feat: Shared Schema Cache (SchemaStore) — Step 6.1

[tnaum-ms]

Step 6.1: Shared Schema Cache (SchemaStore)

Introduces a SchemaStore singleton that accumulates schema data per {clusterId, databaseName, collectionName}, enabling cross-tab and scratchpad schema sharing.

What Changed

New: SchemaStore singleton (src/documentdb/SchemaStore.ts)

• Map<"clusterId::db::coll", SchemaAnalyzer> — shared schema cache
• Debounced onDidChangeSchema event (1s per key) to avoid excessive consumer churn
• clearSchema() / clearCluster() / clearDatabase() for lifecycle cleanup
• Registered as VS Code Disposable in extension.ts

Refactored: ClusterSession delegates to SchemaStore

• Removed private _schemaAnalyzer and _highestPageAccumulated fields
• getKnownFields() and getCurrentPageAsTable() now delegate to SchemaStore
• Removed unused getCurrentSchema() method
• Added _clusterId, _databaseName, _collectionName for SchemaStore key construction
• Page-based dedup dropped — stats are already documented as approximate

New: Scratchpad results feed SchemaStore

• Extended ExecutionResult with source.namespace (extracted from @mongosh ShellResult.source)
• executeScratchpadCode.ts feeds document results (very conservative: only Cursor and Document types)
• Running db.users.find() in the scratchpad now populates field completions for the users collection across all Collection View tabs

Wired: Schema cleanup on disconnect/drop

• removeConnection → clearCluster()
• deleteCollection → clearSchema()
• deleteDatabase → clearDatabase()
• ClustersClient.deleteClient → clearCluster() (lazy import to avoid circular dep)

Tests

• 22 unit tests covering read/write ops, key isolation, event debouncing, lifecycle, singleton behavior
• All existing scratchpad tests pass (52 tests across 4 suites)
• Manual test plan (6 scenarios) in docs/plan/06.1-shared-schema-cache.md

Key Behaviors

Scenario	Before	After
Two tabs on same collection	Independent schemas	Shared — Tab B gets Tab A's field completions
Scratchpad find() result	Documents discarded	Fed to SchemaStore → completions available
Tab closed, new tab opened	Schema lost	Schema persists in SchemaStore
Connection removed	Schema leaked	Schema cleared via clearCluster()
Collection dropped	Schema leaked	Schema cleared via clearSchema()

Plan

Full design doc with architecture, decisions (D1–D7), resolved questions (Q1–Q3), deviations log: docs/plan/06.1-shared-schema-cache.md (gitignored, local only)

[Copilot]

Pull request overview

Introduces a shared, extension-wide schema cache (SchemaStore) keyed by {clusterId, databaseName, collectionName} to enable cross-tab and scratchpad-driven schema sharing, and wires lifecycle cleanup into disconnect/drop flows.

Changes:

• Added SchemaStore singleton with per-collection schema accumulation, debounced change events, and clear/reset APIs.
• Refactored ClusterSession to feed/query schema via SchemaStore instead of per-session accumulation.
• Extended scratchpad execution results with source.namespace and feeds document results into SchemaStore; added schema cleanup hooks on connection/database/collection removal.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/extension.ts	Registers SchemaStore for disposal with extension lifecycle.
src/documentdb/SchemaStore.ts	Implements singleton schema cache, debounced change events, and clear/reset APIs.
src/documentdb/SchemaStore.test.ts	Adds unit tests for SchemaStore behavior (read/write, isolation, debounce, lifecycle).
src/documentdb/ClusterSession.ts	Delegates schema operations to SchemaStore and feeds query results into shared cache.
src/documentdb/ClustersClient.ts	Clears SchemaStore data when a client is deleted.
src/documentdb/scratchpad/types.ts	Extends ExecutionResult with optional source.namespace metadata.
src/documentdb/scratchpad/ScratchpadEvaluator.ts	Extracts source.namespace from @mongosh shell results.
src/commands/scratchpad/executeScratchpadCode.ts	Feeds scratchpad document/cursor results into SchemaStore.
src/commands/removeConnection/removeConnection.ts	Clears cluster schema cache on connection removal.
src/commands/deleteDatabase/deleteDatabase.ts	Clears database schema cache on database deletion.
src/commands/deleteCollection/deleteCollection.ts	Clears collection schema cache on collection deletion.

[tnaum-ms]

Re: lazy import in deleteClient — Fixed in 700a4fd. Replaced the lazy import() with a static import. Confirmed no circular dependency exists: SchemaStore imports from schema-analyzer and vscode; ClustersClient imports from mongodb and local auth/credential modules. The two are independent.

[tnaum-ms]

PR Review Responses

Finding 1 (live cross-tab updates): Acknowledged. Severity assessed as Low. The shared SchemaStore data is consumed on the next query — tabs do not receive push updates. The onDidChangeSchema event was designed for Step 7 (Scratchpad CompletionItemProvider), which will be the first real consumer needing push semantics. Wiring a tRPC subscription for Collection View tabs is out of scope for the infrastructure PR. No code change.

Finding 2 (silent invalidation): Acknowledged. Severity assessed as Low. clearCluster/clearDatabase intentionally do not fire per-key events (plan §D4: "No granular events — consumers should treat cluster disconnect as full invalidation"). No production code subscribes to onDidChangeSchema, so events would have zero effect today. Invalidation contract will be designed with Step 7. No code change.

Finding 3 (WithId<Document> typing): Acknowledged. The type predicate in feedResultToSchemaStore is technically unsound — it does not verify _id exists. However, the underlying SchemaAnalyzer.addDocuments() in the schema-analyzer package also requires WithId<Document> because its primary data source is the driver's find() which always returns WithId<Document>. Loosening SchemaStore alone would still hit the same constraint at SchemaAnalyzer. The scratchpad path almost always returns _id (omitted only with explicit { _id: 0 } projections). Deferring a decision on whether to loosen the schema-analyzer package API. No code change for now.

Finding 4 (namespace.collection optionality): Acknowledged. The defensive guard is intentional. When @mongosh provides a namespace, collection is always present. The guard costs nothing and protects against future upstream changes. No code change.

Finding 5 (lazy import): Fixed in 700a4fd — replaced with static import.

[tnaum-ms]

Re: unsound type predicate — Fixed in d756cb6. Added '_id' in d to the type predicate so the narrowing to WithId<Document> is sound. Documents from { _id: 0 } projections are now excluded — their artificial shapes should not feed schema analysis.