openpond-sdk 0.12.3 → 0.12.4

package/dist/experiment-evaluation-schedules.js

@@ -8,33 +8,34 @@ import {
   scheduledExperimentHash,
   sealExperimentEvaluationSchedule,
   verifyExperimentEvaluationSchedule
-} from "./chunk-G6Y6CQYV.js";
-import "./chunk-YCQVVFE5.js";
-import "./chunk-K5WRVSOL.js";
-import "./chunk-UPS52OPZ.js";
-import "./chunk-HKWX624G.js";
-import "./chunk-YVUIHH6Z.js";
-import "./chunk-FY7NHCEH.js";
-import "./chunk-36WJTS5V.js";
-import "./chunk-BFR3VC25.js";
-import "./chunk-K3JCP4FO.js";
-import "./chunk-ROBAX5TN.js";
-import "./chunk-CWXM57QR.js";
-import "./chunk-FHCSDXIA.js";
-import "./chunk-2U54EOXJ.js";
-import "./chunk-TMUAKJWX.js";
-import "./chunk-MLPPSCDQ.js";
-import "./chunk-G53MA5XX.js";
-import "./chunk-DUKW3R7O.js";
-import "./chunk-KAE7AMYE.js";
-import "./chunk-CTVKI3BV.js";
-import "./chunk-T4APHDES.js";
-import "./chunk-QTJQIHDZ.js";
-import "./chunk-ZJ3H3VS7.js";
-import "./chunk-CK3ONRRC.js";
-import "./chunk-TCHAWJMB.js";
-import "./chunk-3KXDH6OB.js";
-import "./chunk-COPVBJPL.js";
+} from "./chunk-K5LWSWV7.js";
+import "./chunk-3B5KSH3V.js";
+import "./chunk-I54WQBXL.js";
+import "./chunk-DF2QXVZ3.js";
+import "./chunk-BVMR5OFW.js";
+import "./chunk-GPQLBVQC.js";
+import "./chunk-CMWZZZGW.js";
+import "./chunk-B5XTGEBE.js";
+import "./chunk-LR5CTOC5.js";
+import "./chunk-Z252BNAZ.js";
+import "./chunk-WQFU635Q.js";
+import "./chunk-GAIDBPFY.js";
+import "./chunk-MO4HAJZG.js";
+import "./chunk-TK3DXH7A.js";
+import "./chunk-RRKRA237.js";
+import "./chunk-ZHQK3MOM.js";
+import "./chunk-WMIKJGXW.js";
+import "./chunk-Q4TKI2MV.js";
+import "./chunk-H3367ISL.js";
+import "./chunk-53LLNDXF.js";
+import "./chunk-O6FKHAWN.js";
+import "./chunk-QRS7Y3EM.js";
+import "./chunk-VN5V2E2S.js";
+import "./chunk-U7XWIOSB.js";
+import "./chunk-HGXVOFCL.js";
+import "./chunk-CL3IOOVL.js";
+import "./chunk-H3JZ3P5T.js";
+import "./chunk-YLJ4XMA6.js";
 export {
   ExperimentEvaluationFireSchema,
   ExperimentEvaluationScheduleCommandSchema,

package/dist/experiment-evaluation-schedules.js.map

@@ -1,7 +1 @@
-{
-  "version": 3,
-  "sources": [],
-  "sourcesContent": [],
-  "mappings": "",
-  "names": []
-}
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/experiment-improvements.js

