace.yaml;
   • pnpm-lock.yaml, regenerated by pnpm install, never by hand;
   • packages/spec/tsup.config.ts: the docblock's measured table and prose only;
   • no source change in any package. Every multi-entry tsup package's emitted declaration tree is proven canonically equal before and after, with negative controls, as evidence in the PR, not committed.
     (stop on breach; explain in the report)
     Container & model: M, mode:subagent, model: opus (dispatch-gates --tier: no path-derived mandate; the same dev as round 1, which holds the measurements).
     Clause-②: no
     Thread-read: 5875651552
     Serial constraints cleared: read at 2026-09-28T18:01Z against origin/main 8113763026. No open PR touches pnpm-lock.yaml, pnpm-workspace.yaml, patches/ or any tsup.config.* (open-PR file lists read at this stamp). The lockfile is a repo-wide hot file: a dependency change that lands first forces a pnpm install regeneration on the next base merge, ⛔ never a hand-resolved hunk.

   ---

   Generated by Claude Code
10. objectstack-fleet commented on Sep 28, 2026

    

    objectstack-fleetbot

    on Sep 28, 2026

    ContributorAuthor

    More actions

    os-dev-report
    {
    "issue": 20419,
    "status": "done",
    "branch": "claude/issue-20419-dts-one-program",
    "pr": "#20499",
    "session": "session_014EJ1ED8X4MMrT18BhVx4tx — subagent of the domain:spec seat-2 PM session (round 2); the harness id of this run is the parent's",
    "premise_still_valid": true,
    "summary": "Route A is implemented in draft PR #20499, whose body opens with this card's closing line. It touches four files: patches/tsup@8.5.1.patch (5 added lines in tsup's bundled rollup-plugin-dts: on a tsconfig override, key every entry by the tsconfig directory), patchedDependencies in pnpm-workspace.yaml, pnpm-lock.yaml regenerated by pnpm install, and spec's tsup.config.ts docblock. No package source changed and the 6144 ceiling is unchanged. Census: a static census over all 67 tsup packages and a createProgram probe across the 31-package census closure agree on 12 multi-program packages: spec 18, platform-objects 11, metadata 3, metadata-core 1+2, service-cluster 1+2, and core, lint, objectql, plugin-auth, plugin-webhooks, service-datasource and types at 2 each. After the patch each has one program per config item. Spec DTS pass at the 6144 ceiling in an 8192 MB cgroup (8113763, noCheck on): 18 → 1 programs, live heap 4997 → 882 MB (after-mark-compact reading from only 4 mark-compacts, so a lower bound; 1677 MB before a mark-compact), peak RSS 5694 → 3537 MB, 153 → 41 s. The whole spec build went 189 → 87 s. Equality: 29 of 31 closure packages are byte-identical before and after, including 11 of the 12 census packages. spec (130 files) and metadata-protocol are canonically equal with 0 files differing, and one spec chunk is paired by content (data-engine ↔ analytics.zod). Negative controls were caught in all 12 census packages, on an entry file each plus a chunk in lint, metadata-core, objectql and spec. Rot guard, measured with pnpm 10.31.0: tsup moved to 8.5.0 gives ERR_PNPM_UNUSED_PATCH (and a frozen install gives ERR_PNPM_LOCKFILE_CONFIG_MISMATCH); removing the yaml key makes the frozen install refuse.",
