Agent Task subagents spawn on Fable/GPT without user requesting those models

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

In Agent chat, Task/subagents are started on Fable (and related GPT-class allowlist slugs) even when the parent chat is Grok 4.5 and the user never asked for those models. Failed subagents then auto-switch with: API usage limit reached Switched to composer-2.5 after reaching API limit.

Steps to Reproduce

  1. Open Cursor IDE Agent chat with parent model Grok 4.5.
  2. Run a multi-step workflow that causes the agent to spawn parallel Task subagents (example: /janitor fleet PR prep).
  3. Do not request any subagent model override.
  4. Observe Task launches using model: claude-fable-5-thinking-high (and/or UI referring to Fable/GPT).
  5. Observe failed Task runs reporting automatic switch to composer-2.5 after API usage limit.

Expected Behavior

Subagents should inherit the parent model (or only use a different model when the user explicitly asks). Agents should not pick Fable/GPT allowlist slugs unprompted, and the runtime should not silently re-route Task subagents to other models without clear user consent.

Operating System

MacOS

Version Information

IDE:
Version: 3.11.19
VSCode Version: 1.125.0
Commit: bf249e6efb5b097f23d7e21d7283429f0760b740
Date: 2026-07-12T21:39:24.175Z
OS: Darwin arm64 25.5.0

For AI issues: which model did you use?

Parent: Cursor Grok 4.5. Subagents incorrectly launched as claude-fable-5-thinking-high; runtime then reported switch to composer-2.5. Task allowlist also exposes gpt-5.6-sol-high.

For AI issues: add Request ID with privacy disabled

Data sharing is enabled (privacy mode off) – please use these IDs.

Composer / conversation ID:
9c71d064-ed1b-4f46-9861-397d5195a343

Parent Request IDs (Grok 4.5 chat that spawned the Task subagents):

  • 7474a89f-1202-43b8-8fd5-2a26cadeec1b – primary parent turn that launched the Fable Task subagents (2026-07-14T16:35:55Z)
  • 13787516-d75e-453c-b6e7-2c9d6fee876c – follow-up parent turn (2026-07-14T16:43:09Z)
  • 6d6885a6-cc61-480c-a104-3adb6111efdb – parent turn handling subagent completions (2026-07-14T16:43:34Z)

Subagent Request IDs / composer IDs from subagent.request spans under parent request 7474a89f-...:

  • requestId 8f26ba29-d59a-4663-bc8e-2457b29097a1 → composerId 0413f3aa-243c-4a68-8ce7-42380cb77385 (failed: API usage limit → switched to composer-2.5)
  • requestId e662eac2-89de-49e7-846a-65475a9d1533 → composerId cb938239-83f0-4c34-9a3c-a4c8d6208642 (failed: API usage limit → switched to composer-2.5)
  • requestId 8707d0ad-caa8-43c9-bbc7-408e7942ee70 → composerId 80cb3f82-8a9f-4566-962e-1bee67afb519 (completed)

Additional failed Task/subagent composer IDs from this session (same error: API usage limit reached Switched to composer-2.5 after reaching API limit.):

  • e525f4e7-7922-487c-a40a-3dbef7fee870
  • 0a974aec-84ce-47e1-a16f-0fa8ded20d39
  • 6fe2c673-8640-46fb-8814-1c18bb3a1a02

Additional Information

Evidence from this session:

  • Parent agent Task tool calls included explicit model: "claude-fable-5-thinking-high" even though the user did not request a model.
  • Task tool prompt/schema lists allowed slugs including claude-fable-5-thinking-high and gpt-5.6-sol-high, and says to only pass model when the user requests one – but the parent still chose Fable unprompted.
  • Subagent transcripts ended with: API usage limit reached Switched to composer-2.5 after reaching API limit.
  • Related forum threads cover BYOK/OpenAI subagent failures; this report is specifically about unsolicited model selection / switching for Task subagents.
  • The /janitor workflow that spawned the parallel Task subagents is open-sourced here for the Cursor team to inspect: GitHub - jsolly/agent-skills: Public AI agent skills mirrored from jsolly/dotagents (allowlisted) · GitHub

Data sharing is enabled for this account/session, so these Request IDs should be inspectable on your side.

Does this stop you from using Cursor

Yes - when API rate limits have been hit, this skill makes the session fail completely. The /janitor workflow depends on parallel Task subagents; once those are launched on Fable/GPT and hit API usage limit reached Switched to composer-2.5 after reaching API limit., the fleet pass dies and cannot finish, so Cursor is unusable for this workflow until the limit clears.

5

I just encountered what I think is the same issue/bug myself, reported here: Cursor This Mac (Remote Control) Spun Up Phantom GPT-5.5 High Fast Subagent

7

Hey there!

You’re actually seeing two separate things, and only one is a bug:

  1. Task subagents running on models you didn’t pick (Fable/GPT) is a known bug. When the agent spawns parallel Task subagents it can choose a model on its own instead of inheriting your Grok 4.5 parent, and it sometimes reaches for premium slugs unprompted. We’ve had several reports of this and the team is actively looking into the subagent model-selection behavior — no fix has shipped yet.

  2. “Switched to composer-2.5 after reaching API limit” is expected, not a silent re-route. Once you hit your usage limit, Cursor falls back to an included first-party model (Composer) so the session can keep going. The unrequested premium subagents in #1 are what burn usage quickly, which is what tripped the limit here.

Until the fix lands, the most reliable lever is to pin the model explicitly for your Task subagents instead of leaving it to the agent - e.g. set model: inherit (or a first-party model like composer-2.5) in the subagent/Task calls in your /janitor skill. An explicit model is respected; toggling models off in Settings won’t reliably stop subagents from selecting them. In the meantime, you can watch the subagent cards and cancel any that spin up on a model you didn’t want.

Model field details: Subagents → Model configuration

(@durden - the “terminated but still running” part you hit is a separate known issue; good to keep that one in your own thread so we can track it there.)