@strav/brain 1.0.0-alpha.9 → 1.0.2

package/src/drivers/gemini/gemini_brain_driver.ts

@@ -0,0 +1,1064 @@
+/**
+ * `GeminiBrainDriver` — implementation of `Provider` backed by the
+ * official `@google/genai` SDK (Gemini Developer API / Vertex AI).
+ *
+ * Maps framework shapes to Gemini's wire format:
+ *
+ *   - `system` → `config.systemInstruction` (string-joined when
+ *     multi-block). Cache flags on the system prompt are ignored —
+ *     Gemini's prompt caching uses an explicit Caches API rather
+ *     than per-block flags, so `cache: true` becomes a no-op
+ *     consistent with the OpenAI provider.
+ *
+ *   - `Message[]` → `Content[]`. Framework `role: 'user' | 'assistant'`
+ *     maps to Gemini's `role: 'user' | 'model'`. String content
+ *     becomes a single `{text}` part; `ContentBlock[]` content fans
+ *     out:
+ *       - `TextBlock`         → `{text}`
+ *       - `ToolUseBlock`      → `{functionCall: {id, name, args}}`
+ *       - `ToolResultBlock`   → `{functionResponse: {id, name,
+ *                                  response: {result | error}}}`
+ *       - `MCP*` blocks       → silently dropped (Anthropic-only).
+ *
+ *   - `Tool[]` → `[{functionDeclarations: [{name, description,
+ *     parametersJsonSchema: inputSchema}]}]`. We use
+ *     `parametersJsonSchema` (not `parameters`) so JSON-Schema-shaped
+ *     tool inputs pass through verbatim without translation to
+ *     Gemini's `Schema` form.
+ *
+ *   - `MCPServer[]` → resolved via the local MCP client
+ *     (`@strav/brain/mcp`). Discovered tools are namespaced
+ *     `<server>__<tool>` and merged with caller-supplied tools.
+ *     Transports are closed in a `finally` once the loop exits.
+ *     Gemini has no first-party server-side MCP equivalent to
+ *     Anthropic's connector.
+ *
+ *   - `thinking: 'adaptive'` → `thinkingConfig: { thinkingBudget: -1 }`
+ *     (auto). `'disabled'` → `thinkingConfig: { thinkingBudget: 0 }`.
+ *     Explicit `effort` (`low`/`medium`/`high`/`xhigh`/`max`) maps to
+ *     `thinkingConfig.thinkingLevel`. Non-thinking models ignore the
+ *     field upstream — we always emit, the SDK rejects only for
+ *     models that don't support it.
+ *
+ *   - `cache: true` → no-op. Gemini's prompt cache lives behind the
+ *     `Caches` API; same accepted-silently behavior as OpenAI.
+ *
+ *   - `countTokens` IS implemented — `ai.models.countTokens` exists
+ *     and is cheap. Returns `totalTokens`.
+ */
+
+import { GoogleGenAI, ThinkingLevel } from '@google/genai'
+import type {
+  Content,
+  FunctionDeclaration,
+  GenerateContentConfig,
+  GenerateContentParameters,
+  GenerateContentResponse,
+  Part,
+} from '@google/genai'
+import type { AgentResult } from '../../agent_result.ts'
+import { BrainError } from '../../brain_error.ts'
+import type { GeminiProviderConfig } from '../../brain_config.ts'
+import type { MCPServer } from '../../mcp_server.ts'
+import type { AgentGenerateResult } from '../../agent_generate_result.ts'
+import type { AgentStreamEvent } from '../../agent_stream_event.ts'
+import type {
+  AudioSource,
+  EmbedOptions,
+  EmbedResult,
+  ServerTool,
+  TranscribeOptions,
+  TranscribeResult,
+} from '../../types.ts'
+import { resolveMcpTools, type ResolveMcpToolsOptions } from '../../mcp/resolve_mcp_tools.ts'
+import { parseGenerated, type OutputSchema } from '../../output_schema.ts'
+import { runToolWithRecovery } from '../../tool_runner.ts'
+import type {
+  BrainDriver,
+  RunWithToolsOptions,
+  RunWithToolsOptionsWithSuspend,
+} from '../../brain_driver.ts'
+import type { SuspendedRun } from '../../suspended_run.ts'
+import type { Tool } from '../../tool.ts'
+import type {
+  ChatOptions,
+  ChatResult,
+  ChatUsage,
+  ContentBlock,
+  GenerateResult,
+  Message,
+  StreamEvent,
+  SystemPrompt,
+  TextBlock,
+  ToolResultBlock,
+  ToolUseBlock,
+} from '../../types.ts'
+
+const DEFAULT_GEMINI_MODEL = 'gemini-2.5-flash'
+const DEFAULT_GEMINI_EMBED_MODEL = 'text-embedding-004'
+
+/**
+ * The slice of `GoogleGenAI` the provider exercises. Narrowed so
+ * tests can inject a stub without satisfying the full SDK surface.
