stable-harness 0.0.80 → 0.0.81

package/dist/index.d.ts

@@ -1,5 +1,5 @@
 import { createStableHarnessRuntime as createCoreStableHarnessRuntime } from "@stable-harness/core";
-import type { CompiledWorkspace, RuntimeAdapter, RuntimeRequest, RuntimeToolGateway, RuntimeWorkflowAdapter, StableHarnessRuntime, WorkspaceAdapterPolicy } from "@stable-harness/core";
+import type { CompiledWorkspace, ExecutionEvaluatorRule, RuntimeAdapter, RuntimeRequest, RuntimeToolGateway, RuntimeWorkflowAdapter, StableHarnessRuntime, ToolGuardrail, WorkspaceAdapterPolicy } from "@stable-harness/core";
 import type { RuntimeMemoryStore } from "@stable-harness/memory";
 export { createDeepAgentsAdapter, createDeepAgentsMemoryMaintenanceTarget } from "@stable-harness/adapter-deepagents";
 export { createDeepAgentsMiddlewareSkillProvider, createLangGraphRuntimeAdapter, createLangGraphWorkflowAdapter, createRegistrySkillResolverProvider, } from "@stable-harness/adapter-langgraph";
@@ -7,12 +7,12 @@ export type { LangGraphNodeHandler, LangGraphNodeHandlerInput, LangGraphNodeReso
 export type { LangGraphRegistrySkillOutput } from "@stable-harness/adapter-langgraph";
 export { createLangMemServiceProvider } from "@stable-harness/memory";
 export { createInMemoryRuntimeMemoryStore, createJsonFileRuntimeMemoryStore } from "@stable-harness/memory";
-export { applySpecDrivenPhaseTransition, containsRecoverableResultOutput, createSpecDrivenArtifact, createSpecDrivenArtifactEvent, createSpecDrivenPhaseEvent, createSpecDrivenWorkflowPolicy, createSpecDrivenWorkflowState, projectRuntimeTrace, resolveEnabledMemories, } from "@stable-harness/core";
-export type { CompiledWorkspace, RuntimeAdapter, RuntimeEvent, RuntimeWorkflowAdapter, RuntimeRequest, RuntimeResponse, RuntimeRunRecord, RuntimeTraceEntry, StableHarnessRuntime, SpecDrivenPhaseRecord, SpecDrivenPhaseStatus, SpecDrivenPhaseTransition, SpecDrivenWorkflowState, WorkspaceAgent, WorkspaceModel, WorkspaceRuntimePolicy, WorkspaceSpecDrivenPhase, WorkspaceSpecDrivenWorkflowPolicy, WorkspaceTool, } from "@stable-harness/core";
+export { applySpecDrivenPhaseTransition, containsRecoverableResultOutput, createSpecDrivenArtifact, createSpecDrivenArtifactEvent, createSpecDrivenPhaseEvent, createSpecDrivenWorkflowPolicy, createSpecDrivenWorkflowState, defaultExecutionEvaluatorRules, defaultToolGuardrails, evaluateExecutionRules, evaluateToolGuardrails, projectRuntimeTrace, repeatToolGuardrail, resolveEnabledMemories, requiredPlanToolGuardrail, reviewExecutionEvidence, toolDependencyGuardrail, } from "@stable-harness/core";
+export type { CompiledWorkspace, ExecutionEvaluatorRule, RuntimeAdapter, RuntimeEvent, RuntimeWorkflowAdapter, RuntimeRequest, RuntimeResponse, RuntimeRunRecord, RuntimeTraceEntry, StableHarnessRuntime, SpecDrivenPhaseRecord, SpecDrivenPhaseStatus, SpecDrivenPhaseTransition, SpecDrivenWorkflowState, WorkspaceAgent, WorkspaceModel, WorkspaceRuntimePolicy, WorkspaceSpecDrivenPhase, WorkspaceSpecDrivenWorkflowPolicy, WorkspaceTool, ToolGuardrail, ToolGuardrailContext, ToolGuardrailDecision, } from "@stable-harness/core";
 export { loadWorkspaceFromYaml } from "@stable-harness/workspace-yaml";
 export { createInMemoryToolGateway, createModuleToolGateway } from "@stable-harness/tool-gateway";
 export type { ModuleToolDescriptor, ToolGateway, ToolGatewayContext, ToolGatewayInvokeRequest, ToolGatewayInvokeResult, ToolGatewayTool, } from "@stable-harness/tool-gateway";
-type RuntimeAssemblyInput = {
+export type RuntimeAssemblyInput = {
     workspaceRoot: string;
     adapters?: RuntimeAdapter[];
     adapterFactories?: Record<string, RuntimeAdapterFactory>;
@@ -20,6 +20,8 @@ type RuntimeAssemblyInput = {
     workflowAdapterFactories?: Record<string, RuntimeWorkflowAdapterFactory>;
     workflowAdapterOptions?: Record<string, unknown>;
     toolGateway?: RuntimeToolGateway;
+    toolGuardrails?: readonly ToolGuardrail[];
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     memory?: RuntimeMemoryStore;
 };
 type RuntimeAdapterFactory = (input: {

package/dist/index.js

@@ -1 +1 @@
-import{createBackendModel as e,createDeepAgentsAdapter as r}from"@stable-harness/adapter-deepagents";import{createLangGraphRuntimeAdapter as t,createLangGraphWorkflowAdapter as a,createRegistrySkillResolverProvider as o}from"@stable-harness/adapter-langgraph";import{createStableHarnessRuntime as n}from"@stable-harness/core";import{createModuleToolGateway as i}from"@stable-harness/tool-gateway";import{loadWorkspaceFromYaml as s}from"@stable-harness/workspace-yaml";export{createDeepAgentsAdapter,createDeepAgentsMemoryMaintenanceTarget}from"@stable-harness/adapter-deepagents";export{createDeepAgentsMiddlewareSkillProvider,createLangGraphRuntimeAdapter,createLangGraphWorkflowAdapter,createRegistrySkillResolverProvider}from"@stable-harness/adapter-langgraph";export{createLangMemServiceProvider}from"@stable-harness/memory";export{createInMemoryRuntimeMemoryStore,createJsonFileRuntimeMemoryStore}from"@stable-harness/memory";export{applySpecDrivenPhaseTransition,containsRecoverableResultOutput,createSpecDrivenArtifact,createSpecDrivenArtifactEvent,createSpecDrivenPhaseEvent,createSpecDrivenWorkflowPolicy,createSpecDrivenWorkflowState,projectRuntimeTrace,resolveEnabledMemories}from"@stable-harness/core";export{loadWorkspaceFromYaml}from"@stable-harness/workspace-yaml";export{createInMemoryToolGateway,createModuleToolGateway}from"@stable-harness/tool-gateway";export function createStableHarnessRuntime(e){return"string"==typeof e?createStableRuntime({workspaceRoot:e}):"workspaceRoot"in e?createStableRuntime(e):n(e)}export async function createStableRuntime(e){const r=await s(e.workspaceRoot),t=e.toolGateway??await i({tools:r.tools.values()});return n({workspace:r,toolGateway:t,memory:e.memory,qualityReviewModel:createQualityReviewModel(r),adapters:e.adapters??createRuntimeAdapters(r,e),workflowAdapters:e.workflowAdapters??createWorkflowAdapters(r,e)})}function createQualityReviewModel(r){const t=function readQualityModelRef(e){const r=isRecord(e)?e:{};return readString((isRecord(r.reviewer)?r.reviewer:r).modelRef)}(r.runtime.quality),a=t?r.models.get(t):void 0,o=a?e(a):void 0;return function isQualityReviewModel(e){return isRecord(e)&&"function"==typeof e.invoke}(o)?o:void 0}export async function requestStableRuntime(e,r){return e.request(r)}function createRuntimeAdapters(e,a){const o={deepagents:({policy:e})=>r(e.config?{config:e.config}:{}),langgraph:({policy:e})=>t({...readLangGraphOptions(e.config),name:e.name}),...a.adapterFactories},n=function runtimeAdapterPolicies(e){const r=e.runtime.adapters?.filter(e=>!1!==e.enabled);return r&&r.length>0?r:[...new Set([...e.agents.values()].map(e=>e.backend))].map(e=>({name:e}))}(e);return n.map(r=>{const t=o[r.name];if(t)return t({policy:r,workspace:e});throw new Error(`Unsupported runtime adapter: ${r.name}`)})}function createWorkflowAdapters(e,r){const t={langgraph:({name:e,options:r})=>a({...readLangGraphOptions(r),name:e}),...r.workflowAdapterFactories};return[...new Set([...e.workflows.values()].map(e=>e.adapter??"").filter(Boolean))].map(a=>{const o=t[a];return o?.({name:a,workspace:e,options:readWorkflowAdapterOptions(r,a)})}).filter(e=>Boolean(e))}function readWorkflowAdapterOptions(e,r){return e.workflowAdapterOptions?.[r]??{}}function readLangGraphOptions(e){return isRecord(e)?{...e,...void 0!==readLangGraphSkillProvider(e)?{skillProvider:readLangGraphSkillProvider(e)}:{}}:{}}function readLangGraphSkillProvider(e){if(!1===e.skillProvider)return!1;const r=function readSkillProviderConfig(e){return isRecord(e.skills)?e.skills:isRecord(e.skillProvider)?e.skillProvider:void 0}(e);if(!r)return;const t=readString(r.provider)??readString(r.name)??"registry-resolver";if(["none","disabled","false"].includes(t))return!1;if("registry-resolver"!==t)throw new Error(`Unsupported LangGraph skill provider: ${t}`);return o({..."boolean"==typeof r.includeContent?{includeContent:r.includeContent}:{},..."number"==typeof r.maxBytes&&Number.isFinite(r.maxBytes)?{maxBytes:r.maxBytes}:{}})}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}
+import{createBackendModel as e,createDeepAgentsAdapter as r}from"@stable-harness/adapter-deepagents";import{createLangGraphRuntimeAdapter as a,createLangGraphWorkflowAdapter as t,createRegistrySkillResolverProvider as o}from"@stable-harness/adapter-langgraph";import{createStableHarnessRuntime as n}from"@stable-harness/core";import{createModuleToolGateway as i}from"@stable-harness/tool-gateway";import{loadWorkspaceFromYaml as s}from"@stable-harness/workspace-yaml";export{createDeepAgentsAdapter,createDeepAgentsMemoryMaintenanceTarget}from"@stable-harness/adapter-deepagents";export{createDeepAgentsMiddlewareSkillProvider,createLangGraphRuntimeAdapter,createLangGraphWorkflowAdapter,createRegistrySkillResolverProvider}from"@stable-harness/adapter-langgraph";export{createLangMemServiceProvider}from"@stable-harness/memory";export{createInMemoryRuntimeMemoryStore,createJsonFileRuntimeMemoryStore}from"@stable-harness/memory";export{applySpecDrivenPhaseTransition,containsRecoverableResultOutput,createSpecDrivenArtifact,createSpecDrivenArtifactEvent,createSpecDrivenPhaseEvent,createSpecDrivenWorkflowPolicy,createSpecDrivenWorkflowState,defaultExecutionEvaluatorRules,defaultToolGuardrails,evaluateExecutionRules,evaluateToolGuardrails,projectRuntimeTrace,repeatToolGuardrail,resolveEnabledMemories,requiredPlanToolGuardrail,reviewExecutionEvidence,toolDependencyGuardrail}from"@stable-harness/core";export{loadWorkspaceFromYaml}from"@stable-harness/workspace-yaml";export{createInMemoryToolGateway,createModuleToolGateway}from"@stable-harness/tool-gateway";export function createStableHarnessRuntime(e){return"string"==typeof e?createStableRuntime({workspaceRoot:e}):"workspaceRoot"in e?createStableRuntime(e):n(e)}export async function createStableRuntime(e){const r=await s(e.workspaceRoot),a=e.toolGateway??await i({tools:r.tools.values()});return n({workspace:r,toolGateway:a,memory:e.memory,qualityReviewModel:createQualityReviewModel(r),toolGuardrails:e.toolGuardrails,executionEvaluatorRules:e.executionEvaluatorRules,adapters:e.adapters??createRuntimeAdapters(r,e),workflowAdapters:e.workflowAdapters??createWorkflowAdapters(r,e)})}function createQualityReviewModel(r){const a=function readQualityModelRef(e){const r=isRecord(e)?e:{};return readString((isRecord(r.reviewer)?r.reviewer:r).modelRef)}(r.runtime.quality),t=a?r.models.get(a):void 0,o=t?e(t):void 0;return function isQualityReviewModel(e){return isRecord(e)&&"function"==typeof e.invoke}(o)?o:void 0}export async function requestStableRuntime(e,r){return e.request(r)}function createRuntimeAdapters(e,t){const o={deepagents:({policy:e})=>r(e.config?{config:e.config}:{}),langgraph:({policy:e})=>a({...readLangGraphOptions(e.config),name:e.name}),...t.adapterFactories},n=function runtimeAdapterPolicies(e){const r=e.runtime.adapters?.filter(e=>!1!==e.enabled);return r&&r.length>0?r:[...new Set([...e.agents.values()].map(e=>e.backend))].map(e=>({name:e}))}(e);return n.map(r=>{const a=o[r.name];if(a)return a({policy:r,workspace:e});throw new Error(`Unsupported runtime adapter: ${r.name}`)})}function createWorkflowAdapters(e,r){const a={langgraph:({name:e,options:r})=>t({...readLangGraphOptions(r),name:e}),...r.workflowAdapterFactories};return[...new Set([...e.workflows.values()].map(e=>e.adapter??"").filter(Boolean))].map(t=>{const o=a[t];return o?.({name:t,workspace:e,options:readWorkflowAdapterOptions(r,t)})}).filter(e=>Boolean(e))}function readWorkflowAdapterOptions(e,r){return e.workflowAdapterOptions?.[r]??{}}function readLangGraphOptions(e){return isRecord(e)?{...e,...void 0!==readLangGraphSkillProvider(e)?{skillProvider:readLangGraphSkillProvider(e)}:{}}:{}}function readLangGraphSkillProvider(e){if(!1===e.skillProvider)return!1;const r=function readSkillProviderConfig(e){return isRecord(e.skills)?e.skills:isRecord(e.skillProvider)?e.skillProvider:void 0}(e);if(!r)return;const a=readString(r.provider)??readString(r.name)??"registry-resolver";if(["none","disabled","false"].includes(a))return!1;if("registry-resolver"!==a)throw new Error(`Unsupported LangGraph skill provider: ${a}`);return o({..."boolean"==typeof r.includeContent?{includeContent:r.includeContent}:{},..."number"==typeof r.maxBytes&&Number.isFinite(r.maxBytes)?{maxBytes:r.maxBytes}:{}})}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/docs/architecture/runtime-controls.md

@@ -0,0 +1,154 @@
+# Runtime Controls
+
+Stable Harness keeps model choice soft and runtime controls hard. Prompts,
+skills, and specialist descriptions can guide behavior, while guardrails and
+evaluators enforce behavior before tool execution and before final delivery.
+
+## Tool guardrails
+
+Tool guardrails run before gateway tools are invoked. The same pipeline applies
+to model-visible DeepAgents gateway calls and explicit runtime `toolCall`
+requests. Each guardrail receives structured runtime context and either returns
+a blocking decision or `undefined` to let the next guardrail run.
+
+```ts
+import {
+  createStableHarnessRuntime,
+  defaultToolGuardrails,
+  type ToolGuardrail,
+} from "stable-harness";
+
+const approvalGuardrail: ToolGuardrail = (context) => {
+  if (context.toolId !== "deploy_service") return undefined;
+  return {
+    status: "approval_required",
+    reason: "deploy_service requires operator approval",
+    eventOutput: JSON.stringify({
+      status: "approval_required",
+      toolId: context.toolId,
+      reason: "deploy_service requires operator approval",
+    }),
+    modelOutput: "Status: approval_required\nThis tool requires operator approval before execution.",
+  };
+};
+
+const runtime = createStableHarnessRuntime({
+  workspace,
+  adapters,
+  toolGateway,
+  toolGuardrails: [
+    approvalGuardrail,
+    ...defaultToolGuardrails,
+  ],
+});
+```
+
+The default guardrail pipeline is:
+
+- `requiredPlanToolGuardrail`
+- `toolDependencyGuardrail`
+- `repeatToolGuardrail`
+
+Use a custom guardrail when a decision must be enforced before tool execution.
+Do not use prompt text or specialist instructions for hard control.
+
+## Execution evaluator rules
+
+Execution evaluator rules run during final-answer quality review. Each rule
+returns zero or more issues. Non-recoverable issues block delivery; recoverable
+issues trigger continuation or configured recovery.
+
+```ts
+import {
+  createStableHarnessRuntime,
+  defaultExecutionEvaluatorRules,
+  type ExecutionEvaluatorRule,
+} from "stable-harness";
+
+const requireCitationRule: ExecutionEvaluatorRule = (input) => {
+  const text = input.output?.text ?? "";
+  if (text.includes("Source:")) return [];
+  return [{
+    code: "missing_source_citation",
+    message: "Final answer must cite at least one observed source.",
+    recoverable: true,
+  }];
+};
+
+const runtime = createStableHarnessRuntime({
+  workspace,
+  adapters,
+  toolGateway,
+  executionEvaluatorRules: [
+    ...defaultExecutionEvaluatorRules,
+    requireCitationRule,
+  ],
+});
+```
+
+Use evaluator rules when the run already produced output and the runtime must
+decide whether the answer is deliverable.
+
+## Sequence
+
+```mermaid
+sequenceDiagram
+  autonumber
+  actor User
+  participant Product as Product runtime assembly
+  participant Runtime as Stable runtime
+  participant Adapter as Adapter or native tool call
+  participant Guardrails as ToolGuardrail[]
+  participant Gateway as Tool gateway
+  participant Tool as Tool implementation
+  participant Quality as ExecutionEvaluatorRule[]
+
+  User->>Product: Request
+  Product->>Runtime: workspace, adapters, toolGuardrails, executionEvaluatorRules
+  Runtime->>Adapter: Run selected agent or explicit toolCall
+
+  loop For each model-visible or direct tool call
+    Adapter->>Guardrails: evaluate structured tool context
+    alt Guardrail blocks
+      Guardrails-->>Adapter: blocking decision
+      Adapter-->>Runtime: control event and model-readable output
+      Runtime->>Adapter: continue, recover, or synthesize per policy
+    else Guardrail passes
+      Guardrails-->>Adapter: no decision
+      Adapter->>Gateway: invoke tool
+      Gateway->>Tool: execute with runtime context
+      Tool-->>Gateway: evidence or error
+      Gateway-->>Adapter: tool result
+      Adapter-->>Runtime: structured evidence event
+    end
+  end
+
+  Adapter-->>Runtime: candidate final answer
+  Runtime->>Quality: review final answer against evidence and policy
+  alt Final answer passes
+    Quality-->>Runtime: pass
+    Runtime-->>Product: deliver final output
+  else Recoverable issue
+    Quality-->>Runtime: continue_react
+    Runtime->>Adapter: continue with evidence and issue context
+  else Blocking issue
+    Quality-->>Runtime: blocked
+    Runtime-->>Product: blocked output or evidence synthesis
+  end
+```
+
+## Boundary
+
+- Put hard pre-tool controls in `toolGuardrails`.
+- Put hard final-answer controls in `executionEvaluatorRules`.
+- Put deterministic evidence workflows in bounded composite tools.
+- Put workflow preferences and SOP in skills and prompts.
+- Keep product-specific policy in the product runtime assembly, not in generic
+  DeepAgents semantics.
+
+## Release
+
+Stable Harness releases are published by the GitHub Actions release workflow.
+Local release commands are validation tools only: run `npm run release:pack` and
+`npm run release:smoke` before merging release-sensitive changes, then let the
+workflow publish with the configured npm token.

