@ai-sdk/harness 0.0.0-6b196531-20260710185421

package/src/agent/internal/run-prompt.ts

@@ -0,0 +1,901 @@
+import {
+  getHarnessV1BuiltinToolFilteringDenialReason,
+  isHarnessV1BuiltinToolIncluded,
+  type HarnessV1,
+  type HarnessV1BuiltinToolFiltering,
+  type HarnessV1PendingToolApproval,
+  type HarnessV1Prompt,
+  type HarnessV1PromptControl,
+  type HarnessV1Session,
+  type HarnessV1StreamPart,
+  type HarnessV1ToolSpec,
+} from '../../v1';
+import { toHarnessStream } from './to-harness-stream';
+import {
+  executeTool,
+  generateId,
+  isExecutableTool,
+  safeParseJSON,
+  type Context,
+  type Experimental_SandboxSession as SandboxSession,
+  type ToolSet,
+} from '@ai-sdk/provider-utils';
+import type {
+  LanguageModelV4FinishReason,
+  LanguageModelV4ToolCall,
+  LanguageModelV4Usage,
+} from '@ai-sdk/provider';
+import { parseToolCall } from 'ai/internal';
+import type { ContentPart, TelemetryOptions, TextStreamPart } from 'ai';
+import type { HarnessAgentToolApprovalContinuation } from '../harness-agent-tool-approval-continuation';
+import type { HarnessAgentToolApprovalConfiguration } from '../harness-agent-settings';
+import { HarnessStreamTextResult } from './harness-stream-text-result';
+import { translateStreamPart } from './translate-stream-part';
+import { stripWorkDir } from './strip-work-dir';
+import { createTurnTelemetry, type TurnContentPart } from './turn-telemetry';
+import { resolveCustomToolApproval } from './permission-mode';
+import { logBridgeError } from '../../utils/bridge-diagnostics';
+
+/**
+ * Drive one prompt turn end-to-end:
+ *  - call `session.doPromptTurn` via `toHarnessStream`
+ *  - translate harness events to AI SDK `TextStreamPart`s and push into the
+ *    result object
+ *  - execute host-side user tools when their `tool-call` events arrive and
+ *    submit results back to the harness
+ *  - close the result when the harness signals `finish` (or on error)
+ *
+ * Returns the result synchronously after the stream is wired up; callers
+ * await its `PromiseLike` accessors to observe completion.
+ */
+export function runPrompt<
+  TOOLS extends ToolSet,
+  RUNTIME_CONTEXT extends Context,
+>(input: {
+  harness: HarnessV1;
+  session: HarnessV1Session;
+  /**
+   * Turn entry point. `'prompt'` (default) starts a new turn from `prompt`;
+   * `'continue'` continues the in-flight turn via `doContinueTurn` and ignores
+   * `prompt`/`instructions`.
+   */
+  mode?: 'prompt' | 'continue';
+  /** Required for `mode: 'prompt'`; absent for `mode: 'continue'`. */
+  prompt?: HarnessV1Prompt;
+  instructions: string | undefined;
+  tools: TOOLS;
+  activeTools?: ToolSet;
+  toolSpecs: HarnessV1ToolSpec[];
+  builtinToolFiltering?: HarnessV1BuiltinToolFiltering | undefined;
+  sandboxSession: SandboxSession;
+  sessionWorkDir: string;
+  runtimeContext: RUNTIME_CONTEXT;
+  abortSignal: AbortSignal | undefined;
+  telemetry?: TelemetryOptions | undefined;
+  toolApproval?: HarnessAgentToolApprovalConfiguration | undefined;
+  pendingToolApprovals?: readonly HarnessV1PendingToolApproval[];
+  toolApprovalContinuations?:
+    | readonly HarnessAgentToolApprovalContinuation[]
+    | undefined;
+  onPendingToolApproval?: (approval: HarnessV1PendingToolApproval) => void;
+  onToolApprovalSettled?: (approvalId: string) => void;
+  onTurnFinished?: () => void;
+}): {
+  result: HarnessStreamTextResult<TOOLS, RUNTIME_CONTEXT>;
+  done: Promise<void>;
+} {
+  const result = new HarnessStreamTextResult<TOOLS, RUNTIME_CONTEXT>({
+    tools: input.tools,
+    runtimeContext: input.runtimeContext,
+    // toolsContext is not configurable for harnesses; pass undefined cast.
