feat(grok): native ACP parity (catalog, effort, usage, plan, multi-agent, compact)

[EnzoTironi]

What Changed

Grok provider ACP parity (self-contained branch: shared parser commit + Grok adapter commit):

Area	Behavior
Shared ACP	session/update kinds → typed events; set_model _meta pass-through (#5422 same commits)
Catalog	Live commands → slash/skills; empty live catalog clears stale; static /compact always present
Effort	Per-model Reasoning select; apply via session/set_model _meta.reasoningEffort (no process restart)
Usage	Prompt _meta + usage_update → context meter; per-model windows; idempotent
Plan entry	Plan mode → /plan text; Plan/Build toggle for Grok
Plan surface	exit_plan_mode → proposed-plan card + abandoned capture; session plan.md only under ~/.grok/sessions/
Multi-agent	spawn_subagent → task.* with unique stamps + dedupe
Compact	/compact + auto_compact_completed → compacted work-log row

Fixes #5417 #5418 #5419 #5420.

Merge note: Prefer merging #5422 first (shared parser only, size L), then this. If maintainers want one PR, this tip already contains both commits and can land alone.

Why

Grok Build already exposes these on the wire. T3 left them dark while Claude/Codex surface them. One adapter-boundary PR avoids unmergeable half-states from mid-file splits.

Related: #5403 #5405 #5409 #5412 #5131 — this tip absorbs those lakes with review fixes.

UI Changes

Uses existing Traits / meter / proposed-plan / task rows / Plan toggle — no new chrome.

Grok selected with effort menu

Plan mode toggle

Slash menu: /plan (built-in) + /compact (provider)

Checklist

• Small and focused — large by necessity (full Grok adapter). feat(server): ACP sessionUpdate parity for Grok clients #5422 is the small shared slice if you want a smaller first merge.
• Explained what/why
• UI screenshots for effort, plan toggle, slash/compact
• Video N/A

Out of scope

• Desktop / Zoen brand
• Web sticky → feat(web): sticky provider options and Grok composer affordances #5426
• feat(orchestrator): Surface Grok reasoning effort options #5160 dump

Test plan

• 72 unit tests (GrokAcpSupport, XAi, GrokProvider, GrokAdapter)
• Live spawn_subagent probe
• set_model effort without process restart

Model: grok-4.5 (Grok Build)

---

Note

High Risk
Large, central Grok provider/ACP adapter changes affect turn lifecycle, catalog merging, plan-mode prompt rewriting, and token usage—behavior users rely on in every Grok thread; regressions would be user-visible across sessions.

Overview
Brings Grok’s live ACP wire behavior into T3’s existing provider UI: context meter, Plan/Build, slash/skills, proposed-plan cards, task rows, and compaction work-log entries—without new chrome.

Shared ACP parsing now turns more session/update kinds into typed events (reasoning chunks, usage, available commands, config options, session info, user echoes) and passes optional _meta through session/set_model.

Grok provider snapshot enables Plan/Build and in-thread model changes (requiresNewThreadForModelChange → false), discovers models with Reasoning option descriptors from _meta, merges live slash/skills from initialize and available_commands_update (empty catalog clears stale entries), always keeps static /compact, and improves probe auth messaging.

Grok driver publishes catalog changes immediately via PubSub so clients do not wait on the next health probe.

Grok adapter maps those events to runtime semantics: token usage from prompt _meta and usage_update; thread titles from session_info_update; /compact / auto_compact_completed → thread.state.changed compacted; exit_plan_mode and session plan.md writes → turn.proposed.completed with abandoned capture so turns do not hang; spawn_subagent → task.* with dedupe; plan mode via /plan and /default prompt prefixes; mid-thread reasoning effort via session/set_model _meta.reasoningEffort (CLI flags only on spawn); optional onAvailableCommands callback for the driver.

Mock agent and tests are extended to simulate initialize commands, usage meta, compaction, session info, skills-reload response ids, and effort set_model without process restart.

^{Reviewed by Cursor Bugbot for commit 29e5c6b. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add native ACP parity to Grok provider including catalog, reasoning effort, token usage, plan mode, multi-agent, and compact support

• GrokAdapter.ts now emits thread.token-usage.updated from both prompt RPC meta and session/update UsageUpdated events, including per-model context window tracking.
• Plan mode is supported end-to-end: prompt text is prefixed with /plan or /default based on interactionMode, enter_plan_mode tool calls toggle plan state, and turn.proposed.completed events are emitted from plan.md writes and exit_plan_mode.
• Multi-agent support tracks Grok subagent tool calls and emits task.started/task.completed events; /compact prompts trigger thread.state.changed via auto_compact_completed session notifications.
• ACP command catalog is discovered at initialization and live-updated: slash commands and skills are derived from availableCommands in the ACP init result, with /compact always present as a static fallback via ensureGrokStaticSlashCommands.
• Reasoning effort is applied via session/set_model _meta.reasoningEffort (or session/set_config_option when exposed as a config option), and models now surface a Reasoning select in capabilities when advertised.
• session_info_update notifications map to thread.metadata.updated; agent_thought_chunk maps to ContentDelta with streamKind: 'reasoning_text'.
• Provider presentation now shows the interaction mode toggle and no longer requires a new thread on model change.
• Behavioral Change: provider snapshots are pushed immediately on commandCatalogChanges, not only on health probes or setting changes.

^{Macroscope summarized 29e5c6b.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a1107d08-4425-42a6-923f-896cbe02c01a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This feature PR introduces substantial new Grok provider capabilities including reasoning effort configuration, token usage tracking, Plan/Build mode mapping, multi-agent task visibility, compact commands, and slash command catalogs. The scope and complexity of these new runtime behaviors warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}

[EnzoTironi]

Review findings from Bugbot/Macroscope for this slice are fixed on the Zoen product tip (zoen/main @ 16474ab0d). Will restack / re-push these PR heads after further slim-slicing. Open remaining: effort restart without session.exited (or adopt #5403 set_model path).

[EnzoTironi]

UI screenshots

Grok model + effort

Plan mode

Slash: /plan + /compact

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 0792ff8. Configure here.}

[ahmed-besic]

Man I just wanted grok to work normally with some QoL improvements, and you took that and turned it up to 11
hopefully it works, its a ton of changes

[EnzoTironi]

Man I just wanted grok to work normally with some QoL improvements, and you took that and turned it up to 11 hopefully it works, its a ton of changes

I am still working on it. I am now making the stop button work properly. You can checkout to the PR and test it out.

[ahmed-besic]

Oh yeah, i totally forgot about that, the thread still keeps going even if you interrupt it.

There is also steer/queue prompts
(currently the message gets send but its put in the queue, in t3 code it continues working unter that message, you think you steered but you didn't, it only starts workingon that when it replies)