A micro agent harness that exposes agents over the network with REST control plane and WebSocket data plane for multi-turn conversations with permission back-channel
Mani has 14 tools with significant definition gaps. Core filesystem tools (read, edit, write, delete, glob, grep) lack visible input schemas in the provided source, only tool names and brief descriptions are evident. The bash tool is particularly risky: no input schema visible, minimal description ('Execute bash commands in the workspace'), and no error guidance for command failures. Delegation and planning tools are underdocumented. Example tools (add, now, echo) show schema structure but are toy implementations. No tool descriptions explain when to use them vs. alternatives, prerequisites, or what happens on failure. Output schemas are not documented. Error handling is absent, no recovery guidance, no categorization of retryable vs. fatal errors. Security concerns: bash tool accepts arbitrary commands with no validation or sandboxing hints. Overall, definitions are skeletal and would not pass production code review.
Adds two integers
Execute bash commands in the workspace
Delegate tasks to subagents with configurable depth
Delete files from the workspace
Echoes the input text back
Edit file contents in the workspace
Fetch content from HTTP/HTTPS URLs with host allowlist support
Find files matching a glob pattern in the workspace
Search for text patterns in files within the workspace
No input schemas visible for 10 of 14 tools (read, edit, write, delete, glob, grep, bash, planning, delegate, fetch, respond). Cannot verify parameter types, required fields, or constraints. Hard cap: schema score = 0 for these tools.
Bash tool has no input schema, minimal description (20 chars), and no error handling or validation guidance. Accepts arbitrary shell commands with no sandboxing hints or safety constraints documented. Critical security and usability gap.
No tool descriptions explain WHEN to use each tool vs. alternatives, WHAT the output contains, or WHAT happens on failure. Descriptions are 20 - 40 chars (below 50-char baseline for LLM optimization). LLMs cannot reliably select the right tool.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
Returns the current date and time
Create and manage task plans for the agent
Read file contents from the workspace
Return the final result according to the schema. Use this tool once, only at the end.
Write file contents to the workspace
Delete tool has no confirmation step, dry-run option, or error recovery guidance. Destructive operations should support confirmation before execution to prevent accidental data loss.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., file paths, IDs, counts). Breaks tool composition and chaining.