[FEATURE]:Add automatic retry with exponential backoff for LLM API timeouts

[wyp2k]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Problem
When an LLM provider's streaming API accepts a connection but never returns data (e.g. server-side stall, queueing), opencode has no automatic retry mechanism. The request hangs until the configured timeout (default 10 minutes) fires, then errors out and stops the session entirely. The user must manually send a message to resume.
This contrasts with Claude Code and Codex, which both retry transient API failures automatically with exponential backoff — users never notice these incidents.
Reproduction

1. Configure a provider (e.g. volcengine GLM) with timeout: 600000, chunkTimeout: 60000
2. In a long session (~50K tokens/step, 20+ agentic steps), trigger a stream request
3. Provider accepts the request but sends no data (or only SSE keepalive pings that reset chunkTimeout without delivering actual content)
4. Session hangs for the full timeout duration, then errors with TimeoutError: The operation timed out.
5. Agent loop stops; no retry attempted
   Root Cause

• No retry / maxRetries option exists in ProviderConfig.options (confirmed via config schema)
• chunkTimeout is ineffective against providers that send SSE keepalive pings (empty : ping frames) — these reset the chunk timer without delivering data
• Timeout errors are treated as fatal for the current message, not as retryable failures
  Proposed Solution
  Add provider-level retry configuration:
  "options": {
  "maxRetries": 3,
  "retryDelay": 1000,
  "retryBackoff": "exponential",
  "retryOn": ["timeout", "5xx", "connection-error"]
  }
  Behavior:
• On timeout / connection error / 5xx, automatically retry the request up to maxRetries times
• Use exponential backoff (e.g. 1s, 2s, 4s)
• Show a brief "retrying…" indicator in the TUI
• If all retries fail, then surface the error to the user
  Additionally, consider:
• Distinguishing SSE keepalive pings from actual data chunks in chunkTimeout logic (don't reset the timer for empty frames)
• Treating timeout errors as retryable by default (like the AI SDK's built-in maxRetries which opencode currently doesn't expose)
  Environment
• opencode v1.17.20
• Provider: volcengine-plan (GLM-latest), @ai-sdk/openai
• OS: Windows

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [Bug]: Session processor retries indefinitely with unbounded exponential backoff — no max retries or circuit breaker #17648: Session processor retries indefinitely with unbounded exponential backoff — no max retries or circuit breaker (overlapping topic: exponential backoff + max retries for provider errors)
• Implement exponential back-off when hitting rate limits #1712: Implement exponential back-off when hitting rate limits (similar feature request for retry backoff)
• Rate limit retries every ~1 second with no backoff or max limit (OpenAI provider) #30510: Rate limit retries every ~1 second with no backoff or max limit (OpenAI provider) (related retry behavior)
• Retry backoff is uncapped when a retryable error has response headers but no retry-after #33728: Retry backoff is uncapped when a retryable error has response headers but no retry-after (related backoff cap issue)

If your request is distinct (specifically about timeout/stall retries rather than rate-limit retries), please clarify the difference in a comment.

[github-actions]

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.