Deterministic discovery and Ask Registry

Ask Registry checks a verified local artifact before making a backend request. Exact catalog facts feel immediate, work without model reasoning, and avoid unnecessary tokens. Semantic recommendations still use the shared trusted AgentsKit Chat backend and return canonical citations.

flowchart LR
  Q["Your question"] --> E{"Exact declared fact?"}
  E -->|"yes"| L["Local answer + canonical source"]
  E -->|"no"| B["Trusted semantic backend"]
  B --> R["Grounded recommendation + citations"]

What stays local

• exact agent IDs and unambiguous titles;
• npx agentskit add <id> commands;
• exact category and declared capability names;
• canonical agent pages and source links;
• contribution and AgentsKit ecosystem navigation.

Try research, npx agentskit add research, category security, or capability prompt injection defense in Ask Registry. These queries make zero Ask backend calls.

What escalates

Questions that require interpretation, comparison, ranking, or problem-to-agent recommendation are not guessed locally. For example: “Which agent should help triage a noisy incident and draft a handoff?”

The browser sends a validated low-confidence miss with bounded conversation context. Authentication and server configuration select the Registry corpus and persona; query parameters cannot switch authority. The response must cite the catalog or documentation used to ground it.

Machine contract

• /deterministic/site-config.json pins the site ID, artifact path, content hash, and fallback mode.
• /deterministic/knowledge.json is bounded, schema-validated, canonically serialized, and SHA-256 verified.
• /r/index.json remains the installable catalog authority.

If the local artifact is unavailable or corrupt, Ask Registry fails safely to the configured backend. It never treats an unverified local payload as fact.