ReviewOS

also looking at this

stacks/drone

test: observe native execution callback verification cost

#86
Merged chris wants to merge codex/observe-execution-verification into main
8 files +363 -18
8 days ago

The native failure observer can remain pending while the parent verifies a completed recording. This adds bounded timing around the actual core-created Worker's synchronous message callback in the private native browser fixture. It preserves the genuine constructor, frames, callback receiver/argument/return/throw, authorization checks and all existing deadlines. Result/finished markers and the new JSON receipt contain only closed ordinals, kinds, counts and durations; payloads, credentials, recordings and source identifiers are excluded.

The observer is tested with real Bun Workers. Five contract tests pass with 58 assertions, including opaque frame delivery, original return/error identity when diagnostics fail, detached bounded snapshots, a four-marker ceiling per Worker, and deterministic exclusion of marker latency. The retained regression first reported a 7 ms callback as 1,007 ms when a failing marker advanced the external clock by 1,000 ms; the corrected timer reports only 7 ms. Native lint and app/runtime types pass. Independent Spec and Standards reviews are clean after that correction.

The first inline native browser attempt failed when Chrome exited 0 before cleanup at 33.08 seconds during nominal100. It had observed 2,694 event callbacks and no result callback; the original recording had not returned. The failure and exact pinned inputs are retained. The extracted/fixed source has not completed a full native browser run locally, and no unchanged GUI retry was substituted. Exact new-head Linux qualification is required before main.

Related to #44, #64 and #68. Separate local pure verification/HTTP measurements demonstrate a one-second diagnostic blocking possibility, not the historical Linux 40-second completion cause. This is observational test infrastructure, with no production or physical-output change and no claim to fix the unresolved timeout. The workflow adds only one artifact path; no command, job or deadline changes.

This draft is stacked on #85's reviewed execution visibility qualification. Its parent already contains the qualified VTOL workspace now on main in #84. All remaining investigation criteria stay open.

Exact new-source Linux CI, on c1b1569e1bf74a3cfe9c4d1b3b3687ee8026f9ce, is running. Main publication requires its complete qualification.

The exact-source Linux execution-browser job has passed. Its verified artifact measures 3,107.70 ms in the original nominal result callback, with all observed frames/limits intact and Chrome exiting after cleanup. This passing observation does not explain earlier 40-second failures. The same workflow's workspace-browser job failed startup with 400 ENOENT port probes, empty startup stderr and Chrome termination only during cleanup; #72 retains that separate failure. The full workflow is still running, and this draft is not qualified for main.

8 changed files on the files tab, with 0 review threads.

Sign in to comment on this pull request.