How agents use Genomi skills, base MCP tools, genomi.invoke, semantic retrieval hints, defaults, and response profiles.

Genomi is designed for host agents. Skills provide routing rules; MCP tools perform the actual work.

Skill files

File	Role
`SKILL.md`	Umbrella Genomi skill and core operating rules
`AGENTS.md`	Project/session rules for agents working in the repository
`INSTALL_FOR_AGENTS.md`	Canonical install workflow
`skills/<capability>/SKILL.md`	Capability-specific routing and result semantics
`skills/conventions/*.md`	Shared conventions for context routing and evidence quality

The installer symlinks the umbrella skill and per-capability skills into detected host skill directories such as ~/.claude/skills, ~/.codex/skills, ~/.openclaw/skills, ~/.hermes/skills, and ~/.agents/skills.

Static guidance vs live state

Skill files are static guidance. Live state comes from Genomi operations:

genomi.describe_context for selected user, AGI, access readiness, and active response profile.
genomi.check_libraries for installed public libraries.
Operation result envelopes for defaults, source state, missing libraries, and background jobs.

Do not infer current genome context from old chat memory or from a skill file.

Base tools

The default MCP tool list exposes the base capabilities:

genomi.* context, install, user/profile, library, background-job, and dispatcher tools.
journal.* notebook tools.
research.* reviewed-research memory tools because they belong to the journal capability.

Non-base capability tools are invoked through genomi.invoke.

Dispatching a capability tool

Read the capability skill first, then call:

{
  "tool": "pharmacogenomics.review_medication",
  "params": {
    "drug": "clopidogrel",
    "gene": "CYP2C19"
  }
}

genomi.invoke validates the registered operation name and that operation's own input schema before dispatch.

Semantic retrieval hints

Search-like operations can accept semantic_context. The host should always preserve the user's original wording as raw_query and add alternate biomedical wording only when the chat supports it.

Example:

{
  "tool": "prs.search_scores",
  "params": {
    "query": "will I go bald",
    "semantic_context": {
      "raw_query": "will I go bald",
      "host_expansions": [
        "male pattern baldness",
        "androgenetic alopecia",
        "hair loss"
      ],
      "host_entities": [
        {
          "text": "androgenetic alopecia",
          "type": "trait_or_condition"
        }
      ]
    }
  }
}

Semantic expansions are retrieval terms, not evidence. Inspect returned term_matches, term_misses, retrieval streams, and diagnostics.

Response profiles

The install flow asks for a default answer style and persists it in the Genomi registry. Current profiles:

Profile	Use
`eli5`	Teaching style with definitions and step-by-step explanation
`patient`	Concise patient-friendly explanations
`literate`	Genetics-literate answers with genes, variants, zygosity, and evidence classes
`expert`	Methods, quality checks, source provenance, caveats, and structured evidence

Response profiles shape communication style only. They do not override evidence limits, privacy boundaries, or clinical-confirmation language.

Direct CLI vs MCP

Routine use is MCP. The normal shell commands are:

genomi serve
genomi install
genomi --help

Developer/debug shell surfaces such as genomi tools and genomi call are gated by GENOMI_CLI=1 in the current runtime. Agent-facing docs should route tool execution through MCP.

Skills and Tools