Cross-agent memory layer with drift detection · LongMemEval-S 56.6% · MCP + A2A · local-first · write-time-zero-LLM design
nautilus-compass defines 17 tools with schemas and descriptions present, but quality is inconsistent. Most tool descriptions are adequate (50-150 chars), but parameter descriptions are sparse or missing context. Schemas are present but lack depth: many parameters lack type constraints (e.g., 'parameters' object in long_task, submit_platform_task are untyped). No output schemas documented. Error handling is absent, tools do not guide recovery or categorize failures. Naming is verb-forward (recall, drift_check, ingest_obs) but some names are vague (long_task, submit_platform_task). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. STDIO transport caps protocol readiness at 50.
Register a new worker agent in the platform registry
Check alignment/deviation/alert status against stored anchors and persona vectors
Retrieve drift detection timeline (last 30 days)
Record feedback for adaptive anchor retraining (direction + reason)
Retrieve governance audit log for compliance and access tracking
Dispatch governance action (e.g., purge, archive, reindex)
Untyped 'parameters' object in long_task and submit_platform_task. LLMs cannot infer valid structure.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields. Violates pattern:tool and mxe:include-chaining-ids.
No error handling guidance. Tools do not indicate retryability, user-fixable errors, or recovery steps. Violates pattern:recovery-guide.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Check governance lock status for project memory
Create or update governance plan for memory lifecycle and retention
Ingest observation into session memory (write-time-zero-LLM design)
Ingest results from completed platform task
Submit a long-running task for asynchronous processing
Retrieve user/agent profile including persona vectors and metadata
Generate cryptographic proof of memory impact and citation chain
Retrieve top-k memory hits from the corpus based on semantic query
Search across sessions by metadata, timestamp, or content
Submit task to platform queue for distributed processing
Recall memory within a specific thread/conversation context
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels (governance_dispatch marked DESTRUCTIVE, feedback_log marked WRITE). Violates current MCP spec (2026-07-28).
Generic tool names (long_task, submit_platform_task, profile) lack specificity. 'profile' does not indicate read-only intent; 'long_task' is vague about what task types are valid. Violates pattern:tool naming.