+ */
+export interface GeminiModelsClient {
+  generateContent(params: GenerateContentParameters): Promise<GenerateContentResponse>
+  generateContentStream(
+    params: GenerateContentParameters,
+  ): Promise<AsyncIterable<GenerateContentResponse>>
+  countTokens(params: { model: string; contents: Content[] }): Promise<{ totalTokens?: number }>
+  /**
+   * Optional on the test seam — the real SDK always provides it,
+   * but tests that don't exercise embed don't need to stub it.
+   * `embed()` calls this directly; missing it throws a clear
+   * TypeError if invoked.
+   */
+  embedContent?(params: {
+    model: string
+    contents: string[]
+    config?: { outputDimensionality?: number; abortSignal?: AbortSignal }
+  }): Promise<{ embeddings?: Array<{ values?: number[] }> }>
+}
+
+export interface GeminiProviderOptions {
+  client?: { models: GeminiModelsClient }
+  /** Internal seam — tests inject a stub MCP client factory. */
+  mcpClientFactory?: ResolveMcpToolsOptions['clientFactory']
+  /** See `OpenAIProviderOptions.mcpPool` — same semantics. */
+  mcpPool?: ResolveMcpToolsOptions['pool']
+}
+
+export class GeminiBrainDriver implements BrainDriver {
+  readonly name: string
+  private readonly models: GeminiModelsClient
+  private readonly defaultModel: string
+  private readonly defaultMaxTokens: number
+  private readonly defaultEmbedModel: string
+  private readonly mcpClientFactory?: ResolveMcpToolsOptions['clientFactory']
+  private readonly mcpPool?: ResolveMcpToolsOptions['pool']
+
+  constructor(name: string, config: GeminiProviderConfig, options: GeminiProviderOptions = {}) {
+    this.name = name
+    this.defaultModel = config.defaultModel ?? DEFAULT_GEMINI_MODEL
+    this.defaultMaxTokens = config.defaultMaxTokens ?? 4096
+    this.defaultEmbedModel = config.defaultEmbedModel ?? DEFAULT_GEMINI_EMBED_MODEL
+    this.mcpClientFactory = options.mcpClientFactory
+    this.mcpPool = options.mcpPool
+    if (options.client) {
+      this.models = options.client.models
+    } else {
+      const httpOpts =
+        config.baseUrl !== undefined || config.apiVersion !== undefined
+          ? {
+              ...(config.baseUrl !== undefined ? { baseUrl: config.baseUrl } : {}),
+              ...(config.apiVersion !== undefined ? { apiVersion: config.apiVersion } : {}),
+            }
+          : undefined
+      const sdk = new GoogleGenAI({
+        apiKey: config.apiKey,
+        ...(httpOpts ? { httpOptions: httpOpts } : {}),
+      })
+      this.models = sdk.models as unknown as GeminiModelsClient
+    }
+  }
+
+  async chat(messages: readonly Message[], options: ChatOptions = {}): Promise<ChatResult> {
+    const params = this.buildParams(messages, options, [])
+    const response = await this.models.generateContent(params)
+    return this.toChatResult(response, params.model)
+  }
+
+  async *stream(
+    messages: readonly Message[],
+    options: ChatOptions = {},
+  ): AsyncIterable<StreamEvent> {
+    const params = this.buildParams(messages, options, [])
+    const stream = await this.models.generateContentStream(params)
+    let finishReason: string | null = null
+    let lastUsage: ChatUsage | undefined
+    for await (const chunk of stream) {
+      const candidate = chunk.candidates?.[0]
+      const text = candidateText(candidate)
+      if (text.length > 0) yield { type: 'text', delta: text }
+      if (candidate?.finishReason) finishReason = String(candidate.finishReason)
+      if (chunk.usageMetadata) lastUsage = toUsage(chunk.usageMetadata)
+    }
+    yield {
+      type: 'stop',
+      stopReason: finishReason,
+      usage: lastUsage ?? {
+        inputTokens: 0,
+        outputTokens: 0,
+        cacheReadTokens: 0,
+        cacheCreationTokens: 0,
+      },
+    }
+  }
+
+  async countTokens(messages: readonly Message[], options: ChatOptions = {}): Promise<number> {
+    const contents = this.toContents(messages)
+    const model = options.model ?? this.defaultModel
+    const response = await this.models.countTokens({ model, contents })
+    return response.totalTokens ?? 0
+  }
+
+  /**
+   * Gemini embeddings via `ai.models.embedContent`. Returns one
+   * vector per input text. `usage.inputTokens` is `0` — Gemini's
+   * embed endpoint doesn't surface token counts in the response
+   * for the Gemini Developer API tier (Vertex's request-level
+   * metadata exposes billable characters, but that's a different
+   * accounting unit and not the framework's contract). Apps that
+   * need exact embed-token usage call `countTokens` separately
+   * before the call.
