Feature: per-tool execute timeout, a hung blocking tool stalls the turn forever

[HoneyTyagii]

Problem

tool() has no execution timeout. If a blocking tool's execute() never settles (a fetch to a dead endpoint, a third-party SDK that hangs, a forgotten await), the turn stalls indefinitely:

• ToolExecutor.runTool (agents/src/voice/tool_executor.ts) awaits the tool promise with no time bound
• The tool's AbortSignal only fires on interruption / explicit cancel() / drain(), never on elapsed time
• The only related knob, DRAIN_TOOL_TIMEOUT_MS (via LK_DRAIN_PLAYOUT_TIMEOUT_MS), bounds how long drain waits after abort was already signalled; it does nothing for a tool that hangs mid-turn while the user is silent

Requested behavior

An executeTimeout (ms) option on tool():

const getWeather = tool({
  name: 'getWeather',
  description: '...',
  parameters: z.object({ location: z.string() }),
  executeTimeout: 10_000,
  execute: async ({ location }, { abortSignal }) => { ... },
});

On expiry:

1. Abort the tool's existing AbortController (one already exists per running task)
2. Emit a FunctionCallOutput with a timeout ToolError so the model can tell the user the tool timed out (related: fix(tts): restart ChunkedStream retries under fresh request IDs #1994, "Cancelled function tools vanish without a trace")
3. Clean up the RunningTask entry so hasRunningTasks returns to normal

Optionally a default at the Toolset / AgentSession level, with per-tool override.

Why not just AbortSignal.timeout() in every execute?

That's the current workaround and it's:

• Boilerplate repeated in every tool, easy to forget
• Powerless when execute ignores the signal (many SDK calls don't accept one), since nothing races the promise from the outside
• Inconsistent: some tools end up with timeouts, others don't, and a single offender still wedges the whole turn

Environment

• agents-js 1.4.x / current main

Happy to contribute a PR if the approach sounds good.