Loading ToolBench
Governance for AI-written project docs: verify, catch drift, keep them code-grounded, enforce in CI. OKF-compatible, zero-dependency. Exposes read-only wiki validation, audit, and retrieval tools over MCP.
Strong schema coverage and comprehensive tool set (17 tools, all with explicit JSON schemas and descriptions). Tool names follow verb_noun convention (validate, audit, search_docs, get_doc). Descriptions are detailed (100-300 chars typical) and explain WHAT, WHEN, and context. All parameters have type definitions and descriptions. However, several tools lack output schema documentation, and some parameter descriptions could be more prescriptive about constraints. No tool annotations (readOnlyHint, destructiveHint) despite all tools being read-only. Error handling guidance is absent from tool descriptions.
Report detection, structure, frontmatter, evidence, link, adapter, encoding, and sensitive-info findings, plus the wiki knowledge-graph summary.
Check local runtime, package readiness, project detection, and stable safety-policy signals.
Explain a finding rule (for example wiki_link.missing, frontmatter.required, evidence.stale, okf.type_required) and suggest safe remediation steps.
Retrieve a single wiki document by path (repo-relative, wiki-relative, or bare name). Returns frontmatter, body, and optional section extraction. Supports compact mode and character limits.
Retrieve a document plus its related documents (backlinks, forward links, same-domain docs). Returns the target document and up to 3 related docs with optional section extraction and compact mode.
Emit the wiki knowledge graph (documents plus resolved document-to-document links). structuredContent always carries the full structured graph as JSON, regardless of format. The text content is the human-readable summary for text/json, or the rendered diagram for mermaid (fenced graph TD) and dot (Graphviz).
No output schema documentation. Tool descriptions explain inputs but not return types, fields, or structure. LLMs cannot plan downstream calls or extract data without knowing what fields to expect.
Missing tool annotations. All 17 tools are read-only (no state mutation), but none declare readOnlyHint=true. This prevents clients from optimizing caching, retry logic, and safety policies.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2025-06-18+ | v2 |
Produce the next prompt to run in a coding agent after CLI setup, with project-type-specific source-evidence guidance.
Read-only retrieval. Enumerate wiki documents with metadata (path, title, status, doc_type, visibility, last_updated, tags) and optional status/visibility/docType filters. Returns no bodies. Restricted/sensitive documents are excluded unless includeSensitive is true.
Report the project's effective governance mode (lite/standard/strict), where it came from, the rule severities it contributes, and the lite/standard/strict capability matrix. Read-only here: changing the mode is never available over MCP — only the CLI writes it (llm-wiki mode set <level> --write).
Reuse audit coverage and the wiki graph to recommend prioritized next review/repair/setup actions. Advisory; writes nothing.
Guided read-only orientation: learn the wiki and codebase for a domain or the whole project. Returns a curated reading list with optional learning goal focus.
Guided read-only task preparation: assemble the minimal context needed to make a specific change (feature, fix, refactor, docs-sync, okf-extract). Returns a bounded context bundle with the target doc, related docs, and evidence pointers.
Produce a repeatable agent workflow prompt for a task (bootstrap for first-time wiki enrichment; onboard/prepare for the guided read-only workflows; or feature, fix, refactor, docs-sync, okf-extract).
Full-text search the wiki (AND-semantics, zero-dependency keyword/substring matching). Returns matching documents with snippets, metadata, and optional filters. Restricted/sensitive documents are excluded unless includeSensitive is true.
Read-only health snapshot: total documents, a health score (mean of verified %, enrichment %, evidence coverage %), status mix, and stale-verified / orphan counts.
Report whether LLM-WIKI is initialized, document status counts, missing recommended docs, link health, and selected adapter state.
Run audit-backed structure and safety validation (CI-style). Reports missing required docs, broken markdown links, unresolved [[wiki links]], evidence drift on verified docs, and frontmatter contract issues.
No error recovery guidance in tool descriptions. Tools like search_docs, get_doc, and list_docs do not document what happens on missing documents, invalid filters, or empty results. LLMs lack guidance on next steps.
Parameter constraints under-specified. 'profiles' accepts array of strings with no enum or pattern. 'rule' in explain() has no enum of valid rule IDs. 'task' in prompt() has enum but 'domain' in onboard() is free-form with no guidance.
Pagination not documented for list_docs and search_docs. No mention of limit behavior, default page size, or whether results are truncated. search_docs accepts 'limit' but list_docs does not, inconsistent interface.