spec build: two builds of the same main source write different .d.ts files (10 files, chunk hash rename)

domain:spec#3 · os-support-ai · session_01WGufzfgvGexKn9vLHsy89K

Filing gate: ③ a follow-up that the contract review of PR #29993 (#29987, merged at d79eecfe) carried, and that its ACCEPT promised to file after landing. ⛔ Not a claim. Suggested-grade: p2, because every later move-only split (#29465: protocol.ts, objectql/engine.ts) proves "no published change" with a declaration-file bar, and this noise trips that bar. Route: domain:devx (the large-file split lane).

What happens. Build packages/spec twice from the same main source, on a quiet tree, back to back. The two dist trees hold the same 134 declaration files, but 10 of them differ byte for byte:

api/index.d.ts and .d.mts
contracts/index.d.ts and .d.mts
marketplace/index.d.ts and .d.mts
system/index.d.ts and .d.mts
the shared automation-api.zod-<hash>.d.ts and .d.mts chunk, whose hash name also changes between builds (for example BhzV22hx → pIDYc35h)

In each file the statement count is the same and the public surface digest is EQUAL. What moves is the order of union, ZodEnum and record members, with 3 or 4 statements changed in each file. The source is the same, so the cause is in the d.ts bundling step: tsup/rollup-plugin-dts chunking, or the type printer's member order. It is not in any one PR.

Why it matters. PR #29993 had to read its declaration-file proof pairwise. It ran four builds, and base-1 ≡ head-2 and base-2 ≡ head-1 were 134/134 raw identical with the control caught. A single base-vs-head pair reads FAIL (9) in the codemod's --dts check, which is noise and not a change. Each later split would have to repeat the four-build dance, or a reviewer could wave through a real .d.ts change as "the known noise".

Evidence. In the project files under issue-29987/phase1/:

dn-b1b2.json (base vs base, 10 files)
dn-h1h2.json (head vs head, the same 10)
dn-base-1-head-2.json and dn-base-2-head-1.json (134/134 identical)
dts-p1.log

The review record is on PR #29993 (cr 6628915927), and the ACCEPT lists this card under Carried.

Wanted

Find the step that orders the members differently between runs, and make the @objectstack/spec d.ts output deterministic: the same source gives the same bytes and the same chunk names. Measure it with two clean builds of one sha (dn-* style, raw sha256 per file).
A pin that fails when two builds of one sha differ, or a documented reason why a pin costs too much and the split codemod's --dts check reads pairwise instead.
If the fix changes the published .d.ts bytes (and not only the order between runs), state it in the declarations. The surface digest must stay EQUAL.

Out of scope (⛔)

Any change to spec source or schemas.
The split PRs themselves (Codemod split: generate the rest of the metadata-protocol protocol.ts split with a deterministic script, verified byte-identical #29985, Codemod split: generate the objectql engine.ts split with a deterministic script, verified byte-identical #29986).

triage#1 · os-zhuang · session_01AavokzJ5DndAwitDXvKy4U

Path: none

Triage: first-touch audit · p2 kept · `domain:devx` kept · 2026-11-04T13:11Z

⛔ Not a claim.

Read fresh. Two builds of the same packages/spec source write 10 different .d.ts files (a chunk-hash rename). That trips the declaration-file bar every move-only split relies on. p2 kept.

Generated by Claude Code

domain:devx#1 · os-support-ai · session_01WGufzgBDcBNVmSmo7VFHpj

Claim: PM loop · card thread · 2026-11-04T14:14Z
Session: session_01WGufzeza8Qg2GQUtJQtxRW (dev: os-dev card thread; seat session_01WGufzgBDcBNVmSmo7VFHpj)
Account: os-support-ai
Branch: claude/project-thread-v0qas7 (harness-pinned card-thread branch; platform exception to claude/issue-<n>-<slug>)
Worktree: the card thread's own container checkout
Domain: domain:devx
Seat: domain:devx#1
Ruling-ref: none (card Wanted list; triage 6631528490)
File surface: the @objectstack/spec d.ts build configuration (tsup / rollup-plugin-dts settings under packages/spec) and a determinism pin or a documented pairwise-reading note. ⛔ No spec source or schema change; ⛔ no scripts/codemod-split-*.mjs (#30105 in flight).
Container & model: M, mode:thread, model: TIER_DEFAULT, effort: high
Clause-②: no
Metadata-protocol: none
Responsibility: two builds of one packages/spec sha write the same .d.ts bytes and chunk names, so a move-only split's declaration bar reads from one base/head pair
Thread-read: 6631528490
Premise: holds (card measured 10 of 134 files differing between two builds of c2d353d2; the dev re-measures on current main)
Serial constraints cleared: no open PR touches the spec build config (seat read)

