[JacobStephens2]

Summary

thirdshift score
  machine lock → Command record/log + Activity log
  previous window end .. Command start (first: seven days)
  configured owners → merged factory PRs → exclude security-fix
  seeded draw across Harness/Model pairs → private weekly sample
thirdshift scores [--json]
  list retained windows, seeds and samples by ISO week

Adds [scoring] defaults to Setup and documents both commands. Scoring uses the existing Inbox and Stop interfaces; failed searches land in Needs you and advance no window.

Closes #1024

Evidence

• Before: cargo test --test scoring first_window_looks_back_seven_days_and_is_a_recorded_command failed because score was parsed as an Issue URL. The sampling test initially returned 29 unfiltered candidates instead of a three-pair sample; the Stop test initially rejected the displayed Scoring Command name.
  After: Scoring and Setup integration suites passed (10 + 30 tests), including the factory-writer/Scoring-reader round trip. The compact-help guard and complete-config Setup fixture pass. cargo check --all-targets and cargo clippy --all-targets -- -D warnings passed. The full CI Rust job passed on final head 6cd2e01; all applicable PR checks are green.
• Final local selection: NEXTEST_TEST_THREADS=12 /usr/bin/time -p scripts/affected-tests issue-978 passed all 43 site browser tests. The duplicate Rust run was deliberately interrupted once the full Rust CI job passed on that exact head, reusing CI as docs/testing.md requires. This local attempt took 819.55 seconds before cancellation and included 7.14 seconds of incremental Rust compilation; it is not a completed full-suite timing.
• The production GraphQL query was checked read-only against GitHub: 210 matches, 50 returned with a next-page cursor. Its page sizes fit GitHub's node budget.

Test seams

• Whole binary against fake gh: score, scores, status/Inbox and Stop; retained samples, Command records, Command logs and Activity logs.
• User config and Setup through their existing public interfaces.

Review

Fixed point: issue-978 (1ff8b2076d8381ecc3ec1c30b7483ac6aac53aa7). Standards and Spec reviewed independently, with a focused follow-up at 258146a. Standards raised one maintenance judgment, addressed by consolidating Built-with parsing and formatting under harness::built_with; no hard violations or outstanding judgments remain. Spec has no findings. Both axes read 34 of the final 35 changed files. Unread by both axes: src/setup.rs, changed after their follow-up to complete the config fixture with the new Scoring defaults. The later help compaction in src/main.rs was validated after review. Generated-output exclusions: none. Reports, including final files-read lists and coverage accounting, are retained locally as .thirdshift-review-erJbHo/standards.md and .thirdshift-review-erJbHo/spec.md.

Unaddressed findings

Standards

None.

Spec

None.

Security

None. Guidance-mode review of the issue-1024 diff against the pinned merge base found no introduced Security finding: gh runs via argv exec with no shell, configured owners are validated before search interpolation, security-fix label matching is case-insensitive and fail-closed on incomplete label evidence, samples persist privately with an OsRng seed, and over-limit searches advance no window. No fixes committed.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

The commands and configuration can be reverted; Scoring retains local samples without changing repositories.

Blast Radius: Scoring

Existing Runs share the User config reader and Built-with formatter. Samples remain private under ~/.thirdshift/evals/scoring/; Grader sessions and criterion Scores belong to subsequent work. Searches above GitHub's 1,000-result limit or with incomplete issue-label evidence fail without advancing the window.

Built with codex · gpt-6.1-sol · xhigh