+    toolsContext: undefined as never,
+    harnessId: input.harness.harnessId,
+    sessionId: input.session.sessionId,
+  });
+  const pendingToolApprovals = input.pendingToolApprovals ?? [];
+  const onPendingToolApproval = input.onPendingToolApproval ?? (() => {});
+  const onToolApprovalSettled = input.onToolApprovalSettled ?? (() => {});
+  const activeTools = input.activeTools ?? input.tools;
+
+  const telemetry = createTurnTelemetry({
+    telemetry: input.telemetry,
+    harnessId: input.harness.harnessId,
+    modelId: input.session.modelId,
+    instructions: input.instructions,
+    promptText: input.prompt != null ? promptToText(input.prompt) : '',
+    runtimeContext: input.runtimeContext,
+  });
+
+  const done = (async () => {
+    let bridge: Awaited<ReturnType<typeof toHarnessStream>>;
+    try {
+      bridge = await toHarnessStream({
+        invoke:
+          input.mode === 'continue'
+            ? emit =>
+                input.session.doContinueTurn({
+                  tools: input.toolSpecs,
+                  abortSignal: input.abortSignal,
+                  emit,
+                })
+            : emit => {
+                if (input.prompt == null) {
+                  throw new Error(
+                    'runPrompt: `prompt` is required for mode "prompt".',
+                  );
+                }
+                return input.session.doPromptTurn({
+                  prompt: input.prompt,
+                  tools: input.toolSpecs,
+                  instructions: input.instructions,
+                  abortSignal: input.abortSignal,
+                  emit,
+                });
+              },
+      });
+    } catch (err) {
+      telemetry.error(err);
+      logBridgeError({
+        harnessId: input.harness.harnessId,
+        sessionId: input.session.sessionId,
+        context: 'failed to start harness turn',
+        error: err,
+      });
+      result.fail(err);
+      return;
+    }
+
+    const { stream, control } = bridge;
+    const reader = stream.getReader();
+    const toolCallsByToolCallId = new Map<string, ToolCallTextStreamPart>();
+    const rawToolCallsByToolCallId = new Map<
+      string,
+      Extract<HarnessV1StreamPart, { type: 'tool-call' }>
+    >();
+    const pendingApprovalsByApprovalId = new Map(
+      pendingToolApprovals.map(approval => [approval.approvalId, approval]),
+    );
+    const pendingApprovalsByToolCallId = new Map(
+      pendingToolApprovals.map(approval => [approval.toolCallId, approval]),
+    );
+    const continuationsByApprovalId = new Map(
+      (input.toolApprovalContinuations ?? []).map(continuation => [
+        continuation.approvalResponse.approvalId,
+        continuation,
+      ]),
+    );
+    const settledApprovalToolCallIds = new Set<string>();
+    let finalFinish:
+      | Extract<HarnessV1StreamPart, { type: 'finish' }>
+      | undefined;
+
+    // Accumulate the model's output content per step so telemetry can record
+    // `gen_ai.output.messages` and reporters can log what was actually said.
