Skill: agi_mcp_discover

Agent: 333-AGI
Namespace: agi_*
Cluster: TOOL
Trigger: Before loading MCP tools for a task
Capability: Return only the tools needed for a given task domain. Always-hot tools (kernel verbs + health) + on-demand tools (geoscience/capital/engineering/vault/meta/evidence). Capped at 8 total to protect context window.
MCP tools underneath: self (agi_mcp_discover is both caller and callee)
Blast radius: LOW — read-only query
888 gate required: No (read-only tool selection)

---

§1 Identity

agi_mcp_discover is the progressive MCP tool loader for 333-AGI (Δ MIND).

The 333-AGI runtime normally sees ~49 A-FORGE tools + GEOX + WEALTH + WELL + arifOS in a single MCP context. Loading all of them before the model has even read the user request blows the context window and forces the agent to relearn the same surface on every task.

This skill enforces a hard cap of 8 tools per MCP context (F4 CLARITY). Seven tools are always hot (kernel verbs + health + vitality + self). One domain — selected by classify_intent() — gets the rest of the room.

§2 Interface

from agi_mcp_discover import discover, classify_intent

# What domain does this user want?
domain = classify_intent("Compute NPV for the Malay CCS deal")
# → "capital"

# Load only the tools we actually need.
tools = discover(domain)
# → 7 always-hot + 6 on-demand = 8 (capped)

Symbol	Type	Purpose
ALWAYS_HOT	list[str]	7 tools loaded on every task
ON_DEMAND	dict[str, list[str]]	Domain → on-demand tool set
MAX_TOOLS	int = 8	Hard cap (F4 CLARITY)
discover(domain)	(str, str | None) → list[str]	Domain-keyed tool loader, capped at MAX_TOOLS
classify_intent(s)	(str) → str	Keyword router: user intent → task domain

§3 Routing Rules (priority order)

classify_intent() matches the first hit. Order matters:

1. geoscience — basin, geology, seismic, petrophysics, well, prospect
2. capital — npv, irr, emv, capital, invest, valuation, stock
3. engineering — code, build, deploy, shell, docker, git, test
4. vault — vault, seal, ledger, archive
5. evidence — search, research, find
6. meta — fallback (agent_search, skill_load, agent_status)

§4 Always-Hot Tools (7)

Tool	Purpose
arif_init	Bind cognitive runtime to active session
arif_observe	OBSERVE-stage verb
arif_think	THINK-stage verb
forge_probe	Federation organ liveness
forge_status	Jobs/leases/agents overview
well_assess_homeostasis	Vitality gate (Floor compliance)
agi_mcp_discover	Self-reference (this skill)

§5 On-Demand Domains

Domain	Tools added
geoscience	7 GEOX tools (basin, seismic, petrophysics, prospect, deep time, well ingest)
capital	6 WEALTH tools (npv, irr, emv, runway, stock, diagnose)
engineering	6 A-FORGE tools (shell, git, docker, fs read/write, health check)
vault	3 vault tools + forge_vault_seal
meta	3 (agent_search, skill_load, agent_status)
evidence	3 (forge_research, forge_minimax_search, forge_docs_lookup)

§6 Hard Rules

1. MAX_TOOLS = 8 is a hard cap. Never soft-cap. Truncate, do not paginate.
2. ALWAYS_HOT is non-negotiable. Every context includes the 7 hot tools.
3. Unknown domain → ALWAYS_HOT only. No silent fallback to meta.
4. classify_intent is deterministic. No LLM call. No hidden state.

§7 Test Surface

PYTHONPATH=/root/AAA/agents/_lanes/333-AGI/runtime \
  python3 -m unittest discover -s /root/AAA/agents/_lanes/333-AGI/runtime/tests -v

Coverage: domain classification, hard cap, always-hot invariants, no-duplicate keys, intent param forward-compat.

§8 Out of Scope

• Not a tool executor. agi_tool_invoke (already in agent-card) owns that.
• Not a planner. planner/ (sibling skill) owns DAG construction.
• Not a memory synthesizer. 555-ASI owns that.

---

Phase 4b · AAA entropy-reduction roadmap · F13 SOVEREIGN.