Local agency/runtime harness that turns ChatGPT Web into a repository- and computer-capable development agent.
Critical gaps across all dimensions. Tool names lack action verbs (relai_work, relai_exec, relai_search, relai_read, relai_write, snapshot). Descriptions are present but generic (avg ~80 chars, below 194-char baseline). Input schemas visible but lack parameter descriptions, most params have only type and description fields with minimal context. No output schemas documented. No error handling guidance. No parameter constraints (enums, ranges, formats). Security concerns: relai_exec exposes shell command execution without validation or permission gates. No idempotency markers. Tool composition is poor: relai_work and relai_exec both perform actions but lack clear separation of concerns.
Execute shell commands and scripts in project context
Read file contents from project repositories
Search for files and content within project repositories
Primary work orchestration tool for managing coding tasks and project operations
Write or modify files in project repositories
Capture current state snapshot of project files and structure
Tool names lack action verbs. 'relai_work', 'relai_exec', 'snapshot' do not follow verb_noun convention. LLMs cannot infer intent from names alone. Should be: 'begin_work', 'execute_command', 'capture_snapshot'.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. Must document return types, field names, and structure for all 6 tools.
relai_exec exposes arbitrary shell command execution without validation, sanitization, or permission gates. Critical security vulnerability: LLMs can be tricked into command injection. Requires input validation, allowlist of safe commands, and permission checks.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2025-06-18+ | v2 |
Parameter descriptions are minimal (avg ~20 chars). 'Shell interpreter to use' lacks format/constraint info. 'Action to perform' does not explain consequences. Descriptions must be 50-200 chars and include format, range, and valid values.
No error handling guidance. Tools do not indicate which errors are retryable, user-fixable, or fatal. No recovery suggestions. Must add error classification and actionable recovery paths.