stable-harness 0.0.36 → 0.0.38

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

@@ -0,0 +1,38 @@
+import type { WorkspaceBoundaryDiagnostic, WorkspaceBoundaryDiagnosticCode, WorkspaceBoundaryScanPolicy } from "./workspace/types.js";
+export type BoundaryScanLayer = "tool" | "skill" | "subagent";
+export type BoundaryScanResource = {
+    id: string;
+    description?: string;
+    text?: string;
+    tools?: string[];
+    skills?: string[];
+    subagents?: string[];
+    locations?: string[];
+    metadata?: Record<string, unknown>;
+};
+export type BoundaryScanFinding = {
+    code: WorkspaceBoundaryDiagnosticCode;
+    layer: BoundaryScanLayer;
+    resources: string[];
+    message: string;
+    locations: string[];
+    details: Record<string, unknown>;
+};
+export type BoundaryScanLayerResult = {
+    resources: BoundaryScanResource[];
+    findings: BoundaryScanFinding[];
+};
+export type BoundaryScanResult = {
+    tools: BoundaryScanLayerResult;
+    skills: BoundaryScanLayerResult;
+    subagents: BoundaryScanLayerResult;
+    findings: BoundaryScanFinding[];
+};
+export declare function scanAgentBoundary(input: {
+    tools?: BoundaryScanResource[];
+    skills?: BoundaryScanResource[];
+    subagents?: BoundaryScanResource[];
+}): BoundaryScanResult;
+export declare function checkBoundaryScanPolicy(result: BoundaryScanResult, policy: WorkspaceBoundaryScanPolicy | undefined): WorkspaceBoundaryDiagnostic[];
+export declare function assertBoundaryPolicyDiagnostics(diagnostics: WorkspaceBoundaryDiagnostic[]): void;
+export declare function formatBoundaryDiagnostic(diagnostic: WorkspaceBoundaryDiagnostic): string;

package/node_modules/@stable-harness/core/dist/boundary-scan.js

@@ -0,0 +1 @@
+export function scanAgentBoundary(e){const n=scanLayer("tool",e.tools??[]),i=scanLayer("skill",e.skills??[]),o=scanLayer("subagent",e.subagents??[]);return{tools:n,skills:i,subagents:o,findings:[...n.findings,...i.findings,...o.findings]}}export function checkBoundaryScanPolicy(e,n){return e.findings.map(e=>({code:e.code,layer:e.layer,resources:e.resources,severity:!0===n?.failOn?.[e.code]?"error":"warning",message:e.message,locations:e.locations,details:e.details}))}export function assertBoundaryPolicyDiagnostics(e){if(0!==e.filter(e=>"error"===e.severity).length)throw new Error(`Workspace boundary scan failed:\n${e.map(formatBoundaryDiagnostic).join("\n")}`)}export function formatBoundaryDiagnostic(e){const n=e.locations.length>0?` (${e.locations.join(", ")})`:"";return`- ${e.severity} ${e.code} [${e.layer}] ${e.resources.join(", ")}: ${e.message}${n}`}function scanLayer(e,n){const i=n.flatMap(i=>[...scanSingleResource(e,i),...scanHardcodedSiblingRouting(e,i,n)]);for(let o=0;o<n.length;o+=1)for(let t=o+1;t<n.length;t+=1)i.push(...scanResourcePair(e,n[o],n[t]));return{resources:n,findings:i}}function scanSingleResource(e,n){const i=[],o=resourceText(n);return(0===(n.description??"").trim().length||meaningfulTokens(o).length<3)&&i.push(createFinding("weak-positive-boundary",e,[n],`${e} ${n.id} does not declare a clear positive responsibility boundary.`)),function hasBroadFallbackLanguage(e){return positiveSentences(e).some(e=>/\b(any|anything|all|everything|general|generic|fallback|catch[-\s]?all)\b/iu.test(e))}(o)&&!function hasLocalScopeLanguage(e){return/\b(only|specific|scope|focused|owned|responsible|specialized|local|bounded)\b/iu.test(e)}(o)&&i.push(createFinding("broad-fallback-scope",e,[n],`${e} ${n.id} uses broad fallback language without declaring local scope.`)),function hasOnlyExclusionLanguage(e){return/\b(do not|don't|except|exclude|avoid|not for|never)\b/iu.test(e)&&!/\b(handle|provide|manage|analyze|create|run|execute|inspect|summarize|search|own|owns)\b/iu.test(e)}(o)&&i.push(createFinding("weak-positive-boundary",e,[n],`${e} ${n.id} relies on exclusions without a positive capability boundary.`)),i}function scanHardcodedSiblingRouting(e,n,i){return i.filter(e=>e.id!==n.id).filter(e=>function mentionsSiblingRouting(e,n){const i=n.replace(/[.*+?^${}()|[\]\\]/gu,"\\$&"),o=new RegExp(`\\b(route|delegate|send|handoff|assign|use)\\b[^.!?\\n]{0,80}(?:^|[^\\p{L}\\p{N}_-])${i}(?:$|[^\\p{L}\\p{N}_-])`,"iu");return positiveSentences(e).some(e=>o.test(e))}(resourceText(n),e.id)).map(i=>createFinding("hardcoded-sibling-routing",e,[n,i],`${e} ${n.id} hardcodes routing to sibling ${i.id}; use loaded metadata or explicit runtime policy instead.`,{sibling:i.id}))}function scanResourcePair(e,n,i){const o=normalizeScope(n.description??""),t=normalizeScope(i.description??"");if(o&&o===t)return[createFinding("duplicate-scope",e,[n,i],`${e} resources ${n.id} and ${i.id} declare the same responsibility scope.`)];const s=function lexicalSimilarity(e,n){const i=new Set(meaningfulTokens(e)),o=new Set(meaningfulTokens(n));return 0===i.size||0===o.size?0:[...i].filter(e=>o.has(e)).length/new Set([...i,...o]).size}(resourceText(n),resourceText(i));return s>=.72?[createFinding("possible-sibling-overlap",e,[n,i],`${e} resources ${n.id} and ${i.id} have likely overlapping model-visible boundaries.`,{similarity:s})]:[]}function createFinding(e,n,i,o,t={}){return{code:e,layer:n,resources:i.map(e=>e.id),message:o,locations:[...new Set(i.flatMap(e=>e.locations??[]))],details:{...t,resourceCount:i.length}}}function resourceText(e){return[e.description,e.text,e.tools?.join(" "),e.skills?.join(" "),e.subagents?.join(" ")].filter(Boolean).join("\n")}function normalizeScope(e){return meaningfulTokens(e).join(" ")}function meaningfulTokens(n){return[...n.toLowerCase().matchAll(/[a-z0-9][a-z0-9_-]{2,}/gu)].map(e=>e[0]).filter(n=>!e.has(n))}function positiveSentences(e){return e.split(/[.!?\n]+/u).map(e=>e.trim()).filter(e=>e.length>0).filter(e=>!/\b(do not|does not|don't|not for|never|exclude|avoid)\b/iu.test(e))}const e=new Set(["and","are","for","from","into","that","the","this","with","when","workspace","agent","subagent","skill","tool"]);

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

@@ -0,0 +1,18 @@
+export type WorkspaceEvaluation = {
+    id: string;
+    description?: string;
+    sourcePath?: string;
+    suite?: string;
+    cases: WorkspaceEvaluationCase[];
+    config?: Record<string, unknown>;
+};
+export type WorkspaceEvaluationCase = {
+    id: string;
+    description?: string;
+    agentId?: string;
+    workflowId?: string;
+    input?: unknown;
+    tools?: string[];
+    assertions?: Record<string, unknown>;
+    metadata?: Record<string, unknown>;
+};

package/node_modules/@stable-harness/core/dist/evaluations/index.js

@@ -0,0 +1 @@
+export{};

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

@@ -0,0 +1,9 @@
+import type { RuntimeEmit, RuntimeStore, WorkspaceAgent } from "./types.js";
+export declare function assertExecutionContract(input: {
+    store: RuntimeStore;
+    emit: RuntimeEmit;
+    requestId: string;
+    sessionId: string;
+    agent: WorkspaceAgent;
+    metadata?: Record<string, unknown>;
+}): void;

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

@@ -0,0 +1 @@
+export function assertExecutionContract(e){(function contractDisabled(e){const t=isRecord(e?.executionContract)?e.executionContract:void 0;return!1===t?.enabled})(e.metadata)||(function assertRequiredPlan(e){const t=readExecutionContract(e.agent);if(!t.requiresPlan)return;const n=readStringArray(t.planEvidenceTools);if(0===n.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"invalid_plan_contract"}),new Error("Execution contract requires plan evidence tools when requiresPlan is enabled");const r=e.store.getRun(e.requestId),o=new Set((r?.events??[]).flatMap(readEvidenceToolId));if(!n.some(e=>o.has(e)))throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_plan",missingEvidenceTools:n}),new Error(`Execution contract requires a planning checkpoint from one of: ${n.join(", ")}`)}(e),function assertRequiredEvidenceTools(e){const t=function readRequiredEvidenceTools(e){return readStringArray(readExecutionContract(e).requiredEvidenceTools)}(e.agent);if(0===t.length)return;const n=e.store.getRun(e.requestId),r=new Set((n?.events??[]).flatMap(readEvidenceToolId)),o=t.filter(e=>!r.has(e));if(0!==o.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_required_evidence",missingEvidenceTools:o}),new Error(`Execution contract missing required evidence tools: ${o.join(", ")}`)}(e),function assertToolDependencies(e){const t=readExecutionContract(e.agent),n=isRecord(t.toolDependencies)?t.toolDependencies:{},r=Object.entries(n).map(([e,t])=>[e,readStringArray(t)]).filter(e=>e[1].length>0);if(0===r.length)return;const o=e.store.getRun(e.requestId),i=new Set((o?.events??[]).flatMap(readEvidenceToolId)),s=r.filter(([e])=>i.has(e)).flatMap(([e,t])=>t.filter(e=>!i.has(e)).map(t=>`${e} requires ${t}`));if(0!==s.length)throw e.emit({type:"runtime.execution.contract.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,reason:"missing_tool_dependency",missingEvidenceTools:s}),new Error(`Execution contract missing required tool dependencies: ${s.join(", ")}`)}(e))}function readExecutionContract(e){return isRecord(e.config.executionContract)?e.config.executionContract:{}}function readEvidenceToolId(e){return"runtime.tool.direct.completed"===e.type?[e.toolId]:"runtime.adapter.event"===e.type&&isRecord(e.event)?"agent.tool.result"!==e.event.phase||"string"!=typeof e.event.toolId?[]:function isSuccessfulEvidenceEvent(e){const t=function readString(e){return"string"==typeof e&&e.length>0?e:void 0}(e.controlStatus)??function readOutputStatus(e){if("string"!=typeof e)return;const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:e.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e.output);return!t||/^(?:completed|success|ok|recorded)$/iu.test(t)}(e.event)?[e.event.toolId]:[]:[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}

