@poncho-ai/harness 0.31.0 → 0.31.2

package/dist/index.js

@@ -361,9 +361,11 @@ var OVERHEAD_MULTIPLIER = 1.15;
 var MIN_COMPACTABLE_MESSAGES = 4;
 var DEFAULT_COMPACTION_CONFIG = {
   enabled: true,
-  trigger: 0.8,
-  keepRecentMessages: 6
+  trigger: 0.75,
+  keepRecentMessages: 4
 };
+var SUMMARIZATION_MESSAGE_TRUNCATION_CHARS = 1200;
+var SUMMARIZATION_MAX_OUTPUT_TOKENS = 768;
 var SUMMARIZATION_PROMPT = `Summarize the following conversation into a structured working state that allows continuation without re-asking questions. Include:
 
 1. **User intent**: What the user originally asked for and any refinements
@@ -418,7 +420,7 @@ var buildSummarizationMessages = (messagesToCompact, instructions) => {
   const conversationLines = [];
   for (const msg of messagesToCompact) {
     const text = getTextContent(msg);
-    const truncated = text.length > 2e3 ? text.slice(0, 2e3) + "\n...[truncated]" : text;
+    const truncated = text.length > SUMMARIZATION_MESSAGE_TRUNCATION_CHARS ? text.slice(0, SUMMARIZATION_MESSAGE_TRUNCATION_CHARS) + "\n...[truncated]" : text;
     conversationLines.push(`[${msg.role}]: ${truncated}`);
   }
   const prompt = instructions ? `${SUMMARIZATION_PROMPT}
@@ -473,7 +475,7 @@ var compactMessages = async (model, messages, config, options) => {
     const result = await generateText({
       model,
       messages: summarizationMessages,
-      maxOutputTokens: 2048
+      maxOutputTokens: SUMMARIZATION_MAX_OUTPUT_TOKENS
     });
     const summary = result.text.trim();
     if (!summary) {
@@ -1658,7 +1660,7 @@ Logs print to console:
 [event] run:started {"type":"run:started","runId":"run_abc123","agentId":"my-agent"}
 [event] tool:started {"type":"tool:started","tool":"read_file","input":{"path":"README.md"}}
 [event] tool:completed {"type":"tool:completed","tool":"read_file","duration":45,"output":{"path":"README.md","content":"..."}}
-[event] run:completed {"type":"run:completed","runId":"run_abc123","result":{"status":"completed","response":"...","steps":3,"tokens":{"input":1500,"output":840}}}
+[event] run:completed {"type":"run:completed","runId":"run_abc123","result":{"status":"completed","response":"...","steps":3,"tokens":{"input":1500,"output":840,"cached":1200,"cacheWrite":300}}}
 \`\`\`
 
 ### Production telemetry (generic OTLP)
@@ -2110,7 +2112,7 @@ var ponchoDocsTool = defineTool({
 import { randomUUID as randomUUID3 } from "crypto";
 import { readFile as readFile9 } from "fs/promises";
 import { resolve as resolve11 } from "path";
-import { getTextContent as getTextContent2 } from "@poncho-ai/sdk";
+import { defineTool as defineTool7, getTextContent as getTextContent2 } from "@poncho-ai/sdk";
 
 // src/upload-store.ts
 import { createHash as createHash2 } from "crypto";
@@ -3887,6 +3889,7 @@ var MODEL_CONTEXT_WINDOWS = {
 };
 var DEFAULT_CONTEXT_WINDOW = 2e5;
 var OPENAI_CODEX_DEFAULT_INSTRUCTIONS = "You are Codex, based on GPT-5. You are running as a coding agent in Poncho.";
+var OPENAI_CODEX_RESPONSES_URL = process.env.OPENAI_CODEX_RESPONSES_URL ?? "https://chatgpt.com/backend-api/codex/responses";
 var extractSystemInstructionFromInput = (input) => {
   if (!Array.isArray(input)) return void 0;
   for (const message of input) {
@@ -3909,6 +3912,18 @@ var extractSystemInstructionFromInput = (input) => {
   }
   return void 0;
 };
+var normalizeToolParameterSchemas = (tools) => {
+  if (!Array.isArray(tools)) return;
+  for (const tool of tools) {
+    if (!tool || typeof tool !== "object") continue;
+    const entry = tool;
+    if (!entry.parameters || typeof entry.parameters !== "object") continue;
+    const schema = entry.parameters;
+    if (schema.type === "object" && (typeof schema.properties !== "object" || schema.properties === null)) {
+      schema.properties = {};
+    }
+  }
+};
 var getModelContextWindow = (modelName) => {
   if (MODEL_CONTEXT_WINDOWS[modelName] !== void 0) {
     return MODEL_CONTEXT_WINDOWS[modelName];
@@ -3938,7 +3953,7 @@ var createModelProvider = (provider, config) => {
