feat(mastra): migrate v2 LangGraph chat extension to Mastra v3

[rohit-sourcefuse]

Migrate lb4-llm-chat-component from LangGraph + LangChain to Mastra v3

Replaces the LangGraph/LangChain runtime with Mastra v3 (Agent + Memory + Workflow primitives). LangGraph and LangChain are fully removed; ChatGraph + DbQueryGraph + VisualizationGraph and their node classes are deleted and re-expressed as Mastra workflows. The locked SSE wire contract (8 event types) is preserved byte-identical, so existing consumers keep working.

111 commits, +18.6k / −13.1k LOC across 223 files. 372 mocha tests pass; build + eslint + prettier + SonarCloud quality gate all green on HEAD 9649101.

This description is current as of HEAD 9649101. Earlier revisions of this PR body (smaller test counts, "stub steps", "LangChain kept") are obsolete — everything below is authoritative.

---

Runtime

• Agent. Streams the registered mastra.getAgent('chatAgent') (per-request model / tools / instructions resolved from RequestContext), so every /reply is a single root trace. No detached per-request new Agent().
• WorkflowRunner (REQUEST-scoped) bridges LB4 controllers to agent.stream().fullStream. A single AsyncEventQueue enforces total event order across pre-processing, the fullStream pump, and tool-side eventWriter calls.
• Single Mastra instance (SINGLETON) holds storage pools, vector clients, observability exporters; per-request DI flows through agent.stream({requestContext}). Mastra is used as a library, not a server — LB4 owns the HTTP surface.
• UsageAccumulator (REQUEST-scoped) replaces the LangChain-callback token counter.
• Context compression — v2's ContextCompressionNode ported to a Mastra input processor (MAX_TOKEN_COUNT).

Workflows (fully wired — no stub steps)

Three workflows registered on MastraProvider, callable via mastra.getWorkflow(id).createRun().start():

• generateQueryWorkflow — parallel(check-cache, get-tables, check-templates) → post-cache-and-tables → branch(AsIs | FromTemplate | Continue) → get-columns → generate-checklist → dountil(sql-and-validate) → branch(save | failed).
• improveQueryWorkflow — load-existing → dountil(fix-query) → branch(save-improved | failed).
• visualizationWorkflow — select-visualisation → call-query-generation → get-dataset-data → render-visualization.

Behaviour restored to v2 parity (each covered by tests):

• Syntactic (DB EXPLAIN) + semantic validators, plus the table_not_found reclassification — a missing-table/column error re-selects and widens the allowed table set for the next dountil iteration.
• Unanswerable-question gate — get-columns detects when no table can answer the prompt and fast-fails with a user-facing clarification instead of burning the full validation loop and returning an empty (or wrong) dataset.
• Disliked-cache filtering — a cached dataset a user disliked (or that no longer exists) is regenerated, not re-served.
• "Similar" cache hit seeds SQL generation with the matched (non-disliked) query as a worked example, rather than discarding it.
• SQL-gen tier selection — cheap tier for single-table queries and validation-fix retries, smart tier for multi-table first attempts (override via nodes.sqlGenerationNode.useSmartLLMForSingleTableQueries).
• Checklist gate — honours nodes.generateChecklistNode.enabled and skips on ≤2 tables, eliding a planning LLM call per query.
• Failure messages — the failed terminal surfaces a reason (unanswerable clarification ▸ last validation feedback ▸ generic rephrase) instead of a silent empty dataset.
• Cache / template hits short-circuit.

Bindings (BREAKING vs the pre-Mastra names)

• Model tiers are the bare keys ChatLLM / FileLLM / CheapLLM / SmartLLM / SmartNonThinkingLLM (the Mastra* aliases were removed).
• Runtime infra under MastraInternalBindings (Mastra, Storage, Tools, Observability, RunRegistry, ResourceId).
• Tool layer de-prefixed (BREAKING): IGraphTool, ToolStore, GetDataAsDatasetTool / ImproveDatasetTool / AskAboutDatasetTool / GenerateVisualizationTool; files *.tool.ts; LB4 keys services.<X>Tool.
• The Mastra-instance provider class keeps a semantic prefix: MastraProvider.

