Repository navigation
ci: move e2e, real-adapter and windows pod boot to the fleet - #12133
Conversation
|
Intent: Take the browser E2E gate, the real-adapter contract lane and the Windows pod-boot canary off the shared hosted runners and onto the team CodeBuild fleet, leaving hosted only the one E2E step that genuinely needs an unprivileged user namespace. |
UX Review (Fable 5) — ⏭️ skippedRevision |
Design Review (Fable 5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS Title claims the Windows pod boot canary moves to the fleet; the diff pins it hosted — the merge commit will record a false fact permanently. Watch
Suggestions
[DESIGN-REVIEWED] 02f7b9b |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 4.8 Review — ✅ no blocking findingsReviewed Review detailsBoth facts check out: No findings. [OPUS-REVIEWED] 02f7b9b Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
First Principles Review (Fable 5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The new hosted job re-pays checkout/install/sysctl the PR exists to save: Not justified as shipped
What this change shipsInventory (8 items) — 5 justifiedIntent: stop spending ~15 of 27 shared hosted-runner minutes on jobs that don't need the hosted kernel — a FIX of a measured cost.
[FIRST-PRINCIPLES-REVIEWED] 02f7b9b |
e79a3e4 to
d763eaf
Compare
…he namespace step out hosted
d763eaf to
02f7b9b
Compare
|
Rebased: main moved to e8b94aa since this head went green-adjacent; patch content verified byte-identical across the rebase. Also: the E2E suite's own gateway now takes |
…he namespace step out hosted (kirodotdev#12133)
Problem / Motivation
After #11880 the 16 backend shards, lint and the frontend lanes run on the CodeBuild fleet, but the browser gate
E2E (stub ACP backend, offline)still ran on a shared hosted runner. It is 15 of the 27 hosted minutes a CI run has left. Only one of its steps needed the hosted kernel: the ~3.5-minute private-workflow MCP run that needsunshare --map-root-user. The other 11 minutes (SPA build, Playwright specs, memory UI evidence, i18n render gate) sat on the shared pool for no reason.Real Adapter Contract Testswas hosted with no stated fleet blocker.Why it matters
Hosted-runner minutes are the shared queue every PR waits in. Moving these two jobs takes the per-run hosted footprint from ~27 to ~18 minutes and keeps the shared pool for the jobs that genuinely need it.
What changed (motivation → approach → change)
e2eubuntu-latestlinux_runner_large, behindrun-as-runnere2e-private-namespacee2e)ubuntu-latestjob, one test filereal-adapter-contractubuntu-24.04linux_runner, behindrun-as-runnerThe E2E suite's own gateway takes unsandboxed consent for its disposable config. The fleet container refuses
unshare(CLONE_NEWUSER)at the runtime policy level, which no sysctl lifts, and the agent binary this lane runs isfake_acp_backend.py, a stdlib echo stub -- so OS isolation guards nothing the browser specs assert. Without the consent every agent-dependent spec gets a sandbox-refusal card instead of a turn. A sandboxed spawn doing real work stays covered bye2e-private-namespace(hosted, real namespace) and bye2e-boot-matrix, which asserts both sandbox tiers.The E2E job is split along the one hard boundary. The fleet container refuses user namespaces, so
test/e2e/test_private_workflow_memory.pykeeps its exact command, env and the AppArmor sysctl in its own small hosted job. Everything else ine2emoves to the fleet.On the fleet the gateway refuses provider execution as root, so
e2eandreal-adapter-contractuse the samerun-as-runnerboundary as the backend shards: setup and every step that writesGITHUB_OUTPUT/GITHUB_ENVrun first as the runner identity, then the tests run throughci-shellasrunner. The Playwright browser is pinned to$RUNNER_TEMP/ms-playwright(written toGITHUB_ENV) so the root cache restore and the runner-user specs find the same directory.Pod Boot Canary (Windows)stays hosted and its comment now records why instead of leaving it unexplained. The Windows fleet is a container, whereschtasks /Createexits 1 on its own/ST is earlier than current timewarning, so the runner user cannot create the scheduled task this canary exists to prove. Measured on the fleet: all 3 tests fail at that probe before any pod boots.The
docs/ci/ci-and-reviews.mdjob table names the new lane and the runner each job uses.Backwards compatibility
Compatible: no product code changes. Fork PRs and non-allowlisted actors still resolve to the hosted labels through the existing
changesoutputs.Tests
test_ci_nonroot_boundary.py:e2eandreal-adapter-contractroute to the fleet with the boundary after setup and every file-command write before it;e2e-private-namespacekeeps the exact private-workflow command, env and sysctl and stays hosted;e2eno longer mentions the namespace or that file.test/test_playwright_e2e.py: the suite's gateway seedssandbox_allow_unsandboxed_execfor its own throwaway home only, so smoke and the boot matrix keep their existing sandbox expectations.test_ci_fleet_routing_expression_parity.py: the two migrated jobs are added to the pinned resolver-consumer table;test_ci_nonroot_boundary.pykeepspod-boot-windowsin the stays-hosted set.test_e2e_ci_scheduling.py: the private gate is asserted on the new job.Manual verification
This PR runs under the modified workflow, so the CI on this PR is the proof:
e2eandreal-adapter-contractmust showcodebuild-kirocrew-gha-*runner labels and pass, ande2e-private-namespacemust pass onubuntu-latest. Measured:Real Adapter Contract Testspassed on the Linux fleet in 4 min,e2e-private-namespacepassed hosted in 4 min.actionlintpasses onci.yml.Related Issues
no linked issue: follow-up to #11880
Checklist
feat|fix|docs|refactor|perf|test|chore|ci|build|revert: ...)