+    let stepText = '';
+    let stepReasoning = '';
+    let stepToolCalls: TurnContentPart[] = [];
+    const buildStepContent = (): TurnContentPart[] => {
+      const parts: TurnContentPart[] = [];
+      if (stepText) parts.push({ type: 'text', text: stepText });
+      if (stepReasoning) parts.push({ type: 'reasoning', text: stepReasoning });
+      parts.push(...stepToolCalls);
+      return parts;
+    };
+    const resetStepContent = (): void => {
+      stepText = '';
+      stepReasoning = '';
+      stepToolCalls = [];
+    };
+    const zeroUsage: LanguageModelV4Usage = {
+      inputTokens: {
+        total: undefined,
+        noCache: undefined,
+        cacheRead: undefined,
+        cacheWrite: undefined,
+      },
+      outputTokens: {
+        total: undefined,
+        text: undefined,
+        reasoning: undefined,
+      },
+    };
+    const toolCallsFinishReason: LanguageModelV4FinishReason = {
+      unified: 'tool-calls',
+      raw: undefined,
+    };
+    const finishForToolApprovalPause = async (): Promise<void> => {
+      telemetry.stepFinish({
+        finishReason: toolCallsFinishReason,
+        usage: zeroUsage,
+        content: buildStepContent(),
+      });
+      resetStepContent();
+      result.finishStep({
+        finishReason: toolCallsFinishReason,
+        usage: zeroUsage,
+        providerMetadata: undefined,
+        warnings: [],
+      });
+      telemetry.end({
+        finishReason: toolCallsFinishReason,
+        usage: zeroUsage,
+      });
+      await result.finish();
+    };
+    const enqueueApprovalRequest = (approval: {
+      approvalId: string;
+      toolCall: ToolCallTextStreamPart;
+      isAutomatic?: boolean;
+    }): void => {
+      result.enqueue({
+        type: 'tool-approval-request',
+        approvalId: approval.approvalId,
+        toolCall: approval.toolCall,
+        ...(approval.isAutomatic !== undefined
+          ? { isAutomatic: approval.isAutomatic }
+          : {}),
+      } as TextStreamPart<TOOLS>);
+    };
+    const enqueueAutomaticApprovalResponse = (input: {
+      approvalId: string;
+      toolCall: ToolCallTextStreamPart;
+      approved: boolean;
+      reason?: string;
+      providerExecuted?: boolean;
+    }): void => {
+      result.enqueue({
+        type: 'tool-approval-response',
+        approvalId: input.approvalId,
+        toolCall: input.toolCall,
+        approved: input.approved,
+        ...(input.reason !== undefined ? { reason: input.reason } : {}),
+        ...(input.providerExecuted !== undefined
+          ? { providerExecuted: input.providerExecuted }
+          : {}),
+      } as TextStreamPart<TOOLS>);
+    };
+    const enqueueApprovalResponse = (
+      approval: HarnessV1PendingToolApproval,
+      continuation: HarnessAgentToolApprovalContinuation,
+    ): void => {
+      result.enqueue({
+        type: 'tool-approval-response',
+        approvalId: approval.approvalId,
+        toolCall: continuation.toolCall,
+        approved: continuation.approvalResponse.approved,
+        ...(continuation.approvalResponse.reason !== undefined
+          ? { reason: continuation.approvalResponse.reason }
+          : {}),
+        ...(approval.providerExecuted !== undefined
+          ? { providerExecuted: approval.providerExecuted }
+          : {}),
+      } as TextStreamPart<TOOLS>);
+    };
+    const processPendingApprovalContinuation = async (
+      approval: HarnessV1PendingToolApproval,
+      continuation: HarnessAgentToolApprovalContinuation,
+    ): Promise<void> => {
+      enqueueApprovalResponse(approval, continuation);
+      onToolApprovalSettled(approval.approvalId);
+      pendingApprovalsByApprovalId.delete(approval.approvalId);
+      pendingApprovalsByToolCallId.delete(approval.toolCallId);
+      settledApprovalToolCallIds.add(approval.toolCallId);
+
+      if (approval.kind === 'builtin') {
+        if (control.submitToolApproval == null) {
+          throw new Error(
+            `Harness '${input.harness.harnessId}' emitted a built-in tool approval request but does not support approval responses.`,
+          );
+        }
+        await control.submitToolApproval({
+          approvalId: approval.approvalId,
+          approved: continuation.approvalResponse.approved,
+          reason: continuation.approvalResponse.reason,
+        });
+        return;
+      }
+
+      if (!continuation.approvalResponse.approved) {
+        await control.submitToolResult({
+          toolCallId: approval.toolCallId,
+          output: {
+            type: 'execution-denied',
+            reason: continuation.approvalResponse.reason,
+          },
+        });
+        return;
+      }
+
+      const rawToolCall =
+        rawToolCallsByToolCallId.get(approval.toolCallId) ??