    "tests": "Heads: before-trees at 8113763 (unpatched), after-trees at 78cbf5d (patched; the same source for every package), and gates, builds and tests at the merged head 57e5191 (branch plus origin/main fc0db22). (1) Instruments: a cgroup-v1 memory cgroup capped at 8192 MB (round 1 proved the cap kills first); --trace-gc for the largest heap after and before a mark-compact; a createProgram probe preloaded into the tsup DTS worker only (measurement, no argument or return changed). (2) Canonical comparison: a TypeScript re-print sorting union members, property-only type literals, top-level statements and import/export specifiers; chunk hashes stripped; chunks paired by unique chunk-name-agnostic content. canon-dts2 on all 31 packages reported PAIRING problems=0 and FULL-CANON-DIFF 0 for each. For spec, "RENAMED chunks (B → A): analytics.zod → data-engine" and "FULL-CANON-DIFF after rename: 0 file(s)". With statement sort off, spec shows 20 files differing, the same kind as in round 1. (3) Negative controls: the first real string type node became number, located by the parser and never in a comment. Every run named exactly the mutated file: core, lint (+chunk runtime-HASH.d.cts: "no partner"), metadata, metadata-core (+chunk repository), objectql (+chunk util), platform-objects (identity/index.d.ts, because index.d.ts only re-exports; the first attempt pointed at a directory, errored, and was redone), plugin-auth, plugin-webhooks, service-cluster, service-datasource, types, spec data/index.d.mts, and spec's renamed analytics.zod chunk (pairing failed: "no partner in A for analytics.zod-…", diff = data-engine-HASH.d.ts). (4) Rot: scratch projects on packageManager pnpm 10.31.0 with the same patch file and key. On tsup 8.5.1, install exit 0 and the patch marker is present once. On tsup 8.5.0, install exit 1 with "ERR_PNPM_UNUSED_PATCH The following patches were not used: tsup@8.5.1", and the frozen install exits 1 with ERR_PNPM_LOCKFILE_CONFIG_MISMATCH. In-tree, scripts/ablation-replace.mjs deleted the yaml block (anchor 1 → 0) and pnpm install --frozen-lockfile exited 1 with ERR_PNPM_LOCKFILE_CONFIG_MISMATCH; the restore blob equals HEAD b96f61f907d3 and git diff HEAD is empty. (5) spec tests that read tsup.config.ts as text: "Test Files 8 passed (8) / Tests 130 passed (130)". Full spec suites were not re-run: no source changed, and the declarations are canonically equal.",
    "gates": [
    "node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands at 57e5191 :: 60 derived over 4 changed paths; --ran with an exit code per line: "60 derived famil(ies) accounted for — 60 run, 0 NOT-MEASURED (a DERIVED zero — all 60 recorded an exit code and none of them is 3)"",
    "all 60 :: exit 0, including check:override-consistency, check:vendor-export-contract-resolve, check:workspace-manifest-cycles, check:osv-exemptions, check:prerelease-pin-watch, check:changeset-fixed and check:nul-bytes",
    "pnpm check:dts-closure :: exit 0 ("72 built package(s) swept - 166/166 declared declaration file(s) present")",
    "pnpm check:dual-build-cjs-loads :: exit 0 (after a whole-workspace build)",
    "pnpm check:lean-entry-closure :: exit 0",
    "pnpm exec turbo run build (census closure) --force --concurrency=2 :: exit 0, "29 successful, 29 total / 0 cached"",
    "pnpm exec turbo run build --filter=!@objectstack/docs --concurrency=2 :: exit 0, "72 successful, 72 total"",
    "pnpm --filter @objectstack/spec build in the 8192 MB cgroup :: exit 0, peak RSS 3680 MB, 77.6 s, DTS 38650 ms, check-dts-emitted 36/36, check-dts-references 394/394",
    "pnpm --filter @objectstack/spec check:generated :: exit 0 ("All 15 generated artifacts are up to date")",
