[Bug] Responses combo treats completed text as success despite tool_choice=required (NVIDIA target)

Start with src/server/responses/core-combo.ts and src/server/responses/combo-stream-preflight.ts, then trace how the Responses request's required tool choice reaches the upstream adapter and how completed child responses are validated. Add isolated tests for required calls, zero-call completions, safe retries before output, and committed streaming output, including the 151-tool and continuation cases. Done means a completed response with no required function call is never accepted as a successful tool round, while streaming is not replayed after commitment.

Client or integration

Direct HTTP/API client. Also affects Codex CLI when its RHODIZ combo selects the NVIDIA target.

Area

Proxy and routing; Streaming; Tools / MCP / web search.

Summary

A Responses request with a declared function and tool_choice: "required" can return HTTP 200, status: "completed" but zero function_call items when the selected target is NVIDIA Nemotron Ultra. The combo accepts that response as a successful target outcome, instead of identifying that the required-tool contract was unsatisfied. Consequently a Codex tool round-trip may make no progress, despite the client seeing a successful completion.

This differs from the existing issues about forwarding tool-choice settings on the Chat Completions path (e.g. #5211) and DeepSeek rejecting forced choices (#4631). This is a successful-but-semantically-invalid Responses result that a combo marks terminal-success.

Reproduction

Use a separately authorized OpenCodex instance with a configured NVIDIA provider; credential details are not needed in the report.

{
  "model": "nvidia/nvidia/nemotron-3-ultra-550b-a55b",
  "input": [
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "Call get_fixture_marker with key=status. Do not answer with text before the tool result."
        }
      ]
    }
  ],
  "tools": [
    {
      "type": "function",
      "name": "get_fixture_marker",
      "description": "Return a fixed test value",
      "parameters": {
        "type": "object",
        "properties": { "key": { "type": "string" } },
        "required": ["key"],
        "additionalProperties": false
      }
    }
  ],
  "tool_choice": "required",
  "store": false,
  "stream": false,
  "max_output_tokens": 750
}

Observed against the installed OpenCodex 2.89.0:

Route	HTTP	Response status	Output item types	Required function?
NVIDIA direct	200	completed	reasoning, message	No (0 calls)
GPT-6.1 Sol direct	200	completed	function_call	Yes (1 call)
Claude Opus 5.5 direct	200	completed	function_call	Yes (1 call)

With the exact same one-function request sent to the 11-target combo alias RHODIZ, the combo selected NVIDIA and returned completed without a function call. This reproduced 2/2 times. A separate 151-declared-function test reproduced the same outcome via the combo, whereas the GPT-6.1 Sol route completed the required function call and the subsequent function_call_output continuation 2/2 times.

An earlier 25-turn text-response stress test with 127/128/129/151 declared (but unused) tools had 25/25 HTTP 200 success. That test did not establish actual tool-call correctness: a real Codex CLI read-tool test subsequently showed intermittent failures on the combo (1/5 passes in one batch) versus GPT-6.1 Sol direct (5/5). This comparison is corroborating context, not proof of which exact model served every failed Codex CLI turn.

Version

@bitkyc08/opencodex 2.89.0, Codex CLI 0.163.0.

Operating system

Linux, x86_64 (VM; distribution not independently verified).

Provider and model

NVIDIA / nvidia/nemotron-3-ultra-550b-a55b; controls: openai/gpt-6.1-sol and anthropic/claude-opus-5-5.

Logs or error output

In src/server/responses/core-combo.ts, the combo returns a child immediately when response.ok after its stream preflight and success bookkeeping. The existing preflight in src/server/responses/combo-stream-preflight.ts detects certain retryable zero-output terminals, but does not enforce the caller's required-function selection contract for a completed 200 response. The upstream adapter may be disregarding required; the proxy currently does not detect that and makes the misbehavior a successful combo result.

See the same-issue streaming follow-up comment for per-event SSE counts. No raw user prompts or access logs are included.

Screenshots and supporting files

A minimal synthetic JSON reproducer and comparative results are embedded above. The follow-up comment documents streaming response.completed with zero calls and the isolated test counts. No sensitive attachments.

Redacted configuration

{
  "combos": {
    "rhodiz-coding-primary": {
      "alias": "RHODIZ",
      "strategy": "failover",
      "targets": [
        {"provider": "nvidia", "model": "nvidia/nemotron-3-ultra-550b-a55b"},
        {"provider": "openai", "model": "gpt-6.1-sol"},
        {"provider": "anthropic", "model": "claude-opus-5-5"}
      ]
    }
  }
}

The snippet is illustrative: the actual 11-target combo contains additional routes, whose details do not affect the minimal required-function reproduction.

Expected behavior / proposed acceptance

1. Preserve explicit tool_choice: "required" or a named mandatory tool at the upstream adapter boundary, and verify the upstream wire really supports it.
2. For a fully buffered, successful non-streaming Responses child, a completed response with zero required function_call items must not be reported as a valid completed tool round. Either safely select an eligible alternative before any output/effects, or return a typed compatibility error rather than a false success.
3. Streaming handling must remain conservative: never replay or hop after client-visible output, a tool side effect, or ambiguous commitment. Do not turn every text-only 200 into a retry. Handle only requests explicitly requiring a tool and only if it is provably safe.
4. Preserve complete tool catalogs (including 151 tools), tool call IDs, function_call_output continuation, caller model identity, authorization boundaries, and cancellation.
5. Add isolated positive/negative tests for direct and combo paths, no output vs. committed output, one required tool and 151 declared tools. A passing end-to-end sequence is: request with required → exactly one declared function call → caller-provided function result → completed answer using that result.

Privacy and provenance

Reported by the affected operator via their connected GitHub account. Examples use synthetic markers only; no account IDs, internal addresses, tokens, raw user prompts, or private logs are attached. The production installation has not been modified as part of this upstream report.

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.