@@ -5,7 +5,7 @@ import {
   HostedProfileCandidateReceiptSchema,
   verifyHostedProfileCandidateExport,
   verifyHostedProfileCandidateReceipt
-} from "./chunk-77N3QEIL.js";
+} from "./chunk-OR3TMIJN.js";
 import {
   ExperimentImprovementCommandSchema,
   ExperimentImprovementHumanOptionsSchema,
@@ -19,16 +19,17 @@ import {
   sealExperimentImprovementState,
   verifyExperimentImprovementHumanOptions,
   verifyExperimentImprovementState
-} from "./chunk-TSMGMA6R.js";
-import "./chunk-DUKW3R7O.js";
-import "./chunk-CTVKI3BV.js";
-import "./chunk-T4APHDES.js";
-import "./chunk-QTJQIHDZ.js";
-import "./chunk-ZJ3H3VS7.js";
-import "./chunk-CK3ONRRC.js";
-import "./chunk-TCHAWJMB.js";
-import "./chunk-3KXDH6OB.js";
-import "./chunk-COPVBJPL.js";
+} from "./chunk-XON75UFN.js";
+import "./chunk-Q4TKI2MV.js";
+import "./chunk-53LLNDXF.js";
+import "./chunk-O6FKHAWN.js";
+import "./chunk-QRS7Y3EM.js";
+import "./chunk-VN5V2E2S.js";
+import "./chunk-U7XWIOSB.js";
+import "./chunk-HGXVOFCL.js";
+import "./chunk-CL3IOOVL.js";
+import "./chunk-H3JZ3P5T.js";
+import "./chunk-YLJ4XMA6.js";
 export {
   ExperimentImprovementClient,
   ExperimentImprovementCommandSchema,

package/dist/experiment-improvements.js.map

@@ -1,7 +1 @@
-{
-  "version": 3,
-  "sources": [],
-  "sourcesContent": [],
-  "mappings": "",
-  "names": []
-}
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/experiments.js

@@ -27,31 +27,31 @@ import {
   verifyExperimentRunDetails,
   verifyExternalDatasetPackage,
   verifyRunDiagnostics
-} from "./chunk-K5WRVSOL.js";
-import "./chunk-UPS52OPZ.js";
-import "./chunk-HKWX624G.js";
-import "./chunk-YVUIHH6Z.js";
-import "./chunk-FY7NHCEH.js";
-import "./chunk-36WJTS5V.js";
-import "./chunk-BFR3VC25.js";
-import "./chunk-K3JCP4FO.js";
-import "./chunk-ROBAX5TN.js";
-import "./chunk-CWXM57QR.js";
-import "./chunk-FHCSDXIA.js";
-import "./chunk-2U54EOXJ.js";
-import "./chunk-TMUAKJWX.js";
-import "./chunk-MLPPSCDQ.js";
+} from "./chunk-I54WQBXL.js";
+import "./chunk-DF2QXVZ3.js";
+import "./chunk-BVMR5OFW.js";
+import "./chunk-GPQLBVQC.js";
+import "./chunk-CMWZZZGW.js";
+import "./chunk-B5XTGEBE.js";
+import "./chunk-LR5CTOC5.js";
+import "./chunk-Z252BNAZ.js";
+import "./chunk-WQFU635Q.js";
+import "./chunk-GAIDBPFY.js";
+import "./chunk-MO4HAJZG.js";
+import "./chunk-TK3DXH7A.js";
+import "./chunk-RRKRA237.js";
+import "./chunk-ZHQK3MOM.js";
 import {
   RunExperimentSchema
-} from "./chunk-G53MA5XX.js";
-import "./chunk-DUKW3R7O.js";
-import "./chunk-KAE7AMYE.js";
+} from "./chunk-WMIKJGXW.js";
+import "./chunk-Q4TKI2MV.js";
+import "./chunk-H3367ISL.js";
 import {
   ExperimentScoringPassSchema,
   ExperimentScoringRequestSchema,
   verifyExperimentScoringPass
-} from "./chunk-CTVKI3BV.js";
-import "./chunk-T4APHDES.js";
+} from "./chunk-53LLNDXF.js";
+import "./chunk-O6FKHAWN.js";
 import {
   ExperimentDefinitionRefSchema,
   ExperimentDefinitionSchema,
@@ -68,22 +68,23 @@ import {
   experimentDefinitionRef,
   mapExperimentGradingFields,
   verifyExperimentDefinition
-} from "./chunk-QTJQIHDZ.js";
+} from "./chunk-QRS7Y3EM.js";
 import {
   HarnessExperimentPolicySchema,
   HarnessExperimentSamplingSchema,
   resolveHarnessExperimentSampling
-} from "./chunk-ZJ3H3VS7.js";
+} from "./chunk-VN5V2E2S.js";
 import {
   ClaudeCodeControlSchema,
   ClaudeCodeRuntimeSchema,
   ClaudeProcessEvidenceSchema,
   CodexExperimentRuntimeSchema,
   ExperimentAttemptGradeSchema
-} from "./chunk-CK3ONRRC.js";
-import "./chunk-TCHAWJMB.js";
-import "./chunk-3KXDH6OB.js";
-import "./chunk-COPVBJPL.js";
+} from "./chunk-U7XWIOSB.js";
+import "./chunk-HGXVOFCL.js";
+import "./chunk-CL3IOOVL.js";
+import "./chunk-H3JZ3P5T.js";
+import "./chunk-YLJ4XMA6.js";
 export {
   ClaudeCodeControlSchema,
   ClaudeCodeRuntimeSchema,

