What it does
About this page
This documents a stack generated from soliplex-template. A generated
project ships its own copy of this page without this note.
The provisioner is the template's own adapters (its room config, records, ingester fragments and advice) over soliplex-concierge's provisioner framework, which runs the steps below and serves the HTTP API.
Holding a lock (src/.rooms-provisioned.lock) through step 8, the
provisioner:
- refuses a room id that anything already uses (see Before anything changes)
- files a
new-roomtracking issue taggedprovision-start, when a Gitea is configured and the concierge extension is installed - writes
room_config.yaml,REQUEST.md(who asked for what, and what is set up) andprovision.yaml(the same, as data) intosrc/rooms-provisioned/<room-id>/ - creates the room's uploads directory,
uploads/rooms/<room-id>/ - registers the room's database with the ingester, in a fragment of its
own,
src/rooms-provisioned/databases.d/<room-id>.yaml(anfssource over the room's uploads, and its place inlancedb.databases), then has the ingester reload its config, which creates the database - makes the room private, granting READ and WRITE to its owner: by
default, users whose
preferred_username(PROVISIONER_OWNER_CLAIM) is the requester's; the policy is also recorded as the room'sauthz.yaml - audits the installation's rooms
- commits the room (its directory, with all three records) and its
fragment in
src/rooms-provisioned(the requester as author), and pushes it when a Gitea is configured - has the backend reload its room configs, which makes the room live
- tags the tracking issue
provision-complete, orprovision-incompletewhen the steps after the commit left work for an operator (listed in the comment), then comments on and closes it
A failure before the commit undoes what the run itself created, and only
that (and tags the issue provision-failed, comments on it, and leaves it
open). After the commit, the room is valid as
committed: a step that fails there is reported, with how to finish it.
Two steps have time limits:
- the ingester's reload (120 seconds): one that times out may still apply after the run has undone its changes, so the failure says to reload the ingester again once it answers, and to remove the room's database directory if one appears
- the push (60 seconds): one that times out is reported as a step to finish
src/rooms-provisioned is a git repository of its own (the stack's
src/* rule ignores it), listed in room_paths as
/app/src/rooms-provisioned. The backend reads it through its read-only
src/ mount, outside its config watch.
Before anything changes
The provisioner refuses a room id when any resource named for it already exists:
- its directory in the clone
- its database directory (
rag/db/<room-id>.lancedb) - its fragment in the clone (
databases.d/<room-id>.yaml) - an entry for it in either map of
haiku.rag/haiku.rag.yamlor of a fragment (or another database placed at its location)
It also refuses to start when haiku.rag/haiku.rag.yaml cannot take the
fragments:
- its
databases_includemust namerooms-provisioned/databases.d, where the ingester mounts the clone's fragments - every database in its
ingester.databasesmust be placed in itslancedb.databases, as haiku.rag requires once fragments place theirs - a non-empty uploads directory
- a room authorization policy
The refusal names each one found, and changes nothing: remove them, or choose another name.
Limits
- The lock serializes provisioner runs only. Hand edits of
haiku.rag/haiku.rag.yaml, orrag_db.py/add_room.pyruns, are not serialized with it; an ingester reload applies whatever the file says then. - Every room the provisioner creates is a conversational RAG room over its own database; suggestions are left out, for an edit once documents are ingested.