openpond-sdk 0.5.14 → 0.5.16

package/dist/model-starter-catalog.js

@@ -1,8 +1,8 @@
 import {
   ModelStarterSchema,
   validateResolvedModelStarter
-} from "./chunk-6QWV5TEJ.js";
-import "./chunk-4IDTRFW2.js";
+} from "./chunk-SMWHGQ5W.js";
+import "./chunk-NZ35HEEK.js";
 import {
   ModelProjectVersionedRefSchema
 } from "./chunk-ZNFOVV7B.js";

package/dist/model-starters.js

@@ -11,7 +11,7 @@ import {
   previewModelStarter,
   validateModelStarterCreation,
   validateResolvedModelStarter
-} from "./chunk-6QWV5TEJ.js";
+} from "./chunk-SMWHGQ5W.js";
 import {
   ModelStarterEnvironmentAttemptSchema,
   ModelStarterExecutionSchema,
@@ -28,7 +28,7 @@ import {
   resolveModelTasksetExecutionResourcesAsset,
   validateModelTasksetPackage,
   verifyModelStarterEnvironmentAttempt
-} from "./chunk-4IDTRFW2.js";
+} from "./chunk-NZ35HEEK.js";
 import "./chunk-ZNFOVV7B.js";
 export {
   ModelStarterContentSchema,

package/dist/taskset-drafts.js

@@ -27,7 +27,7 @@ import {
   saveTasksetDraftWorkspaceFile,
   tasksetDraftFromTaskset,
   validateTasksetDraftWorkspace
-} from "./chunk-LV4TDE5S.js";
+} from "./chunk-IIZV2N5F.js";
 import {
   AuthoringProvenanceSchema,
   AuthoringRepairSchema,
@@ -97,8 +97,8 @@ import {
   isTrainingSourceRef,
   validatePortability,
   validateTaskset
-} from "./chunk-WSTMQGOY.js";
-import "./chunk-4IDTRFW2.js";
+} from "./chunk-TLXDZ4VG.js";
+import "./chunk-NZ35HEEK.js";
 import {
   ModelTasksetDraftRequestSchema,
   TasksetDraftFileInfoSchema,
@@ -106,7 +106,7 @@ import {
   TasksetDraftFilePathSchema,
   TasksetDraftFileSchema
 } from "./chunk-S7I45WJQ.js";
-import "./chunk-F5JBFRAH.js";
+import "./chunk-GZOXT7YE.js";
 import "./chunk-A4ZDOZKP.js";
 import "./chunk-5FBPNWT4.js";
 import {

package/dist/taskset-packages.js

@@ -13,7 +13,7 @@ import {
   projectPortableTaskRecord,
   projectPreferenceComparisonRelease,
   resolvePortableTasksetRewardExecution
-} from "./chunk-7VZB24CH.js";
+} from "./chunk-NHMZCVZO.js";
 import {
   OpenPondTasksetPackageClient,
   OpenPondTasksetPackageError,
@@ -30,7 +30,7 @@ import {
   publishTasksetDraft,
   renderTasksetDraftManifests,
   validateTasksetDraftWorkspace
-} from "./chunk-LV4TDE5S.js";
+} from "./chunk-IIZV2N5F.js";
 import {
   GraderFixtureSchema,
   MAX_TASKSET_PACKAGE_BYTES,
@@ -53,11 +53,11 @@ import {
   tasksetPackageRewardBinding,
   validateTasksetLearningResources,
   validateTasksetPackage
-} from "./chunk-WSTMQGOY.js";
+} from "./chunk-TLXDZ4VG.js";
 import {
   ModelStarterExecutionSchema,
   modelStarterExecutionAssetId
-} from "./chunk-4IDTRFW2.js";
+} from "./chunk-NZ35HEEK.js";
 import {
   ModelTasksetAuthoringOwnerSchema,
   ModelTasksetAuthoringSchema,
@@ -69,7 +69,7 @@ import {
   TasksetDraftFilePathSchema,
   TasksetDraftFileSchema
 } from "./chunk-S7I45WJQ.js";
-import "./chunk-F5JBFRAH.js";
+import "./chunk-GZOXT7YE.js";
 import "./chunk-A4ZDOZKP.js";
 import "./chunk-5FBPNWT4.js";
 import {

