Loading ToolBench
MCP server implementing USD-inspired LIVRPS composition arcs as a formal argumentation framework for AI critical thinking
Cognitive Bridge demonstrates strong naming conventions (verb_noun pattern: cb_manage_*, cb_decide, cb_probe_user) and comprehensive parameter schemas with type definitions and descriptions. Tool descriptions are detailed and context-aware, explaining WHEN and WHY to use each tool. However, output schemas are not formally documented in the source code, responses are formatted as plain text strings rather than structured JSON objects. Error handling is implicit (validation gates mentioned but not visible in code excerpts). The server implements tool annotations (readOnlyHint, destructiveHint, idempotentHint) correctly. Parameter descriptions are thorough (e.g., arc strength enums, falsifiable_if requirements), but output structure documentation is missing, which limits LLM ability to parse and chain results.
CRITICAL: Record a project decision with FULL accountability. Requires naming alternatives rejected and second-order effects. Second-order effects become INHERITS assertions at the decision path, permanently encoding constraints in the composition stage.
CRITICAL: Use this tool to permanently record a structural decision, verified fact, or domain constraint. Do NOT rely on conversational memory. Manages assertion lifecycle: assert, promote, retract, falsify. Routes through resolver engine with full conflict detection pipeline.
CRITICAL: Use this tool for every conflict lifecycle event. Do NOT ignore active conflicts. Manages conflict resolution, challenges, deferrals, creation, and experiment proposals. Enforces steelman gate for challenges and experiment gate for proposals.
CRITICAL: Surface pending PAYLOADS (known unknowns) at or below a path. PAYLOADS are assertions at arc=50 representing evidence that EXISTS in the world but has not yet been loaded into the composition stage. Read-only tool that never modifies the stage.
IMPORTANT: Record an observation about the user's cognitive style to calibrate the argumentation protocol. Updates the Cognitive Operating Signature (COS) kernel which governs conflict surfacing sensitivity, coworker posture, and epistemic pressure. Probe values are smoothed via exponential moving average (alpha=0.7).
Output schemas not formally documented. Tools return plain-text formatted strings (e.g., _format_resolution_result) rather than structured JSON. LLMs cannot reliably parse multi-line text responses or extract fields for downstream tool chaining.
Error handling strategy not visible in source. Code mentions 'surfaces those errors as error strings' but no explicit error classification (retryable vs user-fixable vs fatal) or recovery guidance is shown. Validation gates exist but error responses are not documented.
Parameter dependencies underdocumented. cb_manage_assertion requires falsifiable_if only for arc=10 (LOCAL), but this conditional constraint is buried in the description. cb_manage_conflict requires different params for different actions (steelman_summary for 'challenge', experiment_protocol for 'propose_experiment'), but mutual exclusivity is not explicitly stated.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
ALWAYS call this tool when the argumentation protocol needs calibration. Exposes all CognitiveParameters knobs for runtime tuning. Call with no arguments to inspect current settings. Every change is recorded as a PARAMETERS_TUNED event in the audit trail.
No pagination or result limits documented. cb_payload_check and cb_manage_conflict may return unbounded lists of conflicts/payloads. No mention of limit, offset, or next_cursor parameters to prevent context window exhaustion.
Tool composition assumes prior state. cb_manage_conflict requires conflict_id, but no discovery tool (list_conflicts, search_conflicts) is provided. Agents cannot find conflict IDs without calling cb_manage_assertion first and parsing text output.