[JacobStephens2]

Summary

Run metrics and Experiments now resolve equivalent trace evidence through one Session usage module. Partial totals preserve omitted token/cache fields and costs; reported zeroes remain authoritative.

Run metrics (normalized events) ─┐
                                ├─ Session usage ─ consumer grouping / presentation
Experiments (retained spans) ────┘

No stored records or trace formats change. Historical metrics use the shared reader without rewriting saved completion snapshots. The glossary, navigation index, guide and hand-written site pages describe the settled rules.

Closes #1319

Evidence

• Before: cargo test --test metrics partial_model_totals_preserve_known_call_usage -- --exact --nocapture failed on the review base: input/cache reads became null after an output-only Model total.
  After: the same regression passes with input 16, output 3, cache reads 5; replacing raw input with 20 yields inclusive input 25, retaining cache reads and call cost.
• Ten paired event/span interface tests pass, exercising sparse/scoped amendments, repeated completions, latest totals, cache conventions, overlapping turns, unknown Models, zeroes, cost precedence and saturating addition.
• Focused metrics, completion-snapshot and Experiment regressions pass. Existing Experiment tests retain Void attempts, Resumes, Grader exclusion and selected-span read counts.
• cargo check --all-targets and cargo clippy --all-targets -- -D warnings pass. All 50 site browser tests pass on the final commit.
• An earlier full run stopped at distinct_suite_task_pairs_keep_distinct_trial_results after its fake Model Grader failed. The exact isolated rerun passed: cargo nextest run --test experiments -E 'test(=distinct_suite_task_pairs_keep_distinct_trial_results)' (1 passed, 32.6 s). No product-code change was made for the unreproduced failure.
• Final validation on f540a16c: scripts/affected-tests main passed all 50 site browser tests and all 3,022 Rust tests (66 skipped). Timed selector wall time: 1,954.34 s, including 52.41 s of Rust compilation; Nextest test execution: 1,715.83 s.

Test seams

• session::usage::from_events and session::usage::from_spans: equivalent normalized events and recorded retained spans, returning one owned Session usage summary; summary addition combines distinct Sessions.
• thirdshift metrics ... --json: historical reconstruction, partial Model totals, and completion snapshots.
• Existing Experiment JSON/text and captured-reader seams: partial totals, unknown Models, Void attempts, Resumes, Grader exclusion and one spans read per selected Session.

Review coverage

Independent thirdshift-code-review Standards and Spec reviews against main (6ecf0715) each found zero findings and inspected all 15 changed files, including generated site/docs/index.html. Neither axis excluded generated output. No changed file was left unread.

Reports, including each axis's complete files-read list, are retained at .thirdshift-review-TcwoHu/standards.md and .thirdshift-review-TcwoHu/spec.md for the Session trace.

Unaddressed findings

Standards

None.

Spec

None.

Security

None. Guidance-mode review of 6ecf0715...HEAD (15 files, accounting consolidation into src/session/usage.rs) found no introduced vulnerability: inputs are the operator's own local session traces and retained spans (no lower-trust principal, no new sink, network, auth, secret, or filesystem-discovery path); the thirdshift.model_usage JSON parse, saturating u64 arithmetic, and cost handling moved verbatim from the prior per-consumer implementations; the trace::set visibility widening stays inside the crate-private session module; and the harness cache-convention check is behavior-identical (Claude | Grok). No proof-of-concept failed, so no Security fix was made. session::usage unit tests (10 passed) and the partial_model_totals_preserve_known_call_usage metrics regression pass. No pre-existing Base-branch findings were tested or recorded in this review.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

Reader/accounting refactor; no data migration. Existing completion snapshots remain fixed.

Blast Radius: accounting

New or historically reconstructed Run metrics and Experiment token/cost summaries use the shared rules.

Built with codex · gpt-6.1-sol · xhigh