package/dist/experiments.js.map

@@ -1,7 +1 @@
-{
-  "version": 3,
-  "sources": [],
-  "sourcesContent": [],
-  "mappings": "",
-  "names": []
-}
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/human-review.js

@@ -1,8 +1,9 @@
 import {
   HumanReviewClientError,
   OpenPondHumanReviewClient
-} from "./chunk-AG5DDL7T.js";
-import "./chunk-COPVBJPL.js";
+} from "./chunk-DH6O4JFB.js";
+import "./chunk-H3JZ3P5T.js";
+import "./chunk-YLJ4XMA6.js";
 export {
   HumanReviewClientError,
   OpenPondHumanReviewClient

package/dist/human-review.js.map

@@ -1,7 +1 @@
-{
-  "version": 3,
-  "sources": [],
-  "sourcesContent": [],
-  "mappings": "",
-  "names": []
-}
+{"version":3,"sources":[],"sourcesContent":[],"mappings":"","names":[]}

package/dist/index.js

@@ -1,13 +1,13 @@
 import {
   OpenPondLearningClient,
   OpenPondLearningError
-} from "./chunk-3LP7DNLQ.js";
+} from "./chunk-H4NTJFEE.js";
 import {
   OpenPondProfileActionsClient
-} from "./chunk-2DMYAJLS.js";
+} from "./chunk-3DWPLYFJ.js";
 import {
   OpenPondWorkflowsClient
-} from "./chunk-MEKWESBF.js";
+} from "./chunk-FBZTDDRC.js";
 import {
   DEFAULT_REFINER_REVIEW_PROFILE,
   REFINER_CORE_VERSION,
@@ -20,19 +20,19 @@ import {
   defineReviewProfile,
   refinerProfilePrompt,
   serializeReviewProfile
-} from "./chunk-O4GE6VTR.js";
+} from "./chunk-PWW5QKZS.js";
 import {
   OpenPondProjectActionsClient
-} from "./chunk-IWT6JUZD.js";
+} from "./chunk-L6EJX2VQ.js";
 import {
   OpenPondApiError,
   apiFetch,
   readApiJson
-} from "./chunk-CEZEVILU.js";
+} from "./chunk-HJ2AIAKX.js";
 import {
   HumanReviewClientError,
   OpenPondHumanReviewClient
-} from "./chunk-AG5DDL7T.js";
+} from "./chunk-DH6O4JFB.js";
 import {
   AgentImportPreviewSchema,
   AgentImportReceiptSchema,
@@ -66,7 +66,7 @@ import {
   hashCollectorEligibleBoundaries,
   summarizeCollectorCoverage,
   verifyConnectedRecordedExecution
-} from "./chunk-VY5AFE7A.js";
+} from "./chunk-24N6A6SF.js";
 import {
   AgentImportCommitRequestSchema,
   AgentImportFileSchema,
@@ -83,7 +83,7 @@ import {
   ConnectedStartingAssetsReceiptSchema,
   ConnectedStartingAssetsRequestSchema,
   verifyConnectedStartingAssetsReceipt
-} from "./chunk-QEIT5VXJ.js";
+} from "./chunk-EFUT3XKW.js";
 import {
   CandidateEvaluationControlSchema,
   CandidateEvaluationOptionSchema,
@@ -91,7 +91,7 @@ import {
   CandidateEvaluationPreparationSchema,
   CandidateEvaluationRequestSchema,
   createCandidateEvaluationClient
-} from "./chunk-BSK3VLM3.js";
+} from "./chunk-ETX25FTB.js";
 import {
   ExperimentImprovementClient,
   HostedProfileCandidateCommandSchema,
@@ -99,7 +99,7 @@ import {
   HostedProfileCandidateReceiptSchema,
   verifyHostedProfileCandidateExport,
   verifyHostedProfileCandidateReceipt
-} from "./chunk-77N3QEIL.js";
