3tears-agent-skills

3tears-agent-skills

Procedural memory for 3tears-based agents. Stores per-agent / per-user labeled markdown procedures (body) and tool-surface modifications (tool_additions, tool_restrictions) plus a prompt_mode enum controlling how the body interacts with the consumer's base system prompt.

This package provides the schema + Collection layer:

• agent_skills -- one row per skill (partition column agent_id).
• agent_skill_invocations -- one row per skill load with synchronous outcome tracking populated by the consumer's post-LLM hook (partition column agent_id; composite FK CASCADE on parent skill).
• Two BaseEntity subclasses (AgentSkillEntity, AgentSkillInvocationEntity).
• Two BaseCollection subclasses exposing the public skill-registry API.
• A trigger-maintained search_vector (FTS for skill_list query-filter ranking; NOT auto-load).

Agent tools and the per-turn composition renderer do not live here.

Migration registration

from threetears.agent.skills import register as register_skills
from threetears.core.data.migrations import MigrationRunner

runner = MigrationRunner()
register_skills(runner)

Migrations are agent-scoped and depend on conversations (the invocation rows carry conversation_id -- ordering on apply, not an FK constraint).

SkillRegistryClient Protocol -- why this package takes no ACL / tools deps

agent-memory (a sibling package) takes direct dependencies on 3tears-agent-acl and 3tears-agent-tools because its memory tools need first-class ACL evaluator + tool-registry types in their public surface. agent-skills deliberately diverges. The tool factories take a thin SkillRegistryClient Protocol (three async methods, acl_permits / list_skill_eligible_tools / get_tool_introspect, plus a ConversationIdResolver / ActiveSkillProbe / ActiveSkillSetter callable triple wired by the consumer) instead of importing the ACL evaluator or the tool registry types directly.

The trade-off: callers must implement a small adapter (~10 lines over their existing NamespaceCollection + ACL cache + in-process tool registry), but agent-skills ships with zero hard deps beyond core and langchain-core. The consumer wires this; tests mock it. The ACL-via-evaluator approach targets the composition renderer, not the tools surface.

Future sibling packages should follow whichever pattern their public surface demands -- direct deps when types are part of the contract, Protocol when the contract is method-shaped and the deps are incidental.