Pickle-Thinker

OpenCode plugin for GLM-4.6 / Big Pickle that auto-adds a steering prefix—set it to “Ultrathink:” or... any other reminder you want.

Two modes (default: tool):

• lite: prefix the latest user message only.
• tool: after every tool call, inject a follow-up prompt; uses two variants (normal vs. failure-heuristic) for post-tool steering.

Config lives in ~/.config/opencode/pickle-thinker.jsonc (auto-created) or .opencode/pickle-thinker.jsonc per project:

{
  // mode: "lite" keeps single-prefix behavior.
  // mode: "tool" adds an extra user turn after each tool result (more turns/tokens).
  "enabled": true,
  "mode": "tool",
  "prefix": "Ultrathink: ",
  "debug": false,
}

Heads-up: tool mode increases turns/tokens and may affect subscription usage.

Steering hacks

• Set prefix to anything you like, e.g. prefix: "Responde solo en español: " to force language, or prefix: "Read the docs before answering: " to nudge behavior. You can drop the word “Ultrathink” entirely if you want. This behaves like a lightweight user-prompt hook (similar to Claude Code’s UserPromptSubmit), letting you bake reminders or constraints into every turn.

Installation

Option 1: npm (when published)

Add to your repository opencode.json or user-level ~/.config/opencode/opencode.json:

{
  "plugin": ["@howaboua/[email protected]"]
}

Option 2: Manual folder install (for testing)

1. Build the plugin: bun run build
2. Copy the plugin folder to your target project:

# Option A: Install to specific project
cp -r /path/to/pickle-thinker/dist /your/project/.opencode/plugins/pickle-thinker

# Option B: Install to user-level plugins
mkdir -p ~/.config/opencode/plugins/pickle-thinker
cp -r /path/to/pickle-thinker/dist/* ~/.config/opencode/plugins/pickle-thinker/

3. Add to your opencode.json:

{
  "plugin": ["./.opencode/plugins/pickle-thinker"]
}

or for user-level:

{
  "plugin": ["~/.config/opencode/plugins/pickle-thinker"]
}

Option 3: Git submodule (for development)

cd your-project
git submodule add https://github.com/howaboua/pickle-thinker .opencode/plugins/pickle-thinker
cd .opencode/plugins/pickle-thinker
bun run build

Then add to opencode.json:

{
  "plugin": ["./.opencode/plugins/pickle-thinker"]
}

How It Works

Hybrid Architecture: Uses dual-approach injection for maximum compatibility:

• Fetch wrapper: Intercepts API calls before they reach the model (master approach)
• OpenCode hooks: Transforms messages in the OpenCode pipeline (interleaved approach)

Model filtering: Only runs for glm-4.6 and big-pickle models.

• Only runs for glm-4.6 and big-pickle.
• Lite mode: prepend prefix to the most recent user message.
• Tool mode: insert prefix after each tool output; if the tool output looks like an error, swap to a “failed” prompt.

Examples

• Lite mode (prefix only):
  User: Ultrathink: Explain quantum computing

• Tool mode (auto-injected after a tool call):
  Ultrathink: Analyze the tool output and continue.

• Tool mode (failure heuristic fired):
  Ultrathink: Tool output failed. Consider re-running the tool or re-reading the file before editing it.

Acknowledgments

This plugin was created using the Dynamic Context Pruning plugin by Dan Tarquinen as a template. The DCP plugin provided the excellent fetch wrapper architecture and API format handling that made this implementation possible.