+        ({
+          type: 'tool-call',
+          toolCallId: approval.toolCallId,
+          toolName: approval.toolName,
+          input: approval.input,
+        } satisfies Extract<HarnessV1StreamPart, { type: 'tool-call' }>);
+
+      const outcome = await maybeExecuteHostTool({
+        event: rawToolCall,
+        tools: activeTools,
+        sandboxSession: input.sandboxSession,
+        abortSignal: input.abortSignal,
+        control,
+        onPreliminaryResult: preliminaryOutput => {
+          const stripped = stripWorkDir(
+            {
+              type: 'tool-result',
+              toolCallId: rawToolCall.toolCallId,
+              toolName: rawToolCall.toolName,
+              result: preliminaryOutput as Extract<
+                HarnessV1StreamPart,
+                { type: 'tool-result' }
+              >['result'],
+            },
+            input.sessionWorkDir,
+          ) as Extract<HarnessV1StreamPart, { type: 'tool-result' }>;
+          result.enqueue({
+            type: 'tool-result',
+            toolCallId: rawToolCall.toolCallId,
+            toolName: rawToolCall.toolName,
+            input: undefined,
+            output: stripped.result,
+            preliminary: true,
+          } as TextStreamPart<TOOLS>);
+        },
+      });
+      telemetry.toolEnd(rawToolCall.toolCallId, outcome);
+    };
+
+    try {
+      for (const approval of pendingToolApprovals) {
+        const continuation = continuationsByApprovalId.get(approval.approvalId);
+        if (continuation != null) {
+          await processPendingApprovalContinuation(approval, continuation);
+        }
+      }
+
+      while (true) {
+        const { value, done } = await reader.read();
+        if (done) break;
+        if (value == null) continue;
+
+        // Begin the operation span on stream-start, using the runtime-resolved
+        // model the adapter reports (falling back to the session's model).
+        if (value.type === 'stream-start') {
+          telemetry.start(value.modelId ?? input.session.modelId);
+        }
+
+        // Open a step span lazily before the first content of each step.
+        if (
+          value.type !== 'stream-start' &&
+          value.type !== 'finish-step' &&
+          value.type !== 'finish' &&
+          value.type !== 'error'
+        ) {
+          telemetry.ensureStepOpen();
+        }
+
+        /*
+         * Strip the session working-directory prefix for everything the
+         * consumer sees. The original `value` is kept intact for host tool
+         * execution below — the tools need the absolute path to resolve
+         * against the sandbox root, so the strip is display-only.
+         */
+        const displayValue = stripWorkDir(value, input.sessionWorkDir);
+        const settledApprovalToolCallReplay =
+          displayValue.type === 'tool-call' &&
+          !displayValue.providerExecuted &&
+          settledApprovalToolCallIds.has(displayValue.toolCallId);
+
+        if (settledApprovalToolCallReplay) {
+          continue;
+        }
+
+        if (displayValue.type === 'tool-approval-request') {
+          const toolCall = toolCallsByToolCallId.get(displayValue.toolCallId);
+          if (toolCall == null) {
+            throw new Error(
+              `Harness '${input.harness.harnessId}' emitted approval request '${displayValue.approvalId}' for unknown tool call '${displayValue.toolCallId}'.`,
+            );
+          }
+          const rawToolCall = rawToolCallsByToolCallId.get(
+            displayValue.toolCallId,
+          );
+          const toolName = rawToolCall?.toolName ?? toolCall.toolName;
+          if (
+            !isHarnessV1BuiltinToolIncluded({
+              toolName,
+              toolFiltering: input.builtinToolFiltering,
+            })
+          ) {
+            if (control.submitToolApproval == null) {
+              throw new Error(
+                `Harness '${input.harness.harnessId}' emitted a built-in tool approval request but does not support approval responses.`,
+              );
+            }
+            await control.submitToolApproval({
+              approvalId: displayValue.approvalId,
+              approved: false,
+              reason: getHarnessV1BuiltinToolFilteringDenialReason({
+                toolName,
+              }),
+            });
+            continue;
+          }
+        }
+
+        // Forward to consumer as soon as possible.
