@odla-ai/harness 0.11.17 → 0.11.19

package/dist/node.d.cts

@@ -1,6 +1,7 @@
-import { v as HarnessTaskSpec, i as HarnessAgentOutput, h as HarnessAgentInput, H as HarnessControlPlane, f as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, C as CodeSessionEventData, y as HarnessToolName, z as HarnessToolRequest } from './types-D-Kv1svf.cjs';
+import { t as HarnessTaskSpec, j as HarnessAgentOutput, i as HarnessAgentInput, H as HarnessControlPlane, g as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, f as HarnessToolName, w as HarnessToolRequest } from './types-DM-UgoBN.cjs';
 import { CodePortableCheckpoint, CodeSourceFile, CodeVerificationReceipt, CodeCheckpointState } from '@odla-ai/camel/code';
 import { TaintLabel, ToolOutput, Skill, Inference, AgentRunBudget, AgentRun, CompactionPolicy } from '@odla-ai/ai';
+import { C as CodeSessionEventData } from './code-session-event-data-C6mQB76s.cjs';
 import { PolicyOutcome } from '@odla-ai/camel/policy';
 import { Graph, PartitionVerdict } from '@odla-ai/graph';
 
@@ -181,8 +182,17 @@ declare class CodeRuntimeReconciler {
     private notify;
 }
 
-/** Version of the outbound Code host heartbeat contract. */
-declare const CODE_RUNTIME_PROTOCOL_VERSION: 3;
+/** Version of the outbound Code host heartbeat contract.
+ *
+ * 2 understood read-only coding sessions, 3 materializes selected sources on
+ * the host, 4 runs `planning` sessions: read-only like a review, inferring on
+ * the planning role's model, with a coding session's collaboration tools. */
+declare const CODE_RUNTIME_PROTOCOL_VERSION: 4;
+/** What a Code session is for, which decides its model, its tools, and whether it may write.
+ *
+ * `coding` implements; `review` judges a candidate and is always read-only; `planning`
+ * coordinates (reads the repository, works the PM board) and is always read-only. */
+type CodeSessionRole = "coding" | "review" | "planning";
 /** Frozen, secret-filtered view of the checkout from which a terminal connected. */
 interface CodeLocalSourceDescriptor {
     kind: "local_checkout";
@@ -853,7 +863,7 @@ declare function verifyCodeCandidate(input: VerifyCodeCandidateInput): Promise<C
 /** Trusted session state, verification policy, and review callback used to stop at a portable checkpoint. */
 interface PrepareRuntimeCheckpointInput {
     sessionId: string;
-    role: "coding" | "review";
+    role: CodeSessionRole;
     workspace: StagedWorkspace;
     baseCommitSha: string;
     trustedBaseDigest: `sha256:${string}`;
@@ -888,7 +898,7 @@ interface RuntimeCheckpointSession {
     trustedBaseDigest: `sha256:${string}`;
     planningInputDigest: `sha256:${string}`;
     conversationRefs: string[];
-    role: "coding" | "review";
+    role: CodeSessionRole;
     /** The session's own recipe list and its digest; the manager's release list stands in when absent. */
     recipes?: readonly CodeBuildRecipe[];
     buildPolicyDigest?: `sha256:${string}`;
@@ -909,11 +919,17 @@ declare class CodeRuntimeCheckpointManager {
     acknowledged(command: CodeRuntimeCommand, result: CodeRuntimeCommandResult): Promise<boolean>;
 }
 
+/** How much of the agent's work the session stream carries. `standard` is the
+ * content-minimized contract every watcher has always read; `verbose` adds the
+ * reasoning, interim prose, tool inputs, and tool outputs, bounded, for a
+ * watcher that renders the work as it happens. */
+type CodeEventDetail = "standard" | "verbose";
 interface CodeCommandMetadata {
-    role: "coding" | "review";
+    role: CodeSessionRole;
     readOnly: boolean;
     title: string;
     prompt: string;
+    eventDetail: CodeEventDetail;
     /** The card's executable proof, argv per command; null when the session carries none. */
     proof: readonly (readonly string[])[] | null;
     maxTokensPerInteraction: number;
@@ -1719,4 +1735,4 @@ declare function extractTapFailingTests(output: string, limit?: number): string[
 /** One line naming the failing tests, or `null` when the output names none. */
 declare function describeTapFailures(output: string): string | null;
 
-export { CODE_PATCH_PATH, CODE_RUNTIME_PROTOCOL_VERSION, type CodeAgentAttemptOptions, type CodeAgentAttemptResult, type CodeAgentRun, type CodeBuildRecipe, type CodeExpectedArtifact, type CodeLocalSourceDescriptor, type CodeMemory, type CodeMemoryEvidence, type CodeMemoryKind, type CodeMemoryStore, type CodeRecipeExecutor, type CodeRecipeResult, type CodeRuntimeAgentControlPlane, type CodeRuntimeBinding, type CodeRuntimeCandidateResponse, type CodeRuntimeCapabilities, CodeRuntimeCheckpointManager, type CodeRuntimeClientOptions, type CodeRuntimeCollaborationSkillManifest, type CodeRuntimeCollaborationToolManifest, type CodeRuntimeCollaborationToolRequest, type CodeRuntimeCommand, type CodeRuntimeCommandEngine, type CodeRuntimeCommandKind, type CodeRuntimeCommandResult, CodeRuntimeControlError, type CodeRuntimeControlPlane, type CodeRuntimeInferenceOptions, type CodeRuntimeLoopOptions, type CodeRuntimeMemory, type CodeRuntimeNewMemory, CodeRuntimeReconciler, type CodeRuntimeReviewRequest, type CodeRuntimeReviewResponse, type CodeRuntimeSnapshot, type CodeRuntimeSourceArchive, type CodeRuntimeSourceSnapshot, type CodeSkillOpts, type CodeSurface, type CodeToolBrokerOptions, type CodeToolCallRecord, type CodeToolDecision, type CodeVerificationEvidence, type CodeVerificationLog, type CodeVerificationPolicy, type ContainerEngine, type ContainerEngineSelectionOptions, type ContainerEngineVerificationOptions, type ContainerLimits, type ContainerRunOptions, type ContainerRunResult, type CreateCodeWorkspaceCheckpointInput, DEPENDENCY_INSTALL_TIMEOUT_MS, type DecomposedRun, DecompositionError, type DependencyCacheInput, type GoalAttempt, type GoalAttemptInput, type GoalAttemptOutcome, type GoalAttemptRecord, type GoalBudget, type GoalEvent, type GoalOutcome, type GoalRun, type GoalRunSpec, type GoalStoppedReason, type GoalStrategy, type HarnessRunnerOptions, type IntegrateOptions, type IntegrationStep, type LentDependencies, type LentDirectory, MAX_MEMORY_BODY, MAX_TAP_FAILURES, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, PROOF_RECIPE_PREFIX, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, REPOSITORY_RECIPES_FILE, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, type RepositoryRecipeEnvelope, type ResolvedCodeRecipes, type RestoreCodeWorkspaceCheckpointInput, type RestoredCodeWorkspaceCheckpoint, type RunCodeAgentOptions, type RuntimeCheckpointSession, SYSTEM_PROMPT_FOR, type StageWorkspaceOptions, type StagedWorkspace, type StrategyChoice, type StrategySignals, type SubGoal, type SubGoalResult, TheseusRuntimeEngine, type TheseusRuntimeEngineOptions, V1_SYSTEM_PROMPT, V2_SYSTEM_PROMPT, V3_SYSTEM_PROMPT, type VerifyCodeCandidateInput, applyCodePatch, applyPatchDialectToDiff, assertCodeBuildRecipe, assertDisjointPlan, assertLentPath, assertPinnedImage, assertReservedMount, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, cloneTree, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeRuntimeSessionSkillLoader, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, dependencyCacheKey, dependencyInstallArgs, describeCodeRecipes, describeGateFailure, describePatchFailure, describeTapFailures, digestStagedWorkspace, extractTapFailingTests, feedbackIsActionable, hasContextFreeHunk, hazardFromAttempt, hostBuildProducts, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, lendBuildProducts, lendDependencies, lentDirectories, materializeCodeRuntimeArchive, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, npmFailureSummary, outcomeCloses, outcomeMemory, parseRepositoryRecipes, patchPaths, pinnedNpmVersion, planReachCollisions, prepareContainerDependencies, prepareRuntimeCheckpoint, racedAttempt, readOnlyNotice, readRepositoryRecipes, recallAbout, recipeOutput, registeredFiles, renderMemories, renderOutcome, resolveCodePath, resolveCodeRecipes, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runContainerCommand, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, sessionRecipesFor, sessionSkillsFor, stageManifests, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withProofRecipes, withRecipeDependencies, workspaceDirs };
+export { CODE_PATCH_PATH, CODE_RUNTIME_PROTOCOL_VERSION, type CodeAgentAttemptOptions, type CodeAgentAttemptResult, type CodeAgentRun, type CodeBuildRecipe, type CodeExpectedArtifact, type CodeLocalSourceDescriptor, type CodeMemory, type CodeMemoryEvidence, type CodeMemoryKind, type CodeMemoryStore, type CodeRecipeExecutor, type CodeRecipeResult, type CodeRuntimeAgentControlPlane, type CodeRuntimeBinding, type CodeRuntimeCandidateResponse, type CodeRuntimeCapabilities, CodeRuntimeCheckpointManager, type CodeRuntimeClientOptions, type CodeRuntimeCollaborationSkillManifest, type CodeRuntimeCollaborationToolManifest, type CodeRuntimeCollaborationToolRequest, type CodeRuntimeCommand, type CodeRuntimeCommandEngine, type CodeRuntimeCommandKind, type CodeRuntimeCommandResult, CodeRuntimeControlError, type CodeRuntimeControlPlane, type CodeRuntimeInferenceOptions, type CodeRuntimeLoopOptions, type CodeRuntimeMemory, type CodeRuntimeNewMemory, CodeRuntimeReconciler, type CodeRuntimeReviewRequest, type CodeRuntimeReviewResponse, type CodeRuntimeSnapshot, type CodeRuntimeSourceArchive, type CodeRuntimeSourceSnapshot, type CodeSessionRole, type CodeSkillOpts, type CodeSurface, type CodeToolBrokerOptions, type CodeToolCallRecord, type CodeToolDecision, type CodeVerificationEvidence, type CodeVerificationLog, type CodeVerificationPolicy, type ContainerEngine, type ContainerEngineSelectionOptions, type ContainerEngineVerificationOptions, type ContainerLimits, type ContainerRunOptions, type ContainerRunResult, type CreateCodeWorkspaceCheckpointInput, DEPENDENCY_INSTALL_TIMEOUT_MS, type DecomposedRun, DecompositionError, type DependencyCacheInput, type GoalAttempt, type GoalAttemptInput, type GoalAttemptOutcome, type GoalAttemptRecord, type GoalBudget, type GoalEvent, type GoalOutcome, type GoalRun, type GoalRunSpec, type GoalStoppedReason, type GoalStrategy, type HarnessRunnerOptions, type IntegrateOptions, type IntegrationStep, type LentDependencies, type LentDirectory, MAX_MEMORY_BODY, MAX_TAP_FAILURES, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, PROOF_RECIPE_PREFIX, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, REPOSITORY_RECIPES_FILE, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, type RepositoryRecipeEnvelope, type ResolvedCodeRecipes, type RestoreCodeWorkspaceCheckpointInput, type RestoredCodeWorkspaceCheckpoint, type RunCodeAgentOptions, type RuntimeCheckpointSession, SYSTEM_PROMPT_FOR, type StageWorkspaceOptions, type StagedWorkspace, type StrategyChoice, type StrategySignals, type SubGoal, type SubGoalResult, TheseusRuntimeEngine, type TheseusRuntimeEngineOptions, V1_SYSTEM_PROMPT, V2_SYSTEM_PROMPT, V3_SYSTEM_PROMPT, type VerifyCodeCandidateInput, applyCodePatch, applyPatchDialectToDiff, assertCodeBuildRecipe, assertDisjointPlan, assertLentPath, assertPinnedImage, assertReservedMount, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, cloneTree, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeRuntimeSessionSkillLoader, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, dependencyCacheKey, dependencyInstallArgs, describeCodeRecipes, describeGateFailure, describePatchFailure, describeTapFailures, digestStagedWorkspace, extractTapFailingTests, feedbackIsActionable, hasContextFreeHunk, hazardFromAttempt, hostBuildProducts, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, lendBuildProducts, lendDependencies, lentDirectories, materializeCodeRuntimeArchive, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, npmFailureSummary, outcomeCloses, outcomeMemory, parseRepositoryRecipes, patchPaths, pinnedNpmVersion, planReachCollisions, prepareContainerDependencies, prepareRuntimeCheckpoint, racedAttempt, readOnlyNotice, readRepositoryRecipes, recallAbout, recipeOutput, registeredFiles, renderMemories, renderOutcome, resolveCodePath, resolveCodeRecipes, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runContainerCommand, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, sessionRecipesFor, sessionSkillsFor, stageManifests, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withProofRecipes, withRecipeDependencies, workspaceDirs };

package/dist/node.d.ts

@@ -1,6 +1,7 @@
-import { v as HarnessTaskSpec, i as HarnessAgentOutput, h as HarnessAgentInput, H as HarnessControlPlane, f as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, C as CodeSessionEventData, y as HarnessToolName, z as HarnessToolRequest } from './types-D-Kv1svf.js';
+import { t as HarnessTaskSpec, j as HarnessAgentOutput, i as HarnessAgentInput, H as HarnessControlPlane, g as HarnessToolBroker, a as HarnessLease, d as HarnessInferenceRequest, e as HarnessInferenceResponse, f as HarnessToolName, w as HarnessToolRequest } from './types-DM-UgoBN.js';
 import { CodePortableCheckpoint, CodeSourceFile, CodeVerificationReceipt, CodeCheckpointState } from '@odla-ai/camel/code';
 import { TaintLabel, ToolOutput, Skill, Inference, AgentRunBudget, AgentRun, CompactionPolicy } from '@odla-ai/ai';
+import { C as CodeSessionEventData } from './code-session-event-data-wCQc2py3.js';
 import { PolicyOutcome } from '@odla-ai/camel/policy';
 import { Graph, PartitionVerdict } from '@odla-ai/graph';
 
@@ -181,8 +182,17 @@ declare class CodeRuntimeReconciler {
     private notify;
 }
 
-/** Version of the outbound Code host heartbeat contract. */
-declare const CODE_RUNTIME_PROTOCOL_VERSION: 3;
+/** Version of the outbound Code host heartbeat contract.
+ *
+ * 2 understood read-only coding sessions, 3 materializes selected sources on
+ * the host, 4 runs `planning` sessions: read-only like a review, inferring on
+ * the planning role's model, with a coding session's collaboration tools. */
+declare const CODE_RUNTIME_PROTOCOL_VERSION: 4;
+/** What a Code session is for, which decides its model, its tools, and whether it may write.
+ *
+ * `coding` implements; `review` judges a candidate and is always read-only; `planning`
+ * coordinates (reads the repository, works the PM board) and is always read-only. */
+type CodeSessionRole = "coding" | "review" | "planning";
 /** Frozen, secret-filtered view of the checkout from which a terminal connected. */
 interface CodeLocalSourceDescriptor {
     kind: "local_checkout";
@@ -853,7 +863,7 @@ declare function verifyCodeCandidate(input: VerifyCodeCandidateInput): Promise<C
 /** Trusted session state, verification policy, and review callback used to stop at a portable checkpoint. */
 interface PrepareRuntimeCheckpointInput {
     sessionId: string;
-    role: "coding" | "review";
+    role: CodeSessionRole;
     workspace: StagedWorkspace;
     baseCommitSha: string;
     trustedBaseDigest: `sha256:${string}`;
@@ -888,7 +898,7 @@ interface RuntimeCheckpointSession {
     trustedBaseDigest: `sha256:${string}`;
     planningInputDigest: `sha256:${string}`;
     conversationRefs: string[];
-    role: "coding" | "review";
+    role: CodeSessionRole;
     /** The session's own recipe list and its digest; the manager's release list stands in when absent. */
     recipes?: readonly CodeBuildRecipe[];
     buildPolicyDigest?: `sha256:${string}`;
@@ -909,11 +919,17 @@ declare class CodeRuntimeCheckpointManager {
     acknowledged(command: CodeRuntimeCommand, result: CodeRuntimeCommandResult): Promise<boolean>;
 }
 
+/** How much of the agent's work the session stream carries. `standard` is the
+ * content-minimized contract every watcher has always read; `verbose` adds the
+ * reasoning, interim prose, tool inputs, and tool outputs, bounded, for a
+ * watcher that renders the work as it happens. */
+type CodeEventDetail = "standard" | "verbose";
 interface CodeCommandMetadata {
-    role: "coding" | "review";
+    role: CodeSessionRole;
     readOnly: boolean;
     title: string;
     prompt: string;
+    eventDetail: CodeEventDetail;
     /** The card's executable proof, argv per command; null when the session carries none. */
     proof: readonly (readonly string[])[] | null;
     maxTokensPerInteraction: number;
@@ -1719,4 +1735,4 @@ declare function extractTapFailingTests(output: string, limit?: number): string[
 /** One line naming the failing tests, or `null` when the output names none. */
 declare function describeTapFailures(output: string): string | null;
 