package/node_modules/@stable-harness/adapter-deepagents/dist/src/internal/gateway-tools.js

@@ -1 +1 @@
-import{ToolMessage as t}from"@langchain/core/messages";import{tool as e}from"@langchain/core/tools";import{afterToolInvoke as o,beforeToolInvoke as r,createToolRepeatState as n,missingRequiredPlanContent as s,missingToolDependencyContent as a,toolInvocationEvents as i}from"@stable-harness/core";import{isSuccessfulEvidenceOutput as u,observedToolEvidence as l,recordObservedToolEvidence as c}from"./gateway/tool-evidence.js";import{emitStructuredToolFailure as d}from"./gateway/tool-failure-events.js";export function buildGatewayTools(l,c,p,f,g=n(l.workspace.runtime.toolGateway)){return l.toolGateway?p.flatMap(n=>{const p=l.toolGateway?.get(n);if(!p)return[];const m=l.workspace.tools.get(n),y=m?.schema??p.schema;return[e(async e=>async function invokeGuardedGatewayTool(e){emitToolResult(e.input,e.agentId,e.toolId,void 0);const n=i({request:e.input.request,getEvents:e.input.getEvents}),l=s({agent:e.input.agent,events:n,toolId:e.toolId});if(l)return emitToolResult(e.input,e.agentId,e.toolId,l),l;const c=a({agent:e.input.agent,events:n,toolId:e.toolId});if(c)return emitToolResult(e.input,e.agentId,e.toolId,c),c;const p=e.repeatState?r(e.toolId,e.args,e.repeatState):void 0;if(p)return emitToolResult(e.input,e.agentId,e.toolId,p.eventOutput),p.modelOutput;const f=await async function invokeGatewayTool(e,o,r,n,s){try{if(e.toolFailureTracker?.isCircuitOpen(r))throw new Error(`Tool circuit is open: ${r}`);const t=await e.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:e.workspace.root,requestId:e.requestId,sessionId:e.sessionId,agentId:o,requestInput:e.request.input,observedEvidence:formatObservedEvidenceForToolContext(e),approvalIds:readApprovalIds(e.request.metadata)}});return e.toolFailureTracker?.recordSuccess(r),t}catch(n){if(d(e,o,r,n),function isToolArgumentValidationError(t){return t instanceof Error&&"ToolArgumentValidationError"===t.name&&"string"==typeof t.toolId&&Array.isArray(t.issues)}(n))return new t({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(n)});if(e.workspace.runtime.retry?.tools?.enabled)throw n;return new t({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(n),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(e.input,e.agentId,e.toolId,e.args,e.repairModel),g=f instanceof t?String(f.content):stringifyDeepAgentResult(f.output),m=e.repeatState?o({toolId:e.toolId,args:e.args,output:g,successful:!(f instanceof t&&"error"===f.status)&&u(g),state:e.repeatState}):{};return emitToolResult(e.input,e.agentId,e.toolId,m.eventOutput??g),void 0!==m.modelOutput?m.modelOutput:f instanceof t?f:g}({input:l,agentId:c,toolId:n,args:e,repairModel:f,repeatState:g}),{name:n,description:buildToolDescription(m?.description??p.description??n,y,l.workspace.runtime.toolGateway,n),schema:{type:"object",additionalProperties:!0}})]}):[]}function emitToolResult(t,e,o,r){void 0!==r&&c(t,e,o,r),t.emit({type:"runtime.adapter.event",requestId:t.requestId,sessionId:t.sessionId,agentId:e,event:void 0===r?{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.start",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.result",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),evidenceOutput:r,...toolControlProjection(r)}})}export function stringifyDeepAgentResult(e){if(e instanceof t)return function stringifyToolMessageContent(t){return"string"==typeof t?t:JSON.stringify(t)}(e.content);if("string"==typeof e)return e;if(isRecord(e)){const t=e.structuredResponse??e.structured_response;if(void 0!==t)return"string"==typeof t?t:JSON.stringify(t);const o=(Array.isArray(e.messages)?e.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(e.update)&&Array.isArray(e.update.messages)?e.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(e)}function buildToolDescription(t,e,o,r){const n=function toolRepeatPolicyDescription(t,e){const o=function repeatGuardConfig(t){return isRecord(t)&&isRecord(t.repeatGuard)?t.repeatGuard:{}}(t),r=function readPositiveIntegerMap(t){return isRecord(t)?new Map(Object.entries(t).map(([t,e])=>[t,readPositiveInteger(e)]).filter(t=>void 0!==t[1])):new Map}(o.maxSuccessfulCallsByTool).get(e)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${t}\n\n${n}`:t;return e?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(e))}`:s}function readPositiveInteger(t){return"number"==typeof t&&Number.isInteger(t)&&t>0?t:void 0}function previewToolOutput(t){const e=t.replace(/\s+/gu," ").trim();return e.length>500?`${e.slice(0,497)}...`:e}export function toolControlProjection(t){const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);if("string"==typeof e?.status)return{controlStatus:e.status};const o=function readTextStatus(t){return String(t).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(t);return o?{controlStatus:o}:"string"==typeof e?.error?{controlStatus:e.error}:t.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function readApprovalIds(t){const e=t?.approvalIds??t?.approvalId;return"string"==typeof e&&e.trim()?[e.trim()]:Array.isArray(e)?e.filter(t=>"string"==typeof t&&t.trim().length>0):void 0}function formatObservedEvidenceForToolContext(t){const e=l(t).map(t=>`Tool: ${t.toolId}\n${t.output}`).join("\n\n---\n\n");return e.length>12e3?`${e.slice(0,12e3)}\n[truncated]`:e}function formatToolArgumentError(t){return JSON.stringify({error:"tool_argument_validation_failed",toolId:t.toolId,issues:t.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function formatError(t){return t instanceof Error?t.message:String(t)}
+import{ToolMessage as t}from"@langchain/core/messages";import{tool as e}from"@langchain/core/tools";import{afterToolInvoke as o,createToolRepeatState as r,evaluateToolGuardrails as n,toolInvocationEvents as s}from"@stable-harness/core";import{isSuccessfulEvidenceOutput as a,observedToolEvidence as i,recordObservedToolEvidence as u}from"./gateway/tool-evidence.js";import{emitStructuredToolFailure as l}from"./gateway/tool-failure-events.js";export function buildGatewayTools(i,u,c,d,p=r(i.workspace.runtime.toolGateway)){return i.toolGateway?c.flatMap(r=>{const c=i.toolGateway?.get(r);if(!c)return[];const f=i.workspace.tools.get(r),g=f?.schema??c.schema;return[e(async e=>async function invokeGuardedGatewayTool(e){emitToolResult(e.input,e.agentId,e.toolId,void 0);const r=s({request:e.input.request,getEvents:e.input.getEvents}),i=n({agent:e.input.agent,args:e.args,events:r,repeatState:e.repeatState,toolId:e.toolId},e.input.toolGuardrails);if(i)return emitToolResult(e.input,e.agentId,e.toolId,i.eventOutput),i.modelOutput;const u=await async function invokeGatewayTool(e,o,r,n,s){try{if(e.toolFailureTracker?.isCircuitOpen(r))throw new Error(`Tool circuit is open: ${r}`);const t=await e.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:e.workspace.root,requestId:e.requestId,sessionId:e.sessionId,agentId:o,requestInput:e.request.input,observedEvidence:formatObservedEvidenceForToolContext(e),approvalIds:readApprovalIds(e.request.metadata)}});return e.toolFailureTracker?.recordSuccess(r),t}catch(n){if(l(e,o,r,n),function isToolArgumentValidationError(t){return t instanceof Error&&"ToolArgumentValidationError"===t.name&&"string"==typeof t.toolId&&Array.isArray(t.issues)}(n))return new t({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(n)});if(e.workspace.runtime.retry?.tools?.enabled)throw n;return new t({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(n),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(e.input,e.agentId,e.toolId,e.args,e.repairModel),c=u instanceof t?String(u.content):stringifyDeepAgentResult(u.output),d=e.repeatState?o({toolId:e.toolId,args:e.args,output:c,successful:!(u instanceof t&&"error"===u.status)&&a(c),state:e.repeatState}):{};return emitToolResult(e.input,e.agentId,e.toolId,d.eventOutput??c),void 0!==d.modelOutput?d.modelOutput:u instanceof t?u:c}({input:i,agentId:u,toolId:r,args:e,repairModel:d,repeatState:p}),{name:r,description:buildToolDescription(f?.description??c.description??r,g,i.workspace.runtime.toolGateway,r),schema:{type:"object",additionalProperties:!0}})]}):[]}function emitToolResult(t,e,o,r){void 0!==r&&u(t,e,o,r),t.emit({type:"runtime.adapter.event",requestId:t.requestId,sessionId:t.sessionId,agentId:e,event:void 0===r?{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.start",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.result",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),evidenceOutput:r,...toolControlProjection(r)}})}export function stringifyDeepAgentResult(e){if(e instanceof t)return function stringifyToolMessageContent(t){return"string"==typeof t?t:JSON.stringify(t)}(e.content);if("string"==typeof e)return e;if(isRecord(e)){const t=e.structuredResponse??e.structured_response;if(void 0!==t)return"string"==typeof t?t:JSON.stringify(t);const o=(Array.isArray(e.messages)?e.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(e.update)&&Array.isArray(e.update.messages)?e.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(e)}function buildToolDescription(t,e,o,r){const n=function toolRepeatPolicyDescription(t,e){const o=function repeatGuardConfig(t){return isRecord(t)&&isRecord(t.repeatGuard)?t.repeatGuard:{}}(t),r=function readPositiveIntegerMap(t){return isRecord(t)?new Map(Object.entries(t).map(([t,e])=>[t,readPositiveInteger(e)]).filter(t=>void 0!==t[1])):new Map}(o.maxSuccessfulCallsByTool).get(e)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${t}\n\n${n}`:t;return e?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(e))}`:s}function readPositiveInteger(t){return"number"==typeof t&&Number.isInteger(t)&&t>0?t:void 0}function previewToolOutput(t){const e=t.replace(/\s+/gu," ").trim();return e.length>500?`${e.slice(0,497)}...`:e}export function toolControlProjection(t){const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);if("string"==typeof e?.status)return{controlStatus:e.status};const o=function readTextStatus(t){return String(t).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(t);return o?{controlStatus:o}:"string"==typeof e?.error?{controlStatus:e.error}:t.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function readApprovalIds(t){const e=t?.approvalIds??t?.approvalId;return"string"==typeof e&&e.trim()?[e.trim()]:Array.isArray(e)?e.filter(t=>"string"==typeof t&&t.trim().length>0):void 0}function formatObservedEvidenceForToolContext(t){const e=i(t).map(t=>`Tool: ${t.toolId}\n${t.output}`).join("\n\n---\n\n");return e.length>12e3?`${e.slice(0,12e3)}\n[truncated]`:e}function formatToolArgumentError(t){return JSON.stringify({error:"tool_argument_validation_failed",toolId:t.toolId,issues:t.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function formatError(t){return t instanceof Error?t.message:String(t)}

package/node_modules/@stable-harness/adapter-deepagents/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/adapter-deepagents",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -15,7 +15,7 @@
     "@langchain/node-vfs": "^0.1.4",
     "@langchain/ollama": "^1.2.7",
     "@langchain/openai": "^1.4.5",
-    "@stable-harness/core": "0.0.80",
+    "@stable-harness/core": "0.0.81",
     "deepagents": "^1.10.1",
     "langchain": "^1.4.0"
   },

package/node_modules/@stable-harness/adapter-langgraph/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/adapter-langgraph",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -11,6 +11,6 @@
   "types": "dist/src/index.d.ts",
   "peerDependencies": {
     "@langchain/langgraph": "^1.3.0",
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }

package/node_modules/@stable-harness/core/dist/quality/execution-review.d.ts

@@ -1,2 +1,5 @@
-import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
-export declare function reviewExecutionEvidence(input: QualityReviewInput, policy: QualityPolicy): QualityReviewResult;
+import type { QualityPolicy, QualityReviewInput, QualityReviewIssue, QualityReviewResult } from "./types.js";
+export type ExecutionEvaluatorRule = (input: QualityReviewInput, policy: QualityPolicy) => QualityReviewIssue[];
+export declare const defaultExecutionEvaluatorRules: readonly ExecutionEvaluatorRule[];
+export declare function reviewExecutionEvidence(input: QualityReviewInput, policy: QualityPolicy, rules?: readonly ExecutionEvaluatorRule[]): QualityReviewResult;
+export declare function evaluateExecutionRules(input: QualityReviewInput, policy: QualityPolicy, rules?: readonly ExecutionEvaluatorRule[]): QualityReviewIssue[];