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

@@ -0,0 +1,20 @@
+export * from "./runtime/persistence/artifacts.js";
+export type { ApprovalRequest, ApprovalRequestStatus } from "@stable-harness/governance";
+export * from "./boundary-scan.js";
+export * from "./execution-contract.js";
+export * from "./recovery/tool-call.js";
+export * from "./runtime/persistence/inspection.js";
+export { createWorkspaceSandboxPolicy } from "./runtime/governance/sandbox.js";
+export * from "./memory-plugins.js";
+export * from "./runtime/persistence/queue.js";
+export * from "./runtime.js";
+export * from "./runtime/selection-repair.js";
+export * from "./runtime/tool-failure.js";
+export * from "./runtime/persistence/stores.js";
+export * from "./trace.js";
+export * from "./types.js";
+export * from "./workspace/tool-quality.js";
+export * from "./evaluations/index.js";
+export * from "./quality/index.js";
+export * from "./spec-driven/index.js";
+export * from "./workflows/index.js";

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

@@ -0,0 +1 @@
+export*from"./runtime/persistence/artifacts.js";export*from"./boundary-scan.js";export*from"./execution-contract.js";export*from"./recovery/tool-call.js";export*from"./runtime/persistence/inspection.js";export{createWorkspaceSandboxPolicy}from"./runtime/governance/sandbox.js";export*from"./memory-plugins.js";export*from"./runtime/persistence/queue.js";export*from"./runtime.js";export*from"./runtime/selection-repair.js";export*from"./runtime/tool-failure.js";export*from"./runtime/persistence/stores.js";export*from"./trace.js";export*from"./types.js";export*from"./workspace/tool-quality.js";export*from"./evaluations/index.js";export*from"./quality/index.js";export*from"./spec-driven/index.js";export*from"./workflows/index.js";

package/node_modules/@stable-harness/core/dist/memory-plugins/maintenance.d.ts

@@ -0,0 +1,42 @@
+import type { MemoryMaintenanceOperation, MemoryProvider } from "@stable-harness/memory";
+import type { ApprovalQueue, SkillCandidateStore } from "@stable-harness/governance";
+import type { CompiledWorkspace, RuntimeEmit, WorkspaceAgent } from "../types.js";
+export type MemoryMaintenanceTarget = {
+    name: string;
+    run(input: MemoryMaintenanceTargetInput): Promise<MemoryMaintenanceResult>;
+};
+export type MemoryMaintenanceTargetInput = {
+    emit: RuntimeEmit;
+    requestId: string;
+    sessionId: string;
+    agent: WorkspaceAgent;
+    workspace: CompiledWorkspace;
+    prompt: string;
+};
+export type MemoryMaintenanceResult = {
+    operations?: MemoryMaintenanceOperation[];
+    metadata?: Record<string, unknown>;
+};
+export type MemoryMaintenanceDaemon = {
+    runOnce(input?: Partial<Pick<MemoryMaintenanceTargetInput, "requestId" | "sessionId" | "agent">>): Promise<MemoryMaintenanceResult[]>;
+    start(): void;
+    stop(): void;
+};
+export declare function createMemoryMaintenanceDaemon(input: {
+    workspace: CompiledWorkspace;
+    emit?: RuntimeEmit;
+    targets: MemoryMaintenanceTarget[];
+    intervalMs?: number;
+    prompt?: string;
+    defaultAgentId?: string;
+}): MemoryMaintenanceDaemon;
+export declare function createLangMemMaintenanceTarget(input: {
+    providers: MemoryProvider[] | undefined;
+    limit?: number;
+}): MemoryMaintenanceTarget;
+export declare function createSkillCandidateMinerTarget(input: {
+    providers: MemoryProvider[] | undefined;
+    store: SkillCandidateStore;
+    approvals?: ApprovalQueue;
+    minEvidenceCount?: number;
+}): MemoryMaintenanceTarget;

package/node_modules/@stable-harness/core/dist/memory-plugins/maintenance.js

@@ -0,0 +1 @@
+import{formatError as e,readRecord as t,readWorkspaceId as n,resolveEnabledMemories as a,resolveMemoryProvider as r,resolvePluginNamespace as s}from"./shared.js";export function createMemoryMaintenanceDaemon(e){let t;const n=e.emit??(()=>{});return{async runOnce(t={}){const a=t.agent??function resolveMaintenanceAgent(e,t){const n=e.agents.get(t??e.runtime.defaultAgentId);if(!n)throw new Error("Memory maintenance requires a configured default agent");return n}(e.workspace,e.defaultAgentId),r={emit:n,requestId:t.requestId??`memory-maintenance-${Date.now()}`,sessionId:t.sessionId??"memory-maintenance",agent:a,workspace:e.workspace,prompt:e.prompt??defaultMaintenancePrompt(e.workspace)},s=[];for(const t of e.targets)s.push(await runMaintenanceTarget(r,t));return s},start(){e.intervalMs&&!t&&(t=setInterval(()=>{this.runOnce()},e.intervalMs))},stop(){t&&(clearInterval(t),t=void 0)}}}export function createLangMemMaintenanceTarget(e){return{name:"LangMem",async run(t){const n=[];for(const o of a(t.workspace,"write")){const a=r(e.providers??[],o.provider);if(!a)continue;const i=s(t.workspace,t.agent,{input:"",sessionId:t.sessionId},o),c=mergeMemoryRecords(await a.search({namespace:i,query:t.prompt,limit:e.limit??100}),await a.search({namespace:i,query:"",limit:e.limit??100}));n.push(...await a.consolidate({records:c}))}return{operations:n}}}}export function createSkillCandidateMinerTarget(e){return{name:"skillCandidateMiner",async run(t){const n=await async function collectMaintenanceRecords(e,t){const n=[];for(const o of a(t.workspace,"all")){const a=r(e??[],o.provider);if(!a)continue;const i=s(t.workspace,t.agent,{input:"",sessionId:t.sessionId},o);n.push(...mergeMemoryRecords(await a.search({namespace:i,query:t.prompt,limit:100}),await a.search({namespace:i,query:"",limit:100})))}return mergeMemoryRecords(n,[])}(e.providers,t),o=function groupSkillCandidateRecords(e){const t=new Map;for(const n of e){const e=readSkillCandidateName(n);e&&t.set(e,[...t.get(e)??[],n])}return t}(n),i=[];for(const[n,a]of o){if(a.length<(e.minEvidenceCount??3))continue;const r=await e.store.upsert(createSkillCandidateInput(t.workspace,n,a));i.push(r),t.emit({type:"runtime.skill.candidate.created",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,candidateId:r.id,name:r.name,confidence:r.confidence,evidenceCount:r.evidenceCount,status:r.status,proposedPath:r.proposedPath}),await(e.approvals?.create({kind:"skill_candidate",reason:`Review mined skill candidate ${r.name}`,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,subject:{candidate:r}}))}return{operations:[],metadata:{candidateCount:i.length}}}}}async function runMaintenanceTarget(t,n){t.emit({type:"runtime.memory.maintenance.started",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,target:n.name});try{const e=await n.run(t);return t.emit({type:"runtime.memory.maintenance.completed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,target:n.name,operationCount:e.operations?.length??0}),e}catch(a){return t.emit({type:"runtime.memory.maintenance.failed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,target:n.name,error:e(a)}),{operations:[],metadata:{error:e(a)}}}}function mergeMemoryRecords(e,t){const n=new Map;for(const a of[...e,...t])n.set(a.id,a);return[...n.values()]}function readSkillCandidateName(e){const n=t(e.metadata?.skillCandidate),a=t(t(e.metadata?.requestMetadata)?.skillCandidate),r=n?.name??a?.name;return"string"==typeof r&&r.trim()?function slugify(e){return e.toLowerCase().replace(/[^a-z0-9]+/gu,"-").replace(/^-+|-+$/gu,"")}(r):void 0}function createSkillCandidateInput(e,t,a){const r=function titleFromName(e){return e.split("-").filter(Boolean).map(e=>`${e[0]?.toUpperCase()??""}${e.slice(1)}`).join(" ")}(t);return{workspaceId:n(e),name:t,status:"review_required",confidence:Math.min(.95,.55+.1*a.length),proposedPath:`resources/skills/${t}/SKILL.md`,title:r,summary:`Repeated stable workflow mined from ${a.length} memory records.`,draftMarkdown:renderSkillDraft(r,a),evidence:a.map(e=>({evidenceType:"memory_record",evidenceRef:e.id,summary:e.summary??e.content.slice(0,160),weight:e.confidence}))}}function renderSkillDraft(e,t){return[`# ${e}`,"","## Purpose","","Apply this stable workflow when the task matches the evidence.","","## Evidence","",t.map(e=>`- ${e.summary??e.content}`).join("\n")].join("\n")}function defaultMaintenancePrompt(e){const n=t(e.runtime.memory?.maintenance);return"string"==typeof n?.prompt?n.prompt:"durable preferences workspace facts reusable procedures prior corrections stale duplicated contradicted memories"}

package/node_modules/@stable-harness/core/dist/memory-plugins/shared.d.ts

@@ -0,0 +1,8 @@
+import type { MemoryProvider } from "@stable-harness/memory";
+import type { CompiledWorkspace, RuntimeRequest, WorkspaceAgent, WorkspaceMemory } from "../types.js";
+export declare function resolveEnabledMemories(workspace: CompiledWorkspace, phase?: "all" | "recall" | "write"): WorkspaceMemory[];
+export declare function resolvePluginNamespace(workspace: CompiledWorkspace, agent: WorkspaceAgent, request: RuntimeRequest, memory: WorkspaceMemory): string;
+export declare function resolveMemoryProvider(providers: MemoryProvider[], configuredProvider: string): MemoryProvider | undefined;
+export declare function readRecord(value: unknown): Record<string, unknown> | undefined;
+export declare function readWorkspaceId(workspace: CompiledWorkspace): string;
+export declare function formatError(error: unknown): string;

package/node_modules/@stable-harness/core/dist/memory-plugins/shared.js

@@ -0,0 +1 @@
+export function resolveEnabledMemories(e,r="all"){const o=e.runtime.memory??{};if(!function isMemoryPhaseEnabled(e,r){return!1!==e.enabled&&("recall"!==r||!1!==e.recall&&!1!==e.read&&!1!==readMemoryFlag(e,"LangMem","read"))&&("write"!==r||!1!==e.write&&!1!==readMemoryFlag(e,"LangMem","write"))}(o,r))return[];const n=Array.isArray(o.refs)?o.refs.filter(e=>"string"==typeof e):[];return(n.length>0?n:[...e.memories.keys()]).map(r=>e.memories.get(r)).filter(e=>Boolean(e?.enabled))}export function resolvePluginNamespace(e,r,o,n){const t=readWorkspaceId(e),a=n.profile??n.id;return"agent"===a?`${t}:agent:${r.id}:${n.id}`:"user"===a?`${t}:user:${function readUserId(e){return e.metadata?.userId??"local"}(o)}:${n.id}`:"session"===a?`${t}:session:${o.sessionId??"default"}:${n.id}`:`${t}:${a}:${n.id}`}export function resolveMemoryProvider(e,r){return e.find(e=>e.name===r||e.name===`${r}-service`)}export function readRecord(e){return"object"!=typeof e||null===e||Array.isArray(e)?void 0:e}export function readWorkspaceId(e){return e.runtime.workspaceId??e.runtime.profile??e.root}export function formatError(e){return e instanceof Error?e.message:String(e)}function readMemoryFlag(e,r,o){const n=readRecord(e[r])??readRecord(e[r.toLowerCase()]);return"boolean"==typeof n?.[o]?n[o]:void 0}

