[JacobStephens2]

Summary

Code review Trial
  prepared historical diff → production code-review skill → retained Standards + Spec reports
Code review Experiment
  fresh or legacy replay import → 3 Trials per Arm per Eval task → pass^3

Adds the review-only Session kind and Session prompt, uses the implement session's reviewer-independence rules, and regenerates the prompts page. Replay imports now supply the required Code review Trial inputs; reimport completes missing legacy inputs while preserving local criteria, comments and custom Trial inputs.

Closes #1016

Evidence

• Before: a_code_review_trial_reviews_the_prepared_diff_and_keeps_both_reports failed with unknown code-review; the Experiment regression failed at Code-review Trials require #1016.
  After: both regressions pass, including mixed outcomes proving pass^3 is 1 for an Arm passing all three Trials and 0 for an Arm passing only two.
• Before: the Spec reviewer's exact reproduction and full import-to-Experiment regression failed with Eval task needs a [trial] table.
  After: the retained regression imports a suite, runs all six Trials, keeps both reports, and reports pass^3. The legacy-import regression also passes and preserves local edits.
• cargo test --test evals --test experiments --test trial: 56 passed, 2 subprocess entrypoints ignored after merging the updated Base branch. Both Code review regressions and the mechanical-score regression pass after the fixture cleanup.
• cargo test --bin thirdshift prompts_page: 12 passed; regeneration and docs checks pass. Site browser checks: 43 passed.
• cargo check --all-targets and cargo fmt --check pass. CI on the exact final head, d3bafe1, passed the full Rust suite, Clippy, formatting, site and Console checks.
• scripts/affected-tests ac8d24494e52f42268e524f716be20e4f71af361 passed all 43 site checks. Its duplicate Rust run was stopped after exact-head CI passed, reusing that result as docs/testing.md directs.

Merge Danger

Door: two-way

Blast Radius: Evals

Enables Code review Trials and completes missing private replay-manifest inputs. Existing custom Trial tables and grading criteria are preserved.

Test seams

• The whole binary's versioned --trial protocol against fake Harnesses: the Code review prompt, reviewer independence, and retained reports.
• The whole binary's evals import-replays, experiment, and experiments --json interfaces against fake Harnesses: imported and prepared review inputs, preservation of local edits on reimport, three Trials per Arm, and pass^3.

Review coverage

Reviewed against the updated issue-977 head, pinned at ac8d24494e52f42268e524f716be20e4f71af361, after merging its Void Trial changes into this branch. Fresh independent reviewers inspected the reconciled diff and the final fixture cleanup.

• Standards: 1 low-priority duplication finding, addressed by sharing the report-setup fixture. Follow-up confirmed resolution.
• Spec: 1 P2 finding, reproduced as written, fixed, and verified with the retained regression. Follow-up confirmed resolution.
• Both reviewers read all 21 changed files, including generated output. Generated-output exclusions: none. Unread changed files: none.

Unaddressed findings

• Standards: None.
• Spec: None.

Built with codex · gpt-6.1-sol · xhigh