package/node_modules/@stable-harness/core/dist/quality/execution-review.js

@@ -1 +1 @@
-import{controlBlockers as e,controlGaps as n,successfulEvidenceOutputs as t,successfulEvidenceToolIds as r}from"./event-evidence.js";const s=/(?<![\w.])(?:\d{1,3}(?:,\d{3})+|\d+)(?:\.\d+)?[%kKmMbBtTxX]?(?!\w)/gu;export function reviewExecutionEvidence(e,n){if(!n.enabled||!n.executionReview.enabled)return{verdict:"pass",issues:[]};const t=[...blockerIssues(e,n),...controlGapIssues(e,n),...emptyFinalIssues(e,n),...toolEvidenceIssues(e,n),...ungroundedNumberIssues(e,n)];return 0===t.length?{verdict:"pass",issues:[]}:{verdict:t.some(e=>!e.recoverable)?"blocked":"continue_react",issues:t}}function ungroundedNumberIssues(e,n){if(!n.executionReview.rejectUngroundedNumbers||!e.output?.text.trim())return[];const r=numberSet(t(e.events).join("\n"));if(0===r.size)return[];const s=[...numberSet(e.output.text)].filter(e=>!function isSupportedNumber(e,n){if(n.has(e))return!0;const t=Number.parseFloat(e);if(!Number.isFinite(t))return!1;for(const e of n){const n=Number.parseFloat(e);if(Number.isFinite(n)&&Math.abs(n-t)<=roundingTolerance(t))return!0}return!1}(e,r));return 0===s.length?[]:[{code:"ungrounded_numeric_claim",message:`Final answer contains numeric claims not found in successful tool evidence: ${s.slice(0,12).join(", ")}`,recoverable:!1}]}function numberSet(e){const n=new Set;for(const t of e.matchAll(s)){const e=normalizeNumber(t[0]);e&&n.add(e)}return n}function normalizeNumber(e){const n=e.replace(/,/gu,"").replace(/^\+/u,"").replace(/[%kKmMbBtTxX]$/u,"").trim();if(n){if(/^\d+$/u.test(n)){const e=Number.parseInt(n,10);if(e>=1&&e<=20)return;return String(e)}return/^\d+\.\d+$/u.test(n)?n.replace(/0+$/u,"").replace(/\.$/u,""):void 0}}function roundingTolerance(e){return Math.abs(e)>=1e3?1:Math.abs(e)>=100?.1:Math.abs(e)>=10?.05:.005}function blockerIssues(n,t){return t.executionReview.stopOnBlocker?e(n.events).map(e=>({code:"control_blocker",message:`Execution produced a control blocker: ${e}`,recoverable:!1})):[]}function controlGapIssues(e,t){if(!t.executionReview.stopOnBlocker||!e.output?.text.trim())return[];const r=n(e.events).filter(n=>!function mentionsGap(e,n){const[t,r]=n.split(":"),s=e.toLowerCase();return s.includes(n.toLowerCase())||Boolean(t&&r&&s.includes(t.toLowerCase())&&s.includes(r.toLowerCase()))}(e.output?.text??"",n));return 0===r.length?[]:[{code:"unresolved_control_gap",message:`Final answer omitted unresolved runtime evidence gap(s): ${r.slice(0,8).join(", ")}`,recoverable:!0}]}function emptyFinalIssues(e,n){return!n.executionReview.rejectEmptyFinal||e.output?.text.trim()?[]:[{code:"empty_final_answer",message:"The final answer is empty.",recoverable:!0}]}function toolEvidenceIssues(e,n){return!n.executionReview.requireToolEvidence||r(e.events).length>0?[]:[{code:"missing_tool_evidence",message:"No successful tool or delegated-task evidence was observed.",recoverable:!0}]}
+import{controlBlockers as e,controlGaps as t,successfulEvidenceOutputs as n,successfulEvidenceToolIds as o}from"./event-evidence.js";const r=/(?<![\w.])(?:\d{1,3}(?:,\d{3})+|\d+)(?:\.\d+)?[%kKmMbBtTxX]?(?!\w)/gu;export const defaultExecutionEvaluatorRules=[function blockerIssues(t,n){return n.executionReview.stopOnBlocker?e(t.events).map(e=>({code:"control_blocker",message:`Execution produced a control blocker: ${e}`,recoverable:!1})):[]},function controlGapIssues(e,n){if(!n.executionReview.stopOnBlocker||!e.output?.text.trim())return[];const o=t(e.events).filter(t=>!function mentionsGap(e,t){const[n,o]=t.split(":"),r=e.toLowerCase();return r.includes(t.toLowerCase())||Boolean(n&&o&&r.includes(n.toLowerCase())&&r.includes(o.toLowerCase()))}(e.output?.text??"",t));return 0===o.length?[]:[{code:"unresolved_control_gap",message:`Final answer omitted unresolved runtime evidence gap(s): ${o.slice(0,8).join(", ")}`,recoverable:!0}]},function emptyFinalIssues(e,t){return!t.executionReview.rejectEmptyFinal||e.output?.text.trim()?[]:[{code:"empty_final_answer",message:"The final answer is empty.",recoverable:!0}]},function toolEvidenceIssues(e,t){return!t.executionReview.requireToolEvidence||o(e.events).length>0?[]:[{code:"missing_tool_evidence",message:"No successful tool or delegated-task evidence was observed.",recoverable:!0}]},function ungroundedNumberIssues(e,t){if(!t.executionReview.rejectUngroundedNumbers||!e.output?.text.trim())return[];const o=numberSet(n(e.events).join("\n"));if(0===o.size)return[];const r=[...numberSet(e.output.text)].filter(e=>!function isSupportedNumber(e,t){if(t.has(e))return!0;const n=Number.parseFloat(e);if(!Number.isFinite(n))return!1;for(const e of t){const t=Number.parseFloat(e);if(Number.isFinite(t)&&Math.abs(t-n)<=roundingTolerance(n))return!0}return!1}(e,o));return 0===r.length?[]:[{code:"ungrounded_numeric_claim",message:`Final answer contains numeric claims not found in successful tool evidence: ${r.slice(0,12).join(", ")}`,recoverable:!1}]}];export function reviewExecutionEvidence(e,t,n=defaultExecutionEvaluatorRules){if(!t.enabled||!t.executionReview.enabled)return{verdict:"pass",issues:[]};const o=evaluateExecutionRules(e,t,n);return 0===o.length?{verdict:"pass",issues:[]}:{verdict:o.some(e=>!e.recoverable)?"blocked":"continue_react",issues:o}}export function evaluateExecutionRules(e,t,n=defaultExecutionEvaluatorRules){return n.flatMap(n=>n(e,t))}function numberSet(e){const t=new Set;for(const n of e.matchAll(r)){const e=normalizeNumber(n[0]);e&&t.add(e)}return t}function normalizeNumber(e){const t=e.replace(/,/gu,"").replace(/^\+/u,"").replace(/[%kKmMbBtTxX]$/u,"").trim();if(t){if(/^\d+$/u.test(t)){const e=Number.parseInt(t,10);if(e>=1&&e<=20)return;return String(e)}return/^\d+\.\d+$/u.test(t)?t.replace(/0+$/u,"").replace(/\.$/u,""):void 0}}function roundingTolerance(e){return Math.abs(e)>=1e3?1:Math.abs(e)>=100?.1:Math.abs(e)>=10?.05:.005}

package/node_modules/@stable-harness/core/dist/quality/runtime.d.ts

@@ -1,4 +1,5 @@
 import type { RuntimeMemoryContext, RuntimeOutput, RuntimeRequest } from "../types.js";
+import { type ExecutionEvaluatorRule } from "./execution-review.js";
 import type { QualityPolicy, QualityReviewInput, QualityReviewModel } from "./types.js";
 export type QualityRuntimeInput = QualityReviewInput & {
     requestId: string;
@@ -7,6 +8,7 @@ export type QualityRuntimeInput = QualityReviewInput & {
     getEvents: () => import("../types.js").RuntimeEvent[];
     runAdapter: (request: RuntimeRequest) => Promise<RuntimeOutput>;
     reviewModel?: QualityReviewModel;
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     memory?: RuntimeMemoryContext;
     pluginMemories: RuntimeMemoryContext[];
 };

package/node_modules/@stable-harness/core/dist/quality/runtime.js

@@ -1 +1 @@
-import{successfulEvidenceOutputs as e}from"./event-evidence.js";import{buildQualityRecoveryRequest as t}from"./recovery-policy.js";import{reviewExecutionEvidence as i}from"./execution-review.js";import{reviewWithLlm as n}from"./llm-review.js";import{reviewPlanningEvidence as r}from"./planning-review.js";import{synthesizeEvidenceOnlyReport as s}from"./synthesis.js";export async function recoverQualityReview(e,t,i,n){if(!n.enabled)return i;let r=t,s=i;for(let t=0;t<n.recovery.maxLoops+1;t+=1){const i=await emitPlanningReview(e,r,s,n);if("blocked"===i.verdict)return qualityFailureOutput("planning",i);const u=buildQualityRecovery(e,r,i,"planning",n,t);if(u){r=u,s=await e.runAdapter(r);continue}const o=await emitExecutionReview(e,r,s,n);if("pass"!==o.verdict){const t=await trySynthesizeExecution(e,r,o,n);if(t)return t}const a=buildQualityRecovery(e,r,o,"execution",n,t);if(!a)return"pass"===o.verdict?s:await trySynthesizeExecution(e,r,o,n)??qualityFailureOutput("execution",o);r=a,s=await e.runAdapter(r)}return qualityFailureOutput("execution",{verdict:"blocked",issues:[{code:"quality_recovery_exhausted",message:`Quality recovery exceeded maxLoops=${n.recovery.maxLoops}.`,recoverable:!1}]})}async function trySynthesizeExecution(e,t,n,r){const u=s({...reviewInputFor(e,t),output:void 0},n,r);if(!u)return;e.emit({type:"runtime.quality.synthesis.created",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,mode:r.synthesis.mode});const o={text:u},a=i({...reviewInputFor(e,t),output:o},r);return emitReviewEvent(e,"execution",a),"pass"===a.verdict?o:void 0}function emitPlanningReview(e,t,i,n){return emitReview(e,"planning",r,t,i,n)}function emitExecutionReview(e,t,n,r){return emitReview(e,"execution",i,t,n,r)}async function emitReview(e,t,i,r,s,u){const o={...reviewInputFor(e,r),output:s},a="planning"===t?u.planningReview.enabled:u.executionReview.enabled;if(!a)return i(o,u);const c=i(o,u),d=await n({phase:t,review:o,policy:u,model:e.reviewModel}),v="pass"===c.verdict?d??c:c;return a&&emitReviewEvent(e,t,v),v}function emitReviewEvent(e,t,i){"planning"!==t?e.emit({type:"runtime.quality.execution.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues}):e.emit({type:"runtime.quality.planning.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues})}function buildQualityRecovery(i,n,r,s,u,o){if(o>=u.recovery.maxLoops)return;const a=t({request:n,result:r,phase:s,policy:u,availableToolIds:i.agent.tools,availableSubagentIds:i.agent.subagents,observedEvidence:"execution"===s?e(i.getEvents()):[]});return a&&i.emit({type:"runtime.quality.recovery.started",requestId:i.requestId,sessionId:i.sessionId,agentId:i.agent.id,phase:s,attempt:o+1,verdict:r.verdict}),a}function reviewInputFor(e,t){return{workspace:e.workspace,agent:e.agent,request:t,events:e.getEvents()}}function qualityFailureOutput(e,t){return{text:[`Stable runtime quality review blocked final delivery during ${e}.`,"",...t.issues.length>0?t.issues.map(e=>`- ${e.code}: ${e.message}`):["- quality_review_failed: Quality review did not pass."]].join("\n")}}
+import{successfulEvidenceOutputs as e}from"./event-evidence.js";import{buildQualityRecoveryRequest as t}from"./recovery-policy.js";import{reviewExecutionEvidence as i}from"./execution-review.js";import{reviewWithLlm as n}from"./llm-review.js";import{reviewPlanningEvidence as r}from"./planning-review.js";import{synthesizeEvidenceOnlyReport as s}from"./synthesis.js";export async function recoverQualityReview(e,t,i,n){if(!n.enabled)return i;let r=t,s=i;for(let t=0;t<n.recovery.maxLoops+1;t+=1){const i=await emitPlanningReview(e,r,s,n);if("blocked"===i.verdict)return qualityFailureOutput("planning",i);const u=buildQualityRecovery(e,r,i,"planning",n,t);if(u){r=u,s=await e.runAdapter(r);continue}const o=await emitExecutionReview(e,r,s,n);if("pass"!==o.verdict){const t=await trySynthesizeExecution(e,r,o,n);if(t)return t}const a=buildQualityRecovery(e,r,o,"execution",n,t);if(!a)return"pass"===o.verdict?s:await trySynthesizeExecution(e,r,o,n)??qualityFailureOutput("execution",o);r=a,s=await e.runAdapter(r)}return qualityFailureOutput("execution",{verdict:"blocked",issues:[{code:"quality_recovery_exhausted",message:`Quality recovery exceeded maxLoops=${n.recovery.maxLoops}.`,recoverable:!1}]})}async function trySynthesizeExecution(e,t,n,r){const u=s({...reviewInputFor(e,t),output:void 0},n,r);if(!u)return;e.emit({type:"runtime.quality.synthesis.created",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,mode:r.synthesis.mode});const o={text:u},a=i({...reviewInputFor(e,t),output:o},r,e.executionEvaluatorRules);return emitReviewEvent(e,"execution",a),"pass"===a.verdict?o:void 0}function emitPlanningReview(e,t,i,n){return emitReview(e,"planning",r,t,i,n)}function emitExecutionReview(e,t,n,r){return emitReview(e,"execution",(t,n)=>i(t,n,e.executionEvaluatorRules),t,n,r)}async function emitReview(e,t,i,r,s,u){const o={...reviewInputFor(e,r),output:s},a="planning"===t?u.planningReview.enabled:u.executionReview.enabled;if(!a)return i(o,u);const c=i(o,u),d=await n({phase:t,review:o,policy:u,model:e.reviewModel}),v="pass"===c.verdict?d??c:c;return a&&emitReviewEvent(e,t,v),v}function emitReviewEvent(e,t,i){"planning"!==t?e.emit({type:"runtime.quality.execution.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues}):e.emit({type:"runtime.quality.planning.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues})}function buildQualityRecovery(i,n,r,s,u,o){if(o>=u.recovery.maxLoops)return;const a=t({request:n,result:r,phase:s,policy:u,availableToolIds:i.agent.tools,availableSubagentIds:i.agent.subagents,observedEvidence:"execution"===s?e(i.getEvents()):[]});return a&&i.emit({type:"runtime.quality.recovery.started",requestId:i.requestId,sessionId:i.sessionId,agentId:i.agent.id,phase:s,attempt:o+1,verdict:r.verdict}),a}function reviewInputFor(e,t){return{workspace:e.workspace,agent:e.agent,request:t,events:e.getEvents()}}function qualityFailureOutput(e,t){return{text:[`Stable runtime quality review blocked final delivery during ${e}.`,"",...t.issues.length>0?t.issues.map(e=>`- ${e.code}: ${e.message}`):["- quality_review_failed: Quality review did not pass."]].join("\n")}}

package/node_modules/@stable-harness/core/dist/runtime/direct-tool-call.d.ts

@@ -1,4 +1,5 @@
 import type { CompiledWorkspace, RuntimeEvent, RuntimeOutput, RuntimeRequest, RuntimeToolFailureTracker, RuntimeToolGateway, WorkspaceAgent } from "../types.js";
+import { type ToolGuardrail } from "./policy/tool-invocation.js";
 export declare function runDirectToolCall(input: {
     gateway: RuntimeToolGateway | undefined;
     workspace: CompiledWorkspace;
@@ -8,4 +9,6 @@ export declare function runDirectToolCall(input: {
     sessionId: string;
     agent: WorkspaceAgent;
     toolFailureTracker?: RuntimeToolFailureTracker;
+    toolGuardrails?: readonly ToolGuardrail[];
+    events?: RuntimeEvent[];
 }): Promise<RuntimeOutput>;

package/node_modules/@stable-harness/core/dist/runtime/direct-tool-call.js

@@ -1 +1 @@
-import{toolCircuitOpenEvent as o,toolFailureEvent as t}from"./tool-failure.js";export async function runDirectToolCall(o){const t=o.request.toolCall;if(!t)throw new Error("Direct tool call request is missing");if(!o.gateway)throw new Error("Runtime tool gateway is not configured");const e=await async function resolveDirectToolCall(o){if(o.agent.tools.includes(o.toolId)&&o.gateway.get(o.toolId))return{toolId:o.toolId,args:o.args};const t=await(o.gateway.repairToolCall?.({toolId:o.toolId,args:o.args,allowedToolIds:o.agent.tools,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,requestInput:o.request.input,approvalIds:readApprovalIds(o.request.metadata)}}));if(t&&o.agent.tools.includes(t.toolId)&&o.gateway.get(t.toolId))return emitToolRepair(o,"repaired",t.toolId),t;if(!o.agent.tools.includes(o.toolId))throw emitToolRepair(o,"blocked",void 0,`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`),new Error(`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`);throw emitToolRepair(o,"blocked",void 0,`Tool is not registered: ${o.toolId}`),new Error(`Tool is not registered: ${o.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,emit:o.emit,request:o.request,toolId:t.toolId,args:t.args});if(o.emit({type:"runtime.tool.direct.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:e.toolId}),o.toolFailureTracker?.isCircuitOpen(e.toolId)){const t=new Error(`Tool circuit is open: ${e.toolId}`);throw emitToolFailure(o,e.toolId,t),t}const r=await async function invokeToolWithFailureEvents(o,t){try{return await o.gateway.invoke({toolId:t.toolId,args:t.args,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,requestInput:o.request.input,approvalIds:readApprovalIds(o.request.metadata)}})}catch(e){throw emitToolFailure(o,t.toolId,e),e}}(o,e);return o.toolFailureTracker?.recordSuccess(r.toolId),o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:r.toolId,output:r.output}),{text:(s=r.output,"string"==typeof s?s:JSON.stringify(s)),metadata:{toolCall:{toolId:r.toolId}}};var s}function emitToolFailure(e,r,s){const a=t({requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,toolId:r,error:s});e.emit(a),e.toolFailureTracker?.recordFailure(r)&&e.emit(o({requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,toolId:r,reason:"runtime.tool.failure"===a.type?a.failure.reason:"unknown"}))}function emitToolRepair(o,t,e,r){o.emit({type:"runtime.inventory.repair",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,status:t,diagnostic:{layer:"tool",owner:"stable_runtime_policy",originalId:o.toolId,repairedId:e,candidateIds:o.agent.tools,reason:r}})}function readApprovalIds(o){const t=o?.approvalIds??o?.approvalId;return"string"==typeof t&&t.trim()?[t.trim()]:Array.isArray(t)?t.filter(o=>"string"==typeof o&&o.trim().length>0):void 0}
+import{evaluateToolGuardrails as t}from"./policy/tool-invocation.js";import{toolCircuitOpenEvent as o,toolFailureEvent as e}from"./tool-failure.js";export async function runDirectToolCall(o){const e=o.request.toolCall;if(!e)throw new Error("Direct tool call request is missing");if(!o.gateway)throw new Error("Runtime tool gateway is not configured");const r=await async function resolveDirectToolCall(t){if(t.agent.tools.includes(t.toolId)&&t.gateway.get(t.toolId))return{toolId:t.toolId,args:t.args};const o=await(t.gateway.repairToolCall?.({toolId:t.toolId,args:t.args,allowedToolIds:t.agent.tools,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,requestInput:t.request.input,approvalIds:readApprovalIds(t.request.metadata)}}));if(o&&t.agent.tools.includes(o.toolId)&&t.gateway.get(o.toolId))return emitToolRepair(t,"repaired",o.toolId),o;if(!t.agent.tools.includes(t.toolId))throw emitToolRepair(t,"blocked",void 0,`Tool ${t.toolId} is not assigned to agent ${t.agent.id}`),new Error(`Tool ${t.toolId} is not assigned to agent ${t.agent.id}`);throw emitToolRepair(t,"blocked",void 0,`Tool is not registered: ${t.toolId}`),new Error(`Tool is not registered: ${t.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,emit:o.emit,request:o.request,toolId:e.toolId,args:e.args});o.emit({type:"runtime.tool.direct.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:r.toolId});const s=t({agent:o.agent,args:r.args,events:o.events??[],toolId:r.toolId},o.toolGuardrails);if(s)return o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:r.toolId,output:s.eventOutput}),{text:s.modelOutput,metadata:{toolCall:{toolId:r.toolId},controlStatus:s.status}};if(o.toolFailureTracker?.isCircuitOpen(r.toolId)){const t=new Error(`Tool circuit is open: ${r.toolId}`);throw emitToolFailure(o,r.toolId,t),t}const a=await async function invokeToolWithFailureEvents(t,o){try{return await t.gateway.invoke({toolId:o.toolId,args:o.args,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,requestInput:t.request.input,approvalIds:readApprovalIds(t.request.metadata)}})}catch(e){throw emitToolFailure(t,o.toolId,e),e}}(o,r);return o.toolFailureTracker?.recordSuccess(a.toolId),o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:a.toolId,output:a.output}),{text:(i=a.output,"string"==typeof i?i:JSON.stringify(i)),metadata:{toolCall:{toolId:a.toolId}}};var i}function emitToolFailure(t,r,s){const a=e({requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,toolId:r,error:s});t.emit(a),t.toolFailureTracker?.recordFailure(r)&&t.emit(o({requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,toolId:r,reason:"runtime.tool.failure"===a.type?a.failure.reason:"unknown"}))}function emitToolRepair(t,o,e,r){t.emit({type:"runtime.inventory.repair",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,status:o,diagnostic:{layer:"tool",owner:"stable_runtime_policy",originalId:t.toolId,repairedId:e,candidateIds:t.agent.tools,reason:r}})}function readApprovalIds(t){const o=t?.approvalIds??t?.approvalId;return"string"==typeof o&&o.trim()?[o.trim()]:Array.isArray(o)?o.filter(t=>"string"==typeof t&&t.trim().length>0):void 0}

package/node_modules/@stable-harness/core/dist/runtime/policy/tool-invocation.d.ts

