feat(coding-agent): add tokens/sec to the working loader

[samsja]

Summary

Adds tokens/sec to the working loader message so the user can see output throughput while the model is streaming.

Changes

• agent-activity.ts: Added tokensPerSecond to AgentActivityStatus, tracked via streamingStartedAt and streamingStartTokens fields. The rate is computed as delta_tokens / elapsed_seconds and only reported after 500ms of streaming to avoid jitter.
• interactive-mode.ts: Updated getWorkingLoaderMessage() to append (N tok/s) after the token count when a rate is available.
• agent-activity.test.ts: Updated toEqual assertions for the new tokensPerSecond field.

Example

Thinking · 5s · ↓ 1.2k tokens (240 tok/s)

Testing

• npx vitest run test/agent-activity.test.ts — 13/13 pass
• npx vitest run test/interactive-mode-status.test.ts — 151/151 pass
• npm run check — clean

---

Note

Low Risk
UI-only activity tracking and loader text; no auth, persistence, or protocol changes.

Overview
The interactive working loader now shows output throughput while the model streams, e.g. Thinking · 5s · ↓ 1.2k tokens (240 tok/s).

AgentActivityTracker gains tokensPerSecond on AgentActivityStatus, measured from the first output of the current turn using the same monotonic token count as the existing loader. The rate stays 0 until at least 500ms of streaming so early numbers do not flicker. getWorkingLoaderMessage() appends (N tok/s) only when a positive rate is available; tests and the changelog are updated accordingly.

^{Reviewed by Cursor Bugbot for commit dfb501a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add tokens/sec display to the coding agent working loader

• Adds a tokensPerSecond field to AgentActivityStatus, computed from streaming start time and token count delta in agent-activity.ts.
• Rate is only reported after ≥500ms of streaming and a positive token delta; returns 0 otherwise.
• The working loader message in interactive-mode.ts appends the rate in parentheses, e.g. ↓ 1.2k tokens (45 tok/s), when the rate is positive.

^{Macroscope summarized dfb501a.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit dfb501a. Configure here.}

[cursor]

Stale tokens/sec across stream gaps

Medium Severity

streamingStartedAt is set on the first output tokens but never cleared on message_end, so computeTokensPerSecond keeps dividing by wall time through tool runs and waits. The loader then shows a decaying tok/s while idle or executing, and later streams inherit that inflated elapsed time instead of measuring active output throughput.

Additional Locations (1)

• packages/coding-agent/src/modes/interactive/agent-activity.ts#L130-L138

 

^{Reviewed by Cursor Bugbot for commit dfb501a. Configure here.}

[alexzhang13]

We probably should only consider TokensPerSecond when generations are actually happening. It might be slightly misleading otw