@strav/brain 1.0.0-alpha.16 → 1.0.0-alpha.17

package/src/providers/openai_responses_provider.ts

@@ -0,0 +1,700 @@
+/**
+ * `OpenAIResponsesProvider` — implementation of `Provider` backed
+ * by the `openai` SDK's Responses API
+ * (`client.responses.create`).
+ *
+ * Use when an app needs:
+ *   - OpenAI's server-side tools: `web_search`, `code_interpreter`
+ *     (via the framework's `ChatOptions.serverTools` union).
+ *   - The Responses API's reasoning surfaces (gpt-5 / o-series).
+ *
+ * For everything else (plain chat, embeddings, transcription,
+ * function calling without server tools), the standard
+ * `OpenAIProvider` (driver `'openai'`) is simpler. Apps that
+ * use both register them as two separate providers and route
+ * per-call.
+ *
+ * Inherits `embed` + `transcribe` from `OpenAIProvider`
+ * (embeddings + Whisper live on different endpoints unchanged).
+ *
+ * V1 coverage:
+ *   - `chat` / `stream` via `responses.create` (with `stream: true`
+ *     for the streaming variant).
+ *   - `runWithTools` / `streamWithTools` — function-calling loop
+ *     against the Responses API. Local tools + MCP tools + server
+ *     tools all combine.
+ *   - `generate` / `runWithToolsAndSchema` /
+ *     `streamWithToolsAndSchema` — throw `BrainError` with
+ *     "structured output via Responses API is a follow-up slice"
+ *     guidance. Apps that need structured output use
+ *     `OpenAIProvider` (driver `'openai'`).
+ *
+ * The Responses API's message shape (`input_items`) is different
+ * from chat completions' `messages`, so this is a separate
+ * provider class rather than a strategy inside `OpenAIProvider`.
+ * Translation lives in this file.
+ */
+
+import OpenAI from 'openai'
+import type { AgentGenerateResult } from '../agent_generate_result.ts'
+import type { AgentResult } from '../agent_result.ts'
+import type { AgentStreamEvent } from '../agent_stream_event.ts'
+import { BrainError } from '../brain_error.ts'
+import type { OpenAIResponsesProviderConfig } from '../brain_config.ts'
+import { resolveMcpTools, type ResolveMcpToolsOptions } from '../mcp/resolve_mcp_tools.ts'
+import type { MCPServer } from '../mcp_server.ts'
+import type { OutputSchema } from '../output_schema.ts'
+import type { Provider, RunWithToolsOptions } from '../provider.ts'
+import type { Tool } from '../tool.ts'
+import { runToolWithRecovery } from '../tool_runner.ts'
+import type {
+  ChatOptions,
+  ChatResult,
+  ChatUsage,
+  ContentBlock,
+  GenerateResult,
+  Message,
+  ServerTool,
+  StreamEvent,
+  SystemPrompt,
+  TextBlock,
+  ToolResultBlock,
+  ToolUseBlock,
+} from '../types.ts'
+import { OpenAIProvider } from './openai_provider.ts'
+
+const DEFAULT_OPENAI_MODEL = 'gpt-5'
+const DEFAULT_OPENAI_MAX_TOKENS = 4096
+
+export interface OpenAIResponsesProviderOptions {
+  client?: OpenAI
+  /** Internal seam — tests inject a stub MCP client factory. */
+  mcpClientFactory?: ResolveMcpToolsOptions['clientFactory']
+}
+
+/** Translation: framework `ServerTool` → Responses API tool entry. */
+type ResponsesTool = Record<string, unknown>
+
+export class OpenAIResponsesProvider extends OpenAIProvider implements Provider {
+  constructor(
+    name: string,
+    config: OpenAIResponsesProviderConfig,
+    options: OpenAIResponsesProviderOptions = {},
+  ) {
+    // Reuse OpenAIProvider's constructor for the SDK client + the
+    // chat / embed / transcribe model defaults. Inheritance keeps
+    // `client`, `defaultEmbedModel`, `defaultTranscribeModel`
+    // working unchanged.
+    super(
+      name,
+      {
+        driver: 'openai',
+        apiKey: config.apiKey,
+        ...(config.baseUrl !== undefined ? { baseUrl: config.baseUrl } : {}),
+        ...(config.organization !== undefined ? { organization: config.organization } : {}),
+        defaultModel: config.defaultModel ?? DEFAULT_OPENAI_MODEL,
+        defaultMaxTokens: config.defaultMaxTokens ?? DEFAULT_OPENAI_MAX_TOKENS,
+        ...(config.defaultEmbedModel !== undefined
+          ? { defaultEmbedModel: config.defaultEmbedModel }
+          : {}),
+        ...(config.defaultTranscribeModel !== undefined
+          ? { defaultTranscribeModel: config.defaultTranscribeModel }
+          : {}),
+      },
+      options,
+    )
+  }
+
+  // ─── chat / stream ──────────────────────────────────────────────────────
+
+  override async chat(
+    messages: readonly Message[],
+    options: ChatOptions = {},
+  ): Promise<ChatResult> {
+    const params = this.buildResponsesParams(messages, options, [])
+    const response = await this.client.responses.create(
+      params,
+      reqOpts(options),
+    )
+    return this.toChatResultFromResponse(response, params.model as string)
+  }
+
+  override async *stream(
+    messages: readonly Message[],
+    options: ChatOptions = {},
+  ): AsyncIterable<StreamEvent> {
+    const params: OpenAI.Responses.ResponseCreateParamsStreaming = {
+      ...this.buildResponsesParams(messages, options, []),
+      stream: true,
+    }
+    const stream = await this.client.responses.create(params, reqOpts(options))
+    let finishReason: string | null = null
+    let usage: OpenAI.Responses.ResponseUsage | undefined
+    for await (const event of stream) {
+      // Text deltas — `output_text.delta` is the streaming chunk
+      // for the model's text output.
+      if (event.type === 'response.output_text.delta') {
+        const delta = (event as { delta: string }).delta
+        if (delta && delta.length > 0) yield { type: 'text', delta }
+      } else if (event.type === 'response.completed') {
+        const completed = (event as { response: OpenAI.Responses.Response }).response
+        usage = completed.usage
+        // Responses API doesn't have a finish_reason field directly;
+        // the response.status === 'completed' is the signal.
+        finishReason = completed.status ?? null
+      }
+    }
+    yield {
+      type: 'stop',
+      stopReason: finishReason,
+      usage: toUsage(usage),
+    }
+  }
+
+  // ─── runWithTools / streamWithTools ─────────────────────────────────────
+
+  override 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 () => {} }
+    try {
+      return await this._runResponsesLoop(messages, [...tools, ...resolved.tools], options)
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async _runResponsesLoop(
+    messages: readonly Message[],
+    tools: readonly Tool[],
+    options: RunWithToolsOptions,
+  ): Promise<AgentResult> {
+    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.buildResponsesParams(workingMessages, options, tools)
+      const response = await this.client.responses.create(params, reqOpts(options))
+      addUsage(aggregated, response.usage)
+
+      const assistantBlocks = fromResponsesOutput(response.output)
+      const toolCalls = response.output.filter(
+        (o): o is OpenAI.Responses.ResponseFunctionToolCall => o.type === 'function_call',
+      )
+      workingMessages.push({ role: 'assistant', content: assistantBlocks })
+
+      if (toolCalls.length === 0) {
+        const text = textFromOutput(response.output)
+        return {
+          text,
+          messages: workingMessages,
+          iterations,
+          stopReason: response.status ?? 'completed',
+          usage: aggregated,
+        }
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of toolCalls) {
+        let parsedInput: unknown = {}
+        let parseFailed: { content: string; isError: boolean } | undefined
+        try {
+          parsedInput = call.arguments ? JSON.parse(call.arguments) : {}
+        } catch (err) {
+          parseFailed = await tryRecoverParseError(
+            call.name,
+            call.call_id,
+            err as Error,
+            options,
+          )
+        }
+        const { content, isError } = parseFailed ?? await runToolWithRecovery(
+          toolMap.get(call.name),
+          call.name,
+          call.call_id,
+          parsedInput,
+          options,
+        )
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.call_id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+      }
+      workingMessages.push({ role: 'user', content: resultBlocks })
+
+      iterations++
+      if (iterations >= maxIterations) {
+        const text = textFromOutput(response.output)
+        return {
+          text,
+          messages: workingMessages,
+          iterations,
+          stopReason: 'max_iterations',
+          usage: aggregated,
+        }
+      }
+    }
+  }
+
+  override async *streamWithTools(
+    messages: readonly Message[],
+    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 () => {} }
+    try {
+      yield* this._streamResponsesLoop(messages, [...tools, ...resolved.tools], options)
+    } finally {
+      await resolved.close()
+    }
+  }
+
+  private async *_streamResponsesLoop(
+    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: OpenAI.Responses.ResponseCreateParamsStreaming = {
+        ...this.buildResponsesParams(workingMessages, options, tools),
+        stream: true,
+      }
+      const stream = await this.client.responses.create(params, reqOpts(options))
+      let finishReason: string | null = null
+      let finalResponse: OpenAI.Responses.Response | undefined
+
+      for await (const event of stream) {
+        if (event.type === 'response.output_text.delta') {
+          const delta = (event as { delta: string }).delta
+          if (delta && delta.length > 0) yield { type: 'text', delta }
+        } else if (event.type === 'response.completed') {
+          const completed = (event as { response: OpenAI.Responses.Response }).response
+          finalResponse = completed
+          finishReason = completed.status ?? null
+        }
+      }
+
+      yield { type: 'iteration_end', iteration: iterations, stopReason: finishReason }
+
+      if (!finalResponse) {
+        // The stream ended without a completion event — surface the
+        // best stop we have and bail.
