Forensic deep-dive of any codebase (9 languages). Persistent code knowledge graph + MCP server (9 composite tools incl. hybrid NL query, Mermaid export & cross-stack trace) for AI coding agents, with 5 durable markdown artifacts (MAP, HOTPATHS, ARCHAEOLOGY, MENTAL_MODEL, AGENT_BRIEF) as the human-readable projection.
Server has 6 tools with mixed quality. Two domain tools (impact, context) have detailed descriptions and proper schemas; four fixture/sample tools (add, get_weather, get_info, orphan_tool) are test stubs with minimal or missing descriptions. Naming is inconsistent: 'impact' and 'context' are nouns (should be verbs like 'analyze_impact', 'get_context'); 'add' lacks domain context; 'orphan_tool' is self-referential. Parameter descriptions vary widely, impact has good detail, but get_info has zero parameters and no output schema documented. No tool declares output structure. Error handling and recovery guidance absent across all tools.
Add two integers
One-tool-call 360° view of a symbol. PRD §4.5 tool 3 (DEC-016). Returns the symbol's definition + callers + callees + members + siblings + recent commits + dominant author. Single composite call replaces the typical 5-6 narrower queries an agent would otherwise chain.
Get info
Get weather for a city
Blast-radius analysis for *symbol*. PRD §4.5 tool 1 (DEC-016). * ``direction="upstream"`` walks INCOMING CALLS (who calls this?). * ``direction="downstream"`` walks OUTGOING CALLS (what does this call?). Returns ``{matches, depth_buckets, summary}`` where ``depth_buckets`` is a list of lists — bucket ``i`` is the set of symbols ``i+1`` hops away. Each entry carries the resolved confidence.
A located provider with no consumer — HANDLES, no ROUTES_TO
Tool names are nouns or generic verbs, not action-oriented. 'impact' should be 'analyze_impact'; 'context' should be 'get_context'; 'add' lacks domain context; 'orphan_tool' is self-referential and unclear. LLMs rely on verb-noun naming to infer intent.
Four of six tools (add, get_weather, get_info, orphan_tool) have minimal or missing descriptions (<30 chars or empty). LLMs cannot determine when to select these tools without clear, actionable descriptions (target 50-200 chars).
No tool documents its output schema or return type. LLMs cannot plan downstream tool calls or extract required fields without knowing what structure to expect. All tools should declare return types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
get_info and orphan_tool have empty input schemas (no parameters). This is valid but unusual; descriptions should explain what these tools return and when to call them. Currently opaque.
No error handling or recovery guidance in any tool. If impact() fails to find a symbol or context() hits a database error, the LLM has no guidance on what to do next (retry, ask user, abort). Add error classification and actionable recovery hints.