fix(web): show Codex round usage metadata

[techotaku39]

Summary

• Mark Codex token usage events explicitly and convert them into round metadata.
• Display Codex usage as Round: <duration> · <turns> turn(s).
• Derive live Codex round duration from HAPI user invocation to final assistant completion.
• Exclude replayed local Codex history from derived Round duration because replay timestamps do not represent the original turn timing.
• Preserve the existing Claude Code usage display behavior.

Validation

• bun typecheck — passed.
• bun run test:e2e -- terminal-wrap-fidelity.spec.ts — 2 passed.
• Codex launcher tests — 105 passed.
• Related Web tests — 103 passed.
• bun run build — passed.
• Live Codex round demonstrations — 4 passed, covering 1.8 seconds, 65 seconds, 65 minutes, and 26 hours.

Related Issues

None

AI Disclosure

Implemented and tested with OpenAI Codex (GPT-5.6).

[github-actions]

Findings

• [Major] Imported local Codex history receives replay-time Round durations. See the inline finding at cli/src/codex/codexLocalLauncher.ts:220.

Summary

• Review mode: initial
• One correctness issue found: replayed transcript usage is promoted to Codex round metadata even though its timestamps are hub replay stamps rather than original turn timing.

Testing

• Not run (automation; PR content was not executed). GitHub test, integration, and drift-gate checks are passing. Add a regression for imported local history that asserts no derived duration unless original timestamps are available.

HAPI Bot

[github-actions]

[MAJOR] Imported history gets fabricated Round durations

This branch now adds flavor: codex to replayed token counts, so the web normalizer marks them as Codex and aggregateResponseGroups derives a duration. However replayed user/agent entries are sent without their original createdAt; the hub therefore gives each no-localId row replay-time createdAt/invokedAt (hub/src/store/messages.ts:145). The displayed value measures replay/loader speed, not the historical turn.

Suggested fix:

? { ...message, model: transcriptModel, hapiUsageScope: imported-history }

Keep the Codex flavor only on managed/live usage, or propagate the scope and skip duration derivation for imported history. Add an import regression covering this path.

[github-actions]

Findings

• None.

Summary
Review mode: follow-up after new commits

• The previous imported-history duration issue is resolved at cli/src/codex/codexLocalLauncher.ts:220: replayed token counts no longer carry the Codex marker, and web/src/chat/reducerTimeline.test.ts:239 covers the no-summary behavior.
• No additional issues found in the latest full diff.
• Residual risk: duration is covered by separate reducer/runtime tests rather than one full normalize-to-render pipeline test.

Testing

• Not run (automation; PR content was not executed). GitHub integration and drift-gate pass; test was still in progress at review time.
• Suggested: add a normalize → reduceTimeline → aggregateResponseGroups regression asserting imported history has no Round while live usage derives duration from the user invocation.

HAPI Bot