+} from "./chunk-OR3TMIJN.js";
 import {
   ExperimentEvaluationFireSchema,
   ExperimentEvaluationScheduleCommandSchema,
@@ -110,7 +110,7 @@ import {
   scheduledExperimentHash,
   sealExperimentEvaluationSchedule,
   verifyExperimentEvaluationSchedule
-} from "./chunk-G6Y6CQYV.js";
+} from "./chunk-K5LWSWV7.js";
 import {
   AdvancedRefinerEvaluationAccountingSchema,
   AdvancedRefinerEvaluationCommandSchema,
@@ -125,7 +125,7 @@ import {
   sealAdvancedRefinerEvaluationPin,
   sealAdvancedRefinerReviewReceipt,
   verifyAdvancedRefinerEvaluationPin
-} from "./chunk-YUKV7RCD.js";
+} from "./chunk-VHGC44HA.js";
 import {
   ExperimentImprovementCommandSchema,
   ExperimentImprovementHumanOptionsSchema,
@@ -139,37 +139,42 @@ import {
   sealExperimentImprovementState,
   verifyExperimentImprovementHumanOptions,
   verifyExperimentImprovementState
-} from "./chunk-TSMGMA6R.js";
-import "./chunk-YCQVVFE5.js";
+} from "./chunk-XON75UFN.js";
+import "./chunk-3B5KSH3V.js";
 import {
   OpenPondExperimentsClient
-} from "./chunk-K5WRVSOL.js";
-import "./chunk-UPS52OPZ.js";
-import "./chunk-HKWX624G.js";
-import "./chunk-YVUIHH6Z.js";
-import "./chunk-FY7NHCEH.js";
-import "./chunk-36WJTS5V.js";
-import "./chunk-BFR3VC25.js";
-import "./chunk-K3JCP4FO.js";
-import "./chunk-ROBAX5TN.js";
-import "./chunk-CWXM57QR.js";
+} from "./chunk-I54WQBXL.js";
+import "./chunk-DF2QXVZ3.js";
+import "./chunk-BVMR5OFW.js";
+import "./chunk-GPQLBVQC.js";
+import "./chunk-CMWZZZGW.js";
+import "./chunk-B5XTGEBE.js";
+import "./chunk-LR5CTOC5.js";
+import "./chunk-Z252BNAZ.js";
+import "./chunk-WQFU635Q.js";
+import "./chunk-GAIDBPFY.js";
 import {
-  ConnectedDatasetSourceRefSchema
-} from "./chunk-FHCSDXIA.js";
-import "./chunk-2U54EOXJ.js";
-import "./chunk-TMUAKJWX.js";
-import "./chunk-MLPPSCDQ.js";
-import "./chunk-G53MA5XX.js";
-import "./chunk-DUKW3R7O.js";
-import "./chunk-KAE7AMYE.js";
-import "./chunk-CTVKI3BV.js";
-import "./chunk-T4APHDES.js";
-import "./chunk-QTJQIHDZ.js";
-import "./chunk-ZJ3H3VS7.js";
-import "./chunk-CK3ONRRC.js";
-import "./chunk-TCHAWJMB.js";
-import "./chunk-3KXDH6OB.js";
-import "./chunk-COPVBJPL.js";
+  ConnectedDatasetSourceRefSchema,
+  serializeStructuredOutput,
+  structuredOutputTools,
+  structuredToolOutput,
+  validateOutputBytes
+} from "./chunk-MO4HAJZG.js";
+import "./chunk-TK3DXH7A.js";
+import "./chunk-RRKRA237.js";
+import "./chunk-ZHQK3MOM.js";
+import "./chunk-WMIKJGXW.js";
+import "./chunk-Q4TKI2MV.js";
+import "./chunk-H3367ISL.js";
+import "./chunk-53LLNDXF.js";
+import "./chunk-O6FKHAWN.js";
+import "./chunk-QRS7Y3EM.js";
+import "./chunk-VN5V2E2S.js";
+import "./chunk-U7XWIOSB.js";
+import "./chunk-HGXVOFCL.js";
+import "./chunk-CL3IOOVL.js";
+import "./chunk-H3JZ3P5T.js";
+import "./chunk-YLJ4XMA6.js";
 
