PR #32343 - manaflow-ai/cmux

The NEXT nightly's build-nightly-app ends with two serial uploads: the unsigned app (413 MB) and the dSYM handoff (221 MB). prepare-nightly-variant-primary and the signers wait for that job to end. actions/upload-artifact sends 8 MB blocks 5 at a time by default. Across five recent runs these two uploads ran at only 2.1-6.4 MB/s, with a block finishing every 3-17 s.

This PR sets ACTIONS_ARTIFACT_UPLOAD_CONCURRENCY: "32" (one YAML anchor) on:

Upload unsigned nightly app
its retry
Keep the nightly-next dSYMs for publication

ci-macos.yml's product upload already uses the same setting (#14375). The action caps the value at 16 per CPU.

Nothing else changes: artifacts, names, paths, compression levels (app 0, dSYMs 6), retention (1 and 21 days), overwrite, step order, timeouts and the retry are all the same. No step was added or renamed, so .github/nightly-step-coverage.json needs no change. No secret, permission, environment or concurrency-group change. nightly.yml grows to 480,200 bytes, still under the 490,000 budget.

Why not move the dSYMs off the job

The dSYMs exist only on the build Mac's disk. A separate job would have to download an artifact this job uploads first, so it saves nothing.

Running the dSYM upload in the background would mean handing the job's Actions runtime token to a run step. That changes how a secret is handled, so it is out of scope here.

The app upload already runs first, at compression-level: 0, and nothing slow comes before it. That leaves the upload's block concurrency as the safe cut.

Performance

Metric: from the start of Upload unsigned nightly app to the start of signer-app (the build job's tail, plus prepare and the signer pick).

run	metric before	projected after	delta	uploads before (app + dSYM)
41444184160	349 s	317 s	-32 s	97 + 34 s
41439348303	311 s	263 s	-48 s	92 + 55 s
41437704107	600 s	410 s	-190 s	181 + 108 s
41432827903	622 s	435 s	-187 s	188 + 98 s
41430382728	463 s	258 s	-205 s	211 + 93 s
median	463 s	317 s	-187 s	286 s

How "after" is projected: the best throughput these builders reached at 5 blocks was 6.4 MB/s (the dSYM upload in 41444184160), so the uplink carries at least that. Each upload is taken at min(measured, size / 6.4 MB/s), which is about 65 s for the app and 35 s for the dSYMs. Everything else in the window is kept as measured.

This is a projection, and it is meant to be conservative. More blocks in flight can go faster than 6.4 MB/s if the uplink has room, and no slower than today if it does not. The first nightly after landing measures it from the steps' own timings.

Blacksmith minutes per day: unchanged. These steps run on the owned build Mac, and no job moved.

Risk

Low. This only changes how many 8 MB blocks the action holds in flight (up to about 256 MB of buffers on the build Mac); the uploaded bytes are the same. If 32 connections stalled, the existing 10-minute timeout and one retry still bound the app upload, and the dSYM step stays continue-on-error. A revert is one commit.

Tests

New test_tail_uploads_send_blocks_in_parallel in tests/test_nightly_stop_if_publish_doomed.py: fails 3/3 on base, passes at head.
Pass: test_nightly_stop_if_publish_doomed, test_nightly_background_steps, test_ci_upload_artifact_retention, test_nightly_step_coverage, test_ci_fork_runner_routing, test_nightly_build_state, test_nightly_signer_jobs, test_nightly_delivery_concurrency (its functions called directly), test_nightly_workflow_size, test_nightly_unsigned_resume, test_nightly_helpers_beside_release_build, test_nightly_trusted_routing, test_nightly_build_runner.
Pass: bash tests/test_nightly_universal_build.sh, bash tests/test_ci_self_hosted_guard.sh.
actionlint 1.7.12-manaflow.1 with SHELLCHECK_OPTS="-S warning" reports nothing on nightly.yml, the same as base.

Agent session: agent session
🤖 Generated with Claude Code

^{Need help on this PR? Tag @codesmith-bot with what you need.}

Summary by cubic

Speeds up the nightly build job's final artifact uploads by raising the artifact block concurrency from 5 to 32, matching ci-macos.yml's product upload.

The build-nightly-app job's unsigned app (413 MB) and dSYM (221 MB) uploads kept the signers waiting; at the default 5 concurrent 8 MB blocks they ran at 2.1-6.4 MB/s. Setting ACTIONS_ARTIFACT_UPLOAD_CONCURRENCY: "32" on the app upload, its retry, and the dSYM handoff shortens the tail without changing what is uploaded: same bytes, artifact names, compression, retention, and step order. A new test pins the setting on all three steps.