LLM tiers + call-time settings

Cheap / Smart / SmartNonThinking tiers wired for cost parity; CLAUDE_THINKING(_BUDGET), per-provider temperatures, OpenRouter reasoning, and MAX_TOKEN_COUNT all threaded at call time. Providers are AI-SDK-native (@ai-sdk/* + @openrouter/ai-sdk-provider + ollama-ai-provider); create{OpenAI,OpenRouter}Model factories.

Memory + storage

• semanticRecall is opt-in (MASTRA_SEMANTIC_RECALL, default OFF) — leaving it on whenever a vector store is bound caused progressive per-request latency. generateTitle likewise opt-in (MASTRA_GENERATE_TITLE).
• Opt-in PostgresStorageProvider (@mastra/pg) alongside the LibSQL default; opt-in MastraVector (pgvector) for the db-query cache.

Chat history API (v2-Chat-compatible)

GET /chats, GET /chats/{id} (thread + messages), GET /chats/{id}/messages, backed by Mastra Memory (replaces the deleted ARC ChatController/ChatStore). Responses match the v2 Chat/Message shape consumers (e.g. BizBook) depend on: tenantId/userId, title (first prompt; New Chat fallback), top-level inputTokens/outputTokens (persisted to thread metadata per run), createdOn/modifiedOn. Each Mastra message is flattened into user/ai/tool messages; the tool message carries metadata.{toolName, args, existingDatasetId, status} — and for the visualization tool metadata.{visualization, config} — so a consumer can re-run / Load Dataset / re-render a chart from history. resourceId derivation (${tenantId}:${principalId}) is shared by the writer (WorkflowRunner) and readers (ChatController) so scopes never drift.

Observability

MastraMultiObservability fans every agent / workflow / tool span to Langfuse and LangSmith at once (whichever env keys are present); workflow tools forward tracing context so each /reply is one root trace. Env: LANGFUSE_*, LANGSMITH_* (LangChain-style names also honoured), OTEL_SAMPLE_RATE.

Developer ergonomics

• Activity logs — every workflow step status streams to a debug channel: DEBUG=ai-integration:* (or ai-integration:steps) shows step progress on the server console (restores v2 step-log visibility; off by default, no console.*).
• Token streaming (opt-in, default OFF) — MASTRA_STREAM_TOKENS=true emits one message SSE event per text delta for progressive rendering. Default OFF preserves the single-event contract (and v2 parity — v2 never streamed). A consumer must append message events to use it; the bundled sandbox does. (The current BizBook UI renders a new bubble per message event, so it stays OFF until that UI appends.)
• Extensibility — subpath exports lb4-llm-chat-component/testing (accuracy harness) and lb4-llm-chat-component/mastra (steps, workflows, MastraProvider); README "Extending the component" with full Tools / Agents / Steps examples.

Dependencies

Added: @mastra/core, @mastra/memory, @mastra/libsql, @mastra/pg, @mastra/observability, @mastra/langfuse, @mastra/langsmith; @ai-sdk/{openai,anthropic,amazon-bedrock,google,groq,cerebras}, ollama-ai-provider, @openrouter/ai-sdk-provider; ai, zod.
Removed: @langchain/langgraph, @langchain/community, @langchain/core, langchain. The only remaining LangChain-family dependency is the LangSmith exporter SDK (via @mastra/langsmith), used purely for trace export.

Tests

• Step-level + helper unit coverage for every workflow step (cache judge incl. dislike/Similar paths, get-tables, get-columns answerability gate, checklist gate, SQL-gen tier selection, save-dataset, failed-message synthesis).
• DAG-level branch coverage for generateQueryWorkflow (AsIs / FromTemplate / Continue / unanswerable / disliked-regenerate).
• WorkflowRunner unit (Init / Message / coalesced-vs-streamed tokens / Tool / ToolStatus / Error / TokenCount / Status / ResourceId).
• Live integration test (no sinon) — boots real Mastra + Memory + LibSQLStore(:memory:) driven by createMockModel; catches fullStream chunk-shape shifts between Mastra minor versions.
• Endpoint-driven accuracy harness (./testing export) for prompt→SQL golden cases.

