Keystone

A desktop review inbox for AI-generated Markdown.

AI agents (Claude Code, Codex, scripts, …) drop Markdown artifacts — plans, designs, reports — into a watched folder. Keystone surfaces them as a review queue: open, read a nicely rendered version, annotate, and your feedback flows back to disk as a .feedback.md file next to the original. Agents watch for that feedback and act on it, closing the human-in-the-loop.

The keystone is the piece that locks the arch — the review that makes the rest hold.

How it works

agent writes  plan.md            ─┐
                                  ├─►  Keystone shows it in the review queue
you review + annotate            ─┘
                                       │  writes
                                       ▼
              plan.feedback.md   ─►  agent picks up  status: pending
                                       │  acts on it, sets status: consumed
                                       ▼
              (loop continues)

• Folder-based, tool-agnostic. Any agent that can write a file participates. No API, no per-agent integration.
• Plain files. Originals and feedback are Markdown on disk — git-friendly, inspectable, no database.
• Feedback loop. A companion agent skill (Claude / Codex) polls for status: pending feedback and applies it.

See docs/specs/feedback-format.md for the file and feedback format, and docs/ for the full documentation set.

Stack

• Tauri — small, native shell (window, tray, notifications, fs-watch)
• React + Tailwind — the review UI
• Markdown rendering in-app

Status

Early scaffolding. Roadmap:

• v0 — watch a folder, list queue, render Markdown, write .feedback.md (feedback fed back to the agent manually).
• v1 — inline / section-level annotations, desktop notifications on new artifacts, and an agent trigger (skill loop that consumes pending feedback).

License

TBD