         const originalUrl = input instanceof URL ? input.toString() : typeof input === "string" ? input : input.url;
         const parsed = new URL(originalUrl);
         const shouldRewrite = parsed.pathname.includes("/v1/responses") || parsed.pathname.includes("/chat/completions");
-        const targetUrl = shouldRewrite ? "https://chatgpt.com/backend-api/codex/responses" : originalUrl;
+        const targetUrl = shouldRewrite ? OPENAI_CODEX_RESPONSES_URL : originalUrl;
         let body = init?.body;
         if (shouldRewrite && typeof body === "string" && headers.get("Content-Type")?.includes("application/json")) {
           try {
@@ -3946,12 +3961,21 @@ var createModelProvider = (provider, config) => {
             if (typeof payload.instructions !== "string" || payload.instructions.trim() === "") {
               payload.instructions = extractSystemInstructionFromInput(payload.input) ?? OPENAI_CODEX_DEFAULT_INSTRUCTIONS;
             }
+            normalizeToolParameterSchemas(payload.tools);
             payload.store = false;
             body = JSON.stringify(payload);
           } catch {
           }
         }
-        return fetch(targetUrl, { ...init, headers, body });
+        try {
+          return await fetch(targetUrl, { ...init, headers, body });
+        } catch (error) {
+          const message = error instanceof Error ? error.message : String(error);
+          if (shouldRewrite && targetUrl.includes("chatgpt.com") && message.includes("ENOTFOUND chatgpt.com")) {
+            return fetch(originalUrl, { ...init, headers, body });
+          }
+          throw error;
+        }
       }
     });
     return (modelName) => openai(modelName);
@@ -4993,6 +5017,9 @@ var TelemetryEmitter = class {
     if (otlp) {
       await this.sendOtlp(event, otlp);
     }
+    if (event.type === "model:chunk") {
+      return;
+    }
     process.stdout.write(`[event] ${event.type} ${sanitizeEventForLog(event)}
 `);
   }
@@ -5119,8 +5146,12 @@ var ToolDispatcher = class {
 
 // src/harness.ts
 var now = () => Date.now();
-var FIRST_CHUNK_TIMEOUT_MS = 3e5;
-var MAX_TRANSIENT_STEP_RETRIES = 2;
+var FIRST_CHUNK_TIMEOUT_MS = 9e4;
+var MAX_TRANSIENT_STEP_RETRIES = 1;
+var COMPACTION_CHECK_INTERVAL_STEPS = 3;
+var TOOL_RESULT_ARCHIVE_PARAM = "__toolResultArchive";
+var TOOL_RESULT_TRUNCATED_PREFIX = "[TRUNCATED_TOOL_RESULT]";
+var TOOL_RESULT_PREVIEW_CHARS = 700;
 var FirstChunkTimeoutError = class extends Error {
   constructor(modelName, timeoutMs) {
     super(
@@ -5179,18 +5210,11 @@ var isRetryableModelError = (error) => {
   if (error instanceof FirstChunkTimeoutError) {
     return true;
   }
-  if (isNoOutputGeneratedError(error)) {
-    return true;
-  }
   const statusCode = getErrorStatusCode(error);
   if (typeof statusCode === "number") {
     return statusCode === 429 || statusCode >= 500;
   }
-  if (!error || typeof error !== "object") {
-    return false;
-  }
-  const maybeMessage = "message" in error ? String(error.message ?? "").toLowerCase() : "";
-  return maybeMessage.includes("internal server error") || maybeMessage.includes("service unavailable") || maybeMessage.includes("gateway timeout") || maybeMessage.includes("rate limit");
+  return false;
 };
 var toRunError = (error) => {
   const statusCode = getErrorStatusCode(error);
@@ -5242,6 +5266,58 @@ var toProviderSafeToolName = (originalName, index, used) => {
   used.add(candidate);
   return candidate;
 };
+var isToolResultRow = (value) => {
+  if (typeof value !== "object" || value === null) return false;
+  const row = value;
+  return typeof row.tool_use_id === "string" && typeof row.tool_name === "string" && typeof row.content === "string";
+};
+var readArchiveFromParameters = (parameters) => {
+  const raw = parameters?.[TOOL_RESULT_ARCHIVE_PARAM];
+  if (typeof raw !== "object" || raw === null) return {};
+  const out = {};
+  for (const [key, value] of Object.entries(raw)) {
+    if (typeof value !== "object" || value === null) continue;
