[VIPKaiser]

Problem

grok-cli/<id> for any model id that is not in the static grok-cli seed (e.g. grok-4.7 on its release day, grok-4.7-build-fast) gets HTTP 400 from Grok Build until the connection has run a model sync.

• GrokCliExecutor always POSTs to Grok Build's /v1/responses, and live discovery only keeps responses backends, so every grok-cli model is a Responses model.
• But getModelTargetFormat("gc", id) returns null for ids missing from the seed, so resolveChatCoreTargetFormat falls back to the provider format ("openai") and the translator builds a chat-completions body (messages / max_tokens) that is posted to /v1/responses.
• Seeded ids (grok-4.5 / 4.6) work only because they carry targetFormat: "openai-responses" explicitly.

Observed live on release/v3.8.51: grok-cli/grok-4.7 returned 400 on both pooled accounts (the upstream still echoed x-grok-context-window: 500000, so the model itself was recognised) while grok-cli/grok-4.6 succeeded. After a sync, the synced entry's apiFormat: "responses" masks the bug, but connections that never synced (existing ones; new ones before the first sync) keep hitting it.

Fix

One fallback in getModelTargetFormat, next to the existing openai *-pro heuristic (#5842): for the gc alias, return "openai-responses" when the seed has no explicit tag. The registry entry's format: "openai" is left alone on purpose (its comment keeps the generic translate-path contract stable).

• Seeded ids keep their explicit targetFormat (unchanged).
• xai / xai-oauth unknown ids still return null (their executors choose /chat/completions vs /responses themselves).

Relationship to open PRs

Complementary, no file overlap:

• fix(providers): keep grok-4.7 reasoning tiers from object catalogs #14525 seeds grok-4.7 in grok-cli (with reasoning tiers). This PR makes the next unseeded id work without waiting for a seed bump.
• feat(providers): surface new Claude, xAI, and Grok Build chat ids #14237 adds live discovery / autoSync for new Grok Build imports. This PR covers existing connections and the window before the first sync.

Tests

• New tests/unit/grok-cli-passthrough-responses-format.test.ts: alias and provider-id lookups for unknown ids, seeded ids unchanged, resolveChatCoreTargetFormat end-to-end for an unsynced id, and a negative check that xai / xai-oauth are unaffected.
• Existing grok-cli-*, *target-format*, xai*, chatcore-upstream-body suites: 287 + 38 pass locally. Typecheck and lint clean for the touched files.

A changelog fragment will follow once the PR number is assigned.

[diegosouzapw]

Nice catch — the live 400 you observed on release/v3.8.51 (with the upstream echoing the
context-window header, confirming the model itself was fine) makes a strong case, and the fix
mirrors the existing *-pro heuristic pattern closely. The negative test for xai/xai-oauth
staying null is a good touch — it proves the scoping claim rather than just asserting it. No
file overlap with #14525 or #14237, so all three can land independently. Couldn't get a
dedicated live test run in during this pass due to heavy shared-devbox contention; marking
PLAUSIBLE on that basis. Merge-ready.