[ScriptedAlchemy]

Cleanup can delete a worktree after treating truncated Git inspection output as complete. The same audit found regressions in citation lifetimes, split-suggestion history, and blind-evaluation confidence.

Changes

• Cleanup (fix(server): cleanup follows the branch an agent checked out #68): keep the worktree when either reflog or worktree-list output is truncated, so removal requires complete recovery information.
• Thread citations (Fix thread citations rendering as "Thread no longer available" #61): retain family rows for the visible citation's lifetime. Decide whether to fetch from the authoritative shell, and retry a cached failure when a still-visible citation becomes missing from that shell. Web also resolves both written and percent-decoded candidate IDs.
• Split suggestions (feat(chat): suggest fresh threads for independent drafts #55): exclude rolled-back runs and requests cancelled before starting, including canonical user messages whose runId is null. Keep cancelled turns that actually started.
• Blind evaluations (feat(labs): add blind model evaluations and orchestration guidance #53): mark omitted composer-context records as missing artifacts across request, fallback, earlier, and interleaved evidence, so incomplete evidence cannot receive high-confidence scores.
• Shared test failures: preserve T3_MOBILE_TEST_PLATFORM in the test environment, and use native Effect paths and actual filesystem casing behavior in cleanup fixtures.

The index-shutdown ordering races specific to open #67 are fixed on that PR. This PR contains the fixes applicable to current main.

Verification

• check:changed: native addon, formatting, lint, both Knip scans, and GraphQL generation checks passed.
• Selected regression tests: 223 passed, with 1 existing permissions-dependent skip.
• Selected React compiler checks: 7 passed.
• All 41 affected typecheck tasks passed with the supported --no-cache option. The default cached launcher cannot initialize its seccomp_unotify supervisor in this executor; diagnostics and repository settings remain enabled.
• Android FilePreviewModal regression: 3/3 passed after reproducing the original two failures.
• Additional audit suites passed: 116 tests covering thread loading, link parsing, Codex computer use/instructions, and native plugin/model handling; the evaluation builder/runner suite passed 13 tests.
• New coverage checks incomplete cleanup inspection, discarded history, omitted context, citation retention, query eligibility, and cached-failure retry.

The same Windows fixture corrections were verified in the full Windows cleanup suite on #67, at commit 6fe77ad584d27bc3d2cc9b4084cfcf893c09919b: 209 passed, 1 existing permissions test skipped, job success. The final three allocation/hidden-file fixture corrections also passed on this PR with file format/lint checks; this follow-up changes no production code. Live UI verification could not run because the repository's test-t3-app skill requires its T3 Browser panel, which is unavailable in this harness.

Model: GPT-6 Astra Pro. Harness: ChatGPT.

[changeset-bot]

⚠️ No Changeset found

Latest commit: 7ad853b

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR