fix(providers): bound the OpenAI response body read, restore retries, and fail on reported generation failures

[waleedlatif1]

Summary

• Bound the /v1/responses body read (60s) so a stalled response fails fast instead of hanging ~4.5min on the runtime's own socket wall. Time-to-headers stays unbounded — the endpoint withholds its 200 until generation finishes, so all think time is time-to-headers (measured: 14545ms to headers, 1ms of body).
• Annotate opaque transport failures with the phase they died in (awaiting-response-headers vs reading-response-body) plus status, ttfb, content-length and x-request-id. The annotation rides the error message into the trace span, which persists even when a task stops shipping logs.
• Restore status-based retries (408/409/429/5xx, 2 retries, backoff + Retry-After). This path lost them in improvement(openai): migrate to responses api #3135 when it moved off the OpenAI SDK onto raw fetch; every other provider still retries via its SDK client.
• Fail the block when a 200 reports a failed generation. status: "failed" / populated error previously returned success with empty content and billed tokens, and a truncated function_call was executed. Now matches the streaming path.

Deliberately no retry on a stalled body: /v1/responses ignores Idempotency-Key (verified live — same key and body returns two distinct response ids), so a retry would generate and bill a second response. The OpenAI SDK and the AI SDK both decline to retry this class.

Type of Change

• Bug fix

Testing

Tested manually against the live API plus 40 new unit tests.

• Replayed real captured /v1/responses payloads (completed, incomplete/max_output_tokens, tool-call) through the provider — none rejected by the new status guard.
• Verified live that Idempotency-Key is ignored on this endpoint, and that headers arrive only when generation completes.
• Verified the body deadline propagates into an already-returned body under the real runtime.
• Every new test verified able to fail by breaking the implementation first.
• Full suite green; one unrelated pre-existing failure (executor/handlers/pi/cloud-review-tools.test.ts) reproduces identically on staging.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Aug 5, 2026 4:17am

[cursor]

PR Summary

^{Cursor Bugbot is generating a summary for commit c0e78c8. Configure here.}

[greptile-apps]

Greptile Summary

The PR bounds OpenAI Responses body reads, restores status-based retries with cancellable backoff, rejects reported generation failures, and improves transport-error diagnostics.

• Adds a 60-second deadline for successful and error response-body reads.
• Retries 408, 409, 429, and 5xx responses while honoring server pacing.
• Validates non-streaming response status before returning content or executing tools.
• Preserves transport phase and response metadata through provider error wrapping.
• Adds focused tests for retries, deadlines, failure statuses, and agent-level error translation.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains, and the current implementation addresses both previously reported issues by bounding non-2xx body reads and making retry backoff respond immediately to cancellation.

Important Files Changed

Filename	Overview
apps/sim/providers/openai/core.ts	Adds bounded body reads, status retries, response-status validation, cancellation-aware backoff, and phase-annotated transport errors; both previously reported issues are addressed.
apps/sim/executor/handlers/agent/agent-handler.ts	Walks wrapped causes to translate transport aborts and deadlines into a diagnostic provider-timeout message.
apps/sim/providers/openai/core.retry.test.ts	Covers retryable statuses, retry limits, server-directed pacing, caller cancellation, and non-retryable body stalls.
apps/sim/providers/openai/core.transport-phase.test.ts	Covers body deadlines, transport-phase metadata, caller abort behavior, and unbounded generation time-to-headers.
apps/sim/providers/openai/core.response-status.test.ts	Covers failed and incomplete generation handling, truncated tool calls, healthy responses, and continuation turns.
apps/sim/executor/handlers/agent/agent-handler.test.ts	Verifies wrapped TimeoutError and AbortError classification while preserving provider diagnostics.
apps/sim/providers/types.ts	Updates provider response typing to support the changed OpenAI response handling.

Sequence Diagram

sequenceDiagram
  participant Agent as Agent Handler
  participant Provider as OpenAI Provider
  participant API as OpenAI Responses API
  Agent->>Provider: Execute request
  Provider->>API: POST /v1/responses
  alt Retryable HTTP status
    API-->>Provider: 408/409/429/5xx
    Provider->>Provider: Cancellable backoff
    Provider->>API: Retry (up to 2 times)
  else Successful headers
    API-->>Provider: 200 response
    Provider->>Provider: Arm body deadline
    Provider->>Provider: Parse and validate status
    Provider-->>Agent: Content/tool result
  else Body stalls
    Provider->>Provider: Abort body read after 60s
    Provider-->>Agent: Annotated transport error
  end

Loading

_{Reviews (2): Last reviewed commit: "fix(providers): bound stalled error bodi..." | Re-trigger Greptile}

[waleedlatif1]

@greptile

[waleedlatif1]

@cursor review

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit f403ceb. Configure here.}

[waleedlatif1]

Superseded by #6283.

The root cause turned out to be different from what this PR was built on. Confirmed from OpenAI's stored responses for the actual incident: the model repeated one tool call until it consumed the full 128,000-token output budget — status: incomplete, reason: max_output_tokens, 0 reasoning tokens, 287 tokens of stored output against 128,000 billed. /v1/responses withholds its 200 until generation completes, so the client was stuck in the headers phase, never the body phase.

The response-body deadline here was built for a body-stall theory the evidence disproved, and the comments and commit messages document that wrong cause. #6283 keeps what holds up — phase annotation, x-request-id, cause propagation, the failed-generation guard, bounded error bodies — off a clean base with the corrected narrative.

The status-based retries from this branch are still worth having and will follow separately; they're unrelated to this incident.