fix(copilot): route responses-only models across runtimes

[intellectronica]

Summary

• route Copilot models advertised as /responses-only through codex_responses, including Grok 4.6
• apply the same catalog-aware decision to primary agent initialisation, auxiliary clients, and fallback activation
• preserve Copilot Claude models on the OpenAI-compatible chat path

Provenance and related work

This salvages the original implementation from #58835 by cherry-picking its commit, preserving Anatoly Svichkarev's authorship, then addresses the maintainer review on that PR: non-GPT Responses-only models also need catalog-aware routing in the auxiliary and fallback activation paths.

Related overlapping PRs: #74377, #85224, #86075. This branch differs by covering all three runtime paths with regression tests while retaining current-main Claude transport behaviour.

Problem

Copilot's live model catalog advertises Grok 4.6 with:

{"supported_endpoints": ["/responses"]}

Hermes classified Copilot transport primarily from the GPT-5 model-name heuristic. Non-GPT Responses-only models therefore used /chat/completions and failed with:

HTTP 400: model "grok-4.6" is not accessible via the /chat/completions endpoint

The primary model helper, auxiliary client wrapper, and fallback activation each made that decision separately, so fixing only copilot_model_api_mode() left sibling paths broken.

Approach

• make copilot_model_api_mode() honour /responses when /chat/completions is absent
• pass the active credential into the shared Copilot resolver during agent initialisation and fallback activation
• use the same resolver before auxiliary request headers are constructed, then wrap Responses-only models with CodexAuxiliaryClient
• keep GPT-5 heuristics and explicit API-mode overrides unchanged

Verification

• regression tests were added before the fix and failed on current main in all three paths
• scripts/run_tests.sh across six affected test files: 182 passed, 0 failed
• Ruff checks and git diff --check: passed
• a live end-to-end Copilot request could not be completed while GitHub Status reported a Copilot major outage and the token-exchange endpoint returned HTTP 503; the runtime failure itself is captured in existing Hermes logs and the regression tests use the catalog contract

Risk

Low and provider-scoped. Catalog metadata only upgrades a model when /responses is explicitly present and /chat/completions is absent. Chat-capable Claude and Gemini models retain their current transport.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

The routing principle is right: the vendor/name pattern check stops being the sole authority, and the catalog's supported_endpoints decides — with two deliberate guardrails kept ahead of it (GPT-5 pattern first, Copilot-Claude pinned to chat regardless of what the catalog lists). Moving the transport resolution before header construction in _wrap_if_needed fixes a real interleaving bug (catalog discovery issuing its own Copilot request after headers were built), threading api_key into _provider_model_requires_responses_api means main-line and fallback activation now share one catalog-aware decision instead of the fallback silently downgrading to the name heuristic, and the tests cover both directions that matter: responses-only Grok upgrades everywhere (aux wrapper, primary, and fallback activation), while Claude stays on chat even when the catalog lists /responses.

1. hermes_cli/models.py:copilot_model_api_mode (dual-endpoint branch untested) — the upgrade fires only when /responses ∈ endpoints AND /chat/completions ∉ endpoints; nothing pins the second conjunct. Why it matters: a model advertising both endpoints must stay on chat_completions (cheaper path, existing behavior), and a future edit dropping the not in clause would flip every dual-endpoint model to Responses with all current tests green. Suggestion: add the parametrized neighbor — supported_endpoints: ["/responses", "/chat/completions"] → chat_completions — next to the two existing catalog cases.

Nit: the pre-existing Claude/chat comment block immediately below the new catalog fallback now explains the same invariant the new early-return above it already enforces; folding them into one statement (or deleting the stale half) would keep the function's contract stated once.

— Reviewed by Hermes AI reviewer (reviewer-f2)

[intellectronica]

Addressed in 7d31824.

• added the dual-endpoint neighbour: supported_endpoints: ["/responses", "/chat/completions"] → chat_completions
• folded the duplicated Claude/chat comment into the early-return so the invariant is stated once

scripts/run_tests.sh on the six Copilot routing files: 180 passed, 0 failed.

[munim]

@alt-glitch when can we expect this to be merged? its actually blocking copilot grok users.