Loading ToolBench
Bun-first compiler-backed FPF spec runtime exposed through hosted Vercel and local MCP surfaces with a Hono server engine.
Strong schema definitions and descriptions across 11 tools. All tools have clear verb-noun naming (query_, ask_, inspect_, read_, browse_, search_, trace_, expand_, refresh_, get_). Input schemas are well-structured with types, enums, and maxLength constraints. Descriptions are detailed (100-200 chars typical). However, output schemas are not explicitly documented in the visible code, only structured content is mentioned in tool-text.ts renderers. Error handling guidance is minimal. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) detected. Risk classifications are present (WRITE, READ_ONLY) but not exposed in MCP schema.
Return an FPF answer in markdown with grounding metadata using the compiler-backed vectorless runtime.
Browse the FPF catalog of compiled patterns, routes, and lexicon entries. Filter by part, status, or kind to discover relevant material before drilling into individual nodes. Results are paged: pass the previous response's `nextOffset` as `offset` to fetch the next page.
Expand multiple exact citation IDs into raw anchor text plus owning node context without adding new semantics.
Inspect whether the current FPF runtime index exists, whether it is fresh against the current source hash, and which artifacts are present.
Inspect one compiled FPF anchor by exact anchor ID and return raw anchor text plus owning node context.
Inspect one compiled FPF node by exact ID, route name, or lexeme and return anchors plus neighboring relations.
Output schemas not documented in visible code. Tool-text.ts shows rendering logic but no formal schema definitions for structured responses. LLMs cannot plan downstream tool calls without knowing response field types.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) in MCP schema. Risk classifications exist in code (WRITE, READ_ONLY) but are not exposed to clients. Agents cannot distinguish safe vs destructive operations.
Error handling lacks recovery guidance. No visible error responses that tell LLMs what to do next (e.g., 'Try search_fpf() with a simpler query'). Errors should be categorized as retryable, user-fixable, or fatal.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 70 | 2025-06-18+ | v2 |
Answer questions against the compiler-backed vectorless FPF runtime with auditable IDs, citations, constraints, and freshness metadata.
Resolve one FPF selector to the canonical generated markdown page and return exact text plus stable markdown/static paths. Pass `mode: "preview"` for headings + size + a short snippet without the full markdown body.
Build or rebuild the compiler-backed vectorless FPF index from the configured spec source and persist the artifact set.
Full-text search across all compiled FPF nodes. Returns ranked hits with contextual snippets. Use this to find patterns, routes, or lexicon entries by keyword or concept.
Return the deterministic retrieval trace showing normalization, candidate scores, graph expansion, and selected slices.
Parameter descriptions lack explicit format/constraint guidance. E.g., 'selector' accepts 'exact ID, route name, or lexeme' but no examples of valid formats. 'kind' enum is documented but selector format is implicit.
No pagination guidance in browse_fpf_catalog description. While offset/limit/nextOffset are present, the description does not explain the pagination contract or when to stop iterating.