Migration note: importing the prototype into the praxis model (BL-016)

The predecessor was a hand-run MCP gateway whose knowledge lived in flat markdown and whose "known-good" was an operator's memory and ad-hoc baselines. This note describes how to bring that material into the praxis model without weakening any invariant. praxis remains self-contained: the import is a one-time data load, not a runtime dependency on the prototype.

What maps to what

Prototype artifact	praxis target	How
CONTEXT-<host>.md, fleet-inventory.md (host knowledge)	host-knowledge skills under skills/<name>/SKILL.md	One bundle per topic; write a precise, router-targetable description (ADR-0010). Keep prose in the body.
The host list and routing	config/inventory.yaml	Fill the schema in config/inventory.example.yaml; set host_type (it gates actuation, SEC-5) and routing (ssh alias, talosctl endpoints/nodes).
Per-host known-good baselines	KNOWN_GOOD facts in the store	For each baseline value, write a Fact(subject="host:<name>", predicate=<...>, fact_type=KNOWN_GOOD, value=...) via store.put_fact. The drift engine diffs observed against these (drift.diff).
Operator procedures / runbooks	tool skills under skills/ and entries in runbooks/ (the companion repo)	Knowledge becomes a tool skill; the actual script stays a runbook invoked through the runbook adapter.

Procedure

1. Inventory first. Translate the host list into config/inventory.yaml. Get host_type right for every host; a mis-typed Talos host as ubuntu would let the SSH adapter be selected (SEC-5 refuses it, but the data should be correct).
2. Host knowledge. For each prototype context document, create a host-knowledge SKILL.md with a specific description. Run make eval and add golden queries so routing stays measured (ADR-0010).
3. Known-good baselines. For each host and predicate (os_version, ssh_config, file_integrity, service states, listening ports), record a KNOWN_GOOD fact. Treat the baseline as operator-blessed: it is the desired state the drift engine compares against. Bitemporality means a later correction supersedes with an actor and a reason, never an in-place edit (SEC-10).
4. Verify. Ingest a current observation (the ingest_observation tool or a collector), then run drift_scan. The findings are the gap between the imported baseline and reality; triage them with the drift-triage skill.

What not to import

• No secrets, tokens, or private keys. Credentials are injected out of band and scoped through the CredentialBroker (invariant 9); they never enter the store or a skill bundle.
• No executable prototype code as a skill contract.py. Untrusted bundles load inert (allow_contract=False, ADR-0010).