@odla-ai/harness 0.5.2 → 0.7.0

package/dist/node.d.cts

@@ -349,12 +349,12 @@ interface FrozenCodeLocalSource {
 }
 
 /** v1's system prompt, verbatim. Do not edit without re-baselining the bench. */
-declare const V1_SYSTEM_PROMPT = "You are Pi, the coding agent inside an odla Code harness.\nUse only the odla_read, odla_apply_git_diff, and odla_run_recipe tools.\nFor mutations, call odla_apply_git_diff with raw git diff text. It must start\nwith \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" file\nheaders and numbered \"@@\" hunks, and never use \"*** Begin Patch\" wrappers.\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
+declare const V1_SYSTEM_PROMPT = "You are Theseus, the coding agent inside an odla Code harness.\nUse only the odla_read, odla_apply_git_diff, and odla_run_recipe tools.\nFor mutations, call odla_apply_git_diff with raw git diff text. It must start\nwith \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" file\nheaders and numbered \"@@\" hunks, and never use \"*** Begin Patch\" wrappers.\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
 /** The v2 prompt. v1's said "Use only the odla_read, odla_apply_git_diff, and
  *  odla_run_recipe tools", so leaving it in place would have told the model not
  *  to touch the tools this milestone exists to test. */
 declare const V2_SYSTEM_PROMPT = "You are the coding agent inside an odla Code harness.\nStart by orienting: odla_list shows the files in the workspace and odla_search\nfinds a literal string across them. Prefer those over guessing a path.\nThen odla_read a bounded range, and odla_apply_git_diff to mutate.\nFor mutations, call odla_apply_git_diff with raw git diff text. It must start\nwith \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" file\nheaders and numbered \"@@\" hunks, and never use \"*** Begin Patch\" wrappers.\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
-/** Which tool surface the agent sees. `v1` reproduces the Pi container's exact
+/** Which tool surface the agent sees. `v1` reproduces the Theseus container's exact
  *  three tools so the recorded baseline stays comparable. */
 type CodeSurface = "v1" | "v2" | "v3";
 /** v3 leads with orientation, because that is where the tokens went. */
@@ -363,7 +363,7 @@ declare const V3_SYSTEM_PROMPT = "You are the coding agent inside an odla Code h
  * The system prompt for each tool surface, keyed by version.
  *
  * Kept addressable so a benchmark can hold everything else fixed and vary only
- * the prompt — v1 is byte-identical to what the retired Pi agent shipped, which
+ * the prompt — v1 is byte-identical to what the retired Theseus agent shipped, which
  * is what makes the comparison against v2 and v3 meaningful.
  */
 declare const SYSTEM_PROMPT_FOR: Record<CodeSurface, string>;
@@ -372,6 +372,8 @@ interface CodeSkillOpts {
     broker: HarnessToolBroker;
     lease: HarnessLease;
     workspaceDir: string;
+    /** Exact CaMeL-registered recipes available to this attempt. */
+    recipeIds?: readonly string[];
     /** Defaults to v1 so an unqualified run reproduces the baseline. */
     surface?: CodeSurface;
     /** Records every brokered call so the report can attribute errors per tool. */
@@ -379,6 +381,7 @@ interface CodeSkillOpts {
         tool: HarnessToolName;
         ok: boolean;
         durationMs: number;
+        error?: string;
     }): void;
 }
 /**
@@ -398,6 +401,8 @@ interface CodeAgentAttemptOptions {
     prompt: string;
     surface?: CodeSurface;
     maxSteps?: number;
+    /** Exact registered build/test recipe IDs available to the runtime. */
+    recipeIds?: readonly string[];
     /** Run-wide ceiling enforced by runAgent against incremental usage. */
     budget?: AgentRunBudget;
     signal?: AbortSignal;
@@ -406,6 +411,7 @@ interface CodeAgentAttemptOptions {
         tool: HarnessToolName;
         ok: boolean;
         durationMs: number;
+        error?: string;
     }): void;
 }
 /** Whether one attempt completed, and the text it ended on. */
@@ -481,7 +487,7 @@ interface CodeToolBrokerOptions {
     maxPatchBytes?: number;
     maxRecipeWorkspaceFiles?: number;
     maxRecipeWorkspaceBytes?: number;
-    /** Paths Pi may read but may never modify. Prefixes are relative directory names. */
+    /** Paths Theseus may read but may never modify. Prefixes are relative directory names. */
     readOnlyPrefixes?: readonly string[];
     consumeApproval?(input: CodeToolDecision & {
         actionDigest: string;
@@ -489,8 +495,8 @@ interface CodeToolBrokerOptions {
     onDecision?(decision: CodeToolDecision): Promise<void> | void;
 }
 
