Nahuali has 33 tools with complete descriptions and input schemas visible in source. Naming is domain-specific but inconsistent: some tools use clear verb_noun patterns (remember, recall, inspect) while others are vague (proactive, reflect). Most descriptions are 50-200 chars and explain WHEN to use the tool, which is good. However, parameter descriptions are sparse, many lack type info or constraints. Output schemas are not documented. Error handling guidance is absent. The server targets a specialized memory/governance domain, which explains the unconventional naming, but LLM discoverability suffers.
Use to acknowledge a detected anomaly: mark it as reviewed and optionally provide notes or a resolution.
Use to run a comprehensive audit of memory: validates the record ledger, checks claim support, detects conflicts, and generates a detailed compliance report.
Use to generate a structured briefing report summarizing memory state, key entities, recent episodes, and authority scores. Useful for periodic reviews and status checks.
Use when an episode explicitly supports a subject-predicate-object assertion you want recallable, after calling `remember` (pass `sourceLast: true` to cite it). A claim is the canonical evidence-backed assertion; `fact` is a deprecated alias of this tool. Then run `inspect` to confirm the claim is supported, not orphaned.
Use to generate a consolidation plan: a report identifying redundant or conflicting claims/links and recommending consolidation actions.
Deprecated alias of `claim`; prefer that. Kept for backward compatibility while clients migrate to `claim`.
Output schemas not documented. Tools like inspect, validate, briefing, reflect, audit, semantic_rebuild, semantic_status, semantic_report, trust_report, projection_status, projection_validation, projection_report return results but LLMs cannot predict field structure. This violates pattern:tool and pattern:response-shaper.
Deprecated tool aliases (fact, relate) exist alongside canonical versions (claim, link). This creates ambiguity for LLM tool selection and violates pattern:tool composition. Recommend deprecation timeline and eventual removal.
Vague tool names (proactive, reflect) do not start with clear action verbs. 'proactive' does not convey what happens; 'reflect' is ambiguous (introspection vs. HTTP reflection?). Rename to trigger_proactive_operations and introspect_memory for clarity.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Use to generate a graph representation of memory: entities, claims, links, and their relationships. Useful for visualization and analysis.
Use when importing an external source document (the structured interchange format) and you need its provenance preserved as it becomes episodes. Set `dryRun: true` first to preflight scope, size, and evidence-gap counts without writing. Then re-run without `dryRun` to append the records.
Use when you have raw UTF-8 text (notes, a transcript, a web page) to capture as source episodes without building the full interchange document. It chunks and preserves provenance but does not infer claims or links. Set `dryRun: true` to preflight; then re-run to append, and derive claims/links separately if needed.
Use to check the health and integrity of memory: supported vs. unsupported claims, conflicting facts, blind spots, and authority scores. Run this after recording claims/links to confirm they are evidence-backed, not orphaned.
Use to record an intention: a goal, plan, or commitment the agent intends to pursue. Intentions are tracked separately from claims and can be reconciled against actual outcomes.
Use to reconcile intentions against actual outcomes: compare recorded intentions with observed results and flag discrepancies for review.
Use to update an existing intention's status, priority, or other metadata.
Use when an episode explicitly supports a typed connection between two entities (from-relation-to) you want recallable, after `remember` (pass `sourceLast: true` to cite it). `link` is canonical; `relate` is a deprecated alias of this tool. Then run `inspect` to confirm the link is supported, not orphaned.
Use to register a memory hook: a callback or trigger that fires when specific memory events occur (e.g., when a claim is added or a review is resolved).
Use to request an operator review: flag memory state or decisions for human operator review and approval.
Use to trigger proactive memory operations: automatic background tasks like semantic index synchronization, stale data cleanup, or periodic health checks.
Use to record a procedure: a repeatable process or workflow the agent follows. Procedures can be referenced in intentions and used to structure agent behavior.
Use to generate a detailed projection report: schema version, event count, entity/claim/link statistics, and health metrics.
Use to check the status of the memory projection: whether it is current, event count, and any validation issues.
Use to validate the memory projection against the record ledger: ensures the projection is replay-equivalent and consistent.
Use when you need to retrieve memory by querying for episodes, claims, links, or entities matching a topic or pattern. Recall is the primary read interface: it searches lexically and semantically (if configured) and returns ranked results with their supporting evidence and authority scores.
Use to trigger a reflection cycle: memory introspection that identifies blind spots, conflicting facts, and unsupported claims, then generates actionable insights and recommendations.
Deprecated alias of `link`; prefer that. Kept for backward compatibility while clients migrate to `link`.
Use when you observe something worth remembering (a decision, event, or fact stated by the user) and want it preserved verbatim as append-only evidence. Episodes are the evidence other memory cites, so record them before asserting any claim or link. Then derive claims/links with `sourceLast: true` to cite this episode.
Use to initiate a governance review: flag claims, links, or procedures for human review and decision-making. Supports priority levels and custom review criteria.
Use to resolve a pending governance review with a decision (approve, reject, or defer) and optional notes.
Use to rebuild the semantic index from scratch. This is a maintenance operation that should be run after switching embedders or when the index becomes corrupted.
Use to generate a detailed semantic index report: embedder configuration, collection statistics, and performance metrics.
Use to check the status of the semantic index: whether it is current, how many points are indexed, and any drift from the expected state.
Use to put memory into sleep mode: suspend automatic operations and reduce resource consumption while maintaining data integrity.
Use to generate a trust report: authority scores, decision modes, and confidence metrics for the current memory state.
Use to validate the append-only record ledger for tamper evidence (if enabled) and replay integrity. Returns a detailed report of any detected issues.
No error handling guidance in any tool description. Tools do not explain what to do if a claim is unsupported, a scope is invalid, or a dry-run preflight fails. Violates pattern:recovery-guide and pattern:error-classification.
Parameter constraints missing. 'confidence' accepts 0.0-1.0 but no min/max declared. 'priority' accepts strings but no enum of valid values (low|medium|high|urgent?). 'decision' in review_resolve lacks enum. Violates pattern:constrained-input.