Skip to content

Skill Service

Cognition Skill Service is the typed authority for skill artifacts — capability blueprints expressed as ordered steps with typed inputs and known references. It is the advisory family that gives an agent reusable patterns of doing things.

What A Skill Is

PropertyValue
TypeAdvisory artifact (not core truth)
ShapeOrdered steps
Input typingTyped
ReferencesKnown refs to other artifacts
LifecycleSave / list / load / lexical-search / explicit-delete / history
ValidationNon-empty step validation; reference target check

A skill is not model state. It is a typed bundle that describes "to do X, follow these steps." The agent can search, load, and follow skills.

Why Skills As Typed Bundles

A model that hallucinates "the right way to do X" might be wrong. A model that consults a typed skill bundle gets a vetted pattern authored under admitted contracts. The agent's reliability benefits.

BenefitWhat it gives
ReusabilityAgent reuses tested patterns instead of re-deriving
AuditabilityThe skill bundle is typed; what the agent did is reconstructible
CleanupStale skills can be archived under digest
Cross-sessionSkills travel with the agent

Skill Operations

OperationBehavior
SaveSave a typed skill bundle
LoadLoad by id
ListList under typed filter
LexicalSearchSearch by lexical query
DeleteExplicit destructive delete
HistoryRead history of this skill

Reference Validation

A skill bundle's references must point to admitted targets.

CheckWhat it does
Reference target existsYes — fail-closed if missing
Cross-scope referenceForbidden — fail-closed
Type compatibilityReference type matches expected

A skill that references a non-existent or cross-scope artifact fails validation at save time. There is no silent acceptance of broken references.

Reader Scenario: An Agent Saves A New Skill

An agent has learned a useful pattern through interaction with a user.

  1. Compose skill bundle. Ordered steps, typed inputs, references to relevant memory / knowledge artifacts.
  2. Submit Save. Skill service receives.
  3. Validate.
    • Non-empty step validation.
    • Reference target check (each ref must exist in scope).
  4. Admitted. Skill is saved with typed shape.
  5. Service-derived metadata. Support, lineage.
  6. Available for future use. Agent can search / load.

The skill is now part of the agent's typed advisory bundle collection.

Reader Scenario: An Agent Searches For A Skill

An agent is about to do something it has done before.

  1. Lexical search. SkillService.LexicalSearch with query.
  2. Service returns matches. Typed skill bundles matching.
  3. Agent loads top match. Reads the typed steps.
  4. Agent follows the pattern. Ordered steps drive reasoning; the agent reuses the pattern.

The agent did not have to re-invent. It consulted the typed skill.

Reader Scenario: A Skill With A Stale Reference

A skill bundle references a knowledge page that has since been deleted.

  1. Refgraph reasoning. The skill's outgoing reference is broken.
  2. Digest pass. Identifies the broken ref as a candidate for cleanup.
  3. Archive proposed. Skill is archived (reversible).
  4. User can restore. If the user restores the missing knowledge page (or amends the skill to remove the ref), the skill can be unarchived.
  5. Or proceed to remove. Later digest pass may remove the archived skill.

Cleanup is explainable — refgraph reasoning gives "why."

Why Skills Stay Advisory

Skills are advisory, not core truth. They cannot demote kernel truth.

ConcernWhy advisory matters
Skill says XModel may consider, may override under reasoning
Kernel says YInviolate truth
ConflictKernel wins; skill cannot demote kernel

A buggy or out-of-date skill cannot corrupt kernel truth. The kernel primacy is structural.

Boundary Summary

ConcernOwner
Skill bundle storageCognition Skill Service
ValidationService-side under completion gates
Reference graphRefgraph
CleanupDigest + explicit delete
Truth weightAdvisory (not core)

Source Basis

Nimi AI open world platform documentation.