Feature request: configurable API response timeout for subagent calls

[oneryalcin]

Summary

When a subagent makes an API call with large context (~83K tokens) and needs to generate large output (~30KB), the API call can hang indefinitely with zero events written to the JSONL transcript. There is no configurable timeout to detect or recover from this state.

Reproduction

1. Launch a background agent (run_in_background=True) with a task requiring large context
2. Agent accumulates ~80K+ tokens of context via multiple Read calls
3. Agent makes a tool call that succeeds (e.g., mkdir)
4. Next API call (to generate a large Write) hangs — zero JSONL events for 4+ minutes
5. Parent agent polls via TaskOutput — always sees status: running
6. No mechanism to detect the hang or kill the subagent

Observed behavior

• JSONL transcript stops at the successful tool_result — no new entries
• TaskOutput always returns running (can't distinguish "actively generating" from "API hung")
• KillShell fails ("not a local bash task")
• TaskStop kills the agent but destroys any in-progress work

Expected behavior

A configurable timeout (e.g., api_timeout_seconds in ClaudeAgentOptions) that:

1. Detects when an API response hasn't produced any events for N seconds
2. Marks the agent as error with a clear timeout message
3. Writes the error to the agent's JSONL for downstream tracking

Workaround

We implemented a 20-minute wall-clock timeout in our stop hook that unblocks the parent task, but the hung agent process itself stays alive consuming resources.

Environment

• claude-agent-sdk: 0.1.48
• Claude Code CLI: 2.1.72 (bundled)
• Observed on: Python 3.12, Linux (GKE)
• Original incident: CLI 2.0.72, but the underlying issue (no API timeout) persists