pi-thinking-prune

Pi coding-agent extension that mirrors the pi-context-prune pattern, but targets assistant thinking blocks instead of toolResult messages.

It keeps the session JSONL file intact. Pruning only changes future LLM context through Pi's context event.

What it does

1. Detects assistant message content blocks where type === "thinking".
2. Stores the original full thinking text in a persistent local index.
3. Summarizes the thinking blocks with a configured model.
4. Appends a hidden thinking-prune-summary custom message that participates in LLM context.
5. Removes indexed raw thinking blocks from future context windows.
6. Registers thinking_tree_query so the agent can recover originals by short refs like th1.

Before pruning:

{
  "role": "assistant",
  "content": [
    { "type": "thinking", "thinking": "raw reasoning..." },
    { "type": "text", "text": "visible answer" }
  ]
}

Future context after pruning:

{
  "role": "assistant",
  "content": [
    { "type": "text", "text": "visible answer" }
  ]
}

The model also sees a hidden summary message with refs:

Summarized thinking refs: `th1`, `th2`
Use `thinking_tree_query` with these refs to retrieve the original full thinking blocks.

Install

Local validation:

pi install -l /path/to/pi-thinking-prune

From GitHub:

pi install -l git:github.com/ch0udry/pi-thinking-prune

Try without installing:

pi -e /path/to/pi-thinking-prune

Commands

/thinking-pruner settings
/thinking-pruner status
/thinking-pruner on
/thinking-pruner off
/thinking-pruner now
/thinking-pruner model [default|provider/model-id]
/thinking-pruner model provider/model-id:low
/thinking-pruner thinking [default|off|minimal|low|medium|high|xhigh]
/thinking-pruner min-raw-chars [1000]
/thinking-pruner prune-on [every-turn|on-demand|agent-message]
/thinking-pruner batching [turn|agent-message]
/thinking-pruner tree
/thinking-pruner stats
/thinking-pruner help

Tool

thinking_tree_query

Retrieves original thinking blocks that were summarized and pruned.

Input:

{
  "thinkingIds": ["th1", "th2"]
}

Configuration

Stored globally at:

~/.pi/agent/thinking-prune/settings.json

Default:

{
  "enabled": false,
  "showStatusLine": true,
  "summarizerModel": "default",
  "summarizerThinking": "default",
  "pruneOn": "agent-message",
  "batchingMode": "turn",
  "minRawCharsToPrune": 1000,
  "skipOversizedSummary": true
}

minRawCharsToPrune defaults to 1000, so small thinking blocks are ignored by default. Change it with /thinking-pruner min-raw-chars 1000, or from /thinking-pruner settings by opening Minimum raw chars and typing a number. skipOversizedSummary defaults to true, matching pi-context-prune: if a generated summary is larger than the raw thinking it would replace, the extension skips indexing/pruning that batch and advances a local thinking-prune-frontier so it is not summarized repeatedly.

Lifecycle behavior also mirrors pi-context-prune: pending batches are restored on summarizer/persistence failure, final-message auto flush uses session delivery, manual/runtime flush can inject hidden steer messages, and stale extension contexts return a stale-context result instead of corrupting state. enabled defaults to false for safety. Run:

/thinking-pruner on

Trigger modes

Mode	Meaning
agent-message	Summarize/prune when the agent sends a final assistant text response. Recommended default.
every-turn	Summarize/prune after every assistant message with thinking. Useful for testing.
on-demand	Never auto-prune. Use /thinking-pruner now.

Privacy note

Raw thinking is sent once to the configured summarizer model. Future main-model calls receive only the summary plus visible assistant text.

For local-only summarization, set:

{
  "summarizerModel": "local-provider/model-id"
}

Development

npm install
npm test
npm run check

Relation to pi-context-prune

pi-context-prune prunes entire toolResult messages.

pi-thinking-prune prunes only thinking blocks inside assistant messages and keeps assistant text/tool calls intact.