@@ -17,6 +17,25 @@ export type ToolRepeatDecision = {
     eventOutput: string;
     modelOutput: string;
 };
+export type ToolGuardrailContext = {
+    agent: WorkspaceAgent;
+    args: unknown;
+    events: RuntimeEvent[];
+    repeatState?: ToolRepeatState;
+    toolId: string;
+};
+export type ToolGuardrailDecision = {
+    eventOutput: string;
+    modelOutput: string;
+    reason: string;
+    status: string;
+};
+export type ToolGuardrail = (context: ToolGuardrailContext) => ToolGuardrailDecision | undefined;
+export declare const requiredPlanToolGuardrail: ToolGuardrail;
+export declare const toolDependencyGuardrail: ToolGuardrail;
+export declare const repeatToolGuardrail: ToolGuardrail;
+export declare const defaultToolGuardrails: readonly ToolGuardrail[];
+export declare function evaluateToolGuardrails(context: ToolGuardrailContext, guardrails?: readonly ToolGuardrail[]): ToolGuardrailDecision | undefined;
 export declare function createToolRepeatState(config: unknown): ToolRepeatState | undefined;
 export declare function beforeToolInvoke(toolId: string, args: unknown, state: ToolRepeatState): ToolRepeatDecision | undefined;
 export declare function afterToolInvoke(input: {

package/node_modules/@stable-harness/core/dist/runtime/policy/tool-invocation.js

@@ -1 +1 @@
-export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,repeatLimitedTools:new Set,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool),returnPreviousOutputOnRepeatLimit:!0===repeatGuardConfig(e).returnPreviousOutputOnRepeatLimit}}export function beforeToolInvoke(e,t,o){const n=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,n+1);const r=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==r&&n>=r){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const l=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==l&&(o.successfulToolCounts.get(e)??0)>=l){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const a=stableToolCallKey(e,t),s=o.successfulCalls.get(a);if(void 0!==s){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return o.repeatLimitedTools.add(e),{eventOutput:t,modelOutput:t}}const n=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,s);return{eventOutput:n,modelOutput:s}}}export function afterToolInvoke(e){return e.successful?(e.state.successfulCalls.set(stableToolCallKey(e.toolId,e.args),e.output),e.state.latestSuccessfulOutputByTool.set(e.toolId,e.output),e.state.successfulToolCounts.set(e.toolId,(e.state.successfulToolCounts.get(e.toolId)??0)+1),{}):{}}export function isToolRepeatLimitReached(e,t){if(!t)return!1;if(t.repeatLimitedTools.has(e))return!0;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const n=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==n&&(t.successfulToolCounts.get(e)??0)>=n}export function missingRequiredPlanContent(e){const t=readRecord(e.agent.config.executionContract);if(!0!==t.requiresPlan)return"";const o=readStringArray(t.planEvidenceTools);if(0===o.length||o.includes(e.toolId))return"";const n=new Set(e.events.flatMap(readEvidenceToolId));return o.some(e=>n.has(e))?"":["Status: plan_required",`Evidence tool: ${e.toolId}`,`Blocker: execution contract requires a planning checkpoint from one of: ${o.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}export function missingToolDependencyContent(e){const t=readRecord(e.agent.config.executionContract),o=readStringArray(readRecord(t.toolDependencies)[e.toolId]);if(0===o.length)return"";const n=new Set(e.events.flatMap(readEvidenceToolId)),r=o.filter(e=>!n.has(e));return 0===r.length?"":["Status: dependency_required",`Evidence tool: ${e.toolId}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${r.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}export function toolInvocationEvents(e){const t=e.getEvents?.();return t??[]}function repeatGuardConfig(e){const t=readRecord(e);return readRecord(t.repeatGuard)}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){const t=readRecord(e);return new Map(Object.entries(t).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1]))}function repeatLimitModelOutput(e,t,o){return o.returnPreviousOutputOnRepeatLimit&&void 0!==t&&0!==t.trim().length?t:e}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)&&function isToolResultEvent(e){return"deepagents.tool_execution.result"===e.eventType||"agent.tool.result"===e.phase}(e.event)&&"string"==typeof e.event.toolId&&function isSuccessfulEvidenceEvent(e){const t=function readString(e){return"string"==typeof e&&e.length>0?e:void 0}(e.controlStatus)??function readOutputStatus(e){if("string"!=typeof e)return;const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:e.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e.output);return!t||/^(?:completed|success|ok|recorded)$/iu.test(t)}(e.event)?[e.event.toolId]:[]}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}
+export const requiredPlanToolGuardrail=e=>{const t=missingRequiredPlanContent({agent:e.agent,events:e.events,toolId:e.toolId});return t?controlDecision("plan_required",t):void 0};export const toolDependencyGuardrail=e=>{const t=missingToolDependencyContent({agent:e.agent,events:e.events,toolId:e.toolId});return t?controlDecision("dependency_required",t):void 0};export const repeatToolGuardrail=e=>{const t=e.repeatState?beforeToolInvoke(e.toolId,e.args,e.repeatState):void 0;return t?{eventOutput:t.eventOutput,modelOutput:t.modelOutput,reason:t.eventOutput,status:readOutputStatus(t.eventOutput)??"repeated_tool_call_limit"}:void 0};export const defaultToolGuardrails=[requiredPlanToolGuardrail,toolDependencyGuardrail,repeatToolGuardrail];export function evaluateToolGuardrails(e,t=defaultToolGuardrails){for(const o of t){const t=o(e);if(t)return t}}export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,repeatLimitedTools:new Set,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool),returnPreviousOutputOnRepeatLimit:!0===repeatGuardConfig(e).returnPreviousOutputOnRepeatLimit}}export function beforeToolInvoke(e,t,o){const n=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,n+1);const r=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==r&&n>=r){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const l=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==l&&(o.successfulToolCounts.get(e)??0)>=l){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const a=stableToolCallKey(e,t),u=o.successfulCalls.get(a);if(void 0!==u){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return o.repeatLimitedTools.add(e),{eventOutput:t,modelOutput:t}}const n=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,u);return{eventOutput:n,modelOutput:u}}}function controlDecision(e,t){return{eventOutput:t,modelOutput:t,reason:t,status:e}}export function afterToolInvoke(e){return e.successful?(e.state.successfulCalls.set(stableToolCallKey(e.toolId,e.args),e.output),e.state.latestSuccessfulOutputByTool.set(e.toolId,e.output),e.state.successfulToolCounts.set(e.toolId,(e.state.successfulToolCounts.get(e.toolId)??0)+1),{}):{}}export function isToolRepeatLimitReached(e,t){if(!t)return!1;if(t.repeatLimitedTools.has(e))return!0;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const n=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==n&&(t.successfulToolCounts.get(e)??0)>=n}export function missingRequiredPlanContent(e){const t=readRecord(e.agent.config.executionContract);if(!0!==t.requiresPlan)return"";const o=readStringArray(t.planEvidenceTools);if(0===o.length||o.includes(e.toolId))return"";const n=new Set(e.events.flatMap(readEvidenceToolId));return o.some(e=>n.has(e))?"":["Status: plan_required",`Evidence tool: ${e.toolId}`,`Blocker: execution contract requires a planning checkpoint from one of: ${o.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}export function missingToolDependencyContent(e){const t=readRecord(e.agent.config.executionContract),o=readStringArray(readRecord(t.toolDependencies)[e.toolId]);if(0===o.length)return"";const n=new Set(e.events.flatMap(readEvidenceToolId)),r=o.filter(e=>!n.has(e));return 0===r.length?"":["Status: dependency_required",`Evidence tool: ${e.toolId}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${r.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}export function toolInvocationEvents(e){const t=e.getEvents?.();return t??[]}function repeatGuardConfig(e){const t=readRecord(e);return readRecord(t.repeatGuard)}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){const t=readRecord(e);return new Map(Object.entries(t).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1]))}function repeatLimitModelOutput(e,t,o){return o.returnPreviousOutputOnRepeatLimit&&void 0!==t&&0!==t.trim().length?t:e}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)&&function isToolResultEvent(e){return"deepagents.tool_execution.result"===e.eventType||"agent.tool.result"===e.phase}(e.event)&&"string"==typeof e.event.toolId&&function isSuccessfulEvidenceEvent(e){const t=function readString(e){return"string"==typeof e&&e.length>0?e:void 0}(e.controlStatus)??readOutputStatus(e.output);return!t||/^(?:completed|success|ok|recorded)$/iu.test(t)}(e.event)?[e.event.toolId]:[]}function readOutputStatus(e){if("string"!=typeof e)return;const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:e.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/node_modules/@stable-harness/core/dist/runtime.d.ts

@@ -1,9 +1,10 @@
 import type { ApprovalQueue } from "@stable-harness/governance";
 import type { MemoryProvider, RuntimeMemoryStore } from "@stable-harness/memory";
-import type { QualityReviewModel } from "./quality/index.js";
+import type { ExecutionEvaluatorRule, QualityReviewModel } from "./quality/index.js";
+import type { ToolGuardrail } from "./runtime/policy/tool-invocation.js";
 import { createLangSmithTracingCapability } from "./runtime/tracing/langsmith.js";
 import type { CompiledWorkspace, RuntimeCapabilityModule, RuntimeToolGateway, RuntimeAdapter, RuntimeArtifactStore, RuntimeSandboxPolicy, RuntimeStore, RuntimeProgressNarrationOptions, RuntimeWorkflowAdapter, StableHarnessRuntime } from "./types.js";
-type RuntimeFactoryInput = {
+export type RuntimeFactoryInput = {
     workspace: CompiledWorkspace;
     adapters: RuntimeAdapter[];
     workflowAdapters?: RuntimeWorkflowAdapter[];
@@ -16,8 +17,9 @@ type RuntimeFactoryInput = {
     artifacts?: RuntimeArtifactStore;
     progressNarration?: RuntimeProgressNarrationOptions | false;
     qualityReviewModel?: QualityReviewModel;
+    toolGuardrails?: readonly ToolGuardrail[];
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     capabilities?: RuntimeCapabilityModule[];
     langSmithTracing?: false | Parameters<typeof createLangSmithTracingCapability>[0]["options"];
 };
 export declare function createStableHarnessRuntime(input: RuntimeFactoryInput): StableHarnessRuntime;
-export {};

package/node_modules/@stable-harness/core/dist/runtime.js

@@ -1 +1 @@
-import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{buildAdapterErrorRecoveryPrompt as r,buildExecutionContractRecoveryRequest as a,isRecoverableAdapterError as s}from"./recovery/tool-call.js";import{recoverQualityReview as n,resolveQualityPolicy as o}from"./quality/index.js";import{recoverAdapterResultOutput as i}from"./runtime/recovery/adapter-result.js";import{completeRun as u,failRun as c}from"./runtime/completion.js";import{createRuntimeAdministrationMethods as p}from"./runtime/admin/administration.js";import{runDirectToolCall as d}from"./runtime/direct-tool-call.js";import{createApprovalGatedToolGateway as m}from"./runtime/governance/approval-gate.js";import{createSandboxedToolGateway as l}from"./runtime/governance/sandbox.js";import{createRuntimeInspectionMethods as w}from"./runtime/inspection/methods.js";import{createRuntimeCapabilityRegistry as g,normalizeAdapterResult as y}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as f}from"./runtime/memory.js";import{resolveToolCallRecoveryPolicy as I}from"./runtime/recovery/tool-call-policy.js";import{createRuntimeMemoryAdministration as q}from"./runtime/admin/memory.js";import{createInMemoryRuntimeStore as R}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as k}from"./runtime/progress-narration.js";import{repairRuntimeSelection as v}from"./runtime/selection-repair.js";import{createLangSmithTracingCapability as b}from"./runtime/tracing/langsmith.js";import{createToolFailureTracker as A}from"./runtime/tool-failure.js";import{runWorkflowRequest as C}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const y=new Set,j=t.store??R(),h=g([f(t),k({options:t.progressNarration,policy:t.workspace.runtime}),b({policy:t.workspace.runtime,store:j,options:t.langSmithTracing}),...t.capabilities??[]]),emitBase=t=>{const r=function enrichRuntimeEvent(t){return{...t,eventId:t.eventId??e(),emittedAt:t.emittedAt??(new Date).toISOString()}}(t);j.appendEvent(r);for(const e of y)e(r)},emit=e=>{emitBase(e),h.emitSideEffects(e,emitBase)},x=l({gateway:m({gateway:t.toolGateway,approvals:t.approvals,workspace:t.workspace,emit:emit}),workspace:t.workspace,sandbox:t.sandbox,emit:emit}),S={...t,toolGateway:x},E=A(function readToolFailurePolicy(e){if("object"!=typeof e||null===e||Array.isArray(e))return;const t=e.failurePolicy;return"object"!=typeof t||null===t||Array.isArray(t)?void 0:t}(t.workspace.runtime.toolGateway));return{request:async t=>async function runRuntimeRequest(t){const p=t.request.requestId??e(),m=t.request.sessionId??e(),l=[],{agent:w,adapter:g}=await async function resolveExecution(e,t,r){const a=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const a=await v({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description})),trace:{...r,agentId:t,layer:"agent",owner:"stable_runtime_policy"}});return a.ok?a.id:t}(e.workspace,t.agentId,r):e.workspace.runtime.defaultAgentId,s=e.workspace.agents.get(a);if(!s)throw new Error(`Agent ${a} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const n=e.adapters.find(e=>e.canRun(s));if(!n)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:n}}(t.input,t.request,{requestId:p,sessionId:m,emit:e=>l.push(e)});t.store.createRun(function createRunRecord(e,t,r,a){return{requestId:t,sessionId:r,agentId:a.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,p,m,w)),l.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:p,sessionId:m,agentId:w.id,input:t.request.input});try{if(t.request.workflow){const e=await C({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:p,sessionId:m,agentId:w.id,emit:t.emit});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}if(t.request.toolCall){const e=await d({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:p,sessionId:m,agent:w,toolFailureTracker:t.toolFailureTracker});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,c=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),p=c.memory,d=c.pluginMemories??[],m=I({workspace:e.input.workspace,agent:e.agent}),l=o(e.input.workspace.runtime,e.agent),w=new Map;let g;try{g=await runAdapterOnce(e,t,e.request,p,d,w,m)}catch(a){if(!s(a,m))throw a;e.emit(repairStarted(e,"adapter_error",1,errorMessage(a))),g=await runAdapterOnce(e,t,r(e.request,a,m),p,d,w,m),e.emit(repairCompleted(e,"adapter_error","retried",1,errorMessage(a)))}g=await i({...e,request:e.request,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),e.request,g,l),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:g});try{assertRequestExecutionContract(e)}catch(r){const s=a({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:m});if(!s)throw r;e.emit(repairStarted(e,"execution_contract",1,errorMessage(r))),g=await runAdapterOnce(e,t,s,p,d,w,m),g=await i({...e,request:s,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),s,g,l),assertRequestExecutionContract(e),e.emit(repairCompleted(e,"execution_contract","retried",1,errorMessage(r)))}const y=u({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:g,artifacts:e.input.artifacts});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:g,response:y}),y}({...t,adapter:g,requestId:p,sessionId:m,agent:w})}catch(e){return c({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,error:e})}}({input:S,capabilities:h,store:j,emit:emit,request:t,toolFailureTracker:E}),subscribe:e=>(y.add(e),()=>y.delete(e)),...w({workspace:t.workspace,store:j,artifacts:t.artifacts,approvals:t.approvals,emit:emit}),...p({store:j,emit:emit}),...q({memory:t.memory}),cancel(e,t){const r=j.getRun(e);r&&"running"===r.state&&(j.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await h.stop(),y.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,a,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:n=>runAdapterOnce(e,e.adapter,n,t,r,a,s),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,a,s,n,o){return y(await t.run({workspace:{...e.input.workspace,runtime:o},agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:a,pluginMemories:s,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,requestState:n,getEvents:()=>e.store.getRun(e.requestId)?.events??[],emit:e.emit}))}function repairStarted(e,t,r,a){return{type:"runtime.repair.started",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,attempt:r,reason:a}}function repairCompleted(e,t,r,a,s){return{type:"runtime.repair.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,outcome:r,attempt:a,reason:s}}function errorMessage(e){return e instanceof Error?e.message:String(e)}
+import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{buildAdapterErrorRecoveryPrompt as r,buildExecutionContractRecoveryRequest as a,isRecoverableAdapterError as s}from"./recovery/tool-call.js";import{recoverQualityReview as n,resolveQualityPolicy as o}from"./quality/index.js";import{recoverAdapterResultOutput as i}from"./runtime/recovery/adapter-result.js";import{completeRun as u,failRun as c}from"./runtime/completion.js";import{createRuntimeAdministrationMethods as p}from"./runtime/admin/administration.js";import{runDirectToolCall as d}from"./runtime/direct-tool-call.js";import{createApprovalGatedToolGateway as m}from"./runtime/governance/approval-gate.js";import{createSandboxedToolGateway as l}from"./runtime/governance/sandbox.js";import{createRuntimeInspectionMethods as w}from"./runtime/inspection/methods.js";import{createRuntimeCapabilityRegistry as g,normalizeAdapterResult as y}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as f}from"./runtime/memory.js";import{resolveToolCallRecoveryPolicy as I}from"./runtime/recovery/tool-call-policy.js";import{createRuntimeMemoryAdministration as q}from"./runtime/admin/memory.js";import{createInMemoryRuntimeStore as R}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as v}from"./runtime/progress-narration.js";import{repairRuntimeSelection as k}from"./runtime/selection-repair.js";import{createLangSmithTracingCapability as b}from"./runtime/tracing/langsmith.js";import{createToolFailureTracker as A}from"./runtime/tool-failure.js";import{runWorkflowRequest as C}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const y=new Set,x=t.store??R(),j=g([f(t),v({options:t.progressNarration,policy:t.workspace.runtime}),b({policy:t.workspace.runtime,store:x,options:t.langSmithTracing}),...t.capabilities??[]]),emitBase=t=>{const r=function enrichRuntimeEvent(t){return{...t,eventId:t.eventId??e(),emittedAt:t.emittedAt??(new Date).toISOString()}}(t);x.appendEvent(r);for(const e of y)e(r)},emit=e=>{emitBase(e),j.emitSideEffects(e,emitBase)},E=l({gateway:m({gateway:t.toolGateway,approvals:t.approvals,workspace:t.workspace,emit:emit}),workspace:t.workspace,sandbox:t.sandbox,emit:emit}),h={...t,toolGateway:E},S=A(function readToolFailurePolicy(e){if("object"!=typeof e||null===e||Array.isArray(e))return;const t=e.failurePolicy;return"object"!=typeof t||null===t||Array.isArray(t)?void 0:t}(t.workspace.runtime.toolGateway));return{request:async t=>async function runRuntimeRequest(t){const p=t.request.requestId??e(),m=t.request.sessionId??e(),l=[],{agent:w,adapter:g}=await async function resolveExecution(e,t,r){const a=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const a=await k({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description})),trace:{...r,agentId:t,layer:"agent",owner:"stable_runtime_policy"}});return a.ok?a.id:t}(e.workspace,t.agentId,r):e.workspace.runtime.defaultAgentId,s=e.workspace.agents.get(a);if(!s)throw new Error(`Agent ${a} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const n=e.adapters.find(e=>e.canRun(s));if(!n)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:n}}(t.input,t.request,{requestId:p,sessionId:m,emit:e=>l.push(e)});t.store.createRun(function createRunRecord(e,t,r,a){return{requestId:t,sessionId:r,agentId:a.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,p,m,w)),l.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:p,sessionId:m,agentId:w.id,input:t.request.input});try{if(t.request.workflow){const e=await C({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:p,sessionId:m,agentId:w.id,emit:t.emit});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}if(t.request.toolCall){const e=await d({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:p,sessionId:m,agent:w,toolFailureTracker:t.toolFailureTracker,toolGuardrails:t.input.toolGuardrails,events:t.store.getRun(p)?.events??[]});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,c=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),p=c.memory,d=c.pluginMemories??[],m=I({workspace:e.input.workspace,agent:e.agent}),l=o(e.input.workspace.runtime,e.agent),w=new Map;let g;try{g=await runAdapterOnce(e,t,e.request,p,d,w,m)}catch(a){if(!s(a,m))throw a;e.emit(repairStarted(e,"adapter_error",1,errorMessage(a))),g=await runAdapterOnce(e,t,r(e.request,a,m),p,d,w,m),e.emit(repairCompleted(e,"adapter_error","retried",1,errorMessage(a)))}g=await i({...e,request:e.request,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),e.request,g,l),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:g});try{assertRequestExecutionContract(e)}catch(r){const s=a({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:m});if(!s)throw r;e.emit(repairStarted(e,"execution_contract",1,errorMessage(r))),g=await runAdapterOnce(e,t,s,p,d,w,m),g=await i({...e,request:s,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),s,g,l),assertRequestExecutionContract(e),e.emit(repairCompleted(e,"execution_contract","retried",1,errorMessage(r)))}const y=u({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:g,artifacts:e.input.artifacts});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:g,response:y}),y}({...t,adapter:g,requestId:p,sessionId:m,agent:w})}catch(e){return c({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,error:e})}}({input:h,capabilities:j,store:x,emit:emit,request:t,toolFailureTracker:S}),subscribe:e=>(y.add(e),()=>y.delete(e)),...w({workspace:t.workspace,store:x,artifacts:t.artifacts,approvals:t.approvals,emit:emit}),...p({store:x,emit:emit}),...q({memory:t.memory}),cancel(e,t){const r=x.getRun(e);r&&"running"===r.state&&(x.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await j.stop(),y.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,a,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:n=>runAdapterOnce(e,e.adapter,n,t,r,a,s),reviewModel:e.input.qualityReviewModel,executionEvaluatorRules:e.input.executionEvaluatorRules,memory:t,pluginMemories:r}}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,a,s,n,o){return y(await t.run({workspace:{...e.input.workspace,runtime:o},agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:a,pluginMemories:s,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,toolGuardrails:e.input.toolGuardrails,executionEvaluatorRules:e.input.executionEvaluatorRules,requestState:n,getEvents:()=>e.store.getRun(e.requestId)?.events??[],emit:e.emit}))}function repairStarted(e,t,r,a){return{type:"runtime.repair.started",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,attempt:r,reason:a}}function repairCompleted(e,t,r,a,s){return{type:"runtime.repair.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,outcome:r,attempt:a,reason:s}}function errorMessage(e){return e instanceof Error?e.message:String(e)}

package/node_modules/@stable-harness/core/dist/types.d.ts

@@ -4,6 +4,8 @@ import type { RuntimeWorkflowAdapter, RuntimeWorkflowRequest, WorkspaceWorkflow
 import type { SpecDrivenWorkflowState } from "./spec-driven/index.js";
 import type { RuntimeEvent, RuntimeEventListener, RuntimeEmit } from "./runtime/events.js";
 import type { RuntimeToolFailureTracker } from "./runtime/tool-failure.js";
+import type { ToolGuardrail } from "./runtime/policy/tool-invocation.js";
+import type { ExecutionEvaluatorRule } from "./quality/execution-review.js";
 import type { RuntimeArtifact, RuntimeArtifactFilter, RuntimeArtifactRecord, RuntimeOutput, RuntimeRecordState, RuntimeRequest, RuntimeResponse, RuntimeReplayBundle, RuntimeDeletionResult, RuntimeRunFilter, RuntimeRunRecord } from "./runtime/types.js";
 import type { RuntimeToolGateway } from "./runtime/tool-gateway.js";
 import type { CompiledWorkspace, WorkspaceAgent, WorkspaceRuntimePolicy } from "./workspace/types.js";
@@ -28,6 +30,8 @@ export type RuntimeAdapterContext = {
     pluginMemories?: RuntimeMemoryContext[];
     toolGateway?: RuntimeToolGateway;
     toolFailureTracker?: RuntimeToolFailureTracker;
+    toolGuardrails?: readonly ToolGuardrail[];
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     requestState?: Map<string, unknown>;
     getEvents?: () => RuntimeEvent[];
     emit: RuntimeEmit;

package/node_modules/@stable-harness/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/core",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -11,7 +11,7 @@
     ".": "./dist/index.js"
   },
   "peerDependencies": {
-    "@stable-harness/governance": "0.0.80",
-    "@stable-harness/memory": "0.0.80"
+    "@stable-harness/governance": "0.0.81",
+    "@stable-harness/memory": "0.0.81"
   }
 }

package/node_modules/@stable-harness/governance/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/governance",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",

package/node_modules/@stable-harness/memory/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/memory",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",

package/node_modules/@stable-harness/protocols/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/protocols",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -10,6 +10,6 @@
   "main": "dist/src/index.js",
   "types": "dist/src/index.d.ts",
   "peerDependencies": {
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }

package/node_modules/@stable-harness/tool-gateway/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/tool-gateway",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",

package/node_modules/@stable-harness/workspace-yaml/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/workspace-yaml",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -11,6 +11,6 @@
     ".": "./dist/index.js"
   },
   "peerDependencies": {
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stable-harness",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "description": "Stable application runtime and operator control plane for agent workspaces.",
   "license": "Apache-2.0",
@@ -82,14 +82,14 @@
     "@langchain/node-vfs": "^0.1.4",
     "@langchain/ollama": "^1.2.7",
     "@langchain/openai": "^1.4.5",
-    "@stable-harness/adapter-deepagents": "0.0.80",
-    "@stable-harness/adapter-langgraph": "0.0.80",
-    "@stable-harness/core": "0.0.80",
-    "@stable-harness/governance": "0.0.80",
-    "@stable-harness/memory": "0.0.80",
-    "@stable-harness/protocols": "0.0.80",
-    "@stable-harness/tool-gateway": "0.0.80",
-    "@stable-harness/workspace-yaml": "0.0.80",
+    "@stable-harness/adapter-deepagents": "0.0.81",
+    "@stable-harness/adapter-langgraph": "0.0.81",
+    "@stable-harness/core": "0.0.81",
+    "@stable-harness/governance": "0.0.81",
+    "@stable-harness/memory": "0.0.81",
+    "@stable-harness/protocols": "0.0.81",
+    "@stable-harness/tool-gateway": "0.0.81",
+    "@stable-harness/workspace-yaml": "0.0.81",
     "deepagents": "^1.10.1",
     "langchain": "^1.4.0",
     "yaml": "^2.8.2",

package/packages/adapter-deepagents/dist/src/internal/gateway-tools.js

@@ -1 +1 @@
-import{ToolMessage as t}from"@langchain/core/messages";import{tool as e}from"@langchain/core/tools";import{afterToolInvoke as o,beforeToolInvoke as r,createToolRepeatState as n,missingRequiredPlanContent as s,missingToolDependencyContent as a,toolInvocationEvents as i}from"@stable-harness/core";import{isSuccessfulEvidenceOutput as u,observedToolEvidence as l,recordObservedToolEvidence as c}from"./gateway/tool-evidence.js";import{emitStructuredToolFailure as d}from"./gateway/tool-failure-events.js";export function buildGatewayTools(l,c,p,f,g=n(l.workspace.runtime.toolGateway)){return l.toolGateway?p.flatMap(n=>{const p=l.toolGateway?.get(n);if(!p)return[];const m=l.workspace.tools.get(n),y=m?.schema??p.schema;return[e(async e=>async function invokeGuardedGatewayTool(e){emitToolResult(e.input,e.agentId,e.toolId,void 0);const n=i({request:e.input.request,getEvents:e.input.getEvents}),l=s({agent:e.input.agent,events:n,toolId:e.toolId});if(l)return emitToolResult(e.input,e.agentId,e.toolId,l),l;const c=a({agent:e.input.agent,events:n,toolId:e.toolId});if(c)return emitToolResult(e.input,e.agentId,e.toolId,c),c;const p=e.repeatState?r(e.toolId,e.args,e.repeatState):void 0;if(p)return emitToolResult(e.input,e.agentId,e.toolId,p.eventOutput),p.modelOutput;const f=await async function invokeGatewayTool(e,o,r,n,s){try{if(e.toolFailureTracker?.isCircuitOpen(r))throw new Error(`Tool circuit is open: ${r}`);const t=await e.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:e.workspace.root,requestId:e.requestId,sessionId:e.sessionId,agentId:o,requestInput:e.request.input,observedEvidence:formatObservedEvidenceForToolContext(e),approvalIds:readApprovalIds(e.request.metadata)}});return e.toolFailureTracker?.recordSuccess(r),t}catch(n){if(d(e,o,r,n),function isToolArgumentValidationError(t){return t instanceof Error&&"ToolArgumentValidationError"===t.name&&"string"==typeof t.toolId&&Array.isArray(t.issues)}(n))return new t({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(n)});if(e.workspace.runtime.retry?.tools?.enabled)throw n;return new t({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(n),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(e.input,e.agentId,e.toolId,e.args,e.repairModel),g=f instanceof t?String(f.content):stringifyDeepAgentResult(f.output),m=e.repeatState?o({toolId:e.toolId,args:e.args,output:g,successful:!(f instanceof t&&"error"===f.status)&&u(g),state:e.repeatState}):{};return emitToolResult(e.input,e.agentId,e.toolId,m.eventOutput??g),void 0!==m.modelOutput?m.modelOutput:f instanceof t?f:g}({input:l,agentId:c,toolId:n,args:e,repairModel:f,repeatState:g}),{name:n,description:buildToolDescription(m?.description??p.description??n,y,l.workspace.runtime.toolGateway,n),schema:{type:"object",additionalProperties:!0}})]}):[]}function emitToolResult(t,e,o,r){void 0!==r&&c(t,e,o,r),t.emit({type:"runtime.adapter.event",requestId:t.requestId,sessionId:t.sessionId,agentId:e,event:void 0===r?{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.start",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.result",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),evidenceOutput:r,...toolControlProjection(r)}})}export function stringifyDeepAgentResult(e){if(e instanceof t)return function stringifyToolMessageContent(t){return"string"==typeof t?t:JSON.stringify(t)}(e.content);if("string"==typeof e)return e;if(isRecord(e)){const t=e.structuredResponse??e.structured_response;if(void 0!==t)return"string"==typeof t?t:JSON.stringify(t);const o=(Array.isArray(e.messages)?e.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(e.update)&&Array.isArray(e.update.messages)?e.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(e)}function buildToolDescription(t,e,o,r){const n=function toolRepeatPolicyDescription(t,e){const o=function repeatGuardConfig(t){return isRecord(t)&&isRecord(t.repeatGuard)?t.repeatGuard:{}}(t),r=function readPositiveIntegerMap(t){return isRecord(t)?new Map(Object.entries(t).map(([t,e])=>[t,readPositiveInteger(e)]).filter(t=>void 0!==t[1])):new Map}(o.maxSuccessfulCallsByTool).get(e)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${t}\n\n${n}`:t;return e?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(e))}`:s}function readPositiveInteger(t){return"number"==typeof t&&Number.isInteger(t)&&t>0?t:void 0}function previewToolOutput(t){const e=t.replace(/\s+/gu," ").trim();return e.length>500?`${e.slice(0,497)}...`:e}export function toolControlProjection(t){const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);if("string"==typeof e?.status)return{controlStatus:e.status};const o=function readTextStatus(t){return String(t).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(t);return o?{controlStatus:o}:"string"==typeof e?.error?{controlStatus:e.error}:t.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function readApprovalIds(t){const e=t?.approvalIds??t?.approvalId;return"string"==typeof e&&e.trim()?[e.trim()]:Array.isArray(e)?e.filter(t=>"string"==typeof t&&t.trim().length>0):void 0}function formatObservedEvidenceForToolContext(t){const e=l(t).map(t=>`Tool: ${t.toolId}\n${t.output}`).join("\n\n---\n\n");return e.length>12e3?`${e.slice(0,12e3)}\n[truncated]`:e}function formatToolArgumentError(t){return JSON.stringify({error:"tool_argument_validation_failed",toolId:t.toolId,issues:t.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function formatError(t){return t instanceof Error?t.message:String(t)}
+import{ToolMessage as t}from"@langchain/core/messages";import{tool as e}from"@langchain/core/tools";import{afterToolInvoke as o,createToolRepeatState as r,evaluateToolGuardrails as n,toolInvocationEvents as s}from"@stable-harness/core";import{isSuccessfulEvidenceOutput as a,observedToolEvidence as i,recordObservedToolEvidence as u}from"./gateway/tool-evidence.js";import{emitStructuredToolFailure as l}from"./gateway/tool-failure-events.js";export function buildGatewayTools(i,u,c,d,p=r(i.workspace.runtime.toolGateway)){return i.toolGateway?c.flatMap(r=>{const c=i.toolGateway?.get(r);if(!c)return[];const f=i.workspace.tools.get(r),g=f?.schema??c.schema;return[e(async e=>async function invokeGuardedGatewayTool(e){emitToolResult(e.input,e.agentId,e.toolId,void 0);const r=s({request:e.input.request,getEvents:e.input.getEvents}),i=n({agent:e.input.agent,args:e.args,events:r,repeatState:e.repeatState,toolId:e.toolId},e.input.toolGuardrails);if(i)return emitToolResult(e.input,e.agentId,e.toolId,i.eventOutput),i.modelOutput;const u=await async function invokeGatewayTool(e,o,r,n,s){try{if(e.toolFailureTracker?.isCircuitOpen(r))throw new Error(`Tool circuit is open: ${r}`);const t=await e.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:e.workspace.root,requestId:e.requestId,sessionId:e.sessionId,agentId:o,requestInput:e.request.input,observedEvidence:formatObservedEvidenceForToolContext(e),approvalIds:readApprovalIds(e.request.metadata)}});return e.toolFailureTracker?.recordSuccess(r),t}catch(n){if(l(e,o,r,n),function isToolArgumentValidationError(t){return t instanceof Error&&"ToolArgumentValidationError"===t.name&&"string"==typeof t.toolId&&Array.isArray(t.issues)}(n))return new t({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(n)});if(e.workspace.runtime.retry?.tools?.enabled)throw n;return new t({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(n),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(e.input,e.agentId,e.toolId,e.args,e.repairModel),c=u instanceof t?String(u.content):stringifyDeepAgentResult(u.output),d=e.repeatState?o({toolId:e.toolId,args:e.args,output:c,successful:!(u instanceof t&&"error"===u.status)&&a(c),state:e.repeatState}):{};return emitToolResult(e.input,e.agentId,e.toolId,d.eventOutput??c),void 0!==d.modelOutput?d.modelOutput:u instanceof t?u:c}({input:i,agentId:u,toolId:r,args:e,repairModel:d,repeatState:p}),{name:r,description:buildToolDescription(f?.description??c.description??r,g,i.workspace.runtime.toolGateway,r),schema:{type:"object",additionalProperties:!0}})]}):[]}function emitToolResult(t,e,o,r){void 0!==r&&u(t,e,o,r),t.emit({type:"runtime.adapter.event",requestId:t.requestId,sessionId:t.sessionId,agentId:e,event:void 0===r?{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.start",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",eventGroup:"tool_execution",eventType:"deepagents.tool_execution.result",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),evidenceOutput:r,...toolControlProjection(r)}})}export function stringifyDeepAgentResult(e){if(e instanceof t)return function stringifyToolMessageContent(t){return"string"==typeof t?t:JSON.stringify(t)}(e.content);if("string"==typeof e)return e;if(isRecord(e)){const t=e.structuredResponse??e.structured_response;if(void 0!==t)return"string"==typeof t?t:JSON.stringify(t);const o=(Array.isArray(e.messages)?e.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(e.update)&&Array.isArray(e.update.messages)?e.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(e)}function buildToolDescription(t,e,o,r){const n=function toolRepeatPolicyDescription(t,e){const o=function repeatGuardConfig(t){return isRecord(t)&&isRecord(t.repeatGuard)?t.repeatGuard:{}}(t),r=function readPositiveIntegerMap(t){return isRecord(t)?new Map(Object.entries(t).map(([t,e])=>[t,readPositiveInteger(e)]).filter(t=>void 0!==t[1])):new Map}(o.maxSuccessfulCallsByTool).get(e)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${t}\n\n${n}`:t;return e?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(e))}`:s}function readPositiveInteger(t){return"number"==typeof t&&Number.isInteger(t)&&t>0?t:void 0}function previewToolOutput(t){const e=t.replace(/\s+/gu," ").trim();return e.length>500?`${e.slice(0,497)}...`:e}export function toolControlProjection(t){const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);if("string"==typeof e?.status)return{controlStatus:e.status};const o=function readTextStatus(t){return String(t).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(t);return o?{controlStatus:o}:"string"==typeof e?.error?{controlStatus:e.error}:t.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function readApprovalIds(t){const e=t?.approvalIds??t?.approvalId;return"string"==typeof e&&e.trim()?[e.trim()]:Array.isArray(e)?e.filter(t=>"string"==typeof t&&t.trim().length>0):void 0}function formatObservedEvidenceForToolContext(t){const e=i(t).map(t=>`Tool: ${t.toolId}\n${t.output}`).join("\n\n---\n\n");return e.length>12e3?`${e.slice(0,12e3)}\n[truncated]`:e}function formatToolArgumentError(t){return JSON.stringify({error:"tool_argument_validation_failed",toolId:t.toolId,issues:t.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function formatError(t){return t instanceof Error?t.message:String(t)}

