@oai404iao/pi-external-thinking

Replace compatible Pi models' native reasoning with a visible Think scratchpad tool.

Compatibility: Pi 0.84.2 or newer; tested against 0.84.2.

npm identity: @oai404iao/pi-external-thinking. This package is approved for npm bootstrap but is not published yet. To load it for one run from a local checkout:

pi -e ./pi-extensions/external-thinking

To install that local package persistently, use:

pi install ./pi-extensions/external-thinking

After initial npm bootstrap, install it with:

pi install npm:@oai404iao/pi-external-thinking

What it does

1. Registers a package-specific tool labelled Think whose arguments are rendered as dim, italic, visible reasoning in the Pi TUI. Do not put secrets or user-sensitive data in thoughts.
2. Sets Pi's thinking level to off, but only after checking that the active model actually supports disabling native reasoning.
3. In hard mode (the default), forces the first provider request of every user turn to call Think. In soft mode, the tool remains available but the model decides whether to call it.

The tool has no promptSnippet, so it is omitted from Pi's system-prompt tool list. It is still sent to the provider and its visible description tells the model that the scratchpad is shown to the user.

Usage

Action	Command
Toggle	/external-thinking
Enable / disable	/external-thinking on / /external-thinking off
Enable in a specific mode	/external-thinking on hard / /external-thinking on soft
Switch mode (persists)	/external-thinking mode hard / /external-thinking mode soft
Show state	/external-thinking status
Enable at startup	pi --external-thinking

State persists in <agentDir>/external-thinking.json (normally ~/.pi/agent/ or ~/.config/pi/agent/). The prior thinking level is restored when you turn the feature off.

Hard vs. soft mode

• Hard: Pins tool_choice to the package-specific Think tool for the first request of a user turn. If the provider payload does not contain a writable extension-tool definition, the extension pauses itself and restores native-thinking settings rather than falsely claiming enforcement.
• Soft: Does not change tool_choice. The model may call Think, but is not required to do so.

Both modes require a provider API that the extension recognizes and a model for which Pi reports the off thinking level as supported.

Supported models

The extension supports these Pi provider APIs:

• openai-responses
• azure-openai-responses
• openai-completions
• anthropic-messages
• google-generative-ai
• google-vertex

It refuses to enable for unsupported APIs, models that cannot disable native reasoning (thinkingLevelMap.off === null), sessions where the extension tool is excluded by --tools or user tool settings, and openai-codex-responses models whose protocol cannot provide the required reasoning-off guarantee.

If a persisted/flag-enabled session later selects an incompatible model, the extension pauses, restores the prior thinking level, and resumes only after a compatible model is selected.

Upstream attribution

This is a modified port of oh-my-pi's externalThinking feature. The exact source revisions and local modification scope are recorded in THIRD_PARTY_NOTICES.md.

License

This modified port is MIT-licensed and preserves the upstream oh-my-pi copyright notices. See LICENSE and THIRD_PARTY_NOTICES.md.