[JacobStephens2]

Summary

Run metrics keep each Run's friction and Session-trace signals separate, in its Command record and local read-only views.

Run completes
  Command record v2: outcome, Repairs, CI rounds, findings, stage times
  Session traces: tokens/cache by Model, reported cost, errors, denied/killed/repeated work
    advisory destructive-command checks (never fail the Run)

Local records + Session traces + confirmed Bug origins
  metrics <command> [--json]: Run and child Runs
  metrics [--json]: weekly signals by repository and session setup
  show --json / Console details: the same available Run metrics

Completion accounting keeps Unaddressed findings by Standards, Spec and Security from Delivery's final PR body, and review counts from machine-readable summaries, counted once across Resume. Stage durations are compared with repository Usual ranges. Every Ticket Run and Base fix keeps its own metrics.

Weekly trends keep each signal independent, split by repository, Harness, Harness version, requested Model, Effort, Run kind and session kind, with week/repository filters. Tokens retain the observed Model and cache split; cost is used only when reported. Historical reads reconstruct available signals, preserve missing values and never rewrite old records. Confirmed Bug origins are joined from local Scoring data when read.

The whole-Spec review fixed two integration gaps: historical metrics and late Bug-origin confirmations now appear in show --json and Console details through the same projection as metrics. Regression tests cover both views and a confirmation changing to a rejection between Console requests. Two misleading test names were corrected. No price table, combined score or remote metrics fetch is added.

Closes #981

Evidence

• Before: targeted regression commands demonstrate that historical metrics and late Bug-origin confirmations disappear from show --json despite appearing in metrics.
  After: both exact commands pass; historical minutes are 30.0 and the late Bug-origin count is 1.

  cargo test --test issue_981_spec_review show_includes_metrics_reconstructed_for_a_past_run -- --exact --nocapture
  cargo test --test issue_981_spec_review show_includes_confirmed_bug_origins_from_read_time_scoring_data -- --exact --nocapture

• Test seams: whole binary against fixture records/fake Harnesses, one normalized Session trace, and Console HTTP/browser details.
• A further regression preserves legacy Usual ranges without inventing unrecorded historical minutes; it failed before the reader ordering fix and now passes alongside the Run-health and Run-metrics tests.
• npm test --prefix console: 60 passed, including historical/unfinished Runs and phone-width details.
• cargo fmt --check and cargo clippy --all-targets -- -D warnings: passed.
• Required scripts/affected-tests main: the site tests pass; the Rust run encounters an unrelated host warning from the installed snapd.service in the strict deployment-unit check. SYSTEMD_UNIT_PATH=/usr/lib/systemd/system cargo test --test site_deploy passes all 7 deployment tests. All 7 targets after deployment (spec_push_race, spec_run, status, stop, toolchain, update, user_config) pass in a separate continuation with that same unit-path isolation.
• The final selector run took 1,120.82 seconds (18m41s), including compilation and site checks, before the host's deployment warning stopped it. The continuation passed 156 tests; all selected Rust targets were verified across these runs.
• Full Rust validation uses GIT_CONFIG_COUNT=1 GIT_CONFIG_KEY_0=tag.gpgSign GIT_CONFIG_VALUE_0=false to isolate tag fixtures from this host's global signing setting. User configuration is unchanged.

Unaddressed findings

Standards

None.

Spec

None.

Review coverage

Standards: 1 P3 naming finding, addressed. Spec: 2 P2 details-view integration findings, reproduced and addressed. Each reviewer read all 57 changed files, including generated output, plus supporting standards, domain definitions and ADRs. Unread changed files: none. Generated-output exclusions: none. Final reports and combined files-read lists are retained as standards.md and spec.md with the Spec-review Session evidence.

Merge Danger

Door: two-way

Readers accept Command records with or without metrics; historical reads do not rewrite them. Older binaries may omit version-2 records from their views, while the records remain on disk. A release carrying the changed Factory code-review skill requires an Experiment once #977 is available, as specified by #981.

Blast Radius: metrics

Completion recording, Session-trace accounting, CLI metrics, Console details and their documentation. Destructive-command flags remain advisory.

Built with codex · gpt-6.1-sol · high

Tickets

• Run metrics from the Run, in its Command record #1033 done
• Unaddressed findings by kind #1034 done
• The review's findings by kind #1035 done
• thirdshift metrics: one Run, and weekly trends #1036 done
• Session-trace signals #1037 done
• Destructive commands flagged on every Run #1038 done
• Confirmed Bug origins count against their Runs #1039 landed with Count confirmed Bug origins against their Runs #1304
• A Run's metrics in the Console #1040 done