Deterministic demo dataset for development and test environments.

One command builds the same baseline everywhere: documents under a demo prefix, one typed memory per taxonomy type, one session committed into the tree, one demo skill, and two indexed Elixir sources for the demo project. Every fixture carries a distinctive seedling-* keyword so demos and tests can assert on keyword search, find/2, and grep/2 hits.

{:ok, report} = AgentDb.Seed.seed()
{:ok, report} = AgentDb.Seed.seed(prefix: "viking://resources/custom", force: true)

Safety

The seed refuses a store that already holds data ({:error, :store_not_empty}) unless force: true (merge: missing seed URIs are created, present ones revised in place, nothing deleted) or clean: true (only the seed scope is removed first: the prefix subtree, the demo-* memories, the demo skill, and the demo code prefix; data outside that scope survives). A refusal writes nothing. An invalid prefix fails with :invalid_uri on the first write, also having written nothing.

Pass env: :prod (the Mix task passes Mix.env()) to engage the production guard: seeding in production without allow_prod: true is refused as {:error, :production_guarded} with nothing written. The library never reads Mix.env/0 itself, so releases without Mix behave the same through the explicit option.

Scope

The :prefix override (default "viking://resources/demo") scopes the documents and the committed session. Memories, the demo skill (user "demo"), and the demo code entries live at fixed global URIs by design, so a custom prefix still revises the same memories, skill, and code.

Report

Success answers {:ok, report} with the prefix used and per-kind counts:

%{prefix: ..., documents: 5, memories: 5, sessions: 1, messages: 3,
  session: "<prefix>/sessions/intro.md", skills: 1, skill_files: 2,
  indexed: 2}

Re-running converges: the same URIs are revised (memories supersede, the skill is replaced whole, the session document is rewritten), so the tree holds one active value per URI. Each run opens a new session row to commit from; only the committed document is part of the dataset.

Rollback is AgentDb.rm(prefix) plus forgetting the demo-* memory URIs; re-seeding afterwards converges. Reads need no model: keyword search, find, and grep reach seeded content immediately, while embeddings and summaries regenerate through the existing job queue.