pi-adaptive-thinking

Bring adaptive reasoning-effort control to Pi agents.

pi-adaptive-thinking is a Pi extension that lets the agent inspect and change Pi's thinking level through tools named get_thinking_level and set_thinking_level.

Requires Node >=22.19.0 and Pi >=0.84.1.

It mirrors the user-facing behavior of opencode-adaptive-thinking while using Pi-native APIs: pi.getThinkingLevel(), pi.setThinkingLevel(), and thinking_level_select.

Installation

pi install npm:pi-adaptive-thinking

For local development:

pnpm install
pi -e ./src/index.ts

Behavior

The extension registers two tools. get_thinking_level returns the current native Pi thinking level and the levels supported by the selected model. The agent should inspect status only when the level is uncertain, not poll it every turn.

set_thinking_level accepts these parameters:

• level: one of the valid Pi thinking levels for the current model.
• persist: optional boolean, default false.

Temporary changes:

{ "level": "high", "persist": false }

This changes thinking level for the current agent turn and restores the prior/baseline level when the turn ends.

Persistent changes:

{ "level": "low", "persist": true }

This changes the session baseline until another persistent change is made or the user changes thinking level manually.

The extension contributes only static tool guidance to Pi's base prompt. It never replaces the per-turn system prompt, so current level, supported levels, model, and session state cannot invalidate the system-prompt cache prefix.

Configuration

Configuration files are loaded in this order:

1. Project: .pi/adaptive-thinking.json
2. Global: ~/.pi/agent/adaptive-thinking.json
3. Built-in defaults

Project configuration takes precedence over global configuration.

{
  "enabled": true,
  "quiet": false,
  "toolName": "set_thinking_level",
  "toolDescription": "Set your thinking level",
  "statusToolName": "get_thinking_level",
  "guidance": "You MUST manage thinking level actively. Lower it before trivial or routine turns; raise it for ambiguity, debugging, risky changes, or multi-step synthesis. Reassess at turn start, after meaningful new evidence, and when the task shifts. NEVER leave the current level unchanged by inertia, and NEVER reply to a trivial turn before considering a downshift."
}

guidance, tool names, and tool descriptions are loaded once per session and remain static. Do not put runtime state in guidance.

The former systemPrompt field remains a deprecated alias for guidance during the 0.x release line. Existing configurations continue to work and produce one UI warning at session start unless quiet is enabled. A configuration containing both fields is invalid.

Development

pnpm install
pnpm format:check
pnpm lint
pnpm typecheck
pnpm test

Run all checks:

pnpm verify