Loading ToolBench
Terminals convergence engine: an offline MCP server (pure Python standard library). Brings messy sets of ideas together into coherent answers with checkable receipts using the Fano plane (7-grid) and Kuramoto phase-locking dynamics.
Six tools with complete input schemas and descriptions. Naming is domain-specific (converge, explore, optimize, recommend, frame, hold) but lacks action-verb clarity, these are nouns/verbs that don't immediately signal what happens. Descriptions are detailed (150 - 250 chars) and explain purpose well. All parameters have types and descriptions. Output schemas are NOT documented, LLMs cannot predict return structure. Error handling is absent from visible code. No parameter validation guidance (e.g., coherence matrix must be NxN, ideas capped at 7). Tool composition is sound (each does one thing), but missing idempotency and recovery guidance.
Bring a messy set of ideas together into one coherent answer with a checkable receipt. Pass the ideas and, ideally, a coherence matrix (how much each pair agrees, 0..1). Returns the locked trios, the agree-meter r, the sure-o-meter phi, done=true or partial, and the witness: a certificate that the answer holds together under the coherence you supplied. It shows internal consistency. It does not prove the decision is correct in the world.
Open a question into its full fan of distinct angles (high entropy, on purpose). Returns 7 spread points and a divergence score. With no ideas yet, returns 7 empty slots to fill.
Translate the user's own context (notes/files already gathered) onto the 7-grid, then converge. Pass the items as a list of strings.
Re-check a settled answer against new state (converge with memory). Pass the prior converge result plus the updated ideas and coherence. Returns which locked trios came loose (drift), which still hold, the new ones, and the fresh verdict, so you fix the drift instead of re-deciding from scratch.
Take an already-converged result and find its cheapest, cleanest equal form without breaking done=true. Pass the whole object that converge returned.
Output schemas not documented. LLMs cannot predict return structure (locked trios, r, phi, witness, done flag). Agents cannot plan downstream calls or extract fields reliably.
Tool names lack action verbs. 'converge', 'explore', 'optimize' are domain-specific but don't signal intent as clearly as 'run_convergence', 'analyze_divergence', 'refine_solution'. LLMs may struggle to select the right tool when many are available.
No parameter validation guidance. 'ideas' capped at 7 per description, but no minItems/maxItems in schema. 'coherence' must be NxN matrix, but no format constraint. LLMs will pass invalid inputs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Run the whole loop for the user: converge, then optimize, then hand back a pick plus the witness and alternatives. Pass the agent's own best lines as ideas and, ideally, a coherence matrix.
No error handling or recovery guidance in visible code. If coherence matrix is malformed or ideas exceed 7, no actionable error message. Agents cannot self-correct.
No idempotency or confirmation pattern for tools that modify reasoning state. If 'converge' is called twice with same input, does it return identical witness? No guidance on retry safety.