-export { CODE_PATCH_PATH, CODE_RUNTIME_PROTOCOL_VERSION, type CodeAgentAttemptOptions, type CodeAgentAttemptResult, type CodeAgentRun, type CodeBuildRecipe, type CodeExpectedArtifact, type CodeLocalSourceDescriptor, type CodeMemory, type CodeMemoryEvidence, type CodeMemoryKind, type CodeMemoryStore, type CodeRecipeExecutor, type CodeRecipeResult, type CodeRuntimeAgentControlPlane, type CodeRuntimeBinding, type CodeRuntimeCandidateResponse, type CodeRuntimeCapabilities, CodeRuntimeCheckpointManager, type CodeRuntimeClientOptions, type CodeRuntimeCollaborationSkillManifest, type CodeRuntimeCollaborationToolManifest, type CodeRuntimeCollaborationToolRequest, type CodeRuntimeCommand, type CodeRuntimeCommandEngine, type CodeRuntimeCommandKind, type CodeRuntimeCommandResult, CodeRuntimeControlError, type CodeRuntimeControlPlane, type CodeRuntimeInferenceOptions, type CodeRuntimeLoopOptions, type CodeRuntimeMemory, type CodeRuntimeNewMemory, CodeRuntimeReconciler, type CodeRuntimeReviewRequest, type CodeRuntimeReviewResponse, type CodeRuntimeSnapshot, type CodeRuntimeSourceArchive, type CodeRuntimeSourceSnapshot, type CodeSkillOpts, type CodeSurface, type CodeToolBrokerOptions, type CodeToolCallRecord, type CodeToolDecision, type CodeVerificationEvidence, type CodeVerificationLog, type CodeVerificationPolicy, type ContainerEngine, type ContainerEngineSelectionOptions, type ContainerEngineVerificationOptions, type ContainerLimits, type ContainerRunOptions, type ContainerRunResult, type CreateCodeWorkspaceCheckpointInput, DEPENDENCY_INSTALL_TIMEOUT_MS, type DecomposedRun, DecompositionError, type DependencyCacheInput, type GoalAttempt, type GoalAttemptInput, type GoalAttemptOutcome, type GoalAttemptRecord, type GoalBudget, type GoalEvent, type GoalOutcome, type GoalRun, type GoalRunSpec, type GoalStoppedReason, type GoalStrategy, type HarnessRunnerOptions, type IntegrateOptions, type IntegrationStep, type LentDependencies, type LentDirectory, MAX_MEMORY_BODY, MAX_TAP_FAILURES, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, PROOF_RECIPE_PREFIX, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, REPOSITORY_RECIPES_FILE, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, type RepositoryRecipeEnvelope, type ResolvedCodeRecipes, type RestoreCodeWorkspaceCheckpointInput, type RestoredCodeWorkspaceCheckpoint, type RunCodeAgentOptions, type RuntimeCheckpointSession, SYSTEM_PROMPT_FOR, type StageWorkspaceOptions, type StagedWorkspace, type StrategyChoice, type StrategySignals, type SubGoal, type SubGoalResult, TheseusRuntimeEngine, type TheseusRuntimeEngineOptions, V1_SYSTEM_PROMPT, V2_SYSTEM_PROMPT, V3_SYSTEM_PROMPT, type VerifyCodeCandidateInput, applyCodePatch, applyPatchDialectToDiff, assertCodeBuildRecipe, assertDisjointPlan, assertLentPath, assertPinnedImage, assertReservedMount, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, cloneTree, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeRuntimeSessionSkillLoader, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, dependencyCacheKey, dependencyInstallArgs, describeCodeRecipes, describeGateFailure, describePatchFailure, describeTapFailures, digestStagedWorkspace, extractTapFailingTests, feedbackIsActionable, hasContextFreeHunk, hazardFromAttempt, hostBuildProducts, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, lendBuildProducts, lendDependencies, lentDirectories, materializeCodeRuntimeArchive, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, npmFailureSummary, outcomeCloses, outcomeMemory, parseRepositoryRecipes, patchPaths, pinnedNpmVersion, planReachCollisions, prepareContainerDependencies, prepareRuntimeCheckpoint, racedAttempt, readOnlyNotice, readRepositoryRecipes, recallAbout, recipeOutput, registeredFiles, renderMemories, renderOutcome, resolveCodePath, resolveCodeRecipes, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runContainerCommand, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, sessionRecipesFor, sessionSkillsFor, stageManifests, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withProofRecipes, withRecipeDependencies, workspaceDirs };
+export { CODE_PATCH_PATH, CODE_RUNTIME_PROTOCOL_VERSION, type CodeAgentAttemptOptions, type CodeAgentAttemptResult, type CodeAgentRun, type CodeBuildRecipe, type CodeExpectedArtifact, type CodeLocalSourceDescriptor, type CodeMemory, type CodeMemoryEvidence, type CodeMemoryKind, type CodeMemoryStore, type CodeRecipeExecutor, type CodeRecipeResult, type CodeRuntimeAgentControlPlane, type CodeRuntimeBinding, type CodeRuntimeCandidateResponse, type CodeRuntimeCapabilities, CodeRuntimeCheckpointManager, type CodeRuntimeClientOptions, type CodeRuntimeCollaborationSkillManifest, type CodeRuntimeCollaborationToolManifest, type CodeRuntimeCollaborationToolRequest, type CodeRuntimeCommand, type CodeRuntimeCommandEngine, type CodeRuntimeCommandKind, type CodeRuntimeCommandResult, CodeRuntimeControlError, type CodeRuntimeControlPlane, type CodeRuntimeInferenceOptions, type CodeRuntimeLoopOptions, type CodeRuntimeMemory, type CodeRuntimeNewMemory, CodeRuntimeReconciler, type CodeRuntimeReviewRequest, type CodeRuntimeReviewResponse, type CodeRuntimeSnapshot, type CodeRuntimeSourceArchive, type CodeRuntimeSourceSnapshot, type CodeSessionRole, type CodeSkillOpts, type CodeSurface, type CodeToolBrokerOptions, type CodeToolCallRecord, type CodeToolDecision, type CodeVerificationEvidence, type CodeVerificationLog, type CodeVerificationPolicy, type ContainerEngine, type ContainerEngineSelectionOptions, type ContainerEngineVerificationOptions, type ContainerLimits, type ContainerRunOptions, type ContainerRunResult, type CreateCodeWorkspaceCheckpointInput, DEPENDENCY_INSTALL_TIMEOUT_MS, type DecomposedRun, DecompositionError, type DependencyCacheInput, type GoalAttempt, type GoalAttemptInput, type GoalAttemptOutcome, type GoalAttemptRecord, type GoalBudget, type GoalEvent, type GoalOutcome, type GoalRun, type GoalRunSpec, type GoalStoppedReason, type GoalStrategy, type HarnessRunnerOptions, type IntegrateOptions, type IntegrationStep, type LentDependencies, type LentDirectory, MAX_MEMORY_BODY, MAX_TAP_FAILURES, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, PROOF_RECIPE_PREFIX, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, REPOSITORY_RECIPES_FILE, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, type RepositoryRecipeEnvelope, type ResolvedCodeRecipes, type RestoreCodeWorkspaceCheckpointInput, type RestoredCodeWorkspaceCheckpoint, type RunCodeAgentOptions, type RuntimeCheckpointSession, SYSTEM_PROMPT_FOR, type StageWorkspaceOptions, type StagedWorkspace, type StrategyChoice, type StrategySignals, type SubGoal, type SubGoalResult, TheseusRuntimeEngine, type TheseusRuntimeEngineOptions, V1_SYSTEM_PROMPT, V2_SYSTEM_PROMPT, V3_SYSTEM_PROMPT, type VerifyCodeCandidateInput, applyCodePatch, applyPatchDialectToDiff, assertCodeBuildRecipe, assertDisjointPlan, assertLentPath, assertPinnedImage, assertReservedMount, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, cloneTree, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeRuntimeSessionSkillLoader, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, dependencyCacheKey, dependencyInstallArgs, describeCodeRecipes, describeGateFailure, describePatchFailure, describeTapFailures, digestStagedWorkspace, extractTapFailingTests, feedbackIsActionable, hasContextFreeHunk, hazardFromAttempt, hostBuildProducts, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, lendBuildProducts, lendDependencies, lentDirectories, materializeCodeRuntimeArchive, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, npmFailureSummary, outcomeCloses, outcomeMemory, parseRepositoryRecipes, patchPaths, pinnedNpmVersion, planReachCollisions, prepareContainerDependencies, prepareRuntimeCheckpoint, racedAttempt, readOnlyNotice, readRepositoryRecipes, recallAbout, recipeOutput, registeredFiles, renderMemories, renderOutcome, resolveCodePath, resolveCodeRecipes, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runContainerCommand, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, sessionRecipesFor, sessionSkillsFor, stageManifests, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withProofRecipes, withRecipeDependencies, workspaceDirs };

package/dist/node.js

@@ -1,7 +1,7 @@
 import {
   runHarnessRunner,
   runLeasedAttempt
-} from "./chunk-EXSEQEZ6.js";
+} from "./chunk-M6BNH25Z.js";
 import {
   CODE_PATCH_PATH,
   CODE_RUNTIME_PROTOCOL_VERSION,
@@ -73,8 +73,8 @@ import {
   verifyCodeCandidate,
   withProofRecipes,
   withRecipeDependencies
-} from "./chunk-RDCU4PIP.js";
-import "./chunk-6T26HEWI.js";
+} from "./chunk-7NP3YOMY.js";
+import "./chunk-SMN2FWQE.js";
 import {
   allowedWorkspacePath,
   assertPinnedImage,
@@ -86,9 +86,9 @@ import {
   stageWorkspace,
   stageWorkspacePair,
   verifyContainerEngineBoundary
-} from "./chunk-3ON6UAOV.js";
-import "./chunk-T2IW7MYQ.js";
-import "./chunk-ONYW2VSB.js";
+} from "./chunk-A5B3CEK2.js";
+import "./chunk-4TG55ATB.js";
+import "./chunk-O4QLHUGO.js";
 
