kagi

🏭 The No-Shop

"Ixian machines are cunning devices. They do not think, but they do something better — they observe."

A Claude Code plugin marketplace. Ixian technology for those who prefer their tools sharp, their feedback loops closed, and their agent sessions rendered in something other than a scrolling terminal.

---

🔮 Plugins

⚡ Foldspace Console

Your Claude Code sessions, rendered beautifully in foldspace.

Foldspace Console captures every assistant response and streams it to a rich browser viewer in real-time. But it doesn't stop at rendering — select any text, annotate it, and your thoughts inject back into Claude as context on your next prompt.

The loop is closed. The machine observes. The navigator annotates. The path folds.

✨ Features

• 📡 Live streaming — responses appear in the browser the instant Claude finishes speaking
• 🎨 Rich markdown — syntax-highlighted code blocks with copy buttons, tables, blockquotes, the works
• ✍️ Annotations — select text, add your thoughts, they become context on your next prompt
• 🗂️ Multi-session tabs — every Claude Code session gets its own tab, complete with metadata header
• 📊 System telemetry — all-time stats, installed plugins, hooks, MCP servers, model token usage
• 🧬 Session metadata — model, permission mode, git branch, token counts, turn counter
• 🔄 Auto-reconnect — WebSocket reconnects seamlessly if the daemon restarts
• 🚀 Zero config — daemon auto-starts on your first session. Just open the browser.

🏗️ Architecture

┌──────────────┐     Stop Hook      ┌─────────────┐   WebSocket    ┌──────────────────┐
│  Claude Code  │ ────────────────▶ │   Daemon     │ ─────────────▶│  Browser SPA     │
│  (terminal)   │                   │  (Bun HTTP)  │               │  (Foldspace UI)  │
│               │ ◀──────────────── │  :3377       │ ◀─────────────│                  │
└──────────────┘   Prompt Hook      └─────────────┘   Annotate     └──────────────────┘
       │                                  │
       │          SessionStart            │
       └──────────────────────────────────┘

Four hooks. One daemon. One browser tab. That's the whole machine.

Hook	Event	Purpose
session-start.sh	SessionStart	Registers session, auto-starts daemon
stop.sh	Stop	Captures last_assistant_message, POSTs to daemon
prompt.sh	UserPromptSubmit	Fetches annotations, injects as context
session-end.sh	SessionEnd	Final transcript scan for token totals

---

🔄 The Feedback Loop: How Annotations Become Context

This is the core trick — the part where passive observation becomes active navigation.

When you select text in the Foldspace browser SPA and add an annotation, it gets stored in the daemon's in-memory annotation queue (keyed by session ID). Nothing happens yet. The annotations just... wait. Patient as a Face Dancer in a crowd.

Then you type your next prompt in the terminal.

The Injection Mechanism

Claude Code fires the UserPromptSubmit hook before the prompt reaches the model. Here's what prompt.sh does:

1. Reads the hook event from stdin (Claude Code passes session_id and other metadata as JSON)
2. Health-checks the daemon — if the daemon isn't running, the hook exits silently. No crash, no noise.
3. Fetches pending annotations via GET /api/annotations?session={SESSION_ID}
4. If annotations exist, formats them as plain text to stdout
5. Deletes the annotations from the daemon via DELETE /api/annotations?session={SESSION_ID} — they are consumed on injection

Claude Code's hook system takes whatever the UserPromptSubmit hook writes to stdout and prepends it as additional context to the user's prompt. The model sees the annotations alongside (and before) whatever you actually typed.

What the Model Sees

The injected context looks like this:

--- Annotations from Foldspace Console ---
IMPORTANT: The user has left annotations from their Foldspace Console (browser viewer). Acknowledge that you received these annotations before proceeding with your response.

[1] On text:
> The selected text from Claude's response,
> with each line blockquoted

Comment: Your annotation comment here

[2] On text:
> Another selection across
> multiple lines

Comment: And your thoughts on this one

--- End annotations ---

Multiple annotations are numbered sequentially. The original selected text is blockquoted with > prefixes. Each annotation carries the exact text you highlighted and the comment you wrote.

Visibility

Two safeguards ensure the injection isn't a ghost operation: