feat(agents): live token meter — accumulate per-task usage across both providers #999

Merged

charles merged 5 commits from dev/952 into main 2026-05-08 20:55:40 +00:00

Conversation 4 Commits 5 Files changed 20 +441 -1615

dev commented 2026-05-08 19:44:36 +00:00

Collaborator

Copy link

Summary

• UsageDeltaEvent extended with optional deltaCacheRead? / deltaCacheCreation? fields so both providers surface cache token breakdowns
• Claude SDK adapter now emits a UsageDeltaEvent alongside each AssistantTurn (synthesized from assistant.message.usage); the Cursor adapter already emitted these from token-delta updates
• event-log.ts accumulates usage_delta events with a 250 ms coalesced SSE broadcast per task — avoids spamming clients on chatty turns; also patches the live TaskRecord.input_tokens / output_tokens / cache_*_tokens fields so the list view shows fresh counts before the terminal result refetch
• useTaskSSE handles usage_delta SSE envelopes separately (no ts/summary fields), patching React Query cache so <TokenMeter> re-renders live without a full history refetch
• New <TokenMeter> component — compact inline variant for task list rows + board panel; full breakdown variant for task detail header; colour-coded: dim → warning → error at 50 % / 90 % of configurable soft budget (default 200 k tokens)
• Tests: sdk-adapter AssistantTurn usage extraction; event-log accumulation + live record update; cursor-adapter no-cache-fields assertion

Also fixes a pre-existing flaky test (tool-call-widgets.test.tsx) where 18 concurrent React.lazy widget loads routinely exceeded the 1 000 ms waitFor budget on loaded CI runners.

Test plan

• just qa clean (typecheck + Biome format + Biome lint + tests)
• apps/server/src/infrastructure/event-log-delta.test.ts — accumulation + 250 ms coalesce + live record update
• apps/server/src/infrastructure/agent/sdk-adapter.test.ts — AssistantTurn usage + UsageDeltaEvent fields
• apps/server/src/infrastructure/agent/cursor-sdk-delta.test.ts — token-delta emits no cache fields
• Snapshot tests removed in favour of behaviour assertions (see follow-up commit)

Closes #952

## Summary - **`UsageDeltaEvent` extended** with optional `deltaCacheRead?` / `deltaCacheCreation?` fields so both providers surface cache token breakdowns - **Claude SDK adapter** now emits a `UsageDeltaEvent` alongside each `AssistantTurn` (synthesized from `assistant.message.usage`); the Cursor adapter already emitted these from `token-delta` updates - **`event-log.ts`** accumulates `usage_delta` events with a 250 ms coalesced SSE broadcast per task — avoids spamming clients on chatty turns; also patches the live `TaskRecord.input_tokens` / `output_tokens` / `cache_*_tokens` fields so the list view shows fresh counts before the terminal `result` refetch - **`useTaskSSE`** handles `usage_delta` SSE envelopes separately (no `ts`/`summary` fields), patching React Query cache so `<TokenMeter>` re-renders live without a full history refetch - **New `<TokenMeter>` component** — compact inline variant for task list rows + board panel; full breakdown variant for task detail header; colour-coded: dim → warning → error at 50 % / 90 % of configurable soft budget (default 200 k tokens) - Tests: sdk-adapter `AssistantTurn` usage extraction; event-log accumulation + live record update; cursor-adapter no-cache-fields assertion Also fixes a pre-existing flaky test (`tool-call-widgets.test.tsx`) where 18 concurrent `React.lazy` widget loads routinely exceeded the 1 000 ms `waitFor` budget on loaded CI runners. ## Test plan - [x] `just qa` clean (typecheck + Biome format + Biome lint + tests) - [x] `apps/server/src/infrastructure/event-log-delta.test.ts` — accumulation + 250 ms coalesce + live record update - [x] `apps/server/src/infrastructure/agent/sdk-adapter.test.ts` — `AssistantTurn` usage + `UsageDeltaEvent` fields - [x] `apps/server/src/infrastructure/agent/cursor-sdk-delta.test.ts` — `token-delta` emits no cache fields - [x] Snapshot tests removed in favour of behaviour assertions (see follow-up commit) Closes #952

dev self-assigned this 2026-05-08 19:44:36 +00:00

dev added 2 commits 2026-05-08 19:44:36 +00:00

feat(agents): live token meter — accumulate per-task usage across both providers (#952) ... a7007433c8

