igorwarzocha__opencode-workflows/plugins/gemini-glm-focused-mode

Gemini/GLM Focused Mode

Injects a rigorous system prompt for GLM and Gemini models to enforce focused, deterministic coding behavior.

What It Does

Prepends a strict execution framework to improve:

• Precision: Execute immediately, output first, rationale second
• Grounding: Verify claims with tools before stating them
• Persistence: Try 3 approaches before escalating to user
• Completeness: Only report done when verified working

The injected prompt enforces concise responses, complete implementations, and proper error recovery.

Targeting

Activates based on model name matching:

• Matches any model ID containing glm-4.7
• Matches any model ID containing gemini

Other models are unaffected.

Installation

Add the plugin to your global opencode.json plugin array:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "@howaboua/opencode-glm-gemini-prompt-enhancer"
  ]
}

Adjust the path to match your actual installation location.

Bonus: Provider Configuration

The opencode.json file included here is my personal setup. It configures GLM models to use the Anthropic adapter endpoint, which I've found to be faster and more reliable with tool calls - the model just seems smarter through this adapter.

Quirks I've run into:

• Token counting is broken, so you need to manually control session lenght compaction (best used for subagents with specific tasks)
• If you stay within OpenCode's normal limits (max output is always 32k, so the reasonable thinking budget would be 24k), everything's compliant but GLM won't attempt hard thinking
• It could theoretically overextend into deep thinking modes and cause issues, but that's never happened to me

The default thinking budget is set to 64000 tokens since I use GLM for my general, compaction, and explore subagents where I prefer thinking enabled. The issue here is that you can't use variants in this part of the config. You define models for these agents in your opencode.json file, e.g.:

    "plan": { "model": "local/antigravity/gemini-3-pro" },
    "general": {
      "model": "zai-coding-plan/glm-4.7"
    },
    "explore": {
      "model": "zai-coding-plan/glm-4.7"
    },
    "compaction": {
      "model": "zai-coding-plan/glm-4.7"
    }

Your mileage may vary, but I've found this tradeoff worth it. Adapt the config to your own risk tolerance.