Cross-provider subagent delegation fails (Codex → GOAT MiniMax)

Đang mở
  • Có gì không ổn?
#1,096 1 bình luận 0 reaction 1 người được giao Xem trên GitHub

Maintainer thường phản hồi trong vòng 1 ngày

Chưa có ai nhận issue này.

Đánh giá

Độ khó
4/5
Thời gian dự kiến
3-5 ngày
Mức phù hợp với người mới
18/100
Loại issue
Lỗi
Độ rõ ràng
Khá rõ ràng
Mức độ hoạt động
Sôi nổi
Công nghệ
nodejs
Lĩnh vực
ai, authentication

Hướng nghiên cứu

The payload names createHarnessSession and runSubAgent, plus provider resolution and auth routing, as the places where the Director's provider is reused for the child. Start by reading those functions to see how a subagent's pinned model is matched to a provider, and reproduce with a custom subagent pinned to a model from another provider. Done means the child returns SUBAGENT_OK across providers, the original provider error surfaces on failure, and regression tests cover both same-provider and cross-provider delegation.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.

Mô tả

Summary

Problem

I'm trying to use two different model providers within Command Code:

  • Director: GPT-6.1 Sol using my existing ChatGPT/Codex subscription.
  • Subagent: MiniMax M3 using my existing Command Code GOAT subscription.

Both models work independently, but the Director cannot delegate tasks to the MiniMax subagent.

I'd like to request an official fix for cross-provider subagent delegation.

Environment

  • Command Code CLI: 1.88.2
  • Node.js: v24.21.0
  • Operating system: Ubuntu VPS
  • OpenAI provider: @byokkit/cmd-provider-openai 0.1.5
  • Subagent model: MiniMaxAI/MiniMax-M3

Steps to reproduce

  1. Connect ChatGPT/Codex and Command Code GOAT.
  2. Create a custom subagent with MiniMax M3 as its pinned model.
  3. Select GPT-6.1 Sol from the ChatGPT/Codex section as the main model.
  4. Ask Sol to delegate the task: "Reply exactly SUBAGENT_OK."
  5. The subagent immediately fails.
  6. Switch the main model to MiniMax M3 through Command Code and repeat. Delegation succeeds.

Actual result

The subagent fails before generating output or using tools.

The underlying error is:

HTTP 400 — Provider not available. Use Command Code provider.

Trace ID: 70c44b34dc2c73ba64de17906f21efaf

Command Code only displays: sub-agent stopped early: the run errored.

Investigation findings

Codex inspected the installed Command Code implementation.

It found that the subagent correctly selects MiniMax M3 but appears to reuse the Director's provider authentication and model transport.

This would explain why MiniMax → MiniMax works, while Sol/Codex → MiniMax/GOAT fails.

The custom-agent configuration also has no documented way to override the provider separately for the child.

These findings have not yet been confirmed by the Command Code maintainers.

Requested fix

Could you please investigate and fix this in an official Command Code release?

Ideally, each subagent should:

  • Use the provider and authentication associated with its own selected model.
  • Support delegation between ChatGPT/Codex and Command Code GOAT subscriptions.
  • Return clear error messages when execution fails.
  • Continue supporting existing single-provider configurations.

If this feature is intentionally unsupported, please consider adding it.

I would also appreciate confirmation of whether an official workaround already exists or a fix is planned.

Thank you!

Expected Behavior

The GPT-6.1 Sol Director, authenticated through my existing ChatGPT/Codex subscription, should be able to delegate tasks to a MiniMax M3 subagent using my Command Code GOAT subscription.

Each agent should use its own model provider and authentication. The MiniMax subagent should execute the task and return its result to the Director without requiring manual provider switching or an additional subscription.

Actual Behavior

Both models work independently, but cross-provider delegation fails before the subagent executes its first turn.

The underlying error is:

HTTP 400 — Provider not available. Use Command Code provider.

Command Code only displays:

sub-agent stopped early: the run errored

The child produces zero tokens and performs zero tool calls.

Investigation of the installed CLI indicates that the subagent selects the correct MiniMax model but inherits the Director's OpenAI/Codex provider authentication and transport.

MiniMax-to-MiniMax delegation works correctly.

Trace ID: 70c44b34dc2c73ba64de17906f21efaf

Steps to reproduce the issue
  1. Connect an existing ChatGPT/Codex subscription and a Command Code GOAT account using /connect.
  2. Create a custom subagent named sub-agent with model: MiniMaxAI/MiniMax-M3 and reasoningEffort: medium.
  3. Select GPT-6.1 Sol from the ChatGPT/Codex section as the main model.
  4. Ask the Director: "Delegate to sub-agent. Do not use tools. Reply exactly SUBAGENT_OK."
  5. Observe that the subagent fails immediately with a generic run error.
  6. Switch the main model to MiniMax M3 through Command Code GOAT.
  7. Repeat the same delegation. The subagent succeeds and returns SUBAGENT_OK.
Command Code Version

1.88.2

Operating System

Linux

Terminal/IDE

Remote SSH terminal (Ubuntu VPS)

Shell

bash

Session file (optional)

No response

Fix prompt (optional)

Fix cross-provider subagent delegation in Command Code.

The Director uses GPT-6.1 Sol through the OpenAI/Codex subscription, while the subagent uses MiniMax M3 through Command Code GOAT.

The current implementation appears to reuse the Director's provider authentication and model transport when launching a subagent, even when the child has a different pinned model.

Please investigate createHarnessSession, runSubAgent, provider resolution, and authentication routing.

Expected fix:

  • Resolve each subagent's provider and authentication independently.
  • Preserve the Director's existing provider and subscription.
  • Support direct Codex → GOAT delegation without a custom bridge.
  • Surface the original provider error when a subagent fails.
  • Preserve compatibility with existing single-provider delegation.

Add regression tests for both same-provider and cross-provider delegation.

Please implement this in an official release if technically supported, or clarify the intended limitation.

Additional context

No response

Ngôn ngữ chính
Không có dữ liệu ngôn ngữ
Star
4.2k
Fork
377
Merge trung bình
1 ngày 3 giờ
Pull request đã merge (30 ngày)
6

Chuẩn bị môi trường

Dự án này không cung cấp dev container, Dockerfile hay hướng dẫn đóng góp, nên bạn cần tự thiết lập môi trường: hãy bắt đầu từ README và xem hướng dẫn đóng góp lần đầu của chúng tôi để biết các bước chung.

Bắt đầu từ đâu

  1. Đọc hết issue, rồi đọc hướng dẫn đóng góp của dự án.
  2. Bình luận trên issue rằng bạn sẽ nhận — tránh hai người làm cùng một việc.
  3. Fork repository và làm thay đổi trên một nhánh.
  4. Mở pull request có tham chiếu số hiệu của issue.

Issue khác của CommandCodeAI/command-code

Tất cả issue của CommandCodeAI/command-code

Issue tương tự

Thêm issue về AI Infra & Agents

Nhận issue mới trong hộp thư của bạn

Bản tóm tắt ngắn những issue GitHub phù hợp với người mới.