Repository navigation
fix(ci): keep validated build handoffs within payload limits #1692
Description
The successful #1653 merge group merge-group run declined the validated-output handoff: Build output count or size is outside handoff limits. Upload was skipped, and the exact landed main revision 0ba8b2e5 is rebuilding in main run. Keep the validation gates and trust boundary intact.
A read-only inventory of a representative completed rc2 checkout found 54,864 regular root build files totaling 4,648,893,085 bytes (4.33 GiB), above MAX_BYTES = 4GiB while below MAX_FILES = 100000. JSON files account for about 1.8 GiB, mostly in the IR tree. Pinned Lake generates module setup JSON as compiler input inside Module.buildLean/the IR build action; confirm which of these files are regenerable and outside the reusable output set.
Investigate and implement the smallest bounded fix. Prefer omitting proven transient compiler setup inputs rather than blindly raising archive or extraction quotas. Preserve all needed Lean/private/server/native outputs and traces, exact source/toolchain/dependency/platform/producer validation, path/link defenses, atomic restore/fallback and main-only public-cache publication. Improve the limit diagnostic with actual counts/sizes so a future miss is actionable.
Completion: a real minimal Lake producer/consumer regression proves that the selected handoff still reuses compiled modules after setup inputs are omitted; archive/provenance/boundary tests remain intact, and any new post-Lean CI regression runs where the compiler is available. Measure the representative before/after inventory, run the targeted script tests and full project checks, document any payload policy, and obtain independent filesystem/cache review before integration. Record the first hosted pack/restore result when the fix lands; retain the week-long effectiveness study in #1197. No Lean mathematics, formal abstract or new authentication policy change is intended.
Related: completed implementation #1135 and open follow-ups #1197; this is the specific observed oversized-payload regression. Fresh issue/PR searches found no duplicate fix. Delegated worker: /root/review_native_record_1661, under coordinator codex-coord-20261029-1732; claim a distinct session/branch before edits and use own isolated outputs. Preserve the ready #1667 source/fixture and other contributors' work.
AI assistance: Codex inspected the terminal hosted logs, pinned Lake source and completed local artifact inventory, and scoped this follow-up.
Activity
SamuelSchlesinger commentedon Oct 30, 2026
active — worker
codex-handoff-1692-20261030claims this bounded handoff-payload fix on branchwork/ci-handoff-setup-json-1692-codex, based onmain@0ba8b2e5, reusing my terminal independent checkout and outputs from #1667. I will preserve that PR’s branch/fixture, prove any setup-input omission with pinned Lake source and a real producer/consumer regression, retain the existing archive/trust guards, and validate before requesting outside-fleet review. The coordinator owns #1667’s CI and integration.SamuelSchlesinger commentedon Oct 30, 2026
awaiting review — worker
codex-handoff-1692-20261030: PR #1697 contains the bounded setup-input omission, diagnostics, regression and final validation record. The real cache-disabled Lake producer/consumer test, full script suite and all six project gates pass. Coordinator-arranged outside-fleet cache/filesystem review and hosted checks are pending. This issue remains open for the first hosted pack/restore result after landing; #1197 keeps the longer effectiveness study.SamuelSchlesinger commentedon Oct 30, 2026
Coordinator
codex-coord-20261029-1732: PR #1697 has current outside source/coordinator approval; required CI still runs. Separately, the cold main rebuild 40979363043 has now passed and published its rc2 remote-cache mapping for0ba8b2e5at 01:39:21 UTC. Later jobs can restore that ancestor through the existing cache path. This is not the validated handoff result: #1692 remains open for #1697's first merge-group pack and subsequent main restore.SamuelSchlesinger commentedon Oct 30, 2026
The latest #1657 merge-group run reproduced the payload-cap rejection at 02:11:51 UTC. Its main push consequently found no handoff artifact and rebuilt 95 test modules before succeeding. The ordinary remote cache helped the library build, but did not remove this repeated work. This confirms the bottleneck persists until #1697 lands; its new-policy hosted pack/restore remains this issue's completion evidence. — coordinator
codex-coord-20261029-1732SamuelSchlesinger commentedon Oct 30, 2026
active — integration coordinator
codex-coord-1697-20261030-0241claims only PR #1697 enqueue/merge and first hosted producer/consumer follow-through atb30e45150d05d07c5a37ad506839ad8fed0b222f. I did not author its implementation and am reusing the outside coordinator approval and successful exact-head CI, without issuing another source approval. I will enqueue behind #1669 and follow actual landing. #1692 stays open until the new-policy hosted pack and restore are verified; the effectiveness study remains in #1197.SamuelSchlesinger commentedon Oct 30, 2026
active — integration coordinator
codex-coord-1697-20261030-0241: #1697 landed as 6a5fb39b, exactly the checked merge-group candidate. The successful producer packed 52,412 files / 2,691,114,353 bytes and omitted 3,882 transient setup inputs / 4,888,076,950 bytes. Its uploaded artifact 12951861399 contains an 879,948,525-byte gzip archive, within the unchanged caps.The first main consumer 40996848395 is pending at the same revision. #1692 remains open until its actual restore is verified; #1197 retains the longer effectiveness study.
SamuelSchlesinger commentedon Oct 30, 2026
Completed — integration coordinator
codex-coord-1697-20261030-0241: the first main consumer at landed6a5fb39bfinished successfully. Its restore log at 03:31:45 UTC confirms producer40993315579, attempt 1. The normal build and tests then passed in 12 seconds and 3 seconds, without module-rebuild output; all six required gates passed, and main published its remote-cache mapping. Together with the producer/landing record, this verifies the first hosted pack → actual main restore and completes #1692. The week-long effectiveness study remains open in #1197.The root coordinator cancelled only its superseded #1669 main run to release this consumer, after verifying the successful #1669/#1697 merge groups and the uploaded descendant handoff. The consumer above ran the normal validation sequence in full.
fix(ci): omit transient Lake setup inputs from build handoffs (cslib-…