@siriuslan/discuss0.1.1

@siriuslan/discuss

A pi extension that runs a moderated, multi-round discussion between AI participants. Each participant is grounded in a project folder or defined by a role, and runs on the agent you choose: pi, Claude Code or Codex. You get one readable markdown record with a final summary and action items.

• Moderator = your current pi session. It picks who speaks next and writes the summary.
• Participants = separate agent sessions, one per participant. Each keeps its own memory for the whole discussion and is read-only.
• Record = discussion-docs/<id>.md. Every reply is appended under [name].

Commands

Command	What it does
/discuss start [topic]	Choose participants, create the doc, start moderating
/discuss continue [id | doc.md]	Resume a discussion, or turn an existing markdown doc into one
/discuss list	List discussions in the current folder
/discuss teams	View or delete saved teams
/discuss	Menu

Participants

Add any mix of:

• Project folder: the participant runs in that folder and answers from its real code and docs. The tag defaults to the folder name. A role instruction is optional.
• Role: a tag name plus an optional role instruction (e.g. "security reviewer, skeptical of new dependencies"). Press Enter to leave the instruction empty. Each role gets its own workspace folder, <cwd>/<YYYYMMDD-HHmm>-<topic>-<role-name>/. Put code or documents there if you want the role to use them.

Agents

Each participant has its own agent, model and thinking level, fixed for the whole discussion. Setup offers each installed agent with its defaults in one click, or Customize… to choose the model and thinking level.

Agent	Default	Read-only access	Web search (role participants)	Usage shown
pi	anthropic/claude-opus-5-5 · xhigh	tools read, grep, find, ls	not available	$ and tokens
Claude Code (claude)	opus · xhigh	tools Read, Grep, Glob	WebSearch, WebFetch	$ and tokens
Codex (codex)	gpt-5.6-terra · xhigh	read-only sandbox (shell commands can read, not write)	live web search	tokens only

• Web search is on for role participants whose agent supports it. It is off for project participants, who stick to their code.
• Agents whose CLI isn't installed are hidden in setup. Claude Code and Codex use your existing logins.
• Discussions and teams saved before agents were configurable run on pi.

Teams

At the end of setup you can save the participants as a team and load it next time. Teams are global: ~/.pi/agent/discussion/teams/<name>.json. Role participants get a fresh workspace in every discussion.

How a discussion runs

1. Participants are shown only what's new in the doc since their previous turn.
2. They @-mention each other (or @Human) to ask questions. A mention makes that participant active again.
3. A participant with nothing more to add replies I AM HAPPY FOR NOW!!! and is "settled".
4. When everyone is settled, the moderator appends ## Summary with decisions and action items.

Checkpoint: every N turns (default: participants × 3, set during setup) you are asked to continue, pause to add something, or wrap up.

Discussion boundaries

• Moderator: read only the shared discussion document, not participants' project folders or reference-material workspaces. Ask the owning participant to inspect their material and report findings in the discussion instead.
• Participants: read, list and search files only within their own working folder. Do not inspect other participants' folders, parent paths or symlinks leading outside that folder. The extension supplies the transcript; participants should not open it themselves.
• Questions: raise every question in the shared discussion, @-mentioning its owner or @Human. Ask for missing information rather than reading another participant's material.
• Public moderator messages: discuss_post accepts speaker: "moderator" for questions and comments under [Moderator]; it still defaults to [Human] for human input. A discuss_dispatch note is also recorded under [Moderator], addressed to its participant, before the turn starts. Notes remain in the discussion even if that turn fails.

These are agent instructions, not filesystem isolation. They are included in participant briefs and repeated every turn, including turns that resume an existing agent session.

Files

<cwd>/
├── discussion-docs/<id>.md          # the transcript
├── .pi/discussion/<id>.json         # state: participants, agents, sessions, turns, usage
└── <id>-<role-name>/                # one workspace per role participant

Requirements

• pi 1.3.0 or later (tested with 1.3.0).
• Optional: the Claude Code CLI (claude) and/or the Codex CLI (codex), installed on your PATH and logged in, for participants that run on them. Without them, every participant runs on pi.

Install

pi install npm:@siriuslan/discuss

To pin a version, or try it for a single session without installing:

pi install npm:@siriuslan/discuss@0.1.0
pi -e npm:@siriuslan/discuss

It also lives in the siriuslan/pi-packages repository; installing that repository from git loads this extension along with the other packages.

Security

This is an extension: pi runs its code with your user's permissions. It starts one agent process (pi, claude or codex) per participant turn, in the participant's folder. Participants are limited to read-only tools (Codex: its read-only sandbox, which still lets shell commands read files anywhere); role participants on Claude Code or Codex can search the web. The extension itself writes only the discussion files listed above and saved teams.

The folder boundaries above are prompt-level rules. The existing CLI permissions may still allow reads outside the working folder; these rules are not a security boundary for secrets.

Development

Load a local checkout in place, so edits apply after /reload:

pi install /path/to/pi-packages/packages/discuss

Or try it for one session: pi -e /path/to/pi-packages/packages/discuss.

Run the regression tests from packages/discuss after installing its peer dependencies:

npm install
npm test

The tests use a scripted UI, temporary files and a fake agent CLI; they don't run real agents or spend tokens.

License

MIT; see LICENSE.