Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
This server has critical definition quality gaps across nearly all tools. While tool names follow a reasonable verb_noun convention (e.g., create_subscriber, get_subscriber), descriptions are sparse and provide minimal context for LLM selection. Most critically, NO input schemas are visible in the provided source code. The server.py registers tools with add_tool() calls, but the actual tool definitions in src/tools/*.py files are not shown. This means I cannot verify parameter types, required fields, enums, constraints, or output schemas. The descriptions shown in the tool list (e.g., 'Create a new subscriber in the OCS system') are extremely brief (under 50 characters), well below the 194-character production baseline and insufficient to guide LLM reasoning. Without seeing the actual tool implementation code, I can only infer schemas exist, which caps per-tool scores at 50 per the hard rules. The account-history-api-tests.sh script shows an external API exists with a JSON interface, but the MCP tool wrapping is opaque. Risk annotations are present (READ_ONLY, WRITE, DESTRUCTIVE) in the metadata, but tool annotations (readOnlyHint/destructiveHint/idempotentHint) are not visible in the server code.
No input schemas visible in source code. Tool definitions in src/tools/*.py are not provided, only tool registration in server.py. Cannot verify parameter types, required fields, enums, or constraints for any of the 15 tools. This violates the critical pattern:constrained-input and pattern:tool-description requirements.
Tool descriptions are critically sparse (under 50 characters each). Baseline production tools average 194 characters. Current descriptions provide no context on when to use a tool, what it returns, dependencies, or error guidance. Examples: 'Create a new subscriber in the OCS system' (48 chars) does not explain what parameters are required, what gets created, or what the response contains.
Recommendations
Provide complete source code for all tool definitions in src/tools/*.py files, including parameter schemas (types, required fields, enums, descriptions) and output schemas.
Expand every tool description to 100-200 characters following the pattern: WHAT the tool does, WHEN to use it (vs similar tools), and WHAT it returns. Example: 'Create a new subscriber in the OCS system. Use this after validating the subscriber does not exist. Returns subscriber_id, creation_timestamp, and status. Supports idempotent creation via idempotency_key parameter.'
Add comprehensive parameter descriptions to every input. For lookup_subscriber, specify: 'msisdn (string, 10-15 digits, E.164 format), international phone number' and 'imsi (string, 15 digits), international mobile subscriber identity. Exactly one of msisdn or imsi is required.'
Add error guidance to WRITE and DESTRUCTIVE tools. Example for delete_subscriber: 'Returns 204 on success. If subscriber has active subscriptions, returns 409 Conflict with message 'Cannot delete subscriber with active subscriptions.' Resolve by calling list_subscriptions and delete_subscription first, then retry delete_subscriber.'
Implement MCP toolAnnotations. Mark get_* and list_* tools with readOnlyHint=true. Mark delete_* and change_subscription_state with destructiveHint=true. Mark tools supporting idempotency_key with idempotentHint=true.
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
No parameter descriptions visible. Cannot verify that every input parameter has a non-empty description explaining its purpose, format, or constraints. Production baseline requires 100% of parameters to have descriptions.
No output schemas documented. Tools like list_subscriptions should declare pagination support (limit, offset, total_count) and return structure. Without documented output, LLMs cannot plan downstream tool calls or extract the right data.
No error handling guidance visible. Tools like delete_subscriber (DESTRUCTIVE) should explain recovery paths and whether errors are retryable. No evidence of 'If X fails, try Y' error guidance per pattern:recovery-guide.
No tool annotations visible. Risk metadata exists (WRITE, DESTRUCTIVE, READ_ONLY) in the metadata list, but MCP toolAnnotations (readOnlyHint, destructiveHint, idempotentHint) are not present in the server code. These enable LLM-aware safety reasoning.
Tools that accept identifiers (e.g., lookup_subscriber) lack clarity on accepted formats. Description says 'Lookup subscriber by MSISDN or IMSI' but does not specify which parameter accepts which format, expected format/length, or how ambiguity is resolved.
No evidence of idempotency guarantees. Tools like create_subscriber and create_subscription should document whether retries with the same input produce duplicate records or are idempotent. This is critical for agent resilience.
Add dependency hints. Example: 'lookup_subscriber' → 'If only the subscriber name is known, call search_subscribers() first. lookup_subscriber requires MSISDN or IMSI.'
Clarify which parameters accept multiple formats (e.g., subscriber_id vs email vs username). Example: 'update_subscriber accepts subscriber_id (UUID), msisdn (E.164), or email. Only one is required.'
Document idempotency strategy for create_* tools. Example: 'create_subscriber is idempotent if idempotency_key is provided. Retrying with the same idempotency_key and parameters will return the existing subscriber without duplicating.'
Add examples in parameter descriptions (but only as part of constraints, not as prose). Example: 'status (enum: active|pending|inactive|suspended), subscriber status. Default: pending.'
Validate and reject invalid input early with actionable error messages. Example: 'Invalid MSISDN: got "123". Must be 10-15 digits in E.164 format. Example: +1-555-0123. Did you mean: +1-555-0123-4567?'