fix(openai): omit stop parameter for o-series and gpt-5+ models

[rodboev]

Fixes #8184

Summary

GPT-5-nano, gpt-5-mini, o1, o1-mini, o3, o3-mini, and related models return HTTP 400
"Unsupported parameter: 'stop' is not supported with this model." for every chat and autocomplete
request. The autocomplete template always supplies stop tokens, and getMaxStopWords() caps them to 4
(for api.openai.com) but never omits the field entirely. Setting stop: [] in config does not help —
the field is overwritten from the template before the request is sent. The fix is to clear stop to
undefined inside the existing isOSeriesOrGpt5PlusModel branch, which already handles the other
API differences for these models.

Root cause

OpenAI._convertArgs() at core/llm/llms/OpenAI.ts:284 sets finalOptions.stop from the autocomplete
template before the model-capability branch:

finalOptions.stop = options.stop?.slice(0, this.getMaxStopWords());  // line 284

if (this.isOSeriesOrGpt5PlusModel(options.model)) {
  // handles max_tokens and messages — but not stop
}

OpenAI.modifyChatBody() at core/llm/llms/OpenAI.ts:440 does the same for the adapter path.
packages/openai-adapters/src/apis/OpenAI.ts:70 has a parallel block for the official OpenAI API
adapter that also does not clear stop. isOSeriesOrGpt5PlusModel (line 221-223) already matches
all affected models — no new model detection is needed.

Changes

• core/llm/llms/OpenAI.ts: add finalOptions.stop = undefined as the first statement inside the
  isOSeriesOrGpt5PlusModel branch in both _convertArgs() and modifyChatBody().
• packages/openai-adapters/src/apis/OpenAI.ts: strip stop in OpenAIApi.modifyChatBody() via a
  new isOSeriesOrGpt5Plus predicate (/^o[0-9]/ or /gpt-[5-9]/i, so openchat doesn't match)
  that runs for all hosts, not just api.openai.com — Azure deployments route through
  AzureApi.modifyChatBody() → super.modifyChatBody() and previously kept sending stop. The
  predicate matches the core path's gpt-[5-9] definition (gpt-6, gpt-7-turbo, etc.), wider than
  the official-API block's includes("gpt-5") check.

What this doesn't change

• getMaxStopWords() is unchanged — it still serves providers that accept stop tokens.
• OpenAI._convertArgsResponses() is unchanged — the Responses API path does not expose a stop field.
• All non-o-series, non-gpt-5+ models (gpt-4, gpt-4o, gpt-4-turbo, etc.) are unaffected; their stop
  is still populated.
• Azure — Azure.ts (core) sets useOpenAIAdapterFor: [], so it routes through
  OpenAI.modifyChatBody(); the adapter-side AzureApi inherits the new model-based strip from
  OpenAIApi.modifyChatBody(). No change to either Azure class is needed.
• The max_completion_tokens conversion stays gated to the official OpenAI API, exactly as before.
• The workaround requestOptions.extraBodyProperties.stop: null noted in the issue continues to work
  but is no longer necessary for standard deployments.

Checklist

• I've read the contributing guide
• The relevant docs, if any, have been updated or created
• The relevant tests, if any, have been updated or created

Screen recording or screenshot

Before (gpt-5-nano autocomplete request body):

{ "model": "gpt-5-nano", "stop": ["<|fim_suffix|>", "\n\n", ...], "max_completion_tokens": 256 }

After:

{ "model": "gpt-5-nano", "max_completion_tokens": 256 }

N/A — no UI change. Error disappears from the network tab.

Tests

Vitest (core/llm/llms/OpenAI.vitest.ts):

• cd core && npx vitest run llm/llms/OpenAI.vitest.ts — 11 tests pass, 0 failures.
• Coverage: 8 existing tests (streamChat/chat/complete, O1 models, tools, max tokens, embeddings) + 3 new cases:
  • gpt-5-nano streamChat: verifies stop is undefined in request body.
  • o3-mini streamChat: verifies stop is undefined in request body.
  • gpt-4o streamChat (regression): verifies non-o-series models still include stop.

Vitest (packages/openai-adapters):

• cd packages/openai-adapters && npx vitest run src/test/openai-adapter.vitest.ts src/test/main.test.ts — 25 tests pass, 0 failures (20 existing + 5 new).
• New coverage: o3 and gpt-5 on an Azure-style apiBase strip stop; gpt-4o on Azure-style apiBase keeps stop; gpt-6 on a non-official host strips stop without the official-API conversions; openchat keeps stop (predicate guard).

Jest (core/llm/llms/OpenAI.test.ts):

• Tests isOSeriesOrGpt5PlusModel() model detection logic. No new assertions needed for this change.

Prettier:

• All modified files pass formatting checks after prettier --write.

---

Summary by cubic

Omit the stop parameter for o-series and gpt-5+ models to prevent HTTP 400 errors and restore chat/autocomplete. Applies to both core and adapter paths, including Azure-style endpoints.

• Bug Fixes
  • Core core/llm/llms/OpenAI.ts: clear stop in _convertArgs() and modifyChatBody() when isOSeriesOrGpt5PlusModel is true.
  • Adapter packages/openai-adapters/src/apis/OpenAI.ts: strip stop for o-series and gpt-5+ on all hosts; keep official-API-only conversions unchanged.
  • Tests: add coverage for gpt-5-nano and o3-mini (no stop), gpt-4o (keeps stop), Azure-style endpoints, non-official hosts (gpt-6 strips only stop), and guard non-matching models like openchat.

^{Written for commit c854a3e. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 4 files

_{Re-trigger cubic}