+    const row = value;
+    if (typeof row.toolResultId !== "string" || typeof row.conversationId !== "string" || typeof row.toolName !== "string" || typeof row.toolCallId !== "string" || typeof row.createdAt !== "number" || typeof row.sizeBytes !== "number" || typeof row.payload !== "string") {
+      continue;
+    }
+    out[key] = {
+      toolResultId: row.toolResultId,
+      conversationId: row.conversationId,
+      toolName: row.toolName,
+      toolCallId: row.toolCallId,
+      createdAt: row.createdAt,
+      sizeBytes: row.sizeBytes,
+      payload: row.payload
+    };
+  }
+  return out;
+};
+var makeTruncatedToolResultNotice = (toolResultId, toolName, payload) => {
+  const preview = payload.slice(0, TOOL_RESULT_PREVIEW_CHARS);
+  const omittedChars = Math.max(0, payload.length - preview.length);
+  return `${TOOL_RESULT_TRUNCATED_PREFIX} id="${toolResultId}" tool="${toolName}" omittedChars=${omittedChars}
+${preview}${omittedChars > 0 ? "\n...[truncated]" : ""}`;
+};
+var hasUntruncatedToolResults = (messages) => {
+  for (const msg of messages) {
+    if (msg.role !== "tool" || typeof msg.content !== "string") continue;
+    let parsed;
+    try {
+      parsed = JSON.parse(msg.content);
+    } catch {
+      continue;
+    }
+    if (!Array.isArray(parsed)) continue;
+    for (const row of parsed) {
+      if (!isToolResultRow(row)) continue;
+      if (!row.content.startsWith(TOOL_RESULT_TRUNCATED_PREFIX)) {
+        return true;
+      }
+    }
+  }
+  return false;
+};
 var DEVELOPMENT_MODE_CONTEXT = `## Development Mode Context
 
 You are running locally in development mode. Treat this as an editable agent workspace.
@@ -5573,6 +5649,7 @@ var AgentHarness = class _AgentHarness {
   agentFileFingerprint = "";
   mcpBridge;
   subagentManager;
+  archivedToolResultsByConversation = /* @__PURE__ */ new Map();
   resolveToolAccess(toolName) {
     const tools = this.loadedConfig?.tools;
     if (!tools) return true;
@@ -5644,6 +5721,56 @@ var AgentHarness = class _AgentHarness {
     if (this.environment === "development" && this.isToolEnabled("poncho_docs")) {
       this.registerIfMissing(ponchoDocsTool);
     }
+    if (this.isToolEnabled("get_tool_result_by_id")) {
+      this.registerIfMissing(this.createGetToolResultByIdTool());
+    }
+  }
+  createGetToolResultByIdTool() {
+    return defineTool7({
+      name: "get_tool_result_by_id",
+      description: "Retrieve a previously archived full tool result by id for the current conversation. Use this when older tool outputs were truncated in prompt history.",
+      inputSchema: {
+        type: "object",
+        properties: {
+          toolResultId: { type: "string", description: "Archived tool result id to retrieve" },
+          offset: { type: "number", description: "Optional character offset for paging large payloads" },
+          limit: { type: "number", description: "Optional maximum characters to return (default 6000, max 20000)" }
+        },
+        required: ["toolResultId"],
+        additionalProperties: false
+      },
+      handler: async (input, context) => {
+        const conversationId = context.conversationId ?? "__default__";
+        const archive = this.archivedToolResultsByConversation.get(conversationId) ?? {};
+        const toolResultId = typeof input.toolResultId === "string" ? input.toolResultId : "";
+        const record = archive[toolResultId];
+        if (!record) {
+          console.info(
+            `[poncho][cost] Archived tool result lookup miss: id="${toolResultId}" conversation="${conversationId}"`
+          );
+          return {
+            error: `No archived tool result found for id "${toolResultId}" in this conversation.`
+          };
+        }
+        const offset = Math.max(0, Number(input.offset) || 0);
+        const limit = Math.min(Math.max(Number(input.limit) || 6e3, 1), 2e4);
+        const end = Math.min(record.payload.length, offset + limit);
+        const chunk = record.payload.slice(offset, end);
