arizuko › concepts › autoviv

autoviv

Onboarding admits a new person into a group. Autoviv handles the other case: a new chat — a Slack channel, a Discord guild, a Telegram group — that should get its own folder the first time it talks to you. The point worth noticing is that this needs no new code: it’s the same group-creation primitive onboarding uses, triggered by a catch-all route and one MCP tool call instead of an invite. No daemon, no env var, no magic.

the recipe

Three primitives, composed:

1. Catch-all routes to a world agent. Two rows at low priority send every message from a platform into one folder, the “world” agent (call it atlas):

   seq  match                                          target
   100  chat_jid=slack:*/channel/* verb=mention        atlas
   200  chat_jid=slack:*/channel/*                     atlas#observe

   The #observe row stores every channel message so the agent has context. The mention row fires a turn when someone calls @atlas.
2. The world agent notices an unrouted channel. When atlas runs a turn, it sees the inbound message’s chat_jid. If no specific child folder is wired to that JID and the sender holds operator grant, it calls the MCP tool register_group with the JID and a sanitized folder name.
3. register_group does the rest. ipc/ipc.go:1309 → routd writes the group row, seeds the folder with skills and settings (SetupGroup), and inserts a high-priority route room=<jid-room> → <folder> in one step. The next message in that channel lands in atlas/eng-support instead of atlas.

worked example

A teammate invites the bot to a new Slack channel #eng-support with JID slack:T1234/channel/C5678 and posts “@atlas can you help here?”

• Route table has no specific row for C5678; the catch-all rule sends the message to atlas with verb=mention.
• atlas runs a turn, sees the JID and the sender’s operator grant, and calls register_group(jid="slack:T1234/channel/C5678", folder="atlas/eng-support").
• routd writes the group, seeds groups/atlas/eng-support/ with the default skills, inserts route room=T1234/channel/C5678 → atlas/eng-support at seq=0.
• The agent replies in-thread from atlas. The next message in that channel routes to atlas/eng-support directly — its own folder, its own session, its own CLAUDE.md.

why no special env var or daemon

Autoviv is not a feature; it is the absence of a feature. The pieces already exist: glob route matching, the register_group MCP tool, group seeding via SetupGroup, the operator grant. An operator who wants per-channel folders writes a catch-all route and a paragraph in the world agent’s CLAUDE.md. An operator who prefers to register each channel by hand omits both. Same code path either way.

what your world agent needs to know

Paste a paragraph like this into the world agent’s CLAUDE.md or PERSONA.md:

When you receive a message from a Slack channel JID
(slack:*/channel/*) that does not yet have its own folder,
and the sender holds operator grant, call register_group
with folder="atlas/<short-name-derived-from-channel>" and
jid="<the channel JID>". Sanitize the channel name:
lowercase, replace spaces and punctuation with '-'.
Reply once in the new folder confirming the channel
is now registered. Do not register channels for non-operator
senders — ask the operator instead.

Tune it: auto-register channels matching a pattern, ask the operator first, only register on explicit request. The agent is making the policy decision; the platform just runs the call.

web slots

A group holding the web:publish grant gets two writable web slots bind-mounted into the container: ~/public_html/ and ~/private_html/ (Apache mod_userdir convention). The agent writes a file into ~/public_html/ and vited serves it at /pub/<folder>/ — one path segment per folder, no per-group subdomain. ~/private_html/ serves at /priv/<folder>/ behind an OAuth/JWT gate. An autoviv’d child atlas/eng-support serves at /pub/atlas/eng-support/; it shares the instance hostname with every other folder.

limits

• A delegated grant is required. register_group is not in the default role, so atlas has to be handed mcp:register_group explicitly. The row's scope glob is the whole containment rule: granted atlas/**, it can register atlas/eng-support and anything else under atlas, but not other/foo. Widening its reach means editing that glob, not moving the folder.
• Idempotency is partial. Re-registering the same folder is safe, but the route insert is unconditional — calling register_group twice with the same JID writes a duplicate route row. Have the agent check list_routes first.
• Folder seeding is best-effort. If SetupGroup fails, the group + route are still written and routd logs a warning. Inspect groups/<folder>/ if the new agent acts blank.

Full register_group signature and the grant it needs: reference / mcp, reference / grants.

go deeper

Route tables, glob matching, and #observe mode: routing. Grants, roles and delegation: grants reference. Full register_group signature and other MCP tools: mcp reference.