Test plan

• npm run build — clean.
• npm test — 372 pass, eslint + prettier clean.
• SonarCloud quality gate — OK (no failing conditions).
• npm ls @langchain/langgraph — empty.
• Live Mastra Agent integration test against LibSQLStore(:memory:) + createMockModel.
• Consumer end-to-end via lb4-llm-sandbox — chat, db-query, improve, ask, visualization (bar/pie/line), datasets, chat-history, playground all verified live.

How to review

• WorkflowRunner (src/mastra/bridge/workflow-runner.ts) — load-bearing; verify the fullStream switch matches the SSE wire contract.
• MastraProvider (src/providers/mastra/mastra.provider.ts) — workflow + agent registration, memory/semantic-recall/title env gating.
• Workflow steps — src/mastra/workflows/db-query/steps/ and .../visualization/steps/; each step has a typed input/output schema and inline doc-comments mapping to the v2 node it replaces.
• Deleted v2 code lives under its pre-deletion commits — see the inline git show <sha>^:... pointers in step files for behaviour-delta review.

Known follow-ups (not blocking this PR)

• HITL suspend/resume (ApprovalController + RunRegistry consumer) lands in v3.1 — tools emit AwaitingApproval but there is no resume flow yet.
• Token-streaming default flips to ON once a consumer UI (BizBook) appends message events instead of rendering one bubble per event.

[Copilot]

Pull request overview

Migrates the lb4-llm-chat-component runtime from LangGraph/LangChain graphs to Mastra v3 primitives (Mastra singleton + per-request WorkflowRunner), removing the legacy graph/node implementations while keeping the SSE event contract intact.

Changes:

• Replaced ChatGraph.execute() usage with a REQUEST-scoped WorkflowRunner.run() that streams Mastra agent.stream().fullStream into the existing SSE event types.
• Added Mastra v3 bindings/providers (Mastra singleton, default LibSQL storage, tool registry, observability exporters, run registry) plus multiple AI SDK-backed LLM/embedding providers.
• Introduced Mastra-compatible tool wrappers and vector store provider; removed legacy LangGraph graphs/nodes/tools and their test suites, adding new unit/integration coverage for the Mastra path.

