[Bug]: Grok subagents do not show correctly in the Agents panel

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Start T3 Code with Grok Build authenticated (grok login) and Grok CLI on PATH.
2. Create a new thread on the Grok provider, model grok-4.6.
3. Send a prompt that causes the parent to call spawn_subagent at least twice, for example: Use spawn_subagent to launch two independent explore subagents, each inspecting a different part of this repo, then summarize what they found.
4. While the children are running, open the thread's Agents panel.
5. Wait until the parent turn settles (end_turn) while the subagents are still working, and look at the Agents panel and composer again.

Expected behavior

Each Grok spawn_subagent child should appear in the Agents panel as a live roster row (title/role, status, activity, tokens), the same way Claude Task and Codex collab children already do via task.started / task.progress / task.completed.

The panel empty-state copy already promises this:

When this thread spawns subagents or runs a workflow, they show up here with live status, activity, and token usage.

After the parent turn ends, in-process Grok children that are still running should keep their rows in a running/idle state instead of making the thread look finished.

spawn_subagent should not be the only (and bulky) representation of that work in the parent transcript.

Actual behavior

Grok children run, but they never show correctly in the Agents panel:

• The parent transcript only gets a fat spawn_subagent tool row (plus later get_command_or_subagent_output polls).
• The Agents panel stays empty (No agents yet), even while several children are in flight.
• Child ACP session updates are dropped, so there is no live child status, model, or token usage in the roster.
• The parent Grok turn often ends (end_turn) while those children keep running inside the Grok process. T3 then looks idle even though work is still happening.

This is reproducible on current main with Grok CLI 1.0.13 / grok-4.6. I hit it on ordinary multi-agent turns, not a one-off.

Impact

Major degradation or frequent failure

Version or commit

T3 Code (Alpha) 0.0.38; main @ 9cb40178a53cca279c67a9079afab3cddf6b6ddb

Environment

macOS 26.5.1 (arm64), T3 Code (Alpha) 0.0.38 desktop, Grok CLI 1.0.13 (5e9a58528b76) [stable], provider Grok Build, model grok-4.6

Logs or stack traces

# Current main has no Grok task.* mapping and no spawn_subagent detector:
$ rg -n 'type: "task\\.' apps/server/src/provider/Layers/GrokAdapter.ts
# (no matches)

$ rg -n 'spawn_subagent|subagent_spawned|GROK_SUBAGENT' apps/server/src/provider -g '*.ts'
# (no matches)

On main @ 9cb40178:

• GrokAdapter consumes ACP session/update as generic tool-call / content / plan events only. ToolCallUpdated is forwarded with makeAcpToolCallEvent and is not classified as a subagent task, so spawn_subagent renders as an ordinary tool row (apps/server/src/provider/Layers/GrokAdapter.ts around the ToolCallUpdated branch).
• The only registered xAI extension methods are _x.ai/ask_user_question, _x.ai/exit_plan_mode, and _x.ai/session/prompt_complete. There is no handler for _x.ai/session/update kinds subagent_spawned / subagent_progress / subagent_finished (or workflow_updated).
• One ACP runtime projects one root session and intentionally drops child-session updates so they are never flattened into the parent stream:

// apps/server/src/provider/acp/AcpSessionRuntime.ts
// One runtime projects one root ACP session. Child-session updates need
// explicit lineage routing and must never be flattened into this stream.
if (
  startState._tag !== "Started" ||
  notification.sessionId !== startState.result.sessionId
) {
  return;
}

That drop is covered by drops session updates emitted for a child ACP session in AcpJsonRpcConnection.test.ts (V1 choice from #3156). It is correct for not merging child transcripts into the parent timeline, but it means Grok children stay invisible unless the adapter emits task.* itself.

• The Agents panel roster is populated only from persisted task.* activities (packages/client-runtime/src/state/subagentRuntime.ts). Claude, Codex, and Antigravity adapters emit task.started; GrokAdapter does not.
• Because Grok's parent turn can end_turn while spawn_subagent / monitor work continues in-process, the missing task.* stream also leaves the thread looking idle after the parent settles.

Workaround

Watch the parent-transcript spawn_subagent / get_command_or_subagent_output tool rows. There is no Agents-panel workaround; the roster never receives those children.

Additional context

Not a duplicate of the closed Grok feature tickets — those never landed.

Item	State	Why it is not this bug
#6468 / discussion #6983	closed as completed / transferred	Feature request for grok-4.6 Agents-panel support. Closed during the issues→discussions move. The implementation PRs were not merged.
#5420 / discussion #7019	closed issue; discussion still open	Earlier mapping request (spawn_subagent → task.*). Same gap, filed as a feature, never shipped.
#6410, #5503, #7070, #5425	closed, unmerged	Adapter PRs that would have emitted task.* / hidden the bulky tool row.
#8412	open PR, no linked issue	feat(grok): show Grok subagents in the Agents panel — maps subagent_* / workflow_updated onto task.*.
#9139	open PR, no linked issue	fix(grok): emit task lifecycle for background subagents and monitors — maps spawn_subagent / monitor / poll / kill onto task.* so the thread stays live after parent end_turn.
#8499	open, Codex	Codex-only Agents panel miss.
#5952	open, OpenCode	OpenCode native Task tool.
#7281	open	Wrong parent model/effort labels on rows that do appear (Claude). Grok never gets the rows.
#5479	closed	CLI-delegated children (codex exec / opencode run) from another provider, not Grok-native spawn_subagent.

Please treat this as a bug against current main, not as a duplicate of the closed/transferred feature requests. The UI contract already exists; Grok is the provider that does not feed it.

Smallest useful scope is still Grok-adapter-only:

1. Emit task.started / task.progress / task.completed from Grok's subagent lifecycle (subagent_spawned / subagent_finished and/or the spawn_subagent tool identity).
2. Keep child transcripts out of the parent timeline (the existing child-session drop can stay).
3. Prefer a roster row over the bulky spawn_subagent tool row.
4. Keep unique event stamps and per-taskId dedupe so polls / late finishes do not clone rows.

#8412 and #9139 already point at that mapping. This issue exists so that work has a current, non-duplicate bug to close against, instead of the completed-but-unshipped feature tickets.

Start in apps/server/src/provider/Layers/GrokAdapter.ts at the ToolCallUpdated branch, then compare its event handling with the task.* lifecycle emitted by other adapters. Read apps/server/src/provider/acp/AcpSessionRuntime.ts and packages/client-runtime/src/state/subagentRuntime.ts to preserve child-session isolation while feeding the roster. Done means Grok children produce deduplicated task.started, task.progress, and task.completed events, avoid a bulky parent tool row, and keep the thread active until background work finishes.