+        console.info(
+          `[poncho][cost] Archived tool result lookup hit: id="${toolResultId}" conversation="${conversationId}" offset=${offset} returned=${chunk.length} total=${record.payload.length}`
+        );
+        return {
+          toolResultId: record.toolResultId,
+          toolName: record.toolName,
+          toolCallId: record.toolCallId,
+          totalChars: record.payload.length,
+          offset,
+          returnedChars: chunk.length,
+          hasMore: end < record.payload.length,
+          payload: chunk
+        };
+      }
+    });
   }
   shouldEnableWriteTool() {
     const override = process.env.PONCHO_FS_WRITE?.toLowerCase();
@@ -5668,6 +5795,116 @@ var AgentHarness = class _AgentHarness {
   get frontmatter() {
     return this.parsedAgent?.frontmatter;
   }
+  getToolResultArchive(conversationId) {
+    const archive = this.archivedToolResultsByConversation.get(conversationId);
+    return archive ? { ...archive } : {};
+  }
+  seedToolResultArchive(conversationId, parameters) {
+    const seeded = readArchiveFromParameters(parameters);
+    const existing = this.archivedToolResultsByConversation.get(conversationId) ?? {};
+    const merged = { ...existing, ...seeded };
+    this.archivedToolResultsByConversation.set(conversationId, merged);
+    return merged;
+  }
+  truncateHistoricalToolResults(messages, conversationId) {
+    let latestRunId;
+    for (let i = messages.length - 1; i >= 0; i -= 1) {
+      const msg = messages[i];
+      const meta = msg.metadata;
+      const runId = typeof meta?.runId === "string" ? meta.runId : void 0;
+      if (runId) {
+        latestRunId = runId;
+        break;
+      }
+    }
+    if (!latestRunId) {
+      return { changed: false, truncatedCount: 0, archivedCount: 0, omittedChars: 0 };
+    }
+    const archive = this.archivedToolResultsByConversation.get(conversationId) ?? {};
+    this.archivedToolResultsByConversation.set(conversationId, archive);
+    let changed = false;
+    let truncatedCount = 0;
+    let archivedCount = 0;
+    let omittedChars = 0;
+    for (const msg of messages) {
+      if (msg.role !== "tool" || typeof msg.content !== "string") continue;
+      const meta = msg.metadata;
+      const runId = typeof meta?.runId === "string" ? meta.runId : void 0;
+      if (runId === latestRunId) continue;
+      let parsed;
+      try {
+        parsed = JSON.parse(msg.content);
+      } catch {
+        continue;
+      }
+      if (!Array.isArray(parsed)) continue;
+      let rowChanged = false;
+      const nextRows = parsed.map((row) => {
+        if (!isToolResultRow(row)) return row;
+        if (row.content.startsWith(TOOL_RESULT_TRUNCATED_PREFIX)) return row;
+        if (this.shouldPreserveSkillToolResult(row)) return row;
+        const toolResultId = row.tool_use_id;
+        if (!archive[toolResultId]) {
+          archive[toolResultId] = {
+            toolResultId,
+            conversationId,
+            toolName: row.tool_name,
+            toolCallId: row.tool_use_id,
+            createdAt: now(),
+            sizeBytes: Buffer.byteLength(row.content, "utf8"),
+            payload: row.content
+          };
+          archivedCount += 1;
+        }
+        const omitted = Math.max(0, row.content.length - TOOL_RESULT_PREVIEW_CHARS);
+        omittedChars += omitted;
+        truncatedCount += 1;
+        rowChanged = true;
+        return {
+          ...row,
+          content: makeTruncatedToolResultNotice(toolResultId, row.tool_name, row.content)
+        };
+      });
+      if (rowChanged) {
+        msg.content = JSON.stringify(nextRows);
+        if (msg.metadata && typeof msg.metadata === "object") {
+          const meta2 = msg.metadata;
+          if ("_richToolResults" in meta2) {
+            delete meta2._richToolResults;
+          }
+        }
+        changed = true;
+      }
+    }
+    return { changed, truncatedCount, archivedCount, omittedChars };
+  }
