@strav/brain 1.0.0-alpha.17 → 1.0.0-alpha.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@strav/brain",
-  "version": "1.0.0-alpha.17",
+  "version": "1.0.0-alpha.18",
   "description": "Strav AI module — unified Provider interface, BrainManager, threads, prompt caching, tools / agents / MCP. Anthropic + OpenAI providers; Gemini / DeepSeek follow.",
   "type": "module",
   "main": "./src/index.ts",
@@ -8,6 +8,7 @@
   "exports": {
     ".": "./src/index.ts",
     "./mcp": "./src/mcp/index.ts",
+    "./persistence": "./src/persistence/index.ts",
     "./zod": "./src/zod/index.ts"
   },
   "files": [
@@ -24,7 +25,8 @@
     "@anthropic-ai/sdk": "^0.100.0",
     "@google/genai": "^2.7.0",
     "@modelcontextprotocol/sdk": "^1.29.0",
-    "@strav/kernel": "1.0.0-alpha.17",
+    "@strav/database": "1.0.0-alpha.18",
+    "@strav/kernel": "1.0.0-alpha.18",
     "openai": "^6.0.0"
   },
   "peerDependencies": {

package/src/agent_generate_result.ts

@@ -27,4 +27,6 @@ export interface AgentGenerateResult<T = unknown> {
   stopReason: string
   /** Token usage from the single underlying `generate` call. */
   usage: ChatUsage
+  /** See `ChatResult.responseId`. */
+  responseId?: string
 }

package/src/agent_result.ts

@@ -29,4 +29,11 @@ export interface AgentResult {
   stopReason: string
   /** Token usage summed across every model call in the loop. */
   usage: ChatUsage
+  /**
+   * Final provider response id when the provider exposes stateful
+   * conversations (OpenAI Responses API). Captured from the last
+   * model turn so apps that persist the conversation can resume
+   * via `ChatOptions.previousResponseId`. Undefined elsewhere.
+   */
+  responseId?: string
 }

package/src/agent_runner.ts

@@ -26,8 +26,17 @@ import type { AgentStreamEvent } from './agent_stream_event.ts'
 import type { BrainManager } from './brain_manager.ts'
 import { BrainError } from './brain_error.ts'
 import type { OutputSchema } from './output_schema.ts'
-import type { ChatOptions, Message } from './types.ts'
+import type {
+  ChatOptions,
+  Message,
+  ToolUseBlock,
+} from './types.ts'
 import type { RunWithToolsOptions } from './provider.ts'
+import type {
+  SuspendedRun,
+  SuspendedState,
+  ToolResultInput,
+} from './suspended_run.ts'
 
 /**
  * Conditional return shape for `AgentRunner.run()`. With the default
@@ -42,16 +51,47 @@ import type { RunWithToolsOptions } from './provider.ts'
  */
 export type AgentRunResult<T> = [T] extends [never] ? AgentResult : AgentGenerateResult<T>
 
-export class AgentRunner<T = never> {
+/**
+ * Conditional return shape that flips when the runner has opted in
+ * to suspension via `.suspend(gate)`. The phantom `S` generic on
+ * `AgentRunner<T, S>` carries the bit; `S extends true` widens the
+ * union so callers must narrow with `isSuspended(...)` before
+ * touching `result.value` / `result.text`.
+ */
+export type AgentRunMaybeSuspended<T, S extends boolean> = [S] extends [true]
+  ? AgentRunResult<T> | SuspendedRun
+  : AgentRunResult<T>
+
+export class AgentRunner<T = never, S extends boolean = false> {
   private prompt: string | undefined
   private contextBag: Record<string, unknown> = {}
   private schema: OutputSchema<T> | undefined
+  private suspendGate:
+    | ((call: ToolUseBlock, context?: Record<string, unknown>) => boolean | Promise<boolean>)
+    | undefined
 
   constructor(
     private readonly brain: BrainManager,
     private readonly agent: Agent<unknown>,
   ) {}
 