-/** Trusted control plane, isolation engine, Pi image, recipes, and resource policy for a Code runtime. */
-interface CodePiRuntimeEngineOptions {
+/** Trusted control plane, isolation engine, recipes, and resource policy for the Theseus Code runtime. */
+interface TheseusRuntimeEngineOptions {
     control: CodeRuntimeAgentControlPlane;
     engine: ContainerEngine;
     recipes: readonly CodeBuildRecipe[];
@@ -505,11 +511,11 @@ interface CodePiRuntimeEngineOptions {
     localSource?: FrozenCodeLocalSource;
 }
 
-/** Execute fenced Code commands with real Pi in a credentialless, networkless container. */
-declare class CodePiRuntimeEngine implements CodeRuntimeCommandEngine {
+/** Execute fenced Theseus Code commands in a credentialless, networkless container. */
+declare class TheseusRuntimeEngine implements CodeRuntimeCommandEngine {
     #private;
     private readonly options;
-    constructor(options: CodePiRuntimeEngineOptions);
+    constructor(options: TheseusRuntimeEngineOptions);
     execute(command: CodeRuntimeCommand): Promise<CodeRuntimeCommandResult>;
     acknowledged(command: CodeRuntimeCommand, result: CodeRuntimeCommandResult): Promise<void>;
     close(): Promise<void>;
@@ -583,7 +589,7 @@ interface Options {
     fallbackPolicyDigest: `sha256:${string}`;
     event(command: CodeRuntimeCommand, event: CodeSessionEventData, refs: string[]): Promise<void>;
 }
-/** Own checkpoint evidence and candidate admission retries outside the Pi process lifecycle. */
+/** Own checkpoint evidence and candidate admission retries outside the Theseus process lifecycle. */
 declare class CodeRuntimeCheckpointManager {
     #private;
     private readonly options;
@@ -636,7 +642,7 @@ declare function materializeCommandWorkspace(input: {
     requestedLocal: CodeLocalSourceDescriptor | null;
 }>;
 
-/** Create the trusted CaMeL boundary for all Pi filesystem and build effects. */
+/** Create the trusted CaMeL boundary for all Theseus filesystem and build effects. */
 declare function createCodeToolBroker(options: CodeToolBrokerOptions): HarnessToolBroker;
 
 /** One brokered tool call, recorded for per-tool error attribution. */
@@ -644,6 +650,9 @@ interface CodeToolCallRecord {
     tool: HarnessToolName;
     ok: boolean;
     durationMs: number;
+    /** Why it failed, bounded, when it did — the same phrasing the model was
+     * given. Absent on success. */
+    error?: string;
 }
 /** One coding run: which tool surface, which model, and what it may spend. */
 interface RunCodeAgentOptions {
@@ -687,6 +696,8 @@ interface RunCodeAgentOptions {
      * of this package's install graph for a capability not every consumer wants.
      */
     extraSkills?: Skill[];
+    /** Exact registered build/test recipe IDs exposed in the tool schema. */
+    recipeIds?: readonly string[];
 }
 /** A finished run — its final text, why it stopped, and what it cost. */
 interface CodeAgentRun {
@@ -1329,4 +1340,4 @@ interface CodeVerificationEvidence {
 /** Rebuild a candidate from an exact trusted base and emit a prose-free clean-verifier receipt. */
 declare function verifyCodeCandidate(input: VerifyCodeCandidateInput): Promise<CodeVerificationEvidence>;
 
