feat(browser): retain cleanup evidence for failed Agent Tasks

[seonghobae]

Partial implementation of #28, stacked on unchanged exact PR #136 head 1deca80bf6a67394e25a2c914b76d490c2aad960.

Buyer/reliability gap

The controlled Agent Task runner previously discarded temporary-profile cleanup evidence whenever either the ordinary browser pass or forced-close probe failed. Outer loops retained only trial number, failure state, and exception type, while cleanup/isolation checks filtered to successful trials. A failed or interrupted task could therefore lack credential-free teardown evidence even though #28 requires deterministic failure/recovery and no leaked task profile.

Dependency and writer alignment

This Draft targets unchanged exact #136 head 1deca80bf6a67394e25a2c914b76d490c2aad960 and must remain Draft while #136/#132 and their prerequisites remain active. Earlier independent prerequisite movements were incorporated through normal non-destructive two-parent merge commits. Current exact head 499a6499112a11094279a1386238e26987d771d5 contains the current base as ancestry. No force-push, destructive rebase, base mutation, or predecessor evidence transfer occurred.

TDD and RCA sequence

1. Exact aligned ordinary-trial test head 7e1db77e82dae1d43fdf57c20c54337678f3799c required _run_agent_task_trial to retain bounded cleanup evidence after a controlled browser-pass failure and required an all-trial cleanup gate. CI 31655752094, Rust contracts 94309699104, failed exactly the two intended Python boundaries: the mocked RuntimeError escaped and aggregate cleanup proof did not exist. This is the first valid RED.
2. Production head c6c491ff794323bec5de19d7ece415f89d708d28 added bounded ordinary-trial failure evidence and all-trial cleanup gating; merge-alignment head 56ed81c52395ff6e61ab9efd76493fea32445af1 was exact-green in CI 31656517616 and MV3 31656517683.
3. A fresh lifecycle audit found the same defect in _run_agent_task_forced_close_trial. Test head b6583d54d5ccbf2d8028e51035fb5fe20100d6a9 added forced-close failure and cleanup-gate requirements. CI 31656775065, Rust contracts 94312873184, passed repository setup and failed exactly because the forced-close exception still escaped and the dedicated evidence/gate were absent. This is the second valid RED.
4. Production head 4b474260e777c84335636b2fbd732a00e48159d4 added forced-close failure cleanup evidence and a separate all-trial forced-close cleanup gate; CI 31657248055 and MV3 31657248068 succeeded.
5. Current exact head 499a6499112a11094279a1386238e26987d771d5 adds only the truthful Unreleased changelog entry covering ordinary and forced-close failure cleanup evidence.

Implemented contract

Both _run_agent_task_trial and _run_agent_task_forced_close_trial now:

• catch only the reviewed browser-run failure classes (OSError, ValueError, RuntimeError, json.JSONDecodeError) inside the TemporaryDirectory lifecycle;
• retain only bounded failure type, never exception message or browser/page content;
• verify profile removal after the context exits;
• fail with a dedicated cleanup error if the profile remains;
• return trial number, passed: false, failure type, profile_cleaned: true, and duration when browser work fails but teardown succeeds;
• preserve full successful-trial evidence plus profile_cleaned: true; and
• leave unexpected exception classes unnormalized.

Aggregate evidence now computes cleanup across every ordinary and forced-close trial, emits both cleanup results, and applies distinct fail-closed ordinary and forced-close cleanup gates before repeatability/recovery acceptance. An outer failure without cleanup proof therefore fails closed rather than inventing teardown success.

Exact-head proof

On unchanged exact head 499a6499112a11094279a1386238e26987d771d5 against unchanged exact prerequisite 1deca80bf6a67394e25a2c914b76d490c2aad960:

• CI 31657489713: success;
• Rust contracts 94315047896: repository contracts, formatting, locked workspace/all-target check, full tests, strict Clippy, and rustdoc success;
• Production coverage 94315047798: exact owned production function/line/region/branch enforcement success;
• Manifest V3 Compatibility 31657489637: success;
• CodeRabbit exact-head status: success;
• GitHub reports the Draft mergeable against its unchanged prerequisite base; and
• no inline review threads are currently returned.

No predecessor-head, prerequisite, synthetic-merge, skipped, cancelled, absent, pending, or stale evidence is promoted as current proof.

Truth boundary

This lane records controlled temporary-profile teardown evidence only. It does not prove process termination, adversarial filesystem erasure, deletion of every browser artifact, operating-system keychain cleanup, renderer-crash attribution, arbitrary-task recovery, browser/adapter authenticity, policy authorization, durable incident evidence, or #28 completion. A production runtime still needs trusted task/process ownership, durable recovery evidence, and separately reviewed cleanup/retention semantics.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 73ead784-6e84-4829-ae15-9d17a19a07e9

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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.}