+   */
+  /**
+   * Gemini has no dedicated transcription endpoint, so we wrap a
+   * chat call: an AudioBlock + a system message that tells the
+   * model to transcribe verbatim. Apps that want OpenAI-style
+   * Whisper transcription with `language` / `duration` metadata
+   * route to OpenAI (or local Whisper via Ollama).
+   *
+   * `options.prompt` threads into the system instruction —
+   * useful for style/vocabulary hints. `options.language` is
+   * surfaced to the model in the system prompt (Gemini doesn't
+   * have a dedicated language field).
+   */
+  async transcribe(
+    audio: AudioSource,
+    options: TranscribeOptions = {},
+  ): Promise<TranscribeResult> {
+    const lines = [
+      'Transcribe the attached audio verbatim. Output ONLY the transcribed text — no preamble, no quotes, no commentary.',
+      options.language ? `Audio language: ${options.language}.` : undefined,
+      options.prompt ? `Style / vocabulary hints: ${options.prompt}` : undefined,
+    ].filter((s): s is string => s !== undefined)
+    const system = lines.join(' ')
+    const chatResult = await this.chat(
+      [
+        {
+          role: 'user',
+          content: [{ type: 'audio', source: audio }],
+        },
+      ],
+      {
+        system,
+        ...(options.model !== undefined ? { model: options.model } : {}),
+        ...(options.signal !== undefined ? { signal: options.signal } : {}),
+      },
+    )
+    return {
+      text: chatResult.text,
+      model: chatResult.model,
+      raw: chatResult.raw,
+    }
+  }
+
+  async embed(
+    texts: readonly string[],
+    options: EmbedOptions = {},
+  ): Promise<EmbedResult<{ embeddings?: Array<{ values?: number[] }> }>> {
+    const model = options.model ?? this.defaultEmbedModel
+    const config: { outputDimensionality?: number; abortSignal?: AbortSignal } = {}
+    if (options.dimensions !== undefined) config.outputDimensionality = options.dimensions
+    if (options.signal !== undefined) config.abortSignal = options.signal
+    if (!this.models.embedContent) {
+      throw new BrainError(
+        `GeminiBrainDriver.embed: underlying SDK does not implement embedContent. This usually means a test stub omitted it.`,
+        { context: { provider: this.name } },
+      )
+    }
+    const response = await this.models.embedContent({
+      model,
+      contents: texts as string[],
+      ...(Object.keys(config).length > 0 ? { config } : {}),
+    })
+    const embeddings = (response.embeddings ?? []).map((e) => e.values ?? [])
+    return {
+      embeddings,
+      model,
+      usage: { inputTokens: 0 },
+      raw: response,
+    }
+  }
+
+  /**
+   * Resolve MCP tool descriptors for `servers`, threading the
+   * provider's optional `clientFactory` (test seam) and `mcpPool`
+   * (long-lived connections) through. Caller invokes
+   * `resolved.close()` in `finally` — a no-op when the pool owns
+   * lifetimes.