-export { CODE_RUNTIME_PROTOCOL_VERSION, type CodeAgentAttemptOptions, type CodeAgentAttemptResult, type CodeAgentRun, type CodeBuildRecipe, type CodeExpectedArtifact, type CodeLocalSourceDescriptor, type CodeMemory, type CodeMemoryEvidence, type CodeMemoryKind, type CodeMemoryStore, CodePiRuntimeEngine, type CodePiRuntimeEngineOptions, type CodeRecipeExecutor, type CodeRecipeResult, type CodeRuntimeAgentControlPlane, type CodeRuntimeBinding, type CodeRuntimeCandidateResponse, type CodeRuntimeCapabilities, CodeRuntimeCheckpointManager, type CodeRuntimeClientOptions, 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 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, type DecomposedRun, DecompositionError, 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, MAX_MEMORY_BODY, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, type RestoreCodeWorkspaceCheckpointInput, type RestoredCodeWorkspaceCheckpoint, type RunCodeAgentOptions, type RuntimeCheckpointSession, SYSTEM_PROMPT_FOR, type StageWorkspaceOptions, type StagedWorkspace, type StrategyChoice, type StrategySignals, type SubGoal, type SubGoalResult, V1_SYSTEM_PROMPT, V2_SYSTEM_PROMPT, V3_SYSTEM_PROMPT, type VerifyCodeCandidateInput, applyCodePatch, assertCodeBuildRecipe, assertDisjointPlan, assertPinnedImage, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, describePatchFailure, digestStagedWorkspace, feedbackIsActionable, hazardFromAttempt, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, outcomeCloses, outcomeMemory, patchPaths, planReachCollisions, prepareRuntimeCheckpoint, racedAttempt, recallAbout, registeredFiles, renderMemories, renderOutcome, resolveCodePath, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withRecipeDependencies };
+export { 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 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 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, type DecomposedRun, DecompositionError, 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, MAX_MEMORY_BODY, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, 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, assertCodeBuildRecipe, assertDisjointPlan, assertPinnedImage, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, describePatchFailure, digestStagedWorkspace, feedbackIsActionable, hazardFromAttempt, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, outcomeCloses, outcomeMemory, patchPaths, planReachCollisions, prepareRuntimeCheckpoint, racedAttempt, recallAbout, registeredFiles, renderMemories, renderOutcome, resolveCodePath, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withRecipeDependencies };