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

@@ -0,0 +1,25 @@
+import type { MemoryProvider, MemoryRecord } from "@stable-harness/memory";
+import type { CompiledWorkspace, RuntimeEmit, RuntimeOutput, RuntimeRequest, WorkspaceAgent } from "./types.js";
+export { resolveEnabledMemories, resolvePluginNamespace, } from "./memory-plugins/shared.js";
+export type { MemoryMaintenanceDaemon, MemoryMaintenanceResult, MemoryMaintenanceTarget, MemoryMaintenanceTargetInput, } from "./memory-plugins/maintenance.js";
+export { createLangMemMaintenanceTarget, createMemoryMaintenanceDaemon, createSkillCandidateMinerTarget, } from "./memory-plugins/maintenance.js";
+export declare function runMemoryPlugins(input: {
+    providers: MemoryProvider[] | undefined;
+    emit: RuntimeEmit;
+    request: RuntimeRequest;
+    requestId: string;
+    sessionId: string;
+    agent: WorkspaceAgent;
+    workspace: CompiledWorkspace;
+    result: RuntimeOutput;
+}): Promise<void>;
+export declare function recallMemoryPlugins(input: {
+    providers: MemoryProvider[] | undefined;
+    request: RuntimeRequest;
+    agent: WorkspaceAgent;
+    workspace: CompiledWorkspace;
+}): Promise<Array<{
+    namespace: string;
+    records: MemoryRecord[];
+    context: string;
+}>>;

package/node_modules/@stable-harness/core/dist/memory-plugins.js

@@ -0,0 +1 @@
+import{formatError as e,readRecord as r,readWorkspaceId as t,resolveEnabledMemories as n,resolveMemoryProvider as o,resolvePluginNamespace as a}from"./memory-plugins/shared.js";export{resolveEnabledMemories,resolvePluginNamespace}from"./memory-plugins/shared.js";export{createLangMemMaintenanceTarget,createMemoryMaintenanceDaemon,createSkillCandidateMinerTarget}from"./memory-plugins/maintenance.js";export async function runMemoryPlugins(e){const r=n(e.workspace,"write");if(e.providers?.length&&0!==r.length&&!1!==e.request.metadata?.memoryWrite)for(const t of r)await runMemoryPlugin(e,t)}export async function recallMemoryPlugins(e){const r=n(e.workspace,"recall");if(!e.providers?.length||0===r.length)return[];const t=await Promise.all(r.map(async r=>{try{return await async function recallMemoryPlugin(e,r){const t=o(e.providers??[],r.provider);if(!t)return;const n=a(e.workspace,e.agent,e.request,r),s=await t.search({namespace:n,query:buildRecallQuery(e),limit:readRecallLimit(e.workspace)});return{namespace:n,records:s,context:formatRecallContext(e.workspace,s)}}(e,r)}catch{return}}));return t.filter(e=>Boolean(e?.context))}async function runMemoryPlugin(r,t){const n=o(r.providers??[],t.provider);if(!n)return;const s=a(r.workspace,r.agent,r.request,t);r.emit({type:"runtime.memory.plugin.started",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:s});try{const e=await n.propose(function createPluginProposeInput(e,r,t){return{namespace:t,content:[`User input:\n${e.request.input}`,`Agent output:\n${e.result.text}`].join("\n\n"),sourceType:"runtime-run",sourceRef:e.requestId,metadata:{memoryId:r.id,provider:r.provider,profile:r.profile,mode:r.mode,prompts:r.prompts,workspaceRoot:e.workspace.root,agentId:e.agent.id,sessionId:e.sessionId,requestMetadata:e.request.metadata}}}(r,t,s));r.emit({type:"runtime.memory.plugin.completed",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:s,candidateCount:e.length})}catch(o){r.emit({type:"runtime.memory.plugin.failed",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:s,error:e(o)})}}function buildRecallQuery(e){const r=[`task: ${e.request.input}`,`workspace: ${t(e.workspace)}`,`agent: ${e.agent.id}`,"memory_needed: durable preferences, workspace facts, reusable procedures, prior corrections"],n=function readRecentContext(e){return"string"==typeof e?.recentContext?e.recentContext.slice(0,3e3):Array.isArray(e?.recentTurns)?e.recentTurns.filter(e=>"string"==typeof e).slice(-6).join("\n").slice(0,3e3):void 0}(e.request.metadata);return n?[...r,`recent_context: ${n}`].join("\n"):r.join("\n")}function readRecallLimit(e){const t=r(e.runtime.memory?.LangMem),n=r(t?.recall),o=n?.topK??n?.limit;return"number"==typeof o&&Number.isFinite(o)?o:10}function formatRecallContext(e,t){const n=function readRecallContextLimits(e){const t=r(e.runtime.memory?.LangMem),n=r(t?.recall);return{maxContextChars:readPositiveNumber(n?.maxContextChars)??4e3,maxRecordChars:readPositiveNumber(n?.maxRecordChars)??1e3}}(e),o=[];let a=n.maxContextChars;for(const e of t){if(a<=0)break;const r=truncateText(formatRecallRecord(e),Math.min(n.maxRecordChars,a));r.trim()&&(o.push(r),a-=r.length+1)}return o.join("\n")}function formatRecallRecord(e){return`- ${e.summary??e.content}\n  ${e.content}`}function readPositiveNumber(e){return"number"==typeof e&&Number.isFinite(e)&&e>0?Math.floor(e):void 0}function truncateText(e,r){return e.length>r?`${e.slice(0,Math.max(0,r-13))}\n  [truncated]`:e}

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

@@ -0,0 +1,11 @@
+import type { RuntimeEvent } from "../types.js";
+export declare function hasPlanningEvidence(events: RuntimeEvent[]): boolean;
+export declare function successfulEvidenceToolIds(events: RuntimeEvent[]): string[];
+export declare function successfulEvidenceOutputs(events: RuntimeEvent[]): string[];
+export type SuccessfulEvidenceItem = {
+    source: string;
+    output: string;
+};
+export declare function successfulEvidenceItems(events: RuntimeEvent[]): SuccessfulEvidenceItem[];
+export declare function controlBlockers(events: RuntimeEvent[]): string[];
+export declare function controlGaps(events: RuntimeEvent[]): string[];

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

@@ -0,0 +1 @@
+export function hasPlanningEvidence(t){return t.some(t=>{const e=readAdapterEvent(t);return!!e&&("plan"===e.traceType||String(e.traceLabel??"").startsWith("plan.")||"write_todos"===e.toolId&&String(e.phase??"").startsWith("agent.tool."))})}export function successfulEvidenceToolIds(t){const e=t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return[t.toolId];const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId&&isSuccessfulEvidenceEvent(e)?[e.toolId]:[]});return[...new Set(e)]}export function successfulEvidenceOutputs(t){return successfulEvidenceItems(t).map(t=>t.output)}export function successfulEvidenceItems(t){return t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return stringifyEvidence(t.output).map(e=>({source:t.toolId,output:e}));const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId?!isSuccessfulEvidenceEvent(e)||function isPlanningTool(t){return/^(?:write_todos|read_todos)$/u.test(t)}(e.toolId)?[]:stringifyEvidence(e.evidenceOutput??e.output).filter(t=>function isUsableEvidenceOutput(t,e){return"task"!==t||!function looksLikeUnexecutedToolIntent(t){const e=t.trim();return!(e.length>4e3)&&[/\b(?:I need to|I will|I'll|I am going to|I'm going to|Let me)\s+(?:call|use|run|invoke|get|fetch|check)\b[\s\S]{0,1200}\b(?:tool|function|system_time|web_search|url_text_fetch|task)\b/iu,/(?:我需要|我要|我会|我将|让我|我来|接下来我(?:会|将)?)\s*(?:先)?(?:调用|使用|运行|执行|检查|获取|查看)\b[\s\S]{0,1200}\b(?:工具|函数|system_time|web_search|url_text_fetch|task)\b/iu,/```(?:json)?\s*(?:system_time|web_search|url_text_fetch|task)\s*```/iu].some(t=>t.test(e))}(e)&&!function looksLikeDelegatedCommentary(t){const e=t.trim();return!(e.length>8e3)&&[/(?:这篇文章|this article)[\s\S]{0,1200}(?:我认为|我认同|打动我|值得深入探讨|特别认同|what I find|I think|I agree)/iu,/(?:你还有什么|还想深入探讨|would you like|do you want).{0,120}[?？]\s*$/iu].some(t=>t.test(e))}(e)}(e.toolId,t)).map(t=>({source:e.toolId,output:t})):[]})}export function controlBlockers(t){const e=function successfulEventIndexesBySource(t){const e=new Map;return t.forEach((t,n)=>{const o=function successfulEventSource(t){if("runtime.tool.direct.completed"===t.type)return t.toolId;const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId&&isSuccessfulEvidenceEvent(e)?e.toolId:void 0}(t);o&&e.set(o,[...e.get(o)??[],n])}),e}(t);return t.flatMap((t,n)=>{const o=readAdapterEvent(t),r=readString(o?.controlStatus)??readOutputStatus(o?.output),s=readString(o?.toolId)??"tool";return r&&isBlockerStatus(r)?isResolvedByLaterCompletion(r)&&function completedAfter(t,e,n){return(t.get(e)??[]).some(t=>t>n)}(e,s,n)?[]:[`${s}:${r}`]:[]})}export function controlGaps(t){const e=new Set(successfulEvidenceItems(t).map(t=>t.source));return t.flatMap(t=>{const n=readAdapterEvent(t),o=readString(n?.controlStatus)??readOutputStatus(n?.output),r=readString(n?.toolId)??"tool";return o&&isGapStatus(o)?e.has(r)&&isResolvedByLaterCompletion(o)?[]:[`${r}:${o}`]:[]})}function stringifyEvidence(t){return"string"==typeof t?t.trim()?[t]:[]:null==t?[]:[JSON.stringify(t)]}function readAdapterEvent(t){return"runtime.adapter.event"===t.type&&isRecord(t.event)?t.event:void 0}function isSuccessfulEvidenceEvent(t){const e=readString(t.controlStatus)??readOutputStatus(t.output);return!e||/^(?:completed|success|ok|recorded)$/iu.test(e)}function isBlockerStatus(t){return/^(?:blocked|approval_required|schema_repair_failed|tool_argument_error|invalid_input)$/iu.test(t)}function isGapStatus(t){return/^(?:dependency_required|plan_required|repeated_tool_call_limit|duplicate_tool_call)$/iu.test(t)}function isResolvedByLaterCompletion(t){return isGapStatus(t)||isBlockerStatus(t)}function readOutputStatus(t){if("string"!=typeof t)return;const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);return"string"==typeof e?.status?e.status:t.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function readString(t){return"string"==typeof t&&t.trim()?t:void 0}

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

@@ -0,0 +1,2 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
+export declare function reviewExecutionEvidence(input: QualityReviewInput, policy: QualityPolicy): QualityReviewResult;

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

@@ -0,0 +1 @@
+import{controlBlockers as e,successfulEvidenceOutputs as n,successfulEvidenceToolIds as t}from"./event-evidence.js";const r=/(?<![\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),...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,t){if(!t.executionReview.rejectUngroundedNumbers||!e.output?.text.trim())return[];const r=numberSet(n(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: ${s.slice(0,12).join(", ")}`,recoverable:!1}]}function numberSet(e){const n=new Set;for(const t of e.matchAll(r)){const e=normalizeNumber(t[0]);e&&n.add(e)}return n}function normalizeNumber(e){const n=e.replace(/,/gu,"").replace(/^\+/u,"").replace(/[%kKmMbBtTxX]$/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 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||t(e.events).length>0?[]:[{code:"missing_tool_evidence",message:"No successful tool or delegated-task evidence was observed.",recoverable:!0}]}

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

@@ -0,0 +1,9 @@
+export * from "./types.js";
+export * from "./profile.js";
+export * from "./planning-review.js";
+export * from "./execution-review.js";
+export * from "./event-evidence.js";
+export * from "./llm-review.js";
+export * from "./recovery-policy.js";
+export * from "./synthesis.js";
+export * from "./runtime.js";

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

@@ -0,0 +1 @@
+export*from"./types.js";export*from"./profile.js";export*from"./planning-review.js";export*from"./execution-review.js";export*from"./event-evidence.js";export*from"./llm-review.js";export*from"./recovery-policy.js";export*from"./synthesis.js";export*from"./runtime.js";

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

@@ -0,0 +1,7 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewModel, QualityReviewResult } from "./types.js";
+export declare function reviewWithLlm(input: {
+    phase: "planning" | "execution";
+    review: QualityReviewInput;
+    policy: QualityPolicy;
+    model?: QualityReviewModel;
+}): Promise<QualityReviewResult | undefined>;

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

@@ -0,0 +1 @@
+export async function reviewWithLlm(e){if(!e.model||"llm"!==e.policy.reviewer?.mode)return;const t="planning"===e.phase?e.policy.reviewer.planningPrompt??function defaultPlanningPrompt(){return["You are Stable Harness planning quality review.","Review only generic agent-work quality: goal coverage, executable steps, and declared inventory fit.",'Do not invent a domain workflow. Return only JSON: {"verdict":"pass|revise_plan|blocked","issues":[{"code":"...","message":"...","recoverable":true}]}'].join("\n")}():e.policy.reviewer.executionPrompt??function defaultExecutionPrompt(){return["You are Stable Harness execution evidence review.","Review only whether execution evidence supports the final answer and whether unresolved blockers or gaps remain.",'Do not answer the user request. Return only JSON: {"verdict":"pass|continue_react|blocked","issues":[{"code":"...","message":"...","recoverable":true}]}'].join("\n")}();return function parseReviewResult(e){const t=function readResponseText(e){if("string"==typeof e)return e;if(!isRecord(e))return;if("string"==typeof e.content)return e.content;const t=isRecord(e.message)?e.message:void 0;return"string"==typeof t?.content?t.content:void 0}(e),n="string"==typeof t?function parseJsonRecord(e){const t=e.trim().replace(/^```(?:json)?\s*/iu,"").replace(/\s*```$/u,"");try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t):isRecord(e)?e:void 0,r=function readVerdict(e){return"pass"===e||"revise_plan"===e||"continue_react"===e||"blocked"===e?e:void 0}(n?.verdict);if(n&&r)return{verdict:r,issues:readIssues(n.issues)}}(await e.model.invoke([{role:"system",content:t},{role:"user",content:JSON.stringify(buildReviewContext(e.review),null,2)}]))}function buildReviewContext(e){return{userRequest:e.request.input,agent:{id:e.agent.id,tools:e.agent.tools,skills:e.agent.skills??[],subagents:e.agent.subagents},finalOutput:e.output?.text,events:e.events.slice(-30).map(projectEvent)}}function projectEvent(e){return"runtime.adapter.event"===e.type?{type:e.type,event:e.event}:{...e}}function readIssues(e){return Array.isArray(e)?e.flatMap(e=>isRecord(e)&&"string"==typeof e.code&&"string"==typeof e.message?[{code:e.code,message:e.message,recoverable:!1!==e.recoverable}]:[]):[]}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

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

@@ -0,0 +1,2 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
+export declare function reviewPlanningEvidence(input: QualityReviewInput, policy: QualityPolicy): QualityReviewResult;

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

@@ -0,0 +1 @@
+import{hasPlanningEvidence as e}from"./event-evidence.js";export function reviewPlanningEvidence(n,s){return s.enabled&&s.planningReview.enabled&&s.planningReview.requirePlan?e(n.events)?{verdict:"pass",issues:[]}:{verdict:"revise_plan",issues:[{code:"missing_plan_evidence",message:"No upstream planning evidence was observed before the agent produced a result.",recoverable:!0}]}:{verdict:"pass",issues:[]}}

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

@@ -0,0 +1,3 @@
+import type { WorkspaceAgent, WorkspaceRuntimePolicy } from "../types.js";
+import type { QualityPolicy } from "./types.js";
+export declare function resolveQualityPolicy(runtime: WorkspaceRuntimePolicy, agent: WorkspaceAgent): QualityPolicy;

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

@@ -0,0 +1 @@
+export function resolveQualityPolicy(e,n){const r=n.config.quality??e.quality,i=function readProfile(e){const n="string"==typeof e?e:isRecord(e)?e.profile:void 0;return"fast"===n||"balanced"===n||"strict"===n?n:"off"}(r),o=function profileDefaults(e,n){const r=n.tools.length>0||n.subagents.length>0||(n.skills?.length??0)>0;return"off"===e?function disabledPolicy(e){return{enabled:!1,profile:e,reviewer:{mode:"deterministic"},planningReview:{enabled:!1,requirePlan:!1},executionReview:{enabled:!1,requireToolEvidence:!1,rejectEmptyFinal:!1,stopOnBlocker:!1,rejectUngroundedNumbers:!1},synthesis:{enabled:!1,mode:"evidence_only",maxEvidenceItems:5},recovery:{enabled:!1,maxLoops:0}}}(e):{enabled:!0,profile:e,planningReview:{enabled:"fast"!==e,requirePlan:"fast"!==e&&r},executionReview:{enabled:!0,requireToolEvidence:"strict"===e&&r,rejectEmptyFinal:!0,stopOnBlocker:!0,rejectUngroundedNumbers:"strict"===e},synthesis:{enabled:!1,mode:"evidence_only",maxEvidenceItems:5},recovery:{enabled:"fast"!==e,maxLoops:"strict"===e?3:2}}}(i,n),t=isRecord(r)?r:{},a=function readReviewer(e){const n=isRecord(e.reviewer)?e.reviewer:e,r="llm"===n.mode||"string"==typeof n.modelRef?"llm":"deterministic",i=isRecord(n.prompts)?n.prompts:{},o=readString(n.modelRef);return{mode:r,...o?{modelRef:o}:{},...readString(n.planningPrompt)??readString(i.planning)?{planningPrompt:readString(n.planningPrompt)??readString(i.planning)}:{},...readString(n.executionPrompt)??readString(i.execution)?{executionPrompt:readString(n.executionPrompt)??readString(i.execution)}:{}}}(t);return{enabled:o.enabled,profile:i,...a?{reviewer:a}:{},planningReview:{enabled:readBoolean(t.planningReview,"enabled")??o.planningReview.enabled,requirePlan:readBoolean(t.planningReview,"requirePlan")??o.planningReview.requirePlan},executionReview:{enabled:readBoolean(t.executionReview,"enabled")??o.executionReview.enabled,requireToolEvidence:readBoolean(t.executionReview,"requireToolEvidence")??o.executionReview.requireToolEvidence,rejectEmptyFinal:readBoolean(t.executionReview,"rejectEmptyFinal")??o.executionReview.rejectEmptyFinal,stopOnBlocker:readBoolean(t.executionReview,"stopOnBlocker")??o.executionReview.stopOnBlocker,rejectUngroundedNumbers:readBoolean(t.executionReview,"rejectUngroundedNumbers")??o.executionReview.rejectUngroundedNumbers},synthesis:{enabled:readBoolean(t.synthesis,"enabled")??o.synthesis.enabled,mode:(d=t.synthesis,("evidence_only"===(isRecord(d)?d:{}).mode?"evidence_only":void 0)??o.synthesis.mode),maxEvidenceItems:readPositiveInteger(t.synthesis,"maxEvidenceItems")??o.synthesis.maxEvidenceItems},recovery:{enabled:readBoolean(t.recovery,"enabled")??o.recovery.enabled,maxLoops:readPositiveInteger(t.recovery,"maxLoops")??o.recovery.maxLoops}};var d}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function readBoolean(e,n){const r=isRecord(e)?e:{};return"boolean"==typeof r[n]?r[n]:void 0}function readPositiveInteger(e,n){const r=(isRecord(e)?e:{})[n];return"number"==typeof r&&Number.isInteger(r)&&r>0?r:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