+   */
+  private resolveMcp(servers: readonly MCPServer[]): Promise<{
+    tools: Tool[]
+    close: () => Promise<void>
+  }> {
+    if (servers.length === 0) {
+      return Promise.resolve({ tools: [], close: async () => {} })
+    }
+    return resolveMcpTools(servers, {
+      ...(this.mcpClientFactory ? { clientFactory: this.mcpClientFactory } : {}),
+      ...(this.mcpPool ? { pool: this.mcpPool } : {}),
+    })
+  }
+
+  runWithTools(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptionsWithSuspend,
+  ): Promise<AgentResult | SuspendedRun>
+  runWithTools(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options?: RunWithToolsOptions,
+  ): Promise<AgentResult>
+  async runWithTools(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptions = {},
+  ): Promise<AgentResult | SuspendedRun> {
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
+    try {
+      return await this._runLoop(messages, [...tools, ...resolved.tools], options)
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async _runLoop(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptions,
+  ): Promise<AgentResult | SuspendedRun> {
+    const maxIterations = options.maxIterations ?? 10
+    const toolMap = new Map<string, Tool>(tools.map((t) => [t.name, t]))
+    const workingMessages: Message[] = [...messages]
+    const aggregated: ChatUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheReadTokens: 0,
+      cacheCreationTokens: 0,
+    }
+    let iterations = 0
+
+    while (true) {
+      checkAborted(options.signal)
+      const params = this.buildParams(workingMessages, options, tools)
+      const response = await this.models.generateContent(params)
+      addUsage(aggregated, response.usageMetadata)
+
+      const candidate = response.candidates?.[0]
+      if (!candidate) {
+        throw new BrainError('GeminiBrainDriver: response had no candidates.')
+      }
+      const parts = candidate.content?.parts ?? []
+      const assistantContent = fromGeminiParts(parts)
+      workingMessages.push({ role: 'assistant', content: assistantContent })
+
+      const toolUses = (Array.isArray(assistantContent) ? assistantContent : []).filter(
+        (b): b is ToolUseBlock => b.type === 'tool_use',
+      )
+
+      if (toolUses.length === 0) {
+        return {
+          text: typeof assistantContent === 'string'
+            ? assistantContent
+            : candidateText(candidate),
+          messages: workingMessages,
+          iterations,
+          stopReason: candidate.finishReason ? String(candidate.finishReason) : 'stop',
+          usage: aggregated,
+        }
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (let i = 0; i < toolUses.length; i++) {
+        const call = toolUses[i]!
+        if (options.shouldSuspend && await options.shouldSuspend(call, options.context)) {
+          return {
+            status: 'suspended',
+            pendingToolCalls: toolUses.slice(i),
+            state: { messages: workingMessages, iterations, usage: aggregated },
+          }
+        }
+        const { content, isError } = await runToolWithRecovery(
+          toolMap.get(call.name),
+          call.name,
+          call.id,
+          call.input,
+          options,
+        )
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        return {
+          text: candidateText(candidate),
+          messages: workingMessages,
+          iterations,
+          stopReason: 'max_iterations',
+          usage: aggregated,
+        }
+      }
+    }
+  }
+
+  async runWithToolsAndSchema<T>(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions = {},
+  ): Promise<AgentGenerateResult<T>> {
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
+    try {
+      return await this._runLoopWithSchema([...tools, ...resolved.tools], messages, schema, options)
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async _runLoopWithSchema<T>(
+    tools: readonly Tool[],
+    messages: readonly Message[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions,
+  ): Promise<AgentGenerateResult<T>> {
+    const maxIterations = options.maxIterations ?? 10
+    const toolMap = new Map<string, Tool>(tools.map((t) => [t.name, t]))
+    const workingMessages: Message[] = [...messages]
+    const aggregated: ChatUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheReadTokens: 0,
+      cacheCreationTokens: 0,
+    }
+    let iterations = 0
+
+    while (true) {
+      const params = this.buildParams(workingMessages, options, tools)
+      params.config = {
+        ...(params.config ?? {}),
+        responseMimeType: 'application/json',
+        responseJsonSchema: schema.jsonSchema,
+      }
+      const response = await this.models.generateContent(params)
+      addUsage(aggregated, response.usageMetadata)
+
+      const candidate = response.candidates?.[0]
+      if (!candidate) {
+        throw new BrainError('GeminiBrainDriver: response had no candidates.')
