pi-auto-reasoning-tool

A small Pi package that lets agents adjust their own reasoning level with a change_reasoning tool.

The tool is intentionally minimal: the agent directly chooses low, medium, or high, and the extension calls Pi's setThinkingLevel API. No scoring rubric, no UI, no commands.

After every successful/non-retryable agent run, the extension restores reasoning to the session's starting level. Retryable transport/provider failures keep the current level so Pi's auto-retry does not accidentally drop the agent back to the baseline.

Why

Pi already supports reasoning levels. This package exposes a narrow agent-callable tool so the agent can raise or lower its budget when a prompt or in-progress discovery actually needs it, while returning to the user's starting level afterward.

The prompt is conservative: agents are told to use the tool sparingly and preferably in parallel with other useful tool calls, avoiding standalone reasoning-change turns.

Install

pi install npm:@howaboua/pi-auto-reasoning-tool

Or add it to your Pi settings:

{
  "packages": ["npm:@howaboua/pi-auto-reasoning-tool"]
}

Then restart Pi or run /reload.

Tool

Registers one tool:

change_reasoning

Parameters:

level: low | medium | high

Behavior:

1. Agent chooses level.
2. Extension calls pi.setThinkingLevel(level).
3. Tool result reports the previous and applied level.
4. If Pi clamps the requested level because of model capability, the result says so.
5. On the first agent_start, the extension captures the session's starting reasoning level.
6. On agent_end, the extension restores reasoning to that starting level unless the last assistant message is a retryable error.

Agent-facing prompt copy

Description:

Set your reasoning level for the current task.

Prompt snippet:

Set reasoning level sparingly: low default/simple/back-and-forth/cleanup; medium complex single task or feature planning; high multiple tasks, architecture-spanning work, or unexpectedly hard issues.

Guidelines:

Use change_reasoning sparingly; it is often unnecessary because low is the default operating mode.

Prefer calling change_reasoning in parallel with other useful tool calls so you do not waste a turn only changing reasoning.

Use low for single simple tasks, back-and-forth conversations, or simple cleanup after harder tasks.

Use medium for complex single tasks or planning features.

Use high for handling multiple tasks in one turn, work spanning different architecture elements, or unexpected hard-to-solve issues during a turn.

Parameter:

level

Reasoning level to use for this task: low, medium, or high.

Development

npm install
npm run check
pi -e ./src/index.ts

Publish

npm publish --access public

License

MIT