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

@@ -0,0 +1,11 @@
+import type { RuntimeRequest } from "../types.js";
+import type { QualityPolicy, QualityReviewResult } from "./types.js";
+export declare function buildQualityRecoveryRequest(input: {
+    request: RuntimeRequest;
+    result: QualityReviewResult;
+    phase: "planning" | "execution";
+    policy: QualityPolicy;
+    availableToolIds: string[];
+    availableSubagentIds: string[];
+    observedEvidence?: string[];
+}): RuntimeRequest | undefined;

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

@@ -0,0 +1 @@
+export function buildQualityRecoveryRequest(e){if(!e.policy.recovery.enabled||"pass"===e.result.verdict||"blocked"===e.result.verdict)return;const n="planning"===e.phase?function planningLines(e){return["Stable runtime quality review: the previous plan was not ready to execute.",...issueLines(e.result),inventoryLine("Available configured tools",e.availableToolIds),inventoryLine("Available configured subagents",e.availableSubagentIds),"Revise the plan using only the declared inventory, then continue through the backend's normal planning and tool-calling flow.","Do not invent tools, subagents, or a separate planning language."].filter(e=>e.length>0)}(e):function executionLines(e){return["Stable runtime quality review: the execution evidence is not sufficient for a final answer.",...issueLines(e.result),inventoryLine("Available configured tools",e.availableToolIds),inventoryLine("Available configured subagents",e.availableSubagentIds),...(n=e.observedEvidence??[],0===n.length?[]:["Completed observed evidence available for synthesis:",...n.slice(-8).map((e,n)=>`Evidence ${n+1}: ${function compact(e){const n=e.replace(/\s+/gu," ").trim();return n.length>900?`${n.slice(0,897)}...`:n}(e)}`)]),"Continue from the current state using ReAct: inspect the evidence gap, then either synthesize directly from completed observed evidence or choose one remaining declared tool/subagent action.","Do not call tools that already produced completed evidence or a repeat-limit control result for the same evidence need.","Preserve blockers explicitly instead of converting control states into unsupported factual answers."].filter(e=>e.length>0);var n}(e);return{...e.request,input:[e.request.input,"",...n].join("\n"),metadata:{...e.request.metadata,stableHarnessRecovery:`quality_${e.phase}`}}}function issueLines(e){return e.issues.map(e=>`- ${e.code}: ${e.message}`)}function inventoryLine(e,n){return n.length>0?`${e}: ${n.join(", ")}`:""}

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

@@ -0,0 +1,13 @@
+import type { RuntimeMemoryContext, RuntimeOutput, RuntimeRequest } from "../types.js";
+import type { QualityPolicy, QualityReviewInput, QualityReviewModel } from "./types.js";
+export type QualityRuntimeInput = QualityReviewInput & {
+    requestId: string;
+    sessionId: string;
+    emit: (event: import("../types.js").RuntimeEvent) => void;
+    getEvents: () => import("../types.js").RuntimeEvent[];
+    runAdapter: (request: RuntimeRequest) => Promise<RuntimeOutput>;
+    reviewModel?: QualityReviewModel;
+    memory?: RuntimeMemoryContext;
+    pluginMemories: RuntimeMemoryContext[];
+};
+export declare function recoverQualityReview(input: QualityRuntimeInput, request: RuntimeRequest, result: RuntimeOutput, policy: QualityPolicy): Promise<RuntimeOutput>;

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

@@ -0,0 +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")}}

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

@@ -0,0 +1,19 @@
+import type { SynthesisLanguage } from "./language.js";
+export type FactKind = "context" | "data" | "timeBound" | "limit" | "other";
+export type EvidenceFact = {
+    source: string;
+    text: string;
+    kind: FactKind;
+};
+export type EvidenceField = {
+    source: string;
+    label: string;
+    value: string;
+    kind: FactKind;
+};
+export declare function evidenceFields(facts: EvidenceFact[]): EvidenceField[];
+export declare function selectSectionFields(fields: EvidenceField[], kind: FactKind, limit: number): EvidenceField[];
+export declare function fieldTableLines(fields: EvidenceField[], language: SynthesisLanguage): string[];
+export declare function hasExtractableFields(text: string): boolean;
+export declare function classifyFact(text: string): FactKind;
+export declare function normalizeFieldLabel(label: string): string;

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

@@ -0,0 +1 @@
+export function evidenceFields(e){const t=new Set,a=[];for(const l of e)for(const e of parseEvidenceFields(l)){const l=`${normalizeFieldLabel(e.label)}\0${e.value.toLowerCase()}`;t.has(l)||(t.add(l),a.push(e))}return a}export function selectSectionFields(e,t,a){return function withSourceCoverage(e,t){const a=[...e].sort((e,t)=>scoreField(t)-scoreField(e)),l=new Map;for(const e of[...new Set(a.map(e=>e.source))]){const t=a.find(t=>t.source===e);t&&l.set(fieldKey(t),t)}for(const e of a)if(l.set(fieldKey(e),e),l.size>=t)break;return[...l.values()].slice(0,t)}(e.filter(e=>e.kind===t),a).sort((e,t)=>scoreField(t)-scoreField(e)||e.label.localeCompare(t.label))}export function fieldTableLines(e,t){return 0===e.length?[]:[..."zh"===t?["| 项目 | 数值 | 来源 |","|---|---:|---|"]:["| Field | Value | Source |","|---|---:|---|"],...e.map(e=>`| ${escapeTableCell(e.label)} | ${escapeTableCell(e.value)} | ${escapeTableCell(function humanSourceLabel(e){const t=e.replace(/([a-z])([A-Z])/gu,"$1 $2").split(/[_:.\-\s]+/u).map(e=>e.trim()).filter(Boolean);return 0===t.length?e:t.map(e=>e.charAt(0).toUpperCase()+e.slice(1)).join(" ")}(e.source))} |`)]}export function hasExtractableFields(e){return extractKeyValueFields(e).some(e=>!isBoilerplateField(e.label)&&e.value.length>0)}export function classifyFact(e){return/(?:gap|blocked|missing|unavailable|unsupported|缺口|阻塞|缺失|不支持|无法|未提供)/iu.test(e)?"limit":/(?:news|headline|recent|latest|filing|event|公告|新闻|近期|最新|披露|\b20\d{2}[-/年])/iu.test(e)?"timeBound":/(?:[$€£¥%]|\b\d[\d,]*(?:\.\d+)?\b)/u.test(e)?"data":/(?:name|company|symbol|resolved|overview|query|名称|公司|代码|概览|识别)/iu.test(e)?"context":"other"}function parseEvidenceFields(e){return extractKeyValueFields(e.text).filter(e=>e.value.length>0&&!isBoilerplateField(e.label)).filter(e=>!function isLowValueField(e,t){return!!/(?:snapshot for|source type|loaded sources)/iu.test(e)||!!/^\d{1,2}\s+(?:jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\b/iu.test(e)||scoreField({source:"",label:e,value:t,kind:"other"})<0}(e.label,e.value)).map(t=>{const a=classifyFact(`${t.label}: ${t.value}`);return{source:e.source,label:titleCaseLabel(t.label),value:formatEvidence(t.value),kind:"other"===a?e.kind:a}})}function scoreField(e){const t=normalizeFieldLabel(e.label);let a=e.value.length>0?10:0;return/(?:price|value|amount|total|ratio|rate|range|target|estimate|revenue|income|earnings|margin|volume|date|rating|status|count|metric|价格|收入|利润|目标|评级|数量|日期|范围|比率)/iu.test(t)&&(a+=8),/(?:pe ratio|p b|market cap|book value|52 week range|day range|target estimate|revenue|net income|operating income|diluted eps)/iu.test(t)&&(a+=12),/(?:name|company|symbol|title|headline|summary|名称|公司|代码|标题|摘要)/iu.test(t)&&(a+=6),/(?:unix|epoch|timestamp|raw|html|url|href|source|id|identifier|internal|debug)/iu.test(t)&&(a-=20),/^(?:date|time|cik)$/iu.test(t)&&(a-=12),/^(?:open|high|low)$/iu.test(t)&&(a-=4),(/^https?:\/\//iu.test(e.value)||"/"===e.value)&&(a-=12),e.value.length>240&&(a-=5),a}function extractKeyValueFields(e){const t=[...e.matchAll(/(^|[\s;|])([\p{L}\p{N}][\p{L}\p{N} _./()%&+\-]{0,48})[:：](?!\/\/)\s*/gu)].filter(e=>void 0!==e.index).map((e,t,a)=>{a[t-1];const l=function cleanFieldLabel(e){const t=e.replace(/^(?:[-*]\s*)+/u,"").replace(/\s+/gu," ").trim(),a=t.match(/^.+\b(PE Ratio TTM)$/iu);if(a?.[1]&&a[1]!==t&&/(?:beta|\b20\d{2}|\b(?:jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\b)/iu.test(t))return a[1];const l=t.split(" ").filter(Boolean);return l.length>1&&/^\d{4}$/u.test(l[0]??"")||l.length>1&&/^\d+(?:\.\d+)?$/u.test(l[0]??"")&&/^(?:\d+|beta|pe|p\/b)$/iu.test(l[1]??"")||l.length>1&&/^[A-Z]$/u.test(l[0]??"")||l.length>1&&/^(?:inc\.?|llc\.?|ltd\.?|corp\.?|corporation|co\.?)$/iu.test(l[0]??"")||l.length>1&&/^(?:usd|eur|gbp|jpy|cny|cad|aud)$/iu.test(l[0]??"")||l.length>1&&/^(?:gmt|utc)$/iu.test(l[0]??"")?l.slice(1).join(" "):function trailingKnownLabel(e){if(/\bPE Ratio TTM$/iu.test(e)&&!/^PE Ratio TTM$/iu.test(e))return"PE Ratio TTM";const t=e.match(/\b(?:Revenue|Net Income|Operating Income|Diluted EPS|EPS|Market Cap|Book Value|Target Estimate|Target Mean Price|Target Median Price|Target High Price|Target Low Price|PE Ratio TTM|P\/B|Beta(?: 5Y Monthly)?|Published|Updated|Date|Time|Summary|Title|Headline)$/iu);return t&&/(?:filed|\b20\d{2}|10-k|10-q|\bfy\b|\b(?:jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\b|\s{3,}|[a-z].{8,}\s)/iu.test(e.slice(0,t.index))?t?.[0]??e:e}(t)}(e[2]??""),i=function droppedLabelPrefix(e,t){const a=e.trim();if(a===t)return"";const l=a.toLowerCase().lastIndexOf(t.toLowerCase());return l>0?a.slice(0,l):""}(e[2]??"",l);return{index:(e.index??0)+String(e[1]??"").length+i.length,start:(e.index??0)+e[0].length,label:l,rawLabel:e[2]??""}}).filter(e=>!isBoundaryNoiseLabel(e.label)&&!isBoundaryNoiseLabel(e.rawLabel));return t.map((a,l)=>{const i=t[l+1]?.index??e.length;return{label:a.label,value:cleanFieldValue(e.slice(a.start,i),a.label)}}).filter(e=>e.label.length>0)}function cleanFieldValue(e,t){const a=e.replace(/^[\s,;|.-]+/u,"").replace(/[\s,;|.-]+$/u,"").replace(/\s+/gu," ").trim(),l=normalizeFieldLabel(t);if(/\brange\b/iu.test(l))return a.replace(/\s+Beta\b.*$/iu,"").trim();if(/^(?:ticker|symbol|resolved symbol|resolved ticker|股票 resolved ticker|代码)$/iu.test(l)){const e=a.match(/^([A-Z0-9._-]{1,24})(?:\s+\d+[.)]\s+|$)/u);if(e?.[1])return e[1]}return a}function isBoilerplateField(e){return/^(?:status|control status|evidence tool|tool|source|来源)$/iu.test(e.trim())}function isBoundaryNoiseLabel(e){return/^\d{1,2}\s+(?:jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\b/iu.test(e)||/^(?:jan|feb|mar|apr|may|jun|jul|aug|sep|oct|nov|dec)\s+\d{1,4}(?:\s+\d{1,2})?$/iu.test(e)||/^\d{4}\s+\d{1,2}$/u.test(e)}export function normalizeFieldLabel(e){return e.toLowerCase().replace(/[^a-z0-9\p{L}\p{N}]+/giu," ").trim()}function titleCaseLabel(e){if(/\p{Script=Han}/u.test(e))return e;const t=e.split(/[^A-Za-z0-9]+/u).filter(Boolean);return normalizeFieldLabel(e).split(" ").filter(Boolean).map((e,a)=>{const l=t[a]??e;return/^[A-Z0-9]{2,}$/u.test(l)?l:e.charAt(0).toUpperCase()+e.slice(1)}).join(" ")}function formatEvidence(e){const t=e.replace(/\s+/gu," ").trim();return t.length>1200?`${t.slice(0,1197)}...`:t}function escapeTableCell(e){return e.replace(/\|/gu,"\\|").replace(/\r?\n/gu," ")}function fieldKey(e){return`${e.source}\0${normalizeFieldLabel(e.label)}\0${e.value.toLowerCase()}`}

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

@@ -0,0 +1,3 @@
+import type { QualityReviewInput } from "../types.js";
+export type SynthesisLanguage = "en" | "zh";
+export declare function detectSynthesisLanguage(input: QualityReviewInput): SynthesisLanguage;

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

@@ -0,0 +1 @@
+export function detectSynthesisLanguage(e){const t=e.workspace.runtime.responseLanguage;if(function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}(t)){const e=readString(t.language)??readString(t.locale)??readString(t.target);if(e&&/^(?:zh|zh-|chinese|中文|汉语|漢語)/iu.test(e))return"zh";if(e&&/^(?:en|en-|english)$/iu.test(e))return"en"}return/\p{Script=Han}/u.test(e.request.input)?"zh":"en"}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}

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

@@ -0,0 +1,3 @@
+import type { EvidenceField } from "./fields.js";
+import type { SynthesisLanguage } from "./language.js";
+export declare function observedResultLines(fields: EvidenceField[], language: SynthesisLanguage): string[];

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

@@ -0,0 +1 @@
+import{normalizeFieldLabel as e}from"./fields.js";export function observedResultLines(e,t){const i=takeFields(e,isIdentityField,3),n=takeFields(e,isDataField,4),a=takeFields(e,isRecentField,2),o=[i.length>0?observationLine("identity",i,t):void 0,n.length>0?observationLine("data",n,t):void 0,a.length>0?observationLine("recent",a,t):void 0].filter(e=>Boolean(e));return o.length>0?o:["zh"===t?"- 只观察到原始证据，无法形成更具体的归纳。":"- Only raw evidence was observed; no more specific synthesis is available."]}function takeFields(t,i,n){const a=[],o=new Set;for(const s of t){const t=e(s.label);if(!o.has(t)&&i(s)&&(o.add(t),a.push(s),a.length>=n))break}return a}function observationLine(e,t,i){const n=t.map(e=>`${e.label} ${e.value}`).join("zh"===i?"；":"; ");return"zh"===i?`- ${"identity"===e?"已确认标识":"data"===e?"关键数据":"近期证据"}：${n}。`:`- ${"identity"===e?"Confirmed identifiers":"data"===e?"Observed data":"Recent evidence"}: ${n}.`}function isIdentityField(t){const i=e(t.label);return/(?:company|name|symbol|ticker|resolved|identifier|entity|名称|公司|代码|标识)/iu.test(i)}function isDataField(t){const i=e(t.label);return"data"===t.kind||/(?:price|value|amount|total|ratio|rate|range|target|estimate|revenue|income|earnings|margin|volume|count|metric|价格|收入|利润|目标|评级|数量|范围|比率)/iu.test(i)}function isRecentField(t){const i=e(t.label);return"timeBound"===t.kind||/(?:date|time|published|updated|headline|summary|event|filing|日期|时间|发布|更新|标题|摘要|事件|披露)/iu.test(i)}

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

@@ -0,0 +1,2 @@
+import type { QualityPolicy, QualityReviewInput, QualityReviewResult } from "./types.js";
+export declare function synthesizeEvidenceOnlyReport(input: QualityReviewInput, review: QualityReviewResult, policy: QualityPolicy): string | undefined;

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

@@ -0,0 +1 @@
+import{controlBlockers as e,controlGaps as t,successfulEvidenceItems as n}from"./event-evidence.js";import{classifyFact as s,evidenceFields as r,fieldTableLines as o,hasExtractableFields as i,selectSectionFields as c}from"./synthesis/fields.js";import{detectSynthesisLanguage as u}from"./synthesis/language.js";import{observedResultLines as a}from"./synthesis/observed.js";export function synthesizeEvidenceOnlyReport(s,r,o){if(!o.enabled||!o.synthesis.enabled||"evidence_only"!==o.synthesis.mode)return;if("pass"===r.verdict||!function hasRecoverableSynthesisIssue(e){return e.issues.some(e=>"control_blocker"!==e.code)}(r))return;const i=n(s.events).slice(-o.synthesis.maxEvidenceItems),c=e(s.events),a=t(s.events);if(0===i.length&&0===c.length&&0===a.length)return;const l=function latestDelegatedReport(e){const t=e.filter(e=>"task"===e.source&&function looksLikeFinalReport(e){const t=e.trim();return!(t.length<80)&&(/^#{1,3}\s+\S/mu.test(t)||/\n#{1,3}\s+\S/mu.test(t)||/\n-{3,}\n/u.test(t))}(e.output)&&!function looksLikeCommentaryInsteadOfEvidence(e){const t=e.trim();return!(t.length>8e3)&&[/(?:这篇文章|this article)[\s\S]{0,1200}(?:我认为|我认同|打动我|值得深入探讨|特别认同|what I find|I think|I agree)/iu,/(?:你还有什么|还想深入探讨|would you like|do you want).{0,120}[?？]\s*$/iu].some(e=>e.test(t))}(e.output)).at(-1)?.output.trim();return t||void 0}(i);return l||("zh"===u(s)?function buildChineseReport(e,t,n){const s=evidenceFacts(e);return["# 调查结果","","## 结论",...resultSummaryLines(e,s,t,n,"zh"),"","## 证据摘要",...summaryLines(e,t,n,"zh"),"",...structuredFactSections(s,"zh"),"","## 证据缺口与阻塞",...gapLines(t,n,"zh"),"","## 使用的证据来源",...sourceLines(sourceSummary(e),"zh")].join("\n")}(i,c,a):function buildEnglishReport(e,t,n){const s=evidenceFacts(e);return["# Result","","## Conclusion",...resultSummaryLines(e,s,t,n,"en"),"","## Evidence summary",...summaryLines(e,t,n,"en"),"",...structuredFactSections(s,"en"),"","## Evidence gaps and blockers",...gapLines(t,n,"en"),"","## Sources used",...sourceLines(sourceSummary(e),"en")].join("\n")}(i,c,a))}function resultSummaryLines(e,t,n,s,o){if(0===e.length)return["zh"===o?"- 没有成功证据可用于形成调查结论。":"- No successful evidence was available for a result."];const i=[...new Set(e.map(e=>humanSourceLabel(e.source)))],c=r(t),u=a(c,o),l=0===n.length&&0===s.length?"zh"===o?"未观察到运行时阻塞；未覆盖的业务、估值或风险细节不应视为已确认。":"No runtime blocker was observed; uncovered business, valuation, or risk details should not be treated as confirmed.":"zh"===o?`仍有 ${n.length+s.length} 个证据缺口或阻塞，结论应按部分调查读取。`:`${n.length+s.length} evidence gap(s) or blocker(s) remain, so read this as a partial result.`;return["zh"===o?`- 已基于 ${i.join("、")} 形成可交付结果。`:`- Result formed from ${i.join(", ")}.`,`- ${l}`,...u]}function sourceSummary(e){const t=new Map;for(const n of e){const e=t.get(n.source)??{source:n.source,count:0,facts:[]};e.count+=1,e.facts.push(...factLines(n.output)),t.set(n.source,e)}return[...t.values()].map(t=>({...t,facts:t.facts.length>0?t.facts.slice(0,8):[fallbackEvidenceText(evidenceTextForSource(e,t.source))]}))}function evidenceFacts(e){return e.flatMap(e=>{const t=factLines(e.output);return(t.length>0?t:[fallbackEvidenceText(e.output)]).map(t=>({source:e.source,text:t,kind:s(t)}))}).slice(0,40)}function summaryLines(e,t,n,s){if(0===e.length)return["zh"===s?"- 未观察到成功的工具或委托任务证据。":"- No successful tool or delegated-task evidence was observed."];const r=new Set(e.map(e=>e.source)).size,o=0===t.length&&0===n.length?"zh"===s?"未观察到未解决的运行时证据缺口或阻塞。":"No unresolved runtime evidence gaps or blockers were observed.":"zh"===s?`仍有 ${t.length+n.length} 个运行时证据缺口或阻塞。`:`${t.length+n.length} runtime evidence gaps or blockers remain.`;return["zh"===s?`- 已使用 ${e.length} 条完成证据，来自 ${r} 个来源。`:`- Used ${e.length} completed evidence item(s) from ${r} source(s).`,`- ${o}`]}function structuredFactSections(e,t){if(0===e.length)return["zh"===t?"- 证据缺口：没有可用于生成事实性结论的成功证据。":"- Evidence gap: no successful tool or delegated-task evidence was available."];const n=r(e);return[{kind:"context",title:"zh"===t?"## 已确认背景":"## Confirmed context"},{kind:"data",title:"zh"===t?"## 观察到的数据":"## Observed data points"},{kind:"timeBound",title:"zh"===t?"## 近期或时间相关证据":"## Recent or time-bound evidence"},{kind:"limit",title:"zh"===t?"## 证据限制":"## Evidence limits"},{kind:"other",title:"zh"===t?"## 其他观察":"## Additional observations"}].flatMap(s=>{const r=c(n,s.kind,12),u=e.filter(e=>e.kind===s.kind&&!i(e.text)).filter(e=>!function isLowValueTextFact(e){const t=e.trim();return/^https?:\/\/\S+$/iu.test(t)||/^source\s*:/iu.test(t)||/^来源\s*[:：]/u.test(t)}(e.text)).slice(0,5);return 0===r.length&&0===u.length?[]:[s.title,...o(r,t),...u.map(e=>function factLine(e,t){const n=humanSourceLabel(e.source);return"zh"===t?`- ${e.text}（来源：${n}）`:`- ${e.text} (source: ${n})`}(e,t)),""]}).filter((e,t,n)=>""!==e||t<n.length-1)}function sourceLines(e,t){return 0===e.length?["zh"===t?"- 未使用成功证据来源。":"- No successful evidence source was used."]:e.map(e=>`- ${humanSourceLabel(e.source)}${e.count>1?` (${e.count})`:""}: ${e.source}`)}function gapLines(e,t,n){return 0===e.length&&0===t.length?["zh"===n?"- 未观察到未解决的运行时证据缺口或阻塞。":"- No unresolved runtime evidence gaps or blockers were observed."]:"zh"===n?[...e.map(e=>`- 阻塞：${e}`),...t.map(e=>`- 证据缺口：${e}`)]:[...e.map(e=>`- Blocked: ${e}`),...t.map(e=>`- Evidence gap: ${e}`)]}function formatEvidence(e){const t=e.replace(/\s+/gu," ").trim();return t.length>1200?`${t.slice(0,1197)}...`:t}function factLines(e){const t=function factLinesFromJson(e){try{const t=JSON.parse(e);return function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}(t)?Object.entries(t).filter(([e])=>!/^(?:status|controlStatus)$/iu.test(e)).slice(0,8).map(([e,t])=>`${e}: ${formatEvidence(function stringifyJsonValue(e){return"string"==typeof e?e:JSON.stringify(e)}(t))}`):[]}catch{return[]}}(e);return t.length>0?t:function splitPlainTextFacts(e){const t=function stripControlPreamble(e){return e.replace(/^(?:Status:\s*(?:completed|success|ok|recorded)\b\.?\s*)+/iu,"").replace(/^(?:Evidence tool:\s*[A-Za-z0-9_.-]+\b\.?\s*)+/iu,"").trim()}(e);return t.split(/\r?\n|;\s*/u).map(e=>e.trim()).map(stripControlFragments).map(normalizePlainFactLine).filter(e=>e&&!function isControlFact(e){return/^Status:\s*(?:completed|success|ok|recorded)$/iu.test(e)||/^Evidence tool:\s*[A-Za-z0-9_.-]+$/iu.test(e)}(e))}(e).slice(0,8).map(formatEvidence)}function stripControlFragments(e){return e.replace(/\bStatus:\s*(?:completed|success|ok|recorded)\b\.?\s*/giu,"").replace(/\bEvidence tool:\s*[A-Za-z0-9_.-]+\b\.?\s*/giu,"").trim()}function normalizePlainFactLine(e){if(/^https?:\/\//iu.test(e))return"";const t=e.match(/^\d+[.)]\s+(.+)$/u);return t?.[1]?`Headline: ${t[1]}`:e}function evidenceTextForSource(e,t){return e.find(e=>e.source===t)?.output??""}function fallbackEvidenceText(e){return formatEvidence(e)}function humanSourceLabel(e){const t=e.replace(/([a-z])([A-Z])/gu,"$1 $2").split(/[_:.\-\s]+/u).map(e=>e.trim()).filter(Boolean);return 0===t.length?e:t.map(e=>e.charAt(0).toUpperCase()+e.slice(1)).join(" ")}

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

@@ -0,0 +1,52 @@
+import type { CompiledWorkspace, RuntimeEvent, RuntimeOutput, RuntimeRequest, WorkspaceAgent } from "../types.js";
+export type QualityProfile = "off" | "fast" | "balanced" | "strict";
+export type QualityReviewVerdict = "pass" | "revise_plan" | "continue_react" | "blocked";
+export type QualityReviewIssue = {
+    code: string;
+    message: string;
+    recoverable: boolean;
+};
+export type QualityReviewResult = {
+    verdict: QualityReviewVerdict;
+    issues: QualityReviewIssue[];
+};
+export type QualityPolicy = {
+    enabled: boolean;
+    profile: QualityProfile;
+    reviewer?: {
+        mode: "deterministic" | "llm";
+        modelRef?: string;
+        planningPrompt?: string;
+        executionPrompt?: string;
+    };
+    planningReview: {
+        enabled: boolean;
+        requirePlan: boolean;
+    };
+    executionReview: {
+        enabled: boolean;
+        requireToolEvidence: boolean;
+        rejectEmptyFinal: boolean;
+        stopOnBlocker: boolean;
+        rejectUngroundedNumbers: boolean;
+    };
+    synthesis: {
+        enabled: boolean;
+        mode: "evidence_only";
+        maxEvidenceItems: number;
+    };
+    recovery: {
+        enabled: boolean;
+        maxLoops: number;
+    };
+};
+export type QualityReviewInput = {
+    workspace: CompiledWorkspace;
+    agent: WorkspaceAgent;
+    request: RuntimeRequest;
+    events: RuntimeEvent[];
+    output?: RuntimeOutput;
+};
+export type QualityReviewModel = {
+    invoke(input: unknown): Promise<unknown> | unknown;
+};

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

@@ -0,0 +1 @@
+export{};

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

@@ -0,0 +1,28 @@
+import type { RuntimeEvent, RuntimeRequest } from "../types.js";
+export declare function toolCallRecoveryEnabled(policy: unknown): boolean;
+export declare function isRecoverableAdapterError(error: unknown, policy: unknown): boolean;
+export declare function buildAdapterErrorRecoveryPrompt(request: RuntimeRequest, error: unknown, policy: unknown): RuntimeRequest;
+export declare function buildResultRecoveryRequest(input: {
+    request: RuntimeRequest;
+    output: string;
+    events: RuntimeEvent[];
+    availableToolIds?: string[];
+    policy: unknown;
+}): RuntimeRequest | undefined;
+export declare function buildExecutionContractRecoveryRequest(input: {
+    request: RuntimeRequest;
+    events: RuntimeEvent[];
+    policy: unknown;
+}): RuntimeRequest | undefined;
+export declare function assertNoRawToolCallOutput(output: string, policy: unknown): void;
+export declare function containsRawToolCallOutput(output: string, policy: unknown): boolean;
+export declare function containsRecoverableResultOutput(output: string, policy: unknown): boolean;
+export declare function assertNoToolExecutionErrorOutput(output: string, policy: unknown): void;
+export declare function rawToolCallFailureMessage(): string;
+export declare function buildEvidenceSynthesisOutput(input: {
+    request: RuntimeRequest;
+    output: string;
+    events: RuntimeEvent[];
+    policy: unknown;
+}): string | undefined;
+export declare function rawToolCallOutputPreview(output: string): string;

