Skip to content

fix(etw): preserve first-read attribution across idle renewal - #986

Merged
antropos17 merged 2 commits into
masterfrom
codex/etw-idle-attribution-20261019
Oct 19, 2026
Merged

antropos17 merged 2 commits into
masterfrom
codex/etw-idle-attribution-20261019

Conversation

@antropos17

Copy link
Copy Markdown
Owner

An idle agent's first read can wait in the ETW queue while issuer preparation runs first. Reopening its live thread/process handles at renewal moved the proof interval forward and could leave that queued read without a verified actor. Renewal now retains the original OS handles and interval after fresh owner, birth and liveness checks. New workers at capacity also no longer skip renewal of existing workers later in the bounded inventory. Missing refresh still expires the lease, replacement handles cannot verify older events, and reads never renew the lease.

The native regression fails against the old implementation and passes with the fix. A clean ordinary Windows x64 package passed genuine installed Claude Code first reads after 65, 70 and 85 seconds, without README warmup. The original handles covered 68.591, 74.308 and 89.627 seconds before the respective reads. Each case verified the exact OS agent identity, visible action/potential risk, content exclusion, normal exit, native stop and independent session absence. The runner now requires a native held interval exceeding thirty seconds, so a fresh replacement interval cannot qualify this scenario.

Qualification and check receipts bind the source, complete normal-package manifest, published native binary and three live journeys. The earlier ordinary-package attribution FAIL remains unchanged: four instrumented baseline attempts passed, so this repair proves the renewal race and corrected behavior without claiming the unique cause of that historical intermittent failure. Collector drops and mapping resets remain recorded; new or unprepared workers can still be unverified. These passes do not establish complete capture.

Validation: 126 published native self-tests; 87 focused oracle/ETW regression tests; full Windows coverage suite (10,000 passed, 20 skipped); renderer build, formatting, lint, types/Svelte, witness and sequence mutation gates, derived counts and production dependency audit. Documentation-only evidence changes follow the qualified package commit; packaged runtime inputs are unchanged.

@antropos17
antropos17 merged commit ae1b841 into master Oct 19, 2026
7 checks passed
@antropos17
antropos17 deleted the codex/etw-idle-attribution-20261019 branch October 19, 2026 02:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant