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

package/src/providers/gemini_provider.ts

@@ -73,7 +73,12 @@ import type {
 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 { Provider, RunWithToolsOptions } from '../provider.ts'
+import type {
+  Provider,
+  RunWithToolsOptions,
+  RunWithToolsOptionsWithSuspend,
+} from '../provider.ts'
+import type { SuspendedRun } from '../suspended_run.ts'
 import type { Tool } from '../tool.ts'
 import type {
   ChatOptions,
@@ -119,6 +124,8 @@ 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 GeminiProvider implements Provider {
@@ -128,6 +135,7 @@ export class GeminiProvider implements Provider {
   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
@@ -135,6 +143,7 @@ export class GeminiProvider implements Provider {
     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 {
@@ -273,18 +282,42 @@ export class GeminiProvider implements Provider {
     }
   }
 
+  /**
+   * 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> {
-    const mcpServers: readonly MCPServer[] = options.mcpServers ?? []
-    const resolved =
-      mcpServers.length > 0
-        ? await resolveMcpTools(mcpServers, {
-            ...(this.mcpClientFactory ? { clientFactory: this.mcpClientFactory } : {}),
-          })
-        : { tools: [] as Tool[], close: async () => {} }
+  ): Promise<AgentResult | SuspendedRun> {
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
     try {
       return await this._runLoop(messages, [...tools, ...resolved.tools], options)
     } finally {
@@ -296,7 +329,7 @@ export class GeminiProvider implements Provider {
     messages: readonly Message[],
     tools: readonly Tool[],
     options: RunWithToolsOptions,
-  ): Promise<AgentResult> {
+  ): Promise<AgentResult | SuspendedRun> {
     const maxIterations = options.maxIterations ?? 10
     const toolMap = new Map<string, Tool>(tools.map((t) => [t.name, t]))
     const workingMessages: Message[] = [...messages]
@@ -339,7 +372,15 @@ export class GeminiProvider implements Provider {
       }
 
       const resultBlocks: ContentBlock[] = []
-      for (const call of toolUses) {
+      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,
@@ -375,13 +416,7 @@ export class GeminiProvider implements Provider {
     schema: OutputSchema<T>,
     options: RunWithToolsOptions = {},
   ): Promise<AgentGenerateResult<T>> {
-    const mcpServers: readonly MCPServer[] = options.mcpServers ?? []
-    const resolved =
-      mcpServers.length > 0
-        ? await resolveMcpTools(mcpServers, {
-            ...(this.mcpClientFactory ? { clientFactory: this.mcpClientFactory } : {}),
-          })
-        : { tools: [] as Tool[], close: async () => {} }
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
     try {
       return await this._runLoopWithSchema([...tools, ...resolved.tools], messages, schema, options)
     } finally {
@@ -480,13 +515,7 @@ export class GeminiProvider implements Provider {
     tools: readonly Tool[],
     options: RunWithToolsOptions = {},
   ): AsyncIterable<AgentStreamEvent> {
-    const mcpServers: readonly MCPServer[] = options.mcpServers ?? []
-    const resolved =
-      mcpServers.length > 0
-        ? await resolveMcpTools(mcpServers, {
-            ...(this.mcpClientFactory ? { clientFactory: this.mcpClientFactory } : {}),
-          })
-        : { tools: [] as Tool[], close: async () => {} }
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
     try {
       yield* this._streamLoop(messages, [...tools, ...resolved.tools], options)
     } finally {
@@ -605,13 +634,7 @@ export class GeminiProvider implements Provider {
     schema: OutputSchema<T>,
     options: RunWithToolsOptions = {},
   ): AsyncIterable<AgentStreamEvent<T>> {
-    const mcpServers: readonly MCPServer[] = options.mcpServers ?? []
-    const resolved =
-      mcpServers.length > 0
-        ? await resolveMcpTools(mcpServers, {
-            ...(this.mcpClientFactory ? { clientFactory: this.mcpClientFactory } : {}),
-          })
-        : { tools: [] as Tool[], close: async () => {} }
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
     try {
       yield* this._streamLoopWithSchema(
         [...tools, ...resolved.tools],

package/src/providers/openai_compat_provider.ts

@@ -45,12 +45,17 @@ import { BrainError } from '../brain_error.ts'
 import { parseGenerated, type OutputSchema } from '../output_schema.ts'
 import type { RunWithToolsOptions } from '../provider.ts'
 import type { Tool } from '../tool.ts'
+import { recoverOrThrow, runToolWithRecovery } from '../tool_runner.ts'
+import { ToolExecutionError } from '../tool_execution_error.ts'
 import type {
   ChatOptions,
   ChatUsage,
+  ContentBlock,
   GenerateResult,
   Message,
   SystemPrompt,
+  ToolResultBlock,
+  ToolUseBlock,
 } from '../types.ts'
 import { OpenAIProvider } from './openai_provider.ts'
 
@@ -116,36 +121,351 @@ export abstract class OpenAICompatProvider extends OpenAIProvider {
   }
 
   /**
-   * Combined tool-loop + structured output isn't supported on
-   * OpenAI-compat providers in V1. The API's `json_object` mode
-   * doesn't carry schema enforcement, and weaving the
-   * schema-instruction into every turn's system prompt during a
-   * tool loop would surprise apps. Apps run `runTools(...)` +
-   * `generate(...)` as two separate calls, or switch to OpenAI /
-   * Anthropic / Gemini for the combined call.
+   * Combined tool-loop + structured output via the **tool-forcing**
+   * pattern. OpenAI-compat endpoints don't support per-turn
+   * `json_schema` enforcement, but they do support OpenAI-style
+   * function calling — so the framework injects a synthetic
+   * `respond_with_<schemaName>` tool whose JSON-Schema
+   * `parameters` IS the desired output schema. The model uses it
+   * (and only it) for its final answer; the args become the
+   * parsed structured value. Regular tools work normally
+   * alongside.
+   *
+   * The model is prompted to call regular tools first, then
+   * `respond_with` exactly once when ready to answer. If it
+   * doesn't (returns plain text instead, or hits `maxIterations`),
+   * the framework throws `BrainError` — apps should reinforce the
+   * pattern via a clearer system prompt, or simplify the task.
+   *
+   * Caveats vs OpenAI's `strict: true`:
+   *   - Smaller models may emit invalid JSON in the tool args.
+   *     `parseGenerated` + the optional `schema.parse` hook catch
+   *     it at the boundary.
+   *   - Schema features beyond OpenAI function-calling's subset
+   *     (recursive refs, advanced keywords) may not be honored.
+   *     Stick to flat object schemas for best results.
    */
   override async runWithToolsAndSchema<T>(
-    _messages: readonly Message[],
-    _tools: readonly Tool[],
-    _schema: OutputSchema<T>,
-    _options?: RunWithToolsOptions,
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions = {},
   ): Promise<AgentGenerateResult<T>> {
-    throw new BrainError(
-      `${this.name}.runWithToolsAndSchema: combined tool use + structured output is not supported on OpenAI-compat providers in V1. Run \`brain.runTools(...)\` and \`brain.generate(...)\` as two separate calls, or switch to OpenAI / Anthropic / Gemini for this combination.`,
-      { context: { provider: this.name } },
-    )
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
+    try {
+      return await this._toolForcingLoop(
+        messages,
+        [...tools, ...resolved.tools],
+        schema,
+        options,
+      )
+    } finally {
+      await resolved.close()
+    }
   }
 
