Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
This server has 3 tools with basic schema definitions and descriptions, but falls short of production quality. Tool names are clear and verb-prefixed (atlas_query, atlas_execute, atlas_status), which is good. However, descriptions are thin (8-10 words), parameters lack depth in specification, and output schemas are not formally documented. The 'context' parameter in atlas_query is typed as 'object' with no properties defined or constraints specified. The atlas_execute tool's 'args' array has no item count limits or validation guidance. Most critically, no tool declares what fields will be returned, forcing LLMs to infer output structure. Error handling exists (try/catch blocks) but error responses are generic JSON stringifications of API responses rather than actionable recovery guidance. The server implements basic logging but no tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite having tools with clearly different risk profiles. Per the rubric baseline, average tool description length should be 194 chars (p10=34, p90=392), these descriptions average ~40 chars, at the bare minimum. Average of per-tool scores: (48 + 52 + 56) / 3 = 52.
Tools (3)
atlas_executewriteauthsource verified52/100
Execute a command on the Atlas Monolith Agent
atlas_queryread onlyauthsource verified48/100
Query the Atlas Monolith Agent via Tailscale network
atlas_statusread onlyauthsource verified56/100
Get the current status of the Atlas Monolith Agent
Descriptions are 30-54 characters, well below the 194-char median for production tools. Lack context on WHEN to use each tool, dependencies, or return structure.
atlas_execute is a WRITE tool (destructive operations) but has no confirmation/dry-run step, no explicit side-effect warning in description, and no destructiveHint tool annotation.
'context' parameter in atlas_query is typed as object with no properties defined, constraints, or examples. LLM cannot infer what keys to pass.
atlas_query
Recommendations
Expand tool descriptions to 150-200 chars each. For atlas_query: 'Query the Atlas Monolith Agent via the Tailscale network. Accepts a natural-language query and optional context object (keys: project, scope, filters). Use this for read-only information retrieval; use atlas_execute for commands. Returns JSON-structured response with results, metadata, and execution time.'
Add explicit output schema for each tool in the handleListTools response. Example for atlas_status: 'outputSchema: {type: "object", properties: {status: {type: "string", enum: ["running", "stopped", "error"]}, uptime_seconds: {type: "number"}, version: {type: "string"}, healthy: {type: "boolean"}}, required: ["status", "healthy"]}'
Expand 'context' parameter schema in atlas_query. Example: 'context: {type: "object", properties: {project_id: {type: "string", description: "Project identifier"}, scope: {type: "string", enum: ["global", "project", "team"], description: "Scope of the query"}, filters: {type: "array", items: {type: "string"}, description: "Optional filter expressions"}}, description: "Optional context object to narrow query scope. All fields optional."}'
Add destructiveHint annotation to atlas_execute and update its description to mention side effects: 'Execute a command on the Atlas Monolith Agent. This tool MODIFIES STATE and has irreversible consequences. Commands may create resources, delete data, or trigger system-wide changes. Typical uses: run scripts, update configurations, trigger workflows.'
Add readOnlyHint annotation to atlas_query and atlas_status.
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
Error responses return raw JSON stringification of API data without recovery guidance. Rubric requires: 'Error responses must tell the LLM what to do next.'
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk differentiation: atlas_query and atlas_status are read-only; atlas_execute is destructive.
atlas_query's 'args' array lacks minItems, maxItems, or validation guidance on argument format, escape sequences, or injection risks.
atlas_execute
Replace raw error responses with actionable guidance. Example: when atlas_query fails with 404, return '{content: [{type: "text", text: "Query failed: Project not found. Available projects: [list]. Try atlas_query with a valid project_id in context, or call atlas_status to check agent health."}], isError: true}'
Add minItems, maxItems, and pattern validation to atlas_execute's 'args' array. Example: 'args: {type: "array", items: {type: "string"}, minItems: 0, maxItems: 20, description: "Command arguments (max 20 items). Do not embed shell operators or secrets."}'
Document idempotency: if atlas_execute commands are idempotent, add idempotentHint. If not, explicitly state 'This tool is NOT idempotent. Repeated calls with the same command may create duplicate records or trigger multiple side effects. Retry only if the prior call failed.'
Add parameter descriptions for all fields. 'query' in atlas_query should say: 'Natural-language query or structured command. Examples: "list all projects", "get user roles". Max 1000 chars.'
Return structured field lists from each tool, matching the documented output schema. Instead of JSON.stringify(data), extract relevant fields: {status: data.status, uptime_seconds: data.uptime, version: data.version, healthy: data.healthy}.