+  shouldPreserveSkillToolResult(row) {
+    if (row.tool_name.startsWith("todo_")) {
+      return true;
+    }
+    if (row.tool_name !== "activate_skill" && row.tool_name !== "deactivate_skill") {
+      return false;
+    }
+    const content = row.content.trim();
+    if (content.startsWith("Tool error:")) {
+      return false;
+    }
+    try {
+      const parsed = JSON.parse(content);
+      const skill = typeof parsed.skill === "string" ? parsed.skill : void 0;
+      if (skill && this.activeSkillNames.has(skill)) {
+        return true;
+      }
+      const activeSkills = Array.isArray(parsed.activeSkills) ? parsed.activeSkills.filter((v) => typeof v === "string") : [];
+      for (const name of activeSkills) {
+        if (this.activeSkillNames.has(name)) {
+          return true;
+        }
+      }
+    } catch {
+    }
+    return false;
+  }
   async getTodos(conversationId) {
     if (!this.todoStore) return [];
     return this.todoStore.get(conversationId);
@@ -6336,12 +6573,31 @@ var AgentHarness = class _AgentHarness {
     let agent = this.parsedAgent;
     const runId = `run_${randomUUID3()}`;
     const start = now();
-    const maxSteps = agent.frontmatter.limits?.maxSteps ?? 50;
+    const maxSteps = agent.frontmatter.limits?.maxSteps ?? 20;
     const configuredTimeout = agent.frontmatter.limits?.timeout;
     const timeoutMs = this.environment === "development" && configuredTimeout == null ? 0 : (configuredTimeout ?? 300) * 1e3;
     const platformMaxDurationSec = Number(process.env.PONCHO_MAX_DURATION) || 0;
     const softDeadlineMs = input.disableSoftDeadline || platformMaxDurationSec <= 0 ? 0 : platformMaxDurationSec * 800;
     const messages = [...input.messages ?? []];
+    const conversationId = input.conversationId ?? "__default__";
+    this.seedToolResultArchive(conversationId, input.parameters);
+    const truncationSummary = this.truncateHistoricalToolResults(messages, conversationId);
+    if (truncationSummary.changed) {
+      console.info(
+        `[poncho][cost] Truncated ${truncationSummary.truncatedCount} historical tool result(s) (archived_new=${truncationSummary.archivedCount}, omitted_chars=${truncationSummary.omittedChars}) for conversation="${conversationId}"`
+      );
+    }
+    const hasFullToolResults = hasUntruncatedToolResults(messages);
+    const enablePromptCache = !hasFullToolResults;
+    if (!enablePromptCache) {
+      console.info(
+        `[poncho][cost] Prompt cache write disabled for run "${runId}" (untruncated tool results present in history).`
+      );
+    } else {
+      console.info(
+        `[poncho][cost] Prompt cache write enabled for run "${runId}" (history has no untruncated tool results).`
+      );
+    }
     const inputMessageCount = messages.length;
     const events = [];
     const renderCurrentAgentPrompt = () => renderAgentPrompt(this.parsedAgent, {
@@ -6421,7 +6677,6 @@ ${this.skillFingerprint}`;
     const browserEventQueue = [];
     const browserCleanups = [];
     const browserSession = this._browserSession;
-    const conversationId = input.conversationId ?? "__default__";
     if (browserSession) {
       browserCleanups.push(
         browserSession.onFrame(conversationId, (frame) => {
@@ -6488,6 +6743,7 @@ ${this.skillFingerprint}`;
     let totalInputTokens = 0;
     let totalOutputTokens = 0;
     let totalCachedTokens = 0;
+    let totalCacheWriteTokens = 0;
     let transientStepRetryCount = 0;
     let latestContextTokens = 0;
     let toolOutputEstimateSinceModel = 0;
@@ -6516,7 +6772,12 @@ ${this.skillFingerprint}`;
             status: "completed",
             response: responseText,
             steps: step - 1,
-            tokens: { input: totalInputTokens, output: totalOutputTokens, cached: totalCachedTokens },
+            tokens: {
+              input: totalInputTokens,
+              output: totalOutputTokens,
+              cached: totalCachedTokens,
+              cacheWrite: totalCacheWriteTokens
+            },
