Synapse has 9 tools with basic schemas and descriptions, but significant gaps prevent higher scores. All tools have descriptions (10-50 chars, mostly short), and input schemas are present with type definitions. However, descriptions are often too terse to guide LLM selection (e.g., 'Full-text search the vault' lacks context on when to use it vs. ask). Output schemas are not documented anywhere, LLMs cannot plan downstream calls. Error handling is absent; no recovery guidance. Parameter descriptions exist but lack constraints (e.g., 'limit' has no min/max). No tool annotations (readOnlyHint, destructiveHint). The 'build' and 'ingest' tools are write operations but lack confirmation/dry-run patterns despite being expensive (LLM calls, file I/O). Tool composition is reasonable (search → get_page → ask forms a chain), but missing IDs in responses would break chaining.

• Document output schemas for all tools. E.g., search() returns [{slug, title, excerpt, score}]; get_page() returns {slug, markdown, created_at, updated_at}. Include field types and descriptions.
• Expand tool descriptions to 50-150 chars. E.g., 'search: Full-text search the vault by keyword. Returns up to N matching pages with titles and excerpts. Use before get_page() to find relevant content.' Include when/why to call it.
• Add parameter constraints: search(limit: 1-100, default 10), ingest(format: enum[auto, markdown, txt, json], min_chars: 1-10000), build(limit: 1-1000).
• Add error handling to tool descriptions. E.g., 'get_page: Returns 404 if slug not found. Try search() first to find valid slugs. Returns 400 if slug contains invalid characters.'
• Add dry-run and confirmation patterns to build() and ingest(). E.g., build(dry_run=true) estimates cost without calling LLM; build(confirmed=true) required if cost > threshold.