By name: analysis-storage-architect description: Use for raw ingestion storage, normalized analysis stores, session storage, object storage, graph projection boundaries, indexing, partitioning, and trace correlation.
Skill: Analysis Storage Architect
Description
Guides evidence-first storage architecture for analysis sessions, raw inputs, normalized stores, artifacts, and projections.
Instructions
- Verify existing analysis store, artifact store, ingestion session, persistence, and workspace storage contracts.
- Separate raw ingestion storage from normalized analysis stores, relational persistence, object storage, graph projection, and event-sourcing-style records.
- Preserve analysis session storage, workspace IDs, repository references, payload identity, checksums, ordering, provenance, trace correlation, and completeness state.
- Plan large payload handling, storage partitioning, indexing strategies, retention, and cleanup as explicit policies.
- Treat graph databases, vector databases, reports, and LLM packages as projections unless repository documentation says otherwise.
- Keep persistence implementation details in outbound adapters or infrastructure.
- Ensure unknown, incomplete, unresolved, or unavailable evidence remains explicit in stored state.
- Apply
.agents/skills/resilience-engineering/SKILL.mdfor storage retry idempotency, dead-letter handling, retry provenance, cleanup, health checks and partial-write diagnostics.
Expected Inputs
- domain analysis and ingestion models
- application storage ports
- persistence adapters
- workspace storage boundaries
- graph and report projection plans
Expected Outputs
- storage responsibility map
- raw and normalized storage plan
- analysis session persistence plan
- payload and artifact indexing notes
- storage quality and migration risks
Boundaries
- Do not choose a database product without an approved architecture slice.
- Do not store LLM output as verified evidence.
- Do not collapse raw payloads, normalized facts, and projections into one ambiguous structure.
Stop Conditions
Stop if:
- storage table, port, or model names cannot be verified;
- persistence would overwrite raw inputs or lose provenance;
- graph projection would become the source of truth;
- large payload handling lacks deterministic identity.