             duration: now() - start,
             continuation: true,
             continuationMessages: [...messages],
@@ -6529,7 +6790,6 @@ ${this.skillFingerprint}`;
         }
         const stepStart = now();
         yield pushEvent({ type: "step:started", step });
-        yield pushEvent({ type: "model:request", tokens: 0 });
         const dispatcherTools = this.dispatcher.list();
         const exposedToolNames = /* @__PURE__ */ new Map();
         const usedProviderToolNames = /* @__PURE__ */ new Set();
@@ -6548,6 +6808,15 @@ ${this.skillFingerprint}`;
             inputSchema: jsonSchemaToZod(tool.inputSchema)
           };
         }
+        const toolDefsJsonForEstimate = JSON.stringify(
+          dispatcherTools.map((t) => ({
+            name: t.name,
+            description: t.description,
+            inputSchema: t.inputSchema
+          }))
+        );
+        const requestTokenEstimate = estimateTotalTokens(integrityPrompt, messages, toolDefsJsonForEstimate);
+        yield pushEvent({ type: "model:request", tokens: requestTokenEstimate });
         const convertMessage = async (msg) => {
           if (msg.role === "tool") {
             const meta = msg.metadata;
@@ -6737,15 +7006,8 @@ ${textContent}` };
         }
         const modelInstance = this.modelProvider(modelName);
         const compactionConfig = resolveCompactionConfig(agent.frontmatter.compaction);
-        if (compactionConfig.enabled && step === 1) {
-          const toolDefsJson = JSON.stringify(
-            dispatcherTools.map((t) => ({
-              name: t.name,
-              description: t.description,
-              inputSchema: t.inputSchema
-            }))
-          );
-          const estimated = estimateTotalTokens(integrityPrompt, messages, toolDefsJson);
+        if (compactionConfig.enabled && (step === 1 || step % COMPACTION_CHECK_INTERVAL_STEPS === 0)) {
+          const estimated = estimateTotalTokens(integrityPrompt, messages, toolDefsJsonForEstimate);
           const lastReportedInput = totalInputTokens > 0 ? totalInputTokens : 0;
           const effectiveTokens = Math.max(estimated, lastReportedInput);
           if (effectiveTokens > compactionConfig.trigger * contextWindow) {
@@ -6758,11 +7020,14 @@ ${textContent}` };
             if (compactResult.compacted) {
               messages.length = 0;
               messages.push(...compactResult.messages);
-              const emittedMessages = [...compactResult.messages];
-              if (emittedMessages.length > 0 && emittedMessages[emittedMessages.length - 1].role === "user") {
-                emittedMessages.pop();
+              let emittedMessages;
+              if (step === 1) {
+                emittedMessages = [...compactResult.messages];
+                if (emittedMessages.length > 0 && emittedMessages[emittedMessages.length - 1].role === "user") {
+                  emittedMessages.pop();
+                }
               }
-              const tokensAfterCompaction = estimateTotalTokens(integrityPrompt, messages, toolDefsJson);
+              const tokensAfterCompaction = estimateTotalTokens(integrityPrompt, messages, toolDefsJsonForEstimate);
               latestContextTokens = tokensAfterCompaction;
               toolOutputEstimateSinceModel = 0;
               yield pushEvent({
@@ -6789,7 +7054,7 @@ ${textContent}` };
         const coreMessages = cachedCoreMessages;
         const temperature = agent.frontmatter.model?.temperature ?? 0.2;
         const maxTokens = agent.frontmatter.model?.maxTokens;
-        const cachedMessages = addPromptCacheBreakpoints(coreMessages, modelInstance);
+        const cachedMessages = enablePromptCache ? addPromptCacheBreakpoints(coreMessages, modelInstance) : coreMessages;
         const telemetryEnabled = this.loadedConfig?.telemetry?.enabled !== false;
         const result = await streamText({
           model: modelInstance,
@@ -6907,7 +7172,12 @@ ${textContent}` };
             status: "completed",
             response: responseText + fullText,
             steps: step,
-            tokens: { input: totalInputTokens, output: totalOutputTokens, cached: totalCachedTokens },
+            tokens: {
+              input: totalInputTokens,
+              output: totalOutputTokens,
+              cached: totalCachedTokens,
+              cacheWrite: totalCacheWriteTokens
+            },
             duration: now() - start,
             continuation: true,
             continuationMessages: [...messages],
@@ -6935,7 +7205,12 @@ ${textContent}` };
             status: "completed",
             response: responseText + fullText,
             steps: step,
-            tokens: { input: totalInputTokens, output: totalOutputTokens, cached: totalCachedTokens },
+            tokens: {
+              input: totalInputTokens,
+              output: totalOutputTokens,
+              cached: totalCachedTokens,
+              cacheWrite: totalCacheWriteTokens
+            },
             duration: now() - start,
             continuation: true,
             continuationMessages: [...messages],
@@ -6975,11 +7250,14 @@ ${textContent}` };
         const fullResult = await result.response;
         const usage = await result.usage;
         const toolCallsResult = await result.toolCalls;
-        const stepCachedTokens = usage.inputTokenDetails?.cacheReadTokens ?? 0;
+        const details = usage.inputTokenDetails ?? {};
+        const stepCachedTokens = typeof details.cacheReadTokens === "number" ? details.cacheReadTokens : 0;
+        const stepCacheWriteTokens = typeof details.cacheWriteTokens === "number" ? details.cacheWriteTokens : typeof details.cacheCreationTokens === "number" ? details.cacheCreationTokens : typeof details.cacheCreationInputTokens === "number" ? details.cacheCreationInputTokens : 0;
         const stepInputTokens = usage.inputTokens ?? 0;
         totalInputTokens += stepInputTokens;
         totalOutputTokens += usage.outputTokens ?? 0;
         totalCachedTokens += stepCachedTokens;
+        totalCacheWriteTokens += stepCacheWriteTokens;
         latestContextTokens = stepInputTokens;
         toolOutputEstimateSinceModel = 0;
         yield pushEvent({
@@ -6987,9 +7265,13 @@ ${textContent}` };
           usage: {
             input: stepInputTokens,
             output: usage.outputTokens ?? 0,
-            cached: stepCachedTokens
+            cached: stepCachedTokens,
+            cacheWrite: stepCacheWriteTokens
           }
         });
+        console.info(
+          `[poncho][cost] model="${modelName}" step=${step} input=${stepInputTokens} output=${usage.outputTokens ?? 0} cached=${stepCachedTokens} cacheWrite=${stepCacheWriteTokens} totals(input=${totalInputTokens}, output=${totalOutputTokens}, cached=${totalCachedTokens}, cacheWrite=${totalCacheWriteTokens})`
+        );
         const toolCalls = toolCallsResult.map((tc) => ({
           id: tc.toolCallId,
           name: tc.toolName,
@@ -7036,7 +7318,8 @@ ${textContent}` };
             tokens: {
               input: totalInputTokens,
               output: totalOutputTokens,
-              cached: totalCachedTokens
+              cached: totalCachedTokens,
+              cacheWrite: totalCacheWriteTokens
             },
             duration: now() - start,
             contextTokens: latestContextTokens + toolOutputEstimateSinceModel,
@@ -7193,7 +7476,12 @@ ${textContent}` };
             status: "completed",
             response: responseText + fullText,
             steps: step,
-            tokens: { input: totalInputTokens, output: totalOutputTokens, cached: totalCachedTokens },
+            tokens: {
+              input: totalInputTokens,
+              output: totalOutputTokens,
+              cached: totalCachedTokens,
+              cacheWrite: totalCacheWriteTokens
+            },
             duration: now() - start,
             continuation: true,
             continuationMessages: [...messages],
@@ -7224,6 +7512,20 @@ ${textContent}` };
               tool_name: result2.tool,
