Loading ToolBench

Reading the current corpus…

codemap MCP server · D · ToolBench

codemap

Local codebase intelligence for AI agents — a SQLite index of symbols, imports, and calls. SQL and recipes instead of scanning the tree.

2026-07-28+STDIO21 tools

57/100

Definition Quality65

Protocol Readiness50

Supportability49

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

Codemap exposes 21 tools with explicit schemas and descriptions. Most tools have clear verb-noun naming (query, audit, validate, trace, explore). However, descriptions vary significantly in quality and depth. Many tools lack sufficient detail about when to use them vs. alternatives, expected output structure, and error recovery paths. Parameter descriptions are present but often generic. Several tools (audit, context, validate, show, snippet, impact, affected, trace, explore, node) have minimal descriptions (under 50 chars) that fail to guide LLM selection. Output schemas are not documented in the tool definitions themselves, LLMs must infer structure from descriptions alone. Error handling guidance is absent across all tools.

Tools (21)

affectedread onlysource verified62/100

Find tests affected by code changes.

applydestructivesource verified63/100

Apply code transformations from a recipe or diff.

apply_diff_inputdestructivesource verified62/100

Apply transformations from a diff input.

apply_rowsdestructivesource verified65/100

Apply transformations to specific rows from a query result.

auditread onlysource verified60/100

Audit the codebase for issues. Supports baseline comparison and various output formats.

contextread onlysource verified52/100

Build a context envelope for the codebase suitable for LLM consumption.

drop_baselinedestructivesource verified73/100

Delete a named baseline.

exploreread only

Issues (5)

HIGHtool-description ↗

Minimal descriptions on 9 tools (audit, context, validate, show, impact, affected, trace, explore, node), under 50 chars, insufficient for LLM tool selection.

auditcontextvalidateshowimpactaffectedtraceexplorenode

HIGHresponse-shaper ↗

No output schemas documented in tool definitions. LLMs cannot infer what fields to expect from query results, audit output, or graph traversal responses. Agents must guess structure, risking failed downstream chaining.

queryquery_batchquery_recipeauditcontexttraceexplorenode

HIGHrecovery-guide ↗

No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. E.g., 'query' with invalid SQL returns an error, but the description does not explain recovery.

Recommendations

Expand descriptions for audit, context, validate, show, impact, affected, trace, explore, node to 100 - 200 chars. Include: what the tool does, when to call it instead of alternatives, and what structure it returns. E.g., 'audit: Scan the codebase for issues (e.g., unused symbols, type errors). Returns an array of {file, line, issue_type, message}. Use after code changes to catch regressions.'
Document output schemas for all tools. Add a 'Returns' section to each description listing key fields and types. E.g., 'query returns {rows: [{file_path, line_start, symbol, kind}], count: number, summary?: {total, by_kind}}'. This enables LLMs to plan downstream chaining.
Add error handling guidance to descriptions of query, apply, and ingest_* tools. E.g., 'If SQL is invalid, returns {error: string}. Retry with corrected SQL. If database is locked, wait 1s and retry.' This teaches LLMs recovery strategies.
Implement dry-run or confirmation for destructive tools (apply, apply_rows, apply_diff_input, drop_baseline). Add an optional 'dry_run: boolean' parameter (default false) that previews changes without committing. Document in descriptions: 'Pass dry_run=true to preview changes before applying.'
Add parameter constraints and examples to context, show, impact, affected, trace, explore. E.g., 'focus: Array of file paths (relative to repo root, e.g. ["src/main.ts", "lib/utils.ts"]) or symbol names (e.g. ["MyClass", "myFunction"]). At least one required.' This prevents LLM guessing.

Score history

Overall score trend

First recorded score · v2 rubric

57/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	57	2026-07-28+	v2