package/packages/adapter-deepagents/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/adapter-deepagents",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -15,7 +15,7 @@
     "@langchain/node-vfs": "^0.1.4",
     "@langchain/ollama": "^1.2.7",
     "@langchain/openai": "^1.4.5",
-    "@stable-harness/core": "0.0.80",
+    "@stable-harness/core": "0.0.81",
     "deepagents": "^1.10.1",
     "langchain": "^1.4.0"
   },

package/packages/adapter-langgraph/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/adapter-langgraph",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -11,6 +11,6 @@
   "types": "dist/src/index.d.ts",
   "peerDependencies": {
     "@langchain/langgraph": "^1.3.0",
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }

package/packages/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/cli",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -14,12 +14,12 @@
   "types": "dist/src/index.d.ts",
   "peerDependencies": {
     "@langchain/langgraph-api": "^1.2.1",
-    "@stable-harness/adapter-deepagents": "0.0.80",
-    "@stable-harness/adapter-langgraph": "0.0.80",
-    "@stable-harness/core": "0.0.80",
-    "@stable-harness/memory": "0.0.80",
-    "@stable-harness/protocols": "0.0.80",
-    "@stable-harness/tool-gateway": "0.0.80",
-    "@stable-harness/workspace-yaml": "0.0.80"
+    "@stable-harness/adapter-deepagents": "0.0.81",
+    "@stable-harness/adapter-langgraph": "0.0.81",
+    "@stable-harness/core": "0.0.81",
+    "@stable-harness/memory": "0.0.81",
+    "@stable-harness/protocols": "0.0.81",
+    "@stable-harness/tool-gateway": "0.0.81",
+    "@stable-harness/workspace-yaml": "0.0.81"
   }
 }

package/packages/core/dist/quality/execution-review.d.ts

@@ -1,2 +1,5 @@
-import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
-export declare function reviewExecutionEvidence(input: QualityReviewInput, policy: QualityPolicy): QualityReviewResult;
+import type { QualityPolicy, QualityReviewInput, QualityReviewIssue, QualityReviewResult } from "./types.js";
+export type ExecutionEvaluatorRule = (input: QualityReviewInput, policy: QualityPolicy) => QualityReviewIssue[];
+export declare const defaultExecutionEvaluatorRules: readonly ExecutionEvaluatorRule[];
+export declare function reviewExecutionEvidence(input: QualityReviewInput, policy: QualityPolicy, rules?: readonly ExecutionEvaluatorRule[]): QualityReviewResult;
+export declare function evaluateExecutionRules(input: QualityReviewInput, policy: QualityPolicy, rules?: readonly ExecutionEvaluatorRule[]): QualityReviewIssue[];

package/packages/core/dist/quality/execution-review.js

@@ -1 +1 @@
-import{controlBlockers as e,controlGaps as n,successfulEvidenceOutputs as t,successfulEvidenceToolIds as r}from"./event-evidence.js";const s=/(?<![\w.])(?:\d{1,3}(?:,\d{3})+|\d+)(?:\.\d+)?[%kKmMbBtTxX]?(?!\w)/gu;export function reviewExecutionEvidence(e,n){if(!n.enabled||!n.executionReview.enabled)return{verdict:"pass",issues:[]};const t=[...blockerIssues(e,n),...controlGapIssues(e,n),...emptyFinalIssues(e,n),...toolEvidenceIssues(e,n),...ungroundedNumberIssues(e,n)];return 0===t.length?{verdict:"pass",issues:[]}:{verdict:t.some(e=>!e.recoverable)?"blocked":"continue_react",issues:t}}function ungroundedNumberIssues(e,n){if(!n.executionReview.rejectUngroundedNumbers||!e.output?.text.trim())return[];const r=numberSet(t(e.events).join("\n"));if(0===r.size)return[];const s=[...numberSet(e.output.text)].filter(e=>!function isSupportedNumber(e,n){if(n.has(e))return!0;const t=Number.parseFloat(e);if(!Number.isFinite(t))return!1;for(const e of n){const n=Number.parseFloat(e);if(Number.isFinite(n)&&Math.abs(n-t)<=roundingTolerance(t))return!0}return!1}(e,r));return 0===s.length?[]:[{code:"ungrounded_numeric_claim",message:`Final answer contains numeric claims not found in successful tool evidence: ${s.slice(0,12).join(", ")}`,recoverable:!1}]}function numberSet(e){const n=new Set;for(const t of e.matchAll(s)){const e=normalizeNumber(t[0]);e&&n.add(e)}return n}function normalizeNumber(e){const n=e.replace(/,/gu,"").replace(/^\+/u,"").replace(/[%kKmMbBtTxX]$/u,"").trim();if(n){if(/^\d+$/u.test(n)){const e=Number.parseInt(n,10);if(e>=1&&e<=20)return;return String(e)}return/^\d+\.\d+$/u.test(n)?n.replace(/0+$/u,"").replace(/\.$/u,""):void 0}}function roundingTolerance(e){return Math.abs(e)>=1e3?1:Math.abs(e)>=100?.1:Math.abs(e)>=10?.05:.005}function blockerIssues(n,t){return t.executionReview.stopOnBlocker?e(n.events).map(e=>({code:"control_blocker",message:`Execution produced a control blocker: ${e}`,recoverable:!1})):[]}function controlGapIssues(e,t){if(!t.executionReview.stopOnBlocker||!e.output?.text.trim())return[];const r=n(e.events).filter(n=>!function mentionsGap(e,n){const[t,r]=n.split(":"),s=e.toLowerCase();return s.includes(n.toLowerCase())||Boolean(t&&r&&s.includes(t.toLowerCase())&&s.includes(r.toLowerCase()))}(e.output?.text??"",n));return 0===r.length?[]:[{code:"unresolved_control_gap",message:`Final answer omitted unresolved runtime evidence gap(s): ${r.slice(0,8).join(", ")}`,recoverable:!0}]}function emptyFinalIssues(e,n){return!n.executionReview.rejectEmptyFinal||e.output?.text.trim()?[]:[{code:"empty_final_answer",message:"The final answer is empty.",recoverable:!0}]}function toolEvidenceIssues(e,n){return!n.executionReview.requireToolEvidence||r(e.events).length>0?[]:[{code:"missing_tool_evidence",message:"No successful tool or delegated-task evidence was observed.",recoverable:!0}]}
+import{controlBlockers as e,controlGaps as t,successfulEvidenceOutputs as n,successfulEvidenceToolIds as o}from"./event-evidence.js";const r=/(?<![\w.])(?:\d{1,3}(?:,\d{3})+|\d+)(?:\.\d+)?[%kKmMbBtTxX]?(?!\w)/gu;export const defaultExecutionEvaluatorRules=[function blockerIssues(t,n){return n.executionReview.stopOnBlocker?e(t.events).map(e=>({code:"control_blocker",message:`Execution produced a control blocker: ${e}`,recoverable:!1})):[]},function controlGapIssues(e,n){if(!n.executionReview.stopOnBlocker||!e.output?.text.trim())return[];const o=t(e.events).filter(t=>!function mentionsGap(e,t){const[n,o]=t.split(":"),r=e.toLowerCase();return r.includes(t.toLowerCase())||Boolean(n&&o&&r.includes(n.toLowerCase())&&r.includes(o.toLowerCase()))}(e.output?.text??"",t));return 0===o.length?[]:[{code:"unresolved_control_gap",message:`Final answer omitted unresolved runtime evidence gap(s): ${o.slice(0,8).join(", ")}`,recoverable:!0}]},function emptyFinalIssues(e,t){return!t.executionReview.rejectEmptyFinal||e.output?.text.trim()?[]:[{code:"empty_final_answer",message:"The final answer is empty.",recoverable:!0}]},function toolEvidenceIssues(e,t){return!t.executionReview.requireToolEvidence||o(e.events).length>0?[]:[{code:"missing_tool_evidence",message:"No successful tool or delegated-task evidence was observed.",recoverable:!0}]},function ungroundedNumberIssues(e,t){if(!t.executionReview.rejectUngroundedNumbers||!e.output?.text.trim())return[];const o=numberSet(n(e.events).join("\n"));if(0===o.size)return[];const r=[...numberSet(e.output.text)].filter(e=>!function isSupportedNumber(e,t){if(t.has(e))return!0;const n=Number.parseFloat(e);if(!Number.isFinite(n))return!1;for(const e of t){const t=Number.parseFloat(e);if(Number.isFinite(t)&&Math.abs(t-n)<=roundingTolerance(n))return!0}return!1}(e,o));return 0===r.length?[]:[{code:"ungrounded_numeric_claim",message:`Final answer contains numeric claims not found in successful tool evidence: ${r.slice(0,12).join(", ")}`,recoverable:!1}]}];export function reviewExecutionEvidence(e,t,n=defaultExecutionEvaluatorRules){if(!t.enabled||!t.executionReview.enabled)return{verdict:"pass",issues:[]};const o=evaluateExecutionRules(e,t,n);return 0===o.length?{verdict:"pass",issues:[]}:{verdict:o.some(e=>!e.recoverable)?"blocked":"continue_react",issues:o}}export function evaluateExecutionRules(e,t,n=defaultExecutionEvaluatorRules){return n.flatMap(n=>n(e,t))}function numberSet(e){const t=new Set;for(const n of e.matchAll(r)){const e=normalizeNumber(n[0]);e&&t.add(e)}return t}function normalizeNumber(e){const t=e.replace(/,/gu,"").replace(/^\+/u,"").replace(/[%kKmMbBtTxX]$/u,"").trim();if(t){if(/^\d+$/u.test(t)){const e=Number.parseInt(t,10);if(e>=1&&e<=20)return;return String(e)}return/^\d+\.\d+$/u.test(t)?t.replace(/0+$/u,"").replace(/\.$/u,""):void 0}}function roundingTolerance(e){return Math.abs(e)>=1e3?1:Math.abs(e)>=100?.1:Math.abs(e)>=10?.05:.005}

package/packages/core/dist/quality/runtime.d.ts

@@ -1,4 +1,5 @@
 import type { RuntimeMemoryContext, RuntimeOutput, RuntimeRequest } from "../types.js";
+import { type ExecutionEvaluatorRule } from "./execution-review.js";
 import type { QualityPolicy, QualityReviewInput, QualityReviewModel } from "./types.js";
 export type QualityRuntimeInput = QualityReviewInput & {
     requestId: string;
@@ -7,6 +8,7 @@ export type QualityRuntimeInput = QualityReviewInput & {
     getEvents: () => import("../types.js").RuntimeEvent[];
     runAdapter: (request: RuntimeRequest) => Promise<RuntimeOutput>;
     reviewModel?: QualityReviewModel;
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     memory?: RuntimeMemoryContext;
     pluginMemories: RuntimeMemoryContext[];
 };

package/packages/core/dist/quality/runtime.js

@@ -1 +1 @@
-import{successfulEvidenceOutputs as e}from"./event-evidence.js";import{buildQualityRecoveryRequest as t}from"./recovery-policy.js";import{reviewExecutionEvidence as i}from"./execution-review.js";import{reviewWithLlm as n}from"./llm-review.js";import{reviewPlanningEvidence as r}from"./planning-review.js";import{synthesizeEvidenceOnlyReport as s}from"./synthesis.js";export async function recoverQualityReview(e,t,i,n){if(!n.enabled)return i;let r=t,s=i;for(let t=0;t<n.recovery.maxLoops+1;t+=1){const i=await emitPlanningReview(e,r,s,n);if("blocked"===i.verdict)return qualityFailureOutput("planning",i);const u=buildQualityRecovery(e,r,i,"planning",n,t);if(u){r=u,s=await e.runAdapter(r);continue}const o=await emitExecutionReview(e,r,s,n);if("pass"!==o.verdict){const t=await trySynthesizeExecution(e,r,o,n);if(t)return t}const a=buildQualityRecovery(e,r,o,"execution",n,t);if(!a)return"pass"===o.verdict?s:await trySynthesizeExecution(e,r,o,n)??qualityFailureOutput("execution",o);r=a,s=await e.runAdapter(r)}return qualityFailureOutput("execution",{verdict:"blocked",issues:[{code:"quality_recovery_exhausted",message:`Quality recovery exceeded maxLoops=${n.recovery.maxLoops}.`,recoverable:!1}]})}async function trySynthesizeExecution(e,t,n,r){const u=s({...reviewInputFor(e,t),output:void 0},n,r);if(!u)return;e.emit({type:"runtime.quality.synthesis.created",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,mode:r.synthesis.mode});const o={text:u},a=i({...reviewInputFor(e,t),output:o},r);return emitReviewEvent(e,"execution",a),"pass"===a.verdict?o:void 0}function emitPlanningReview(e,t,i,n){return emitReview(e,"planning",r,t,i,n)}function emitExecutionReview(e,t,n,r){return emitReview(e,"execution",i,t,n,r)}async function emitReview(e,t,i,r,s,u){const o={...reviewInputFor(e,r),output:s},a="planning"===t?u.planningReview.enabled:u.executionReview.enabled;if(!a)return i(o,u);const c=i(o,u),d=await n({phase:t,review:o,policy:u,model:e.reviewModel}),v="pass"===c.verdict?d??c:c;return a&&emitReviewEvent(e,t,v),v}function emitReviewEvent(e,t,i){"planning"!==t?e.emit({type:"runtime.quality.execution.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues}):e.emit({type:"runtime.quality.planning.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues})}function buildQualityRecovery(i,n,r,s,u,o){if(o>=u.recovery.maxLoops)return;const a=t({request:n,result:r,phase:s,policy:u,availableToolIds:i.agent.tools,availableSubagentIds:i.agent.subagents,observedEvidence:"execution"===s?e(i.getEvents()):[]});return a&&i.emit({type:"runtime.quality.recovery.started",requestId:i.requestId,sessionId:i.sessionId,agentId:i.agent.id,phase:s,attempt:o+1,verdict:r.verdict}),a}function reviewInputFor(e,t){return{workspace:e.workspace,agent:e.agent,request:t,events:e.getEvents()}}function qualityFailureOutput(e,t){return{text:[`Stable runtime quality review blocked final delivery during ${e}.`,"",...t.issues.length>0?t.issues.map(e=>`- ${e.code}: ${e.message}`):["- quality_review_failed: Quality review did not pass."]].join("\n")}}
+import{successfulEvidenceOutputs as e}from"./event-evidence.js";import{buildQualityRecoveryRequest as t}from"./recovery-policy.js";import{reviewExecutionEvidence as i}from"./execution-review.js";import{reviewWithLlm as n}from"./llm-review.js";import{reviewPlanningEvidence as r}from"./planning-review.js";import{synthesizeEvidenceOnlyReport as s}from"./synthesis.js";export async function recoverQualityReview(e,t,i,n){if(!n.enabled)return i;let r=t,s=i;for(let t=0;t<n.recovery.maxLoops+1;t+=1){const i=await emitPlanningReview(e,r,s,n);if("blocked"===i.verdict)return qualityFailureOutput("planning",i);const u=buildQualityRecovery(e,r,i,"planning",n,t);if(u){r=u,s=await e.runAdapter(r);continue}const o=await emitExecutionReview(e,r,s,n);if("pass"!==o.verdict){const t=await trySynthesizeExecution(e,r,o,n);if(t)return t}const a=buildQualityRecovery(e,r,o,"execution",n,t);if(!a)return"pass"===o.verdict?s:await trySynthesizeExecution(e,r,o,n)??qualityFailureOutput("execution",o);r=a,s=await e.runAdapter(r)}return qualityFailureOutput("execution",{verdict:"blocked",issues:[{code:"quality_recovery_exhausted",message:`Quality recovery exceeded maxLoops=${n.recovery.maxLoops}.`,recoverable:!1}]})}async function trySynthesizeExecution(e,t,n,r){const u=s({...reviewInputFor(e,t),output:void 0},n,r);if(!u)return;e.emit({type:"runtime.quality.synthesis.created",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,mode:r.synthesis.mode});const o={text:u},a=i({...reviewInputFor(e,t),output:o},r,e.executionEvaluatorRules);return emitReviewEvent(e,"execution",a),"pass"===a.verdict?o:void 0}function emitPlanningReview(e,t,i,n){return emitReview(e,"planning",r,t,i,n)}function emitExecutionReview(e,t,n,r){return emitReview(e,"execution",(t,n)=>i(t,n,e.executionEvaluatorRules),t,n,r)}async function emitReview(e,t,i,r,s,u){const o={...reviewInputFor(e,r),output:s},a="planning"===t?u.planningReview.enabled:u.executionReview.enabled;if(!a)return i(o,u);const c=i(o,u),d=await n({phase:t,review:o,policy:u,model:e.reviewModel}),v="pass"===c.verdict?d??c:c;return a&&emitReviewEvent(e,t,v),v}function emitReviewEvent(e,t,i){"planning"!==t?e.emit({type:"runtime.quality.execution.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues}):e.emit({type:"runtime.quality.planning.reviewed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,verdict:i.verdict,issues:i.issues})}function buildQualityRecovery(i,n,r,s,u,o){if(o>=u.recovery.maxLoops)return;const a=t({request:n,result:r,phase:s,policy:u,availableToolIds:i.agent.tools,availableSubagentIds:i.agent.subagents,observedEvidence:"execution"===s?e(i.getEvents()):[]});return a&&i.emit({type:"runtime.quality.recovery.started",requestId:i.requestId,sessionId:i.sessionId,agentId:i.agent.id,phase:s,attempt:o+1,verdict:r.verdict}),a}function reviewInputFor(e,t){return{workspace:e.workspace,agent:e.agent,request:t,events:e.getEvents()}}function qualityFailureOutput(e,t){return{text:[`Stable runtime quality review blocked final delivery during ${e}.`,"",...t.issues.length>0?t.issues.map(e=>`- ${e.code}: ${e.message}`):["- quality_review_failed: Quality review did not pass."]].join("\n")}}

