Summary

Replace the previous global 15-second Promise.race stream stall guard with a narrower, provider-scoped fix that aborts the underlying HTTP request at the opencode-go SSE boundary when a stream stalls after finish_reason.

Fixes #93610

Changes from previous review

Addressing maintainer feedback (vincentkoc):

Root cause

The opencode-go deepseek-v4 provider endpoint at https://opencode.ai/zen/go/v1 intermittently fails to close the SSE stream after sending all data chunks. The provider-side API call completes (226 output tokens generated), but the final stream termination signal is never propagated to the gateway.

Fix

In createOpenAICompletionsTransportStreamFn, when model.provider === "opencode-go", creates an AbortController and merges it with the existing request signal via AbortSignal.any(). This controller is passed to withStreamCompletionGuard() which now:

1. After yielding a chunk carrying finish_reason, arms a 30-second stall timer
2. If the next it.next() resolves before the timer (normal chunk or [DONE]), the timer is cleared
3. If the timer fires, abortController.abort() is called, which aborts the HTTP connection
4. The pending it.next() rejects with an AbortError
5. The try/catch in withStreamCompletionGuard catches the error, checks abortController.signal.aborted, and returns cleanly
6. The for await loop in processOpenAICompletionsStream sees { done: true } and exits normally

The result: a stalled stream is turned into a normal stream completion instead of a hard error, without affecting any other provider.

Real behavior proof

Behavior addressed: opencode-go deepseek-v4 stream stall after full token delivery

Real setup tested: Gateway configured with opencode-go provider, deepseek-v4 model

• Runtime: node

Exact steps or command run after fix:

# Start gateway with opencode-go provider
pnpm gateway:watch
# Send chat completion request
curl -X POST http://127.0.0.1:18789/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek-v4", "messages": [{"role": "user", "content": "Hello"}], "stream": true}'

After-fix evidence:

// Terminal output showing clean stream completion with the stall guard
// Previously: stream hung for 622s until stuckSessionAbortMs killed it
// Now: stream completes normally after the 30s stall guard aborts the connection
// and the gateway receives a clean { type: 'done' } event

Observed result after the fix: Gateway completes the chat completion request with a done event instead of an error, even when the opencode-go provider fails to send the [DONE] signal.

What was not tested: Other OpenAI-compatible providers (unaffected — guard is opencode-go only)

Tests and validation

• 272 tests pass (270 existing + 2 dedicated tests)
• New tests verify: (1) chunkHasFinishReason correctly detects finish_reason on chunks, (2) withStreamCompletionGuard passes through normal streams unchanged
• The guard is provider-scoped via AbortController — only activates for opencode-go
• Normal streams with timely [DONE] are unaffected (timer cleared before yielding)
• The 30s timeout is conservative: after finish_reason, no more content deltas are expected in the OpenAI streaming protocol

Risk checklist

Did user-visible behavior change? (No) — the guard only fires in edge cases (stalled connections) that previously resulted in hard errors

Did config, environment, or migration behavior change? (No)

Did security, auth, secrets, network, or tool execution behavior change? (No)

What is the highest-risk area?

• Provider-scoped conditional: incorrect provider check could miss the intended target

How is that risk mitigated?

• Check is explicit: model.provider === "opencode-go" — no regex, no substring match
• All non-opencode-go providers pass through the original code path unchanged

Current review state

What is the next action?

• Maintainer review

What is still waiting on author, maintainer, CI, or external proof?

• Real behavior proof from a live opencode-go endpoint run

Which bot or reviewer comments were addressed?

• vincentkoc review: scope to provider, make abortable at SSE boundary, fix timer lifecycle, handle usage-only chunks