Reviewed changes

Copilot reviewed 142 out of 146 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/types.ts	Removes LangGraph checkpointer provider types tied to @langchain/langgraph.
src/keys.ts	Adds Mastra v3 binding keys (Mastra, storage, tools, LLM configs, observability, run registry, resourceId) and removes LangGraph checkpointer binding.
src/services/generation.service.ts	Routes generation through WorkflowRunner.run() instead of ChatGraph.execute().
src/services/usage-accumulator.service.ts	Adds a token-usage accumulator to replace the removed LangChain callback counter.
src/services/token-counter.service.ts	Deletes legacy LangChain callback-based token counter.
src/services/index.ts	Updates service exports to remove TokenCounter and export UsageAccumulator.
src/mastra/bridge/async-event-queue.ts	Introduces an async queue to preserve total ordering of SSE events across concurrent producers.
src/mastra/bridge/run-registry.ts	Adds default in-process run registry for HITL approval flow resumption.
src/providers/mastra/storage.provider.ts	Adds default Mastra LibSQL storage provider (file-backed SQLite by default).
src/providers/mastra/mastra.provider.ts	Registers singleton Mastra instance with Memory + workflows.
src/providers/mastra/mastra-tools.provider.ts	Adds default Mastra tool registry provider for the 4 internal tools.
src/providers/mastra/observability/langsmith.provider.ts	Adds optional LangSmith exporter wiring for Mastra observability.
src/providers/mastra/observability/langfuse.provider.ts	Adds optional Langfuse exporter wiring for Mastra observability.
src/observers/mastra-lifecycle.observer.ts	Adds app lifecycle observer to shutdown Mastra cleanly.
src/mastra/bridge/workflow-runner.ts	Implements the main Mastra streaming bridge that maps fullStream chunks to the locked SSE contract.
src/sub-modules/db/postgresql/vector-store/pgvector.mastra.store.ts	Adds Mastra PgVector store provider for Mastra vector storage.
src/sub-modules/db/postgresql/vector-store/index.ts	Exports the new Mastra PgVector store alongside the legacy store.
src/sub-modules/providers/openai/llms/openai.mastra.provider.ts	Adds AI SDK/Mastra-shaped OpenAI provider.
src/sub-modules/providers/openai/llms/index.ts	Exports the new Mastra OpenAI provider.
src/sub-modules/providers/openrouter/llms/openrouter.mastra.provider.ts	Adds AI SDK/Mastra-shaped OpenRouter provider.
src/sub-modules/providers/openrouter/llms/index.ts	Exports the new Mastra OpenRouter provider.
src/sub-modules/providers/ollama/llms/ollama.mastra.provider.ts	Adds AI SDK/Mastra-shaped Ollama provider.
src/sub-modules/providers/ollama/llms/index.ts	Exports the new Mastra Ollama provider.
src/sub-modules/providers/ollama/embedding/ollama-embedding.mastra.provider.ts	Adds AI SDK/Mastra-shaped Ollama embedding provider.
src/sub-modules/providers/ollama/embedding/index.ts	Exports the new Mastra Ollama embedding provider.
src/sub-modules/providers/groq/llms/groq.mastra.provider.ts	Adds AI SDK/Mastra-shaped Groq provider.
src/sub-modules/providers/groq/llms/index.ts	Exports the new Mastra Groq provider.
src/sub-modules/providers/google/llms/gemini.mastra.provider.ts	Adds AI SDK/Mastra-shaped Gemini provider.
src/sub-modules/providers/google/llms/index.ts	Exports the new Mastra Gemini provider.
src/sub-modules/providers/google/embedding/gemini-embedding.mastra.provider.ts	Adds AI SDK/Mastra-shaped Gemini embedding provider.
src/sub-modules/providers/google/embedding/index.ts	Exports the new Mastra Gemini embedding provider.
src/sub-modules/providers/cerebras/llm/cerebras.mastra.provider.ts	Adds AI SDK/Mastra-shaped Cerebras provider.
src/sub-modules/providers/cerebras/llm/index.ts	Exports the new Mastra Cerebras provider.
src/sub-modules/providers/aws/llms/bedrock.mastra.provider.ts	Adds AI SDK/Mastra-shaped Bedrock provider.
src/sub-modules/providers/aws/llms/index.ts	Exports the new Mastra Bedrock provider.
src/sub-modules/providers/aws/embedding/bedrock-embedding.mastra.provider.ts	Adds AI SDK/Mastra-shaped Bedrock embedding provider.
src/sub-modules/providers/aws/embedding/index.ts	Exports the new Mastra Bedrock embedding provider.
src/sub-modules/providers/anthropic/llms/anthropic.mastra.provider.ts	Adds AI SDK/Mastra-shaped Anthropic Claude provider.
src/sub-modules/providers/anthropic/llms/index.ts	Exports the new Mastra Claude provider.
src/graphs/event.types.ts	Extends SSE event union with an explicit Error event type.
src/graphs/types.ts	Removes LangGraph node/edge types; adds Mastra-shaped tool interfaces and tool status enum value for awaiting approval.
src/graphs/message-metadata.type.ts	Introduces framework-free message metadata types (moved out of deleted LangGraph state annotations).
src/graphs/index.ts	Updates exports to remove graph/state exports and add message metadata exports.
src/controllers/index.ts	Removes ChatController export.
src/component.ts	Wires Mastra providers/services/tools + lifecycle observer; removes ChatGraph/ChatController registration.
src/components/visualization/types.ts	Replaces LangGraph annotation-based state with a plain interface for visualizers.
src/components/visualization/visualizer.component.ts	Removes visualization graph/node/tool service registrations; retains visualizers for workflow dispatch.
src/components/visualization/tools/generate-visualization.mastra.tool.ts	Adds Mastra createTool wrapper that invokes visualizationWorkflow.
src/components/visualization/tools/index.ts	Switches export from legacy tool to Mastra tool.
src/components/visualization/visualizers/bar.visualizer.ts	Updates imports to use new visualization types module.
src/components/visualization/visualizers/line.visualizer.ts	Updates imports to use new visualization types module.
src/components/visualization/visualizers/pie.visualizer.ts	Updates imports to use new visualization types module.
src/components/db-query/tools/get-data-as-dataset.mastra.tool.ts	Adds Mastra tool wrapper invoking generateQueryWorkflow.
src/components/db-query/tools/improve-dataset.mastra.tool.ts	Adds Mastra tool wrapper invoking improveQueryWorkflow.
src/components/db-query/tools/ask-about-dataset.mastra.tool.ts	Adds Mastra tool wrapper that runs a one-shot Mastra Agent for dataset Q&A.
src/components/db-query/tools/index.ts	Switches db-query tool exports to Mastra tool wrappers.
src/components/db-query/db-query.component.ts	Removes legacy DbQueryGraph/nodes/tools registration; keeps helper services for workflow step bodies.
src/tests/unit/workflows.unit.ts	Adds workflow smoke tests verifying DAG completion on stubbed paths.
src/tests/integration/workflow-runner-agent.integration.ts	Adds end-to-end integration test against real Mastra + Memory + LibSQL store using Mastra mock model.
src/tests/integration/mastra-test-utils.d.ts	Adds ambient module declaration to compensate for missing Mastra test-utils typings.
src/tests/integration/generation.service.integration.ts	Updates integration tests to stub/use WorkflowRunner instead of ChatGraph.
package.json	Adds Mastra/AI SDK deps, removes @langchain/langgraph, registers a backfill-mastra-threads bin, and drops the removed db-query/testing export.
.gitignore	Ignores default local Mastra sqlite DB files (mastra.db*).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[sf-sahil-jassal]

