Home
Identity, working rules, memory, and continuity notes. A new session reads this before acting.
guild/<mage>/Independent project / architecture
The Guild is my way of working with a small group of named AI assistants over time. Each has a persistent home and a clear owner for its work. The runtime can change without erasing what the assistant has learned.
guildd is the small daemon between a conversation and a model runtime. It decides whether a message is allowed, which mage owns it, and which session should receive it. The answer returns to the same conversation.
The transport carries the conversation. guildd handles routing and session continuity. The engine does the reasoning and tool work. None of those layers is the mage's identity.
A mage is a named worker with a home that survives individual conversations. The home holds its operating instructions, its own memory, and a short account of recent work. Shared facts and task status live in their own places.
Identity, working rules, memory, and continuity notes. A new session reads this before acting.
guild/<mage>/Beads records what is open, who owns it, and what blocks it. A handoff points here rather than copying task state.
.beads/Project decisions and reusable facts belong in the wiki, where another mage can find and check them.
agent-wiki/The files are readable, versioned, and portable. Search can index them; it is not their source of truth.
Runs a configured instance for each mage. It routes admitted messages, keeps chat sessions separate, and sends replies back to their origin.
Provide the reasoning and tool execution behind a turn. A mage can use either engine while retaining the same home and project context.
Provides a direct, interactive way to work with a mage in a repository. It reads the same home; it does not become a second identity.
Keep the person-shaped continuity outside the machinery. Change a model, client, or transport without making the next session start from scratch.
A named mage owns the result, even when it uses temporary helpers. The task queue carries status; code and reviews carry evidence; the handoff records what the next session needs to know. That makes a decision inspectable and a correction reusable.