+        for (const part of translateStreamPart<TOOLS>(displayValue)) {
+          result.enqueue(part);
+        }
+
+        // Tool-call validation lives here (not in translateStreamPart) because
+        // schema parsing is async and needs the merged tool set in scope.
+        if (displayValue.type === 'tool-call') {
+          const parsed = await validateToolCall<TOOLS>({
+            event: displayValue,
+            tools: input.tools,
+          });
+          const parsedToolCall = asToolCallTextStreamPart({ part: parsed });
+          rawToolCallsByToolCallId.set(displayValue.toolCallId, displayValue);
+          toolCallsByToolCallId.set(displayValue.toolCallId, parsedToolCall);
+          result.enqueue(parsed);
+        }
+
+        // Accumulate output content for telemetry / reporters.
+        if (value.type === 'text-delta') {
+          stepText += value.delta;
+        } else if (value.type === 'reasoning-delta') {
+          stepReasoning += value.delta;
+        }
+
+        // Telemetry: a tool execution begins on its `tool-call`.
+        if (value.type === 'tool-call') {
+          stepToolCalls.push({
+            type: 'tool-call',
+            toolCallId: value.toolCallId,
+            toolName: value.toolName,
+            input: value.input,
+          });
+          telemetry.toolStart({
+            toolCallId: value.toolCallId,
+            toolName: value.toolName,
+            input: value.input,
+          });
+        }
+
+        // Telemetry: close a tool span when its provider-executed result lands.
+        if (value.type === 'tool-result') {
+          telemetry.toolEnd(
+            value.toolCallId,
+            value.isError
+              ? { ok: false, error: value.result }
+              : { ok: true, output: value.result },
+          );
+        }
+
+        if (value.type === 'tool-approval-request') {
+          const toolCall = toolCallsByToolCallId.get(value.toolCallId);
+          if (toolCall == null) {
+            throw new Error(
+              `Harness '${input.harness.harnessId}' emitted approval request '${value.approvalId}' for unknown tool call '${value.toolCallId}'.`,
+            );
+          }
+
+          const rawToolCall = rawToolCallsByToolCallId.get(value.toolCallId);
+          const pendingApproval =
+            pendingApprovalsByApprovalId.get(value.approvalId) ??
+            ({
+              approvalId: value.approvalId,
+              toolCallId: value.toolCallId,
+              toolName: toolCall.toolName,
+              input: rawToolCall?.input ?? JSON.stringify(toolCall.input),
+              kind: 'builtin',
+              providerExecuted: rawToolCall?.providerExecuted ?? true,
+              ...(rawToolCall?.nativeName !== undefined
+                ? { nativeName: rawToolCall.nativeName }
+                : {}),
+            } satisfies HarnessV1PendingToolApproval);
+          pendingApprovalsByApprovalId.set(
+            pendingApproval.approvalId,
+            pendingApproval,
+          );
+          pendingApprovalsByToolCallId.set(
+            pendingApproval.toolCallId,
+            pendingApproval,
+          );
+
+          const continuation = continuationsByApprovalId.get(
+            pendingApproval.approvalId,
+          );
+          if (continuation != null) {
+            await processPendingApprovalContinuation(
+              pendingApproval,
+              continuation,
+            );
+            continue;
+          }
+
+          onPendingToolApproval(pendingApproval);
+          enqueueApprovalRequest({
+            approvalId: pendingApproval.approvalId,
+            toolCall,
+          });
+          await finishForToolApprovalPause();
+          return;
+        }
+
+        // Drive step boundaries.
+        if (value.type === 'finish-step') {
+          telemetry.stepFinish({
+            finishReason: value.finishReason,
+            usage: value.usage,
+            providerMetadata: value.harnessMetadata,
+            content: buildStepContent(),
+          });
+          resetStepContent();
+          result.finishStep({
+            finishReason: value.finishReason,
+            usage: value.usage,
+            providerMetadata: value.harnessMetadata,
+            warnings: [],
+          });
+        }
+
+        if (value.type === 'finish') {
+          finalFinish = value;
+          telemetry.end({
+            finishReason: value.finishReason,
+            usage: value.totalUsage,
+          });
+        }
+
+        // Execute host-side tools when the harness asks for one.
