Codex ignores an explicit known-good tool entrypoint and invents incompatible API calls

What issue are you seeing?

A repository's durable instructions specified one exact local executable wrapper as the only supported entrypoint for calling a local OpenAI-compatible model gateway. The wrapper was already known to work because another installed coding agent used the same provider configuration successfully.

Codex repeatedly bypassed that explicit entrypoint and improvised alternate request code, parameters, and result parsing.

Observed failure modes included:

• sending an unsupported explicit temperature to one model, causing HTTP 400;
• reading only message.content while a provider returned the useful answer in reasoning_content, causing a false "empty response";
• sending an oversized review packet that timed out after 300 seconds;
• treating an HTTP/process-level success as a completed review even when no usable answer or artifact existed;
• continuing to reason about a special Codex-specific protocol after the user explicitly instructed it to copy the known-good provider path.

After the shared wrapper was corrected, direct calls through that wrapper succeeded for all three target models, but another Codex task still failed to reliably choose and execute the canonical entrypoint.

Steps to reproduce

1. Provide a known-good local wrapper command for an OpenAI-compatible endpoint.
2. Put an explicit instruction in AGENTS.md that Codex must use only that command and must not create a new dispatcher or client.
3. Ask Codex to send a small review task through the wrapper.
4. In affected runs, Codex reasons about or constructs a different invocation path instead of executing the supplied command.

Expected behavior

When the user supplies a precise, executable, known-good tool entrypoint, Codex should call it as written. It should not invent provider-specific parameters or a parallel protocol unless the supplied command actually fails and the user authorizes debugging.

Tool success should also require usable response content/artifacts, not merely HTTP 200 or process exit 0.

Environment:

• OpenAI Codex Desktop 26.727.6591.0 (x64 MSIX)
• Windows 10.0.26200.8875
• Model observed: GPT-5.6 Sol

No private workspace files or transcripts are attached.