 // ../cloud/dist/sandbox/client-handles.js
 function createSandboxRuntimeNamespace(client) {
@@ -1981,6 +1986,9 @@ async function* streamHostedChatTurn(options) {
   }
 }
 
+// src/work.ts
+import { randomUUID as randomUUID2 } from "node:crypto";
+
 // src/work-sandbox.ts
 function workSandbox(client, custom) {
   return {
@@ -2082,6 +2090,7 @@ var OpenPondWorkClient = class {
     const prompt = input.prompt.trim();
     if (!prompt) throw new Error("Work prompt is required");
     input.signal?.throwIfAborted();
+    const submissionTools = structuredOutputTools(input.requiredOutputs ?? []);
     const emit = async (event) => input.onEvent?.(event);
     const cleanupPolicy = input.cleanup ?? "keep";
     const lifecycle = initialLifecycle(cleanupPolicy);
@@ -2106,7 +2115,7 @@ var OpenPondWorkClient = class {
           content: systemPrompt(sandbox.id, (input.inputs?.length ?? 0) > 0)
         },
         ...(input.history ?? []).map((message) => ({ ...message })),
-        { role: "user", content: prompt }
+        { role: "user", content: prompt + (input.requiredOutputs?.length ? "\nRequired outputs under outputs/: " + JSON.stringify(input.requiredOutputs) : "") }
       ];
       for (let step = 1; step <= maxSteps; step += 1) {
         input.signal?.throwIfAborted();
@@ -2116,7 +2125,7 @@ var OpenPondWorkClient = class {
           token: this.#apiKey,
           model: input.model?.trim() || this.#defaultModel,
           messages,
-          tools: WORK_TOOLS,
+          tools: [...WORK_TOOLS, ...submissionTools],
           toolChoice: "auto",
           signal: input.signal,
           metadata: {
@@ -2139,7 +2148,13 @@ var OpenPondWorkClient = class {
           await emit({ type: "assistant", text: assistantText });
         }
         if (toolCalls.length === 0) {
-          const outputs = await this.#collectOutputs(sandbox.id, outputBaseline);
+          let outputs;
+          try {
+            outputs = await this.#collectOutputs(sandbox.id, outputBaseline, input.requiredOutputs);
+          } catch (error) {
+            messages.push({ role: "user", content: "Output acceptance failed. Correct the required outputs and submit again: " + errorMessage(error) });
+            continue;
+          }
           for (const output of outputs) await emit({ type: "output", output });
           finalizationAttempted = true;
           await this.#finalizeSandbox(sandbox.id, outputs, input, lifecycle, emit);
@@ -2160,7 +2175,8 @@ var OpenPondWorkClient = class {
             toolCall,
             index,
             input.timeoutSeconds,
-            emit
+            emit,
+            input.requiredOutputs
           );
           messages.push(result);
         }
@@ -2169,7 +2185,7 @@ var OpenPondWorkClient = class {
     } catch (error) {
       if (sandbox && outputBaseline && !finalizationAttempted) {
         try {
-          const outputs = await this.#collectOutputs(sandbox.id, outputBaseline);
+          const outputs = await this.#collectOutputs(sandbox.id, outputBaseline, input.requiredOutputs);
           for (const output of outputs) await emit({ type: "output", output });
           await this.#finalizeSandbox(sandbox.id, outputs, input, lifecycle, emit);
         } catch (finalizationError) {
@@ -2241,7 +2257,16 @@ var OpenPondWorkClient = class {
       for (const output of outputs) {
         await emit({ type: "persistence", output, status: "started" });
         try {
-          const download = lazyOutputDownload(this.#sandboxes, sandboxId, output);
+          const originalDownload = lazyOutputDownload(this.#sandboxes, sandboxId, output);
+          const download = async () => {
+            const result = await originalDownload();