+        yield {
+          type: 'stop',
+          stopReason: finishReason ?? 'incomplete',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+        }
+        return
+      }
+
+      addUsage(aggregated, finalResponse.usage)
+      const assistantBlocks = fromResponsesOutput(finalResponse.output)
+      workingMessages.push({ role: 'assistant', content: assistantBlocks })
+
+      const toolCalls = finalResponse.output.filter(
+        (o): o is OpenAI.Responses.ResponseFunctionToolCall => o.type === 'function_call',
+      )
+      if (toolCalls.length === 0) {
+        yield {
+          type: 'stop',
+          stopReason: finishReason ?? 'completed',
+          iterations,
+          usage: aggregated,
+          messages: workingMessages,
+        }
+        return
+      }
+
+      const resultBlocks: ContentBlock[] = []
+      for (const call of toolCalls) {
+        let parsedInput: unknown = {}
+        let parseFailed: { content: string; isError: boolean } | undefined
+        try {
+          parsedInput = call.arguments ? JSON.parse(call.arguments) : {}
+        } catch (err) {
+          parseFailed = await tryRecoverParseError(
+            call.name,
+            call.call_id,
+            err as Error,
+            options,
+          )
+        }
+        yield { type: 'tool_use', id: call.call_id, name: call.name, input: parsedInput }
+        const { content, isError } = parseFailed ?? await runToolWithRecovery(
+          toolMap.get(call.name),
+          call.name,
+          call.call_id,
+          parsedInput,
+          options,
+        )
+        resultBlocks.push({
+          type: 'tool_result',
+          toolUseId: call.call_id,
+          content,
+          ...(isError ? { isError: true } : {}),
+        } satisfies ToolResultBlock)
+        yield {
+          type: 'tool_result',
+          id: call.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
+      }
+    }
+  }
+
+  // ─── Schema variants throw — deferred ──────────────────────────────────
+
+  override async generate<T>(
+    _messages: readonly Message[],
+    _schema: OutputSchema<T>,
+    _options: ChatOptions = {},
+  ): Promise<GenerateResult<T>> {
+    throw new BrainError(
+      'OpenAIResponsesProvider.generate: structured output via the Responses API is a follow-up slice. For json-schema structured output today, route the call to the chat completions provider (driver: "openai").',
+      { context: { provider: this.name } },
+    )
+  }
+
+  override async runWithToolsAndSchema<T>(
+    _messages: readonly Message[],
+    _tools: readonly Tool[],
+    _schema: OutputSchema<T>,
+    _options?: RunWithToolsOptions,
+  ): Promise<AgentGenerateResult<T>> {
+    throw new BrainError(
+      'OpenAIResponsesProvider.runWithToolsAndSchema: combined tools + schema on the Responses API is a follow-up slice. Run runTools + generate as separate calls, or route to the chat completions provider for this combination.',
+      { context: { provider: this.name } },
+    )
+  }
+
+  override async *streamWithToolsAndSchema<T>(
+    _messages: readonly Message[],
+    _tools: readonly Tool[],
+    _schema: OutputSchema<T>,
+    _options?: RunWithToolsOptions,
+  ): AsyncIterable<AgentStreamEvent<T>> {
+    throw new BrainError(
+      'OpenAIResponsesProvider.streamWithToolsAndSchema: streaming + tools + schema on the Responses API is a follow-up slice. Use streamTools without schema, or route to the chat completions provider.',
+      { context: { provider: this.name } },
+    )
+  }
+
+  // ─── Param translation ──────────────────────────────────────────────────
+
+  private buildResponsesParams(
+    messages: readonly Message[],
+    options: ChatOptions,
+    tools: readonly Tool[],
+  ): OpenAI.Responses.ResponseCreateParamsNonStreaming {
+    const model = options.model ?? this.defaultModel
+    const params: OpenAI.Responses.ResponseCreateParamsNonStreaming = {
+      model,
+      input: messages.flatMap((m) => {
+        const r = toResponsesInputItem(m)
+        return Array.isArray(r) ? r : [r]
+      }) as unknown as OpenAI.Responses.ResponseInput,
+      max_output_tokens: options.maxTokens ?? this.defaultMaxTokens,
+    }
+    const systemText = systemPromptText(options.system)
+    if (systemText.length > 0) params.instructions = systemText
+
+    const toolEntries: ResponsesTool[] = []
+    for (const t of tools) {
+      toolEntries.push({
+        type: 'function',
+        name: t.name,
+        description: t.description,
+        parameters: t.inputSchema,
+        strict: false,
+      })
+    }
+    if (options.serverTools && options.serverTools.length > 0) {
+      toolEntries.push(...responsesServerTools(options.serverTools))
+    }
+    if (toolEntries.length > 0) {
+      params.tools = toolEntries as unknown as OpenAI.Responses.ResponseCreateParams['tools']
+    }
+
+    // Reasoning controls — gpt-5 and o-series only. Emit when set;
+    // non-reasoning models reject.
