fix: add timeout and abort retry for LLM API requests

[ycls2002]

Summary

• LLM API calls via fetchWithProxy had no timeout set, causing requests to hang indefinitely when the remote server drops the connection without responding
• This manifested as The socket connection was closed unexpectedly errors that crashed the entire pipeline with no recovery

Root Cause

In packages/core/src/llm/provider.ts, both /responses and /chat/completions endpoint calls to fetchWithProxy lacked a signal parameter, meaning Node.js fetch would wait forever if the server never responded or dropped the TCP connection mid-stream.

Additionally, AbortSignal.timeout() errors ("The operation was aborted") were not in the transient error detection list, so even if a timeout were added, it would not trigger the existing retry mechanism.

Changes

1. Add AbortSignal.timeout(180_000) to LLM fetch calls

Both fetchWithProxy calls in chatCompletionViaCustomOpenAICompatible now include a 3-minute timeout:

// /responses endpoint (line ~792)
const response = await fetchWithProxy(url, {
  method: "POST",
  headers,
  body: JSON.stringify(payload),
  signal: AbortSignal.timeout(180_000),  // NEW
}, client.proxyUrl);

// /chat/completions endpoint (line ~880)
const response = await fetchWithProxy(url, {
  method: "POST",
  headers,
  body: JSON.stringify(payload),
  signal: AbortSignal.timeout(180_000),  // NEW
}, client.proxyUrl);

2. Add abort errors to transient error detection list

// isTransientLLMTransportError (line ~442)
"aborted",
"The operation was aborted",

This ensures timeout-triggered aborts are caught by withTransientLLMRetry and retried up to 2 times (consistent with existing retry behavior for ECONNRESET, socket hang up, etc.).

Behavior After Fix

Scenario	Before	After
Server drops socket	Crash immediately	Retry up to 2 times
Server hangs (no response)	Hang forever	Timeout at 3 min, retry
Transient network glitch	May crash	Auto-retry

Notes

• AbortSignal.timeout() requires Node.js 17.3+ (InkOS already requires Node.js 18+)
• The 180s timeout is generous for LLM generation; adjust if needed for slower models
• No backoff delay between retries (consistent with existing behavior)

---

Related: encountered this issue when using mimo-v2.5-pro via custom endpoint (token-plan-cn.xiaomimimo.com), where the server periodically drops connections during long-running stream completions.

Co-Authored-By: Claude Opus 4.7 noreply@anthropic.com