grok-4.5/grok-4.6 wire_api miscategorized as "completions" — 400 not accessible via /chat/completions

[jaroslawgajewski]

Summary

grok-4.5 and grok-4.6 are catalogued in pkg/cli/data/models.json (provider github-copilot) with "wire_api": "completions", but the live GitHub Copilot backend rejects requests for these models on the legacy Chat Completions endpoint. Every invocation fails with:

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

This reproduces deterministically (observed 4/4 retries failing identically in a Copilot-engine agent job) whenever a workflow selects grok-4.5 or grok-4.6 as the engine model.

Root cause

In pkg/cli/data/models.json → providers.github-copilot.models:

"grok-4.5": { "cost": {...}, "provider_type": "openai", "wire_api": "completions" },
"grok-4.6": { "cost": {...}, "provider_type": "openai", "wire_api": "completions" }

Every other contemporary reasoning/agentic model in the same catalog (gpt-5.5, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, mai-code-1.1-flash, mai-code-1-flash-picker, gpt-5.2, gpt-5.4*) is catalogued with "wire_api": "responses". Only older/simpler chat models (gemini-2.5-pro, gpt-4.1, kimi-k2.7-code, kimi-k3, raptor-mini) use "completions".

The copilot harness reads this catalog entry to auto-configure COPILOT_PROVIDER_WIRE_API before invoking the Copilot CLI:

[copilot-harness] auto-configuring COPILOT_PROVIDER_WIRE_API=completions for model grok-4.6
[copilot-harness] inference routing: mode=cli configuredModel="grok-4.6" endpoint=managed-by-copilot-cli
400 model "grok-4.6" is not accessible via the /chat/completions endpoint

Since the live backend only serves grok-4.5/4.6 via the Responses API (consistent with every other current-generation reasoning-capable model in the same catalog), the wire_api value for these two entries appears to be stale/incorrect — likely inherited from an earlier grok generation that was completions-only, and not updated when the model was migrated.

Note: the automated [model-inventory] bot (e.g. #59825) cross-checks pricing for grok-4.5/grok-4.6 against Copilot SDK/reflect sources and finds no discrepancy, but it does not appear to validate wire_api/endpoint compatibility, so this drift isn't caught by that audit.

Proposed fix

Update the two entries in pkg/cli/data/models.json (and its mirror actions/setup/js/models.json, per the existing "remain identical mirrors" invariant noted in #59825) to:

"grok-4.5": { "cost": {...}, "provider_type": "openai", "wire_api": "responses" },
"grok-4.6": { "cost": {...}, "provider_type": "openai", "wire_api": "responses" }

(Costs unchanged — only wire_api needs correcting.) Ideally the model-inventory automation would also gain a check that flags wire_api mismatches the same way it already flags pricing discrepancies, so this class of drift doesn't require a live-run failure to surface.

Repro

Any workflow using the Copilot engine with model: grok-4.6 (or grok-4.5) fails at the Execute GitHub Copilot CLI step with the 400 above, on every retry, with no other configuration changes.