Permission-safe prompt refinement and autonomous graph orchestration for Codex, Claude Code, Cursor, GitHub Copilot, and Hermes Agent.
This server has well-structured tool definitions with proper JSON schemas and clear naming conventions. However, parameter descriptions are largely absent or minimal, and output schemas are not documented. Tools follow verb_noun naming (get_, plan_, validate_, run_, start_) which is excellent. The 9 tools are well-scoped and single-purpose. Critical gap: no parameter descriptions visible in the schema definitions, LLMs cannot understand what each parameter controls. Output structures are not documented, forcing agents to guess response shapes. Error handling descriptions exist but lack recovery guidance.
Run the built-in adversarial graph planner, validator, scheduler, gate, and budget evaluation suite.
Run the built-in deterministic safety and intent-preservation evaluation suite.
Read the status and eventual result of a graph job created by start_graph.
Report the active GraphVigil installation identity and whether this host must reload after an upgrade.
Compile a request into a bounded autonomous graph without executing it.
Convert a user request into an intent-preserving, permission-safe execution brief.
Execute a validated short graph synchronously. Prefer start_graph and get_graph_run when execution may approach the MCP client timeout. This MCP surface never approves human gates.
Parameter descriptions missing or absent. Schema shows types (string, object, boolean, enum) but no descriptions explaining what each parameter controls. LLMs cannot infer meaning from names alone, 'prompt' could be a user query or a system instruction; 'mode' is ambiguous without context.
Output schemas not documented. Tools return GraphRunResult, GraphJob, validation objects, etc., but response structures are not visible in the MCP definitions. Agents cannot plan downstream calls or extract required fields.
Error handling lacks recovery guidance. Descriptions mention validation failures and job limits but do not tell the LLM what to do next (retry, ask user, call alternative tool). E.g., 'graph validation failed' should suggest 'Call validate_graph() first to debug.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2025-06-18+ | v2 |
Start a validated graph in the background and immediately return a job ID for polling. Use this for multi-agent or potentially long-running graphs.
Validate graph topology, budgets, permissions, isolation, and convergence.
Enum constraints present but undocumented. 'autonomy' and 'primaryExecutor' parameters use enums (plan_only, read_only, workspace, consequential; codex, claude, local) but descriptions do not explain what each value means or when to use it.
Generic 'graph' parameter type. Tools accept a 'graph' object but do not specify its internal structure (required fields, nested types, constraints). LLMs cannot construct valid graph objects without seeing the schema.