0.40.0 Cursor picker: only Grok keeps a real reasoning ladder (Sol/Opus/Fable dummy Medium)

[GusevV1987]

Summary

On bb 0.40.0, the Cursor (acp-cursor) New Thread picker still lets me change reasoning for Grok 4.6 (Low / Medium / High / Extra High). For GPT-5.6 Sol, Claude Opus 5, Claude Fable 5, and Composer 2.5 the same row is stuck on a single Medium described as “Reasoning effort is managed by the connected ACP agent.” Yesterday on 0.39.0 those families advertised a real ladder (none–max for Sol / Opus / Fable). Cursor still exposes Sol’s full effort list over ACP after the model is selected; bb’s catalog probe does not pick it up.

Versions and environment

• bb 0.40.0 (desktop app, auto-updated from 0.39.0 on 2026-08-26). bb --version → 0.40.0.
• macOS 26 (darwin 25.6.0), local daemon, no Connect, default data dir ~/.bb.
• Cursor CLI cursor-agent / agent 2026.08.11-e8db854 (cursor-agent --version).
• Provider: acp-cursor. Reproduced with bb provider models acp-cursor --json (no thread required).
• Still present on main at ead4c605d0fe16066baef12fa544ab14243c9f78 (same Cursor probe constants as tag desktop-v0.40.0 / f3cab2dd8c5c4be6d450be318550f3a04c8c3a1f).

Steps to reproduce

1. Install bb desktop 0.40.0 with cursor-agent 2026.08.11 on PATH and Cursor logged in (cursor-agent login).
2. Run:

bb provider models acp-cursor --json

3. Open New Thread, provider Cursor, select GPT-5.6 Sol, open the Reasoning control.
4. (Optional, to show Cursor itself still has the ladder.) Start cursor-agent acp, initialize with clientCapabilities._meta.parameterizedModelPicker: true, session/new, then session/set_config_option model=gpt-5.6-sol and read configOptions with category: "thought_level".

Did not reproduce as a Cursor-account / billing problem: the same CLI login still returns Sol’s none–max list in step 4. Did not reproduce on 0.39.0 (catalog used cursor-agent --list-models and family ids such as gpt-5.6-sol-medium).

Expected vs actual

Actual (bb provider models acp-cursor --json on 0.40.0, trimmed):

grok-4.6        default=xhigh   efforts=low, medium, high, xhigh
gpt-5.6-sol     default=medium  efforts=medium   # description: Reasoning effort is managed by the connected ACP agent.
claude-opus-5   default=medium  efforts=medium
claude-fable-5  default=medium  efforts=medium
composer-2.5    default=medium  efforts=medium

The New Thread Reasoning control matches that catalog: Grok has a real list; Sol does not.

Expected: Sol / Opus / Fable keep a real reasoning list in the picker, as on 0.39.0 when the catalog came from cursor-agent --list-models (family ids gpt-5.6-sol-medium, claude-opus-5-thinking-medium, claude-fable-5-thinking-medium with none–max).

Evidence

Observed (this machine, 2026-08-26):

• 0.40.0 catalog as above. Fast remains a separate supported tier on acp-cursor.
• Cursor ACP with parameterized picker: session/new returns 35 model options. Switching model via session/set_config_option takes ~1.7 s each here.
• After switching to gpt-5.6-sol, Cursor returns id: "reasoning", category: "thought_level", values none, low, medium, high, xhigh, max (current medium), plus fast off/on.
• After switching to grok-4.6: id: "effort", values low, medium, high, xhigh.
• After switching to claude-opus-5 / claude-fable-5: two thought_level options — thinking true/false first, then effort low–max.
• Composer 2.5: Fast only; no thought_level.

