bug: Cursor thinking menu shows Off/On for Grok 4.6

[fidelix]

What's broken

On a Cursor agent, I selected Grok 4.6 and opened the thinking control (brain icon). I expected Grok's thinking levels (Off, Low, High, Extra high, Max). The menu only had Off and On, with Off selected.

This is intermittent. After a daemon restart, the same Grok 4.6 control sometimes shows the full list. I did not change the model for that check. Later it can go back to Off/On.

Steps to reproduce

Shortest path that shows the bug:

1. Open Paseo Desktop and connect to the local daemon.
2. Create a Cursor agent.
3. Select model Grok 4.6.
4. Click the thinking control (brain icon).

Expected: thinking levels for Grok 4.6 (Off, Low, High, Extra high, Max).

Actual: only Off and On.

If step 4 already shows the full list, restart the Paseo daemon, then repeat steps 3–4. After a restart the menu can drop back to Off/On. I cannot force Off/On on every start.

Where did this happen

Desktop (Electron)

Paseo version

0.7.0-beta.3 (git main)

OS version

Linux (WSL2, Arch)

Agent provider

Custom provider

Provider configuration

Built-in Cursor provider (cursor-agent ACP).

• cursor-agent: 2026.08.25-3e8eec8
• Composer model: Cursor Grok 4.6
• Fast control: Off

Logs

No crash and no error in the composer. This is the thinking menu showing the wrong options.

Screenshots or video

Screenshot of the Cursor composer:

• Model control: Cursor Grok 4.6
• Thinking control open: Off (checked) and On
• Fast control: Off

I could not attach the PNG through the GitHub CLI. I will add it in a comment.

Related issues

Searched open and closed issues. Related, but not the same:

• bug: ACP provider thinking options don't follow the selected model #2823 — Kimi ACP thinking options stay on the first model. Kimi, not Cursor.
• bug(cursor): create_agent fails with thinkingOptionId:"max" — the model's own advertised default #3466 — Cursor create_agent fails with thinkingOptionId: "max". Different failure.
• bug: Grok provider drops CLI-supported effort and permission modes #3469 — Grok CLI provider, not Cursor's Grok 4.6 model.

[fidelix]

Screenshot from the composer: Cursor Grok 4.6 selected, thinking menu only Off / On.

(The GitHub CLI cannot attach the PNG to the issue form. If this comment has no image, drag cursor-thinking-off-on.png onto the issue.)

[edward821220]

Independent reproduction, 2026-09-04, macOS desktop.

• Paseo 0.7.2 (packaged app)
• cursor-agent 2026.09.02-c22c1a3 (unchanged between the two agents)
• Same Cursor model on both: grok-4.6

Two Cursor agents, two minutes apart, different stored thinking:

local time	agent	cwd	config.model	stored thinkingOptionId
22:48	a6c78c29	hrornah-line	grok-4.6	xhigh
22:50	c3715aeb	paseo	grok-4.6	true

The 22:50 agent JSON still has the boolean on Grok:

"model": "grok-4.6",
"thinkingOptionId": "true"

After the 22:48 daemon log rotate / restart, applying Grok effort on that agent:

23:00:55 agent.config.apply.request {"modelId":"grok-4.6","modeId":"agent","thinkingOptionId":"xhigh"}
23:00:57 error: acp does not expose ACP thought-level selection

Same agent later, ACP still holding the boolean while Paseo asked for xhigh:

23:16:14 set_agent_thinking_request thinkingOptionId=xhigh
error: Invalid params
data: Invalid value for effort: true

Did not change Paseo or cursor-agent between 22:48 and 22:50. Opening the model picker did not re-probe.

Related: #4180 for the catalog stamp, #1735 for the live session still using the session/new config after unstable_setSessionModel.