fix(agent): only auto-upgrade api.openai.com requests to codex_responses for Responses-API model families

[Kewe63]

Summary

The chat-completions → codex-responses auto-upgrade for direct OpenAI API calls (api.openai.com) was incorrectly triggered by _is_direct_openai_url(base_url) — a URL-only check. This caused gpt-4o-mini and gpt-4-turbo (non-Responses-API models on direct OpenAI) to be routed to the codex endpoint, which doesn't recognize them, producing 400: model not found errors (#52023).

The correct condition is the model family itself: only gpt-5.x and newer Responses-API models get the auto-upgrade. Non-Responses-API models like gpt-4o-mini, gpt-4-turbo, o1, and o3 must stay on the standard chat-completions endpoint regardless of the base URL.

This PR removes the URL-only OR-branch and replaces it with the model-family check that was already executing for the Azure and Nous paths.

---

Changes

File: agent/agent_init.py (+20/-11)

• Dropped the _is_direct_openai_url(self.base_url) OR agent._provider_model_requires_responses_api(...) condition.
• Simplified to a single _provider_model_requires_responses_api(...) check.
• Added comment block documenting the regression and the model-family-only invariant.

File: tests/run_agent/test_run_agent.py (+49/-10)

• The three pre-existing parametric tests in TestGpt5ApiModeRouting re-aligned to the simplified decision.
• Two new regression tests pin the failure mode reported in [Bug]: GPT-4o-mini and GPT-4.1 (and more) fail with "Encrypted content is not supported with this model" after clean install #52023:
  • test_gpt4o_mini_on_openai_direct_stays_on_chat_completions — gpt-4o-mini must not auto-upgrade to codex on api.openai.com
  • test_gpt4_1_on_openai_direct_stays_on_chat_completions — gpt-4-turbo must not auto-upgrade

---

How to Test

# Run the specific routing tests
.venv/bin/python -m pytest tests/run_agent/test_run_agent.py::TestGpt5ApiModeRouting -v
# ✅ 5 existing tests (re-aligned) + 2 new regression tests

# Or via the test suite wrapper
scripts/run_tests.sh tests/run_agent/test_run_agent.py -k "TestGpt5ApiModeRouting"

---

Checklist

• Diff scoped — 2 files, +59/-21
• URL-only OR removed — comment documents why
• Tests added — 2 new regression tests + 3 existing tests re-aligned
• No force-push, no destructive commands
• Follows Conventional Commits
• Cross-platform impact — pure routing condition in Python, no POSIX-only primitives

---

Risk & Impact

Low. The change is subtractive: it removes an OR-half the model-name check was already evaluating true for in every relevant case. The three pre-existing tests cover the routes that behavior has to preserve (Azure, gpt-5.x on api.openai.com, Nous gpt-5.x); the two new tests directly fail on the bug if it ever regresses.

Type: 🐛 Bug fix
Closes: #52023

[alt-glitch]

This was generated by AI during triage.

Competing fix cluster for the codex_responses URL-only auto-upgrade misroute (#52023): #23734 (earliest open, removes _is_direct_openai_url entirely) and #23895 (relaxes the custom-provider GPT-5.x guard, #23893). This PR keeps a single model-family check and adds regression tests. Same goal, different code paths — related, not duplicate; flagging the cluster so a maintainer can pick one.

[Morad37]

One thing I noticed: the _is_direct_openai_url() path was the only signal that made the upgrade apply for non-GPT-5 models served through api.openai.com. Dropping it entirely means those models stay on chat completions, which is correct. But the comment says the URL signal is still useful as a "model-family amplification" -- and the code still reads api.openai.com URLs in _is_direct_openai_url(), it just doesn't use it as an independent gate anymore. Might be worth cleaning up the comment to not imply the URL signal is still doing anything functional. It's just explanatory now.