package/dist/training-bundle.js

@@ -2,13 +2,14 @@ import {
   learningVerifierModule,
   projectLearningBatchGraders,
   resolvePortableTasksetRewardExecution
-} from "./chunk-7VZB24CH.js";
+} from "./chunk-NHMZCVZO.js";
 import {
   LearningBatchDatasetSourceRefSchema,
+  MAX_TASKSET_PACKAGE_BYTES,
   TasksetSchema,
   computeTasksetHash
-} from "./chunk-WSTMQGOY.js";
-import "./chunk-4IDTRFW2.js";
+} from "./chunk-TLXDZ4VG.js";
+import "./chunk-NZ35HEEK.js";
 import "./chunk-S7I45WJQ.js";
 import {
   TRAINING_EVALUATION_SOURCE_PATH,
@@ -19,7 +20,7 @@ import {
   trainingEvaluationSourceRef,
   trainingInputArtifactUploadHash,
   trainingJobSubmissionHash
-} from "./chunk-F5JBFRAH.js";
+} from "./chunk-GZOXT7YE.js";
 import {
   HarnessSourceSelectionSchema,
   createAgentSnapshot,
@@ -39,7 +40,7 @@ import "./chunk-ZNFOVV7B.js";
 
 // src/training-bundle.ts
 import { compileBoundGraders } from "@openpond/evals/rewards";
-import { verifyLearningTextAsset as verifyLearningTextAsset2 } from "@openpond/evals/learning";
+import { verifyLearningTextAsset as verifyLearningTextAsset3 } from "@openpond/evals/learning";
 
 // src/training-bundle-contracts.ts
 import { z } from "zod";
@@ -179,12 +180,50 @@ function scanAndRedactEvidence(text) {
 }
 
