[JacobStephens2]

Summary

Pause Experiments at live Claude warnings/rejections and Codex quota reports at 85% in either window, including Trials from older Arm builds. Ordinary Runs continue.

plan-limit warning
  → mark a cut-short Trial void
  → stop both workers from starting Trials or retries
  → pause the Experiment
rerun the same Experiment
  → retain finished Trials and resume remaining work

Reconciles the latest issue-977 Graders and Verdicts while preserving limit-pause attempts and their fresh-retry budget.

Closes #1008

Evidence

• Before: a_live_plan_limit_cuts_short_a_trial_and_keeps_its_evidence returned succeeded after a warning; the delayed-stop regression started four Trials instead of two.
  After: both regressions pass, alongside native Codex report detection, pause/resume and ordinary Run coverage.
• Before: the Spec review's exact reproducer, built against baseline b8b9ec9, completed all six Trials despite the Baseline warning.
  After: the retained a_baseline_built_before_this_change_pauses_at_its_live_warning test passes with that actual baseline binary and with its fixed version-1 adapter for CI.
• Before: repository-integrity grading replaced a warning's pause cause, so rerunning with --void-retries 0 exhausted the retry budget.
  After: a_plan_limit_pause_keeps_its_retry_budget_even_if_the_repository_changed passes: the warning remains a resumable pause and rerunning starts a fresh Trial.
• Local validation: cargo test --test experiments --test trial --test fresh_run passes (56 tests; three subprocess entrypoints ignored). The actual pinned older-baseline regression also passes. cargo clippy --all-targets -- -D warnings, cargo fmt --check, and all 43 site tests selected by scripts/affected-tests issue-977 pass.
• CI passed for exact HEAD 969d133, including the full Linux Rust suite, site and Console checks. Release builds and smoke gate passed.
• Timed local selector: NEXTEST_TEST_THREADS=2 NEXTEST_STATUS_LEVEL=pass /usr/bin/time -p scripts/affected-tests issue-977 took 574.95 seconds, including compilation and site checks. After 43 site tests and 134 Rust tests passed, the redundant Rust run was deliberately interrupted when exact-head CI passed, as docs/testing.md directs. This timing is not a completed full-suite benchmark; CI supplies the full-suite result.

Test seams

• Harness Interpretation's live decoding and completion interface.
• Version-1 Trial CLI against fake Harness processes.
• Experiment CLI, saved results and two-worker pause/resume behavior, including older Arms.
• Ordinary Run CLI and Delivery after quota warnings.

Review coverage

Standards and Spec each reviewed all 27 changed files against issue-977, including generated site/docs/index.html. Fresh independent reviews after upstream reconciliation found no further findings. No changed file was left unread; no generated output was excluded.

The Standards P3 was addressed by separating pause policy from quota decoding. The Spec P1 was reproduced, fixed in Experiment supervision, and retained as a regression test.

Unaddressed findings

• Standards: None. The one initial finding was addressed.
• Spec: None. The one initial finding was verified and addressed.

Merge Danger

Door: two-way

No Trial request/result version change or data migration.

Blast Radius: Experiments

Trials now stop early at reported limits. An already-running peer may finish; the same Experiment resumes on rerun.

Built with codex · gpt-6.1-sol · xhigh