Skip to content

README.md

Latest commit

 

History

History
102 lines (68 loc) · 4.1 KB

README.md

File metadata and controls

102 lines (68 loc) · 4.1 KB

Dynamic Context Pruning Plugin

npm version

Automatically reduces token usage in OpenCode by removing obsolete tool outputs from conversation history.

DCP in action

Installation

Add to your OpenCode config:

// opencode.jsonc
{
  "plugin": ["@tarquinen/opencode-dcp@0.3.30"]
}

When a new version is available, DCP will show a toast notification. Update by changing the version number in your config.

Restart OpenCode. The plugin will automatically start optimizing your sessions.

Pruning Strategies

DCP implements two complementary strategies:

Deduplication — Fast, zero-cost pruning that identifies repeated tool calls (e.g., reading the same file multiple times) and keeps only the most recent output. Runs instantly with no LLM calls.

AI Analysis — Uses a language model to semantically analyze conversation context and identify tool outputs that are no longer relevant to the current task. More thorough but incurs LLM cost.

Context Pruning Tool

When strategies.onTool is enabled, DCP exposes a prune tool to Opencode that the AI can call to trigger pruning on demand.

When nudge_freq is enabled, injects reminders (every nudge_freq tool results) prompting the AI to consider pruning when appropriate.

How It Works

Your session history is never modified. DCP replaces pruned outputs with a placeholder before sending requests to your LLM.

Impact on Prompt Caching

LLM providers like Anthropic and OpenAI cache prompts based on exact prefix matching. When DCP prunes a tool output, it changes the message content, which invalidates cached prefixes from that point forward.

Trade-off: You lose some cache read benefits but gain larger token savings from reduced context size. In most cases, the token savings outweigh the cache miss cost—especially in long sessions where context bloat becomes significant.

Configuration

DCP uses its own config file (~/.config/opencode/dcp.jsonc or .opencode/dcp.jsonc), created automatically on first run.

Options

Option Default Description
enabled true Enable/disable the plugin
debug false Log to ~/.config/opencode/logs/dcp/
model (session) Model for analysis (e.g., "anthropic/claude-haiku-4-5")
showModelErrorToasts true Show notifications on model fallback
strictModelSelection false Only run AI analysis with session or configured model (disables fallback models)
pruning_summary "detailed" "off", "minimal", or "detailed"
nudge_freq 10 How often to remind AI to prune (lower = more frequent)
protectedTools ["task", "todowrite", "todoread", "prune"] Tools that are never pruned
strategies.onIdle ["deduplication", "ai-analysis"] Strategies for automatic pruning
strategies.onTool ["deduplication", "ai-analysis"] Strategies when AI calls prune

Strategies: "deduplication" (fast, zero LLM cost) and "ai-analysis" (maximum savings). Empty array disables that trigger.

{
  "enabled": true,
  "strategies": {
    "onIdle": ["deduplication", "ai-analysis"],
    "onTool": ["deduplication", "ai-analysis"]
  },
  "protectedTools": ["task", "todowrite", "todoread", "prune"]
}

Config Precedence

Settings are merged in order: Defaults → Global (~/.config/opencode/dcp.jsonc) → Project (.opencode/dcp.jsonc). Each level overrides the previous, so project settings take priority over global, which takes priority over defaults.

Restart OpenCode after making config changes.

Subagents

DCP automatically skips processing for subagent sessions (general, explore, etc.), but subagents can still invoke the prune tool. To prevent this, disable the tool in your OpenCode config. Any custom agents you've defined should also have prune disabled:

// opencode.jsonc
{
  "agent": {
    "general": { "tools": { "prune": false } },
    "explore": { "tools": { "prune": false } }
  }
}

License

MIT