0.39.0 vs 0.40.0 in tree (permalinks):

• 0.39.0 Cursor catalog was CLI modelCli.listArgs: ["--list-models"] with family ids including gpt-5.6-sol-medium: acp-launch-specs.ts on desktop-v0.39.0.
• 0.40.0 opts into Cursor’s parameterized picker and probes Grok first in a bounded window: known-agents.ts#L74-L86 (reasoningProbePriorityModelIds: ["grok-4.6", "grok-4.5"]).
• Probe timeout is 5 s, then the connection is killed and whatever was collected is returned: bridge.ts#L540, discoverAcpNativeReasoningByModel.
• Unprobed models get the dummy Medium: model-catalog.ts#L43-L47.
• findAcpThoughtLevelConfigOption returns the first category === "thought_level" option: model-catalog.ts#L140-L145.

Hypothesized cause (labeled): 35 models × ~1.7 s cannot finish in 5 s. Priority is Grok 4.6 then Grok 4.5, so the window fills with Grok. Sol is later in the remaining-option walk and keeps the dummy Medium. This matches the 0.40.0 comments in known-agents.ts (“Probe the Grok families first so the bounded discovery window captures their full effort ladders”).

Related (same probe, not required to repro Sol): if Opus/Fable were probed, the first thought_level option is Thinking on/off (true/false), which bb cannot map to low–max; the real Effort list is the second option. Composer has no thought-level list to discover.

Workaround: bb thread spawn --provider acp-cursor --model gpt-5.6-sol --reasoning-level max is accepted by the CLI. Whether Cursor then applies max on Sol was not verified with a live spawn in this report; ACP after a model switch does include max as a reasoning value.

There is no customAgents / customAcpAgents knob that can override reserved acp-cursor probe priority or the 5 s timeout (docs/configuration.md).

What you ruled out

• Not a missing Sol ladder in Cursor ACP — live thought_level after model=gpt-5.6-sol is none–max.
• Not Cursor Fast / parameterized-picker support in general — Fast still appears; Grok’s effort list is complete.
• Not feat(acp): let agents supply their own reasoning levels #1113 (Amp custom level names) — this is the built-in Cursor catalog after 0.40.0.
• Not Recursive tool schemas (z.json()) break Grok 4.6: every turn fails with a generic provider error #1612 (Grok workflow $ref) — picker catalog, no prompt.
• Searched open and closed issues for acp-cursor reasoning / parameterized picker / “managed by the connected ACP agent” / gpt-5.6-sol; no duplicate of this 0.40.0 picker regression.
• customAcpAgents cannot replace acp-cursor (reserved always-listed id).

Suggested priority and effort

High for anyone who picks Cursor Sol/Opus/Fable from New Thread and thinks Medium is the only effort. Workaround: CLI --reasoning-level (Sol likely; Opus/Fable likely still wrong until the first-thought_level match is fixed). Effort looks Low–Medium: probe primaryModels (five Cube families) instead of all 35, give that walk enough time (~12 s at 1.7 s/switch here), and when several thought_level options exist pick the one whose values map to bb reasoning levels (effort / reasoning), not Thinking on/off.

• Reproduced on latest release 0.40.0; same constants still on main at ead4c605.
• Searched open and closed issues.

AGENT GENERATED: by Cursor Grok 4.6
Investigation: local bb thread thr_vkbcc7h6uk (not publicly readable).

[SawyerHood]

Reproduction & root-cause report: https://get-bb.github.io/reports/issues/2503.html

• Verdict: REPRODUCED · root-cause confidence: high
• Root cause (short): The five-second Grok-first ACP probe times out before Sol and Claude primary models, and bb selects Thinking instead of Effort.
• Proposed fix: Probe Cursor primary models before selected-only models with an adequate budget. Select the thought_level option with mapped reasoning values. Add delayed 35-model and dual-option tests.
• Caveats: Cursor ACP supplies no reasoning list for Composer 2.5. The report did not run bb 0.39.0 or start a paid Cursor turn. The raw GitHub fallback works after publication; the local artifact path works before publication.

The report has copy-paste repro steps, screenshots where the bug is visual, code permalinks at the base commit, and repro artifacts. It was independently re-run by a second agent (see its Verification section).

AGENT GENERATED: by GPT-5.6-Sol