majiayu000 name: sparta-intent deprecated: true deprecated_by: "graph_memory.intent + graph_memory.classifiers + graph_memory.ambiguity + graph_memory.hybrid_search" deprecated_date: "2026-02-19" description: "DEPRECATED — Use /memory intent, /memory recall, /memory clarify instead. All capabilities absorbed into graph_memory." triggers: [] composable: memory: true provides: - sparta-intent composes: [, task-monitor]
Sparta Intent Mapper (DEPRECATED 2026-02-19)
This skill has been absorbed into
/memory. All capabilities now live ingraph_memory.*:
Old (sparta-intent) New (graph_memory) sparta_intent.inferencegraph_memory.intent.IntentMappersparta_intent.query_specgraph_memory.intent.QuerySpecclassifiers/predictor.pygraph_memory.classifiers.AmbiguityPredictor/IntentPredictorambiguity_oracle.pygraph_memory.ambiguity.AmbiguityOracleresult_stats.pygraph_memory.ambiguity.ResultStatsclarifier.pygraph_memory.ambiguity.Clarifieraql_compiler.py+arango_exec.pygraph_memory.hybrid_search.hybrid_search_sparta_qra()Why deprecated: This skill was a silo that bypassed
/memory's hybrid search (BM25+vector+graph), taxonomy extraction, multihop traversal, and RecallSource infrastructure. The stress test hit a B-grade ceiling (avg 0.784) because it retrieved individual QRAs without graph context.Migration: Replace
from sparta_intent.inference import IntentMapperwithfrom graph_memory.intent import IntentMapper. Replace raw AQL queries withhybrid_search_sparta_qra().Classifier models still used: The trained DistilBERT models at
/mnt/storage12tb/media/agents/shared/sparta-intent/classifiers/models/are still loaded bygraph_memory.classifiers.
Original Architecture (for reference)
Query → IntentMapper → QuerySpec → AQLLinter → AQLCompiler → PlanGate → ArangoExecutor → ResultStats → AmbiguityOracle → Clarifier
Usage
Query (rule-based, <5ms)
./run.sh query "How does firmware verification prevent attacks?"
./run.sh query "What controls mitigate T1071?" --json
./run.sh query "How do I detect RF jamming?" --execute # runs against ArangoDB
Query with LLM (scillm via Chutes.ai)
./run.sh query "How do I detect RF jamming?" --llm
Diagnose ambiguity
./run.sh diagnose "spacecraft security"
Training (delegates to create-intent-map)
./run.sh train-sft # prints delegation instructions
./run.sh train-grpo # prints delegation instructions
Composability
| Integration | How |
|---|---|
| Memory | recall() before inference for cached QuerySpecs, learn() after successful queries |
| scillm | LLM-based inference via quick_completion() when --llm flag is set |
| Taxonomy | Canonical BRIDGE_KEYWORDS imported from taxonomy/taxonomy.py for Tier0 bridge classification |
| Embedding | graph_memory.embeddings.encode_texts() for vector lane AQL queries |
| Telemetry | track_skill("sparta-intent") wraps query execution for outcome logging |
Safety
- 7-template AQL allowlist — no dynamic query construction
- Path traversal prevention —
os.path.realpath()check on template paths - Bind parameters only — no string interpolation in AQL
- Depth cap — max 2 hops in graph traversal
- k cap — max 25 results (1000 for diagnostics)
- Plan gate — rejects full collection scans before execution
- AQL linter — validates QuerySpec constraints pre-compilation
Oracle Signals
| Signal | Threshold | Meaning |
|---|---|---|
| Margin | < 0.08 | Top results too close in score |
| Entropy | > 1.35 | Score distribution too flat |
| Tag disagreement | > 0.45 | Result set spans multiple tactics |
| Entity disagreement | > 0.55 | Result set spans multiple controls |
Decision: >= 2 triggers = ambiguous → generate clarification question.