Session service -32603 errors under concurrent batch load

Bug Description

OpenCode session service returns JSON-RPC error -32603 ("Internal error: OpenCode service failure") when processing multiple concurrent sessions or long-running batch operations.

Steps to Reproduce

1. Submit 3+ concurrent session resume requests (e.g., batch process 3+ files in parallel)
2. Observe that 1-3 of the sessions fail with error code -32603
3. Retry the failed session; it succeeds (indicating service recovered, not a permanent error)

Expected Behavior

All valid session requests should either succeed or fail with a specific, actionable error code (e.g., 429 Too Many Requests for rate limits).

Impact

• Batch processing becomes unreliable under load (3+ concurrent sessions)
• Long-running operations fail after 20+ minutes
• Users lose work and have no clear recovery path

Current Workarounds (Tested)

• Reduce concurrency: Process 1–2 sessions at a time (slow but reliable)
• Exponential backoff: Retry with 1m, 2m, 4m delays (reduces failures to ~5%)
• Pre-flight health check: Skip batch if service is unavailable

Workarounds reduce frequency but don't eliminate the root cause.

Additional Context

We've validated this issue in production batch processing across 50+ sessions. The error is reproducible and consistently fails the same way under load.

Testing fork available at https://github.com/blogtheristo/opencode-all if upstream fix needs validation.

Error Example

{
  "code": -32603,
  "message": "Internal error: OpenCode service failure",
  "data": {
    "service": "session"
  }
}