 // src/training-learning-batch.ts
-import { compileTaskBatch, learningRef, taskBatchPackageMetadata, verifyLearningTextAsset } from "@openpond/evals";
+import { compileTaskBatch, learningRef, taskBatchPackageMetadata, verifyLearningTextAsset as verifyLearningTextAsset2 } from "@openpond/evals";
+
+// src/learning-batch-assets.ts
+import { verifyLearningTextAsset } from "@openpond/evals/learning";
+function materializeLearningBatchAssets(input) {
+  const tasksetAssetBytes = /* @__PURE__ */ new Map();
+  const tasks = input.tasks.map(({ artifactRefs, ...task }) => ({
+    ...task,
+    assets: artifactRefs.map((asset2) => {
+      if (asset2.visibility !== "policy") throw new Error("Learning task inputs must be policy-visible assets.");
+      const stored = input.assets?.find((value) => value.id === asset2.id);
+      const bytes2 = input.assetBytes?.get(asset2.id) ?? (stored ? new TextEncoder().encode(verifyLearningTextAsset(stored, asset2)) : void 0);
+      if (!bytes2 || bytes2.byteLength !== asset2.sizeBytes || sha256(bytes2) !== asset2.contentHash) throw new Error(`Learning task input bytes are missing or changed: ${asset2.id}.`);
+      const artifactRef = `learning/assets/${contentHash({ task: task.id, asset: asset2.id }).slice(0, 40)}/${asset2.path}`;
+      tasksetAssetBytes.set(artifactRef, new Uint8Array(bytes2));
+      return {
+        id: asset2.id,
+        sourceRefId: input.sourceId,
+        artifactRef,
+        fileName: asset2.path.split("/").at(-1),
+        mediaType: asset2.mediaType,
+        sha256: asset2.contentHash,
+        sizeBytes: asset2.sizeBytes,
+        split: task.split,
+        metadata: { portableAsset: asset2 }
+      };
+    }),
+    ...task.requiredOutputs ? { requiredOutputs: task.requiredOutputs.map((output) => ({
+      ...output,
+      schemaRef: output.schemaRef?.id ?? null,
+      maxBytes: output.maxBytes ?? void 0
+    })) } : {}
+  }));
+  return { tasks, tasksetAssetBytes };
+}
+
+// src/training-learning-batch.ts
 function materializeLearningBatchTaskset(input) {
   const release = compileTaskBatch(input);
   const metadata = taskBatchPackageMetadata(release);
-  if (release.environment.kind !== "text" || release.tools.length || release.tasks.some((task) => task.artifactRefs.length)) throw new Error("This learning batch requires an environment or asset adapter before training preparation.");
+  const work = release.environment.kind === "work";
+  if (!work && (release.environment.kind !== "text" || release.tools.length || release.tasks.some((task) => task.artifactRefs.length)) || work && release.tasks.some((task) => !task.requiredOutputs?.length)) throw new Error("This learning batch requires a supported text or Work environment with declared outputs.");
   const sourceId = input.source.id;
+  const materialized = materializeLearningBatchAssets({ tasks: release.tasks, sourceId, assets: input.assets, assetBytes: input.assetBytes });
   const signals = { demonstrations: [], preferences: [], corrections: [], feedback: [], rewards: [], labels: [] };
   const fixtures = [];
   for (const admission of metadata.admissions) {
@@ -219,7 +258,7 @@ function materializeLearningBatchTaskset(input) {
     const asset2 = input.assets?.find((asset3) => asset3.id === grader.verifierRef.id);
     if (!asset2) throw new Error(`Reward ${grader.id} is missing its immutable verifier source.`);
     const module = learningVerifierModule(grader.verifierRef.contentHash);
-    if (!generatedFiles.some((file) => file.path === module)) generatedFiles.push({ path: module, role: "verifier", content: verifyLearningTextAsset(asset2, grader.verifierRef) });
+    if (!generatedFiles.some((file) => file.path === module)) generatedFiles.push({ path: module, role: "verifier", content: verifyLearningTextAsset2(asset2, grader.verifierRef) });
   }
   const taskset = TasksetSchema.parse({
     schemaVersion: "openpond.taskset.v1",
@@ -236,9 +275,9 @@ function materializeLearningBatchTaskset(input) {
     sourceRefs: [input.source],
     datasetArtifact: null,
     policy: release.policy,
-    environment: { protocolVersion: "openpond.taskEnvironment.v1", kind: "chat", entrypoint: "openpond.text.v1", stateful: false, deterministicSeeds: release.environment.deterministicSeeds, toolNames: [], lifecycle: ["create", "reset", "step", "grade", "cleanup"], defaultTimeoutMs: release.environment.defaultTimeoutMs, networkPolicy: release.environment.networkPolicy, metadata: { portableEnvironment: release.environment } },
-    capabilities: { schemaVersion: "openpond.tasksetCapabilities.v1", taskKind: "chat", supportedSignals: signals.demonstrations.length ? ["demonstration", "reward"] : ["reward"], compatibleMethods: input.batch.purpose === "supervised_training" ? ["sft"] : input.batch.purpose === "reward_training" ? ["grpo", "ppo"] : ["none"], rewardKinds: ["deterministic"], requiresTools: false, requiresState: false, requiresPrivilegedGrading: true, environmentPlacements: ["local", "remote"], exportable: true, portabilityBlockers: [] },