🟠 [BUG]: FileLLM tier injected but never used — file summarisation silently uses ChatLLM

Issue: fileLlm is injected at L265 (@inject(AiIntegrationBindings.FileLLM)) but resolveFileSummaryModelConfig() only ever consults this.chatLlm. Consumers (e.g. reporting-service) that bind a dedicated FileLLM (a vision/file-capable model) expecting it to handle attachments get it ignored — files are summarised by the chat model, which may not even accept file parts. The fileLlm ctor param becomes dead.

Fix:

private resolveFileSummaryModelConfig(): MastraModelConfig | undefined {
  if (this.fileLlm) return this.fileLlm;
  if (this.chatLlm) return this.chatLlm;
  const defaultModel = process.env.MASTRA_DEFAULT_CHAT_MODEL;
  return defaultModel ? toModelRouterFallbackConfig(defaultModel) : undefined;
}

Credit: Open Code

[sf-sahil-jassal]

🟡 [DEAD_CODE]: runRegistry injected but never referenced

Issue: runRegistry is injected (non-optional) but never read or written anywhere in this class — the HITL resume/suspend path that would call .set() is deferred to v3.1 per the handleChunk comment. A non-optional inject for an unused, not-yet-wired dependency is a boot-time coupling with no payoff and risks failing construction if the binding is absent.

Fix: Drop the param until the resume path lands, or mark {optional: true} to avoid a hard boot dependency on an unused binding.

Credit: Open Code

[sf-sahil-jassal]

🟡 [TYPE_SAFETY]: double cast memory as unknown as ThreadMemory hides the real Memory contract

Issue: getMemory() returns a typed Memory; double-casting through unknown to a hand-rolled ThreadMemory to reach updateThread papers over the actual API and will silently break if Mastra renames/moves updateThread. The optional-updateThread guard then makes a missing method a silent no-op (usage never persisted) rather than a visible error.