+      }
+      const parts = candidate.content?.parts ?? []
+      const assistantContent = fromGeminiParts(parts)
+      workingMessages.push({ role: 'assistant', content: assistantContent })
+
+      const toolUses = (Array.isArray(assistantContent) ? assistantContent : []).filter(
+        (b): b is ToolUseBlock => b.type === 'tool_use',
+      )
+
+      if (toolUses.length === 0) {
+        const text = typeof assistantContent === 'string'
+          ? assistantContent
+          : candidateText(candidate)
+        return {
+          value: parseGenerated(text, schema),
+          text,
+          messages: workingMessages,
+          iterations,
+          stopReason: candidate.finishReason ? String(candidate.finishReason) : 'stop',
+          usage: aggregated,
+        }
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of toolUses) {
+        const { content, isError } = await runToolWithRecovery(
+          toolMap.get(call.name),
+          call.name,
+          call.id,
+          call.input,
+          options,
+        )
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        const text = candidateText(candidate)
+        return {
+          value: parseGenerated(text, schema),
+          text,
+          messages: workingMessages,
+          iterations,
+          stopReason: 'max_iterations',
+          usage: aggregated,
+        }
+      }
+    }
+  }
+
+  async *streamWithTools(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptions = {},
+  ): AsyncIterable<AgentStreamEvent> {
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
+    try {
+      yield* this._streamLoop(messages, [...tools, ...resolved.tools], options)
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async *_streamLoop(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptions,
+  ): AsyncIterable<AgentStreamEvent> {
+    const maxIterations = options.maxIterations ?? 10
+    const toolMap = new Map<string, Tool>(tools.map((t) => [t.name, t]))
+    const workingMessages: Message[] = [...messages]
+    const aggregated: ChatUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheReadTokens: 0,
+      cacheCreationTokens: 0,
+    }
+    let iterations = 0
+
+    while (true) {
+      checkAborted(options.signal)
+      yield { type: 'iteration_start', iteration: iterations }
+
+      const params = this.buildParams(workingMessages, options, tools)
+      const stream = await this.models.generateContentStream(params)
+
+      const accumulatedParts: Part[] = []
+      let finishReason: string | null = null
+      let lastUsage: ChatUsage | undefined
+
+      for await (const chunk of stream) {
+        const candidate = chunk.candidates?.[0]
+        const chunkParts = candidate?.content?.parts ?? []
+        for (const part of chunkParts) {
+          if (typeof part.text === 'string' && part.text.length > 0) {
+            yield { type: 'text', delta: part.text }
+          }
+        }
+        accumulatedParts.push(...chunkParts)
+        if (candidate?.finishReason) finishReason = String(candidate.finishReason)
+        if (chunk.usageMetadata) lastUsage = toUsage(chunk.usageMetadata)
+      }
+      if (lastUsage) {
+        aggregated.inputTokens += lastUsage.inputTokens
+        aggregated.outputTokens += lastUsage.outputTokens
+        aggregated.cacheReadTokens += lastUsage.cacheReadTokens
+      }
+
+      yield { type: 'iteration_end', iteration: iterations, stopReason: finishReason }
+
+      const assistantContent = fromGeminiParts(accumulatedParts)
+      workingMessages.push({ role: 'assistant', content: assistantContent })
+
+      const toolUses = (Array.isArray(assistantContent) ? assistantContent : []).filter(
+        (b): b is ToolUseBlock => b.type === 'tool_use',
+      )
+
+      if (toolUses.length === 0) {
+        yield {
+          type: 'stop',
+          stopReason: finishReason ?? 'stop',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+        }
+        return
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of toolUses) {
+        yield { type: 'tool_use', id: call.id, name: call.name, input: call.input }
+        const { content, isError } = await runToolWithRecovery(
+          toolMap.get(call.name),
+          call.name,
+          call.id,
+          call.input,
+          options,
+        )
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+        yield {
+          type: 'tool_result',
+          id: call.id,
+          name: call.name,
+          content,
+          isError,
+        }
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        yield {
+          type: 'stop',
+          stopReason: 'max_iterations',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+        }
+        return
+      }
+    }
+  }
+
+  async *streamWithToolsAndSchema<T>(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions = {},
+  ): AsyncIterable<AgentStreamEvent<T>> {
