Loading ToolBench
MCP server for recursive deep reasoning — architecture analysis, code review, debugging, incident triage, and brainstorming
Server has 7 tools with complete schemas and descriptions. Naming follows verb_noun pattern (deep_scan, deep_explore, deep_synthesize, deep_list_sessions, deep_resume, deep_save_session, deep_delete_session). All tools have detailed descriptions (100-300 chars) and comprehensive input schemas with types, enums, and constraints. However, output schemas are not documented, responses are JSON strings without declared structure. Error handling is minimal; no recovery guidance or error categorization. Tool composition is reasonable but deep_explore and deep_synthesize lack explicit output field documentation, forcing LLMs to infer structure from examples.
Deletes a saved reasoning session by session ID (exact match required).
Explore specific branches of the reasoning tree with recursive what-if chains. Takes the tree from deep_scan (or a previous deep_explore) and digs deeper into specific findings. Can be called multiple times to progressively deepen the analysis. In orchestrator mode, returns structured prompts for Claude Code to reason through. In autonomous mode, calls the API and returns updated tree with results.
Lists saved reasoning sessions with optional filtering by analysis type, status, and limit.
Resumes a saved reasoning session by session ID (supports partial matching).
Manually saves a reasoning session to disk.
Perform initial reconnaissance on a problem, code, or architecture. Maps critical areas, identifies key findings as IF/THEN chains, and may ask clarifying questions if the context is ambiguous. This is always the FIRST tool to call in a deep analysis workflow. Workflow: deep_scan -> deep_explore (repeat) -> deep_synthesize.
Output schemas not documented. deep_scan, deep_explore, and deep_synthesize return JSON strings without declared field structure. LLMs cannot plan downstream operations or extract specific fields reliably.
No error handling guidance. Tools return raw error objects without recovery hints. E.g., 'Invalid session_id' should suggest 'Call deep_list_sessions() to find valid session IDs.'
Destructive operations (deep_delete_session) lack confirmation/dry-run pattern. Agents can permanently delete sessions without explicit confirmation, risking data loss.
deep_resume accepts partial session_id match but deep_delete_session requires exact match. Inconsistent behavior across similar tools increases LLM confusion and error rates.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
Synthesize the completed reasoning tree into a final report. Produces executive summary, action plan, residual risks, decision log, and quick wins.
Parameter descriptions lack format/constraint details. E.g., 'session_id' (minLength=1, maxLength=200) should state format in description: 'Session ID (1-200 chars, alphanumeric + hyphens)' to guide LLM input generation.