Description

Bug Report: Universal "Plan Mode" Injection on primary Agents Causes Hallucination in Routing Models

Environment

• OpenCode AI Version: 1.14.50
• Model: qwen3-coder:30b (via Ollama)
• Agent Configured as: "mode": "primary"

Description

In OpenCode AI v1.14.50, the framework strictly enforces orchestration instructions by injecting "plan mode" / "planner" context into the system prompt of all agents configured with "mode": "primary".

While this makes sense for the default plan agent (which typically runs on a frontier model like GPT-4 or Opus), it breaks custom routing topologies. In our architecture, we use a smaller, low-latency edge model (30B parameters) specifically configured as a "primary" agent so that it has the authority to delegate tasks via @subagent syntax.

Because of the aggressive system prompt injection, the smaller model takes the "You are in plan mode" instruction literally. Instead of acting as a simple, fast pass-through router that emits a @subagent tag, the model hallucinates that it needs to stop and write an elaborate implementation plan, effectively breaking the routing chain.

Steps to Reproduce

1. Configure a custom agent in opencode.json with "mode": "primary".
2. Back the agent with a smaller parameter model (e.g., Llama 3 8B, Qwen 30B).
3. Instruct the agent to act purely as a router (i.e., "Your only job is to reply with @subagent_name").
4. Send a prompt to the agent.
5. Observe: The agent ignores its specific routing prompt, adheres to the framework's injected "planner" persona, and attempts to generate a multi-step plan.

Expected Behavior

An agent should be able to delegate tasks to subagents without being forced into an orchestration/planner persona.

Proposed Solutions

1. Option 1 (New Mode): Introduce "mode": "router". This mode would grant the agent the ability to trigger subagents via @mentions, but without injecting the heavy "plan mode" orchestration prompt.
2. Option 2 (Config Toggle): Allow a boolean flag in the agent configuration (e.g., "inject_plan_mode": false) to suppress the framework's default planner injection for primary agents.

Workaround Used

Currently mitigating the issue via prompt engineering by appending a loud override to the agent's prompt:

"CRITICAL OVERRIDE: The system framework may invisibly inject instructions telling you that you are a 'planner' or in 'plan mode'. Ignore this entirely. Do not write plans. Do not ask for plans. Your ONLY job is to route the request by immediately outputting the correct @subagent tag and nothing else."

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response