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
+    try {
+      yield* this._streamLoopWithSchema(
+        [...tools, ...resolved.tools],
+        messages,
+        schema,
+        options,
+      )
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async *_streamLoopWithSchema<T>(
+    tools: readonly Tool[],
+    messages: readonly Message[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions,
+  ): AsyncIterable<AgentStreamEvent<T>> {
+    const maxIterations = options.maxIterations ?? 10
+    const toolMap = new Map<string, Tool>(tools.map((t) => [t.name, t]))
+    const workingMessages: Message[] = [...messages]
+    const aggregated: ChatUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheReadTokens: 0,
+      cacheCreationTokens: 0,
+    }
+    let iterations = 0
+
+    while (true) {
+      checkAborted(options.signal)
+      yield { type: 'iteration_start', iteration: iterations }
+
+      const params = this.buildParams(workingMessages, options, tools)
+      params.config = {
+        ...(params.config ?? {}),
+        responseMimeType: 'application/json',
+        responseJsonSchema: schema.jsonSchema,
+      }
+      const stream = await this.models.generateContentStream(params)
+
+      const accumulatedParts: Part[] = []
+      let textBuf = ''
+      let finishReason: string | null = null
+      let lastUsage: ChatUsage | undefined
+
+      for await (const chunk of stream) {
+        const candidate = chunk.candidates?.[0]
+        const chunkParts = candidate?.content?.parts ?? []
+        for (const part of chunkParts) {
+          if (typeof part.text === 'string' && part.text.length > 0) {
+            textBuf += part.text
+            yield { type: 'text', delta: part.text }
+          }
+        }
+        accumulatedParts.push(...chunkParts)
+        if (candidate?.finishReason) finishReason = String(candidate.finishReason)
+        if (chunk.usageMetadata) lastUsage = toUsage(chunk.usageMetadata)
+      }
+      if (lastUsage) {
+        aggregated.inputTokens += lastUsage.inputTokens
+        aggregated.outputTokens += lastUsage.outputTokens
+        aggregated.cacheReadTokens += lastUsage.cacheReadTokens
+      }
+
+      yield { type: 'iteration_end', iteration: iterations, stopReason: finishReason }
+
+      const assistantContent = fromGeminiParts(accumulatedParts)
+      workingMessages.push({ role: 'assistant', content: assistantContent })
+
+      const toolUses = (Array.isArray(assistantContent) ? assistantContent : []).filter(
+        (b): b is ToolUseBlock => b.type === 'tool_use',
+      )
+
+      if (toolUses.length === 0) {
+        const text = textBuf
+        const value = parseGenerated(text, schema)
+        yield {
+          type: 'stop',
+          stopReason: finishReason ?? 'stop',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+          value,
+          text,
+        } as AgentStreamEvent<T>
+        return
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of toolUses) {
+        yield { type: 'tool_use', id: call.id, name: call.name, input: call.input }
+        const { content, isError } = await runToolWithRecovery(
+          toolMap.get(call.name),
+          call.name,
+          call.id,
+          call.input,
+          options,
+        )
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+        yield {
+          type: 'tool_result',
+          id: call.id,
+          name: call.name,
+          content,
+          isError,
+        }
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        const text = textBuf
+        const value = parseGenerated(text, schema)
+        yield {
+          type: 'stop',
+          stopReason: 'max_iterations',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+          value,
+          text,
+        } as AgentStreamEvent<T>
+        return
+      }
+    }
+  }
+
+  async generate<T>(
+    messages: readonly Message[],
+    schema: OutputSchema<T>,
+    options: ChatOptions = {},
+  ): Promise<GenerateResult<T>> {
+    const params = this.buildParams(messages, options, [])
+    params.config = {
+      ...(params.config ?? {}),
+      responseMimeType: 'application/json',
+      responseJsonSchema: schema.jsonSchema,
+    }
+    const response = await this.models.generateContent(params)
+    const candidate = response.candidates?.[0]
+    const text = candidateText(candidate)
+    const value = parseGenerated(text, schema)
+    return {
+      value,
+      text,
+      model: response.modelVersion ?? params.model,
+      stopReason: candidate?.finishReason ? String(candidate.finishReason) : null,
+      usage: toUsage(response.usageMetadata),
+      raw: response,
+    }
+  }
+
+  // ─── Param translation ──────────────────────────────────────────────────
+
+  private buildParams(
+    messages: readonly Message[],
+    options: ChatOptions,
+    tools: readonly Tool[],
+  ): GenerateContentParameters {