+  /**
+   * Install a human-in-the-loop gate. Called before each tool
+   * execution inside the agent loop; when it returns `true`, the
+   * run pauses and `.run()` resolves with a `SuspendedRun` instead
+   * of `AgentResult`. Apps obtain results out-of-band and call
+   * `.resume(state, results)` to continue.
+   *
+   * Throws `BrainError` if the runner is also in structured-output
+   * mode (`.output(schema)`) — schema + suspend is a deferred slice.
+   */
+  suspend(
+    gate: (call: ToolUseBlock, context?: Record<string, unknown>) => boolean | Promise<boolean>,
+  ): AgentRunner<T, true> {
+    this.suspendGate = gate
+    return this as unknown as AgentRunner<T, true>
+  }
+
   /** Set the user input. Required before `run()`. */
   input(text: string): this {
     this.prompt = text
@@ -127,10 +167,15 @@ export class AgentRunner<T = never> {
     >
   }
 
-  async run(): Promise<AgentRunResult<T>> {
+  async run(): Promise<AgentRunMaybeSuspended<T, S>> {
     if (this.prompt === undefined) {
       throw new BrainError('AgentRunner.run: input() must be called before run().')
     }
+    if (this.suspendGate !== undefined && this.schema !== undefined) {
+      throw new BrainError(
+        'AgentRunner.run: `.suspend(...)` and `.output(schema)` cannot be combined in V1 — the schema variants don\'t yet model pause/resume. Run tools first with suspension, then call brain.generate(...) on the result for the structured summary.',
+      )
+    }
     const messages: Message[] = [{ role: 'user', content: this.prompt }]
 
     if (this.schema !== undefined) {
@@ -172,8 +217,39 @@ export class AgentRunner<T = never> {
       context: this.contextBag,
     }
     if (this.agent.mcpServers.length > 0) options.mcpServers = this.agent.mcpServers
+    if (this.suspendGate !== undefined) options.shouldSuspend = this.suspendGate
     const result = await this.brain.runTools(messages, this.agent.tools, options)
-    return result as AgentRunResult<T>
+    return result as AgentRunMaybeSuspended<T, S>
+  }
+
+  /**
+   * Resume a previously-suspended run. Takes the `SuspendedRun.state`
+   * snapshot and the results gathered for each `pendingToolCalls`
+   * entry; the loop continues from where it paused.
+   *
+   * The runner's `suspend()` gate carries over so the same
+   * approval logic applies to any further tool calls — pass a
+   * fresh gate via `suspend()` before `resume()` to change the
+   * policy.
+   */
+  async resume(
+    state: SuspendedState,
+    results: readonly ToolResultInput[],
+  ): Promise<AgentRunMaybeSuspended<T, true>> {
+    if (this.schema !== undefined) {
+      throw new BrainError(
+        'AgentRunner.resume: structured-output runners cannot be resumed in V1 — `.output(schema)` is incompatible with pause/resume.',
+      )
+    }
+    const options: RunWithToolsOptions = {
+      ...this.buildChatOptions(),
+      maxIterations: this.agent.maxIterations,
+      context: this.contextBag,
+    }
+    if (this.agent.mcpServers.length > 0) options.mcpServers = this.agent.mcpServers
+    if (this.suspendGate !== undefined) options.shouldSuspend = this.suspendGate
+    const result = await this.brain.resumeTools(state, results, this.agent.tools, options)
+    return result as AgentRunMaybeSuspended<T, true>
   }
 
   private buildChatOptions(): ChatOptions {

package/src/brain_manager.ts

@@ -40,7 +40,12 @@ import type {
   TranscribeOptions,
   TranscribeResult,
 } from './types.ts'
-import type { Provider, RunWithToolsOptions } from './provider.ts'
+import type {
+  Provider,
+  RunWithToolsOptions,
+  RunWithToolsOptionsWithSuspend,
+} from './provider.ts'
+import { appendResumeResults, type SuspendedRun, type SuspendedState, type ToolResultInput } from './suspended_run.ts'
 import type { Tool } from './tool.ts'
 import { DEFAULT_TIERS } from './brain_config.ts'
 
@@ -152,11 +157,21 @@ export class BrainManager {
    * implement `runWithTools` (V1: OpenAI / Gemini / DeepSeek providers
    * don't yet — only `AnthropicProvider`).
    */
+  runTools(
+    input: string | readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptionsWithSuspend,
+  ): Promise<AgentResult | SuspendedRun>
+  runTools(
+    input: string | readonly Message[],
+    tools: readonly Tool[],
+    options?: RunWithToolsOptions,
+  ): Promise<AgentResult>
   async runTools(
     input: string | readonly Message[],
     tools: readonly Tool[],
     options: RunWithToolsOptions = {},
-  ): Promise<AgentResult> {
+  ): Promise<AgentResult | SuspendedRun> {
     const provider = this.provider(options.provider)
     if (!provider.runWithTools) {
       throw new BrainError(
@@ -175,6 +190,42 @@ export class BrainManager {
     return provider.runWithTools(messages, tools, resolved)
   }
 