               content: `Tool error: ${result2.error}`
             });
+            {
+              const archive = this.archivedToolResultsByConversation.get(conversationId);
+              if (archive) {
+                archive[result2.callId] = {
+                  toolResultId: result2.callId,
+                  conversationId,
+                  toolName: result2.tool,
+                  toolCallId: result2.callId,
+                  createdAt: now(),
+                  sizeBytes: Buffer.byteLength(`Tool error: ${result2.error}`, "utf8"),
+                  payload: `Tool error: ${result2.error}`
+                };
+              }
+            }
             richToolResults.push({
               type: "tool-result",
               toolCallId: result2.callId,
@@ -7249,6 +7551,21 @@ ${textContent}` };
               tool_name: result2.tool,
               content: JSON.stringify(strippedOutput ?? null)
             });
+            {
+              const archive = this.archivedToolResultsByConversation.get(conversationId);
+              if (archive) {
+                const payload = JSON.stringify(result2.output ?? null);
+                archive[result2.callId] = {
+                  toolResultId: result2.callId,
+                  conversationId,
+                  toolName: result2.tool,
+                  toolCallId: result2.callId,
+                  createdAt: now(),
+                  sizeBytes: Buffer.byteLength(payload, "utf8"),
+                  payload
+                };
+              }
+            }
             if (mediaItems.length > 0) {
               richToolResults.push({
                 type: "tool-result",
@@ -7283,9 +7600,15 @@ ${textContent}` };
         messages.push({
           role: "assistant",
           content: assistantContent,
-          metadata: { timestamp: now(), id: randomUUID3(), step }
+          metadata: { timestamp: now(), id: randomUUID3(), step, runId }
         });
-        const toolMsgMeta = { timestamp: now(), id: randomUUID3(), step, _richToolResults: richToolResults };
+        const toolMsgMeta = {
+          timestamp: now(),
+          id: randomUUID3(),
+          step,
+          runId,
+          _richToolResults: richToolResults
+        };
         messages.push({
           role: "tool",
           content: JSON.stringify(toolResultsForModel),
@@ -7296,7 +7619,12 @@ ${textContent}` };
             status: "completed",
             response: responseText + fullText,
             steps: step,
-            tokens: { input: totalInputTokens, output: totalOutputTokens, cached: totalCachedTokens },
+            tokens: {
+              input: totalInputTokens,
+              output: totalOutputTokens,
+              cached: totalCachedTokens,
+              cacheWrite: totalCacheWriteTokens
+            },
             duration: now() - start,
             continuation: true,
             continuationMessages: [...messages],
@@ -7357,7 +7685,12 @@ ${this.skillFingerprint}`;
         status: "completed",
         response: responseText,
         steps: maxSteps,
-        tokens: { input: totalInputTokens, output: totalOutputTokens, cached: totalCachedTokens },
+        tokens: {
+          input: totalInputTokens,
+          output: totalOutputTokens,
+          cached: totalCachedTokens,
+          cacheWrite: totalCacheWriteTokens
+        },
         duration: now() - start,
         continuation: true,
         continuationMessages: [...messages],
@@ -8677,7 +9010,7 @@ var createConversationStore = (config, options) => {
 };
 
 // src/index.ts
-import { defineTool as defineTool7 } from "@poncho-ai/sdk";
+import { defineTool as defineTool8 } from "@poncho-ai/sdk";
 export {
   AgentHarness,
   InMemoryConversationStore,
@@ -8710,7 +9043,7 @@ export {
   createSubagentTools,
   createUploadStore,
   createWriteTool,
-  defineTool7 as defineTool,
+  defineTool8 as defineTool,
   deleteOpenAICodexSession,
   deriveUploadKey,
   ensureAgentIdentity,