Larch provides 4 read-only tools with basic schemas and descriptions. All tools have input schemas with typed parameters and descriptions present. However, descriptions are minimal (10-50 chars), lacking context on WHEN to use each tool or dependencies between them. Parameter descriptions are present but terse. No output schemas documented. Error handling is absent, no guidance on recovery or retryable vs. fatal errors. Tool naming is verb-based (search, document, tree, tags) but lacks clarity on distinctions (e.g., 'tags' could mean list all tags or filter by tag). No tool annotations (readOnlyHint, idempotentHint). Composition is reasonable, tools are single-purpose, but output structure is not documented, forcing LLMs to infer response shape.

• Expand tool descriptions to 100-200 chars. Example: 'Search the Markdown knowledge base by keywords. Returns up to 10 matching documents with snippets. Use this to find relevant notes before reading full documents with the document tool. Supports filtering by tag or directory.'
• Document output schemas in tool definitions. Declare that search returns [{path, title, snippet, tags, score}], document returns {path, start_line, end_line, content}, tree returns {directories, files}, tags returns {tags} or {tag, documents}.
• Add error handling guidance. E.g., 'If search returns no results, try broader keywords or omit the tag filter. If document path is invalid, call tree first to list available files.'
• Disambiguate 'tag' parameter. In search, rename to 'filter_tag' with description 'Optional tag to narrow results'. In tags, keep 'tag' but clarify: 'Optional tag name. If omitted, returns all tags; if provided, returns documents with that tag.'
• Add tool annotations: set readOnlyHint=true for all tools (they only read the vault, never modify). This signals to LLMs that these tools are safe to call speculatively and can be retried without side effects.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry