@poncho-ai/harness 0.14.0 → 0.14.1

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.14.0 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.14.1 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
 > tsup src/index.ts --format esm --dts
 
 CLI Building entry: src/index.ts
@@ -7,8 +7,8 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js 173.62 KB
-ESM ⚡️ Build success in 98ms
+ESM dist/index.js 175.01 KB
+ESM ⚡️ Build success in 76ms
 DTS Build start
-DTS ⚡️ Build success in 5281ms
-DTS dist/index.d.ts 20.92 KB
+DTS ⚡️ Build success in 5663ms
+DTS dist/index.d.ts 21.38 KB

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # @poncho-ai/harness
 
+## 0.14.1
+
+### Patch Changes
+
+- [`e000b96`](https://github.com/cesr/poncho-ai/commit/e000b96837cbbb8d95c868c91a614f458868c444) Thanks [@cesr](https://github.com/cesr)! - Durable approval checkpoints, email conversation improvements, and web UI fixes
+  - Simplify approval system to checkpoint-only (remove legacy blocking approvalHandler)
+  - Optimize checkpoint storage with delta messages instead of full history
+  - Add sidebar sections for conversations awaiting approval with status indicator
+  - Fix nested checkpoint missing baseMessageCount in resumeRunFromCheckpoint
+  - Improve email conversation titles (sender email + subject)
+  - Remove email threading — each incoming email creates its own conversation
+  - Fix streaming after approval to preserve existing messages (liveOnly mode)
+  - Preserve newlines in user messages in web UI
+
+- Updated dependencies [[`e000b96`](https://github.com/cesr/poncho-ai/commit/e000b96837cbbb8d95c868c91a614f458868c444)]:
+  - @poncho-ai/sdk@1.0.2
+
 ## 0.14.0
 
 ### Minor Changes

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-import { Message, ToolDefinition, RunInput, AgentEvent, RunResult, JsonSchema, ToolContext } from '@poncho-ai/sdk';
+import { Message, ToolDefinition, ToolContext, RunInput, AgentEvent, RunResult, JsonSchema } from '@poncho-ai/sdk';
 export { ToolDefinition, defineTool } from '@poncho-ai/sdk';
 import { LanguageModel } from 'ai';
 import { z } from 'zod';
@@ -84,7 +84,15 @@ interface Conversation {
         approvalId: string;
         runId: string;
         tool: string;
+        toolCallId?: string;
         input: Record<string, unknown>;
+        checkpointMessages?: Message[];
+        baseMessageCount?: number;
+        pendingToolCalls?: Array<{
+            id: string;
+            name: string;
+            input: Record<string, unknown>;
+        }>;
     }>;
     ownerId: string;
     tenantId: string | null;
@@ -362,17 +370,33 @@ declare const getModelContextWindow: (modelName: string) => number;
  */
 declare const createModelProvider: (provider?: string) => ModelProviderFactory;
 
+interface ToolCall {
+    id: string;
+    name: string;
+    input: Record<string, unknown>;
+}
+interface ToolExecutionResult {
+    callId: string;
+    tool: string;
+    output?: unknown;
+    error?: string;
+}
+declare class ToolDispatcher {
+    private readonly tools;
+    register(tool: ToolDefinition): void;
+    registerMany(tools: ToolDefinition[]): void;
+    unregister(name: string): void;
+    unregisterMany(names: Iterable<string>): void;
+    list(): ToolDefinition[];
+    get(name: string): ToolDefinition | undefined;
+    execute(call: ToolCall, context: ToolContext): Promise<ToolExecutionResult>;
+    executeBatch(calls: ToolCall[], context: ToolContext): Promise<ToolExecutionResult[]>;
+}
+
 interface HarnessOptions {
     workingDir?: string;
     environment?: "development" | "staging" | "production";
     toolDefinitions?: ToolDefinition[];
-    approvalHandler?: (request: {
-        tool: string;
-        input: Record<string, unknown>;
-        runId: string;
-        step: number;
-        approvalId: string;
-    }) => Promise<boolean> | boolean;
     modelProvider?: ModelProviderFactory;
     uploadStore?: UploadStore;
 }
@@ -388,7 +412,6 @@ declare class AgentHarness {
     private modelProvider;
     private readonly modelProviderInjected;
     private readonly dispatcher;
-    private readonly approvalHandler?;
     readonly uploadStore?: UploadStore;
     private skillContextWindow;
     private memoryStore?;
@@ -439,6 +462,19 @@ declare class AgentHarness {
      */
     runWithTelemetry(input: RunInput): AsyncGenerator<AgentEvent>;
     run(input: RunInput): AsyncGenerator<AgentEvent>;
+    executeTools(calls: ToolCall[], context: ToolContext): Promise<ToolExecutionResult[]>;
+    continueFromToolResult(input: {
+        messages: Message[];
+        toolResults: Array<{
+            callId: string;
+            toolName: string;
+            result?: unknown;
+            error?: string;
+        }>;
+        conversationId?: string;
+        parameters?: Record<string, unknown>;
+        abortSignal?: AbortSignal;
+    }): AsyncGenerator<AgentEvent>;
     runToCompletion(input: RunInput): Promise<HarnessRunOutput>;
 }
 
@@ -563,27 +599,4 @@ declare class TelemetryEmitter {
     private sendOtlp;
 }
 
-interface ToolCall {
-    id: string;
-    name: string;
-    input: Record<string, unknown>;
-}
-interface ToolExecutionResult {
-    callId: string;
-    tool: string;
-    output?: unknown;
-    error?: string;
-}
-declare class ToolDispatcher {
-    private readonly tools;
-    register(tool: ToolDefinition): void;
-    registerMany(tools: ToolDefinition[]): void;
-    unregister(name: string): void;
-    unregisterMany(names: Iterable<string>): void;
-    list(): ToolDefinition[];
-    get(name: string): ToolDefinition | undefined;
-    execute(call: ToolCall, context: ToolContext): Promise<ToolExecutionResult>;
-    executeBatch(calls: ToolCall[], context: ToolContext): Promise<ToolExecutionResult[]>;
-}
-
 export { type AgentFrontmatter, AgentHarness, type AgentIdentity, type AgentLimitsConfig, type AgentModelConfig, type BuiltInToolToggles, type Conversation, type ConversationState, type ConversationStore, type CronJobConfig, type HarnessOptions, type HarnessRunOutput, InMemoryConversationStore, InMemoryStateStore, LatitudeCapture, type LatitudeCaptureConfig, LocalMcpBridge, LocalUploadStore, type MainMemory, type McpConfig, type MemoryConfig, type MemoryStore, type MessagingChannelConfig, type ModelProviderFactory, PONCHO_UPLOAD_SCHEME, type ParsedAgent, type PonchoConfig, type RemoteMcpServerConfig, type RuntimeRenderContext, S3UploadStore, STORAGE_SCHEMA_VERSION, type SkillContextEntry, type SkillMetadata, type StateConfig, type StateProviderName, type StateStore, type StorageConfig, type TelemetryConfig, TelemetryEmitter, type ToolAccess, type ToolCall, ToolDispatcher, type ToolExecutionResult, type UploadStore, type UploadsConfig, VercelBlobUploadStore, buildAgentDirectoryName, buildSkillContextWindow, createConversationStore, createDefaultTools, createMemoryStore, createMemoryTools, createModelProvider, createSkillTools, createStateStore, createUploadStore, createWriteTool, deriveUploadKey, ensureAgentIdentity, generateAgentId, getAgentStoreDirectory, getModelContextWindow, getPonchoStoreRoot, jsonSchemaToZod, loadPonchoConfig, loadSkillContext, loadSkillInstructions, loadSkillMetadata, normalizeScriptPolicyPath, parseAgentFile, parseAgentMarkdown, readSkillResource, renderAgentPrompt, resolveAgentIdentity, resolveMemoryConfig, resolveSkillDirs, resolveStateConfig, slugifyStorageComponent };

package/dist/index.js

@@ -2851,7 +2851,6 @@ var AgentHarness = class {
   modelProvider;
   modelProviderInjected;
   dispatcher = new ToolDispatcher();
-  approvalHandler;
   uploadStore;
   skillContextWindow = "";
   memoryStore;
@@ -2925,7 +2924,6 @@ var AgentHarness = class {
     this.environment = options.environment ?? "development";
     this.modelProviderInjected = !!options.modelProvider;
     this.modelProvider = options.modelProvider ?? createModelProvider("anthropic");
-    this.approvalHandler = options.approvalHandler;
     this.uploadStore = options.uploadStore;
     if (options.toolDefinitions?.length) {
       this.dispatcher.registerMany(options.toolDefinitions);
@@ -3317,6 +3315,7 @@ var AgentHarness = class {
     const platformMaxDurationSec = Number(process.env.PONCHO_MAX_DURATION) || 0;
     const softDeadlineMs = platformMaxDurationSec > 0 ? platformMaxDurationSec * 800 : 0;
     const messages = [...input.messages ?? []];
+    const inputMessageCount = messages.length;
     const events = [];
     const systemPrompt = renderAgentPrompt(agent, {
       parameters: input.parameters,
@@ -3366,41 +3365,43 @@ ${boundedMainMemory.trim()}` : "";
       agentId: agent.frontmatter.id ?? agent.frontmatter.name,
       contextWindow
     });
-    if (input.files && input.files.length > 0) {
-      const parts = [
-        { type: "text", text: input.task }
-      ];
-      for (const file of input.files) {
-        if (this.uploadStore) {
-          const buf = Buffer.from(file.data, "base64");
-          const key = deriveUploadKey(buf, file.mediaType);
-          const ref = await this.uploadStore.put(key, buf, file.mediaType);
-          parts.push({
-            type: "file",
-            data: ref,
-            mediaType: file.mediaType,
-            filename: file.filename
-          });
-        } else {
-          parts.push({
-            type: "file",
-            data: file.data,
-            mediaType: file.mediaType,
-            filename: file.filename
-          });
+    if (input.task != null) {
+      if (input.files && input.files.length > 0) {
+        const parts = [
+          { type: "text", text: input.task }
+        ];
+        for (const file of input.files) {
+          if (this.uploadStore) {
+            const buf = Buffer.from(file.data, "base64");
+            const key = deriveUploadKey(buf, file.mediaType);
+            const ref = await this.uploadStore.put(key, buf, file.mediaType);
+            parts.push({
+              type: "file",
+              data: ref,
+              mediaType: file.mediaType,
+              filename: file.filename
+            });
+          } else {
+            parts.push({
+              type: "file",
+              data: file.data,
+              mediaType: file.mediaType,
+              filename: file.filename
+            });
+          }
         }
+        messages.push({
+          role: "user",
+          content: parts,
+          metadata: { timestamp: now(), id: randomUUID3() }
+        });
+      } else {
+        messages.push({
+          role: "user",
+          content: input.task,
+          metadata: { timestamp: now(), id: randomUUID3() }
+        });
       }
-      messages.push({
-        role: "user",
-        content: parts,
-        metadata: { timestamp: now(), id: randomUUID3() }
-      });
-    } else {
-      messages.push({
-        role: "user",
-        content: input.task,
-        metadata: { timestamp: now(), id: randomUUID3() }
-      });
     }
     let responseText = "";
     let totalInputTokens = 0;
@@ -3843,45 +3844,34 @@ ${textContent}` };
               input: call.input,
               approvalId
             });
-            const approved = this.approvalHandler ? await this.approvalHandler({
+            const assistantContent2 = JSON.stringify({
+              text: fullText,
+              tool_calls: toolCalls.map((tc) => ({
+                id: tc.id,
+                name: exposedToolNames.get(tc.name) ?? tc.name,
+                input: tc.input
+              }))
+            });
+            const assistantMsg = {
+              role: "assistant",
+              content: assistantContent2,
+              metadata: { timestamp: now(), id: randomUUID3(), step }
+            };
+            const deltaMessages = [...messages.slice(inputMessageCount), assistantMsg];
+            yield pushEvent({
+              type: "tool:approval:checkpoint",
+              approvalId,
               tool: runtimeToolName,
+              toolCallId: call.id,
               input: call.input,
-              runId,
-              step,
-              approvalId
-            }) : false;
-            if (isCancelled()) {
-              yield emitCancellation();
-              return;
-            }
-            if (!approved) {
-              if (this.insideTelemetryCapture && this.latitudeTelemetry) {
-                const deniedSpan = this.latitudeTelemetry.span.tool({
-                  name: runtimeToolName,
-                  call: { id: call.id, arguments: call.input }
-                });
-                deniedSpan.end({ result: { value: "Tool execution denied by approval policy", isError: true } });
-              }
-              yield pushEvent({
-                type: "tool:approval:denied",
-                approvalId,
-                reason: "No approval handler granted execution"
-              });
-              yield pushEvent({
-                type: "tool:error",
-                tool: call.name,
-                error: "Tool execution denied by approval policy",
-                recoverable: true
-              });
-              toolResultsForModel.push({
-                type: "tool_result",
-                tool_use_id: call.id,
-                tool_name: runtimeToolName,
-                content: "Tool error: Tool execution denied by approval policy"
-              });
-              continue;
-            }
-            yield pushEvent({ type: "tool:approval:granted", approvalId });
+              checkpointMessages: deltaMessages,
+              pendingToolCalls: toolCalls.map((tc) => ({
+                id: tc.id,
+                name: exposedToolNames.get(tc.name) ?? tc.name,
+                input: tc.input
+              }))
+            });
+            return;
           }
           approvedCalls.push({
             id: call.id,
@@ -4015,12 +4005,64 @@ ${textContent}` };
       });
     }
   }
+  async executeTools(calls, context) {
+    return this.dispatcher.executeBatch(calls, context);
+  }
+  async *continueFromToolResult(input) {
+    const messages = [...input.messages];
+    const lastMsg = messages[messages.length - 1];
+    if (!lastMsg || lastMsg.role !== "assistant") {
+      throw new Error("continueFromToolResult: last message must be an assistant message with tool calls");
+    }
+    let allToolCalls = [];
+    try {
+      const parsed = JSON.parse(typeof lastMsg.content === "string" ? lastMsg.content : "");
+      allToolCalls = parsed.tool_calls ?? [];
+    } catch {
+      throw new Error("continueFromToolResult: could not parse tool calls from last assistant message");
+    }
+    const providedMap = new Map(
+      input.toolResults.map((r) => [r.callId, r])
+    );
+    const toolResultsForModel = [];
+    for (const tc of allToolCalls) {
+      const provided = providedMap.get(tc.id);
+      if (provided) {
+        toolResultsForModel.push({
+          type: "tool_result",
+          tool_use_id: tc.id,
+          tool_name: provided.toolName,
+          content: provided.error ? `Tool error: ${provided.error}` : JSON.stringify(provided.result ?? null)
+        });
+      } else {
+        toolResultsForModel.push({
+          type: "tool_result",
+          tool_use_id: tc.id,
+          tool_name: tc.name,
+          content: "Tool error: Tool execution deferred (pending approval checkpoint)"
+        });
+      }
+    }
+    messages.push({
+      role: "tool",
+      content: JSON.stringify(toolResultsForModel),
+      metadata: { timestamp: Date.now(), id: randomUUID3() }
+    });
+    yield* this.runWithTelemetry({
+      messages,
+      conversationId: input.conversationId,
+      parameters: input.parameters,
+      abortSignal: input.abortSignal
+    });
+  }
   async runToCompletion(input) {
     const events = [];
     let runId = "";
     let finalResult;
     const messages = [...input.messages ?? []];
-    messages.push({ role: "user", content: input.task });
+    if (input.task != null) {
+      messages.push({ role: "user", content: input.task });
+    }
     for await (const event of this.runWithTelemetry(input)) {
       events.push(event);
       if (event.type === "run:started") {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.14.0",
+  "version": "0.14.1",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",
@@ -31,7 +31,7 @@
     "redis": "^5.10.0",
     "yaml": "^2.4.0",
     "zod": "^3.22.0",
-    "@poncho-ai/sdk": "1.0.1"
+    "@poncho-ai/sdk": "1.0.2"
   },
   "devDependencies": {
     "@types/mustache": "^4.2.6",

package/src/harness.ts

@@ -37,20 +37,13 @@ import {
   matchesSlashPattern,
   normalizeRelativeScriptPattern,
 } from "./tool-policy.js";
-import { ToolDispatcher } from "./tool-dispatcher.js";
+import { ToolDispatcher, type ToolCall, type ToolExecutionResult } from "./tool-dispatcher.js";
 import { ensureAgentIdentity } from "./agent-identity.js";
 
 export interface HarnessOptions {
   workingDir?: string;
   environment?: "development" | "staging" | "production";
   toolDefinitions?: ToolDefinition[];
-  approvalHandler?: (request: {
-    tool: string;
-    input: Record<string, unknown>;
-    runId: string;
-    step: number;
-    approvalId: string;
-  }) => Promise<boolean> | boolean;
   modelProvider?: ModelProviderFactory;
   uploadStore?: UploadStore;
 }
@@ -414,7 +407,6 @@ export class AgentHarness {
   private modelProvider: ModelProviderFactory;
   private readonly modelProviderInjected: boolean;
   private readonly dispatcher = new ToolDispatcher();
-  private readonly approvalHandler?: HarnessOptions["approvalHandler"];
   readonly uploadStore?: UploadStore;
   private skillContextWindow = "";
   private memoryStore?: MemoryStore;
@@ -500,7 +492,6 @@ export class AgentHarness {
     this.environment = options.environment ?? "development";
     this.modelProviderInjected = !!options.modelProvider;
     this.modelProvider = options.modelProvider ?? createModelProvider("anthropic");
-    this.approvalHandler = options.approvalHandler;
     this.uploadStore = options.uploadStore;
 
     if (options.toolDefinitions?.length) {
@@ -963,6 +954,7 @@ export class AgentHarness {
       ? platformMaxDurationSec * 800
       : 0;
     const messages: Message[] = [...(input.messages ?? [])];
+    const inputMessageCount = messages.length;
     const events: AgentEvent[] = [];
 
     const systemPrompt = renderAgentPrompt(agent, {
@@ -1024,41 +1016,43 @@ ${boundedMainMemory.trim()}`
       contextWindow,
     });
 
-    if (input.files && input.files.length > 0) {
-      const parts: ContentPart[] = [
-        { type: "text", text: input.task } satisfies TextContentPart,
-      ];
-      for (const file of input.files) {
-        if (this.uploadStore) {
-          const buf = Buffer.from(file.data, "base64");
-          const key = deriveUploadKey(buf, file.mediaType);
-          const ref = await this.uploadStore.put(key, buf, file.mediaType);
-          parts.push({
-            type: "file",
-            data: ref,
-            mediaType: file.mediaType,
-            filename: file.filename,
-          } satisfies FileContentPart);
-        } else {
-          parts.push({
-            type: "file",
-            data: file.data,
-            mediaType: file.mediaType,
-            filename: file.filename,
-          } satisfies FileContentPart);
+    if (input.task != null) {
+      if (input.files && input.files.length > 0) {
+        const parts: ContentPart[] = [
+          { type: "text", text: input.task } satisfies TextContentPart,
+        ];
+        for (const file of input.files) {
+          if (this.uploadStore) {
+            const buf = Buffer.from(file.data, "base64");
+            const key = deriveUploadKey(buf, file.mediaType);
+            const ref = await this.uploadStore.put(key, buf, file.mediaType);
+            parts.push({
+              type: "file",
+              data: ref,
+              mediaType: file.mediaType,
+              filename: file.filename,
+            } satisfies FileContentPart);
+          } else {
+            parts.push({
+              type: "file",
+              data: file.data,
+              mediaType: file.mediaType,
+              filename: file.filename,
+            } satisfies FileContentPart);
+          }
         }
+        messages.push({
+          role: "user",
+          content: parts,
+          metadata: { timestamp: now(), id: randomUUID() },
+        });
+      } else {
+        messages.push({
+          role: "user",
+          content: input.task,
+          metadata: { timestamp: now(), id: randomUUID() },
+        });
       }
-      messages.push({
-        role: "user",
-        content: parts,
-        metadata: { timestamp: now(), id: randomUUID() },
-      });
-    } else {
-      messages.push({
-        role: "user",
-        content: input.task,
-        metadata: { timestamp: now(), id: randomUUID() },
-      });
     }
 
     let responseText = "";
@@ -1597,47 +1591,35 @@ ${boundedMainMemory.trim()}`
             input: call.input,
             approvalId,
           });
-          const approved = this.approvalHandler
-            ? await this.approvalHandler({
-                tool: runtimeToolName,
-                input: call.input,
-                runId,
-                step,
-                approvalId,
-              })
-            : false;
-          if (isCancelled()) {
-            yield emitCancellation();
-            return;
-          }
-          if (!approved) {
-            if (this.insideTelemetryCapture && this.latitudeTelemetry) {
-              const deniedSpan = this.latitudeTelemetry.span.tool({
-                name: runtimeToolName,
-                call: { id: call.id, arguments: call.input },
-              });
-              deniedSpan.end({ result: { value: "Tool execution denied by approval policy", isError: true } });
-            }
-            yield pushEvent({
-              type: "tool:approval:denied",
-              approvalId,
-              reason: "No approval handler granted execution",
-            });
-            yield pushEvent({
-              type: "tool:error",
-              tool: call.name,
-              error: "Tool execution denied by approval policy",
-              recoverable: true,
-            });
-            toolResultsForModel.push({
-              type: "tool_result",
-              tool_use_id: call.id,
-              tool_name: runtimeToolName,
-              content: "Tool error: Tool execution denied by approval policy",
-            });
-            continue;
-          }
-          yield pushEvent({ type: "tool:approval:granted", approvalId });
+
+          const assistantContent = JSON.stringify({
+            text: fullText,
+            tool_calls: toolCalls.map(tc => ({
+              id: tc.id,
+              name: exposedToolNames.get(tc.name) ?? tc.name,
+              input: tc.input,
+            })),
+          });
+          const assistantMsg: Message = {
+            role: "assistant",
+            content: assistantContent,
+            metadata: { timestamp: now(), id: randomUUID(), step },
+          };
+          const deltaMessages = [...messages.slice(inputMessageCount), assistantMsg];
+          yield pushEvent({
+            type: "tool:approval:checkpoint",
+            approvalId,
+            tool: runtimeToolName,
+            toolCallId: call.id,
+            input: call.input,
+            checkpointMessages: deltaMessages,
+            pendingToolCalls: toolCalls.map(tc => ({
+              id: tc.id,
+              name: exposedToolNames.get(tc.name) ?? tc.name,
+              input: tc.input,
+            })),
+          });
+          return;
         }
         approvedCalls.push({
           id: call.id,
@@ -1790,12 +1772,87 @@ ${boundedMainMemory.trim()}`
     }
   }
 
