fix(responses): preserve explicit reasoning none

[andyst-dev]

Fixes #75227.

Summary

• Preserve an explicit reasoning_effort: none through the OpenAI Responses transport for GPT-5.6 models by sending reasoning.effort: "none" instead of omitting reasoning.
• Fail fast when none is requested on a Responses route/model where Hermes does not know that value is supported, instead of silently falling back to the model default.
• Add focused transport coverage for the GPT-5.6 wire payload and the unsupported-model error path.

Validation

• scripts/run_tests.sh tests/agent/test_codex_responses_adapter.py -q
• scripts/run_tests.sh tests/tui_gateway/test_reasoning_session_scope.py -q
• python -m py_compile agent/transports/codex.py tests/agent/test_codex_responses_adapter.py
• git diff --check

[teknium1]

Thanks for isolating a real wire-level regression. Current main drops the explicit disablement: agent/transports/codex.py:219-220 sets reasoning_enabled = False, and the disabled branch at :365-366 emits only include=[].

Problems

• agent/transports/codex.py:253-260 treats every non-xAI/non-GitHub gpt-5.6* route as known to support reasoning.effort: "none", but _openai_responses_supports_reasoning_none() at :56-63 checks only the model string. The same transport receives arbitrary base_url values from agent/chat_completion_helpers.py:1218-1235; current issuer classification labels those routes other:<base_url> in agent/codex_responses_adapter.py:49-51. This does not satisfy the proposed fail-fast behavior for unknown Responses relays.

Suggested changes

• Include endpoint/provider capability in the allowlist, rejecting unknown other:<base_url> routes unless support is explicitly known.
• Add a gpt-5.6 unknown-base-url rejection test in addition to the direct supported-route test.

This is an automated hermes-sweeper review.

[teknium1]

This guard only excludes xAI and GitHub; the support helper checks the model name but not base_url or provider. ResponsesApiTransport also serves arbitrary other:<base_url> routes, so a GPT-5.6-named custom relay is treated as known-supported rather than failing fast. Please make endpoint capability part of this predicate and cover an unknown-base-url rejection.

[andyst-dev]

Addressed the review feedback in b4ac12f93.

Changes:

• Explicit reasoning_effort: none is now allowed only for GPT-5.6 models on known OpenAI Responses routes: the official api.openai.com endpoint or the classified Codex backend.
• Unknown Responses relays now fail closed even when the model string is gpt-5.6*.
• Added positive coverage for both official OpenAI and Codex-backend routes, plus the requested unknown-base-URL rejection case.

Validation:

• scripts/run_tests.sh tests/agent/test_codex_responses_adapter.py -q — 12 passed
• scripts/run_tests.sh tests/tui_gateway/test_reasoning_session_scope.py -q — 8 passed
• python -m py_compile agent/transports/codex.py tests/agent/test_codex_responses_adapter.py
• git diff --check