🧠 pi-auto-thinking — Automatic Thinking Level for Pi

@narumitw/pi-auto-thinking is a native Pi coding agent extension that selects Pi's thinking level for each user task from the active model capability and a deterministic task-difficulty score.

Use it when you want simple prompts to stay cheap and fast, while design, debugging, migration, security, or refactor tasks automatically get more reasoning budget.

✨ Features

• Automatically selects minimal, low, medium, high, or xhigh before each agent turn.
• Turns thinking off for models that do not advertise reasoning support.
• Respects model thinkingLevelMap entries that mark levels unsupported.
• Uses conservative defaults: enabled, minLevel: "minimal", maxLevel: "high".
• Avoids extra LLM calls; task difficulty is scored with local deterministic heuristics.
• Detects likely manual thinking-level changes and pauses automation for a few turns.
• Adds /auto-thinking commands for status, enable/disable, and explaining the last decision.
• Shows a compact statusline item after decisions without notifying every turn.

📦 Install

pi install npm:@narumitw/pi-auto-thinking

Try without installing permanently:

pi -e npm:@narumitw/pi-auto-thinking

Try this package locally from the repository root:

pi -e ./extensions/pi-auto-thinking

🚀 Usage

/auto-thinking status
/auto-thinking on
/auto-thinking off
/auto-thinking explain

• status shows whether automation is enabled, the loaded config path, bounds, and warnings.
• on enables automatic thinking selection for subsequent user prompts.
• off disables automation and clears the extension statusline item.
• explain shows the score, selected level, and matched signals from the last decision.

🧮 How scoring works

The extension scores each prompt locally before Pi starts the agent loop. The base level mapping is:

| Score | Base thinking level | | ---: | --- | | <= 0 | minimal | | 1..2 | low | | 3..5 | medium | | 6..8 | high | | >= 9 | xhigh |

Default weights:

| Signal | Weight | | --- | ---: | | Quick, brief, concise, simple, or equivalent Chinese wording | -2 | | Simple explanation, translation, or summary | -1 | | Implementation, editing, or code-change request | +2 | | Debugging, exception, stack trace, failure, or broken behavior | +2 | | Tests, lint, typecheck, TypeScript, CI, or similar | +1 | | Design, architecture, plan, proposal, or tradeoff work | +3 | | Migration, refactor, compatibility, or breaking-change work | +3 | | Security, auth, secrets, concurrency, transaction, rollback, or data-loss topic | +3 | | Explicit deep reasoning request such as “think hard”, “carefully”, “深入”, or “仔細” | +3 | | Code block included | +1 | | One file path mentioned | +1 | | Three or more file paths mentioned | +2 | | Long prompt | +1 | | Image input attached | +1 |

After scoring, config bounds and model support are applied. For example, with the default maxLevel: "high", a score that maps to xhigh is capped at high.

⚙️ Configuration

Optional config file:

$PI_CODING_AGENT_DIR/pi-auto-thinking.json

When PI_CODING_AGENT_DIR is unset, the extension reads:

~/.pi/agent/pi-auto-thinking.json

Example:

{
  "enabled": true,
  "minLevel": "minimal",
  "maxLevel": "high",
  "respectManualTurns": 3,
  "modelOverrides": {
    "anthropic/claude-haiku-4-5": {
      "maxLevel": "medium"
    },
    "anthropic/claude-opus-4-5": {
      "maxLevel": "xhigh"
    },
    "local/small-fast-model": {
      "enabled": false
    }
  }
}

Supported levels are:

off, minimal, low, medium, high, xhigh

respectManualTurns must be an integer from 0 to 20. When a thinking level changes outside this extension, automation pauses for that many future user prompts. Pi's event does not expose the change source, so this detection is intentionally conservative.

🤖 Model capability behavior

• If ctx.model is unavailable, the extension selects off.
• If ctx.model.reasoning is false, the extension selects off.
• If ctx.model.thinkingLevelMap marks a level as null, the extension skips that level.
• Pi still performs its own final clamp when pi.setThinkingLevel() is called.

💸 Cost and latency caveats

Higher thinking levels can increase latency and token usage depending on the provider. The default maximum is high, not xhigh, to avoid unexpectedly expensive turns. Use /auto-thinking explain to inspect why a level was selected, or /auto-thinking off to disable automation for the session.

🗂️ Package layout

extensions/pi-auto-thinking/
├── src/
│   └── auto-thinking.ts
├── README.md
├── LICENSE
├── tsconfig.json
└── package.json

The package exposes its Pi extension through package.json:

{
  "pi": {
    "extensions": ["./src/auto-thinking.ts"]
  }
}

🔎 Keywords

Pi extension, Pi coding agent, automatic thinking, reasoning level, task difficulty, TypeScript Pi package, npm Pi extension.

📄 License

MIT. See LICENSE.