+  async executeTools(
+    calls: ToolCall[],
+    context: ToolContext,
+  ): Promise<ToolExecutionResult[]> {
+    return this.dispatcher.executeBatch(calls, context);
+  }
+
+  async *continueFromToolResult(input: {
+    messages: Message[];
+    toolResults: Array<{ callId: string; toolName: string; result?: unknown; error?: string }>;
+    conversationId?: string;
+    parameters?: Record<string, unknown>;
+    abortSignal?: AbortSignal;
+  }): AsyncGenerator<AgentEvent> {
+    const messages = [...input.messages];
+    const lastMsg = messages[messages.length - 1];
+    if (!lastMsg || lastMsg.role !== "assistant") {
+      throw new Error("continueFromToolResult: last message must be an assistant message with tool calls");
+    }
+
+    let allToolCalls: Array<{ id: string; name: string; input: Record<string, unknown> }> = [];
+    try {
+      const parsed = JSON.parse(typeof lastMsg.content === "string" ? lastMsg.content : "");
+      allToolCalls = parsed.tool_calls ?? [];
+    } catch {
+      throw new Error("continueFromToolResult: could not parse tool calls from last assistant message");
+    }
+
+    const providedMap = new Map(
+      input.toolResults.map(r => [r.callId, r]),
+    );
+    const toolResultsForModel: Array<{
+      type: "tool_result";
+      tool_use_id: string;
+      tool_name: string;
+      content: string;
+    }> = [];
+
+    for (const tc of allToolCalls) {
+      const provided = providedMap.get(tc.id);
+      if (provided) {
+        toolResultsForModel.push({
+          type: "tool_result",
+          tool_use_id: tc.id,
+          tool_name: provided.toolName,
+          content: provided.error
+            ? `Tool error: ${provided.error}`
+            : JSON.stringify(provided.result ?? null),
+        });
+      } else {
+        toolResultsForModel.push({
+          type: "tool_result",
+          tool_use_id: tc.id,
+          tool_name: tc.name,
+          content: "Tool error: Tool execution deferred (pending approval checkpoint)",
+        });
+      }
+    }
+
+    messages.push({
+      role: "tool",
+      content: JSON.stringify(toolResultsForModel),
+      metadata: { timestamp: Date.now(), id: randomUUID() },
+    });
+
+    yield* this.runWithTelemetry({
+      messages,
+      conversationId: input.conversationId,
+      parameters: input.parameters,
+      abortSignal: input.abortSignal,
+    });
+  }
+
   async runToCompletion(input: RunInput): Promise<HarnessRunOutput> {
     const events: AgentEvent[] = [];
     let runId = "";
     let finalResult: RunResult | undefined;
     const messages: Message[] = [...(input.messages ?? [])];
-    messages.push({ role: "user", content: input.task });
+    if (input.task != null) {
+      messages.push({ role: "user", content: input.task });
+    }
 
     for await (const event of this.runWithTelemetry(input)) {
       events.push(event);

package/src/state.ts

@@ -30,7 +30,11 @@ export interface Conversation {
     approvalId: string;
     runId: string;
     tool: string;
+    toolCallId?: string;
     input: Record<string, unknown>;
+    checkpointMessages?: Message[];
+    baseMessageCount?: number;
+    pendingToolCalls?: Array<{ id: string; name: string; input: Record<string, unknown> }>;
   }>;
   ownerId: string;
   tenantId: string | null;