fix: retry on timeout errors instead of failing

[morgaesis]

Fixes #13138

Problem

When provider requests timeout (default 5 min), AbortSignal.timeout() throws a DOMException with name: 'TimeoutError'. This error was not handled in fromError(), causing it to fall through to Unknown error type which is not retryable. Sessions stopped immediately instead of retrying.

Solution

Handle TimeoutError DOMException in fromError() and convert it to APIError with isRetryable: true. This enables automatic retry with exponential backoff via the existing retry mechanism.

Testing

Verified the fix handles timeout errors correctly. The session will now retry on provider timeouts instead of failing immediately.

---

Related issues/PRs (not duplicates):

• Bug: AbortSignal.timeout is not a function #7929 - AbortSignal.timeout is not a function (missing API on Windows/web)
• OPENCODE_EXPERIMENTAL_BASH_DEFAULT_TIMEOUT_MS is being ignored #12762 - Bash tool timeout config being ignored
• feat: add 'never' option for mcp_timeout to disable timeout #8633 / feat(mcp): allow 'never' option for mcp_timeout to disable timeout #8630 - MCP timeout configuration
• feat: add model fallback support with TTFT-based timeout #13189 - Model fallback with TTFT-based timeout

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found.

[rekram1-node]

Hm it shoudln't timeout on it's own, are you passing a custom fetch somewhere?

[morgaesis]

Huh, yeah, I had forgotten I had set the timeout option for the provider...

Why does it matter that I set a custom timeout? Shouldn't opencode be resilient to whatever the timeout is?
IIRC, I set the timeout because of new and slow models like GLM-5, where the wait time is unpredictable and can actually be very slow. In some cases, indistinguishable from full-on hanging for hours.

{
    "openrouter": {
      "options": {
        "timeout": 300000, // 5 minutes
      },
      "models": {
        "openrouter/auto": {
          "reasoning": true,
          "limit": {
            "output": 4000,
            "context": 128000,
          },
          "options": {
            "reasoningEffort": "high",
          },
        }
    }
}

[morgaesis]

@rekram1-node, as a concrete example of when this becomes an issue. Today I was closing my laptop after having been working for a while, but when I reopen, network connectivity has to resume. However, network hiccup causes the model to hang practically forever. Without this timeout, opencode and the model keeps hanging, waiting for some network call to return, but it never will.

This way, at least, I get a failure instead of hopes and prayers the job ever continues.