[takashi-matsuyama]

Summary

Billing-grade work-measurement v2: improve the fidelity of Codex active time and add a model-compute measure. Builds directly on the basou stats work shipped in 0.5.0.

The Codex import adapter previously derived active time the same way as the Claude adapter — gap-capped point timestamps (the engaged-turns method). But a Codex rollout records explicit per-turn boundaries (task_started / task_complete, paired by turn_id, with a reported duration_ms), so the in-turn duration can be the log's real wall-clock span instead of a gap-capped approximation.

What changed

• Codex active time now uses real per-turn intervals. Active time is the union of the real per-turn intervals (in-turn, uncapped, crediting the session's final turn) and the gap-capped engagement series (between-turn bridging). active_time_method becomes turn-intervals when any turn interval contributed, else engaged-turns.
  • The active-time semantics are unchanged — still human-engaged time with idle gaps over 5 minutes excluded. Only the in-turn precision improves, so a re-imported Codex session's active time typically rises (long turns are no longer truncated at the gap cap, and the final turn is counted).
  • Claude imports are unchanged: their transcripts carry no explicit per-turn duration, so no false precision is introduced.
• New session.metrics.machine_active_time_ms (additive, optional): model compute time = summed per-turn duration. A subset of a session's active time. A plain sum, not wall-clock-deduplicated, so two concurrent sessions can sum past their billable (union) active wall-clock — intended (two models working at once is two machine-hours in one wall-clock hour). Surfaced as a Model working line in basou stats (and --by-source / --by-day / --json), a machine segment on basou session show's Work: line, and the basou view Stats tab. Re-import (--force, or basou refresh) to backfill.

Robustness (from adversarial review)

• De-duplicate task_complete by turn_id so a repeated completion cannot double-count machine time and break the machine <= active subset.
• Emit machine_active_time_ms only when every completed turn carried a duration; a session mixing duration-bearing and duration-less completions would otherwise report a partial compute total as if complete.
• Clamp a turn start reconstructed from duration_ms (when task_started is absent) to the session floor, so no active interval predates the session, and bound each turn's machine contribution to its in-session span so a clamped turn stays a subset of active time.

Compatibility

machine_active_time_ms and the turn-intervals method are additive and optional; no schema_version bump. Existing imported sessions are unaffected until re-imported (consistent with how active_* was introduced).

Test plan

• pnpm -r build && pnpm typecheck && pnpm test && pnpm lint all green (707 core + 417 cli + 1 sdk; lint clean apart from pre-existing unrelated warnings).
• New unit tests cover: real turn intervals beating in-turn gap-cap, between-turn bridging vs over-cap idle, start reconstruction, turn_id de-dup, partial-duration omission, reconstructed-start clamping, schema validation, work-stats roll-up (per-session / by-source / by-day / totals), and the new rendered lines in stats / session show.
• Validated against real local Codex rollouts: machine <= active holds for every session, no active interval precedes started_at, and older rollouts without duration_ms correctly omit machine time while still deriving turn-interval active time.

🤖 Generated with Claude Code