+  /**
+   * Resume a previously-suspended tool-use loop. Takes the
+   * `SuspendedRun.state` snapshot plus the results the integrator
+   * gathered for each `pendingToolCalls` entry; appends a `tool_result`
+   * block per entry; re-enters `runTools` so the model can continue
+   * (potentially suspending again on the next tool).
+   *
+   * Mid-batch invariant: every pending call MUST get a result —
+   * otherwise the provider rejects the next request because the
+   * assistant turn's `tool_use` blocks are no longer balanced.
+   * `resumeTools` throws `BrainError` when results are missing.
+   *
+   * The `previousResponseId` carried on the snapshot (when the
+   * provider supports stateful conversations) is threaded back via
+   * `options.previousResponseId` automatically — per-call
+   * `options.previousResponseId` wins if supplied explicitly.
+   */
+  async resumeTools(
+    state: SuspendedState,
+    results: readonly ToolResultInput[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptions = {},
+  ): Promise<AgentResult | SuspendedRun> {
+    const resumed = appendResumeResults(state, results)
+    const merged: RunWithToolsOptions = { ...options }
+    if (merged.previousResponseId === undefined && state.responseId !== undefined) {
+      merged.previousResponseId = state.responseId
+    }
+    const out = await this.runTools(
+      resumed,
+      tools,
+      merged as RunWithToolsOptionsWithSuspend,
+    )
+    return mergeResumeCounters(out, state)
+  }
+
   /**
    * Streaming variant of `generateWithTools`. Yields
    * `AgentStreamEvent<T>`s as the loop progresses; the terminal
@@ -189,6 +240,7 @@ export class BrainManager {
     tools: readonly Tool[],
     options: RunWithToolsOptions = {},
   ): AsyncIterable<AgentStreamEvent<T>> {
+    rejectShouldSuspend(options, 'streamGenerateWithTools')
     const provider = this.provider(options.provider)
     if (!provider.streamWithToolsAndSchema) {
       throw new BrainError(
@@ -220,6 +272,7 @@ export class BrainManager {
     tools: readonly Tool[],
     options: RunWithToolsOptions = {},
   ): Promise<AgentGenerateResult<T>> {
+    rejectShouldSuspend(options, 'generateWithTools')
     const provider = this.provider(options.provider)
     if (!provider.runWithToolsAndSchema) {
       throw new BrainError(
@@ -250,6 +303,7 @@ export class BrainManager {
     tools: readonly Tool[],
     options: RunWithToolsOptions = {},
   ): AsyncIterable<AgentStreamEvent> {
+    rejectShouldSuspend(options, 'streamTools')
     const provider = this.provider(options.provider)
     if (!provider.streamWithTools) {
       throw new BrainError(
@@ -411,3 +465,66 @@ function normalizeInput(input: string | readonly Message[]): readonly Message[]
   }
   return input
 }
+
+/**
+ * V1 scope guard. `shouldSuspend` is wired only into the non-
+ * streaming `runWithTools` loop; the streaming and schema variants
+ * don't yet model pause / resume, so silently ignoring would be
+ * worse than throwing. Apps that need both should run tools first
+ * (suspending as needed), then call `generate` for the structured
+ * summary in a separate step.
+ */
+/**
+ * Carry forward the pre-suspension iteration count + token usage so
+ * `result.iterations` / `result.usage` reflect the full run, not
+ * just the post-resume portion. When the resumed call suspends
+ * again, the new state's iterations + usage also get the carry-
+ * forward so apps see a running total across an arbitrary number
+ * of suspension cycles.
+ */
+function mergeResumeCounters(
+  out: AgentResult | SuspendedRun,
+  state: SuspendedState,
+): AgentResult | SuspendedRun {
+  // +1 accounts for the suspended round itself — at suspension time
+  // the loop hadn't yet incremented `iterations` (we paused mid-
+  // batch, before tool execution). Supplying results to resume
+  // effectively completes that round.
+  const carryIter = state.iterations + 1
+  if ('status' in out) {
+    return {
+      ...out,
+      state: {
+        ...out.state,
+        iterations: out.state.iterations + carryIter,
+        usage: addUsage(out.state.usage, state.usage),
+      },
+    }
+  }
+  return {
+    ...out,
+    iterations: out.iterations + carryIter,
+    usage: addUsage(out.usage, state.usage),
+  }
+}
+
+function addUsage(
+  a: SuspendedState['usage'],
+  b: SuspendedState['usage'],
+): SuspendedState['usage'] {
+  return {
+    inputTokens: a.inputTokens + b.inputTokens,
+    outputTokens: a.outputTokens + b.outputTokens,
+    cacheReadTokens: a.cacheReadTokens + b.cacheReadTokens,
+    cacheCreationTokens: a.cacheCreationTokens + b.cacheCreationTokens,
+  }
+}
+
+function rejectShouldSuspend(options: RunWithToolsOptions, entry: string): void {
+  if (options.shouldSuspend !== undefined) {
+    throw new BrainError(
+      `BrainManager.${entry}: \`shouldSuspend\` is only supported on \`runTools\` (the non-streaming + no-schema entrypoint) in V1. Run tools first with suspension, then call \`generate\` for the structured summary as a separate step.`,
+      { context: { entry } },
+    )
+  }
+}

