Lumina-MSK/AGENT_SKILL/agent_tool_execution/SKILL.md

Skill: agent_tool_execution

Trigger Conditions

Any change involving Gemma-4-E2B agent tools, Exa search, Supabase RAG RPC, or MedGemma escalation.

Rules

1. Tool catalog is authoritative

• Only use: exa_search, supabase_query, escalate_medgemma.
• Never invent tool names in prompts or handlers.
• Bump TOOL_CATALOG_VERSION on breaking schema changes.

2. Secrets and egress

• EXA_API_KEY, SUPABASE_SERVICE_ROLE_KEY, MedGemma keys: backend only.
• Browser calls BFF routes; never https://api.exa.ai directly from the client.
• PHI-scrub queries before any network egress (secrets_and_phi_safety).

3. Exa defaults (Pattern 1 — raw retrieval)

{
  "type": "auto",
  "numResults": 10,
  "contents": { "highlights": true }
}

• Canonical reference: https://docs.exa.ai/reference/search-api-guide-for-coding-agents
• Do not use deprecated params (useAutoprompt, top-level highlights, livecrawl: "always").
• outputSchema synthesis is BFF-only enrichment; not the default agent-loop path.

4. Retrieval policy (NFR-18)

• supabase_query first for in-corpus MOH guideline grounding.
• exa_search for external/recency queries.
• Final clinical answers must cite chunk IDs (Supabase) or URLs (Exa).

5. MedGemma escalation

• Requires consent + audit (NFR-16a) before escalate_medgemma.
• No direct Modal URLs from the browser.

6. MediaPipe agent output

• Parse tool calls from the answer channel when CoT is enabled.
• One block per turn: <tool_call> OR <final>.

Package location

ml/implementation/nlp/agent_runtime/