fix(sdk): give an LLM call 300 s and retry a timeout only once

[matej21]

Problem

Provider requests are not streamed, so the request timeout bounds the whole response. With the 120 s default and ~80 output tokens/s, one call can produce at most ~9.5k output tokens.

In production, an agent that wrote several large files in one turn hit this cap. Each retry sent the identical request and timed out the same way: 5 × 120 s, then the agent errored. A replacement agent failed the same way.

Change

• DEFAULT_PROVIDER_REQUEST_TIMEOUT_MS = 300_000 in provider-request.ts, used by the OpenRouter and Anthropic providers. An explicit timeout in config still overrides it.
• withRetry accepts an optional getMaxAttempts(error), which can only lower the attempt cap. withLLMRetry uses it to cap timeout errors at 2 attempts in total. rate_limit, server_error and network_error keep 5.

Worst case for a stuck call stays ~10 min (2 × 300 s instead of 5 × 120 s), but each attempt now has 2.5× longer to finish.

Tests

• retry.test.ts: a timeout gives up after 2 attempts; server_error keeps the full 5.
• The existing maxAttempts test now uses server_error, because a timeout is capped at 2.
• bun test src/core/agents src/core/llm, tsc --noEmit and biome lint pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MmH33W8tV2jhB25w5GQ4E4