-    tasks: release.tasks.map(({ artifactRefs: _assets, ...task }) => ({ ...task, schemaVersion: "openpond.taskData.v1", sourceRefs: [sourceId], metadata: { admission: metadata.admissions.find((entry) => entry.taskId === task.id), exampleOrigin: input.evidence.find((evidence) => evidence.id === task.id)?.correctionFeedbackId ? "corrected" : "extracted" } })),
+    environment: { protocolVersion: "openpond.taskEnvironment.v1", kind: work ? "work" : "chat", entrypoint: release.environment.entrypoint, stateful: release.environment.stateful, deterministicSeeds: release.environment.deterministicSeeds, toolNames: release.tools.map((tool) => tool.name), lifecycle: ["create", "reset", "step", "grade", "cleanup"], defaultTimeoutMs: release.environment.defaultTimeoutMs, networkPolicy: release.environment.networkPolicy, metadata: { portableEnvironment: release.environment, portableTools: release.tools } },
+    capabilities: { schemaVersion: "openpond.tasksetCapabilities.v1", taskKind: work ? "single_agent" : "chat", supportedSignals: signals.demonstrations.length ? ["demonstration", "reward"] : ["reward"], compatibleMethods: input.batch.purpose === "supervised_training" ? ["sft"] : input.batch.purpose === "reward_training" ? ["grpo", "ppo"] : ["none"], rewardKinds: ["deterministic"], requiresTools: release.tools.length > 0, requiresState: release.environment.stateful, requiresPrivilegedGrading: true, environmentPlacements: ["local", "remote"], exportable: true, portabilityBlockers: [] },
+    tasks: materialized.tasks.map((task) => ({ ...task, schemaVersion: "openpond.taskData.v1", sourceRefs: [sourceId], metadata: { admission: metadata.admissions.find((entry) => entry.taskId === task.id), exampleOrigin: input.evidence.find((evidence) => evidence.id === task.id)?.correctionFeedbackId ? "corrected" : "extracted" } })),
     graders,
     graderFixtures: fixtures,
     learningSignals: signals,
@@ -247,9 +286,9 @@ function materializeLearningBatchTaskset(input) {
     contentHash: "00000000",
     createdAt: input.batch.sealedAt,
     updatedAt: input.batch.sealedAt,
-    metadata: { learning: metadata, learningRelease: learningRef(release), trainingMethod: input.batch.purpose === "supervised_training" ? "sft" : input.batch.purpose === "reward_training" ? "grpo" : "none", tasksetOutputContract: { mode: "structured_json", jsonSchema: input.definition.outputSchema, renderer: null } }
+    metadata: { learning: metadata, learningRelease: learningRef(release), trainingMethod: input.batch.purpose === "supervised_training" ? "sft" : input.batch.purpose === "reward_training" ? "grpo" : "none", tasksetOutputContract: work ? { mode: "artifacts", requiredOutputSource: "task.requiredOutputs" } : { mode: "structured_json", jsonSchema: input.definition.outputSchema, renderer: null } }
   });
-  return { taskset: TasksetSchema.parse({ ...taskset, contentHash: computeTasksetHash(taskset) }), release, generatedFiles };
+  return { taskset: TasksetSchema.parse({ ...taskset, contentHash: computeTasksetHash(taskset) }), release, generatedFiles, tasksetAssetBytes: materialized.tasksetAssetBytes };
 }
 function fixtureOutcome(grade) {
   if (!grade) return null;
@@ -258,10 +297,22 @@ function fixtureOutcome(grade) {
   return { passed: outcome.passed === true, rewardEligible: grade.training.status === "scored" };
 }
 function prepareReviewedLearningBatch(input) {
+  const inputAssetText = [];
+  let inputAssetSize = 0;
+  for (const ref of new Map(input.evidence.flatMap((item) => item.submission.assets.map((ref2) => [ref2.id, ref2]))).values()) {
+    const bytes2 = input.assetBytes?.get(ref.id);
+    const stored = input.assets?.find((asset2) => asset2.id === ref.id);
+    const size = bytes2?.byteLength ?? (stored ? new TextEncoder().encode(stored.text).byteLength : 0);
+    inputAssetSize += size;
+    if (inputAssetSize > MAX_TASKSET_PACKAGE_BYTES) throw new Error("Learning input assets exceed the training package limit.");
+    if (bytes2) inputAssetText.push(new TextDecoder().decode(bytes2));
+    else if (stored) inputAssetText.push(stored.text);
+  }
   const scan = scanAndRedactEvidence(JSON.stringify({
     tasks: input.evidence.map((item) => item.submission),
     targets: input.decisions.map((item) => item.approvedTarget),
-    definition: input.definition
+    definition: input.definition,
+    inputAssets: inputAssetText
   }));
   if (scan.secretStatus !== "passed" || scan.piiStatus !== "passed") throw new Error(`Learning batch contains unresolved data findings (${scan.findings.join(", ")}). Correct the source examples and seal a revised batch before training.`);
   const source = LearningBatchDatasetSourceRefSchema.parse({
@@ -704,7 +755,7 @@ function buildTasksetTrainingBundle(input) {
     for (const reference of references) {
       const asset2 = verifierAssets.find((asset3) => asset3.id === reference.id);
       if (!asset2 || reference.visibility !== "verifier") throw new Error("Training bundle requires its private verifier asset.");
-      verifyLearningTextAsset2(asset2, reference);
+      verifyLearningTextAsset3(asset2, reference);
     }
     addJsonAsset(assets, "reward-binding.json", { kind: "reward_binding_v1", binding: rewardExecution.binding, rewards: rewardExecution.rewards, assets: verifierAssets });
   } else if (input.verifierAssets?.length) {