Architecture

KGM Skill Index

Skills are indexed into KGM as AgentSkill entities alongside existing KnowledgeBundle and KnowledgeConcept entities.

Entity Format

{"type": "entity", "name": "skill:skupper-model-provider",
 "entityType": "AgentSkill",
 "observations": [
   "Signals: setup skupper, stop skupper, skupper status, ...",
   "Action: load_skill(name: \"skupper-model-provider\")"
 ]}

search_nodes("skupper status") matches the Signals observation and returns the entity with the exact load_skill call.

Separate Indexes, Merged Output

File	Contains	Generator
~/.agents/knowledge/.kgm-knowledge.jsonl	KnowledgeBundle + KnowledgeConcept	reindex-kgm.sh
~/.agents/skills/.kgm-skills.jsonl	AgentSkill	reindex-skills-kgm.sh
~/.agents/knowledge/.kgm-index.jsonl	Combined	Concatenation at end of reindex-kgm.sh

Symlink Workaround

Goose does not reliably pass the MEMORY_FILE_PATH env var from config.yaml to the MCP server-memory process (the session extension data has envs: {}). The MCP server falls back to its default dist/memory.jsonl inside the npm package cache.

Fix: setup-kgm.sh creates a symlink:

<npm-cache>/server-memory/dist/memory.jsonl → ~/.agents/knowledge/.kgm-index.jsonl

This breaks on npm package upgrades — setup-kgm.sh re-creates it.

Dispatch Flow

User: "skupper status"
  → Model calls search_nodes("skupper status")
  → KGM returns: skill:skupper-model-provider
     Action: load_skill(name: "skupper-model-provider")
  → Model calls load_skill(name: "skupper-model-provider")
  → Skill content loaded into context
  → Model follows skill instructions

No prose scanning. No pattern matching by the model. One tool call to find, one tool call to load.