Context Manager

The Context Manager is a Tauri plugin that owns all client-side state: the EstateView (local RAG), conversation history, session state, artifacts, and evidence. It assembles the context payload sent to backend agents on every request.

---

What it owns

Data	Storage	Description
EstateView (local RAG)	SurrealDB graph (local)	All cloud resources, edges, overlays, embeddings
Conversation history	Serial DB (local)	Full message history across all sessions
Session state	Serial DB (local)	Active session metadata, current scope, active profiles
Step summaries	Serial DB (local)	Per-step summaries for long-running multi-step operations
Artifacts	Serial DB (local)	Persisted outputs — findings, reports, plans, bundles
Evidence	Serial DB (local)	Execution records, before/after state, approval logs

Nothing in this list leaves the customer's machine unless explicitly exported.

---

Context assembly

On every request to a backend agent, the Context Manager:

1. Reads the current session state (active profiles, scope filters)
2. Queries the local EstateView RAG for relevant resources (vector search + graph traversal)
3. Applies overlay annotations (tags, labels, deployment groups added by users)
4. Serialises the assembled context as a JSON payload
5. Attaches conversation history from the serial DB
6. Sends the complete, self-contained request to the Gateway

The backend agent receives everything it needs. It never needs to query the estate directly.

---

Estate-injected context types

Some context types are always available from the local estate without a fetch step. These do not need to appear in a skill's execution_plan.steps:

Context type	Description
account_scope	Active account IDs and regions from enabled profiles
environment_scope	Environment labels inferred from tags and naming patterns
resource_scope_summary	High-level resource counts by service across enabled profiles
active_profiles	Currently enabled ObserverContext profiles

Declare these in supported_context_types in skill.yaml for ADK validation, but omit them from execution_plan.steps.

---

Local RAG pipeline

The Context Manager's local RAG is built on the EstateView graph in SurrealDB:

Estate scan (Rust plugins) → ingest_resource() → EstateView nodes
                           → write_edge()      → EstateView edges
                           → commit_snapshot() → versioned snapshot

Query time:
  embed_query(user_question)     → 384-dim vector (ONNX, all-MiniLM-L6-v2)
  search_resources(vector)       → cosine similarity over resource embeddings
  get_neighbors(resource_id)     → graph traversal for related resources
  build_rag_context(results)     → structured JSON context for agent

The embedding model (all-MiniLM-L6-v2) runs entirely on-device via ONNX Runtime. No embeddings are sent to external services.

---

Next steps

• Context Engine — The server-side counterpart
• Core Concepts — EstateView and ObserverContext explained
• Platform Overview — Thin backend, stateful frontend principle