Graph-native MCP/tool registry, policy compiler, and impact analyzer
Toolgraph exposes 13 tools with explicit schemas and descriptions. Most tools (9/13) are READ_ONLY governance/analysis queries with clear, domain-specific purposes. Naming is verb-forward and unambiguous (check_access, blast_radius, unmapped_tools). Descriptions are detailed (100 - 300 chars) and explain WHAT the tool does and WHEN to use it. However, 3 tools (read_note, publish_note, read_file, write_file) are example/demo tools with minimal descriptions (20 - 40 chars), and parameter descriptions are sparse or missing for several tools. Input schemas are present but lack explicit type constraints (enums, min/max). Error handling guidance is absent, tools return verdicts/counts but no recovery hints for edge cases.
Authored WRITES edges on tools hinting ``readOnlyHint: true``.
Agents/tools affected if `node` (a resource URI or a policy id) changes.
Can `agent` call `tool`, and does the call path cross a DENY policy? Returns a verdict (AGENT_NOT_FOUND | TOOL_NOT_FOUND | AMBIGUOUS_TOOL | NOT_GRANTED | DENY | ALLOW) with the exact resource/policy evidence path. Every response carries ``agent_found: bool`` and ``found`` for the tool (True/False when resolution ran; null on AGENT_NOT_FOUND). On DENY, the response also carries ``all_authorized: bool`` — True when every deny_evidence row is an operator-declared exception.
Tools hinting ``destructiveHint: true`` with no GOVERNED_BY path — neither directly on the tool nor via any resource it reads/writes. The hint is the server's self-claim (``crawled``/``medium``, ADR-0006): these rows are an authoring priority queue, not violations. Annotations never auto-create edges; the operator authors the policy binding.
Tools with authored governance but no live EXPOSES edge — re-ingest needed.
Demo/example tools (read_note, publish_note, read_file, write_file) have minimal descriptions (<40 chars). Descriptions must be 50 - 200 chars to guide LLM selection and explain when to use the tool.
Parameter descriptions are missing or generic. E.g., 'agent' in check_access lacks context; 'note_id' in read_note has no format hint. Every parameter needs a description explaining what it controls and expected format.
No error handling guidance. Tools return verdicts (AGENT_NOT_FOUND, TOOL_NOT_FOUND, DENY) but do not explain what the LLM should do next. E.g., on AGENT_NOT_FOUND, should the agent retry with a different name or ask the user?
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
Policies no agent currently reaches — likely obsolete or over-narrow.
Publish a demo note to the governed draft area.
Read a file.
Read a deterministic demo note.
Authored edges with no usable evidence pointer. Default scope is READS/WRITES/GOVERNED_BY — load-bearing claims about the world. Pass ``include_grants=True`` to also surface CAN_CALL grants (intent declarations, usually noise unless your policy requires every grant to cite a change-request / runbook).
Granted tools (or all tools if `only_granted=False`) with no authored READS/WRITES — data-flow effects unclassified.
Tools `agent` is allowed to call that reach a DENY policy. Returns an object so the MCP structured-content shape is stable: ``{"agent", "agent_found", "count", "tools": [...]}``. Defaults to ``classification='violation'`` rows only — matching the CLI semantics so MCP consumers don't see authorized_but_governed rows mixed in by default. Pass ``include_authorized=True`` for the full list. ``agent_found=False`` distinguishes a typo'd agent (returns empty) from an agent with zero unsafe tools (also returns empty).
Write a file.
Input schemas lack constraints. 'agent' and 'tool' parameters are bare strings with no enum, pattern, or length limits. Enums or regex patterns would prevent hallucinated invalid values.
Output schemas not documented in tool descriptions. Tools return complex objects (e.g., check_access returns agent_found, found, all_authorized, verdict) but LLMs cannot infer the structure. Document return types explicitly.