+        if (value.type === 'tool-call' && !value.providerExecuted) {
+          const toolCall = value;
+          const parsedToolCall = toolCallsByToolCallId.get(toolCall.toolCallId);
+          if (parsedToolCall == null) {
+            throw new Error(
+              `Harness '${input.harness.harnessId}' could not find parsed tool call '${toolCall.toolCallId}' for custom tool approval.`,
+            );
+          }
+          if (!hasTool({ tools: activeTools, toolName: toolCall.toolName })) {
+            const output = {
+              type: 'execution-denied',
+              reason: getHarnessV1BuiltinToolFilteringDenialReason({
+                toolName: toolCall.toolName,
+              }),
+            };
+            await control.submitToolResult({
+              toolCallId: toolCall.toolCallId,
+              output,
+            });
+            telemetry.toolEnd(toolCall.toolCallId, { ok: true, output });
+            continue;
+          }
+          const customToolApprovalDecision = resolveCustomToolApproval({
+            toolName: toolCall.toolName,
+            toolApproval: input.toolApproval,
+          });
+          if (customToolApprovalDecision.type === 'deny') {
+            const approvalId = generateId();
+            enqueueApprovalRequest({
+              approvalId,
+              toolCall: parsedToolCall,
+              isAutomatic: true,
+            });
+            enqueueAutomaticApprovalResponse({
+              approvalId,
+              toolCall: parsedToolCall,
+              approved: false,
+              reason: customToolApprovalDecision.reason,
+              providerExecuted: false,
+            });
+            const output = {
+              type: 'execution-denied',
+              reason: customToolApprovalDecision.reason,
+            };
+            await control.submitToolResult({
+              toolCallId: toolCall.toolCallId,
+              output,
+            });
+            telemetry.toolEnd(toolCall.toolCallId, { ok: true, output });
+            continue;
+          }
+          const pendingApproval =
+            pendingApprovalsByToolCallId.get(toolCall.toolCallId) ??
+            (customToolApprovalDecision.type === 'request'
+              ? ({
+                  approvalId: generateId(),
+                  toolCallId: toolCall.toolCallId,
+                  toolName: toolCall.toolName,
+                  input: toolCall.input,
+                  kind: 'custom',
+                  providerExecuted: false,
+                  ...(toolCall.nativeName !== undefined
+                    ? { nativeName: toolCall.nativeName }
+                    : {}),
+                } satisfies HarnessV1PendingToolApproval)
+              : undefined);
+          if (pendingApproval != null) {
+            pendingApprovalsByApprovalId.set(
+              pendingApproval.approvalId,
+              pendingApproval,
+            );
+            pendingApprovalsByToolCallId.set(
+              pendingApproval.toolCallId,
+              pendingApproval,
+            );
+            const continuation = continuationsByApprovalId.get(
+              pendingApproval.approvalId,
+            );
+            if (continuation != null) {
+              await processPendingApprovalContinuation(
+                pendingApproval,
+                continuation,
+              );
+              continue;
+            }
+            const pendingParsedToolCall = toolCallsByToolCallId.get(
+              pendingApproval.toolCallId,
+            );
+            if (pendingParsedToolCall == null) {
+              throw new Error(
+                `Harness '${input.harness.harnessId}' could not find parsed tool call '${pendingApproval.toolCallId}' for approval request '${pendingApproval.approvalId}'.`,
+              );
+            }
+            onPendingToolApproval(pendingApproval);
+            enqueueApprovalRequest({
+              approvalId: pendingApproval.approvalId,
+              toolCall: pendingParsedToolCall,
+            });
+            await finishForToolApprovalPause();
+            return;
+          }
+          const outcome = await maybeExecuteHostTool({
+            event: toolCall,
+            tools: activeTools,
+            sandboxSession: input.sandboxSession,
+            abortSignal: input.abortSignal,
+            control,
+            onPreliminaryResult: preliminaryOutput => {
+              /*
+               * Project a `yield`ed value as a preliminary AI SDK
+               * `tool-result` part. Unlike the final result — which is
+               * submitted to the runtime, echoed back as a `tool-result`
+               * event, and stripped on its way through the loop above —
+               * preliminary values never reach the runtime, so strip the
+               * working directory here to match the final result's projection.