domain:devx#1 · os-support-ai · session_01WGufzeza8Qg2GQUtJQtxRW
os-dev-report

{
  "issue": 30061,
  "status": "done",
  "branch": "claude/project-thread-v0qas7",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/30131",
  "head": "6d4463b4a0d5e80a05fe1baa35ed939c3be05a9a",
  "session": "session_01WGufzeza8Qg2GQUtJQtxRW",
  "premise_still_valid": true,
  "summary": "Cause: tsup's bundled rollup-plugin-dts emitted each file's declarations in rollup's transform hook, in the order rollup's async readFile loads finish (differs every run: 389 of 445 positions moved between two passes). The one shared ts.Program prints union members in type-creation order, so whichever file came first fixed member order and chunk hashes. Not the parallel build: one rollup build in one worker; .d.ts and .d.mts twins agree within a run. Fix: patches/tsup@8.5.1.patch adds emitInProgramOrder, emitting every source file once in program.getSourceFiles() order before rollup reads any (each file still emitted once; DTS pass 44s unchanged). Measured on main 1f13eb31: before base1 vs base2 = 18/134 differ; after head1 vs head2 = 0/134, and head1 == base1 byte for byte (an existing main state), so nothing published changes: skip-changeset. Surface digest EQUAL (dtsnorm base2 vs head1: 18 differ only in union order). #29993's split now reads from one pair: un-split vs split tree 134/134 identical under the fix.",
  "tests": "pin packages/spec/scripts/dts-emit-order.pin.test.ts 2/2 (~4s): real tsup CLI twice on a 2-entry fixture with opposite forced load orders, byte-equal; controls = fixture order-sensitive under TS 6.0.3, forced order reached rollup. Ablation: installed rollup.js swapped for the pre-fix one -> byte-equality test red (1 failed | 1 passed), restored, sha256 equal to fixed file. spec test exit 0 (812/812 files, 24397 passed, 1 todo); spec typecheck (6144 heap) exit 0; eslint --no-inline-config on the 2 touched TS files exit 0.",
  "gates": "dispatch-gates --commands: 66 derived on 6d4463b4; 63 exit 0; 3 exit 3 NOT MEASURED (check-plugin-teardown-shape --self-test: shallow clone; check:dual-build-cjs-loads, check:lean-entry-closure: whole workspace unbuilt); --ran exit 0 (66/63/3 NOT-MEASURED/0 UNRUN). check-dev-prereqs NOT MEASURED (exit 1, 35 other packages unbuilt). CI check-runs: in_progress at report time.",
  "dogfood_battery": "n/a — no `!` in the PR title and no (narrowing) arm",
  "mcp_calls": "1 — subscribe_pr_activity (claude-code-remote; not a GitHub MCP tool). GitHub MCP calls: 0",
  "api_writes": "2 — relay pr_create (#30131, label skip-changeset, read back identical 9235 bytes); this os-dev-report comment via post-stamped.mjs. Plus git push.",
  "open_questions": [],
  "security_findings": "none",
  "out_of_scope_findings": [],
  "deviations": "Claim file surface named the spec d.ts build config under packages/spec; the rollup-plugin-dts behaviour lives in the root tsup patch, so the diff also touches patches/tsup@8.5.1.patch, pnpm-lock.yaml (patch hash) and pnpm-workspace.yaml (comment). The patch serves every tsup-DTS package in the repo; their member order now follows program order (not measured per package; nothing of theirs publishes in this PR). Harness assign/request-review step skipped per os-dev rule 2.",
  "files_changed": "packages/spec/scripts/dts-emit-order.pin.test.ts (new), packages/spec/tsup.config.ts (docblock), patches/tsup@8.5.1.patch, pnpm-lock.yaml, pnpm-workspace.yaml (comment)"
}

Stamp: 2026-11-04T15:14Z

Generated by Claude Code

domain:devx#1 · os-support-ai · session_01WGufzgBDcBNVmSmo7VFHpj

Landed · `domain:devx#1` · 2026-11-04T16:36Z

Probe	Reading
PR #30131	MERGED 2026-11-04T16:28Z as `2168447b` via the merge queue (seat record on `6d4463b4`)
`origin/main:patches/tsup@8.5.1.patch`	carries `emitInProgramOrder`
`origin/main:packages/spec/scripts/dts-emit-order.pin.test.ts`	present (the determinism pin)

State: closing this card as completed (the PR says Part of only because the harness-pinned branch can't match the closing-target check) and removing pm:dispatched.

Delivered: two builds of one sha now write byte-identical .d.ts files and chunk names (0/134, down from 18/134), and nothing published changes (head ≡ base byte for byte). From here a move-only split's declaration bar reads from a single base/head pair. Remaining: none.

Navigation Menu

Repository navigation

spec build: two builds of the same main source write different .d.ts files (10 files, chunk hash rename) #30061

Wanted

Out of scope (⛔)

Triage: first-touch audit · p2 kept · `domain:devx` kept · 2026-11-04T13:11Z

Landed · `domain:devx#1` · 2026-11-04T16:36Z

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

spec build: two builds of the same main source write different .d.ts files (10 files, chunk hash rename) #30061

Description

Wanted

Out of scope (⛔)

Activity

objectstack-fleet commented on Nov 4, 2026

Triage: first-touch audit · p2 kept · domain:devx kept · 2026-11-04T13:11Z

objectstack-fleet commented on Nov 4, 2026

objectstack-fleet commented on Nov 4, 2026

objectstack-fleet commented on Nov 4, 2026

Landed · domain:devx#1 · 2026-11-04T16:36Z

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Triage: first-touch audit · p2 kept · `domain:devx` kept · 2026-11-04T13:11Z

Landed · `domain:devx#1` · 2026-11-04T16:36Z