zeekay
0e9c5a94da
@hanzo/persona already owns what a persona IS — the JSON Schema and the
PERSONA.md frontmatter convention. This pack does not restate it: it loads the
profile, checks it against that schema, and files it in the store.
Filing it in the store is the point. brain.db is read from TS, Python, Rust and
Go, so a persona written once is the same persona in all of them. The
alternative is each runtime loading its own copy of a file and drifting, which
is the duplication the brain exists to end.
Stored twice, deliberately, because they answer different questions: as a PAGE
(the document, for search and for a human) and as FACTS, one row per trait, so
recall("persona/<id>") answers "who am I" without parsing markdown. Nested
traits become dotted predicates; an array is joined rather than exploded,
because a persona's tags is one answer and not five rows.
A half-written profile still loads — a missing required field is NAMED in a
warning, never thrown, since a persona you are still writing should not stop a
brain from opening. No PERSONA.md at all is the ordinary case and returns null.
Soft contracts throughout, like every other pack here.
2.5 KiB
@hanzo/bot-persona
The agent's persona, attached to the brain.
A persona is who an agent is — its name, its OCEAN traits, its technical
preferences, the way it talks. @hanzo/persona
owns that shape: the JSON Schema and the PERSONA.md frontmatter convention.
This package does not restate it. It loads the profile, checks it against
that schema, and files it in the brain store.
Filing it in the store is the point. brain.db is read by every Hanzo runtime —
TS, Python, Rust, Go — so a persona written once is the same persona in all of
them. The alternative is each runtime loading its own copy of a file and
drifting, which is what the brain exists to end.
Use
It comes with the meta-pack; you do not register it separately.
import registerBrain from "@hanzo/bot-brain";
const { persona } = await registerBrain(api, {
persona: { path: "~/.hanzo/workspace/PERSONA.md" }, // the default
});
persona?.profile; // the loaded profile, or null
await persona?.reload(); // re-read after an edit
persona: false disables it.
What it writes
The profile is stored twice, because they answer different questions:
| as | where | answers |
|---|---|---|
| a page | persona/<id> |
the document — for search, and for a human to read |
| facts | one row per trait | "who am I", without parsing markdown |
Nested traits become dotted predicates, so a runtime can ask for one without knowing the document's shape:
await store.recall("persona/ada");
// → { predicate: "ocean.openness", object: "0.9", … }
// → { predicate: "tags", object: "math, engines", … }
An array is joined rather than exploded: a persona's tags is one answer, not
five rows.
A PERSONA.md
---
id: ada
name: Ada
ocean:
openness: 0.9
tags: [math, engines]
---
I build things, and I check them before I say they work.
Behaviour worth knowing
- No
PERSONA.mdis normal, not a failure — the pack returnsprofile: nulland stays inert. - A half-written profile still loads. A missing required field is named in a warning, never thrown: a persona you are still writing should not stop a brain from opening.
- A body with no frontmatter is a persona written as prose, and is kept.
- Everything is a soft contract: where the host exposes no store or no tool registry, the pack is inert for that surface rather than throwing.
Tool
Registers brain.persona — reads the profile — so an MCP-native agent can ask
who it is.