fix(providers): name the failing phase of a stalled OpenAI call, and reject a failed generation

[waleedlatif1]

Summary

• An agent block hung ~4.5 min with an empty trace, surfacing only the runtime's TimeoutError: The operation timed out. Root cause: the model repeated one tool call until it consumed the full 128,000-token output budget. /v1/responses withholds its 200 until generation finishes, so the client waited — bounded only by an undocumented runtime socket deadline — and gave up before the response existed. Confirmed against OpenAI's stored response (status: incomplete, reason: max_output_tokens).
• Name the phase a transport failure died in — awaiting-response-headers vs reading-response-body — with status, ttfb, content-length and x-request-id. Precedent: undici splits these into UND_ERR_HEADERS_TIMEOUT / UND_ERR_BODY_TIMEOUT; the OpenAI SDK captures x-request-id. It rides the error message because that reaches the trace span, which survives when a task stops shipping logs.
• Carry the cause through ProviderError so a transport timeout still classifies after wrapping overwrites name.
• Reject a 200 reporting a failed or unusable generation instead of returning empty content with billed tokens, and stop truncated tool calls executing. Matches streamResponsesTurn (already did this) and @ai-sdk/openai (throws on the same condition).
• Bound non-JSON error bodies so a gateway error page can't become the block error.

Deliberately excluded: a response-body deadline (the observed failure is in the headers phase; body transfer measured at ~1 ms) and status-based retries (worth doing, unrelated, separable).

Type of Change

• Bug fix

Testing

Tested manually against the live API plus 14 new unit tests.

• Root cause confirmed from OpenAI's stored responses for the actual incident — all three turns, the third at exactly 128,000 output tokens.
• Verified live that /v1/responses withholds its 200 until generation completes (14,545 ms to headers, 1 ms of body), which is why only the headers phase matters here.
• Replayed real captured /v1/responses payloads (completed, incomplete/max_output_tokens, tool-call) through the provider — none rejected by the new guard.
• Every new test verified able to fail by breaking the implementation first.
• Full suite green; boundary audit, typecheck and lint clean.

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 6:05am

[cursor]

PR Summary

^{Cursor Bugbot is generating a summary for commit 76c8e91. Configure here.}

[greptile-apps]

Greptile Summary

The PR improves OpenAI failure diagnostics and rejects unsuccessful or unusable non-streaming generations.

• Annotates transport timeouts and aborts with the request phase, timing, and available response metadata.
• Preserves transport causes through ProviderError so agent handling can classify wrapped timeouts.
• Rejects failed, non-tolerable incomplete, and truncated tool-call responses while retaining usable truncated prose.
• Bounds unstructured HTTP error bodies and adds focused regression coverage.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains within the scope of this follow-up review.

Important Files Changed

Filename	Overview
apps/sim/providers/openai/core.ts	Adds transport-phase diagnostics, bounded error parsing, generation-status validation, and cause-preserving provider errors.
apps/sim/providers/types.ts	Extends ProviderError to accept standard ErrorOptions and retain wrapped causes.
apps/sim/executor/handlers/agent/agent-handler.ts	Classifies timeout and abort failures through the cause chain while preserving provider diagnostics.
apps/sim/providers/openai/core.response-status.test.ts	Covers failed and incomplete response handling, truncated tool calls, and continuation turns.
apps/sim/providers/openai/core.transport-phase.test.ts	Covers header/body phase annotation, request metadata, bounded error bodies, and fallback behavior.
apps/sim/executor/handlers/agent/agent-handler.test.ts	Verifies wrapped timeout and abort classification at the agent-handler boundary.

Sequence Diagram

sequenceDiagram
  participant Agent as Agent Handler
  participant Provider as OpenAI Provider
  participant API as OpenAI Responses API
  Agent->>Provider: Execute model request
  Provider->>API: Fetch response
  alt Transport failure
    API--xProvider: Timeout or abort
    Provider->>Provider: Attach phase and response metadata
    Provider--xAgent: ProviderError with preserved cause
    Agent->>Agent: Classify cause-chain timeout
  else HTTP 200 response
    API-->>Provider: Response payload
    Provider->>Provider: Validate generation status
    alt Failed or unusable generation
      Provider--xAgent: Reject generation
    else Completed or usable truncated prose
      Provider-->>Agent: Return provider response
    end
  end

Loading

_{Reviews (4): Last reviewed commit: "fix(providers): name the body phase when..." | Re-trigger Greptile}

[waleedlatif1]

@greptile

[waleedlatif1]

@cursor review

[waleedlatif1]

@greptile

[waleedlatif1]

@cursor review

[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 678384b. Configure here.}