   override async *streamWithToolsAndSchema<T>(
-    _messages: readonly Message[],
-    _tools: readonly Tool[],
-    _schema: OutputSchema<T>,
-    _options?: RunWithToolsOptions,
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions = {},
   ): AsyncIterable<AgentStreamEvent<T>> {
-    throw new BrainError(
-      `${this.name}.streamWithToolsAndSchema: combined streaming + tool use + structured output is not supported on OpenAI-compat providers in V1. Use \`brain.streamTools(...)\` and \`brain.generate(...)\` separately, or switch to OpenAI / Anthropic / Gemini for this combination.`,
-      { context: { provider: this.name } },
-    )
+    const resolved = await this.resolveMcp(options.mcpServers ?? [])
+    try {
+      yield* this._toolForcingStream(
+        messages,
+        [...tools, ...resolved.tools],
+        schema,
+        options,
+      )
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async _toolForcingLoop<T>(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions,
+  ): Promise<AgentGenerateResult<T>> {
+    const { respondTool, respondName, augmented } = prepareToolForcing(schema, options, tools)
+    const maxIterations = options.maxIterations ?? 10
+    const toolMap = new Map<string, Tool>(tools.map((t) => [t.name, t]))
+    const workingMessages: Message[] = [...messages]
+    const aggregated: ChatUsage = {
+      inputTokens: 0,
+      outputTokens: 0,
+      cacheReadTokens: 0,
+      cacheCreationTokens: 0,
+    }
+    let iterations = 0
+
+    while (true) {
+      checkAborted(options.signal)
+      const params = this.buildParams(workingMessages, augmented, tools)
+      params.tools = [...(params.tools ?? []), respondTool]
+      const response = await this.client.chat.completions.create(
+        params,
+        reqOpts(options),
+      )
+      addUsageHere(aggregated, response.usage, this)
+
+      const choice = response.choices[0]
+      if (!choice) {
+        throw new BrainError(
+          `${this.name}.runWithToolsAndSchema: response had no choices.`,
+        )
+      }
+      const assistantMessage = choice.message
+      workingMessages.push({
+        role: 'assistant',
+        content: fromOpenAIAssistant(assistantMessage),
+      })
+
+      const toolCalls = assistantMessage.tool_calls ?? []
+      const respond = toolCalls.find(
+        (c) => c.type === 'function' && c.function.name === respondName,
+      )
+      if (respond && respond.type === 'function') {
+        const text = respond.function.arguments ?? ''
+        const value = parseGenerated(text, schema)
+        return {
+          value,
+          text,
+          messages: workingMessages,
+          iterations,
+          stopReason: choice.finish_reason ?? 'stop',
+          usage: aggregated,
+        }
+      }
+
+      if (toolCalls.length === 0 || choice.finish_reason !== 'tool_calls') {
+        throw new BrainError(
+          `${this.name}.runWithToolsAndSchema: model returned without calling \`${respondName}\`. Add a stronger instruction in the system prompt — apps must steer the model to use the synthetic respond tool for its final answer.`,
+          { context: { provider: this.name, text: assistantMessage.content ?? '' } },
+        )
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of toolCalls) {
+        if (call.type !== 'function') continue
+        let parsedInput: unknown
+        let parseFailed: { content: string; isError: boolean } | undefined
+        try {
+          parsedInput = call.function.arguments ? JSON.parse(call.function.arguments) : {}
+        } catch (err) {
+          parseFailed = recoverOrThrow(
+            new ToolExecutionError(
+              call.function.name,
+              call.id,
+              new Error(`Failed to parse tool input JSON: ${(err as Error).message}`),
+            ),
+            options,
+          )
+        }
+        const { content, isError } = parseFailed
+          ?? (await runToolWithRecovery(
+            toolMap.get(call.function.name),
+            call.function.name,
+            call.id,
+            parsedInput,
+            options,
+          ))
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        throw new BrainError(
+          `${this.name}.runWithToolsAndSchema: hit maxIterations (${maxIterations}) without the model calling \`${respondName}\`. Bump maxIterations, simplify the task, or strengthen the system-prompt nudge.`,
+          { context: { provider: this.name } },
+        )
+      }
+    }
+  }
+
+  private async *_toolForcingStream<T>(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    schema: OutputSchema<T>,
+    options: RunWithToolsOptions,
+  ): AsyncIterable<AgentStreamEvent<T>> {
+    const { respondTool, respondName, augmented } = prepareToolForcing(schema, options, tools)
+    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 baseParams = this.buildParams(workingMessages, augmented, tools)
+      baseParams.tools = [...(baseParams.tools ?? []), respondTool]
+      const params: OpenAI.Chat.ChatCompletionCreateParamsStreaming = {
+        ...baseParams,
+        stream: true,
+        stream_options: { include_usage: true },
+      }
