Skip to content

AGENTS.md

Latest commit

 

History

History
66 lines (54 loc) · 3.37 KB

AGENTS.md

File metadata and controls

66 lines (54 loc) · 3.37 KB
Raw

Pantheon

A desktop cockpit for running several AI coding agents side by side, coordinated rather than siloed. Each agent gets a live terminal pane; panes can share a context store, and one can be promoted to conductor to fan work out and collect results.

Runs on Windows and Linux: the terminal layer is portable-pty, which is ConPTY on Windows and a Unix PTY elsewhere. macOS is unexercised rather than ruled out. Status is working prototype, so read the Known gaps section of README.md before relying on any part of it.

This file is orientation for this project only. The shared working-style rules, the guarded-change policy, and the greenfield stack defaults live in ~/Projects/global-llm-configs/AGENTS.md, which every agent host loads. They are not restated here on purpose; if they seem to be missing, the host's symlink to that file is broken, so fix the symlink rather than copying the rules back in.

Where things are

  • src/ : the React and TypeScript front end, built with Vite.
  • src-tauri/src/ : the Rust side. mcp.rs is the in-process MCP server the agents talk to, worktree.rs is the git worktree isolation, lib.rs and main.rs are the app.
  • ui-gallery/ : component previews.
  • README.md : what Pantheon is and what it cannot yet do.
  • CONTRIBUTING.md : setup, building, testing, branching, review, and pull requests. It is the authority on all of those; do not duplicate it here.
  • BACKLOG.md : the reasoning behind the open work. Entries record what was measured, which hypotheses were refuted, and what was tried and did not work. That is the part worth keeping, and it is why it stays prose.

Build with pnpm. dev.cmd and build.cmd on Windows, or dev.sh and build.sh on Linux, wrap the usual commands.

Tasks

Before picking up Orion-scoped work, read the canonical outcome order: docs/orion.md in G-Hensley/projects (or ~/Projects/docs/orion.md on this machine's primary checkout). It states which accepted outcome is next across Lexicon, Daedalus, Crucible, Pantheon and ai-kbase; this repository's own issues below remain authoritative for that outcome's scope, acceptance and status.

Read docs/github-issues.md before picking up work. It records a staged handoff, so first verify its activation checkpoint:

  • Before verified activation, .tasks/ remains authoritative. Use lexicon task for an authorized update, except while the conductor has explicitly paused legacy writers for the handoff.
  • After verified activation, GitHub Issues owns deliverable identity, acceptance and resolution, and the repository's selected GitHub Project owns stage and priority. Never run lexicon task or edit .tasks/; the files become a read-only historical ledger.
  • A staged issue or Ready field is not permission to start. Read its full body, dependencies, assignee, comments and latest conductor/session claim.

docs/github-tracking-migration.md maps every reviewed source identity to its owning issue or historical disposition. It is provenance, not another tracker: the active issue wins if the map and issue later disagree. Reasoning stays in BACKLOG.md; an issue links to it rather than restating it.

This manual tracker handoff installs no MCP tool, dispatch behavior, or runtime change. Pantheon's own dispatch/session journal (brain.jsonl) is unrelated and untouched by it.