package/src/index.ts

@@ -10,7 +10,11 @@
 export { Agent } from './agent.ts'
 export type { AgentGenerateResult } from './agent_generate_result.ts'
 export type { AgentResult } from './agent_result.ts'
-export { AgentRunner, type AgentRunResult } from './agent_runner.ts'
+export {
+  AgentRunner,
+  type AgentRunMaybeSuspended,
+  type AgentRunResult,
+} from './agent_runner.ts'
 export type { AgentStreamEvent } from './agent_stream_event.ts'
 export {
   type AnthropicProviderConfig,
@@ -33,6 +37,7 @@ export {
 } from './brain_manager.ts'
 export { BrainProvider } from './brain_provider.ts'
 export { defineTool, type DefineToolSpec } from './define_tool.ts'
+export { MCPClientPool, type MCPClientFactory } from './mcp/pool.ts'
 export type { MCPServer, MCPServerToolConfig } from './mcp_server.ts'
 export type { OutputSchema } from './output_schema.ts'
 export { AnthropicProvider } from './providers/anthropic_provider.ts'
@@ -42,7 +47,18 @@ export { OllamaProvider } from './providers/ollama_provider.ts'
 export { OpenAICompatProvider } from './providers/openai_compat_provider.ts'
 export { OpenAIProvider } from './providers/openai_provider.ts'
 export { OpenAIResponsesProvider } from './providers/openai_responses_provider.ts'
-export type { Provider, RunWithToolsOptions } from './provider.ts'
+export type {
+  Provider,
+  RunWithToolsOptions,
+  RunWithToolsOptionsWithSuspend,
+} from './provider.ts'
+export {
+  appendResumeResults,
+  isSuspended,
+  type SuspendedRun,
+  type SuspendedState,
+  type ToolResultInput,
+} from './suspended_run.ts'
 export { Thread, type ThreadOptions, type ThreadState } from './thread.ts'
 export type { Tool, ToolContext } from './tool.ts'
 export { ToolExecutionError } from './tool_execution_error.ts'
@@ -50,6 +66,8 @@ export type {
   ChatOptions,
   ChatResult,
   ChatUsage,
+  CompactConfig,
+  CompactionBlock,
   ContentBlock,
   AudioBlock,
   AudioSource,

package/src/mcp/client.ts

@@ -66,6 +66,15 @@ export class MCPClient {
   readonly server: MCPServer
   private readonly _client: Client
   private _connected = false
+  /**
+   * In-flight connect promise — set on the first concurrent
+   * `connect()` and cleared on settle. Subsequent callers that
+   * race against the first one await the same promise instead of
+   * each kicking off their own transport handshake. Necessary for
+   * pooled clients: a fresh `borrow()` followed by parallel
+   * `listTools()` + `callTool()` calls both hit the same connect.
+   */
+  private _connecting: Promise<void> | undefined
   private _transport: StreamableHTTPClientTransport | undefined
   private _authProvider: StoreBackedOAuthProvider | undefined
 