+               */
+              const stripped = stripWorkDir(
+                {
+                  type: 'tool-result',
+                  toolCallId: toolCall.toolCallId,
+                  toolName: toolCall.toolName,
+                  result: preliminaryOutput as Extract<
+                    HarnessV1StreamPart,
+                    { type: 'tool-result' }
+                  >['result'],
+                },
+                input.sessionWorkDir,
+              ) as Extract<HarnessV1StreamPart, { type: 'tool-result' }>;
+              result.enqueue({
+                type: 'tool-result',
+                toolCallId: toolCall.toolCallId,
+                toolName: toolCall.toolName,
+                input: undefined,
+                output: stripped.result,
+                preliminary: true,
+              } as TextStreamPart<TOOLS>);
+            },
+          });
+          telemetry.toolEnd(toolCall.toolCallId, outcome);
+        }
+
+        if (value.type === 'error') {
+          telemetry.error(value.error);
+          logBridgeError({
+            harnessId: input.harness.harnessId,
+            sessionId: input.session.sessionId,
+            context: 'harness stream error',
+            error: value.error,
+          });
+          result.fail(value.error);
+          return;
+        }
+      }
+      input.onTurnFinished?.();
+      await result.finish(
+        finalFinish
+          ? {
+              finishReason: finalFinish.finishReason,
+              totalUsage: finalFinish.totalUsage,
+              providerMetadata: finalFinish.harnessMetadata,
+            }
+          : undefined,
+      );
+    } catch (err) {
+      telemetry.error(err);
+      logBridgeError({
+        harnessId: input.harness.harnessId,
+        sessionId: input.session.sessionId,
+        context: 'harness turn failed',
+        error: err,
+      });
+      result.fail(err);
+    } finally {
+      reader.releaseLock();
+    }
+  })();
+
+  // Swallow the loop's rejection at the top level — failures are observable
+  // via the result's `fullStream` `error` part and rejected promise
+  // accessors. We do not want the orphan promise to become an unhandled
+  // rejection.
+  done.catch(() => {});
+
+  return { result, done };
+}
+
+type HostToolOutcome =
+  | { ok: true; output: unknown }
+  | { ok: false; error: unknown };
+
+function asToolCallTextStreamPart<TOOLS extends ToolSet>(input: {
+  part: TextStreamPart<TOOLS>;
+}): ToolCallTextStreamPart {
+  if (input.part.type !== 'tool-call') {
+    throw new Error(
+      `Expected parsed tool-call stream part, got '${input.part.type}'.`,
+    );
+  }
+  return input.part as ToolCallTextStreamPart;
+}
+
+type ToolCallTextStreamPart = {
+  readonly type: 'tool-call';
+  readonly toolCallId: string;
+  readonly toolName: string;
+  readonly input: unknown;
+  readonly providerExecuted?: boolean;
+  readonly providerMetadata?: unknown;
+  readonly dynamic?: boolean;
+  readonly invalid?: boolean;
+  readonly error?: unknown;
+  readonly title?: string;
+};
+
+function hasTool(input: { tools: ToolSet; toolName: string }): boolean {
+  return Object.prototype.hasOwnProperty.call(input.tools, input.toolName);
+}
+
+async function maybeExecuteHostTool<TOOLS extends ToolSet>(input: {
+  event: { toolCallId: string; toolName: string; input: string };
+  tools: TOOLS;
+  sandboxSession: SandboxSession;
+  abortSignal: AbortSignal | undefined;
+  control: HarnessV1PromptControl;
+  /**
+   * Called for each value a generator `execute` `yield`s before its last. The
+   * caller surfaces these as preliminary `tool-result` parts on the consumer
+   * stream. Never called for a plain (non-generator) `execute`.
