A Python-based autonomous agent framework with tool execution, LLM integration, and desktop automation capabilities
OpenWand exposes 11 tools with basic schemas and minimal descriptions. All tools have input schemas with typed parameters, but descriptions are extremely terse (5-20 chars), violating the 10-1024 char baseline and failing to guide LLM selection. No output schemas documented. Parameter descriptions are present but generic. No error handling guidance, no idempotency markers, no permission gates. Tools like run_command and delete_file are high-risk but lack confirmation patterns or dry-run support. Naming is verb-first (good), but composition issues exist: create_file and write_file overlap; patch_file and edit_file are near-duplicates. No pagination support on list_files despite risk of large results.
Capture a screenshot of the current screen.
Create file.
Create a binary file from already-decoded bytes.
Delete file.
Edit a file by replacing one exact text block.
List files.
Handle patch file for agent toolbox.
Read file.
All tool descriptions are 5-25 characters, far below the 10-1024 char baseline. Descriptions like 'List files.' and 'Read file.' provide zero context for LLM selection or when to use the tool vs alternatives.
No output schemas documented. LLMs cannot plan downstream calls or extract required fields (e.g., does list_files return file sizes, modification times, or just names?).
Destructive tools (delete_file, run_command) lack confirmation patterns, dry-run support, or error recovery guidance. Agents can irreversibly delete files or execute arbitrary commands without safeguards.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
Run command.
Write file.
Create or replace a binary file without altering its bytes.
Overlapping tool responsibilities: create_file vs write_file (both write content); edit_file vs patch_file (both replace text blocks). LLMs waste reasoning cycles choosing between near-duplicates.
list_files accepts a 'limit' parameter but no pagination (offset/cursor) or total count. Large directories will be truncated silently, and agents cannot iterate through results.