Store schema

Each workspace's knowledge base is one SQLite file:

Workspace	Config	Store
default	<base>/config.yaml	<base>/knowledgebase/ground.db
acme	<base>/workspaces/acme/config.yaml	<base>/workspaces/acme/knowledgebase/ground.db

Tables are created on open (CREATE TABLE IF NOT EXISTS), and a missing column is added with an ALTER. Everything that varies by source lives in a JSON details column, so a new source never changes the schema.

Table	Key	Holds
nodes	(source, kind, id)	title, uri, updated_at, ingested_at, details JSON, content JSON
links	(src..., dst..., kind)	attributes JSON; the two ends may be in different sources
ingest_runs	id	a run log per source: status, counts, message
flow_runs	id	one run of a flow: flow and prompt version, trigger, work item, where the source stood, stages reported, status, counts
chats, chat_messages	id / (chat, seq)	a chat: subject, branch, the AI's last call; the messages with the pieces they showed
orchestration_instances	id	one walk through an orchestration: subject, branch, state, status, open run, history
branches	id	name, description, work item, git branch and commits, status (open / merged / abandoned), what its merge applied
branch_nodes	(branch, source, kind, id)	a branch's copy of a node; deleted hides it in the branch; base_at records when main's version was written
branch_links	(branch, src..., dst..., kind)	links the branch added, or removed (deleted)

Reads go through two SQL views, v_nodes and v_links, built for the branch the request is in, so the same queries serve main and every branch. Merging a branch applies its copies to nodes and links in one transaction.

content is null on a pointer node, and a write that carries no content leaves the stored content alone, so refreshing a node's label can never wipe what somebody wrote. Re-ingesting updates rows in place, so ids stay stable and links keep working.