Tool Family: replay-simulator

• Registration source: tools/mcp_tools/replay_simulator.py:22 register_replay_simulator_tools(mcp, *, ctx) — auto-discovered but conditionally registers only when (config["replay_simulator"]["enabled"]: true) (replay_simulator.py:25-28). When false (default), no tools.
• Gate: @audit_tool — local-only dry-run, no target touch, pure simulation. LLM never touches target through this path.

Tools Exported (1) — conditional

Tool	Params	Result Shape	Notes
replay_simulate	plan_json: str, recon_json: str (both JSON strings, must be JSON objects)	render_simulation_result(result) → confidence 0..1, critique text, branch proposals	Validates non-empty else BLOCKED: plan_json is required. / recon_json is required.; json.loads on each else BLOCKED: ... is not valid JSON (...); type check isinstance(..., dict) else BLOCKED: ... must be JSON objects. Tries model_client, model_alias = _get_model_client(config) (best-effort try/except → None on failure); calls simulate(plan, recon, model_client, model_alias) (tools/replay_simulator.py): LLM critiques plan against ReconAssessment, or degrades to rule-based scoring when LLM unavailable. Zero target touch — reads ReconAssessment JSON + plan DAG only.

Dependencies

• tools/replay_simulator.simulate, render_simulation_result
• tools/mcp_tools/registry._get_model_client — router seam for critique LLM
• tools/kernel/audit.make_audit_tool

Config

• replay_simulator.enabled: bool (default false) — registration gate
• models.registry / ollama.host via build_router for critique model; embeddings not used

Auditing

• @audit_tool — plan_json/recon_json not redacted (no secret field), but large payloads are started/completed recorded with truncated not full JSON (audit stores args dict of strings, length capped by caller).

Validation

• JSON parse errors surfaced as BLOCKED with exception text; not an exception propagation.
• Degrades gracefully to rule-based when model_client is None.

Tests

• tests/test_replay_simulator.py — enabled/disabled registration, valid JSON path, invalid JSON BLOCKED, degrade to rules, model client path mock

Related Docs

• docs/mcp/tool-families/assessment-state.md — plan DAG that replay_simulate critiques
• tools/replay_simulator.py — simulation confidence rubric