Allows mid-generation steering of OpenCode sessions by intercepting prompt submissions during active turns and exposing them to the LLM via a checkpoint tool.
# Clone the repo
git clone https://github.com/matchaxnb/opencode-soft-steering.git
cd steering-vllm-implem
# Ensure the @opencode-ai/plugin dependency is available globally
cd ~/.config/opencode && bun add @opencode-ai/[email protected]
# Symlink the plugin into global plugins directory
mkdir -p ~/.config/opencode/plugins
ln -s "$(pwd)/.opencode/plugins/steering-checkpoint.js" ~/.config/opencode/plugins/# Clone the repo
git clone https://github.com/matchaxnb/opencode-soft-steering.git
cd steering-vllm-implem
# Copy the plugin and install dependency
mkdir -p ~/.config/opencode/plugins
cp .opencode/plugins/steering-checkpoint.js ~/.config/opencode/plugins/
cd ~/.config/opencode && bun add @opencode-ai/[email protected]The plugin:
- Injects a checkpoint protocol into the system prompt, instructing the LLM to call the
opencode_checkpointtool after every paragraph - Intercepts prompt submissions during active generation (
tui.prompt.submithook) and writes them to a temp file - Exposes
opencode_checkpointtool that reads the temp file, delivering any pending steering message back to the LLM
This enables "steering" — you can type feedback while the model is mid-response, and it picks it up at the next checkpoint call.