ACP prompt() returns stopReason: end_turn + zero usage on user cancel; should be cancelled with no usage #25899

Summary

Agent.prompt() in packages/opencode/src/acp/agent.ts returns stopReason: "end_turn" whether the turn finished naturally or was interrupted by session/cancel. The two outcomes are indistinguishable on the wire. ACP defines stopReason: "cancelled" for exactly this case; reporting end_turn for a user-cancel makes a stopped turn look like a clean completion.

The same paths can also report usage: { totalTokens: 0, inputTokens: 0, outputTokens: 0 } for cancelled turns. The LLM stream is killed before the AI SDK emits its finish-step usage event, so assistantMessage.tokens is still at its zero initializer when prompt() returns. The provider may already have consumed the prompt tokens, but the exact count is not available. Reporting 0 is misleading; omitted usage is more accurate.

Reproducer

1. Send a session/prompt that triggers a long-running tool, such as a 60s shell loop.
2. Once the tool starts, send session/cancel.
3. Observe the prompt RPC reply.

Expected:

{ "id": 100, "result": { "stopReason": "cancelled", "_meta": {} } }

Actual:

{
  "id": 100,
  "result": {
    "stopReason": "end_turn",
    "usage": { "totalTokens": 0, "inputTokens": 0, "outputTokens": 0 },
    "_meta": {}
  }
}

Trimmed event sequence:

T1> agent_thought_chunk x 4
T1> tool_call title="bash" status="pending" kind="execute"
-- send session/cancel --
CX> tool_call_update title="bash" status="failed" kind="execute"
CX> usage_update
CX> RESULT id=100 stop=end_turn totalTokens=0 inputTokens=0 outputTokens=0

Root cause

When the runner is interrupted by session/cancel, the Effect.onInterrupt path in packages/opencode/src/session/processor.ts runs:

yield* halt(new DOMException("Aborted", "AbortError"))

halt calls MessageV2.fromError, which maps DOMException("AbortError") to a MessageAbortedError wire-level name tag and assigns it to ctx.assistantMessage.error. The runner interrupt callback then returns the in-flight assistant message, so sdk.session.prompt(...) resolves with that message instead of rejecting.

finish-step is what populates ctx.assistantMessage.tokens from the AI SDK usage event. It only fires when the streaming response completes, which does not happen on a mid-stream interrupt. The tokens therefore remain at the zero initializer set when the message was created.

Then acp/agent.ts unconditionally returns:

return {
  stopReason: "end_turn" as const,
  usage: msg ? buildUsage(msg) : undefined,
  _meta: {},
}

Neither the MessageAbortedError tag nor the all-zero token state is checked.

Impact

• ACP clients that special-case stopReason: "cancelled" cannot distinguish a user stop from a clean completion.
• Cost / usage telemetry can undercount cancelled turns by treating unknown usage as zero.
• Downstream heuristics that use token counts can incorrectly treat cancelled turns as zero work.

Proposed fix

In agent.ts, detect msg?.error?.name === "MessageAbortedError" and:

1. Return stopReason: "cancelled" instead of "end_turn".
2. Omit usage rather than reporting all-zero usage, so clients treat it as unknown.

A separate, larger fix could recover provider usage during interrupt cleanup, but the stop-reason fix is correct independently.