package/dist/node.d.ts

@@ -349,12 +349,12 @@ interface FrozenCodeLocalSource {
 }
 
 /** v1's system prompt, verbatim. Do not edit without re-baselining the bench. */
-declare const V1_SYSTEM_PROMPT = "You are Pi, the coding agent inside an odla Code harness.\nUse only the odla_read, odla_apply_git_diff, and odla_run_recipe tools.\nFor mutations, call odla_apply_git_diff with raw git diff text. It must start\nwith \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" file\nheaders and numbered \"@@\" hunks, and never use \"*** Begin Patch\" wrappers.\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
+declare const V1_SYSTEM_PROMPT = "You are Theseus, the coding agent inside an odla Code harness.\nUse only the odla_read, odla_apply_git_diff, and odla_run_recipe tools.\nFor mutations, call odla_apply_git_diff with raw git diff text. It must start\nwith \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" file\nheaders and numbered \"@@\" hunks, and never use \"*** Begin Patch\" wrappers.\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
 /** The v2 prompt. v1's said "Use only the odla_read, odla_apply_git_diff, and
  *  odla_run_recipe tools", so leaving it in place would have told the model not
  *  to touch the tools this milestone exists to test. */
 declare const V2_SYSTEM_PROMPT = "You are the coding agent inside an odla Code harness.\nStart by orienting: odla_list shows the files in the workspace and odla_search\nfinds a literal string across them. Prefer those over guessing a path.\nThen odla_read a bounded range, and odla_apply_git_diff to mutate.\nFor mutations, call odla_apply_git_diff with raw git diff text. It must start\nwith \"diff --git a/<path> b/<path>\", include matching \"---\" and \"+++\" file\nheaders and numbered \"@@\" hunks, and never use \"*** Begin Patch\" wrappers.\nThe workspace, model, and tool effects are controlled by the host broker.\nNever claim a build or test passed unless odla_run_recipe returned that result.";
-/** Which tool surface the agent sees. `v1` reproduces the Pi container's exact
+/** Which tool surface the agent sees. `v1` reproduces the Theseus container's exact
  *  three tools so the recorded baseline stays comparable. */
 type CodeSurface = "v1" | "v2" | "v3";
 /** v3 leads with orientation, because that is where the tokens went. */
@@ -363,7 +363,7 @@ declare const V3_SYSTEM_PROMPT = "You are the coding agent inside an odla Code h
  * The system prompt for each tool surface, keyed by version.
  *
  * Kept addressable so a benchmark can hold everything else fixed and vary only
- * the prompt — v1 is byte-identical to what the retired Pi agent shipped, which
+ * the prompt — v1 is byte-identical to what the retired Theseus agent shipped, which
  * is what makes the comparison against v2 and v3 meaningful.
  */
 declare const SYSTEM_PROMPT_FOR: Record<CodeSurface, string>;
@@ -372,6 +372,8 @@ interface CodeSkillOpts {
     broker: HarnessToolBroker;
     lease: HarnessLease;
     workspaceDir: string;
+    /** Exact CaMeL-registered recipes available to this attempt. */
+    recipeIds?: readonly string[];
     /** Defaults to v1 so an unqualified run reproduces the baseline. */
     surface?: CodeSurface;
     /** Records every brokered call so the report can attribute errors per tool. */
@@ -379,6 +381,7 @@ interface CodeSkillOpts {
         tool: HarnessToolName;
         ok: boolean;
         durationMs: number;
+        error?: string;
     }): void;
 }
 /**
@@ -398,6 +401,8 @@ interface CodeAgentAttemptOptions {
     prompt: string;
     surface?: CodeSurface;
     maxSteps?: number;
+    /** Exact registered build/test recipe IDs available to the runtime. */
+    recipeIds?: readonly string[];
     /** Run-wide ceiling enforced by runAgent against incremental usage. */
     budget?: AgentRunBudget;
     signal?: AbortSignal;
@@ -406,6 +411,7 @@ interface CodeAgentAttemptOptions {
         tool: HarnessToolName;
         ok: boolean;
         durationMs: number;
+        error?: string;
     }): void;
 }
 /** Whether one attempt completed, and the text it ended on. */
@@ -481,7 +487,7 @@ interface CodeToolBrokerOptions {
     maxPatchBytes?: number;
     maxRecipeWorkspaceFiles?: number;
     maxRecipeWorkspaceBytes?: number;
-    /** Paths Pi may read but may never modify. Prefixes are relative directory names. */
+    /** Paths Theseus may read but may never modify. Prefixes are relative directory names. */
     readOnlyPrefixes?: readonly string[];
     consumeApproval?(input: CodeToolDecision & {
         actionDigest: string;
@@ -489,8 +495,8 @@ interface CodeToolBrokerOptions {
     onDecision?(decision: CodeToolDecision): Promise<void> | void;
 }
 
-/** Trusted control plane, isolation engine, Pi image, recipes, and resource policy for a Code runtime. */
-interface CodePiRuntimeEngineOptions {
+/** Trusted control plane, isolation engine, recipes, and resource policy for the Theseus Code runtime. */
+interface TheseusRuntimeEngineOptions {
     control: CodeRuntimeAgentControlPlane;
     engine: ContainerEngine;
     recipes: readonly CodeBuildRecipe[];
@@ -505,11 +511,11 @@ interface CodePiRuntimeEngineOptions {
     localSource?: FrozenCodeLocalSource;
 }
 