@@ -87,6 +96,14 @@ export class MCPClient {
 
   async connect(): Promise<void> {
     if (this._connected) return
+    if (this._connecting) return this._connecting
+    this._connecting = this._doConnect().finally(() => {
+      this._connecting = undefined
+    })
+    return this._connecting
+  }
+
+  private async _doConnect(): Promise<void> {
     const transport = this._buildTransport()
     this._transport = transport
     try {

package/src/mcp/index.ts

@@ -15,6 +15,7 @@ export {
   type MCPOAuthStore,
   MemoryOAuthStore,
 } from './oauth.ts'
+export { MCPClientPool, type MCPClientFactory } from './pool.ts'
 export {
   resolveMcpTools,
   type ResolveMcpToolsOptions,

package/src/mcp/pool.ts

@@ -0,0 +1,106 @@
+/**
+ * `MCPClientPool` — long-lived, per-server `MCPClient` cache.
+ *
+ * Default `resolveMcpTools` flow constructs a fresh `MCPClient` per
+ * call to `runTools` / `runWithTools` / etc., handshakes the
+ * Streamable HTTP transport, lists tools, executes them, then
+ * closes the transport in a `finally`. For one-shot calls that's
+ * fine. For long-running agent workers — chat servers, background
+ * job processors — the per-call handshake adds noticeable
+ * latency and burns connection slots upstream.
+ *
+ * The pool keeps one connected `MCPClient` per `(server.name,
+ * server.url)` pair for the lifetime of the pool. `borrow(server)`
+ * returns the pooled client (lazily creating + connecting on
+ * first use). When the pool is in play, `resolveMcpTools` skips
+ * the per-call `close()` — the pool owns the lifetime — so
+ * subsequent calls reuse the existing transport.
+ *
+ * Apps own the pool's lifetime. Construct one at app boot, hand it
+ * to every provider (or to `BrainProvider` if using the DI
+ * helper), and call `pool.close()` on shutdown.
+ *
+ * ```ts
+ * const pool = new MCPClientPool()
+ *
+ * const openai = new OpenAIProvider(
+ *   'openai',
+ *   { driver: 'openai', apiKey: ... },
+ *   { mcpPool: pool },
+ * )
+ *
+ * // ... many runTools calls later, on graceful shutdown:
+ * await pool.close()
+ * ```
+ *
+ * Concurrency: `borrow()` is synchronous; `MCPClient.connect()`
+ * itself dedupes concurrent calls. Two parallel `runTools` calls
+ * sharing the same pooled client both await one handshake.
+ *
+ * Re-auth: when a borrowed client throws `MCPAuthRequiredError`,
+ * the pool keeps the (still un-authorized) client. Apps call
+ * `pool.evict(server)` after running `completeAuthorization` on
+ * a fresh client so subsequent borrows see the renewed state —
+ * or just reuse the same client the app authorized via the
+ * standard `MCPClient.completeAuthorization` flow.
+ */
+
+import type { MCPServer } from '../mcp_server.ts'
+import { MCPClient } from './client.ts'
+
+/** Internal — factory injection for tests. Defaults to `new MCPClient(server)`. */
+export type MCPClientFactory = (server: MCPServer) => MCPClient
+
+export class MCPClientPool {
+  private readonly clients: Map<string, MCPClient> = new Map()
+  private readonly factory: MCPClientFactory
+
+  constructor(factory: MCPClientFactory = (s) => new MCPClient(s)) {
+    this.factory = factory
+  }
+
+  /**
+   * Return the pooled client for `server`, constructing + caching it on
+   * first call. The client is NOT eagerly connected — the first
+   * `listTools` / `callTool` invocation triggers `connect()` once.
+   */
+  borrow(server: MCPServer): MCPClient {
+    const key = poolKey(server)
+    const existing = this.clients.get(key)
+    if (existing) return existing
+    const client = this.factory(server)
+    this.clients.set(key, client)
+    return client
+  }
+
+  /**
+   * Drop the cached client for `server` and close its transport.
+   * Useful after the app re-authorizes an OAuth server, or after a
+   * transient failure where the connection state is suspect and a
+   * fresh handshake on next borrow is preferable.
+   */
+  async evict(server: MCPServer): Promise<void> {
+    const key = poolKey(server)
+    const client = this.clients.get(key)
+    if (!client) return
+    this.clients.delete(key)
+    await client.close()
+  }
+
+  /** Close every pooled client. Call on app shutdown. */
+  async close(): Promise<void> {
+    const all = [...this.clients.values()]
+    this.clients.clear()
+    await Promise.all(all.map((c) => c.close()))
+  }
+
+  /** Whether the pool currently holds a client for `server`. Used by tests. */
+  has(server: MCPServer): boolean {
+    return this.clients.has(poolKey(server))
+  }
+}
+
+/** Pool key: name + url, so two `MCPServer`s with the same name but different URLs don't collide. */
+function poolKey(server: MCPServer): string {
+  return `${server.name}|${server.url}`
+}