Fix: Type against the published Memory interface (import type {MastraMemory} / the storage thread API) and call updateThread directly; if the installed version truly lacks it, surface that at boot, not per-request.

Credit: Open Code

[sf-sahil-jassal]

🟠 [PERFORMANCE]: maxSize overflow hard-closes mid-stream → silent SSE truncation

Issue: On overflow push() calls close(), which resolves the consumer's pending next() with done:true and discards every already-queued-but-undrained event. For a fast tool emitting >10000 events (or a slow SSE consumer), the user sees the stream end cleanly with no Error event and a partial answer — indistinguishable from success. That's data loss disguised as completion.

Fix: Before hard-closing, enqueue a terminal Error event so the consumer can distinguish overflow from normal completion:

if (this.queue.length >= this.maxSize) {
  this.queue.push(OVERFLOW_ERROR_EVENT as T); // or surface via a callback
  this.close();
  return;
}

…and/or apply real backpressure (await) for producers that can tolerate it.

Credit: Open Code

[rohit-sourcefuse]

Agreed — overflow silently looked identical to a normal stream end (clean done:true, partial answer, no signal to the client). Fixed in 8581610: AsyncEventQueue now accepts an optional overflowValue in its constructor; on maxSize overflow it pushes that value before closing so the consumer sees an explicit event before done. workflow-runner wires an LLMStreamEventType.Error event as the sentinel, so the client receives the error before the stream closes.

[sf-sahil-jassal]

🔴 [STRUCTURE]: non-atomic createThread + saveMessages breaks idempotency — partial failures lose messages forever

Issue: Idempotency is keyed solely on thread existence (L172-176). But the thread is created (L189) and messages saved (L202) as two separate awaits with no transaction. If the process dies or saveMessages throws between them, the thread exists with zero/partial messages; the next run sees the thread, hits summary.skipped++, and never retries the messages — permanent silent data loss in a migration script.

Fix: Make the operation atomic (single transaction over thread+messages if the storage adapter supports it), or write messages first / verify message count on the idempotency check so a half-written thread is detected and completed on re-run rather than skipped.

Credit: Open Code

[rohit-sourcefuse]

Agreed — non-atomic create+save could strand a thread empty and skip it forever. Fixed in 824b7b5: an existing thread is now skipped only when it already holds all its messages; otherwise messages are re-saved (idempotent upsert on the stable source message id, so no duplicates), repairing a thread left behind by a run that died before saveMessages. Added an optional Memory.query probe to count persisted messages (falls back to re-save when unavailable).

[sf-sahil-jassal]

🟡 Improve: Stubbing the visualizer's own callGen via 'callGen' as any couples the test to a private member name

Issue: sinon.stub(visualizer, 'callGen' as any) stubs a protected method of the class under test, with two as any casts to defeat type-checking. callGen exists in production purely as a test seam wrapping ai.generateObject; the test reaches past the public getConfig surface into an internal name, so renaming callGen (a pure refactor) breaks every visualizer test. (Same pattern in line/pie visualizer specs.)

Fix: Mock the actual boundary — the ai module's generateObject — so the test exercises the real callGen indirection and survives internal renames:

import * as ai from 'ai';
sinon.stub(ai, 'generateObject').resolves({object: mockLLMResponse});

If the seam must stay, at least drop the as any by typing the stub against the real key.

Credit: Open Code

[sf-sahil-jassal]

🟡 Improve: Test drives Mastra's internal __registerMastra instead of the public registration path

Issue: Calling generateQueryWorkflow.__registerMastra(mastra) reaches into a double-underscore internal of @mastra/core. The workflows are already passed to the new Mastra({workflows}) constructor on line 15, which is the documented registration mechanism; the manual __registerMastra call duplicates that via a private API and will break on a Mastra minor bump. (Repeated at the improve/visualization cases.)