    "pnpm install --frozen-lockfile :: exit 0",
    "CI: not awaited, per contract"
    ],
    "line_budget": "4 files, +89 / −41 (130 changed lines vs merge base fc0db22): packages/spec/tsup.config.ts +38/−30, patches/tsup@8.5.1.patch +16, pnpm-lock.yaml +16/−11, pnpm-workspace.yaml +19. No skills/** or ratchet ledger touched; no governed surface.",
    "files_changed": [
    "patches/tsup@8.5.1.patch",
    "pnpm-workspace.yaml",
    "pnpm-lock.yaml",
    "packages/spec/tsup.config.ts"
    ],
    "deviations": [
    "Census correction: round 1 estimated about 11 besides spec, including create-objectstack. Measured: create-objectstack has one DTS entry (dts.entry) and one program, and platform-objects (object-form entry) had 11 programs. The census is 12 including spec.",
    "pnpm patch-commit wrote patchedDependencies into package.json's pnpm field. It was moved to pnpm-workspace.yaml per the ruling, and package.json is unchanged vs base. The lockfile was byte-identical before and after the move, and pnpm install --frozen-lockfile exits 0.",
    "The before/after tree comparison ran at 8113763 (source-identical to 78cbf5d). Gates and builds ran at the merged head 57e5191, after origin/main fc0db22 brought core/objectql/driver-sql source changes. Canonical equality is therefore claimed for the pair that differs only by the patch.",
    "Spec live heap after the patch (882 MB) comes from 4 mark-compacts, so it is a lower bound; the largest heap before a mark-compact, 1677 MB, is also reported.",
    "Merged origin/main (2 commits, not touching the lockfile or yaml) as a merge commit before the gates (AGENTS.md multi-agent section 10).",
    "Commit trailers use the model-free pair AGENTS.md mandates, not the harness reminder's model-named Co-Authored-By. The pre-push trailer check passed.",
    "Lock: 3 calls ended in exit 99 queue-timeouts under contention; the slot issue-20419-r2 was resumed each time. OS_VERIFY_LOCK_WAIT was lowered per call to keep wait plus run under the foreground cap."
    ],
    "mcp_calls": "0 — no MCP GitHub tool was called",
    "api_writes": "3 — all through scripts/pm as objectstack-fleet[bot] via the fleet-write relay (one POST /repos/objectstack-ai/objectstack/dispatches each): (1) with-fleet.sh --via dispatch pr_create → POST /repos/objectstack-ai/objectstack/pulls (draft) = #20499, relay run 36473711139 success, body read back byte-identical; (2) label-write.mjs labels_add skip-changeset + assign os-tesla → POST /issues/20499/labels and /assignees, relay run 36473787916 success, read-back matches (dependencies and size/m were added by the labeler bots and preserved); (3) post-stamped.mjs this os-dev-report comment → POST /issues/20419/comments. git push x4 (not REST): empty-branch probe, the patch commit, the docblock commit, the merge commit. Reads were single-card REST GETs (card comments, PR 20499).",
    "open_questions": [],
    "out_of_scope_findings": [
    "carrier: the maintainer (upstream filing) · noted, not filed — the rollup-plugin-dts createPrograms per-directory grouping is still present in 6.5.1. The draft upstream report is in PR #20499's body; no seat writes outside this org",
    "carrier: PR #20499 Acceptance notes · noted, not filed — every tsup bump now owes the patch a decision (pnpm refuses the unused patch, measured). The bump PR is where it is re-derived or retired"
    ]
    }
11. objectstack-fleet commented on Sep 28, 2026

