Cancellation

The SDK accepts a standard cancellation primitive — an AbortSignal in TS, a context.Context in Go, an internal cancel handle in Python.

const ac = new AbortController();

const stream = client.apps.generateStream(
  { prompt: 'pricing page' },
  { signal: ac.signal },
);

// Cancel after 30s
setTimeout(() => ac.abort(), 30_000);

try {
  for await (const ev of stream) {
    if (shouldStop(ev)) ac.abort();
  }
} catch (e) {
  if (ac.signal.aborted) console.log('cancelled cleanly');
  else throw e;
}

For long-lived runs (or when you don't have the original connection), explicitly cancel by run id:

await client.runs.cancel('run_abc');

Both the streaming client and the explicit endpoint converge on the same effect — the server halts the pipeline at the next safe checkpoint and emits a final { type: "cancelled" } event before closing.

Cancellation timing affects whether quota is refunded:

• Before planner finishes — full refund.
• After planner, before code agents start — partial refund (planner tokens are committed).
• After any code agent has emitted a file — no refund. The artifacts are kept; you can resume from this state.

Files committed before cancel stay in the workspace. The chat remains valid — you can send a follow-up prompt to continue or fix what was started. To discard everything from a cancelled run, use the rollback endpoint or simply send a new full prompt.