pi-brainstorm

Fork of @paulmupeters/pi-brainstorm with expanded read-only tool support.

A pi extension that adds a read-only /brainstorm mode with dynamic tool detection. During brainstorm, the agent can use read, find, and any of the following optional tools if registered in the environment:

• grep — for pattern search within files
• ask_user_question — for structured clarifying questions
• subagent / get_subagent_result — for deep codebase exploration

What it does

When brainstorm mode is active:

• allows read, find and any detected optional tools (grep, ask_user_question, subagent, get_subagent_result)
• blocks shell commands and file edits/writes
• keeps the conversation exploratory
• avoids unsolicited "you should do X next" suggestions
• gives a clear recommendation when you ask for the best option
• shows a visible reminder in the UI
• drafts a decision-oriented markdown brief when you finish
• can replace the brainstorm transcript with the reviewed brief in LLM context when you finish without saving or when you save and choose the context-preserving option

Key difference from upstream

The original extension blocks all non-read tools during brainstorm mode. This fork:

• dynamically detects which tools are available in the current pi environment (grep, ask_user_question, subagent, get_subagent_result)
• allows all detected tools alongside read and find, giving the agent richer capabilities during brainstorm (e.g. searching code, asking structured questions, exploring codebases in background)
• silently omits tools that are not registered — no configuration needed
• shows a dynamic allowed-tool list in the system prompt so the agent knows exactly what it can use

UX

• /brainstorm starts brainstorm mode
• /brainstorm again opens a small menu:
  • Continue brainstorming
  • Finish and summarize
  • Cancel and discard
• /brainstorm finish finishes directly
• /brainstorm cancel exits immediately without a summary
• /brainstorm-summary-model configures an optional dedicated summary model
• Ctrl+Alt+B is a shortcut for the same flow

While active, the footer/widget reminds you how to finish or cancel.

Install / test

Quick test

pi --no-extensions -e /path/to/brainstorm.ts

Use from your normal pi setup

Install from npm:

pi install npm:@aalalice233/pi-brainstorm

Or copy extensions/brainstorm.ts into ~/.pi/agent/extensions/.

Optional: enable additional brainstorm tools

The extension detects these optional tools at runtime — install any of the packages below and the tool is automatically available in brainstorm mode:

Tool	Package
grep	Built-in to pi, no extra install needed
ask_user_question	@juicesharp/rpiv-ask-user-question
subagent / get_subagent_result	Built-in to pi, no extra install needed

For example:

pi install npm:@juicesharp/rpiv-ask-user-question

All tool detection is automatic — no config needed.

Brief export

When you finish a brainstorm, the extension:

1. collects the conversation since brainstorm mode started
2. asks the current model, or an optional dedicated summary-model override, to draft a concise decision brief
3. opens that brief in an editor so you can tweak it
4. then offers:
   • Brief to context
   • Brief to markdown
   • Brief to markdown and context
   • Continue brainstorming
   • Exit

The generated brief uses # Decision Brief: <topic> when the session reached a clear decision, recommendation, or strong preference. If no firm conclusion emerged, it uses # Brainstorm Brief: <topic> and calls out the strongest current leaning without inventing certainty. It leads with the recommendation/current leaning, then covers rationale, alternatives, risks/open questions, and a transcript summary capped at 5 sentences.

Default save path:

brainstorms/YYYY-MM-DD-topic.md

Optional summary model override

By default, brainstorm summaries use the currently active Pi model.

You can persist a separate user-level summary model with:

/brainstorm-summary-model

Or set it directly:

/brainstorm-summary-model google/gemini-2.5-flash
/brainstorm-summary-model clear

The preference is stored globally in ~/.pi/agent/settings.json under piBrainstorm.summaryModel and falls back to the active model when unset or unavailable.

If you choose Brief to context or Brief to markdown and context, the brainstorm transcript stays in session history, but future LLM context uses the reviewed brief instead of the full brainstorm exchange.

Notes

• During brainstorm mode, read and find are always enabled; grep, ask_user_question, subagent, and get_subagent_result are enabled when their respective packages/tools are registered.
• The extension restores your previously active tools after finishing/canceling.
• If model-based brief generation is unavailable, the extension falls back to a simple markdown transcript.
• Tool detection happens at runtime — no config or restart needed when installing/removing packages or when tools become available/unavailable.