+    const model = options.model ?? this.defaultModel
+    const contents = this.toContents(messages)
+    const config: GenerateContentConfig = {
+      maxOutputTokens: options.maxTokens ?? this.defaultMaxTokens,
+    }
+
+    const systemText = systemPromptText(options.system)
+    if (systemText.length > 0) {
+      config.systemInstruction = systemText
+    }
+
+    const configTools: NonNullable<GenerateContentConfig['tools']> = []
+    if (tools.length > 0) {
+      const functionDeclarations: FunctionDeclaration[] = tools.map((t) => ({
+        name: t.name,
+        description: t.description,
+        parametersJsonSchema: t.inputSchema,
+      }))
+      configTools.push({ functionDeclarations })
+    }
+    if (options.serverTools && options.serverTools.length > 0) {
+      configTools.push(...geminiServerTools(options.serverTools))
+    }
+    if (configTools.length > 0) {
+      config.tools = configTools
+    }
+
+    const thinking = buildThinkingConfig(options)
+    if (thinking !== undefined) config.thinkingConfig = thinking
+
+    if (options.signal !== undefined) config.abortSignal = options.signal
+
+    return { model, contents, config }
+  }
+
+  private toContents(messages: readonly Message[]): Content[] {
+    return messages.map((m) => ({
+      role: m.role === 'assistant' ? 'model' : 'user',
+      parts: toGeminiParts(m.content),
+    }))
+  }
+
+  private toChatResult(
+    response: GenerateContentResponse,
+    requestedModel: string,
+  ): ChatResult<GenerateContentResponse> {
+    const candidate = response.candidates?.[0]
+    return {
+      text: candidateText(candidate),
+      model: response.modelVersion ?? requestedModel,
+      stopReason: candidate?.finishReason ? String(candidate.finishReason) : null,
+      usage: toUsage(response.usageMetadata),
+      raw: response,
+    }
+  }
+}
+
+// ─── Shape converters ─────────────────────────────────────────────────────
+
+/** Throw a DOMException-shaped abort error if the signal has fired. */
+function checkAborted(signal: AbortSignal | undefined): void {
+  if (signal?.aborted) {
+    throw signal.reason ?? new DOMException('Aborted', 'AbortError')
+  }
+}
+
+function systemPromptText(system: SystemPrompt | undefined): string {
+  if (system === undefined) return ''
+  if (typeof system === 'string') return system
+  if (Array.isArray(system)) return system.map((b) => b.text).join('\n')
+  return system.text
+}
+
+function toGeminiParts(content: string | ContentBlock[]): Part[] {
+  if (typeof content === 'string') return [{ text: content }]
+  const parts: Part[] = []
+  for (const block of content) {
+    if (block.type === 'text') {
+      parts.push({ text: block.text })
+    } else if (block.type === 'image' || block.type === 'document' || block.type === 'audio') {
+      // All three media block types share Gemini's inlineData /
+      // fileData wire shape; only the MIME differs. Base64 →
+      // inlineData. URL → fileData with fileUri. Gemini's
+      // fileData accepts public HTTPS and gs:// URIs; arbitrary
+      // private URLs need to be fetched and converted to base64
+      // by the app.
+      if (block.source.type === 'base64') {
+        parts.push({
+          inlineData: { mimeType: block.source.mediaType, data: block.source.data },
+        })
+      } else {
+        parts.push({
+          fileData: {
+            fileUri: block.source.url,
+            mimeType: guessMimeFromUrl(block.source.url, block.type),
+          },
+        })
+      }
+    } else if (block.type === 'tool_use') {
+      parts.push({
+        functionCall: {
+          id: block.id,
+          name: block.name,
+          args: (block.input ?? {}) as Record<string, unknown>,
+        },
+      })
+    } else if (block.type === 'tool_result') {
+      const text = typeof block.content === 'string'
+        ? block.content
+        : block.content.map((t) => t.text).join('')
+      parts.push({
+        functionResponse: {
+          id: block.toolUseId,
+          name: '',
+          response: block.isError ? { error: text } : { result: text },
+        },
+      })
+    }
+    // MCP blocks (Anthropic-only) silently dropped.
+  }
+  return parts
+}
+
+/**
+ * Gemini's `fileData.mimeType` is required, but our media-block
+ * URL-source variants don't carry it (the app may not know).
+ * Best-effort from the file extension. Default falls back to the
+ * block type's most-common MIME (jpeg for images, pdf for
+ * documents, mp3 for audio).
+ */
+/**
+ * Translate framework `ServerTool[]` into Gemini's typed entries
+ * (`googleSearch` / `codeExecution` / `urlContext`). Anthropic-
+ * specific tools (`web_fetch`) throw with clear guidance.
+ *
+ * Gemini's server tools have no per-tool config — they're enabled
+ * with empty `{}` objects. Domain allowlists / max_uses /
+ * blocked_domains on `web_search` are silently dropped (Gemini
+ * doesn't accept them).