 // src/code-runtime-memory.ts
 async function mutationId(memory) {

package/dist/testing.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/testing.ts","../src/types.ts"],"sourcesContent":["import type { OracleResponse } from \"@odla-ai/ai\";\nimport {\n  HARNESS_PROTOCOL_VERSION,\n  type HarnessCompletion,\n  type HarnessControlPlane,\n  type HarnessEventInput,\n  type HarnessInferenceRequest,\n  type HarnessInferenceResponse,\n  type HarnessLease,\n} from \"./types\";\n\n/** Create a deterministic, valid lease for harness unit and integration tests. */\nexport function fixtureLease(overrides: Partial<HarnessLease> = {}): HarnessLease {\n  return {\n    protocolVersion: HARNESS_PROTOCOL_VERSION,\n    leaseId: \"lease_fixture\",\n    generation: 1,\n    expiresAt: Date.now() + 60_000,\n    task: {\n      taskId: \"task_fixture\",\n      attemptId: \"attempt_fixture\",\n      title: \"Fixture task\",\n      prompt: \"Create HARNESS_RESULT.md\",\n      workspace: \"fixture\",\n      aiRoute: \"coding\",\n      policy: {\n        network: \"none\",\n        timeoutMs: 30_000,\n        maxOutputBytes: 1_000_000,\n        maxPatchBytes: 1_000_000,\n      },\n    },\n    ...overrides,\n  };\n}\n\n/** In-memory control plane that records runner interactions without network access. */\nexport class FakeHarnessControlPlane implements HarnessControlPlane {\n  readonly leases: HarnessLease[] = [];\n  readonly events: HarnessEventInput[] = [];\n  readonly completions: HarnessCompletion[] = [];\n  readonly inferenceRequests: HarnessInferenceRequest[] = [];\n  cancelRequested = false;\n  response: OracleResponse = {\n    id: \"oracle_fixture\",\n    model: \"fixture-model\",\n    provider: \"openai\",\n    role: \"assistant\",\n    content: [{ type: \"text\", text: \"deterministic fixture response\" }],\n    stopReason: \"end_turn\",\n    usage: { inputTokens: 5, outputTokens: 4 },\n  };\n\n  constructor(...leases: HarnessLease[]) { this.leases.push(...leases); }\n\n  async lease(workspaces: string[]): Promise<HarnessLease | null> {\n    const index = this.leases.findIndex((candidate) => workspaces.includes(candidate.task.workspace));\n    return index < 0 ? null : this.leases.splice(index, 1)[0]!;\n  }\n  async heartbeat(): Promise<{ cancelRequested: boolean; expiresAt: number }> {\n    return { cancelRequested: this.cancelRequested, expiresAt: Date.now() + 60_000 };\n  }\n  async appendEvents(_attemptId: string, _leaseId: string, events: HarnessEventInput[]): Promise<void> {\n    this.events.push(...events);\n  }\n  async infer(_attemptId: string, _leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse> {\n    this.inferenceRequests.push(request);\n    return {\n      requestId: request.requestId,\n      response: this.response,\n      receipt: {\n        provider: this.response.provider,\n        model: this.response.model,\n        policyVersion: 1,\n        inputTokens: this.response.usage.inputTokens ?? 0,\n        outputTokens: this.response.usage.outputTokens ?? 0,\n      },\n    };\n  }\n  async complete(_attemptId: string, _leaseId: string, completion: HarnessCompletion): Promise<void> {\n    this.completions.push(completion);\n  }\n}\n","import type { ChatInput, OracleResponse } from \"@odla-ai/ai\";\nimport type { CodeToolPresentation } from \"./code-session-event-types\";\nexport type { CodeToolLocationPreview, CodeToolPresentation } from \"./code-session-event-types\";\n\n/** Current JSONL protocol version exchanged between a runner and an agent container. */\nexport const HARNESS_PROTOCOL_VERSION = 1 as const;\n\n/** Default control-plane route used for model inference requested by coding agents. */\nexport const DEFAULT_AI_ROUTE = \"coding\" as const;\n\n/** Lifecycle state reported for a harness task and its active attempt. */\nexport type HarnessTaskStatus =\n  | \"queued\"\n  | \"running\"\n  | \"cancel_requested\"\n  | \"completed\"\n  | \"failed\"\n  | \"cancelled\";\n\n/** Lifecycle state of one execution attempt for a task. */\nexport type HarnessAttemptStatus = HarnessTaskStatus;\n\n/** Trusted or untrusted participant that emitted a harness event. */\nexport type HarnessActor = \"operator\" | \"runner\" | \"agent\" | \"model\" | \"system\";\n\n/** Resource and isolation limits enforced while an untrusted task executes. */\nexport interface HarnessPolicy {\n  network: \"none\";\n  timeoutMs: number;\n  maxOutputBytes: number;\n  maxPatchBytes: number;\n}\n\n/** Immutable task instructions and execution policy delivered with a lease. */\nexport interface HarnessTaskSpec {\n  taskId: string;\n  attemptId: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  policy: HarnessPolicy;\n  parentAttemptId?: string | null;\n  checkpointSeq?: number | null;\n}\n\n/** Time-bound assignment authorizing a runner to execute one task attempt. */\nexport interface HarnessLease {\n  protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n  leaseId: string;\n  generation: number;\n  expiresAt: number;\n  task: HarnessTaskSpec;\n}\n\n/** Runner-supplied event before the control plane assigns sequence and task metadata. */\nexport interface HarnessEventInput {\n  eventId: string;\n  kind: string;\n  actor: HarnessActor;\n  payload: unknown;\n  createdAt: number;\n}\n\n/** Persisted, ordered event associated with a specific task attempt. */\nexport interface HarnessEvent extends HarnessEventInput {\n  seq: number;\n  taskId: string;\n  attemptId: string;\n}\n\n/** List-view metadata for a task and its current attempt. */\nexport interface HarnessTaskSummary {\n  taskId: string;\n  attemptId: string;\n  appId: string;\n  env: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  status: HarnessTaskStatus;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** List-view metadata for one attempt, including retry ancestry and runner ownership. */\nexport interface HarnessAttemptSummary {\n  attemptId: string;\n  taskId: string;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  status: HarnessAttemptStatus;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** Complete task view including attempts, events, result, and generated patch. */\nexport interface HarnessTaskDetail extends HarnessTaskSummary {\n  attempts: HarnessAttemptSummary[];\n  events: HarnessEvent[];\n  patch: string | null;\n  result: unknown;\n}\n\n/** Public control-plane view of a registered harness runner. */\nexport interface HarnessRunnerView {\n  runnerId: string;\n  appId: string;\n  env: string;\n  name: string;\n  createdAt: number;\n  lastSeenAt: number | null;\n  revokedAt: number | null;\n}\n\n/** Credential-free normalized model request sent from an agent through the runner. */\nexport interface HarnessInferenceRequest {\n  requestId: string;\n  /** Exact initial, follow-up, or resume command whose budget this call consumes. */\n  interactionId?: string;\n  call: ChatInput;\n}\n\n/** Normalized model response plus auditable provider, policy, and token metadata. */\nexport interface HarnessInferenceResponse {\n  requestId: string;\n  response: OracleResponse;\n  receipt: {\n    provider: string;\n    model: string;\n    policyVersion: number;\n    inputTokens: number;\n    outputTokens: number;\n    /**\n     * USD charged for this call, priced against the model the control plane\n     * actually resolved.\n     *\n     * ABSENT when the live catalog has no price for that model — never zero.\n     * An unpriced call is unknown spend, and reporting it as free is what let\n     * a goal's `maxUsd` look enforced while nothing enforced it. The runtime\n     * cannot compute this itself: it asks for `brokered` and only the control\n     * plane knows which model answered.\n     */\n    costUsd?: number;\n  };\n}\n\n/** Bounded, content-minimized session activity emitted by a Code runtime.\n * Message bodies are projected separately into the app's owner-private\n * odla-db chat. `interactionId` is optional so stored v1 events remain valid. */\nexport type CodeSessionEventData = (\n  | { type: \"message\"; actor: \"agent\" | \"system\"; body: string }\n  | { type: \"diagnostic\"; level: \"error\"; message: string }\n  | { type: \"thinking\"; available: true; durationMs: number }\n  | {\n      type: \"tool\";\n      phase: \"started\";\n      tool: HarnessToolName;\n      operationId?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"tool\";\n      phase: \"completed\";\n      tool: HarnessToolName;\n      ok: boolean;\n      durationMs: number;\n      operationId?: string;\n      /** Why the call failed, bounded. Absent when `ok`. Without this a watcher\n       * saw that a tool failed and never why, which is what made an 84%\n       * apply_patch failure rate impossible to diagnose (PM bug 515655ec). */\n      failureReason?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"started\";\n      skill: string;\n      tool: string;\n      operationId: string;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"completed\";\n      skill: string;\n      tool: string;\n      ok: boolean;\n      durationMs: number;\n      operationId: string;\n    }\n  | {\n      type: \"usage\"; provider: string; model: string;\n      inputTokens: number; outputTokens: number; durationMs: number;\n      interactionTokens?: number; interactionMaxTokens?: number;\n      /** USD for this call; absent when the model is unpriced, never zero. */\n      costUsd?: number;\n      /** Cumulative USD for this owner interaction, when every call in it was\n       *  priced. Absent the moment one was not, so a partial total can never be\n       *  mistaken for the whole. */\n      interactionCostUsd?: number;\n    }\n  | {\n      type: \"status\"; status: \"running\" | \"idle\" | \"failed\" | \"checkpointed\";\n      durationMs?: number;\n    }\n) & { interactionId?: string };\n\n/** Registry-assigned cursor and timestamp for an owner-visible Code event. */\nexport type CodeSessionEvent = CodeSessionEventData & {\n  eventId: string; sequence: number; createdAt: number;\n};\n\n/** Closed set of effects an agent container may request from its trusted broker. */\nexport type HarnessToolName =\n  | \"sandbox.read\"\n  | \"sandbox.list\"\n  | \"sandbox.search\"\n  | \"sandbox.overview\"\n  | \"sandbox.where_is\"\n  | \"sandbox.who_imports\"\n  | \"sandbox.who_touches\"\n  | \"sandbox.apply_patch\"\n  | \"sandbox.edit\"\n  | \"sandbox.write\"\n  | \"sandbox.run_recipe\";\n/** Correlated, structured tool request emitted by an untrusted agent container. */\nexport interface HarnessToolRequest {\n  requestId: string;\n  tool: HarnessToolName;\n  input: Record<string, unknown>;\n}\n/** Bounded tool result returned to an agent after trusted policy evaluation. */\nexport interface HarnessToolResponse {\n  requestId: string;\n  ok: boolean;\n  content: string;\n  details?: Record<string, unknown>;\n}\n\n/** Validated JSONL message emitted by an untrusted agent container. */\nexport type HarnessAgentOutput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"event\";\n      kind: string;\n      payload?: unknown;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.request\";\n      requestId: string;\n      call: ChatInput;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.request\" } & HarnessToolRequest)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.complete\";\n      status: \"completed\" | \"failed\" | \"cancelled\";\n      result?: unknown;\n    };\n\n/** JSONL command or inference result written by the trusted runner to an agent. */\nexport type HarnessAgentInput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"task.start\";\n      task: HarnessTaskSpec;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.response\";\n      requestId: string;\n      response: OracleResponse;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.response\" } & HarnessToolResponse)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.cancel\";\n      reason: string;\n    };\n\n/** Terminal attempt report submitted by a runner to the control plane. */\nexport interface HarnessCompletion {\n  status: \"completed\" | \"failed\" | \"cancelled\";\n  result?: unknown;\n  patch?: string;\n  error?: string;\n}\n\n/** Operations a credentialed runner may perform against the harness control plane. */\nexport interface HarnessControlPlane {\n  lease(workspaces: string[]): Promise<HarnessLease | null>;\n  heartbeat(attemptId: string, leaseId: string): Promise<{ cancelRequested: boolean; expiresAt: number }>;\n  appendEvents(attemptId: string, leaseId: string, events: HarnessEventInput[]): Promise<void>;\n  infer(attemptId: string, leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n  complete(attemptId: string, leaseId: string, completion: HarnessCompletion): Promise<void>;\n}\n\n/** Trusted inference bridge used to keep model credentials outside agent containers. */\nexport interface HarnessAiConnection {\n  infer(request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n}\n\n/** Trusted tool boundary. Implementations must evaluate CaMeL policy before effects. */\nexport interface HarnessToolBroker {\n  execute(\n    context: { lease: HarnessLease; workspaceDir: string; signal?: AbortSignal },\n    request: HarnessToolRequest,\n  ): Promise<HarnessToolResponse>;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKO,IAAM,2BAA2B;;;ADOjC,SAAS,aAAa,YAAmC,CAAC,GAAiB;AAChF,SAAO;AAAA,IACL,iBAAiB;AAAA,IACjB,SAAS;AAAA,IACT,YAAY;AAAA,IACZ,WAAW,KAAK,IAAI,IAAI;AAAA,IACxB,MAAM;AAAA,MACJ,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,OAAO;AAAA,MACP,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,SAAS;AAAA,MACT,QAAQ;AAAA,QACN,SAAS;AAAA,QACT,WAAW;AAAA,QACX,gBAAgB;AAAA,QAChB,eAAe;AAAA,MACjB;AAAA,IACF;AAAA,IACA,GAAG;AAAA,EACL;AACF;AAGO,IAAM,0BAAN,MAA6D;AAAA,EACzD,SAAyB,CAAC;AAAA,EAC1B,SAA8B,CAAC;AAAA,EAC/B,cAAmC,CAAC;AAAA,EACpC,oBAA+C,CAAC;AAAA,EACzD,kBAAkB;AAAA,EAClB,WAA2B;AAAA,IACzB,IAAI;AAAA,IACJ,OAAO;AAAA,IACP,UAAU;AAAA,IACV,MAAM;AAAA,IACN,SAAS,CAAC,EAAE,MAAM,QAAQ,MAAM,iCAAiC,CAAC;AAAA,IAClE,YAAY;AAAA,IACZ,OAAO,EAAE,aAAa,GAAG,cAAc,EAAE;AAAA,EAC3C;AAAA,EAEA,eAAe,QAAwB;AAAE,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAAG;AAAA,EAEtE,MAAM,MAAM,YAAoD;AAC9D,UAAM,QAAQ,KAAK,OAAO,UAAU,CAAC,cAAc,WAAW,SAAS,UAAU,KAAK,SAAS,CAAC;AAChG,WAAO,QAAQ,IAAI,OAAO,KAAK,OAAO,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,EAC1D;AAAA,EACA,MAAM,YAAsE;AAC1E,WAAO,EAAE,iBAAiB,KAAK,iBAAiB,WAAW,KAAK,IAAI,IAAI,IAAO;AAAA,EACjF;AAAA,EACA,MAAM,aAAa,YAAoB,UAAkB,QAA4C;AACnG,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAC5B;AAAA,EACA,MAAM,MAAM,YAAoB,UAAkB,SAAqE;AACrH,SAAK,kBAAkB,KAAK,OAAO;AACnC,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,SAAS;AAAA,QACP,UAAU,KAAK,SAAS;AAAA,QACxB,OAAO,KAAK,SAAS;AAAA,QACrB,eAAe;AAAA,QACf,aAAa,KAAK,SAAS,MAAM,eAAe;AAAA,QAChD,cAAc,KAAK,SAAS,MAAM,gBAAgB;AAAA,MACpD;AAAA,IACF;AAAA,EACF;AAAA,EACA,MAAM,SAAS,YAAoB,UAAkB,YAA8C;AACjG,SAAK,YAAY,KAAK,UAAU;AAAA,EAClC;AACF;","names":[]}