-/** Execute fenced Code commands with real Pi in a credentialless, networkless container. */
-declare class CodePiRuntimeEngine implements CodeRuntimeCommandEngine {
+/** Execute fenced Theseus Code commands in a credentialless, networkless container. */
+declare class TheseusRuntimeEngine implements CodeRuntimeCommandEngine {
     #private;
     private readonly options;
-    constructor(options: CodePiRuntimeEngineOptions);
+    constructor(options: TheseusRuntimeEngineOptions);
     execute(command: CodeRuntimeCommand): Promise<CodeRuntimeCommandResult>;
     acknowledged(command: CodeRuntimeCommand, result: CodeRuntimeCommandResult): Promise<void>;
     close(): Promise<void>;
@@ -583,7 +589,7 @@ interface Options {
     fallbackPolicyDigest: `sha256:${string}`;
     event(command: CodeRuntimeCommand, event: CodeSessionEventData, refs: string[]): Promise<void>;
 }
-/** Own checkpoint evidence and candidate admission retries outside the Pi process lifecycle. */
+/** Own checkpoint evidence and candidate admission retries outside the Theseus process lifecycle. */
 declare class CodeRuntimeCheckpointManager {
     #private;
     private readonly options;
@@ -636,7 +642,7 @@ declare function materializeCommandWorkspace(input: {
     requestedLocal: CodeLocalSourceDescriptor | null;
 }>;
 
-/** Create the trusted CaMeL boundary for all Pi filesystem and build effects. */
+/** Create the trusted CaMeL boundary for all Theseus filesystem and build effects. */
 declare function createCodeToolBroker(options: CodeToolBrokerOptions): HarnessToolBroker;
 
 /** One brokered tool call, recorded for per-tool error attribution. */
@@ -644,6 +650,9 @@ interface CodeToolCallRecord {
     tool: HarnessToolName;
     ok: boolean;
     durationMs: number;
+    /** Why it failed, bounded, when it did — the same phrasing the model was
+     * given. Absent on success. */
+    error?: string;
 }
 /** One coding run: which tool surface, which model, and what it may spend. */
 interface RunCodeAgentOptions {
@@ -687,6 +696,8 @@ interface RunCodeAgentOptions {
      * of this package's install graph for a capability not every consumer wants.
      */
     extraSkills?: Skill[];
+    /** Exact registered build/test recipe IDs exposed in the tool schema. */
+    recipeIds?: readonly string[];
 }
 /** A finished run — its final text, why it stopped, and what it cost. */
 interface CodeAgentRun {
@@ -1329,4 +1340,4 @@ interface CodeVerificationEvidence {
 /** Rebuild a candidate from an exact trusted base and emit a prose-free clean-verifier receipt. */
 declare function verifyCodeCandidate(input: VerifyCodeCandidateInput): Promise<CodeVerificationEvidence>;
 
