Read-only runtime diagnostics for the ARIE backend (Adaptive Revenue Intelligence Engine). Exposes system health, job queue inspection, API contract introspection, migration status, error logs, and configuration metadata.
Strong foundation with 11 well-named, read-only diagnostic tools. All tools have descriptions (avg 156 chars, within baseline 194) and explicit input schemas with types. Naming follows verb_noun pattern consistently (get_*, list_*, inspect_*). However, several parameter descriptions lack actionable detail (e.g., 'UUID of the job' without format hints), and output schemas are not documented in the source. Error handling guidance is minimal, tools describe what they return but not how to recover from failures. No tool annotations (readOnlyHint, etc.) visible. Composition is sound: each tool has one responsibility, and chaining is supported via returned IDs.
Cost ledger summary: total spend by category (provider calls, model calls, etc.) within a time window. Useful for understanding system resource consumption.
Organization configuration: execution mode, provider settings, and feature flags. Useful for understanding how an organization is configured.
Status of all configured providers: enabled/disabled, last test result, last error (if any). Useful for debugging provider connectivity issues.
Bounded, aggregated recent-failure summary: job-queue failures and provider-call errors/suppressions within window_minutes (default 60, max 1440). Counts by category plus up to 20 most recent rows of each — never an unbounded raw log dump.
Current routing configuration: which providers are active, their priority order, and cost/latency tradeoffs. Useful for understanding how lead enrichment is routed.
Consolidated system health: database reachability, schema migration status, job queue depth, and worker fleet liveness. Works even when the API process is down.
Output schemas not documented. Tools describe what they return in prose (e.g., 'job queue status, attempt count, last error') but no structured schema is visible in source. LLMs cannot plan downstream tool calls or extract fields reliably without documented return types.
Parameter descriptions lack actionable constraints. 'UUID of the job' does not specify format (uuid4 vs uuid5), length, or validation rules. 'Filter by job type' does not enumerate valid job types. LLMs cannot self-correct invalid inputs without explicit constraints.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Read the running API process's live OpenAPI schema (GET /openapi.json against ARIE_MCP_API_BASE_URL, default http://localhost:8000). The only tool that talks to the API process rather than the database. Optional path_prefix filters routes; include_schemas=true also returns component schemas (bounded — truncated=true if they'd exceed the response cap).
Look up a single job by id: its queue status, attempt count, last error (truncated), and its parent lead's status/is_shadow/organization_id. Returns NOT_FOUND if the job_id doesn't exist.
Applied vs. pending migrations, with checksum-integrity status for each applied one. Read-only — no apply capability exists here or anywhere in this server.
List jobs in 'failed' or 'dead_letter' status, optionally filtered by status/job_type/since, newest first. Capped at 100 rows regardless of the requested limit; `truncated=true` means more rows exist beyond the cap.
List all organizations: id, name, slug, created_at, and member count. Useful for understanding the multi-tenant landscape.
No error recovery guidance. Tools describe what they do but not what to do if they fail. E.g., 'Returns NOT_FOUND if the job_id doesn't exist', but no guidance on how to recover (search for job? list all jobs?). Agents cannot self-correct without recovery hints.
No tool annotations visible. Tools are read-only but lack readOnlyHint annotation. This prevents clients from optimizing caching, parallelization, or safety policies. Current MCP spec (2026-07-28) supports tool annotations, they should be present.
Pagination not fully specified. list_failed_jobs and list_organizations accept limit but descriptions do not clarify if results are sorted, what 'truncated' means, or how to fetch the next page. Baseline pattern requires total count or next_cursor.