package/src/mcp/resolve_mcp_tools.ts

@@ -21,6 +21,7 @@
 import type { MCPServer } from '../mcp_server.ts'
 import type { Tool, ToolContext } from '../tool.ts'
 import { MCPClient } from './client.ts'
+import type { MCPClientPool } from './pool.ts'
 
 export interface ResolvedMcpTools {
   tools: Tool[]
@@ -30,6 +31,14 @@ export interface ResolvedMcpTools {
 export interface ResolveMcpToolsOptions {
   /** Override the client factory — tests inject mock clients per server here. */
   clientFactory?(server: MCPServer): MCPClient
+  /**
+   * When set, clients are borrowed from the pool instead of being
+   * constructed fresh per call, and the returned `close` becomes a
+   * no-op — the pool owns the lifetime, and apps call
+   * `pool.close()` on shutdown. Mutually beneficial with
+   * `clientFactory` (tests pass a factory to the pool itself).
+   */
+  pool?: MCPClientPool
 }
 
 const NAME_SEPARATOR = '__'
@@ -40,13 +49,16 @@ export async function resolveMcpTools(
 ): Promise<ResolvedMcpTools> {
   const clients: MCPClient[] = []
   const tools: Tool[] = []
+  const pooled = options.pool !== undefined
 
   for (const server of servers) {
     if (server.tools?.enabled === false) continue
-    const client = options.clientFactory
-      ? options.clientFactory(server)
-      : new MCPClient(server)
-    clients.push(client)
+    const client = options.pool
+      ? options.pool.borrow(server)
+      : options.clientFactory
+        ? options.clientFactory(server)
+        : new MCPClient(server)
+    if (!pooled) clients.push(client)
 
     const allowed = server.tools?.allowedTools
     const allowedSet = allowed ? new Set(allowed) : null
@@ -60,9 +72,15 @@ export async function resolveMcpTools(
 
   return {
     tools,
-    close: async () => {
-      await Promise.all(clients.map((c) => c.close()))
-    },
+    // Pooled clients live across calls — `close` becomes a no-op
+    // and the pool owns the lifetime. Non-pooled clients close
+    // here so each `runWithTools` invocation cleans up its own
+    // transports.
+    close: pooled
+      ? async () => {}
+      : async () => {
+          await Promise.all(clients.map((c) => c.close()))
+        },
   }
 }
 

package/src/persistence/brain_message.ts

@@ -0,0 +1,34 @@
+/**
+ * `BrainMessage` — the typed row of `brain_message`. One per turn.
+ *
+ * `content` mirrors `Message.content` — string for plain text or
+ * `ContentBlock[]` when the turn carries structured blocks
+ * (tool_use, tool_result, image, compaction, ...). JSONB hydration
+ * is automatic.
+ *
+ * Assistant turns carry `model` / `usage` / `stop_reason` /
+ * `response_id`; user turns leave them NULL. The repository's
+ * `appendTurn` helper writes the right shape per role.
+ */
+
+import { Model } from '@strav/database'
+import type { ChatUsage, ContentBlock } from '../types.ts'
+import { brainMessageSchema } from './schema/brain_message_schema.ts'
+
+export type BrainMessageRole = 'user' | 'assistant'
+
+export class BrainMessage extends Model {
+  static override readonly schema = brainMessageSchema
+
+  id!: string
+  tenant_id!: string
+  thread_id!: string
+  turn_index!: number
+  role!: BrainMessageRole
+  content!: string | ContentBlock[]
+  model!: string | null
+  usage!: ChatUsage | null
+  stop_reason!: string | null
+  response_id!: string | null
+  created_at!: Date
+}