+{"version":3,"sources":["../src/testing.ts","../src/types.ts"],"sourcesContent":["import type { OracleResponse } from \"@odla-ai/ai\";\nimport {\n  HARNESS_PROTOCOL_VERSION,\n  type HarnessCompletion,\n  type HarnessControlPlane,\n  type HarnessEventInput,\n  type HarnessInferenceRequest,\n  type HarnessInferenceResponse,\n  type HarnessLease,\n} from \"./types\";\n\n/** Create a deterministic, valid lease for harness unit and integration tests. */\nexport function fixtureLease(overrides: Partial<HarnessLease> = {}): HarnessLease {\n  return {\n    protocolVersion: HARNESS_PROTOCOL_VERSION,\n    leaseId: \"lease_fixture\",\n    generation: 1,\n    expiresAt: Date.now() + 60_000,\n    task: {\n      taskId: \"task_fixture\",\n      attemptId: \"attempt_fixture\",\n      title: \"Fixture task\",\n      prompt: \"Create HARNESS_RESULT.md\",\n      workspace: \"fixture\",\n      aiRoute: \"coding\",\n      policy: {\n        network: \"none\",\n        timeoutMs: 30_000,\n        maxOutputBytes: 1_000_000,\n        maxPatchBytes: 1_000_000,\n      },\n    },\n    ...overrides,\n  };\n}\n\n/** In-memory control plane that records runner interactions without network access. */\nexport class FakeHarnessControlPlane implements HarnessControlPlane {\n  readonly leases: HarnessLease[] = [];\n  readonly events: HarnessEventInput[] = [];\n  readonly completions: HarnessCompletion[] = [];\n  readonly inferenceRequests: HarnessInferenceRequest[] = [];\n  cancelRequested = false;\n  response: OracleResponse = {\n    id: \"oracle_fixture\",\n    model: \"fixture-model\",\n    provider: \"openai\",\n    role: \"assistant\",\n    content: [{ type: \"text\", text: \"deterministic fixture response\" }],\n    stopReason: \"end_turn\",\n    usage: { inputTokens: 5, outputTokens: 4 },\n  };\n\n  constructor(...leases: HarnessLease[]) { this.leases.push(...leases); }\n\n  async lease(workspaces: string[]): Promise<HarnessLease | null> {\n    const index = this.leases.findIndex((candidate) => workspaces.includes(candidate.task.workspace));\n    return index < 0 ? null : this.leases.splice(index, 1)[0]!;\n  }\n  async heartbeat(): Promise<{ cancelRequested: boolean; expiresAt: number }> {\n    return { cancelRequested: this.cancelRequested, expiresAt: Date.now() + 60_000 };\n  }\n  async appendEvents(_attemptId: string, _leaseId: string, events: HarnessEventInput[]): Promise<void> {\n    this.events.push(...events);\n  }\n  async infer(_attemptId: string, _leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse> {\n    this.inferenceRequests.push(request);\n    return {\n      requestId: request.requestId,\n      response: this.response,\n      receipt: {\n        provider: this.response.provider,\n        model: this.response.model,\n        policyVersion: 1,\n        inputTokens: this.response.usage.inputTokens ?? 0,\n        outputTokens: this.response.usage.outputTokens ?? 0,\n      },\n    };\n  }\n  async complete(_attemptId: string, _leaseId: string, completion: HarnessCompletion): Promise<void> {\n    this.completions.push(completion);\n  }\n}\n","import type { ChatInput, OracleResponse } from \"@odla-ai/ai\";\nimport type { CodeToolPresentation } from \"./code-session-event-types\";\nexport type { CodeToolLocationPreview, CodeToolPresentation } from \"./code-session-event-types\";\n\n/** Current JSONL protocol version exchanged between a runner and an agent container. */\nexport const HARNESS_PROTOCOL_VERSION = 1 as const;\n\n/** Default control-plane route used for model inference requested by coding agents. */\nexport const DEFAULT_AI_ROUTE = \"coding\" as const;\n\n/** Lifecycle state reported for a harness task and its active attempt. */\nexport type HarnessTaskStatus =\n  | \"queued\"\n  | \"running\"\n  | \"cancel_requested\"\n  | \"completed\"\n  | \"failed\"\n  | \"cancelled\";\n\n/** Lifecycle state of one execution attempt for a task. */\nexport type HarnessAttemptStatus = HarnessTaskStatus;\n\n/** Trusted or untrusted participant that emitted a harness event. */\nexport type HarnessActor = \"operator\" | \"runner\" | \"agent\" | \"model\" | \"system\";\n\n/** Resource and isolation limits enforced while an untrusted task executes. */\nexport interface HarnessPolicy {\n  network: \"none\";\n  timeoutMs: number;\n  maxOutputBytes: number;\n  maxPatchBytes: number;\n}\n\n/** Immutable task instructions and execution policy delivered with a lease. */\nexport interface HarnessTaskSpec {\n  taskId: string;\n  attemptId: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  policy: HarnessPolicy;\n  parentAttemptId?: string | null;\n  checkpointSeq?: number | null;\n}\n\n/** Time-bound assignment authorizing a runner to execute one task attempt. */\nexport interface HarnessLease {\n  protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n  leaseId: string;\n  generation: number;\n  expiresAt: number;\n  task: HarnessTaskSpec;\n}\n\n/** Runner-supplied event before the control plane assigns sequence and task metadata. */\nexport interface HarnessEventInput {\n  eventId: string;\n  kind: string;\n  actor: HarnessActor;\n  payload: unknown;\n  createdAt: number;\n}\n\n/** Persisted, ordered event associated with a specific task attempt. */\nexport interface HarnessEvent extends HarnessEventInput {\n  seq: number;\n  taskId: string;\n  attemptId: string;\n}\n\n/** List-view metadata for a task and its current attempt. */\nexport interface HarnessTaskSummary {\n  taskId: string;\n  attemptId: string;\n  appId: string;\n  env: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  status: HarnessTaskStatus;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** List-view metadata for one attempt, including retry ancestry and runner ownership. */\nexport interface HarnessAttemptSummary {\n  attemptId: string;\n  taskId: string;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  status: HarnessAttemptStatus;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** Complete task view including attempts, events, result, and generated patch. */\nexport interface HarnessTaskDetail extends HarnessTaskSummary {\n  attempts: HarnessAttemptSummary[];\n  events: HarnessEvent[];\n  patch: string | null;\n  result: unknown;\n}\n\n/** Public control-plane view of a registered harness runner. */\nexport interface HarnessRunnerView {\n  runnerId: string;\n  appId: string;\n  env: string;\n  name: string;\n  createdAt: number;\n  lastSeenAt: number | null;\n  revokedAt: number | null;\n}\n\n/** Credential-free normalized model request sent from an agent through the runner. */\nexport interface HarnessInferenceRequest {\n  requestId: string;\n  /** Exact initial, follow-up, or resume command whose budget this call consumes. */\n  interactionId?: string;\n  call: ChatInput;\n}\n\n/** Normalized model response plus auditable provider, policy, and token metadata. */\nexport interface HarnessInferenceResponse {\n  requestId: string;\n  response: OracleResponse;\n  receipt: {\n    provider: string;\n    model: string;\n    policyVersion: number;\n    inputTokens: number;\n    outputTokens: number;\n    /**\n     * USD charged for this call, priced against the model the control plane\n     * actually resolved.\n     *\n     * ABSENT when the live catalog has no price for that model — never zero.\n     * An unpriced call is unknown spend, and reporting it as free is what let\n     * a goal's `maxUsd` look enforced while nothing enforced it. The runtime\n     * cannot compute this itself: it asks for `brokered` and only the control\n     * plane knows which model answered.\n     */\n    costUsd?: number;\n  };\n}\n\nexport type { CodeSessionEvent, CodeSessionEventData } from \"./code-session-event-data\";\n\n/** Closed set of effects an agent container may request from its trusted broker. */\nexport type HarnessToolName =\n  | \"sandbox.read\"\n  | \"sandbox.list\"\n  | \"sandbox.search\"\n  | \"sandbox.overview\"\n  | \"sandbox.where_is\"\n  | \"sandbox.who_imports\"\n  | \"sandbox.who_touches\"\n  | \"sandbox.apply_patch\"\n  | \"sandbox.edit\"\n  | \"sandbox.write\"\n  | \"sandbox.run_recipe\";\n/** Correlated, structured tool request emitted by an untrusted agent container. */\nexport interface HarnessToolRequest {\n  requestId: string;\n  tool: HarnessToolName;\n  input: Record<string, unknown>;\n}\n/** Bounded tool result returned to an agent after trusted policy evaluation. */\nexport interface HarnessToolResponse {\n  requestId: string;\n  ok: boolean;\n  content: string;\n  details?: Record<string, unknown>;\n}\n\n/** Validated JSONL message emitted by an untrusted agent container. */\nexport type HarnessAgentOutput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"event\";\n      kind: string;\n      payload?: unknown;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.request\";\n      requestId: string;\n      call: ChatInput;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.request\" } & HarnessToolRequest)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.complete\";\n      status: \"completed\" | \"failed\" | \"cancelled\";\n      result?: unknown;\n    };\n\n/** JSONL command or inference result written by the trusted runner to an agent. */\nexport type HarnessAgentInput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"task.start\";\n      task: HarnessTaskSpec;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.response\";\n      requestId: string;\n      response: OracleResponse;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.response\" } & HarnessToolResponse)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.cancel\";\n      reason: string;\n    };\n\n/** Terminal attempt report submitted by a runner to the control plane. */\nexport interface HarnessCompletion {\n  status: \"completed\" | \"failed\" | \"cancelled\";\n  result?: unknown;\n  patch?: string;\n  error?: string;\n}\n\n/** Operations a credentialed runner may perform against the harness control plane. */\nexport interface HarnessControlPlane {\n  lease(workspaces: string[]): Promise<HarnessLease | null>;\n  heartbeat(attemptId: string, leaseId: string): Promise<{ cancelRequested: boolean; expiresAt: number }>;\n  appendEvents(attemptId: string, leaseId: string, events: HarnessEventInput[]): Promise<void>;\n  infer(attemptId: string, leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n  complete(attemptId: string, leaseId: string, completion: HarnessCompletion): Promise<void>;\n}\n\n/** Trusted inference bridge used to keep model credentials outside agent containers. */\nexport interface HarnessAiConnection {\n  infer(request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n}\n\n/** Trusted tool boundary. Implementations must evaluate CaMeL policy before effects. */\nexport interface HarnessToolBroker {\n  execute(\n    context: { lease: HarnessLease; workspaceDir: string; signal?: AbortSignal },\n    request: HarnessToolRequest,\n  ): Promise<HarnessToolResponse>;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKO,IAAM,2BAA2B;;;ADOjC,SAAS,aAAa,YAAmC,CAAC,GAAiB;AAChF,SAAO;AAAA,IACL,iBAAiB;AAAA,IACjB,SAAS;AAAA,IACT,YAAY;AAAA,IACZ,WAAW,KAAK,IAAI,IAAI;AAAA,IACxB,MAAM;AAAA,MACJ,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,OAAO;AAAA,MACP,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,SAAS;AAAA,MACT,QAAQ;AAAA,QACN,SAAS;AAAA,QACT,WAAW;AAAA,QACX,gBAAgB;AAAA,QAChB,eAAe;AAAA,MACjB;AAAA,IACF;AAAA,IACA,GAAG;AAAA,EACL;AACF;AAGO,IAAM,0BAAN,MAA6D;AAAA,EACzD,SAAyB,CAAC;AAAA,EAC1B,SAA8B,CAAC;AAAA,EAC/B,cAAmC,CAAC;AAAA,EACpC,oBAA+C,CAAC;AAAA,EACzD,kBAAkB;AAAA,EAClB,WAA2B;AAAA,IACzB,IAAI;AAAA,IACJ,OAAO;AAAA,IACP,UAAU;AAAA,IACV,MAAM;AAAA,IACN,SAAS,CAAC,EAAE,MAAM,QAAQ,MAAM,iCAAiC,CAAC;AAAA,IAClE,YAAY;AAAA,IACZ,OAAO,EAAE,aAAa,GAAG,cAAc,EAAE;AAAA,EAC3C;AAAA,EAEA,eAAe,QAAwB;AAAE,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAAG;AAAA,EAEtE,MAAM,MAAM,YAAoD;AAC9D,UAAM,QAAQ,KAAK,OAAO,UAAU,CAAC,cAAc,WAAW,SAAS,UAAU,KAAK,SAAS,CAAC;AAChG,WAAO,QAAQ,IAAI,OAAO,KAAK,OAAO,OAAO,OAAO,CAAC,EAAE,CAAC;AAAA,EAC1D;AAAA,EACA,MAAM,YAAsE;AAC1E,WAAO,EAAE,iBAAiB,KAAK,iBAAiB,WAAW,KAAK,IAAI,IAAI,IAAO;AAAA,EACjF;AAAA,EACA,MAAM,aAAa,YAAoB,UAAkB,QAA4C;AACnG,SAAK,OAAO,KAAK,GAAG,MAAM;AAAA,EAC5B;AAAA,EACA,MAAM,MAAM,YAAoB,UAAkB,SAAqE;AACrH,SAAK,kBAAkB,KAAK,OAAO;AACnC,WAAO;AAAA,MACL,WAAW,QAAQ;AAAA,MACnB,UAAU,KAAK;AAAA,MACf,SAAS;AAAA,QACP,UAAU,KAAK,SAAS;AAAA,QACxB,OAAO,KAAK,SAAS;AAAA,QACrB,eAAe;AAAA,QACf,aAAa,KAAK,SAAS,MAAM,eAAe;AAAA,QAChD,cAAc,KAAK,SAAS,MAAM,gBAAgB;AAAA,MACpD;AAAA,IACF;AAAA,EACF;AAAA,EACA,MAAM,SAAS,YAAoB,UAAkB,YAA8C;AACjG,SAAK,YAAY,KAAK,UAAU;AAAA,EAClC;AACF;","names":[]}