+            const required = input.requiredOutputs?.find((item) => normalizedOutputPath(output.path) === `${WORK_OUTPUT_DIRECTORY}/${item.path}`);
+            if (required) {
+              const validation = validateOutputBytes(required, Buffer.from(result.file.contentsBase64, "base64"));
+              if (output.validation && validation?.artifactHash !== output.validation.artifactHash) throw new Error("work_saved_output_changed");
+            }
+            return result;
+          };
           await input.persistOutput({ output, download });
           lifecycle.persistence.persistedCount += 1;
           await emit({ type: "persistence", output, status: "succeeded" });
@@ -2322,9 +2347,17 @@ var OpenPondWorkClient = class {
     });
     return outputSignatures(await this.#listOutputFiles(sandboxId));
   }
-  async #collectOutputs(sandboxId, baseline) {
+  async #collectOutputs(sandboxId, baseline, requiredOutputs) {
     const files = await this.#listOutputFiles(sandboxId);
-    return files.filter((file) => baseline.get(file.path) !== outputSignature(file)).sort((left, right) => left.path.localeCompare(right.path)).slice(0, MAX_WORK_OUTPUTS).map(workOutputFromFile);
+    const outputs = files.filter((file) => baseline.get(file.path) !== outputSignature(file)).sort((left, right) => left.path.localeCompare(right.path)).slice(0, MAX_WORK_OUTPUTS).map(workOutputFromFile);
+    for (const required of requiredOutputs ?? []) {
+      const output = outputs.find((item) => normalizedOutputPath(item.path) === `${WORK_OUTPUT_DIRECTORY}/${required.path}`);
+      if (!output) throw new Error("work_required_output_missing:" + required.path);
+      const downloaded = await this.#sandboxes.downloadFileResponse(sandboxId, output.path);
+      if (downloaded.file.truncated) throw new Error("work_output_download_truncated");
+      output.validation = validateOutputBytes(required, Buffer.from(downloaded.file.contentsBase64, "base64"));
+    }
+    return outputs;
   }
   async #listOutputFiles(sandboxId) {
     const listed = await this.#sandboxes.listFiles(sandboxId, {
@@ -2366,9 +2399,24 @@ var OpenPondWorkClient = class {
     }
     throw new Error(`Sandbox ${sandboxId} did not become ready before timeout`);
   }
-  async #executeTool(sandboxId, toolCall, index, timeoutSeconds, emit) {
+  async #executeTool(sandboxId, toolCall, index, timeoutSeconds, emit, requiredOutputs) {
     const toolCallId = toolCall.id || `work_tool_${index}`;
     const name = toolCall.function?.name;
+    const structured = structuredToolOutput(requiredOutputs ?? [], name ?? "");
+    if (structured) {
+      try {
+        const bytes = serializeStructuredOutput(structured, JSON.parse(toolCall.function?.arguments || "{}"));
+        const path = `${WORK_OUTPUT_DIRECTORY}/${structured.path}`;
+        const temporary = `${WORK_OUTPUT_DIRECTORY}/.openpond-${randomUUID2()}`;
+        await this.#sandboxes.uploadFileBase64(sandboxId, temporary, Buffer.from(bytes).toString("base64"));
+        const quote = (value) => "'" + value.replaceAll("'", "'\\''") + "'";
+        const moved = await this.#sandboxes.exec(sandboxId, { command: `mkdir -p -- ${quote(path.slice(0, path.lastIndexOf("/")))} && mv -f -- ${quote(temporary)} ${quote(path)}`, timeoutSeconds: 30 });
+        if (moved.command.exitCode !== 0) throw new Error("work_output_registration_failed");
+        return toolMessage(toolCallId, { ok: true, path, validation: validateOutputBytes(structured, bytes) });
+      } catch (error) {
+        return toolMessage(toolCallId, { ok: false, error: errorMessage(error) });
+      }
+    }
     if (name !== "run_command") {
       return toolMessage(toolCallId, { error: `Unknown tool: ${name ?? "missing"}` });
     }