Grok provider: native ACP parity (catalog, effort, usage)

[EnzoTironi]

Before submitting

• I searched existing issues and did not find a duplicate.
• I am describing a concrete problem or use case, not just a vague idea.

Area

apps/server

Problem or use case

Grok Build already exposes slash/skills catalog, per-model reasoning effort, and context usage over ACP, but T3 leaves those dark for Grok while Claude/Codex surface them.

Proposed solution

Provider core for grok agent stdio: command catalog, reasoning effort application, token usage → context meter, auth/status, set_model continuity.

Depends on sessionUpdate parser (#5417 / #5422).

Why this matters

Users cannot pick effort, see context fill, or use installed skills from T3 on Grok threads.

Smallest useful scope

Server provider adapter only. Split catalog / effort / usage further if maintainers require size:L max.

Alternatives considered

• fix(grok): surface per-model reasoning effort in the composer #5403: effort via session/set_model _meta (no process restart) — re-probe vs CLI --reasoning-effort
• feat(orchestrator): Surface Grok reasoning effort options #5160: spawn-flag effort inside a huge multi-file dump — ignore packaging
• fix(providers): show Grok and OpenCode skills/slash commands in the picker #5131: skills/slash picker only ([Bug]: Grok provider: skills/slash commands empty; ACP crash on skills-reload (BigInt RequestId) #4109)

Risks or tradeoffs

Process restart for effort may emit session.exited; prefer set_model if live Grok accepts mid-thread _meta.

Examples or references

PR #5423. Overlaps #5403, #5405, #5131.

[Amadeus415]

Still missing on T3 Code Nightly 0.0.34-nightly.20260812.1079 + Grok CLI 1.0.3 (macOS).

Composer has no Reasoning / thinking control for Grok 4.6. Claude and Codex show the control; Grok does not. Turns stay on the Grok CLI default.

Workaround: set default_reasoning_effort in ~/.grok/config.toml and start a new T3 Grok thread. That is global, not a per-thread toggle.

Grok itself advertises the effort menu (low / medium / high / xhigh for 4.6). This looks like T3 publishing Grok models with empty capabilities instead of mapping ACP _meta.reasoningEfforts into the existing composer picker.

Related: #5403, #5503.

[juliusmarminge]

Thanks for documenting the Grok catalog, effort and usage gaps. The orchestrator V2 work has now merged in #2829.

V2's ACP integration now exposes the provider's command catalog, advertised model/reasoning options and usage updates. These feed the command picker, model options and context presentation through the shared ACP path.

Closing this as delivered. If a specific part is still missing in a build containing V2, please open a focused follow-up with the provider/version and the behavior you are seeing.