Session-native Minerva

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

A harness that already speaks MCPHub (sonar, local-agent, Grok) should be able to ask two session questions without loading 36 look-alike tools:

1. Which skill should I load for this task?
2. What is the honest next library or stack action?

Mutations stay on the CLI so the harness can approval-gate them.

Authority

Domain	Owner
Prompt assembly, /skill, load_skill, /agent	sonar / local-agent
MCP trust, AUTO, ledger	the harness
Lazy catalog, resolve_tool, call stats	MCPHub
Task evidence	Cortex
Files under ~/.agents	disk SSOT — Minerva writes, harnesses read
.minerva-skills.json	Minerva-local only — harnesses do not read it
learn, resolve_skill, presence vs readiness, suggest	Minerva

minerva skill activate does not inject into a live session. Put skills on a profile (minerva profile add-skills) for durable loading. Use the harness load_skill for a one-shot body.

Nine read-only MCP tools

Tool	Use when
minerva_learn	First time / how Minerva works
minerva_status	Unified doctor — is the stack actually ready?
minerva_suggest	Ranked next library or stack actions
minerva_resolve_skill	Which skill should I load for this task?
minerva_skill	action=list|show|compare
minerva_profile	action=list|show|compare
minerva_library	action=lint
minerva_stack_check	PATH presence only
minerva_evidence	action=docs|search

There are no MCP aliases for the old CRUD names. Create, activate, delete, add-skills, export/import, and evidence save/close are CLI-only.

MCPHub config (safe fragment)

Do not commit a live mcphub.yaml — it often holds vault refs. Copy this fragment into ~/.config/mcphub/mcphub.yaml and point command at the Minerva you actually built.

expose: lazy
pin:
  - minerva__learn
  - minerva__resolve_skill
  - minerva__status
  - minerva__suggest
servers:
  minerva:
    command: minerva   # or an absolute path to bin/minerva
    args: [mcp, serve]
    enabled: true
    description: Agent library operator and stack readiness
    tags: [agent, skills, profiles, readiness]
    use_when:
      - Resolve which skill to load for the current task
      - Check whether the shared agent library or companion stack is ready
      - Get a one-page Minerva onboarding brief
      - Rank next library or stack actions without mutating disk

Restart the MCPHub gateway after changing pins or the Minerva binary.

Sonar

Sonar already loads ~/.agents (/agent, load_skill). It should call Minerva through MCPHub, not ./bin/minerva, except when dogfooding the Minerva repo itself.

Default trust lists the nine minerva__* routes as read-only. AUTO may run query-only ./bin/minerva learn|status|skill show|skill resolve|library lint from a workspace-built bin/minerva. Activate, apply, delete, mcp serve, and stack deep stay approval-gated.

minerva learn --json
minerva skill resolve "review this pull request" --json
minerva bridge show <profile> --harness sonar

Profiles

Optional kind: workspace|role on agent.yaml. Names default, dev, and workspace infer workspace. Suggest only auto-adds skills to workspace profiles — never to a specialist like code-reviewer.

Out of scope here

Cross-harness skill sync (~/.grok/skills, ~/.claude/skills, .skill-lock.json), a Studio TUI, and mid-session injection that bypasses load_skill.