fix(agent): do not execute tool calls from a max_output_tokens-truncated response

[w0nche0l]

Problem

When the provider stops a turn at max_output_tokens, the response can still carry a function_call item whose arguments are whatever prefix fit in the budget (production capture: a reasoning model spent 3,000 tokens thinking and was cut off at {"commands":). responseHasToolCalls and extractToolCallsFromResponse only look at output[].type, so the loop parsed the fragment, logged Failed to parse tool call arguments, fed the error back to the model, and made another request against the same exhausted budget. Each truncated turn cost one wasted round trip, and the run only ended when a stop condition fired.

Fix

New outputItemsWithoutTruncatedCalls(response): on a response with status === 'incomplete' and incompleteDetails.reason === 'max_output_tokens', drop every function_call whose item status is not completed (when the provider omits status, the arguments must parse). On any other response it is response.output unchanged. extractToolCallsFromResponse and responseHasToolCalls read through it, and model-result.ts uses it wherever a response's output is written to state or echoed into the next request, so a dropped call is never sent back without an output. Calls that completed before the cut-off execute normally, matching the Vercel AI SDK and OpenAI Agents SDK loops. With no surviving call the loop finalizes on the truncated turn and the caller sees incompleteDetails, the right signal to raise the budget.

Tests

tests/unit/max-output-tokens-truncation.test.ts:

• extractors: fully truncated response yields no calls; mixed response yields only the completed call.
• real callModel against a mocked betaResponsesSend: fully truncated turn finalizes after 1 request with nothing executed (without the fix: 5 requests before stepCountIs(3)); mixed turn executes the completed weather call once, never the shell call, and the follow-up request's input carries function_call/function_call_output for the weather call only.

Context: OpenRouterTeam/openrouter-web#43149 carries this as a patch on @openrouter/agent@0.10.0 until the next release.