ci(e2e): start builds from main's DerivedData so test-only changes skip the app compile

[teamleaderleo]

A test-e2e.yml dispatch that changes one test file recompiles the whole app host. The build starts from an empty DerivedData, and a fresh checkout makes every source file newer than any earlier build. In run 35868678521 the compile took 14 min and the selected test ran for 18 s. CLAUDE.md's "Reading CI cost" puts the app scheme at 691 of the 735 s build-for-testing spends.

With this change, the build job adopts the newest DerivedData (build database, intermediates and products) that a main run of this workflow published. It then restores the recorded modification time onto every input file whose content is byte-identical. Changed files keep their checkout time, so Xcode rebuilds only what differs. Correctness does not depend on how far the adopted DerivedData is from the selected revision; the distance only affects compile time.

• Producer: a run on main whose selected revision main already contains publishes its DerivedData, minus Logs and Index.noindex, as a 3-day artifact. This is the trust rule the compilation cache already uses. It is an artifact rather than actions/cache because the repository cache is at 10.5 GB, over its 10 GB cap.
• Consumer: it adopts only artifacts from workflow_dispatch runs of test-e2e.yml on main, and archive members or links that escape DerivedData are rejected. The key is the existing compilation-cache fingerprint: Xcode version plus workspace and DerivedData paths, which is what the intermediates embed.
• Failure: adopt, package and publish are optional, allowlisted in test_ci_self_hosted_guard.sh alongside the compilation-cache steps. A miss, expired artifact or bad archive leaves an empty DerivedData, which is today's cold build.

Tradeoffs. Adding steps to the build job changes the E2E product recipe fingerprint, so exact-product reuse misses once after merge. Archive size and transfer time are not measured yet; publishing is skipped above 12 GiB. The first producer appears when someone dispatches test-e2e.yml on a main-contained revision (for example ref=main).

Validation. tests/test_e2e_warm_derived_data.py (5 tests, registered in the linux-guard lane) covers: unchanged inputs take the producer's time and changed or new inputs do not; DerivedData and .git are not inputs; only main dispatches of this workflow are trusted; escaping archive members are rejected. All 137 linux-guard tests pass locally and actionlint is clean. Not yet verified on a Mac: that Xcode 26 in fact rebuilds only the changed test file from an adopted DerivedData. The cost report in the build summary now records the adoption, so the first two dispatches after merge (one ref=main producer, then a test-only branch) measure it.

— Dulcinea g1 🎐
run: run_cmux_main_app_host_baseline_inherited_failures_blocking_full_ci_20260923_576a13f4

🤖 Generated with Claude Code

---

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

---

Summary by cubic

Starts E2E builds from the newest DerivedData a main run of test-e2e.yml published, so a change to one test file rebuilds only that test instead of the whole app host. The build job adopts the archive, restores the recorded modification time onto every input whose content is byte-identical, and stamps changed or new inputs with the current time so they cannot keep an archive's older mtime and escape a rebuild. A miss, expired artifact, or bad archive falls back to today's cold build.

Tradeoffs and validation. Exact-product reuse misses once after merge because the added steps change the E2E product recipe fingerprint. Publishing is skipped above 12 GiB, requires a recorded input manifest, and archive transfer time is not measured yet. The adopt step times out after 10 minutes. tests/test_e2e_warm_derived_data.py covers unchanged, changed, and new inputs, trust rules, archive escape rejection (including absolute in-tree links), and all 137 linux-guard tests pass locally. Not yet verified on a Mac that Xcode 26 rebuilds only the changed test file from an adopted DerivedData.

^{Written for commit eadec1a. Summary will update on new commits.}

[cursor]

Bugbot is paused — on-demand spend limit reached

Bugbot uses usage-based billing for this team and has hit its on-demand spend limit.

A team admin can raise the spend limit in the Cursor dashboard, or wait for the next billing cycle to continue.

[coderabbitai]

Warning

Review limit reached

Next included review available in 17 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used all 10 included reviews currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Repository: manaflow-ai/cmux/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 43064714-7a17-47d8-aeb2-559bb44d129f

📥 Commits

Reviewing files that changed from the base of the PR and between 3ca19ad and eadec1a.

📒 Files selected for processing (8)

• .github/workflows/ci-guards.yml
• .github/workflows/test-e2e.yml
• scripts/ci/e2e_warm_derived_data.py
• scripts/ci/product_input_identity.py
• tests/test-execution.toml
• tests/test_ci_guard_workflow_structure.py
• tests/test_ci_self_hosted_guard.sh
• tests/test_e2e_warm_derived_data.py

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[teamleaderleo]

Independent review (a separate agent, read-only, at a6c46c35db). It re-ran the new test, 10 related guard tests and actionlint, all passing.

#	Finding	Disposition in HEAD
1	Blocker. A changed input unpacked from an archive (GhosttyKit, Sparkle, Sentry) keeps the archive's older mtime. swift-driver treats external dependencies as changed only when newer than the prior build, so importers would not be rebuilt.	Fixed. replay stamps every changed or new input with the current time. Regression test test_a_changed_input_unpacked_with_an_old_time_is_still_rebuilt.
2	Adoption has no size cap or timeout.	Fixed. Artifacts over 12 GiB are skipped, and the step has a 10-minute timeout.
3	The helper is outside product identity.	Fixed. Added to PRODUCT_CI_INPUTS.
4	Every absolute symlink is rejected, so Xcode's in-tree absolute links would force a cold build every time.	Fixed. Links are accepted when their target resolves inside DerivedData, and extraction uses the tar filter because Python 3.14's data default refuses absolute links. Regression test added.
5	A corrupt gzip escapes the handler.	Fixed. Any exception now means a cold build.
6	Restore runs after "Resolve Swift packages".	Kept. Replay needs the resolved package tree, and the failure path rebuilds resolver state.
7	"Record build input times" can fail the build.	Fixed. It is now optional and allowlisted, and publishing requires it to have succeeded.
8	The newest artifact by creation time may be an old revision.	Kept. That costs speed, not correctness.
9	llbuild compares inodes, and a fresh checkout changes every inode.	Adoption sets com.apple.dt.XCBuild IgnoreFileSystemDeviceInodeChanges. The effect is measured after merge, as the PR body says.

Trust, tar bounds, if: wiring and output names were checked and hold. There are no remaining blockers. All 137 linux-guard tests pass locally on HEAD, and actionlint is clean.

— Dulcinea g1 🎐