+      const stream = await this.client.chat.completions.create(params, reqOpts(options))
+
+      let textBuf = ''
+      const toolCallsByIndex = new Map<
+        number,
+        { id?: string; name?: string; args: string; started: boolean }
+      >()
+      let finishReason: string | null = null
+      let lastUsage: OpenAI.CompletionUsage | undefined
+
+      for await (const chunk of stream) {
+        const choice = chunk.choices[0]
+        const delta = choice?.delta
+        if (delta?.content && typeof delta.content === 'string' && delta.content.length > 0) {
+          textBuf += delta.content
+          yield { type: 'text', delta: delta.content }
+        }
+        if (delta?.tool_calls) {
+          for (const tc of delta.tool_calls) {
+            const entry = toolCallsByIndex.get(tc.index) ?? { args: '', started: false }
+            if (tc.id) entry.id = tc.id
+            if (tc.function?.name) entry.name = tc.function.name
+            toolCallsByIndex.set(tc.index, entry)
+            if (!entry.started && entry.id !== undefined && entry.name !== undefined) {
+              entry.started = true
+              if (entry.name !== respondName) {
+                yield { type: 'tool_use_start', id: entry.id, name: entry.name }
+              }
+            }
+            if (tc.function?.arguments) {
+              entry.args += tc.function.arguments
+              if (
+                entry.started &&
+                entry.id !== undefined &&
+                entry.name !== respondName
+              ) {
+                yield {
+                  type: 'tool_use_delta',
+                  id: entry.id,
+                  argsDelta: tc.function.arguments,
+                }
+              }
+            }
+          }
+        }
+        if (choice?.finish_reason) finishReason = choice.finish_reason
+        if (chunk.usage) lastUsage = chunk.usage
+      }
+
+      addUsageHere(aggregated, lastUsage, this)
+      yield { type: 'iteration_end', iteration: iterations, stopReason: finishReason }
+
+      const assistantBlocks: ContentBlock[] = []
+      if (textBuf.length > 0) assistantBlocks.push({ type: 'text', text: textBuf })
+      const orderedCalls = [...toolCallsByIndex.entries()]
+        .sort(([a], [b]) => a - b)
+        .map(([, v]) => v)
+      for (const call of orderedCalls) {
+        if (!call.id || !call.name) continue
+        let parsedInput: unknown = {}
+        try {
+          parsedInput = call.args ? JSON.parse(call.args) : {}
+        } catch {
+          parsedInput = call.args
+        }
+        assistantBlocks.push({
+          type: 'tool_use',
+          id: call.id,
+          name: call.name,
+          input: parsedInput,
+        } satisfies ToolUseBlock)
+      }
+      const assistantContent: string | ContentBlock[] =
+        assistantBlocks.length === 1 && assistantBlocks[0]?.type === 'text'
+          ? assistantBlocks[0].text
+          : assistantBlocks
+      workingMessages.push({ role: 'assistant', content: assistantContent })
+
+      const respond = orderedCalls.find((c) => c.name === respondName)
+      if (respond && respond.id) {
+        const text = respond.args
+        const value = parseGenerated(text, schema)
+        yield {
+          type: 'stop',
+          stopReason: finishReason ?? 'stop',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+          value,
+          text,
+        } as AgentStreamEvent<T>
+        return
+      }
+
+      if (finishReason !== 'tool_calls' || orderedCalls.length === 0) {
+        throw new BrainError(
+          `${this.name}.streamWithToolsAndSchema: model returned without calling \`${respondName}\`. Strengthen the system-prompt nudge.`,
+          { context: { provider: this.name, text: textBuf } },
+        )
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of orderedCalls) {
+        if (!call.id || !call.name) continue
+        let parsedInput: unknown
+        let parseFailed: { content: string; isError: boolean } | undefined
+        try {
+          parsedInput = call.args ? JSON.parse(call.args) : {}
+        } catch (err) {
+          parseFailed = recoverOrThrow(
+            new ToolExecutionError(
+              call.name,
+              call.id,
+              new Error(`Failed to parse tool input JSON: ${(err as Error).message}`),
+            ),
+            options,
+          )
+          parsedInput = call.args
+        }
+        yield { type: 'tool_use', id: call.id, name: call.name, input: parsedInput }
+        const { content, isError } = parseFailed
+          ?? (await runToolWithRecovery(
+            toolMap.get(call.name),
+            call.name,
+            call.id,
+            parsedInput,
+            options,
+          ))
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+        yield { type: 'tool_result', id: call.id, name: call.name, content, isError }
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        throw new BrainError(
+          `${this.name}.streamWithToolsAndSchema: hit maxIterations (${maxIterations}) without the model calling \`${respondName}\`.`,
+          { context: { provider: this.name } },
+        )
+      }
+    }
   }
 