+ */
+function geminiServerTools(
+  serverTools: readonly ServerTool[],
+): NonNullable<GenerateContentConfig['tools']> {
+  const out: NonNullable<GenerateContentConfig['tools']> = []
+  for (const t of serverTools) {
+    if (t.type === 'web_search') {
+      out.push({ googleSearch: {} })
+    } else if (t.type === 'code_execution') {
+      out.push({ codeExecution: {} })
+    } else if (t.type === 'url_context') {
+      out.push({ urlContext: {} })
+    } else if (t.type === 'web_fetch') {
+      throw new BrainError(
+        'GeminiBrainDriver: server tool `web_fetch` is Anthropic-only. Use `url_context` for Gemini or route the call to Anthropic.',
+        { context: { provider: 'google' } },
+      )
+    }
+  }
+  return out
+}
+
+function guessMimeFromUrl(
+  url: string,
+  kind: 'image' | 'document' | 'audio',
+): string {
+  const lower = url.toLowerCase().split('?')[0] ?? ''
+  if (kind === 'image') {
+    if (lower.endsWith('.png')) return 'image/png'
+    if (lower.endsWith('.webp')) return 'image/webp'
+    if (lower.endsWith('.gif')) return 'image/gif'
+    if (lower.endsWith('.jpg') || lower.endsWith('.jpeg')) return 'image/jpeg'
+    return 'image/jpeg'
+  }
+  if (kind === 'document') {
+    return 'application/pdf'
+  }
+  // audio
+  if (lower.endsWith('.mp3')) return 'audio/mp3'
+  if (lower.endsWith('.wav')) return 'audio/wav'
+  if (lower.endsWith('.ogg')) return 'audio/ogg'
+  if (lower.endsWith('.flac')) return 'audio/flac'
+  if (lower.endsWith('.webm')) return 'audio/webm'
+  if (lower.endsWith('.aac') || lower.endsWith('.m4a')) return 'audio/aac'
+  return 'audio/mp3'
+}
+
+function fromGeminiParts(parts: readonly Part[]): string | ContentBlock[] {
+  const blocks: ContentBlock[] = []
+  for (const part of parts) {
+    if (typeof part.text === 'string' && part.text.length > 0) {
+      blocks.push({ type: 'text', text: part.text })
+    } else if (part.functionCall) {
+      const fc = part.functionCall
+      blocks.push({
+        type: 'tool_use',
+        id: fc.id ?? `gemini_${cryptoRandomId()}`,
+        name: fc.name ?? '',
+        input: fc.args ?? {},
+      } satisfies ToolUseBlock)
+    }
+  }
+  if (blocks.length === 1 && blocks[0]?.type === 'text') return blocks[0].text
+  return blocks
+}
+
+function candidateText(candidate: { content?: { parts?: Part[] } } | undefined): string {
+  const parts = candidate?.content?.parts ?? []
+  return parts
+    .filter((p) => typeof p.text === 'string' && p.text.length > 0)
+    .map((p) => p.text as string)
+    .join('')
+}
+
+function buildThinkingConfig(options: ChatOptions): GenerateContentConfig['thinkingConfig'] {
+  if (options.effort !== undefined) {
+    const level = effortToThinkingLevel(options.effort)
+    return level !== undefined ? { thinkingLevel: level } : { thinkingBudget: -1 }
+  }
+  if (options.thinking === 'adaptive') return { thinkingBudget: -1 }
+  if (options.thinking === 'disabled') return { thinkingBudget: 0 }
+  return undefined
+}
+
+function effortToThinkingLevel(
+  effort: NonNullable<ChatOptions['effort']>,
+): ThinkingLevel | undefined {
+  switch (effort) {
+    case 'low': return ThinkingLevel.LOW
+    case 'medium': return ThinkingLevel.MEDIUM
+    case 'high':
+    case 'xhigh':
+    case 'max':
+      return ThinkingLevel.HIGH
+  }
+}
+
+function toUsage(u: { promptTokenCount?: number; candidatesTokenCount?: number; cachedContentTokenCount?: number } | undefined): ChatUsage {
+  return {
+    inputTokens: u?.promptTokenCount ?? 0,
+    outputTokens: u?.candidatesTokenCount ?? 0,
+    cacheReadTokens: u?.cachedContentTokenCount ?? 0,
+    cacheCreationTokens: 0,
+  }
+}
+
+function addUsage(
+  acc: ChatUsage,
+  u: { promptTokenCount?: number; candidatesTokenCount?: number; cachedContentTokenCount?: number } | undefined,
+): void {
+  if (!u) return
+  acc.inputTokens += u.promptTokenCount ?? 0
+  acc.outputTokens += u.candidatesTokenCount ?? 0
+  acc.cacheReadTokens += u.cachedContentTokenCount ?? 0
+}
+
+function cryptoRandomId(): string {
+  // Stable, low-entropy fallback for synthesizing tool-use ids when
+  // Gemini omits them. Uniqueness within a single response is all the
+  // loop requires — the id only travels back paired with its result
+  // and never escapes to the caller.
+  return Math.random().toString(36).slice(2, 12)
+}