package/packages/core/dist/runtime/direct-tool-call.d.ts

@@ -1,4 +1,5 @@
 import type { CompiledWorkspace, RuntimeEvent, RuntimeOutput, RuntimeRequest, RuntimeToolFailureTracker, RuntimeToolGateway, WorkspaceAgent } from "../types.js";
+import { type ToolGuardrail } from "./policy/tool-invocation.js";
 export declare function runDirectToolCall(input: {
     gateway: RuntimeToolGateway | undefined;
     workspace: CompiledWorkspace;
@@ -8,4 +9,6 @@ export declare function runDirectToolCall(input: {
     sessionId: string;
     agent: WorkspaceAgent;
     toolFailureTracker?: RuntimeToolFailureTracker;
+    toolGuardrails?: readonly ToolGuardrail[];
+    events?: RuntimeEvent[];
 }): Promise<RuntimeOutput>;

package/packages/core/dist/runtime/direct-tool-call.js

@@ -1 +1 @@
-import{toolCircuitOpenEvent as o,toolFailureEvent as t}from"./tool-failure.js";export async function runDirectToolCall(o){const t=o.request.toolCall;if(!t)throw new Error("Direct tool call request is missing");if(!o.gateway)throw new Error("Runtime tool gateway is not configured");const e=await async function resolveDirectToolCall(o){if(o.agent.tools.includes(o.toolId)&&o.gateway.get(o.toolId))return{toolId:o.toolId,args:o.args};const t=await(o.gateway.repairToolCall?.({toolId:o.toolId,args:o.args,allowedToolIds:o.agent.tools,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,requestInput:o.request.input,approvalIds:readApprovalIds(o.request.metadata)}}));if(t&&o.agent.tools.includes(t.toolId)&&o.gateway.get(t.toolId))return emitToolRepair(o,"repaired",t.toolId),t;if(!o.agent.tools.includes(o.toolId))throw emitToolRepair(o,"blocked",void 0,`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`),new Error(`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`);throw emitToolRepair(o,"blocked",void 0,`Tool is not registered: ${o.toolId}`),new Error(`Tool is not registered: ${o.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,emit:o.emit,request:o.request,toolId:t.toolId,args:t.args});if(o.emit({type:"runtime.tool.direct.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:e.toolId}),o.toolFailureTracker?.isCircuitOpen(e.toolId)){const t=new Error(`Tool circuit is open: ${e.toolId}`);throw emitToolFailure(o,e.toolId,t),t}const r=await async function invokeToolWithFailureEvents(o,t){try{return await o.gateway.invoke({toolId:t.toolId,args:t.args,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,requestInput:o.request.input,approvalIds:readApprovalIds(o.request.metadata)}})}catch(e){throw emitToolFailure(o,t.toolId,e),e}}(o,e);return o.toolFailureTracker?.recordSuccess(r.toolId),o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:r.toolId,output:r.output}),{text:(s=r.output,"string"==typeof s?s:JSON.stringify(s)),metadata:{toolCall:{toolId:r.toolId}}};var s}function emitToolFailure(e,r,s){const a=t({requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,toolId:r,error:s});e.emit(a),e.toolFailureTracker?.recordFailure(r)&&e.emit(o({requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,toolId:r,reason:"runtime.tool.failure"===a.type?a.failure.reason:"unknown"}))}function emitToolRepair(o,t,e,r){o.emit({type:"runtime.inventory.repair",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,status:t,diagnostic:{layer:"tool",owner:"stable_runtime_policy",originalId:o.toolId,repairedId:e,candidateIds:o.agent.tools,reason:r}})}function readApprovalIds(o){const t=o?.approvalIds??o?.approvalId;return"string"==typeof t&&t.trim()?[t.trim()]:Array.isArray(t)?t.filter(o=>"string"==typeof o&&o.trim().length>0):void 0}
+import{evaluateToolGuardrails as t}from"./policy/tool-invocation.js";import{toolCircuitOpenEvent as o,toolFailureEvent as e}from"./tool-failure.js";export async function runDirectToolCall(o){const e=o.request.toolCall;if(!e)throw new Error("Direct tool call request is missing");if(!o.gateway)throw new Error("Runtime tool gateway is not configured");const r=await async function resolveDirectToolCall(t){if(t.agent.tools.includes(t.toolId)&&t.gateway.get(t.toolId))return{toolId:t.toolId,args:t.args};const o=await(t.gateway.repairToolCall?.({toolId:t.toolId,args:t.args,allowedToolIds:t.agent.tools,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,requestInput:t.request.input,approvalIds:readApprovalIds(t.request.metadata)}}));if(o&&t.agent.tools.includes(o.toolId)&&t.gateway.get(o.toolId))return emitToolRepair(t,"repaired",o.toolId),o;if(!t.agent.tools.includes(t.toolId))throw emitToolRepair(t,"blocked",void 0,`Tool ${t.toolId} is not assigned to agent ${t.agent.id}`),new Error(`Tool ${t.toolId} is not assigned to agent ${t.agent.id}`);throw emitToolRepair(t,"blocked",void 0,`Tool is not registered: ${t.toolId}`),new Error(`Tool is not registered: ${t.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,emit:o.emit,request:o.request,toolId:e.toolId,args:e.args});o.emit({type:"runtime.tool.direct.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:r.toolId});const s=t({agent:o.agent,args:r.args,events:o.events??[],toolId:r.toolId},o.toolGuardrails);if(s)return o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:r.toolId,output:s.eventOutput}),{text:s.modelOutput,metadata:{toolCall:{toolId:r.toolId},controlStatus:s.status}};if(o.toolFailureTracker?.isCircuitOpen(r.toolId)){const t=new Error(`Tool circuit is open: ${r.toolId}`);throw emitToolFailure(o,r.toolId,t),t}const a=await async function invokeToolWithFailureEvents(t,o){try{return await t.gateway.invoke({toolId:o.toolId,args:o.args,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,requestInput:t.request.input,approvalIds:readApprovalIds(t.request.metadata)}})}catch(e){throw emitToolFailure(t,o.toolId,e),e}}(o,r);return o.toolFailureTracker?.recordSuccess(a.toolId),o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:a.toolId,output:a.output}),{text:(i=a.output,"string"==typeof i?i:JSON.stringify(i)),metadata:{toolCall:{toolId:a.toolId}}};var i}function emitToolFailure(t,r,s){const a=e({requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,toolId:r,error:s});t.emit(a),t.toolFailureTracker?.recordFailure(r)&&t.emit(o({requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,toolId:r,reason:"runtime.tool.failure"===a.type?a.failure.reason:"unknown"}))}function emitToolRepair(t,o,e,r){t.emit({type:"runtime.inventory.repair",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,status:o,diagnostic:{layer:"tool",owner:"stable_runtime_policy",originalId:t.toolId,repairedId:e,candidateIds:t.agent.tools,reason:r}})}function readApprovalIds(t){const o=t?.approvalIds??t?.approvalId;return"string"==typeof o&&o.trim()?[o.trim()]:Array.isArray(o)?o.filter(t=>"string"==typeof t&&t.trim().length>0):void 0}

package/packages/core/dist/runtime/policy/tool-invocation.d.ts

@@ -17,6 +17,25 @@ export type ToolRepeatDecision = {
     eventOutput: string;
     modelOutput: string;
 };
+export type ToolGuardrailContext = {
+    agent: WorkspaceAgent;
+    args: unknown;
+    events: RuntimeEvent[];
+    repeatState?: ToolRepeatState;
+    toolId: string;
+};
+export type ToolGuardrailDecision = {
+    eventOutput: string;
+    modelOutput: string;
+    reason: string;
+    status: string;
+};
+export type ToolGuardrail = (context: ToolGuardrailContext) => ToolGuardrailDecision | undefined;
+export declare const requiredPlanToolGuardrail: ToolGuardrail;
+export declare const toolDependencyGuardrail: ToolGuardrail;
+export declare const repeatToolGuardrail: ToolGuardrail;
+export declare const defaultToolGuardrails: readonly ToolGuardrail[];
+export declare function evaluateToolGuardrails(context: ToolGuardrailContext, guardrails?: readonly ToolGuardrail[]): ToolGuardrailDecision | undefined;
 export declare function createToolRepeatState(config: unknown): ToolRepeatState | undefined;
 export declare function beforeToolInvoke(toolId: string, args: unknown, state: ToolRepeatState): ToolRepeatDecision | undefined;
 export declare function afterToolInvoke(input: {

package/packages/core/dist/runtime/policy/tool-invocation.js

@@ -1 +1 @@
-export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,repeatLimitedTools:new Set,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool),returnPreviousOutputOnRepeatLimit:!0===repeatGuardConfig(e).returnPreviousOutputOnRepeatLimit}}export function beforeToolInvoke(e,t,o){const n=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,n+1);const r=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==r&&n>=r){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const l=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==l&&(o.successfulToolCounts.get(e)??0)>=l){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const a=stableToolCallKey(e,t),s=o.successfulCalls.get(a);if(void 0!==s){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return o.repeatLimitedTools.add(e),{eventOutput:t,modelOutput:t}}const n=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,s);return{eventOutput:n,modelOutput:s}}}export function afterToolInvoke(e){return e.successful?(e.state.successfulCalls.set(stableToolCallKey(e.toolId,e.args),e.output),e.state.latestSuccessfulOutputByTool.set(e.toolId,e.output),e.state.successfulToolCounts.set(e.toolId,(e.state.successfulToolCounts.get(e.toolId)??0)+1),{}):{}}export function isToolRepeatLimitReached(e,t){if(!t)return!1;if(t.repeatLimitedTools.has(e))return!0;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const n=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==n&&(t.successfulToolCounts.get(e)??0)>=n}export function missingRequiredPlanContent(e){const t=readRecord(e.agent.config.executionContract);if(!0!==t.requiresPlan)return"";const o=readStringArray(t.planEvidenceTools);if(0===o.length||o.includes(e.toolId))return"";const n=new Set(e.events.flatMap(readEvidenceToolId));return o.some(e=>n.has(e))?"":["Status: plan_required",`Evidence tool: ${e.toolId}`,`Blocker: execution contract requires a planning checkpoint from one of: ${o.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}export function missingToolDependencyContent(e){const t=readRecord(e.agent.config.executionContract),o=readStringArray(readRecord(t.toolDependencies)[e.toolId]);if(0===o.length)return"";const n=new Set(e.events.flatMap(readEvidenceToolId)),r=o.filter(e=>!n.has(e));return 0===r.length?"":["Status: dependency_required",`Evidence tool: ${e.toolId}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${r.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}export function toolInvocationEvents(e){const t=e.getEvents?.();return t??[]}function repeatGuardConfig(e){const t=readRecord(e);return readRecord(t.repeatGuard)}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){const t=readRecord(e);return new Map(Object.entries(t).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1]))}function repeatLimitModelOutput(e,t,o){return o.returnPreviousOutputOnRepeatLimit&&void 0!==t&&0!==t.trim().length?t:e}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)&&function isToolResultEvent(e){return"deepagents.tool_execution.result"===e.eventType||"agent.tool.result"===e.phase}(e.event)&&"string"==typeof e.event.toolId&&function isSuccessfulEvidenceEvent(e){const t=function readString(e){return"string"==typeof e&&e.length>0?e:void 0}(e.controlStatus)??function readOutputStatus(e){if("string"!=typeof e)return;const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:e.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e.output);return!t||/^(?:completed|success|ok|recorded)$/iu.test(t)}(e.event)?[e.event.toolId]:[]}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}
+export const requiredPlanToolGuardrail=e=>{const t=missingRequiredPlanContent({agent:e.agent,events:e.events,toolId:e.toolId});return t?controlDecision("plan_required",t):void 0};export const toolDependencyGuardrail=e=>{const t=missingToolDependencyContent({agent:e.agent,events:e.events,toolId:e.toolId});return t?controlDecision("dependency_required",t):void 0};export const repeatToolGuardrail=e=>{const t=e.repeatState?beforeToolInvoke(e.toolId,e.args,e.repeatState):void 0;return t?{eventOutput:t.eventOutput,modelOutput:t.modelOutput,reason:t.eventOutput,status:readOutputStatus(t.eventOutput)??"repeated_tool_call_limit"}:void 0};export const defaultToolGuardrails=[requiredPlanToolGuardrail,toolDependencyGuardrail,repeatToolGuardrail];export function evaluateToolGuardrails(e,t=defaultToolGuardrails){for(const o of t){const t=o(e);if(t)return t}}export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,repeatLimitedTools:new Set,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool),returnPreviousOutputOnRepeatLimit:!0===repeatGuardConfig(e).returnPreviousOutputOnRepeatLimit}}export function beforeToolInvoke(e,t,o){const n=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,n+1);const r=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==r&&n>=r){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const l=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==l&&(o.successfulToolCounts.get(e)??0)>=l){const t=o.latestSuccessfulOutputByTool.get(e),n=repeatedToolCallLimitContent(e,t);return o.repeatLimitedTools.add(e),{eventOutput:n,modelOutput:repeatLimitModelOutput(n,t,o)}}const a=stableToolCallKey(e,t),u=o.successfulCalls.get(a);if(void 0!==u){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return o.repeatLimitedTools.add(e),{eventOutput:t,modelOutput:t}}const n=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,u);return{eventOutput:n,modelOutput:u}}}function controlDecision(e,t){return{eventOutput:t,modelOutput:t,reason:t,status:e}}export function afterToolInvoke(e){return e.successful?(e.state.successfulCalls.set(stableToolCallKey(e.toolId,e.args),e.output),e.state.latestSuccessfulOutputByTool.set(e.toolId,e.output),e.state.successfulToolCounts.set(e.toolId,(e.state.successfulToolCounts.get(e.toolId)??0)+1),{}):{}}export function isToolRepeatLimitReached(e,t){if(!t)return!1;if(t.repeatLimitedTools.has(e))return!0;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const n=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==n&&(t.successfulToolCounts.get(e)??0)>=n}export function missingRequiredPlanContent(e){const t=readRecord(e.agent.config.executionContract);if(!0!==t.requiresPlan)return"";const o=readStringArray(t.planEvidenceTools);if(0===o.length||o.includes(e.toolId))return"";const n=new Set(e.events.flatMap(readEvidenceToolId));return o.some(e=>n.has(e))?"":["Status: plan_required",`Evidence tool: ${e.toolId}`,`Blocker: execution contract requires a planning checkpoint from one of: ${o.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}export function missingToolDependencyContent(e){const t=readRecord(e.agent.config.executionContract),o=readStringArray(readRecord(t.toolDependencies)[e.toolId]);if(0===o.length)return"";const n=new Set(e.events.flatMap(readEvidenceToolId)),r=o.filter(e=>!n.has(e));return 0===r.length?"":["Status: dependency_required",`Evidence tool: ${e.toolId}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${r.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}export function toolInvocationEvents(e){const t=e.getEvents?.();return t??[]}function repeatGuardConfig(e){const t=readRecord(e);return readRecord(t.repeatGuard)}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){const t=readRecord(e);return new Map(Object.entries(t).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1]))}function repeatLimitModelOutput(e,t,o){return o.returnPreviousOutputOnRepeatLimit&&void 0!==t&&0!==t.trim().length?t:e}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)&&function isToolResultEvent(e){return"deepagents.tool_execution.result"===e.eventType||"agent.tool.result"===e.phase}(e.event)&&"string"==typeof e.event.toolId&&function isSuccessfulEvidenceEvent(e){const t=function readString(e){return"string"==typeof e&&e.length>0?e:void 0}(e.controlStatus)??readOutputStatus(e.output);return!t||/^(?:completed|success|ok|recorded)$/iu.test(t)}(e.event)?[e.event.toolId]:[]}function readOutputStatus(e){if("string"!=typeof e)return;const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:e.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/packages/core/dist/runtime.d.ts

