Loading ToolBench
Serves the structured cross-repo map to coding agents. Read-only access to the data store (~/.atlas) with four map queries: context, callers, impact, endpoints. Makes no network calls (stdio is local).
Atlas provides 6 well-named, read-only code-analysis tools with clear descriptions (avg 180 chars) and complete JSON Schema input definitions. All tools follow verb_noun naming (atlas_context, atlas_callers, atlas_impact, atlas_endpoints, atlas_path, atlas_neighborhood). Descriptions are specific and actionable, explaining WHAT each tool does and WHEN to use it. However, parameter descriptions are sparse (many lack detail on format/constraints), output schemas are not documented, and error handling is minimal. No tool annotations (readOnlyHint, idempotentHint) despite all being read-only. No pagination guidance for potentially large result sets.
Direct callers of a symbol (reverse call edges). Use to gauge the immediate blast radius of changing a function.
Focused context pack for a symbol or file: the target plus its direct callers and callees. Use when orienting in unfamiliar code or before editing a function.
The workspace HTTP surface: resolved cross-repo links (FE consume ↔ BE expose) and external endpoints (consumed but not exposed by any repo here). Use for cross-repo / backend-dependency questions.
What a change affects: all transitive intra-repo callers, plus cross-repo consumers of any endpoint the target handles. Use before refactors, renames, or deletes.
The local call subgraph around a symbol (its callers + callees to a given depth) in one bounded call — use to understand 'what's around this' without separate context/impact queries.
Shortest connection between two symbols/files across the whole workspace — spanning call edges and cross-repo HTTP contracts. Use to answer 'how does A reach B?' in one query instead of reading files.
Output schemas not documented. LLMs cannot plan downstream calls or extract fields without knowing what each tool returns. All 6 tools lack return type documentation.
Parameter descriptions incomplete. 'query', 'symbol', 'target', 'from', 'to' lack format guidance (e.g., 'A symbol name, file path, or full node id' is stated in tool description but not in parameter schema). 'workspace' and 'repo' descriptions are generic. LLMs cannot infer valid input formats.
No tool annotations despite all tools being read-only and idempotent. Missing readOnlyHint and idempotentHint in tool definitions. Agents cannot infer safety properties without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
No pagination or result-limit guidance. Tools like atlas_endpoints and atlas_impact may return large result sets (all transitive callers, all cross-repo consumers). No limit, offset, or cursor parameters documented. Risk of context window exhaustion.
Error handling minimal. catch block returns raw error messages without recovery guidance. LLMs receive 'Symbol not found' but no hint to try search or alternative queries. No error classification (retryable vs fatal).