-export { CODE_RUNTIME_PROTOCOL_VERSION, type CodeAgentAttemptOptions, type CodeAgentAttemptResult, type CodeAgentRun, type CodeBuildRecipe, type CodeExpectedArtifact, type CodeLocalSourceDescriptor, type CodeMemory, type CodeMemoryEvidence, type CodeMemoryKind, type CodeMemoryStore, CodePiRuntimeEngine, type CodePiRuntimeEngineOptions, type CodeRecipeExecutor, type CodeRecipeResult, type CodeRuntimeAgentControlPlane, type CodeRuntimeBinding, type CodeRuntimeCandidateResponse, type CodeRuntimeCapabilities, CodeRuntimeCheckpointManager, type CodeRuntimeClientOptions, 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 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, type DecomposedRun, DecompositionError, 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, MAX_MEMORY_BODY, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, type RestoreCodeWorkspaceCheckpointInput, type RestoredCodeWorkspaceCheckpoint, type RunCodeAgentOptions, type RuntimeCheckpointSession, SYSTEM_PROMPT_FOR, type StageWorkspaceOptions, type StagedWorkspace, type StrategyChoice, type StrategySignals, type SubGoal, type SubGoalResult, V1_SYSTEM_PROMPT, V2_SYSTEM_PROMPT, V3_SYSTEM_PROMPT, type VerifyCodeCandidateInput, applyCodePatch, assertCodeBuildRecipe, assertDisjointPlan, assertPinnedImage, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, describePatchFailure, digestStagedWorkspace, feedbackIsActionable, hazardFromAttempt, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, outcomeCloses, outcomeMemory, patchPaths, planReachCollisions, prepareRuntimeCheckpoint, racedAttempt, recallAbout, registeredFiles, renderMemories, renderOutcome, resolveCodePath, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withRecipeDependencies };
+export { 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 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 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, type DecomposedRun, DecompositionError, 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, MAX_MEMORY_BODY, MEASURED_PREMIUM, type MaterializedCodeSource, type MaterializedGitTree, type NewCodeMemory, type PrepareRuntimeCheckpointInput, type PreparedRuntimeCheckpoint, type RacedAttemptOptions, type RacedOutcome, type RecallOptions, type RecipeDependencies, 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, assertCodeBuildRecipe, assertDisjointPlan, assertPinnedImage, attachCodeRuntimeReferences, buildContainerRunArgs, buildRecipeContainerArgs, chooseStrategy, codeSkill, createCodeRuntimeControlClient, createCodeRuntimeInference, createCodeToolBroker, createCodeWorkspaceCheckpoint, createContainerRecipeExecutor, describePatchFailure, digestStagedWorkspace, feedbackIsActionable, hazardFromAttempt, installedDependencies, integrateSubGoals, isCheckpointEffectCompleted, materializeCodeRuntimeSource, materializeCommandWorkspace, materializeGitTree, outcomeCloses, outcomeMemory, patchPaths, planReachCollisions, prepareRuntimeCheckpoint, racedAttempt, recallAbout, registeredFiles, renderMemories, renderOutcome, resolveCodePath, restoreCodeWorkspaceCheckpoint, runCodeAgent, runCodeAgentAttempt, runCodeRuntimeHeartbeatLoop, runContainerAttempt, runGoal, runHarnessRunner, runLeasedAttempt, runtimeMemoryStore, safeWorkspaceLabel, selectContainerEngine, selectWinner, stageWorkspace, stageWorkspacePair, straySubGoalFiles, stripPatchEnvelope, validateCodePatch, validateMemory, validateRelativePath, verifyCodeCandidate, verifyContainerEngineBoundary, withRecipeDependencies };

package/dist/node.js

@@ -4,12 +4,12 @@ import {
 } from "./chunk-XWXRNPGY.js";
 import {
   CODE_RUNTIME_PROTOCOL_VERSION,
-  CodePiRuntimeEngine,
   CodeRuntimeCheckpointManager,
   CodeRuntimeControlError,
   CodeRuntimeReconciler,
   MAX_MEMORY_BODY,
   SYSTEM_PROMPT_FOR,
+  TheseusRuntimeEngine,
   V1_SYSTEM_PROMPT,
   V2_SYSTEM_PROMPT,
   V3_SYSTEM_PROMPT,
@@ -44,7 +44,7 @@ import {
   validateMemory,
   validateRelativePath,
   verifyCodeCandidate
-} from "./chunk-5HX5LWTG.js";
+} from "./chunk-QZXCQSPZ.js";
 import {
   assertPinnedImage,
   buildContainerRunArgs,
@@ -361,7 +361,6 @@ async function installedDependencies(repoRoot) {
 }
 export {
   CODE_RUNTIME_PROTOCOL_VERSION,
-  CodePiRuntimeEngine,
   CodeRuntimeCheckpointManager,
   CodeRuntimeControlError,
   CodeRuntimeReconciler,
@@ -369,6 +368,7 @@ export {
   MAX_MEMORY_BODY,
   MEASURED_PREMIUM,
   SYSTEM_PROMPT_FOR,
+  TheseusRuntimeEngine,
   V1_SYSTEM_PROMPT,
   V2_SYSTEM_PROMPT,
   V3_SYSTEM_PROMPT,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@odla-ai/harness",
-  "version": "0.5.2",
+  "version": "0.7.0",
   "description": "Safe, inspectable coding-task protocol and credentialless container runner for odla Studio.",
   "license": "MIT",
   "homepage": "https://odla.ai/docs/packages/harness",