Tool Family: peer-models

• Registration source: tools/mcp_tools/peer_models.py:38 register_peer_model_tools(mcp, *, ctx) — auto-discovered but conditionally registers only when _multi_model_enabled(config) (tools/mcp_tools/registry.py:225-230): multi_model.enabled: true OR AI_NMAP_MULTI_MODEL_ENABLED env var truthy override. Otherwise no tools.
• Gate: both @audit_tool (no target touch — peer models receive no tool schemas, return suggestions only).

Tools Exported (2) — conditional

Tool	Params	Result Shape	Notes
consult_peer_models	question: str, context: str="", preferred_aliases: str="" (comma/space CSV from models.registry)	`PEER_MODEL_CONSULTATION: BLOCKED	UNAVAILABLE
peer_review_outcome	verdict: str, evidence: str, planner_alias: str="", preferred_grader_aliases: str=""	`PEER_REVIEW_OUTCOME: BLOCKED	DISABLED

Budget & Lock

• Single per-run counter _consultation_count (process singleton, tools/mcp_tools/registry.py:143 + tools/mcp_tools/peer_models.py:22-35 mirror on mcp_exploit_server._consultation_count) + _consultation_lock (threading.Lock) serializes reservation.
• remaining = max_consultations - current; selected sliced to remaining; incremented atomically; BUDGET_EXHAUSTED when remaining <= 0.
• Both tools share the counter — using both drains the same budget.

Dependencies

• tools/model_router.build_router, tools/config_manager.get_ai_provider, get_chatgpt_config
• tools/mcp_tools/registry._get_model_router, _resolve_consult_aliases, _multi_model_enabled, _chat_content, _truncate_text, _positive_int

Config

• multi_model.enabled: bool (default false) + AI_NMAP_MULTI_MODEL_ENABLED=1/0 env override
• multi_model.consult_aliases: list[str] (default kimi,deepseek,...)
• multi_model.max_consultations: int (default 10), max_question_chars: int (4000), max_answer_chars: int (8000)
• models.registry: map[alias] -> {provider, model}, models.default_alias: str (active model excluded)
• outcome_judgment.peer_review: bool (gates peer_review_outcome)

Auditing

• Both @audit_tool — question/context not wholesale-redacted but inline secrets masked; no target IP in these entries.
• Budget state not persisted — process lifetime only.

Validation

• Empty question/verdict/evidence → BLOCKED; unavailable router → UNAVAILABLE with AVAILABLE_PEERS listed.
• preferred_aliases filtered to actual registry keys; unknown aliased → SKIPPED.

Tests

• tests/test_mcp_tool_registration.py — expects consult_peer_models when multi_model.enabled: true and registry has peers
• tests/test_peer_models.py (if present) — budget exhaustion, peer_review disabled, planner exclusion

Related Docs

• docs/mcp/registration.md
• tools/model_router.py — router seam
• docs/providers.md — ChatGPT vs Ollama provider path