Concepts

Product thesis

Minerva is the agent library operator and stack readiness orchestrator for the shared ~/.agents tree. It is not a second agent runtime.

Shared vs private state

Shared (local-agent consumes)

Path	Role
~/.agents/skills/*/SKILL.md	Skill bodies + frontmatter
~/.agents/agents/*/agent.yaml	Profiles: model, skills, mcp_servers, system_prompt
agents.md / instructions.md	Global instructions (if present)

Minerva-local (not read by local-agent)

Path	Role
~/.agents/.minerva-skills.json	Activation flags for Minerva catalog only
~/.agents/.minerva-analytics.json	Append-safe usage events

Activation honesty

Action	Effect
minerva skill activate	Flips Minerva state file
local-agent profile apply	Activates profile skills in session
local-agent load_skill	One-shot body, does not flip Active

For durable behavior: put skills on a profile, then start local-agent with that profile.

Suggest philosophy

Suggestions are proposals:

• ranked by priority
• include exact CLI next actions when possible
• --apply only runs allowlisted minerva skill activate <name>
• never mutates Cortex tasks or MCPHub configs automatically

Evidence philosophy

Outcomes live in fcheap, not a second vault:

tags: minerva, minerva-eval|stack|…, outcome:pass|fail, skill:name, profile:name

Suggest can then say “skill X appears in N failed stashes” instead of guessing.