package/dist/testing.d.cts

@@ -1,5 +1,5 @@
 import { OracleResponse } from '@odla-ai/ai';
-import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-D-Kv1svf.cjs';
+import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-DM-UgoBN.cjs';
 
 /** Create a deterministic, valid lease for harness unit and integration tests. */
 declare function fixtureLease(overrides?: Partial<HarnessLease>): HarnessLease;

package/dist/testing.d.ts

@@ -1,5 +1,5 @@
 import { OracleResponse } from '@odla-ai/ai';
-import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-D-Kv1svf.js';
+import { H as HarnessControlPlane, a as HarnessLease, b as HarnessEventInput, c as HarnessCompletion, d as HarnessInferenceRequest, e as HarnessInferenceResponse } from './types-DM-UgoBN.js';
 
 /** Create a deterministic, valid lease for harness unit and integration tests. */
 declare function fixtureLease(overrides?: Partial<HarnessLease>): HarnessLease;

package/dist/testing.js

@@ -1,6 +1,6 @@
 import {
   HARNESS_PROTOCOL_VERSION
-} from "./chunk-ONYW2VSB.js";
+} from "./chunk-O4QLHUGO.js";
 
 // src/testing.ts
 function fixtureLease(overrides = {}) {

package/dist/{types-D-Kv1svf.d.ts → types-DM-UgoBN.d.cts}

@@ -1,47 +1,5 @@
 import { OracleResponse, ChatInput } from '@odla-ai/ai';
 
-/** One bounded result location that an owner may inspect without loading the
- * complete broker result into the activity stream. */
-interface CodeToolLocationPreview {
-    path: string;
-    line?: number;
-    text?: string;
-}
-/** Deliberately small, tool-specific owner presentation. These projections are
- * built after the trusted broker has bounded the corresponding request/result;
- * they are not arbitrary model narration or complete tool output. */
-type CodeToolPresentation = {
-    kind: "query";
-    query?: string;
-    scope?: string;
-    count?: number;
-    results?: CodeToolLocationPreview[];
-    excerpt?: string;
-} | {
-    kind: "read";
-    path: string;
-    startLine?: number;
-    endLine?: number;
-    excerpt?: string;
-} | {
-    kind: "list";
-    scope?: string;
-    count?: number;
-    paths?: string[];
-} | {
-    kind: "patch";
-    paths?: string[];
-    additions?: number;
-    deletions?: number;
-} | {
-    kind: "recipe";
-    recipeId: string;
-    exitCode?: number;
-    timedOut?: boolean;
-    outputLimitExceeded?: boolean;
-    excerpt?: string;
-};
-
 /** Current JSONL protocol version exchanged between a runner and an agent container. */
 declare const HARNESS_PROTOCOL_VERSION: 1;
 /** Default control-plane route used for model inference requested by coding agents. */
@@ -170,81 +128,7 @@ interface HarnessInferenceResponse {
         costUsd?: number;
     };
 }
-/** Bounded, content-minimized session activity emitted by a Code runtime.
- * Message bodies are projected separately into the app's owner-private
- * odla-db chat. `interactionId` is optional so stored v1 events remain valid. */
-type CodeSessionEventData = ({
-    type: "message";
-    actor: "agent" | "system";
-    body: string;
-} | {
-    type: "diagnostic";
-    level: "error";
-    message: string;
-} | {
-    type: "thinking";
-    available: true;
-    durationMs: number;
-} | {
-    type: "tool";
-    phase: "started";
-    tool: HarnessToolName;
-    operationId?: string;
-    presentation?: CodeToolPresentation;
-} | {
-    type: "tool";
-    phase: "completed";
-    tool: HarnessToolName;
-    ok: boolean;
-    durationMs: number;
-    operationId?: string;
-    /** Why the call failed, bounded. Absent when `ok`. Without this a watcher
-     * saw that a tool failed and never why, which is what made an 84%
-     * apply_patch failure rate impossible to diagnose (PM bug 515655ec). */
-    failureReason?: string;
-    presentation?: CodeToolPresentation;
-} | {
-    type: "collaboration";
-    phase: "started";
-    skill: string;
-    tool: string;
-    operationId: string;
-} | {
-    type: "collaboration";
-    phase: "completed";
-    skill: string;
-    tool: string;
-    ok: boolean;
-    durationMs: number;
-    operationId: string;
-} | {
-    type: "usage";
-    provider: string;
-    model: string;
-    inputTokens: number;
-    outputTokens: number;
-    durationMs: number;
-    interactionTokens?: number;
-    interactionMaxTokens?: number;
-    /** USD for this call; absent when the model is unpriced, never zero. */
-    costUsd?: number;
-    /** Cumulative USD for this owner interaction, when every call in it was
-     *  priced. Absent the moment one was not, so a partial total can never be
-     *  mistaken for the whole. */
-    interactionCostUsd?: number;
-} | {
-    type: "status";
-    status: "running" | "idle" | "failed" | "checkpointed";
-    durationMs?: number;
-}) & {
-    interactionId?: string;
-};
-/** Registry-assigned cursor and timestamp for an owner-visible Code event. */
-type CodeSessionEvent = CodeSessionEventData & {
-    eventId: string;
-    sequence: number;
-    createdAt: number;
-};
+
 /** Closed set of effects an agent container may request from its trusted broker. */
 type HarnessToolName = "sandbox.read" | "sandbox.list" | "sandbox.search" | "sandbox.overview" | "sandbox.where_is" | "sandbox.who_imports" | "sandbox.who_touches" | "sandbox.apply_patch" | "sandbox.edit" | "sandbox.write" | "sandbox.run_recipe";
 /** Correlated, structured tool request emitted by an untrusted agent container. */
@@ -329,4 +213,4 @@ interface HarnessToolBroker {
     }, request: HarnessToolRequest): Promise<HarnessToolResponse>;
 }
 