   /**
@@ -185,3 +505,112 @@ function schemaInstruction(schema: OutputSchema<unknown>): string {
   ].filter((s): s is string => s !== undefined)
   return lines.join('\n')
 }
+
+// ─── Tool-forcing helpers ────────────────────────────────────────────────
+
+const RESPOND_TOOL_PREFIX = 'respond_with_'
+
+/**
+ * Build the synthetic respond-tool entry + the system-prompt nudge
+ * apps inject alongside their own system message. Validates that
+ * the chosen tool name doesn't collide with any user tool — that
+ * would make the loop's terminal detection ambiguous.
+ */
+function prepareToolForcing(
+  schema: OutputSchema<unknown>,
+  options: ChatOptions,
+  userTools: readonly Tool[],
+): {
+  respondTool: OpenAI.Chat.ChatCompletionTool
+  respondName: string
+  augmented: ChatOptions
+} {
+  const respondName = `${RESPOND_TOOL_PREFIX}${schema.name}`
+  if (userTools.some((t) => t.name === respondName)) {
+    throw new BrainError(
+      `OpenAICompatProvider.runWithToolsAndSchema: synthetic tool name "${respondName}" collides with a user-supplied tool. Rename your tool or the OutputSchema.name to avoid the clash.`,
+      { context: { conflictingName: respondName } },
+    )
+  }
+  const respondTool: OpenAI.Chat.ChatCompletionTool = {
+    type: 'function',
+    function: {
+      name: respondName,
+      description:
+        `Submit your final answer. Call this exactly once, after using any other tools you need. ` +
+        `The arguments MUST conform to the schema below. Do not return prose alongside or in place of this call.` +
+        (schema.description ? ` (${schema.description})` : ''),
+      parameters: schema.jsonSchema as Record<string, unknown>,
+    },
+  }
+  const augmented: ChatOptions = {
+    ...options,
+    system: combineSystem(options.system, toolForcingInstruction(respondName)),
+  }
+  return { respondTool, respondName, augmented }
+}
+
+function toolForcingInstruction(respondName: string): string {
+  return [
+    `When you are ready to give the final answer, call the \`${respondName}\` function with the structured arguments.`,
+    `Use any other available tools first to gather what you need. Once you have enough information, call \`${respondName}\` exactly once and do NOT also return prose text.`,
+  ].join(' ')
+}
+
+function reqOpts(options: { signal?: AbortSignal }): { signal?: AbortSignal } | undefined {
+  return options.signal !== undefined ? { signal: options.signal } : undefined
+}
+
+function checkAborted(signal: AbortSignal | undefined): void {
+  if (signal?.aborted) {
+    throw signal.reason ?? new DOMException('Aborted', 'AbortError')
+  }
+}
+
+function fromOpenAIAssistant(
+  msg: OpenAI.Chat.ChatCompletionMessage,
+): string | ContentBlock[] {
+  const blocks: ContentBlock[] = []
+  if (msg.content) blocks.push({ type: 'text', text: msg.content })
+  if (msg.tool_calls) {
+    for (const call of msg.tool_calls) {
+      if (call.type !== 'function') continue
+      let parsedInput: unknown = {}
+      try {
+        parsedInput = call.function.arguments ? JSON.parse(call.function.arguments) : {}
+      } catch {
+        parsedInput = call.function.arguments ?? {}
+      }
+      blocks.push({
+        type: 'tool_use',
+        id: call.id,
+        name: call.function.name,
+        input: parsedInput,
+      } satisfies ToolUseBlock)
+    }
+  }
+  if (blocks.length === 1 && blocks[0]?.type === 'text') return blocks[0].text
+  return blocks
+}
+
+/**
+ * Add provider-mapped usage onto an accumulator. Calls `mapUsage`
+ * on the provider instance so subclasses (e.g., DeepSeek) honor
+ * their vendor-specific cache fields.
+ */
+function addUsageHere(
+  acc: ChatUsage,
+  u: OpenAI.CompletionUsage | undefined,
+  provider: OpenAICompatProvider,
+): void {
+  if (!u) return
+  // Cast: `mapUsage` is protected on the abstract class; we're
+  // inside the module so the access is valid at runtime.
+  const mapped = (provider as unknown as {
+    mapUsage(u: OpenAI.CompletionUsage | undefined): ChatUsage
+  }).mapUsage(u)
+  acc.inputTokens += mapped.inputTokens
+  acc.outputTokens += mapped.outputTokens
+  acc.cacheReadTokens += mapped.cacheReadTokens
+  acc.cacheCreationTokens += mapped.cacheCreationTokens
+}