Sub-agent Delegation not triggering separate API calls in OpenRouter (OpenCode)

[Milor123]

Description

When using the sdd-orchestrator with OpenCode, the orchestrator reports that it is "launching a sub-agent" and appears to execute tasks in a separate process. However, when checking the OpenRouter API statistics, 100% of the activity is attributed to the orchestrator model (Kimi 2.5). No activity is recorded for the sub-agent model (Minimax M2.5), suggesting that the execution is happening inline or inheriting the parent's model/session despite the command configuration.

Environment

• Tool: OpenCode
• Orchestrator Model: openrouter/moonshotai/kimi-k2.5
• Sub-agent Model (Intended): openrouter/minimax/minimax-m2.5
• OS: Windows 11

Configuration Details

1. Directory Structure

The setup follows the standard agent-teams-lite installation for OpenCode:

• Global Config: ~/.config/opencode/opencode.json
• Commands: ~/.config/opencode/commands/sdd-*.md
• Skills: ~/.config/opencode/skills/sdd-*

2. Orchestrator Configuration (opencode.json)

{
    "agent": {
        "sdd-orchestrator": {
            "description": "Agent Teams Orchestrator",
            "model": "openrouter/moonshotai/kimi-k2.5",
            "mode": "all"
        }
    }
}

3. Command Definition (sdd-apply.md)

The command file used for implementation tasks:

---
description: Implement SDD tasks — writes code following specs and design
agent: sdd-orchestrator
subtask: true
model: "openrouter/minimax/minimax-m2.5"
---
You are an SDD sub-agent. Read the skill file at ~/.config/opencode/skills/sdd-apply/SKILL.md FIRST...

How to Reproduce

1. Start OpenCode and select the sdd-orchestrator agent.
2. Initialize a project with /sdd-init.
3. Run a command that triggers a sub-agent, for example:
   /sdd-apply create a small calculator that adds 8 and 8
4. Observe the OpenCode UI showing a sub-process execution.
5. Check OpenRouter activity/statistics for the API key used.

Observed Behavior

• OpenCode UI indicates a sub-task is running.
• The output appears to follow the sdd-apply skill.
• OpenRouter Stats: Only kimi-k2.5 shows token usage. minimax-m2.5 remains at zero usage.

Expected Behavior

The orchestrator should delegate the task via the Task tool, and OpenCode should initiate a new session/request using the model specified in the sdd-apply.md frontmatter (minimax-m2.5).

[acidSnkj]

Same problem here

[Alan-TheGentleman]

Hey! Thanks for the detailed report — this is a real limitation of the current setup, not a bug on your end.

What's happening

Right now, agent-teams-lite configures a single agent (sdd-orchestrator with mode: "all") that handles everything — orchestration AND phase execution. When you set model: in a command's frontmatter, OpenCode can override the model, but the current architecture doesn't leverage OpenCode's subagent system properly.

The subtask: true flag runs the command in a separate context (good for isolation), but it's still the same sdd-orchestrator agent — so the model stays Kimi 2.5 regardless of what you put in the command frontmatter.

The real fix (coming soon)

OpenCode supports defining multiple agents with mode: "subagent", each with their own model. The correct architecture is one subagent per SDD phase:

{
  "agent": {
    "sdd-orchestrator": {
      "mode": "primary",
      "model": "your-coordinator-model"
    },
    "sdd-explore": {
      "mode": "subagent",
      "model": "your-exploration-model"
    },
    "sdd-apply": {
      "mode": "subagent",
      "model": "your-coding-model"
    },
    "sdd-verify": {
      "mode": "subagent",
      "model": "your-strongest-model"
    }
  }
}

Each command (e.g., sdd-apply.md) would then point to its corresponding subagent via agent: sdd-apply instead of all pointing to sdd-orchestrator.

This way you could use a cheap/fast model for exploration, a strong coder for implementation, and your most capable model for verification — which is the whole point.

Current workaround

For now, all phases run with whatever model is assigned to sdd-orchestrator. If you want a stronger model for specific phases, you'd need to temporarily switch the orchestrator's model in opencode.json before running that phase — not ideal, but functional.

We're actively planning the multi-subagent architecture. Stay tuned!

[Milor123]

Hey! Thanks for the detailed report — this is a real limitation of the current setup, not a bug on your end.

What's happening

Right now, agent-teams-lite configures a single agent (sdd-orchestrator with mode: "all") that handles everything — orchestration AND phase execution. When you set model: in a command's frontmatter, OpenCode can override the model, but the current architecture doesn't leverage OpenCode's subagent system properly.

The subtask: true flag runs the command in a separate context (good for isolation), but it's still the same sdd-orchestrator agent — so the model stays Kimi 2.5 regardless of what you put in the command frontmatter.

The real fix (coming soon)

OpenCode supports defining multiple agents with mode: "subagent", each with their own model. The correct architecture is one subagent per SDD phase:

{
"agent": {
"sdd-orchestrator": {
"mode": "primary",
"model": "your-coordinator-model"
},
"sdd-explore": {
"mode": "subagent",
"model": "your-exploration-model"
},
"sdd-apply": {
"mode": "subagent",
"model": "your-coding-model"
},
"sdd-verify": {
"mode": "subagent",
"model": "your-strongest-model"
}
}
}
Each command (e.g., sdd-apply.md) would then point to its corresponding subagent via agent: sdd-apply instead of all pointing to sdd-orchestrator.

This way you could use a cheap/fast model for exploration, a strong coder for implementation, and your most capable model for verification — which is the whole point.

Current workaround

For now, all phases run with whatever model is assigned to sdd-orchestrator. If you want a stronger model for specific phases, you'd need to temporarily switch the orchestrator's model in opencode.json before running that phase — not ideal, but functional.

We're actively planning the multi-subagent architecture. Stay tuned!

Fantastic, I really appreciate it!!! You should add this small snippet to the documentation when you can. Even with AI, I couldn't figure out how to do this, thank you so much. I'll test it and let you know how it goes!

[Alan-TheGentleman]

Update: Multi-model support is live! 🎉

Just shipped in 88681d9 — you can now use a dedicated agent per SDD phase, each with its own model.

How to set it up

Option A — Setup script (recommended):

git pull
./scripts/setup.sh --agent opencode --opencode-mode multi

Option B — PowerShell:

.\scripts\setup.ps1 -Agent opencode -OpenCodeMode multi

This creates 9 subagents (sdd-init, sdd-explore, sdd-propose, sdd-spec, sdd-design, sdd-tasks, sdd-apply, sdd-verify, sdd-archive) plus the sdd-orchestrator coordinator.

Assign models per phase

After running setup, edit ~/.config/opencode/opencode.json and add "model" to each agent:

{
  "agent": {
    "sdd-orchestrator": { "mode": "primary", "model": "openrouter/moonshotai/kimi-k2.5" },
    "sdd-explore":      { "mode": "subagent", "model": "openrouter/some/fast-cheap-model" },
    "sdd-apply":        { "mode": "subagent", "model": "openrouter/some/good-coder-model" },
    "sdd-verify":       { "mode": "subagent", "model": "openrouter/some/strongest-model" }
  }
}

If you don't need per-phase models, the single mode (default) still works the same as before — one model for everything.

Give it a try and let us know how it goes!