+   */
+  onPreliminaryResult: (output: unknown) => void;
+}): Promise<HostToolOutcome> {
+  const tool = input.tools[input.event.toolName];
+
+  if (!isExecutableTool(tool)) return { ok: true, output: undefined };
+
+  const parsed = await safeParseJSON({ text: input.event.input });
+  const args = parsed.success ? parsed.value : input.event.input;
+
+  try {
+    /*
+     * Normalize the tool's return value through `executeTool`, the same helper
+     * the non-harness AI SDK uses, so generator `execute` functions behave
+     * identically here: each `yield`ed value arrives as a `preliminary` part
+     * and the last `yield` is re-emitted as the `final` part; a plain value or
+     * Promise arrives as a single `final` part. The underlying runtimes accept
+     * exactly one tool result per call, so only the final value is submitted
+     * back to the model — preliminary values are surfaced to the consumer
+     * stream alone, matching how the AI SDK treats `onPreliminaryToolResult`.
+     */
+    let output: unknown;
+    const stream = executeTool({
+      tool,
+      input: args as never,
+      options: {
+        toolCallId: input.event.toolCallId,
+        messages: [],
+        abortSignal: input.abortSignal,
+        context: undefined as never,
+        experimental_sandbox: input.sandboxSession,
+      },
+    });
+    for await (const part of stream) {
+      if (part.type === 'preliminary') {
+        input.onPreliminaryResult(part.output);
+      } else {
+        output = part.output;
+      }
+    }
+
+    await input.control.submitToolResult({
+      toolCallId: input.event.toolCallId,
+      output,
+    });
+    return { ok: true, output };
+  } catch (err) {
+    await input.control.submitToolResult({
+      toolCallId: input.event.toolCallId,
+      output: { error: String(err) },
+      isError: true,
+    });
+    return { ok: false, error: err };
+  }
+}
+
+/*
+ * Validate an inbound `tool-call` event against the merged tool set's schema
+ * using the AI SDK's canonical `parseToolCall`. Returns an AI SDK `tool-call`
+ * stream part with parsed input on success, or a `dynamic + invalid: true`
+ * part on failure (unknown tool, schema mismatch, malformed JSON).
+ *
+ * The harness `tool-call` event is structurally a `LanguageModelV4ToolCall`
+ * (plus an optional harness-only `nativeName`). `providerExecuted` already
+ * lives on the V4 type — `true` for adapter builtins (Claude Code's `Bash`,
+ * Codex's `shell`), false/undefined for host tools — and is passed through
+ * to the AI SDK part by `parseToolCall`.
+ */
+export async function validateToolCall<TOOLS extends ToolSet>(args: {
+  event: Extract<HarnessV1StreamPart, { type: 'tool-call' }>;
+  tools: TOOLS;
+}): Promise<TextStreamPart<TOOLS>> {
+  const { event, tools } = args;
+  const toolCall: LanguageModelV4ToolCall = {
+    type: 'tool-call',
+    toolCallId: event.toolCallId,
+    toolName: event.toolName,
+    input: event.input,
+    ...(event.providerExecuted !== undefined
+      ? { providerExecuted: event.providerExecuted }
+      : {}),
+    ...(event.providerMetadata !== undefined
+      ? { providerMetadata: event.providerMetadata }
+      : {}),
+  };
+
+  const parsed = await parseToolCall<TOOLS>({
+    toolCall,
+    tools,
+    repairToolCall: undefined,
+    refineToolInput: undefined,
+    instructions: undefined,
+    messages: [],
+  });
+
+  return parsed as TextStreamPart<TOOLS>;
+}
+
+/** Best-effort plain text of the turn's prompt, for telemetry input messages. */
+function promptToText(prompt: HarnessV1Prompt): string {
+  if (typeof prompt === 'string') return prompt;
+  const content = (prompt as { content?: unknown }).content;
+  if (typeof content === 'string') return content;
+  if (Array.isArray(content)) {
+    return content
+      .filter(
+        (part): part is { type: 'text'; text: string } =>
+          typeof part === 'object' &&
+          part != null &&
+          (part as { type?: unknown }).type === 'text',
+      )
+      .map(part => part.text)
+      .join('');
+  }
+  return '';
+}
+
+// keep import bound so unused-but-needed type stays cited
+export type _ContentPartMarker<T extends ToolSet> = ContentPart<T>;