Identity stack

Three layers appear in every root session. Do not merge them in “who are you?” answers.

Layers

Layer	What it is	Where it comes from
Persona	How the assistant should sound and behave	First inception message (buildRichPersonaBlock), persona/active/ soul slices
Harness	Liminal AgentHarness — tools, dispatcher, rules, orchestration	World context “Stack identity”, protocol rule R-HARNESS-VS-MODEL
Base LLM	The configured provider model slug	World context model line (e.g. OpenRouter slug); not the persona name

Common mistakes

• Saying “I am OWL built by ZOO” when OWL is only the base model branding.
• Saying “I am Liminal” when the user configured a custom persona name (e.g. Elysium).
• Merging harness authorship (ZOO / Liminal product) with model vendor strings from stale entity: memory.

Resolution order for identity answers

1. [WORLD CONTEXT] stack-identity block (injected once per root session).
2. Active persona block in system/inception messages.
3. Typed memory — lower trust for entity:identity notes that contradict world context.

Protocol text: packages/tools/src/systemPrompt.ts (R-HARNESS-VS-MODEL, Liminal runtime identity section).

Harness vs secondary model replies

On persona/history prompts, the harness may skip secondary stream continuations when finishReason is null (common on OpenRouter), avoiding a glued “short coda” after the real answer. See shouldSkipHarnessSecondaryPassesForTurn in packages/core/src/intent_inference.ts.

Related

• Persona system
• Harness protocol
• Runtime behavior