feat(workspace): per-turn token + cost meter and reasoning blocks #602

Merged

charles merged 1 commit from dev/566 into main

2026-04-30 20:56:41 +00:00

dev commented

2026-04-30 20:40:54 +00:00

Collaborator

Adds token / cost / duration visibility and extended-thinking display to the workspace transcript.

Test plan

Send a foreman chat message; verify the completed assistant bubble shows a footer pill: Xs · Yk+Zk tok · $N.
Hover / focus the pill; verify the tooltip shows cache-creation and cache-read tok counts.
Session with multiple turns: verify the totals strip appears at the top of the transcript and sums correctly.
With an Opus session using extended thinking: verify a collapsible "Reasoning" section appears above the assistant text, default collapsed.
Legacy sessions (no meter data): verify no pill renders and no totals strip.

Closes #566

Adds token / cost / duration visibility and extended-thinking display to the workspace transcript. ## Test plan - Send a foreman chat message; verify the completed assistant bubble shows a footer pill: `Xs · Yk+Zk tok · $N`. - Hover / focus the pill; verify the tooltip shows cache-creation and cache-read tok counts. - Session with multiple turns: verify the totals strip appears at the top of the transcript and sums correctly. - With an Opus session using extended thinking: verify a collapsible "Reasoning" section appears above the assistant text, default collapsed. - Legacy sessions (no meter data): verify no pill renders and no totals strip. Closes #566

dev self-assigned this

2026-04-30 20:40:54 +00:00

dev added 1 commit

2026-04-30 20:40:54 +00:00

feat(workspace): per-turn token + cost meter and reasoning blocks (#566 )

qa / dockerfile (pull_request) Successful in 6s

Details

qa / qa (pull_request) Successful in 2m18s

Details

183c672565

- Extend ForemanMessage with meter fields (input/output/cache tokens,
  cost_usd, duration_ms) and optional thinking[] for reasoning blocks
- Persist meter data from the result event when appendMessage is called
  at the end of each foreman chat task
- Extract thinking blocks in sdk-adapter; propagate through AssistantTurn
  and event-log assistant detail
- Transcript: TurnMeter footer pill (duration · tokens · $cost) with
  hover tooltip for cache breakdown; ReasoningBlock collapsible section
  above assistant text; session totals header strip

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

dev requested review from reviewer

2026-04-30 20:43:17 +00:00

reviewer approved these changes

2026-04-30 20:47:17 +00:00

reviewer left a comment

CI green (qa + dockerfile). All AC met:

Per-turn pill renders on completed assistant bubbles with duration · tokens · cost.
Hover/focus tooltip shows cache-creation + cache-read breakdown.
Session totals strip sums correctly; hidden on legacy sessions.
ReasoningBlock is collapsible, default collapsed.
extractAssistantText already filters only type === "text" blocks, so thinking blocks are excluded at source.

Nit (non-blocking): TurnMeter generates tooltipId via Math.random() on every render — the IDs stay in sync within each commit so there's no functional breakage, but useId() is the idiomatic React 18 pattern and avoids needless churn on parent-triggered re-renders.

CI green (`qa` + `dockerfile`). All AC met: - Per-turn pill renders on completed assistant bubbles with duration · tokens · cost. - Hover/focus tooltip shows cache-creation + cache-read breakdown. - Session totals strip sums correctly; hidden on legacy sessions. - `ReasoningBlock` is collapsible, default collapsed. - `extractAssistantText` already filters only `type === "text"` blocks, so thinking blocks are excluded at source. Nit (non-blocking): `TurnMeter` generates `tooltipId` via `Math.random()` on every render — the IDs stay in sync within each commit so there's no functional breakage, but `useId()` is the idiomatic React 18 pattern and avoids needless churn on parent-triggered re-renders.