-export { type CodeSessionEventData as C, DEFAULT_AI_ROUTE as D, type HarnessControlPlane as H, type HarnessLease as a, type HarnessEventInput as b, type HarnessCompletion as c, type HarnessInferenceRequest as d, type HarnessInferenceResponse as e, type HarnessToolBroker as f, type HarnessToolResponse as g, type HarnessAgentInput as h, type HarnessAgentOutput as i, type HarnessAiConnection as j, type CodeSessionEvent as k, type CodeToolLocationPreview as l, type CodeToolPresentation as m, HARNESS_PROTOCOL_VERSION as n, type HarnessActor as o, type HarnessAttemptStatus as p, type HarnessAttemptSummary as q, type HarnessEvent as r, type HarnessPolicy as s, type HarnessRunnerView as t, type HarnessTaskDetail as u, type HarnessTaskSpec as v, type HarnessTaskStatus as w, type HarnessTaskSummary as x, type HarnessToolName as y, type HarnessToolRequest as z };
+export { DEFAULT_AI_ROUTE as D, type HarnessControlPlane as H, type HarnessLease as a, type HarnessEventInput as b, type HarnessCompletion as c, type HarnessInferenceRequest as d, type HarnessInferenceResponse as e, type HarnessToolName as f, type HarnessToolBroker as g, type HarnessToolResponse as h, type HarnessAgentInput as i, type HarnessAgentOutput as j, type HarnessAiConnection as k, HARNESS_PROTOCOL_VERSION as l, type HarnessActor as m, type HarnessAttemptStatus as n, type HarnessAttemptSummary as o, type HarnessEvent as p, type HarnessPolicy as q, type HarnessRunnerView as r, type HarnessTaskDetail as s, type HarnessTaskSpec as t, type HarnessTaskStatus as u, type HarnessTaskSummary as v, type HarnessToolRequest as w };

package/dist/{types-D-Kv1svf.d.cts → types-DM-UgoBN.d.ts}

@@ -1,47 +1,5 @@
 import { OracleResponse, ChatInput } from '@odla-ai/ai';
 
-/** One bounded result location that an owner may inspect without loading the
- * complete broker result into the activity stream. */
-interface CodeToolLocationPreview {
-    path: string;
-    line?: number;
-    text?: string;
-}
-/** Deliberately small, tool-specific owner presentation. These projections are
- * built after the trusted broker has bounded the corresponding request/result;
- * they are not arbitrary model narration or complete tool output. */
-type CodeToolPresentation = {
-    kind: "query";
-    query?: string;
-    scope?: string;
-    count?: number;
-    results?: CodeToolLocationPreview[];
-    excerpt?: string;
-} | {
-    kind: "read";
-    path: string;
-    startLine?: number;
-    endLine?: number;
-    excerpt?: string;
-} | {
-    kind: "list";
-    scope?: string;
-    count?: number;
-    paths?: string[];
-} | {
-    kind: "patch";
-    paths?: string[];
-    additions?: number;
-    deletions?: number;
-} | {
-    kind: "recipe";
-    recipeId: string;
-    exitCode?: number;
-    timedOut?: boolean;
-    outputLimitExceeded?: boolean;
-    excerpt?: string;
-};
-
 /** Current JSONL protocol version exchanged between a runner and an agent container. */
 declare const HARNESS_PROTOCOL_VERSION: 1;
 /** Default control-plane route used for model inference requested by coding agents. */
@@ -170,81 +128,7 @@ interface HarnessInferenceResponse {
         costUsd?: number;
     };
 }
-/** Bounded, content-minimized session activity emitted by a Code runtime.
- * Message bodies are projected separately into the app's owner-private
- * odla-db chat. `interactionId` is optional so stored v1 events remain valid. */
-type CodeSessionEventData = ({
-    type: "message";
-    actor: "agent" | "system";
-    body: string;
-} | {
-    type: "diagnostic";
-    level: "error";
-    message: string;
-} | {
-    type: "thinking";
-    available: true;
-    durationMs: number;
-} | {
-    type: "tool";
-    phase: "started";
-    tool: HarnessToolName;
-    operationId?: string;
-    presentation?: CodeToolPresentation;
-} | {
-    type: "tool";
-    phase: "completed";
-    tool: HarnessToolName;
-    ok: boolean;
-    durationMs: number;
-    operationId?: string;
-    /** Why the call failed, bounded. Absent when `ok`. Without this a watcher
-     * saw that a tool failed and never why, which is what made an 84%
-     * apply_patch failure rate impossible to diagnose (PM bug 515655ec). */
-    failureReason?: string;
-    presentation?: CodeToolPresentation;
-} | {
-    type: "collaboration";
-    phase: "started";
-    skill: string;
-    tool: string;
-    operationId: string;
-} | {
-    type: "collaboration";
-    phase: "completed";
-    skill: string;
-    tool: string;
-    ok: boolean;
-    durationMs: number;
-    operationId: string;
-} | {
-    type: "usage";
-    provider: string;
-    model: string;
-    inputTokens: number;
-    outputTokens: number;
-    durationMs: number;
-    interactionTokens?: number;
-    interactionMaxTokens?: number;
-    /** USD for this call; absent when the model is unpriced, never zero. */
-    costUsd?: number;
-    /** Cumulative USD for this owner interaction, when every call in it was
-     *  priced. Absent the moment one was not, so a partial total can never be
-     *  mistaken for the whole. */
-    interactionCostUsd?: number;
-} | {
-    type: "status";
-    status: "running" | "idle" | "failed" | "checkpointed";
-    durationMs?: number;
-}) & {
-    interactionId?: string;
-};
-/** Registry-assigned cursor and timestamp for an owner-visible Code event. */
-type CodeSessionEvent = CodeSessionEventData & {
-    eventId: string;
-    sequence: number;
-    createdAt: number;
-};
+
 /** Closed set of effects an agent container may request from its trusted broker. */
 type HarnessToolName = "sandbox.read" | "sandbox.list" | "sandbox.search" | "sandbox.overview" | "sandbox.where_is" | "sandbox.who_imports" | "sandbox.who_touches" | "sandbox.apply_patch" | "sandbox.edit" | "sandbox.write" | "sandbox.run_recipe";
 /** Correlated, structured tool request emitted by an untrusted agent container. */
@@ -329,4 +213,4 @@ interface HarnessToolBroker {
     }, request: HarnessToolRequest): Promise<HarnessToolResponse>;
 }
 
-export { type CodeSessionEventData as C, DEFAULT_AI_ROUTE as D, type HarnessControlPlane as H, type HarnessLease as a, type HarnessEventInput as b, type HarnessCompletion as c, type HarnessInferenceRequest as d, type HarnessInferenceResponse as e, type HarnessToolBroker as f, type HarnessToolResponse as g, type HarnessAgentInput as h, type HarnessAgentOutput as i, type HarnessAiConnection as j, type CodeSessionEvent as k, type CodeToolLocationPreview as l, type CodeToolPresentation as m, HARNESS_PROTOCOL_VERSION as n, type HarnessActor as o, type HarnessAttemptStatus as p, type HarnessAttemptSummary as q, type HarnessEvent as r, type HarnessPolicy as s, type HarnessRunnerView as t, type HarnessTaskDetail as u, type HarnessTaskSpec as v, type HarnessTaskStatus as w, type HarnessTaskSummary as x, type HarnessToolName as y, type HarnessToolRequest as z };
+export { DEFAULT_AI_ROUTE as D, type HarnessControlPlane as H, type HarnessLease as a, type HarnessEventInput as b, type HarnessCompletion as c, type HarnessInferenceRequest as d, type HarnessInferenceResponse as e, type HarnessToolName as f, type HarnessToolBroker as g, type HarnessToolResponse as h, type HarnessAgentInput as i, type HarnessAgentOutput as j, type HarnessAiConnection as k, HARNESS_PROTOCOL_VERSION as l, type HarnessActor as m, type HarnessAttemptStatus as n, type HarnessAttemptSummary as o, type HarnessEvent as p, type HarnessPolicy as q, type HarnessRunnerView as r, type HarnessTaskDetail as s, type HarnessTaskSpec as t, type HarnessTaskStatus as u, type HarnessTaskSummary as v, type HarnessToolRequest as w };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@odla-ai/harness",
-  "version": "0.11.17",
+  "version": "0.11.19",
   "description": "Safe, inspectable coding-task protocol and credentialless container runner for odla Studio.",
   "license": "MIT",
   "homepage": "https://odla.ai/docs/packages/harness",