    

    objectstack-fleetbot

    on Sep 28, 2026

    ContributorAuthor

    More actions

    ACCEPT (round 2, root cause) — PR #20499 at head 57e5191f21 · domain:spec seat 2 (session_014EJ1ED8X4MMrT18BhVx4tx) · 2026-09-28T19:46Z

    The seat reviewed the dev report on this card (round 2) against GitHub, the diff and tsup's own source.

    • PR shape:
      • draft, base main, first line Fixes #20419, Clause-②: no at line start, assignee os-tesla, skip-changeset;
      • 4 files (+89 / −41): patches/tsup@8.5.1.patch, pnpm-workspace.yaml (patchedDependencies), pnpm-lock.yaml (regenerated by pnpm install), and the docblock of packages/spec/tsup.config.ts;
      • no package source changes. It is not a contract surface by path, so this is the seat's review.
    • The patch, read against tsup 8.5.1's dist/rollup.js (getCompilerOptions):
      • On a cache MISS with a tsconfig override, dirName becomes the tsconfig's directory. On a HIT it stayed dirname(input), which is the one-program-per-entry defect.
      • The patch sets the HIT's dirName to the value the MISS computes. On a MISS it re-assigns the same value, and the no-override path is untouched.
      • Five lines, correct as written.
    • Evidence:
      • Census: 12 multi-program packages, measured two ways that agree.
      • Spec DTS pass, 18 → 1 programs. Live heap 4997 → 882 MB (a lower bound; 1677 MB before a mark-compact), peak RSS 5694 → 3537 MB, 153 → 41 s.
      • Equality: 29 of 31 closure packages are byte-identical. spec and metadata-protocol are canonically equal, with one renamed chunk paired by content.
      • Negative controls were caught in all 12 census packages.
      • Rot guard, measured: another tsup version gives ERR_PNPM_UNUSED_PATCH; removing the key fails the frozen install.
      • Gates: 60 derived, 60 run, all exit 0, including check:dual-build-cjs-loads after a whole-workspace build.
    • skip-changeset: right under the seat's Q2 answer (5874957193). No JS or shipped source moves, and the declarations are canonically equal.
    • Findings:
      • The upstream report to rollup-plugin-dts is drafted in the PR body, for the maintainer to file (carrier: the maintainer).
      • Every tsup bump now owes the patch a decision → carrier: the PR's Acceptance notes and the pnpm-workspace.yaml comment.
    • Veto window: route A was the seat's call with the maintainer's veto open until this PR lands (5874957193). No veto has arrived. Landing on green closes the window.
    • Landing: when every check on the head is green or a roster skip, this seat runs the pre-landing checks, flips it ready and arms auto-merge.

    ---

    Generated by Claude Code
12. objectstack-fleet

    closed this as completedin #20499on Sep 28, 2026
13. objectstack-fleet commented on Sep 28, 2026

    

    objectstack-fleetbot

    on Sep 28, 2026

    ContributorAuthor

    More actions

    Landing record — round 2 (root cause) · PR #20499 MERGED · domain:spec seat 2 (session_014EJ1ED8X4MMrT18BhVx4tx) · 2026-09-28T20:53Z

    • Merged through the merge queue at 2026-09-28T20:52Z as c5ad1de027, after the round-2 ACCEPT 5877254372. The maintainer's veto window on route A (5874957193) closed with the landing, and no veto was received.
    • Verified by content on origin/main c5ad1de027, not by the merged flag:
      • patches/tsup@8.5.1.patch, pnpm-workspace.yaml and packages/spec/tsup.config.ts are blob-identical to the reviewed head.
      • patchedDependencies names tsup@8.5.1: patches/tsup@8.5.1.patch.
    • Card outcome: both rounds landed.
      • Round 1, PR build(spec): drop the DTS pass's duplicate type check — 93% to 83% of its 6144 MB ceiling #20483 (noCheck), dropped the duplicate type check.
      • Round 2, PR build: one ts.Program per DTS pass — patch tsup's bundled rollup-plugin-dts grouping (spec 4997 → 882 MB live heap) #20499, gives one ts.Program per DTS pass. Spec's pass went from 18 programs to 1, with a live heap of 4997 → 882 MB.
      • Fixes #20419 closed this card as completed; pm:dispatched is removed in this act.
    • Carried forward (not filed):
      • The upstream report to rollup-plugin-dts is drafted in PR build: one ts.Program per DTS pass — patch tsup's bundled rollup-plugin-dts grouping (spec 4997 → 882 MB live heap) #20499's body, for the maintainer to file.
      • Every tsup bump now owes the patch a decision; pnpm refuses an unused patch at install.

    ---

    Generated by Claude Code
14. objectstack-fleet

    removed

    pm:dispatched

    on Sep 28, 2026
15. veigajoao

    added 2 commits that reference this issue on Sep 29, 2026

    build(spec): drop the DTS pass's duplicate type check — 93% to 83% of…

    ...

    8113763

    build: one ts.Program per DTS pass — patch tsup's bundled rollup-plug…

    ...

    c5ad1de