SuperAgent exposes 4 tools with STDIO transport only, limiting remote accessibility. Tool naming lacks action verbs (codex, gemini, continue, list-agents are acceptable but generic). Descriptions are present and reasonably detailed (100-200 chars), which is above baseline. However, schema quality is inconsistent: schemas ARE defined via Zod but parameter descriptions are sparse or missing for critical fields. All tools accept dynamic agent names but provide no enum constraints or validation examples. The codex, gemini, and continue tools share identical or nearly-identical schemas with copy-pasted descriptions, signaling a composition problem, these should either be consolidated or differentiated. Error handling is present (status='ok'|'failed') but does not guide recovery ('User not found, try search_users()' style guidance is absent). Security concerns: the 'extraArgs' parameter accepts arbitrary CLI arguments, creating potential command injection vectors. No evidence of rate limiting, permission gating, or audit logging. Output is formatted as plain text with ANSI colors rather than structured JSON, forcing LLMs to parse unstructured responses. Schema definitions use zod-to-json-schema conversion (correct approach) but descriptions are embedded in schema generation, making them harder to audit.

• Consolidate codex, gemini, continue into a single 'run-agent' tool with an 'agent_backend' enum parameter (codex|gemini|continue). This reduces naming confusion and gives LLMs one canonical interface.
• Add detailed descriptions for every parameter: agent (enum of valid agent names), prompt (max length, expected format), extraArgs (warn: use only for safe flags; no shell metacharacters), timeoutMs (min 1000, max 3600000), workingDirectory (must be within allowed paths, resolved relative to server root).
• Implement input validation: reject extraArgs containing pipes, semicolons, backticks, or other shell metacharacters. Return 'Invalid argument: extraArgs cannot contain shell metacharacters. Use: [--flag, --key=value]'.
• Change output format from ANSI-colored plain text to structured JSON: {"results": [{"taskIndex": 0, "agent": "name", "status": "ok"|"error", "durationMs": 123, "response": "...", "errorCode": "..."}]}. This allows LLMs to reliably extract and chain results.
• Add error recovery guidance to failed tool calls: e.g., 'Task failed: timeout after 30s. Retry with higher timeoutMs (current: 30000), or check that workingDirectory exists and agent is installed.'
• Implement rate limiting: cap concurrency to 1-10 per request and 100 requests per minute server-wide. Return 429 with backoff guidance when exceeded.
• Add a 'confirm_before_execute' mode for destructive agent operations (those that modify files, delete data, etc.). Return result with status 'input_required' asking user for confirmation before proceeding.