@@ -1,9 +1,10 @@
 import type { ApprovalQueue } from "@stable-harness/governance";
 import type { MemoryProvider, RuntimeMemoryStore } from "@stable-harness/memory";
-import type { QualityReviewModel } from "./quality/index.js";
+import type { ExecutionEvaluatorRule, QualityReviewModel } from "./quality/index.js";
+import type { ToolGuardrail } from "./runtime/policy/tool-invocation.js";
 import { createLangSmithTracingCapability } from "./runtime/tracing/langsmith.js";
 import type { CompiledWorkspace, RuntimeCapabilityModule, RuntimeToolGateway, RuntimeAdapter, RuntimeArtifactStore, RuntimeSandboxPolicy, RuntimeStore, RuntimeProgressNarrationOptions, RuntimeWorkflowAdapter, StableHarnessRuntime } from "./types.js";
-type RuntimeFactoryInput = {
+export type RuntimeFactoryInput = {
     workspace: CompiledWorkspace;
     adapters: RuntimeAdapter[];
     workflowAdapters?: RuntimeWorkflowAdapter[];
@@ -16,8 +17,9 @@ type RuntimeFactoryInput = {
     artifacts?: RuntimeArtifactStore;
     progressNarration?: RuntimeProgressNarrationOptions | false;
     qualityReviewModel?: QualityReviewModel;
+    toolGuardrails?: readonly ToolGuardrail[];
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     capabilities?: RuntimeCapabilityModule[];
     langSmithTracing?: false | Parameters<typeof createLangSmithTracingCapability>[0]["options"];
 };
 export declare function createStableHarnessRuntime(input: RuntimeFactoryInput): StableHarnessRuntime;
-export {};

package/packages/core/dist/runtime.js

@@ -1 +1 @@
-import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{buildAdapterErrorRecoveryPrompt as r,buildExecutionContractRecoveryRequest as a,isRecoverableAdapterError as s}from"./recovery/tool-call.js";import{recoverQualityReview as n,resolveQualityPolicy as o}from"./quality/index.js";import{recoverAdapterResultOutput as i}from"./runtime/recovery/adapter-result.js";import{completeRun as u,failRun as c}from"./runtime/completion.js";import{createRuntimeAdministrationMethods as p}from"./runtime/admin/administration.js";import{runDirectToolCall as d}from"./runtime/direct-tool-call.js";import{createApprovalGatedToolGateway as m}from"./runtime/governance/approval-gate.js";import{createSandboxedToolGateway as l}from"./runtime/governance/sandbox.js";import{createRuntimeInspectionMethods as w}from"./runtime/inspection/methods.js";import{createRuntimeCapabilityRegistry as g,normalizeAdapterResult as y}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as f}from"./runtime/memory.js";import{resolveToolCallRecoveryPolicy as I}from"./runtime/recovery/tool-call-policy.js";import{createRuntimeMemoryAdministration as q}from"./runtime/admin/memory.js";import{createInMemoryRuntimeStore as R}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as k}from"./runtime/progress-narration.js";import{repairRuntimeSelection as v}from"./runtime/selection-repair.js";import{createLangSmithTracingCapability as b}from"./runtime/tracing/langsmith.js";import{createToolFailureTracker as A}from"./runtime/tool-failure.js";import{runWorkflowRequest as C}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const y=new Set,j=t.store??R(),h=g([f(t),k({options:t.progressNarration,policy:t.workspace.runtime}),b({policy:t.workspace.runtime,store:j,options:t.langSmithTracing}),...t.capabilities??[]]),emitBase=t=>{const r=function enrichRuntimeEvent(t){return{...t,eventId:t.eventId??e(),emittedAt:t.emittedAt??(new Date).toISOString()}}(t);j.appendEvent(r);for(const e of y)e(r)},emit=e=>{emitBase(e),h.emitSideEffects(e,emitBase)},x=l({gateway:m({gateway:t.toolGateway,approvals:t.approvals,workspace:t.workspace,emit:emit}),workspace:t.workspace,sandbox:t.sandbox,emit:emit}),S={...t,toolGateway:x},E=A(function readToolFailurePolicy(e){if("object"!=typeof e||null===e||Array.isArray(e))return;const t=e.failurePolicy;return"object"!=typeof t||null===t||Array.isArray(t)?void 0:t}(t.workspace.runtime.toolGateway));return{request:async t=>async function runRuntimeRequest(t){const p=t.request.requestId??e(),m=t.request.sessionId??e(),l=[],{agent:w,adapter:g}=await async function resolveExecution(e,t,r){const a=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const a=await v({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description})),trace:{...r,agentId:t,layer:"agent",owner:"stable_runtime_policy"}});return a.ok?a.id:t}(e.workspace,t.agentId,r):e.workspace.runtime.defaultAgentId,s=e.workspace.agents.get(a);if(!s)throw new Error(`Agent ${a} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const n=e.adapters.find(e=>e.canRun(s));if(!n)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:n}}(t.input,t.request,{requestId:p,sessionId:m,emit:e=>l.push(e)});t.store.createRun(function createRunRecord(e,t,r,a){return{requestId:t,sessionId:r,agentId:a.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,p,m,w)),l.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:p,sessionId:m,agentId:w.id,input:t.request.input});try{if(t.request.workflow){const e=await C({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:p,sessionId:m,agentId:w.id,emit:t.emit});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}if(t.request.toolCall){const e=await d({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:p,sessionId:m,agent:w,toolFailureTracker:t.toolFailureTracker});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,c=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),p=c.memory,d=c.pluginMemories??[],m=I({workspace:e.input.workspace,agent:e.agent}),l=o(e.input.workspace.runtime,e.agent),w=new Map;let g;try{g=await runAdapterOnce(e,t,e.request,p,d,w,m)}catch(a){if(!s(a,m))throw a;e.emit(repairStarted(e,"adapter_error",1,errorMessage(a))),g=await runAdapterOnce(e,t,r(e.request,a,m),p,d,w,m),e.emit(repairCompleted(e,"adapter_error","retried",1,errorMessage(a)))}g=await i({...e,request:e.request,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),e.request,g,l),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:g});try{assertRequestExecutionContract(e)}catch(r){const s=a({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:m});if(!s)throw r;e.emit(repairStarted(e,"execution_contract",1,errorMessage(r))),g=await runAdapterOnce(e,t,s,p,d,w,m),g=await i({...e,request:s,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),s,g,l),assertRequestExecutionContract(e),e.emit(repairCompleted(e,"execution_contract","retried",1,errorMessage(r)))}const y=u({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:g,artifacts:e.input.artifacts});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:g,response:y}),y}({...t,adapter:g,requestId:p,sessionId:m,agent:w})}catch(e){return c({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,error:e})}}({input:S,capabilities:h,store:j,emit:emit,request:t,toolFailureTracker:E}),subscribe:e=>(y.add(e),()=>y.delete(e)),...w({workspace:t.workspace,store:j,artifacts:t.artifacts,approvals:t.approvals,emit:emit}),...p({store:j,emit:emit}),...q({memory:t.memory}),cancel(e,t){const r=j.getRun(e);r&&"running"===r.state&&(j.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await h.stop(),y.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,a,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:n=>runAdapterOnce(e,e.adapter,n,t,r,a,s),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,a,s,n,o){return y(await t.run({workspace:{...e.input.workspace,runtime:o},agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:a,pluginMemories:s,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,requestState:n,getEvents:()=>e.store.getRun(e.requestId)?.events??[],emit:e.emit}))}function repairStarted(e,t,r,a){return{type:"runtime.repair.started",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,attempt:r,reason:a}}function repairCompleted(e,t,r,a,s){return{type:"runtime.repair.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,outcome:r,attempt:a,reason:s}}function errorMessage(e){return e instanceof Error?e.message:String(e)}
+import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{buildAdapterErrorRecoveryPrompt as r,buildExecutionContractRecoveryRequest as a,isRecoverableAdapterError as s}from"./recovery/tool-call.js";import{recoverQualityReview as n,resolveQualityPolicy as o}from"./quality/index.js";import{recoverAdapterResultOutput as i}from"./runtime/recovery/adapter-result.js";import{completeRun as u,failRun as c}from"./runtime/completion.js";import{createRuntimeAdministrationMethods as p}from"./runtime/admin/administration.js";import{runDirectToolCall as d}from"./runtime/direct-tool-call.js";import{createApprovalGatedToolGateway as m}from"./runtime/governance/approval-gate.js";import{createSandboxedToolGateway as l}from"./runtime/governance/sandbox.js";import{createRuntimeInspectionMethods as w}from"./runtime/inspection/methods.js";import{createRuntimeCapabilityRegistry as g,normalizeAdapterResult as y}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as f}from"./runtime/memory.js";import{resolveToolCallRecoveryPolicy as I}from"./runtime/recovery/tool-call-policy.js";import{createRuntimeMemoryAdministration as q}from"./runtime/admin/memory.js";import{createInMemoryRuntimeStore as R}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as v}from"./runtime/progress-narration.js";import{repairRuntimeSelection as k}from"./runtime/selection-repair.js";import{createLangSmithTracingCapability as b}from"./runtime/tracing/langsmith.js";import{createToolFailureTracker as A}from"./runtime/tool-failure.js";import{runWorkflowRequest as C}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const y=new Set,x=t.store??R(),j=g([f(t),v({options:t.progressNarration,policy:t.workspace.runtime}),b({policy:t.workspace.runtime,store:x,options:t.langSmithTracing}),...t.capabilities??[]]),emitBase=t=>{const r=function enrichRuntimeEvent(t){return{...t,eventId:t.eventId??e(),emittedAt:t.emittedAt??(new Date).toISOString()}}(t);x.appendEvent(r);for(const e of y)e(r)},emit=e=>{emitBase(e),j.emitSideEffects(e,emitBase)},E=l({gateway:m({gateway:t.toolGateway,approvals:t.approvals,workspace:t.workspace,emit:emit}),workspace:t.workspace,sandbox:t.sandbox,emit:emit}),h={...t,toolGateway:E},S=A(function readToolFailurePolicy(e){if("object"!=typeof e||null===e||Array.isArray(e))return;const t=e.failurePolicy;return"object"!=typeof t||null===t||Array.isArray(t)?void 0:t}(t.workspace.runtime.toolGateway));return{request:async t=>async function runRuntimeRequest(t){const p=t.request.requestId??e(),m=t.request.sessionId??e(),l=[],{agent:w,adapter:g}=await async function resolveExecution(e,t,r){const a=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const a=await k({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description})),trace:{...r,agentId:t,layer:"agent",owner:"stable_runtime_policy"}});return a.ok?a.id:t}(e.workspace,t.agentId,r):e.workspace.runtime.defaultAgentId,s=e.workspace.agents.get(a);if(!s)throw new Error(`Agent ${a} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const n=e.adapters.find(e=>e.canRun(s));if(!n)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:n}}(t.input,t.request,{requestId:p,sessionId:m,emit:e=>l.push(e)});t.store.createRun(function createRunRecord(e,t,r,a){return{requestId:t,sessionId:r,agentId:a.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,p,m,w)),l.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:p,sessionId:m,agentId:w.id,input:t.request.input});try{if(t.request.workflow){const e=await C({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:p,sessionId:m,agentId:w.id,emit:t.emit});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}if(t.request.toolCall){const e=await d({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:p,sessionId:m,agent:w,toolFailureTracker:t.toolFailureTracker,toolGuardrails:t.input.toolGuardrails,events:t.store.getRun(p)?.events??[]});return u({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,result:e,artifacts:t.input.artifacts})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,c=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),p=c.memory,d=c.pluginMemories??[],m=I({workspace:e.input.workspace,agent:e.agent}),l=o(e.input.workspace.runtime,e.agent),w=new Map;let g;try{g=await runAdapterOnce(e,t,e.request,p,d,w,m)}catch(a){if(!s(a,m))throw a;e.emit(repairStarted(e,"adapter_error",1,errorMessage(a))),g=await runAdapterOnce(e,t,r(e.request,a,m),p,d,w,m),e.emit(repairCompleted(e,"adapter_error","retried",1,errorMessage(a)))}g=await i({...e,request:e.request,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),e.request,g,l),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:g});try{assertRequestExecutionContract(e)}catch(r){const s=a({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:m});if(!s)throw r;e.emit(repairStarted(e,"execution_contract",1,errorMessage(r))),g=await runAdapterOnce(e,t,s,p,d,w,m),g=await i({...e,request:s,result:g,recoveryPolicy:m,runAdapter:r=>runAdapterOnce(e,t,r,p,d,w,m)}),g=await n(createQualityRuntimeInput(e,p,d,w,m),s,g,l),assertRequestExecutionContract(e),e.emit(repairCompleted(e,"execution_contract","retried",1,errorMessage(r)))}const y=u({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:g,artifacts:e.input.artifacts});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:g,response:y}),y}({...t,adapter:g,requestId:p,sessionId:m,agent:w})}catch(e){return c({store:t.store,emit:t.emit,requestId:p,sessionId:m,agent:w,error:e})}}({input:h,capabilities:j,store:x,emit:emit,request:t,toolFailureTracker:S}),subscribe:e=>(y.add(e),()=>y.delete(e)),...w({workspace:t.workspace,store:x,artifacts:t.artifacts,approvals:t.approvals,emit:emit}),...p({store:x,emit:emit}),...q({memory:t.memory}),cancel(e,t){const r=x.getRun(e);r&&"running"===r.state&&(x.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await j.stop(),y.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,a,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:n=>runAdapterOnce(e,e.adapter,n,t,r,a,s),reviewModel:e.input.qualityReviewModel,executionEvaluatorRules:e.input.executionEvaluatorRules,memory:t,pluginMemories:r}}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,a,s,n,o){return y(await t.run({workspace:{...e.input.workspace,runtime:o},agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:a,pluginMemories:s,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,toolGuardrails:e.input.toolGuardrails,executionEvaluatorRules:e.input.executionEvaluatorRules,requestState:n,getEvents:()=>e.store.getRun(e.requestId)?.events??[],emit:e.emit}))}function repairStarted(e,t,r,a){return{type:"runtime.repair.started",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,attempt:r,reason:a}}function repairCompleted(e,t,r,a,s){return{type:"runtime.repair.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,layer:t,outcome:r,attempt:a,reason:s}}function errorMessage(e){return e instanceof Error?e.message:String(e)}

package/packages/core/dist/types.d.ts

@@ -4,6 +4,8 @@ import type { RuntimeWorkflowAdapter, RuntimeWorkflowRequest, WorkspaceWorkflow
 import type { SpecDrivenWorkflowState } from "./spec-driven/index.js";
 import type { RuntimeEvent, RuntimeEventListener, RuntimeEmit } from "./runtime/events.js";
 import type { RuntimeToolFailureTracker } from "./runtime/tool-failure.js";
+import type { ToolGuardrail } from "./runtime/policy/tool-invocation.js";
+import type { ExecutionEvaluatorRule } from "./quality/execution-review.js";
 import type { RuntimeArtifact, RuntimeArtifactFilter, RuntimeArtifactRecord, RuntimeOutput, RuntimeRecordState, RuntimeRequest, RuntimeResponse, RuntimeReplayBundle, RuntimeDeletionResult, RuntimeRunFilter, RuntimeRunRecord } from "./runtime/types.js";
 import type { RuntimeToolGateway } from "./runtime/tool-gateway.js";
 import type { CompiledWorkspace, WorkspaceAgent, WorkspaceRuntimePolicy } from "./workspace/types.js";
@@ -28,6 +30,8 @@ export type RuntimeAdapterContext = {
     pluginMemories?: RuntimeMemoryContext[];
     toolGateway?: RuntimeToolGateway;
     toolFailureTracker?: RuntimeToolFailureTracker;
+    toolGuardrails?: readonly ToolGuardrail[];
+    executionEvaluatorRules?: readonly ExecutionEvaluatorRule[];
     requestState?: Map<string, unknown>;
     getEvents?: () => RuntimeEvent[];
     emit: RuntimeEmit;

package/packages/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/core",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -11,7 +11,7 @@
     ".": "./dist/index.js"
   },
   "peerDependencies": {
-    "@stable-harness/governance": "0.0.80",
-    "@stable-harness/memory": "0.0.80"
+    "@stable-harness/governance": "0.0.81",
+    "@stable-harness/memory": "0.0.81"
   }
 }

package/packages/evaluation/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/evaluation",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -10,6 +10,6 @@
   "main": "dist/src/index.js",
   "types": "dist/src/index.d.ts",
   "peerDependencies": {
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }

package/packages/governance/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/governance",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",

package/packages/memory/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/memory",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",

package/packages/protocols/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/protocols",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -10,6 +10,6 @@
   "main": "dist/src/index.js",
   "types": "dist/src/index.d.ts",
   "peerDependencies": {
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }

package/packages/tool-gateway/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/tool-gateway",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",

package/packages/workspace-yaml/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stable-harness/workspace-yaml",
-  "version": "0.0.80",
+  "version": "0.0.81",
   "type": "module",
   "files": [
     "dist/**/*.js",
@@ -11,6 +11,6 @@
     ".": "./dist/index.js"
   },
   "peerDependencies": {
-    "@stable-harness/core": "0.0.80"
+    "@stable-harness/core": "0.0.81"
   }
 }