Two tools with basic schemas and descriptions, but significant gaps in parameter documentation and error guidance. 'agent' tool has a good description (174 chars, within baseline 34-392 range) and complete schema with three parameters, all described. 'glob' tool has minimal description (31 chars, below baseline) and lacks depth on constraints. Neither tool documents output schemas, error recovery paths, or includes actionable error messages. No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Parameter descriptions are present but lack format/range constraints. The 'agent' tool's WRITE risk and worktree fallback logic suggest error handling exists, but it's not surfaced to the LLM via structured error responses.

• Expand glob description to 100-150 chars: 'Find files matching a glob pattern. Returns up to 200 file paths sorted alphabetically. Excludes node_modules, .git, and target directories. Use this to discover source files before reading or editing them.'
• Add output schema documentation: agent returns {agent_id: string, duration_ms: number, output: string, is_error: boolean}; glob returns {matches: string[], total_count: number, truncated: boolean}.
• Add parameter constraints to glob: pattern (required, 1-256 chars, no null bytes, regex-safe), path (optional, 1-4096 chars, must be valid directory).
• Implement structured error responses with recovery hints: 'Pattern cannot be empty. Try: glob(pattern="**/*.rs", path=".")' instead of bare string errors.
• Add tool annotations: agent with destructiveHint=true (spawns sub-agents, may modify files); glob with readOnlyHint=true (no side effects).