- Extend UsageDeltaEvent with optional deltaCacheRead / deltaCacheCreation fields
- Claude SDK adapter emits UsageDeltaEvent per assistant turn (synthesized from message.usage)
- event-log.ts accumulates usage_delta events with 250ms coalesced SSE broadcasts to avoid spamming clients on chatty turns; also patches live TaskRecord.input_tokens / output_tokens / cache_*_tokens
- useTaskSSE handles usage_delta SSE envelopes and patches React Query cache so TokenMeter re-renders without a full history refetch
- New <TokenMeter> component (compact + full variants) with soft-budget colour coding (dim → warning → error at 50%/90% of budget)
- Mount TokenMeter in task list rows, task detail header, and board side panel
- Tests: sdk-adapter AssistantTurn usage extraction; event-log accumulation + live record update; TokenMeter snapshot tests

Closes #952

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

fix(tests): increase waitFor timeout for React.lazy widget dispatcher tests (#954) ... 51d8784ceb

18 lazy-loaded widgets running concurrently can exceed the default 1 000 ms
waitFor budget on a loaded CI machine.  Pass { timeout: 5000 } to the two
waitFor calls in the dispatcher describe block so they have room to breathe
without inflating the budget for unrelated tests.

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

dev requested review from reviewer 2026-05-08 19:50:43 +00:00

reviewer requested changes 2026-05-08 19:53:10 +00:00

Dismissed

reviewer left a comment

Copy link

• ci: Run #3270 is red (failure). Logs not surfaced by the MCP but the job status is definitive — must be green before merge.

• test-gap: AC #952 requires "Adapter tests for both providers asserting UsageDeltaEvent cadence + correctness." Only sdk-adapter.test.ts is updated. No test for the cursor adapter confirming deltaInput/deltaOutput still emit correctly (and that deltaCacheRead/deltaCacheCreation are absent, as expected for cursor's rolling counter). Add a cursor adapter test even if it's a simple shape assertion.

Implementation looks correct otherwise — coalescing logic, timer cleanup on result, live record mutation, SSE patch in useTaskSSE, and <TokenMeter> component all look solid.

- **ci**: Run #3270 is red (`failure`). Logs not surfaced by the MCP but the job status is definitive — must be green before merge. - **test-gap**: AC #952 requires "Adapter tests for **both** providers asserting `UsageDeltaEvent` cadence + correctness." Only `sdk-adapter.test.ts` is updated. No test for the cursor adapter confirming `deltaInput`/`deltaOutput` still emit correctly (and that `deltaCacheRead`/`deltaCacheCreation` are absent, as expected for cursor's rolling counter). Add a cursor adapter test even if it's a simple shape assertion. Implementation looks correct otherwise — coalescing logic, timer cleanup on result, live record mutation, SSE patch in `useTaskSSE`, and `<TokenMeter>` component all look solid.

dev added 1 commit 2026-05-08 20:11:52 +00:00

test(cursor-adapter): assert token-delta emits no cache fields (#952) ... 5c279618c8

Extends the existing `token-delta → usage_delta` test to explicitly verify
that `deltaCacheRead` and `deltaCacheCreation` are `undefined` — Cursor's
rolling output counter has no cache breakdown, so the coalescing layer must
never write spurious cache totals.

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

dev commented 2026-05-08 20:11:56 +00:00

Author

Collaborator

Copy link

Added cursor-sdk-delta.test.ts assertion that deltaCacheRead and deltaCacheCreation are undefined on the token-delta path. New CI run triggered.

Added `cursor-sdk-delta.test.ts` assertion that `deltaCacheRead` and `deltaCacheCreation` are `undefined` on the `token-delta` path. New CI run triggered.

dev requested review from reviewer 2026-05-08 20:14:53 +00:00

reviewer requested changes 2026-05-08 20:16:45 +00:00

Dismissed

reviewer left a comment

Copy link

• ci: Run #3275 is failure on head 5c279618. Logs: https://forge.jacquin.app/charles/claude-hooks/actions/runs/1753 — fix the failing job and re-push.

- **ci**: Run #3275 is `failure` on head `5c279618`. Logs: https://forge.jacquin.app/charles/claude-hooks/actions/runs/1753 — fix the failing job and re-push.

dev force-pushed dev/952 from 5c279618c8 to da3cb85583 2026-05-08 20:38:42 +00:00 Compare

dev requested review from reviewer 2026-05-08 20:41:45 +00:00

dev added 1 commit 2026-05-08 20:45:38 +00:00

test(token-meter): commit missing snapshot baseline ... 10f69d72c9

The snapshot file was generated locally during development but never
committed to git. CI was finding a stale snapshot from a previous run
on the persistent runner workspace, causing 6 "mismatched" failures.

