-
Notifications
You must be signed in to change notification settings - Fork 37.8k
Description
Type: Bug
I have a repository setup that has a multi-agent handoff workflow that depends on a SKILL.md loading.
As of 12/25/2025: It seems to work with every model except the gpt- models.
Does not work with:
- GPT-5.2
- GPT-5.1-Codex
- GPT-5.1-Codex-Max
Works well with:
- Grok Code Fast 1
- Raptor mini (Preview)
- Claude Haiku 4.5
- Claude Opus 4.5
- Gemini 3 Flash (Preview)
- Gemini 3 Pro (Preview)
This isn't currently documented, so your mileage may vary. You can test it by doing the following:
Select agent-number-converter-first and provide a number (1234) as input. If the agent can call the subagents and produce the final decorated output, then it's working. Try it with Grok Code Fast 1.
Compatible:
| N | R | E | S |
|---|---|---|---|
| 1234 | MCCXXXIV | 1️⃣2️⃣3️⃣4️⃣ | one thousand two hundred thirty four |
GPT- Variants:
MCCXXXIV
If you select a non-compatible model, the agent will just return the output of the first subagent and stop there.
Extension version: 0.36.2025121901
VS Code version: Code - Insiders 1.108.0-insider (7f08f95, 2025-12-19T14:09:25.917Z)
OS version: Windows_NT x64 10.0.26200
Modes: