[JacobStephens2]

Summary

metrics --json, /api/v1/metrics and the Console Metrics page now carry four paired signal groups. All three surfaces read the same document, built from local files only: Command records, github-inbox.json (read from the file alone, with no quota check and no gh), and the Scoring checkpoints.

 trends()
   facts = outcomes::facts(snapshot)           # each counted Command + its Stop request
   outcomes::read(facts)                       # existing stacks and comparison
+  groups::read(facts, window, Sources {
+      origins, kept_pull_requests(logs_dir), read_scores(), filters })
+    yield.pull_requests      merged / closed / open / not kept, survival rate
+    quality.bug_origins      merged PRs by merge bucket, confirmed Bug origins
+    quality.scores           weekly Grader pass rates, shown once calibrated (blind Scores withheld)
+    quality.interventions    Needs you and Stop counts
+    efficiency               Succeeded-only minutes/Repairs/CI rounds, first-pass clean,
+                             minutes and $ per success, wasted time/spend share (by Harness and Model)
+    friction                 % stages above Usual range, per-Session friction, Security holds
   weeks|days rows            repairs / ci_rounds / total_minutes now cover Succeeded only

Each group line holds its whole-window totals, a value for each bucket (drawn as a static SVG sparkline) and, for a bounded window, current/prior/delta. The Change column also shows each window's n. Raw Standards and Spec finding counts appear in no group; only security > 0 holds do.

Evidence

Execution evidence

• Before (review findings, on 495bc84):

  • cargo test --test console metrics_page_heads_the_yield_group_without_a_bounded_window → panicked at tests/console/metrics.rs:296:5 … test result: FAILED. 0 passed; 1 failed
  • cargo test --bin thirdshift logs::metrics::scope::tests → panicked at src/logs/metrics/scope.rs:256:9 … FAILED. 2 passed; 1 failed
  • cargo test --test metrics confirmed_bug_origins_show_without_a_kept_answer → FAILED
  • cargo test --test console metrics_page_heads_the_yield_group_when_unbounded → FAILED

  After (6cd1945):

  • cargo test --test metrics → test result: ok. 44 passed; 0 failed; 1 ignored
  • cargo test --test console metrics → ok. 9 passed; 0 failed
  • cargo test --bin thirdshift metrics → ok. 14 passed; 0 failed

• Acceptance tests (tests/metrics.rs), all pass:

  • fast_failures_and_crashes_raise_waste_without_flattering_succeeded_means: after 2 fast failures and a crash, the Succeeded means are unchanged, failures rise to 3, and both wasted shares rise.
  • kept_pull_requests_bug_origins_and_calibrated_scores_read_no_github: gh_calls() is empty.
  • stage_bottlenecks_friction_and_interventions_compare_with_the_preceding_window.

• Parity: tests/console/api.rs still checks that /api/v1/metrics equals metrics --json for every window. metrics_page_shows_the_four_signal_groups_as_the_api_reports_them checks the page against the API.

• scripts/affected-tests origin/issue-2036:

  • npm test --prefix site: PASS
  • cargo build --locked: PASS
  • npm test --prefix console: PASS. The Metrics browser test checks all four headings and sparkline strokes in every Theme at 390px, with no motion.
  • cargo test: FAIL, only in tests/process.rs findings::*. Which test fails changes from run to run (merged_fix_is_rechecked_after_a_later_fix_fails, fixing_dispatches_each_reproduced_finding_…), and they pass alone. The same binary fails the same way on the base commit: in a temporary worktree at origin/issue-2036, cargo test --test process findings:: -- --test-threads=4 gave FAILED. 17 passed; 1 failed. Every test target after process passed with cargo test --no-fail-fast --test <each>.

Visual evidence

I made captures at 1440×900 and 390×844 of /metrics?window=7d with the same fixture: 42 Ticket and Review runs, kept PR states, and Session signals. Publishing them to a secret Gist was refused by this session's permission settings, so they are not embedded here. They are on the build machine at /tmp/before-desktop.VuQdxp.png, /tmp/after-desktop.PLooVg.png, /tmp/before-mobile.1aNEdW.png and /tmp/after-mobile.JO7U1Z.png.

• Before: Outcomes, plus one Signals table with outcome-blind means.
• After: four group sections, each with sparkline and Change columns. At 390px the tables scroll inside their regions and the page does not overflow.

Merge Danger

Door: two-way

The change is read-only. Nothing is written to records, github-inbox.json or the Scoring checkpoints.

Blast Radius: consumers

• The weeks/days rows' repairs, ci_rounds and total_minutes now cover Succeeded Commands only. A consumer that relied on outcome-blind means will see different numbers.
• groups is a new required field in the OpenAPI document.

Review coverage

• Reports: .thirdshift-review-QXR9uO/standards.md and spec.md. Counts: Standards 11, Spec 6.
• Excluded as generated: site/docs/index.html, whose source is docs/guide.md; npm run docs:check --prefix site passes.
• Changed files the review left unread: src/logs/reading.rs, src/logs/reading/attribution.rs and docs/code-navigation.md. These changed only in the review-fix commit 6cd1945, which shares the pull-request URL identity and updates the code-navigation rows.

Unaddressed findings

Standards

• 1 (part): needs_you is narrower than the glossary's Needs you. I renamed the "attention" group to interventions. I kept the narrower count on purpose: Q1.4 in docs/grill-with-docs/2036-console-metrics.md leaves live operational state (Needs you's Silent and Slow rows) on the existing Inbox pages. docs/guide.md now says the count covers Commands that entered Needs you when they ended.
• 6 (part): kept_pull_requests repeats load's decode, and Store::scores repeats measure's pair closure.
  • load also asks GitHub::quota_waits(), and the issue requires that metrics make zero GitHub calls, so the file-only read is the point of the separate function. I did fix the legacy row-only History keys this finding exposed: kept_pull_requests now skips keys without ::.
  • The two pair closures build pairs from different types (the full Week and the narrow Sampled projection), and both now feed the one shared calibration::measure_pairs.
• 8: message chains into command.record.run_metrics. These follow how the existing reader in src/logs/metrics/trends.rs (collect) walks the same fields. Moving them onto Command would be a refactor outside Survivorship-safe efficiency, session friction, kept PR survival, and post-merge quality signals #2066.
• 9: attempt_minutes reads the filesystem, and trends() calls read_scores() directly. This matches the existing ConfirmedCounts::read in src/logs/metrics/bug_origins.rs, which calls crate::scoring::read_confirmed_origins() the same way. The reviewer rated it minor.

Spec

• 2: attempt time is missing for ended Commands without Run metrics. The reviewer's own test passes as written: cargo test --test metrics a_failed_command_without_run_metrics_still_counts_its_attempt_time → test … ok. Snapshot reading already rebuilds run_metrics (with total_minutes from ended_at) for every ended record that lacks them; see src/logs/reading.rs RunMetrics::historical. The test stays in the suite.
• 5: no single fixture runs through all three surfaces. tests/console/api.rs (the parity(...) loop over /api/v1/metrics windows) already asserts that the API equals metrics --json for the whole document, groups included. metrics_page_shows_the_four_signal_groups_as_the_api_reports_them then checks the page against the API.

Closes #2066

Built with claude 2.1.314 · opus · default effort

Security review outcome

No unaddressed introduced findings.