+    if (options.effort !== undefined) {
+      params.reasoning = { effort: options.effort } as OpenAI.Responses.ResponseCreateParams['reasoning']
+    } else if (options.thinking === 'adaptive') {
+      params.reasoning = { effort: 'medium' } as OpenAI.Responses.ResponseCreateParams['reasoning']
+    } else if (options.thinking === 'disabled') {
+      params.reasoning = { effort: 'minimal' } as OpenAI.Responses.ResponseCreateParams['reasoning']
+    }
+
+    return params
+  }
+
+  private toChatResultFromResponse(
+    response: OpenAI.Responses.Response,
+    requestedModel: string,
+  ): ChatResult<OpenAI.Responses.Response> {
+    return {
+      text: textFromOutput(response.output),
+      model: response.model ?? requestedModel,
+      stopReason: response.status ?? null,
+      usage: toUsage(response.usage),
+      raw: response,
+    }
+  }
+}
+
+// ─── Translation helpers ──────────────────────────────────────────────────
+
+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
+}
+
+/**
+ * Translate a framework `Message` into a Responses API input item.
+ * V1 covers text + tool_use + tool_result; other content blocks
+ * (image / document / audio) fall back to text concatenation until
+ * the Responses API multimodal slice ships.
+ */
+function toResponsesInputItem(message: Message): unknown {
+  if (typeof message.content === 'string') {
+    return { role: message.role, content: message.content }
+  }
+  // For user-role tool results, emit one `function_call_output` per
+  // tool_result block. The Responses API wants each result as its
+  // own input item, NOT bundled in a message turn.
+  if (message.role === 'user') {
+    const toolResults = message.content.filter((b): b is ToolResultBlock => b.type === 'tool_result')
+    if (toolResults.length > 0) {
+      // Multi-item return — caller handles arrays in input.
+      const items: unknown[] = []
+      const remainingText: string[] = []
+      for (const block of message.content) {
+        if (block.type === 'tool_result') {
+          const content = typeof block.content === 'string'
+            ? block.content
+            : block.content.map((t) => t.text).join('')
+          items.push({
+            type: 'function_call_output',
+            call_id: block.toolUseId,
+            output: content,
+          })
+        } else if (block.type === 'text') {
+          remainingText.push(block.text)
+        }
+      }
+      if (remainingText.length > 0) {
+        items.unshift({ role: 'user', content: remainingText.join('') })
+      }
+      return items
+    }
+    // Plain user message with mixed blocks → flatten text.
+    return {
+      role: 'user',
+      content: message.content
+        .filter((b): b is TextBlock => b.type === 'text')
+        .map((b) => b.text)
+        .join(''),
+    }
+  }
+  // Assistant turn with tool_use blocks → emit function_call items.
+  const items: unknown[] = []
+  const textParts: string[] = []
+  for (const block of message.content) {
+    if (block.type === 'text') {
+      textParts.push(block.text)
+    } else if (block.type === 'tool_use') {
+      items.push({
+        type: 'function_call',
+        call_id: block.id,
+        name: block.name,
+        arguments: JSON.stringify(block.input ?? {}),
+      })
+    }
+  }
+  if (textParts.length > 0) {
+    items.unshift({ role: 'assistant', content: textParts.join('') })
+  }
+  return items.length === 1 ? items[0] : items
+}
+
+/**
+ * Extract framework `ContentBlock[]` from a Responses API output
+ * array — text from `output_message.content[].text`, tool calls
+ * from `function_call` items. Server-tool calls (web_search,
+ * code_interpreter) are not surfaced as blocks; they live on
+ * `response.output` and apps inspect via `raw` for now.