Fix: Resolve workflows through the singleton you already constructed — mastra.getWorkflow('generateQueryWorkflow') after the constructor registration — and drop the __registerMastra calls entirely.

Credit: Open Code

[sf-sahil-jassal]

🔴 CONSUMER REGRESSION: No upgrade / breaking-changes section for 3.x → 4.0

Issue: This PR is a major LangChain/LangGraph → Mastra migration with hard breaking changes for the two downstream services (both pin ^3.0.0), yet the README has no consumer-facing "Upgrading from 3.x" / breaking-changes section. The only migration text is the internal src/mastra/README.md (a maintainer v2→v3 node map) and a single testing note at L864. Consumers that today do @inject(AiIntegrationBindings.SmartLLM) llm: BaseChatModel, import {RunnableConfig} from the package, bind services.GetTablesNode / use GetTablesNode, or implement IDataSetStore will break at compile/runtime with no documented path:

• BaseChatModel → the tier bindings now resolve to AI-SDK LanguageModelV2 (breaks resource-management cv-data-extractor.service.ts, skill-query.service.ts; reporting consumers).
• RunnableConfig is no longer exported (breaks reporting sow-recommendation.tool.ts) → call-site must move to the Mastra tool ctx + re-exported createTool/Tool.
• GetTablesNode class + services.GetTablesNode binding deleted → replaced by exported pickRelevantTables (breaks reporting table-selection.acceptance.ts).
• IDataSetStore gains new required methods.

Fix: Add a top-level ## Upgrading from 3.x (Breaking changes) section enumerating each removed/changed symbol and its replacement, e.g.:

## Upgrading from 3.x (Breaking changes)
- Model tiers (`SmartLLM`/`CheapLLM`/`FileLLM`/...) now resolve to AI-SDK `LanguageModelV2`, **not** LangChain `BaseChatModel`. Replace `BaseChatModel` injections + LangChain call APIs with the AI-SDK `generate`/`stream` surface.
- `RunnableConfig` is removed. Migrate custom tools to Mastra `createTool` (re-exported from the package root) and read context from the tool execute ctx.
- `GetTablesNode` / `services.GetTablesNode` are removed. Use the exported `pickRelevantTables` seam for host table-selection tests.
- `IDataSetStore` requires new methods: <list>.

Credit: Open Code

[rohit-sourcefuse]

Done — added the "Upgrading from 3.x to 4.0 (Breaking changes)" section in 824b7b5, enumerating: binding value type BaseChatModel → AI-SDK LanguageModelV2; removed RunnableConfig/./nodes/./state exports + the pickRelevantTables replacement; new required IDataSetStore methods; Node 22 || 24 engines; and the dropped provider inference knobs (with a pointer to track the providerOptions re-wiring).

[sf-sahil-jassal]

🟡 IMPROVE: subpath exports use "type" instead of the "types" condition key

Issue: Every subpath conditional export points its declaration file at a "type" key (here and in the new ./db-query/testing, ./mastra, ./mastra-* entries). "type" is not a resolution condition Node or TypeScript recognises — TS only honours the "types" condition inside exports. Type resolution currently only works because of the parallel typesVersions fallback; the exports block itself contributes nothing to .d.ts resolution and is misleading. Under moduleResolution: "bundler"/"node16" consumers that don't consult typesVersions, subpath types could fail to resolve.

Fix: rename the condition key and order it before default:

"./mastra": {
  "types": "./dist/mastra/index.d.ts",
  "default": "./dist/mastra/index.js"
}

Apply to all subpath entries (pre-existing ./aws/./openrouter etc. too).

Credit: Open Code

[sf-sahil-jassal]

why is this removed ?? why it is not needed now ?

[sf-sahil-jassal]

50+ Sonar issues, conflicts as well

[sonarqubecloud]

SonarQube reviewer guide

Quality Gate passed

Issues
54 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
1.0% Duplication on New Code

See analysis details on SonarQube Cloud