package/dist/chunk-ONYW2VSB.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/types.ts"],"sourcesContent":["import type { ChatInput, OracleResponse } from \"@odla-ai/ai\";\nimport type { CodeToolPresentation } from \"./code-session-event-types\";\nexport type { CodeToolLocationPreview, CodeToolPresentation } from \"./code-session-event-types\";\n\n/** Current JSONL protocol version exchanged between a runner and an agent container. */\nexport const HARNESS_PROTOCOL_VERSION = 1 as const;\n\n/** Default control-plane route used for model inference requested by coding agents. */\nexport const DEFAULT_AI_ROUTE = \"coding\" as const;\n\n/** Lifecycle state reported for a harness task and its active attempt. */\nexport type HarnessTaskStatus =\n  | \"queued\"\n  | \"running\"\n  | \"cancel_requested\"\n  | \"completed\"\n  | \"failed\"\n  | \"cancelled\";\n\n/** Lifecycle state of one execution attempt for a task. */\nexport type HarnessAttemptStatus = HarnessTaskStatus;\n\n/** Trusted or untrusted participant that emitted a harness event. */\nexport type HarnessActor = \"operator\" | \"runner\" | \"agent\" | \"model\" | \"system\";\n\n/** Resource and isolation limits enforced while an untrusted task executes. */\nexport interface HarnessPolicy {\n  network: \"none\";\n  timeoutMs: number;\n  maxOutputBytes: number;\n  maxPatchBytes: number;\n}\n\n/** Immutable task instructions and execution policy delivered with a lease. */\nexport interface HarnessTaskSpec {\n  taskId: string;\n  attemptId: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  policy: HarnessPolicy;\n  parentAttemptId?: string | null;\n  checkpointSeq?: number | null;\n}\n\n/** Time-bound assignment authorizing a runner to execute one task attempt. */\nexport interface HarnessLease {\n  protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n  leaseId: string;\n  generation: number;\n  expiresAt: number;\n  task: HarnessTaskSpec;\n}\n\n/** Runner-supplied event before the control plane assigns sequence and task metadata. */\nexport interface HarnessEventInput {\n  eventId: string;\n  kind: string;\n  actor: HarnessActor;\n  payload: unknown;\n  createdAt: number;\n}\n\n/** Persisted, ordered event associated with a specific task attempt. */\nexport interface HarnessEvent extends HarnessEventInput {\n  seq: number;\n  taskId: string;\n  attemptId: string;\n}\n\n/** List-view metadata for a task and its current attempt. */\nexport interface HarnessTaskSummary {\n  taskId: string;\n  attemptId: string;\n  appId: string;\n  env: string;\n  title: string;\n  prompt: string;\n  workspace: string;\n  aiRoute: string;\n  status: HarnessTaskStatus;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** List-view metadata for one attempt, including retry ancestry and runner ownership. */\nexport interface HarnessAttemptSummary {\n  attemptId: string;\n  taskId: string;\n  parentAttemptId: string | null;\n  checkpointSeq: number | null;\n  status: HarnessAttemptStatus;\n  runnerId: string | null;\n  generation: number;\n  createdAt: number;\n  updatedAt: number;\n}\n\n/** Complete task view including attempts, events, result, and generated patch. */\nexport interface HarnessTaskDetail extends HarnessTaskSummary {\n  attempts: HarnessAttemptSummary[];\n  events: HarnessEvent[];\n  patch: string | null;\n  result: unknown;\n}\n\n/** Public control-plane view of a registered harness runner. */\nexport interface HarnessRunnerView {\n  runnerId: string;\n  appId: string;\n  env: string;\n  name: string;\n  createdAt: number;\n  lastSeenAt: number | null;\n  revokedAt: number | null;\n}\n\n/** Credential-free normalized model request sent from an agent through the runner. */\nexport interface HarnessInferenceRequest {\n  requestId: string;\n  /** Exact initial, follow-up, or resume command whose budget this call consumes. */\n  interactionId?: string;\n  call: ChatInput;\n}\n\n/** Normalized model response plus auditable provider, policy, and token metadata. */\nexport interface HarnessInferenceResponse {\n  requestId: string;\n  response: OracleResponse;\n  receipt: {\n    provider: string;\n    model: string;\n    policyVersion: number;\n    inputTokens: number;\n    outputTokens: number;\n    /**\n     * USD charged for this call, priced against the model the control plane\n     * actually resolved.\n     *\n     * ABSENT when the live catalog has no price for that model — never zero.\n     * An unpriced call is unknown spend, and reporting it as free is what let\n     * a goal's `maxUsd` look enforced while nothing enforced it. The runtime\n     * cannot compute this itself: it asks for `brokered` and only the control\n     * plane knows which model answered.\n     */\n    costUsd?: number;\n  };\n}\n\n/** Bounded, content-minimized session activity emitted by a Code runtime.\n * Message bodies are projected separately into the app's owner-private\n * odla-db chat. `interactionId` is optional so stored v1 events remain valid. */\nexport type CodeSessionEventData = (\n  | { type: \"message\"; actor: \"agent\" | \"system\"; body: string }\n  | { type: \"diagnostic\"; level: \"error\"; message: string }\n  | { type: \"thinking\"; available: true; durationMs: number }\n  | {\n      type: \"tool\";\n      phase: \"started\";\n      tool: HarnessToolName;\n      operationId?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"tool\";\n      phase: \"completed\";\n      tool: HarnessToolName;\n      ok: boolean;\n      durationMs: number;\n      operationId?: string;\n      /** Why the call failed, bounded. Absent when `ok`. Without this a watcher\n       * saw that a tool failed and never why, which is what made an 84%\n       * apply_patch failure rate impossible to diagnose (PM bug 515655ec). */\n      failureReason?: string;\n      presentation?: CodeToolPresentation;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"started\";\n      skill: string;\n      tool: string;\n      operationId: string;\n    }\n  | {\n      type: \"collaboration\";\n      phase: \"completed\";\n      skill: string;\n      tool: string;\n      ok: boolean;\n      durationMs: number;\n      operationId: string;\n    }\n  | {\n      type: \"usage\"; provider: string; model: string;\n      inputTokens: number; outputTokens: number; durationMs: number;\n      interactionTokens?: number; interactionMaxTokens?: number;\n      /** USD for this call; absent when the model is unpriced, never zero. */\n      costUsd?: number;\n      /** Cumulative USD for this owner interaction, when every call in it was\n       *  priced. Absent the moment one was not, so a partial total can never be\n       *  mistaken for the whole. */\n      interactionCostUsd?: number;\n    }\n  | {\n      type: \"status\"; status: \"running\" | \"idle\" | \"failed\" | \"checkpointed\";\n      durationMs?: number;\n    }\n) & { interactionId?: string };\n\n/** Registry-assigned cursor and timestamp for an owner-visible Code event. */\nexport type CodeSessionEvent = CodeSessionEventData & {\n  eventId: string; sequence: number; createdAt: number;\n};\n\n/** Closed set of effects an agent container may request from its trusted broker. */\nexport type HarnessToolName =\n  | \"sandbox.read\"\n  | \"sandbox.list\"\n  | \"sandbox.search\"\n  | \"sandbox.overview\"\n  | \"sandbox.where_is\"\n  | \"sandbox.who_imports\"\n  | \"sandbox.who_touches\"\n  | \"sandbox.apply_patch\"\n  | \"sandbox.edit\"\n  | \"sandbox.write\"\n  | \"sandbox.run_recipe\";\n/** Correlated, structured tool request emitted by an untrusted agent container. */\nexport interface HarnessToolRequest {\n  requestId: string;\n  tool: HarnessToolName;\n  input: Record<string, unknown>;\n}\n/** Bounded tool result returned to an agent after trusted policy evaluation. */\nexport interface HarnessToolResponse {\n  requestId: string;\n  ok: boolean;\n  content: string;\n  details?: Record<string, unknown>;\n}\n\n/** Validated JSONL message emitted by an untrusted agent container. */\nexport type HarnessAgentOutput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"event\";\n      kind: string;\n      payload?: unknown;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.request\";\n      requestId: string;\n      call: ChatInput;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.request\" } & HarnessToolRequest)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.complete\";\n      status: \"completed\" | \"failed\" | \"cancelled\";\n      result?: unknown;\n    };\n\n/** JSONL command or inference result written by the trusted runner to an agent. */\nexport type HarnessAgentInput =\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"task.start\";\n      task: HarnessTaskSpec;\n    }\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"inference.response\";\n      requestId: string;\n      response: OracleResponse;\n    }\n  | ({ protocolVersion: typeof HARNESS_PROTOCOL_VERSION; type: \"tool.response\" } & HarnessToolResponse)\n  | {\n      protocolVersion: typeof HARNESS_PROTOCOL_VERSION;\n      type: \"attempt.cancel\";\n      reason: string;\n    };\n\n/** Terminal attempt report submitted by a runner to the control plane. */\nexport interface HarnessCompletion {\n  status: \"completed\" | \"failed\" | \"cancelled\";\n  result?: unknown;\n  patch?: string;\n  error?: string;\n}\n\n/** Operations a credentialed runner may perform against the harness control plane. */\nexport interface HarnessControlPlane {\n  lease(workspaces: string[]): Promise<HarnessLease | null>;\n  heartbeat(attemptId: string, leaseId: string): Promise<{ cancelRequested: boolean; expiresAt: number }>;\n  appendEvents(attemptId: string, leaseId: string, events: HarnessEventInput[]): Promise<void>;\n  infer(attemptId: string, leaseId: string, request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n  complete(attemptId: string, leaseId: string, completion: HarnessCompletion): Promise<void>;\n}\n\n/** Trusted inference bridge used to keep model credentials outside agent containers. */\nexport interface HarnessAiConnection {\n  infer(request: HarnessInferenceRequest): Promise<HarnessInferenceResponse>;\n}\n\n/** Trusted tool boundary. Implementations must evaluate CaMeL policy before effects. */\nexport interface HarnessToolBroker {\n  execute(\n    context: { lease: HarnessLease; workspaceDir: string; signal?: AbortSignal },\n    request: HarnessToolRequest,\n  ): Promise<HarnessToolResponse>;\n}\n"],"mappings":";AAKO,IAAM,2BAA2B;AAGjC,IAAM,mBAAmB;","names":[]}