3tears-agent-identity

Versioned identity blocks for LLM agents — the substrate for self-evolution: the agent proposes changes to its own identity, a human consents, and nothing is lost, only superseded.

Model

An identity block (persona, reinforcement, anti-sycophancy guard, self-improvement scratch, presence voice) is a linear parent-pointer version chain. Exactly one active version exists per (agent_id, customer_id, user_id, block_key, variant); each version carries its content, a rationale, a content_hash, and its lineage.

A block can have variants: copies named by the consumer, each with its own chain and its own active version, at the block's tier. "" is the block itself and the default everywhere, so a consumer that names no variant sees one chain per block. The package gives a name no meaning; which variant applies where, and which one a proposer may write, is the consumer's rule (load_identity_propose_tool(..., variants=...) binds the tool's).

Lifecycle: propose (agent) → consent (user) → apply, with non-destructive rollback. Consent is tiered — tier-1 (identity-shaping) blocks require human consent before applying; tier-2 (routine) blocks auto-apply with an async veto.

This differs from knowledge's shadow-chain (which is scope-layering, not temporal version history): identity versions are a temporal chain with rollback.

Contents (v0.15.0)

• T2.1a: identity_versions table + IdentityVersionsCollection + the block / status / tier value types + migration v001.
• T2.1b: lifecycle ops (propose/consent/reject/rollback) + FrameworkEvents + owner RBAC + the identity_propose tool.

Consumers own the apply path (subscribe/authorize/write); this package provides the store, the version chain, the lifecycle, and the events.