+ */
+function fromResponsesOutput(
+  output: readonly OpenAI.Responses.ResponseOutputItem[],
+): string | ContentBlock[] {
+  const blocks: ContentBlock[] = []
+  for (const item of output) {
+    if (item.type === 'message' && item.role === 'assistant') {
+      for (const part of item.content) {
+        if (part.type === 'output_text') {
+          blocks.push({ type: 'text', text: part.text })
+        }
+      }
+    } else if (item.type === 'function_call') {
+      let parsed: unknown = {}
+      try {
+        parsed = item.arguments ? JSON.parse(item.arguments) : {}
+      } catch {
+        parsed = item.arguments ?? {}
+      }
+      blocks.push({
+        type: 'tool_use',
+        id: item.call_id,
+        name: item.name,
+        input: parsed,
+      } satisfies ToolUseBlock)
+    }
+    // Server-tool result items (web_search_call, code_interpreter_call,
+    // etc.) are surfaced on `raw` — V1 doesn't add framework blocks
+    // for them; apps inspect raw when they care.
+  }
+  if (blocks.length === 1 && blocks[0]?.type === 'text') return blocks[0].text
+  return blocks
+}
+
+function textFromOutput(output: readonly OpenAI.Responses.ResponseOutputItem[]): string {
+  const parts: string[] = []
+  for (const item of output) {
+    if (item.type === 'message' && item.role === 'assistant') {
+      for (const p of item.content) {
+        if (p.type === 'output_text') parts.push(p.text)
+      }
+    }
+  }
+  return parts.join('')
+}
+
+function responsesServerTools(serverTools: readonly ServerTool[]): ResponsesTool[] {
+  const out: ResponsesTool[] = []
+  for (const t of serverTools) {
+    if (t.type === 'web_search') {
+      out.push({ type: 'web_search' })
+    } else if (t.type === 'code_execution') {
+      out.push({ type: 'code_interpreter', container: { type: 'auto' } })
+    } else if (t.type === 'web_fetch') {
+      throw new BrainError(
+        'OpenAIResponsesProvider: server tool `web_fetch` is Anthropic-only. Use `web_search` for OpenAI, or route to Anthropic.',
+        { context: { provider: 'openai-responses' } },
+      )
+    } else if (t.type === 'url_context') {
+      throw new BrainError(
+        'OpenAIResponsesProvider: server tool `url_context` is Gemini-only. Route to Gemini, or include the URL in the prompt and use `web_search`.',
+        { context: { provider: 'openai-responses' } },
+      )
+    }
+  }
+  return out
+}
+
+function toUsage(u: OpenAI.Responses.ResponseUsage | undefined): ChatUsage {
+  return {
+    inputTokens: u?.input_tokens ?? 0,
+    outputTokens: u?.output_tokens ?? 0,
+    cacheReadTokens: u?.input_tokens_details?.cached_tokens ?? 0,
+    cacheCreationTokens: 0,
+  }
+}
+
+function addUsage(acc: ChatUsage, u: OpenAI.Responses.ResponseUsage | undefined): void {
+  if (!u) return
+  acc.inputTokens += u.input_tokens ?? 0
+  acc.outputTokens += u.output_tokens ?? 0
+  acc.cacheReadTokens += u.input_tokens_details?.cached_tokens ?? 0
+}
+
+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')
+  }
+}
+
+/**
+ * Handle a JSON.parse failure on a `function_call.arguments` field
+ * through the standard `onToolError` recovery hook. Returns a
+ * recovery result or rethrows as ToolExecutionError. Kept inline
+ * (not in tool_runner.ts) because the call shape — error-only,
+ * pre-execute — differs from the standard path.
+ */
+async function tryRecoverParseError(
+  toolName: string,
+  callId: string,
+  cause: Error,
+  options: RunWithToolsOptions,
+): Promise<{ content: string; isError: boolean }> {
+  const { ToolExecutionError } = await import('../tool_execution_error.ts')
+  const err = new ToolExecutionError(
+    toolName,
+    callId,
+    new Error(`Failed to parse tool input JSON: ${cause.message}`),
+  )
+  const recovered = options.onToolError?.(err)
+  if (typeof recovered !== 'string') throw err
+  return { content: recovered, isError: true }
+}