fix(models): route Responses-only Copilot models to the Responses API

[SvichkarevAnatoly]

What does this PR do?

Routes Responses-only Copilot models to the Responses API so they stop failing with HTTP 400 "model … is not accessible via the /chat/completions endpoint".

copilot_model_api_mode() picks the Copilot API mode primarily from a model-id regex (_should_use_copilot_responses_api) that only recognises GPT-5+ models as Responses-API models. A non-GPT model that is Responses-only — e.g. mai-code-1-flash-picker — falls through to chat_completions and the request fails.

This PR keeps the opencode-parity regex first (so GPT-5+ reasoning models stay on their native Responses endpoint even when the catalog also advertises /chat/completions), then consults the catalog's supported_endpoints to rescue models the regex doesn't cover: when a model does not advertise /chat/completions at all, it is routed by its declared endpoint — /responses → codex_responses, /v1/messages → anthropic_messages. Models that do advertise /chat/completions (Claude, Gemini on Copilot) keep using it, so their behaviour is unchanged.

Model	supported_endpoints	before	after
mai-code-1-flash-picker	/responses	chat_completions → 400	codex_responses ✅
claude-* / gemini-* (Copilot)	/chat/completions, …	chat_completions	chat_completions (unchanged)
Messages-only model	/v1/messages	anthropic_messages	anthropic_messages (unchanged)

Follow-up to #58830 (raw-token exchange). #58830 is required first — without it every Copilot model fails with the integrator error, so this change is only observable once tokens are exchanged correctly. Together they make every model advertised by a Copilot seat usable.

Related Issue

Relates to #45813

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• hermes_cli/models.py: in copilot_model_api_mode, replace the "messages-only" catalog check with a general "no /chat/completions" fallback that honours /responses (→ codex_responses) and /v1/messages (→ anthropic_messages).
• tests/hermes_cli/test_model_validation.py: add 3 unit tests — Responses-only non-GPT model → codex_responses, Messages-only → anthropic_messages, chat-capable non-GPT stays chat_completions.

How to Test

1. On top of fix(runtime): exchange raw Copilot OAuth token in runtime resolution #58830, configure the Copilot provider and select mai-code-1-flash-picker.
2. On main: HTTP 400 … not accessible via the /chat/completions endpoint.
3. With this patch: returns 200 via the Responses API and the model answers. All other Copilot models keep their previous api_mode.
4. pytest tests/hermes_cli/test_model_validation.py -q → 90 passed.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(models):)
• I searched for existing PRs (no duplicate; follow-up to fix(runtime): exchange raw Copilot OAuth token in runtime resolution #58830)
• My PR contains only changes related to this fix
• I've run the tests and they pass (tests/hermes_cli/test_model_validation.py: 90 passed)
• I've added tests for my changes
• I've tested on my platform: macOS 15

Documentation & Housekeeping

• No user-facing docs/config keys changed — N/A for README / cli-config.yaml.example
• No architecture/workflow change — N/A for CONTRIBUTING.md / AGENTS.md
• Cross-platform safe (pure-Python catalog/string handling)

[tonydwb]

Code Review Summary

Verdict: Comment (LGTM)

Routing fix: routes Responses-only Copilot models to the Responses API. Correctly distinguishes between models that use the Responses API endpoint vs. the Chat Completions endpoint. 2 files changed — model routing logic and tests.

[teknium1]

Thanks for tracing the catalog-routing gap. The primary defect is present on current main: hermes_cli/models.py:3408-3425 routes non-GPT catalog entries through a fallback that recognizes only Messages-only models, so a Responses-only entry reaches chat_completions. The proposed catalog fallback addresses that path while retaining the existing GPT-first rule.

Problems

• agent/chat_completion_helpers.py:1503-1534 independently computes fallback API mode. Its Copilot branch ultimately uses the GPT-name-only predicate in run_agent.py:1430-1433, so a non-GPT Copilot fallback advertising only /responses still activates as chat_completions.

Suggested changes

• Reuse copilot_model_api_mode() or a shared catalog-aware resolver in the fallback route, and add a fallback-activation regression test for a non-GPT /responses-only catalog entry.

Automated hermes-sweeper review.

[teknium1]

This fixes the primary resolver, but fallback activation still derives Copilot mode separately in agent/chat_completion_helpers.py:1503-1534 through the GPT-name-only predicate. Please route that sibling path through the same catalog-aware decision and add a fallback regression test; otherwise a non-GPT /responses-only fallback remains on Chat Completions.

[intellectronica]

I reproduced this with Copilot Grok 4.6 and opened #88538 by cherry-picking 0df0765 (preserving your authorship), then adding the catalog-aware auxiliary and fallback activation paths requested in the maintainer review. Thank you for the original fix.