The 'conflicts' tool has a well-structured input schema with proper types and a clear, domain-specific description. However, the tool name 'conflicts' is a noun, not a verb-based action name, which violates the Agentic Tool Pattern for naming. The description is adequate at ~240 chars but could be optimized for LLM clarity. Output schema is not documented in the tool definition, forcing LLMs to infer the response structure from the description alone. Parameter descriptions are present and reasonably detailed, but lack validation guidance (e.g., file path format constraints, context_lines range). No error handling guidance is provided in the description, LLMs have no recovery path if a file cannot be read or conflict parsing fails. Security is adequate for a read-only tool but could benefit from explicit permission declaration. Overall, this is a competent but not production-grade tool definition.

• Rename tool from 'conflicts' to 'parse_conflicts' to align with verb-based action naming convention (pattern:tool). This makes intent clear to LLMs and avoids ambiguity with the 'conflicts' noun.
• Document the complete output schema in the tool definition. Include a description like: 'Returns a Result object with fields: files (array of FileResult), total (int, count of conflicts), has_diff3 (bool, whether any conflicts use diff3 format), summary (string, summary message). Each FileResult contains: file (string, path), conflicts (array of Conflict objects), count (int). Each Conflict contains: file, line, end_line, ours_ref, theirs_ref, ours, theirs, base (optional), context_above (optional), context_below (optional).'
• Fix schema/description mismatch for 'file' parameter. Either update the InputSchema.Properties['file'].Type to accept both 'string' and 'array' (e.g., oneOf), or remove the phrase 'Can also accept a single string' from the description and require callers to always pass an array.
• Add validation constraints to context_lines description: 'Number of context lines (0 - 100, defaults to 1). 0 returns no context; values >100 are clamped to 100 to prevent excessive output.'
• Enhance error handling in tool description: 'If a file cannot be read (not found, permission denied, or unreadable), the tool returns an error. Verify file paths are absolute or relative to the working directory. Files must be readable by the process.'
• Add permission hints: 'This is a read-only tool. It requires file read access but does not modify files.' This signals to LLMs that the tool is safe to call and invokes no side effects.