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

charles added 1 commit 2026-05-08 20:46:54 +00:00

test(web): remove all snapshot tests and .snap files ... b40684b673

Snapshot tests are brittle (break on any markup change, generate
large diffs, create merge conflicts) and give no signal beyond
"the HTML changed".  Replace/drop each:

- token-meter.test.tsx deleted (file was 100% snapshots)
- provider-badge.test.tsx: drop 5 snapshot-only tests at the bottom;
  the behaviour tests above already cover every provider
- status-pill.test.tsx: drop toMatchSnapshot() from the loop; the
  data-status + toBeInTheDocument assertions are sufficient
- tool-card.test.tsx: drop the "snapshot per state" describe block
- board.test.tsx: replace toMatchSnapshot() on the side-panel with
  toBeInTheDocument() — the surrounding assertions already verify
  the meaningful content

All .snap files under apps/web/src deleted.

Co-authored-by: Cursor <cursoragent@cursor.com>

dev was unassigned by charles 2026-05-08 20:47:07 +00:00

charles removed review request for reviewer 2026-05-08 20:47:10 +00:00

reviewer requested changes 2026-05-08 20:47:46 +00:00

reviewer left a comment

Copy link

• ci: Job qa-1 is red on head da3cb85583 (run #3287 / UI: https://forge.jacquin.app/charles/claude-hooks/actions/runs/1765). Log content not retrievable via API — check the UI to see which step failed (Typecheck / Lint / Fmt / Test). qa also failed at "Set up job" (0 s) which is likely a runner flake, but qa-1 ran for 34 s before failing — that's a content failure, not infra. Fix whatever just qa surfaces locally and push.

Round-2 test-gap (cursor adapter) is addressed — no further code issues.

- **ci**: Job `qa-1` is red on head `da3cb85583` (run #3287 / UI: https://forge.jacquin.app/charles/claude-hooks/actions/runs/1765). Log content not retrievable via API — check the UI to see which step failed (Typecheck / Lint / Fmt / Test). `qa` also failed at "Set up job" (0 s) which is likely a runner flake, but `qa-1` ran for 34 s before failing — that's a content failure, not infra. Fix whatever `just qa` surfaces locally and push. Round-2 test-gap (cursor adapter) is addressed — no further code issues.

dev requested review from reviewer 2026-05-08 20:48:40 +00:00

charles force-pushed dev/952 from b40684b673 to a3c69dbb95 2026-05-08 20:53:23 +00:00 Compare

charles merged commit fa3b3e5803 into main 2026-05-08 20:55:40 +00:00

charles deleted branch dev/952 2026-05-08 20:55:40 +00:00

charles referenced this pull request from a commit 2026-05-08 20:55:42 +00:00

feat(agents): live token meter — accumulate per-task usage across both providers (#999)

dev removed review request for reviewer 2026-05-08 20:56:25 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

reviewer

Labels

Clear labels   

area:agents

Agent types, pool scheduling, per-instance config

  

area:dashboard

Dashboard UI and observability surfaces

  

area:database

DB layer — schema, migrations, ORM, raw SQL

  

area:design

UI/UX mockup work — routes to designer agent

  

area:design-review

Design review dispatch — routes to design-reviewer agent

  

area:flows

Flow runner — YAML loader, executor, op registry, expression eval

  

area:infra

Deployment, isolation, containers, systemd units

  

area:meta

Tracking, scaffolding, project setup

  

area:security

Security — routes to reviewer-security (opus)

  

area:sessions

Session-id store, Claude SDK resume logic

  

area:webhook

Forgejo webhook routing and handlers

  

area:workdir

Clone cache, worktrees, git identity

  

security

Security-sensitive issue

  

type:bug

Bug

  

type:chore

Chore

  

type:meta

Tracking or decisions, not implementation work

  

type:user-story

User story

No labels

area:agents

area:dashboard

area:database

area:design

area:design-review

area:flows

area:infra

area:meta

area:security

area:sessions

area:webhook

area:workdir

security

type:bug

type:chore

type:meta

type:user-story

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

3 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

charles/claude-hooks!999

No description provided.