[JacobStephens2]

Summary

Trial
  agent failure / timeout / unexplained kill → failed
  external cause → void → up to two fresh retries
    exhausted → drop the Eval task from both Arms
    more than a tenth dropped → inconclusive Verdict
experiments <id>
  show Void totals and causes per Arm
  report repeated Trial-only memory-cap stops as an Arm finding

Provider-limit pauses consume the fixed retry budget across resumes, including #1008's live-warning supervision. Existing Experiments retain their saved Void rules across runner upgrades. Host-kill and memory-cap causes require the outer launcher's evidence; the sandbox and cap launcher remain #1014's work.

Closes #1007

Evidence

• Before: the reviewer's exact SIGKILL and Resume probes each failed: agent work was incorrectly classified as void. The exact old-Bar probe failed because an existing Experiment could not resume.
  After: all three exact probes pass. Permanent regressions are an_agent_killing_its_own_harness_fails_without_a_retry, a_server_error_on_resume_after_trial_work_is_failed, and resuming_an_experiment_preserves_its_preexisting_void_rules in the Experiment/Trial integration tests. The existing test rejecting an explicit Bar change also passes.
• Before: a timed-out agent that changed files became void, and repeated provider-limit pauses granted unlimited retries.
  After: focused regressions confirm a timeout stays failed and provider-limit attempts exhaust the Bar's three total attempts.
• Login failures retry in distinct fresh repositories, then remove the task from both Arms. Tests cover per-Arm causes, the two-retry ceiling, and repeated-cap findings. All ten standalone Verdict tests passed, including the exact one-tenth boundary.
• cargo check and cargo fmt --check pass. CI on the exact head passed the full cargo test suite, Clippy, formatting, Python, site and Console checks. The local affected-test selection passed all 43 site checks; its duplicate Rust run was stopped after CI passed, reusing that result as docs/testing.md directs.

Test seams

• Version-1 Trial protocol (thirdshift --trial): observable classification and artifacts from fake Harness failures.
• Experiment CLI (thirdshift experiment / thirdshift experiments): fresh retries, fixed Bar, paired drop-outs, per-Arm Void counts, causes and memory-cap findings.
• Standalone Verdict (verdict::judge): existing paired drop-out and one-tenth threshold tests.

Review coverage

Reviewed against issue-977, pinned at 105f581b6e2d6603594b08dce724c321c53a3480, through head b7fe9d0. Standards: 0 findings. Spec: 3 P2 findings, all reproduced and fixed with retained regression tests. Both axes read all 29 changed files since that fixed point, including the merged #1008 work and generated HTML. The PR against the updated base changes 14 files; all were read. No changed file was left unread and no generated output was excluded.

Unaddressed findings

• Standards: None.
• Spec: None.

Merge Danger

Door: two-way

Blast Radius: Experiments

Changes Trial classification, retry accounting and saved Experiment views.

Built with codex · gpt-6.1-sol · xhigh