Autonomous AI agent for console with advanced tool use made by another agent
Two tools with partial schema visibility. search_codebase has a complete, well-described schema with proper types and defaults (query: string, top_k: integer with default=5). github tool has a flat schema with operation as a discriminator but lacks conditional parameter documentation, many parameters (pr_number, issue_number, etc.) are optional and their applicability depends on the operation value, which is undocumented. Both descriptions are adequate (100+ chars) but lack actionable guidance on when to use each tool or how they compose. No output schemas documented. No error handling guidance visible. Tool naming follows verb_noun convention (search_codebase, github) but 'github' is generic, operation-specific names (list_prs, create_issue) would be clearer. No tool annotations (readOnlyHint, destructiveHint) despite github being a WRITE tool.
Perform GitHub operations like listing PRs, creating issues, etc.
Search the workspace codebase using BM25 to find relevant files. Works across Python, JS/TS, Go, Rust, Java, C/C++, and more. Understands camelCase and snake_case. Returns paths, relevance scores, and snippets.
github tool uses operation discriminator without documenting parameter dependencies. Parameters like pr_number, issue_number, title, body, head, base are conditionally required based on operation value, but descriptions do not state which operation requires which parameters.
No output schemas documented for either tool. LLMs cannot plan downstream calls or extract required fields (e.g., PR IDs, issue IDs) without knowing the response structure.
github tool lacks tool annotations. It is marked WRITE risk but has no destructiveHint or idempotentHint to signal to agents which operations are safe to retry.
github tool name is generic. 'github' does not convey the specific action. Operation-specific tool names (list_prs, create_issue, add_comment) or a more descriptive name (github_api) would improve LLM tool selection clarity.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
No error handling guidance visible. Tools do not document what errors are retryable, what user-fixable errors look like, or how to recover from failures (e.g., 'GitHub API rate limit exceeded, wait 60 seconds and retry').