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

@@ -0,0 +1 @@
+export function toolCallRecoveryEnabled(e){return!0===readToolCallRecovery(e).enabled}export function isRecoverableAdapterError(e,t){const o=readToolCallRecovery(t);if(!0!==o.enabled)return!1;const n=e instanceof Error?e.message:String(e);return readRegexps(o.adapterErrorPatterns,[/XML syntax error|tool.?call.*syntax|malformed.*(?:XML|tool)|Non string tool message content|repeat limit reached for tool/iu]).some(e=>e.test(n))}export function buildAdapterErrorRecoveryPrompt(e,t,o){const n=t instanceof Error?t.message:String(t),r=readToolCallRecovery(o).instruction;return recoverRequest(e,["Stable runtime recovery: the backend failed while parsing a tool call.",`Parser error: ${n}`,"string"==typeof r?r:"Continue the same user request using the backend's normal tool-calling mechanism, then return a final human-readable answer.","Do not print raw tool-call markup, JSON tool-call envelopes, or pseudo tool-call text in the final answer."])}export function buildResultRecoveryRequest(e){const t=readToolCallRecovery(e.policy);if(!0!==t.enabled)return;if(containsRawToolCallText(e.output,t)){const t=recentToolEvidence(e.events,1e3);return recoverRequest(e.request,["Stable runtime recovery: your previous final answer printed raw tool-call markup instead of executing the tool.","Continue the same user request by calling the available upstream tool normally when more evidence is required.","If you call a tool, the next assistant action must be the backend's structured tool call itself, with no prose before it.",...e.availableToolIds?.length?[`Available configured tools: ${e.availableToolIds.join(", ")}`,"Do not invent, print, or call tools that are not in this list."]:[],"If the conversation context already contains enough evidence to answer, synthesize the final answer from that context instead.","Do not print XML, JSON, markdown fences, pseudo tool-call text, plans, or future-intent text such as saying you will call or wait for a tool.",...t.length>0?["","Recent executed tool evidence:",...t]:[],"","Previous invalid final answer:",e.output])}if(containsToolExecutionErrorText(e.output,t)){const t=recentToolEvidence(e.events,1e3);return recoverRequest(e.request,["Stable runtime recovery: your previous final answer exposed a backend tool execution error instead of handling it.","Continue the same user request using the backend's normal structured tool-calling mechanism.","Do not retry the same invalid tool arguments. If the failed tool is not required to answer the user, synthesize the final answer from the available context instead.",...e.availableToolIds?.length?[`Available configured tools: ${e.availableToolIds.join(", ")}`,"Do not invent, print, or call tools that are not in this list."]:[],"Do not print tool error stacks, schema validation diagnostics, raw tool-call markup, JSON tool-call envelopes, or pseudo tool-call text in the final answer.",...t.length>0?["","Recent executed tool evidence:",...t]:[],"","Previous invalid final answer:",e.output])}const o=function lastConfiguredEventHint(e,t){const o=function readEventRecoveryHints(e){return(Array.isArray(e)?e:[]).flatMap(e=>isRecord(e)&&"string"==typeof e.instruction?[{..."string"==typeof e.toolId?{toolId:e.toolId}:{},..."string"==typeof e.phase?{phase:e.phase}:{},..."string"==typeof e.outputIncludes?{outputIncludes:e.outputIncludes}:{},..."string"==typeof e.outputMatches?{outputMatches:e.outputMatches}:{},instruction:e.instruction}]:[])}(t.eventRecoveryHints);if(0!==o.length)return e.flatMap(e=>function readMatchingHints(e,t){const o="runtime.adapter.event"===e.type&&isRecord(e.event)?e.event:void 0;return o?t.filter(e=>function eventMatchesHint(e,t){return(!t.toolId||e.toolId===t.toolId)&&(!t.phase||e.phase===t.phase)&&(t.outputIncludes?"string"==typeof e.output&&e.output.includes(t.outputIncludes):!t.outputMatches||"string"==typeof e.output&&new RegExp(t.outputMatches,"u").test(e.output))}(o,e)).map(e=>({output:"string"==typeof o.output?o.output:"Adapter event matched configured recovery hint.",instruction:e.instruction})):[]}(e,o)).at(-1)}(e.events,t);return o?recoverRequest(e.request,["Stable runtime recovery: a previous adapter event matched a configured recovery hint.",o.output,o.instruction]):void 0}export function buildExecutionContractRecoveryRequest(e){if(!0!==readToolCallRecovery(e.policy).enabled)return;const t=function lastMissingEvidenceTools(e){for(let t=e.length-1;t>=0;t-=1){const o=e[t];if("runtime.execution.contract.failed"===o?.type)return readStringArray(o.missingEvidenceTools)}return[]}(e.events);return 0!==t.length?recoverRequest(e.request,["Stable runtime recovery: the execution contract was not satisfied.",`Required evidence tool(s) were missing: ${t.join(", ")}`,"Continue the same user request by calling the missing required evidence tool(s) through the backend's normal structured tool-calling mechanism.","Do not produce a final answer until the required evidence tool call has executed and you have synthesized its result.","Do not print XML, JSON, markdown fences, pseudo tool-call text, plans, or future-intent text in the final answer."]):void 0}export function assertNoRawToolCallOutput(e,t){if(containsRawToolCallOutput(e,t))throw new Error(`Adapter returned raw tool-call text as the final answer after recovery. The backend must execute tools instead of printing tool-call markup. Output preview: ${previewOutput(e)}`)}export function containsRawToolCallOutput(e,t){const o=readToolCallRecovery(t);return!0===o.enabled&&containsRawToolCallText(e,o)}export function containsRecoverableResultOutput(e,t){const o=readToolCallRecovery(t);return!0===o.enabled&&containsRecoverableResultFailureText(e,o)}export function assertNoToolExecutionErrorOutput(e,t){const o=readToolCallRecovery(t);if(!0===o.enabled&&containsToolExecutionErrorText(e,o))throw new Error(`Adapter returned a tool execution error as the final answer after recovery. Output preview: ${previewOutput(e)}`)}export function rawToolCallFailureMessage(){return["The model attempted to call a tool but returned the tool call as text instead of executing it.","Please retry the request or use a model/backend configuration with reliable tool calling for this workspace."].join(" ")}export function buildEvidenceSynthesisOutput(e){const t=readToolCallRecovery(e.policy);if(!0!==t.enabled||!1===t.synthesizeFromEvidenceOnFailure||!containsRecoverableResultFailureText(e.output,t))return;const o=function latestDelegatedTaskReport(e){return e.flatMap(e=>{const t="runtime.adapter.event"===e.type&&isRecord(e.event)?e.event:void 0;if(!t||"agent.tool.result"!==t.phase||"task"!==t.toolId)return[];const o="string"==typeof t.output?t.output.trim():"";return function looksLikeFinalReport(e){const t=e.trim();return!(t.length<80)&&(/^#{1,3}\s+\S/mu.test(t)||/\n#{1,3}\s+\S/mu.test(t)||/\n-{3,}\n/u.test(t))}(o)?[o]:[]}).at(-1)||void 0}(e.events);if(o)return o;const n=recentToolEvidence(e.events,6e3);return 0!==n.length?"zh"==(/\p{Script=Han}/u.test(e.request.input)?"zh":"en")?function buildChineseEvidenceSynthesis(e,t){return["上游模型在已有工具证据后仍输出了伪工具调用；runtime 已拒绝该 raw 输出，并直接交付已执行工具返回的证据结果。","","已执行的工具证据：",...t,"",`被拒绝的最终输出预览：${previewRejectedOutput(e)}`].join("\n")}(e.output,n):function buildEnglishEvidenceSynthesis(e,t){return["The upstream model still returned pseudo tool-call text after tool evidence was available. The runtime rejected that raw output and is returning the executed tool evidence directly.","","Executed tool evidence:",...t,"",`Rejected final output preview: ${previewRejectedOutput(e)}`].join("\n")}(e.output,n):void 0}function previewOutput(e){const t=e.replace(/\s+/gu," ").trim();return t.length>300?`${t.slice(0,297)}...`:t}function previewRejectedOutput(e){return previewOutput(e).replace(/[<>]/gu,"")}export function rawToolCallOutputPreview(e){return previewOutput(e)}function recoverRequest(e,t){return{...e,input:[e.input,"",...t].join("\n"),metadata:{...e.metadata,stableHarnessRecovery:"tool_call"}}}function containsRawToolCallText(e,t){const o=readRegexps(t.rawOutputPatterns,[/\{\s*"name"\s*:\s*"[^"]+"\s*,\s*"arguments"\s*:/iu,/\{\s*"tool_name"\s*:\s*"[^"]+"\s*,\s*"parameters"\s*:/iu,/\{\s*"type"\s*:\s*"[^"]+"\s*,\s*"args"\s*:/iu,/^\s*[A-Za-z_][\w.-]*\s*\([^)]{0,2000}\)\s*$/iu,/^\s*[A-Za-z_][\w.-]*(?:_command|_tool|_analysis|_investigate|task)\s*$/iu,/```(?:json)?[\s\S]{0,4000}"(?:tool_name|tool|name|subagent_type)"\s*:/iu,/```(?:json)?[\s\S]{0,4000}"(?:tool_name|tool|name|subagent_type)"\s*:[\s\S]{0,4000}"(?:arguments|parameters|task)"\s*:/iu,/```(?:json)?[\s\S]{0,2000}"query"\s*:[\s\S]{0,2000}"(?:max_results|count|freshness|market)"\s*:/iu]);return!![/<\s*(?:tool_call|task)\b[^>]*>/iu,/<\s*\/\s*(?:tool_call|task)\s*>/iu,/<\s*\/?\s*tool_code\b[^>]*>/iu,/<\s*\/?\s*[A-Za-z_][\w.-]*(?:_command|_tool|_analysis|_investigate|_todos|task)\b[^>]*>/iu].some(t=>t.test(e))||function looksLikeStandaloneRecoveryCandidate(e){const t=e.trim();return t.length<=6e3||/^\s*(?:```|\{|\[|[A-Za-z_][\w.-]*\s*\()/u.test(t)}(e)&&(o.some(t=>t.test(e))||[/^[\s\S]{0,2400}\b(?:I need to|I will|I'll|I am going to|I'm going to)\s+(?:call|use|invoke|delegate)\b[\s\S]{0,1200}\b(?:tool|function|specialist|subagent)\b[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}\b(?:I will|I'll|I am going to|I'm going to)\s+(?:investigate|gather|check)\b[\s\S]{0,1200}\b(?:evidence|cluster|system|results?)\b[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}\b(?:waiting for|wait for)\b[\s\S]{0,1200}\b(?:tool|function|specialist|subagent)\b[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}\b(?:Would you like me to|Do you want me to|Should I|I can help with)\b[\s\S]{0,1200}\?[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}\bCould you please provide\b[\s\S]{0,1200}\b(?:tool|function|specialist|subagent|task)\b[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}\b(?:I don't|I do not) have enough information\b[\s\S]{0,1200}\b(?:tool|function|specialist|subagent|task|context)\b[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}\bLet me\s+(?:call|use|invoke|delegate|check|run|verify|gather|inspect)\b[\s\S]{0,1200}\b(?:tool|function|specialist|subagent|results?|data|evidence|commands?)\b[\s\S]{0,400}$/iu,/^[\s\S]{0,2400}(?:我需要|我要|我会|我将|让我|我来|接下来我(?:会|将)?)\s*(?:先)?(?:调用|使用|运行|执行|检查|读取|收集|调查|验证|查看)[\s\S]{0,1200}$/iu,/^[\s\S]{0,2400}(?:要不要|是否需要|需要我|你想让我)[\s\S]{0,1200}(?:继续|进一步|帮你|分析|检查)[\s\S]{0,1200}[?？][\s\S]{0,400}$/iu].some(t=>t.test(e)))}function containsRecoverableResultFailureText(e,t){return containsRawToolCallText(e,t)||containsToolExecutionErrorText(e,t)}function containsToolExecutionErrorText(e,t){return readRegexps(t.toolFailureOutputPatterns,[/^Error invoking tool ['"][^'"]+['"] with kwargs /iu,/Received tool input did not match expected schema/iu,/ToolMessage.*status.*error/iu]).some(t=>t.test(e))}function recentToolEvidence(e,t){return e.flatMap(e=>{const o="runtime.adapter.event"===e.type&&isRecord(e.event)?e.event:void 0;return o&&"agent.tool.result"===o.phase&&"string"==typeof o.toolId?function isControlToolOutput(e){if("string"!=typeof e||!e.trim().startsWith("{"))return!1;try{const t=JSON.parse(e),o=isRecord(t)?t.status:void 0;return"duplicate_tool_call"===o||"repeated_tool_call_limit"===o||"tool_argument_error"===o}catch{return!1}}(o.output)?[]:[`- ${o.toolId}: ${formatToolEvidence(o,t)}`]:[]}).slice(-5)}function formatToolEvidence(e,t=1e3){return"string"==typeof e.output&&e.output.trim()?e.output.slice(0,t):"string"==typeof e.error&&e.error.trim()?`error: ${e.error.slice(0,t)}`:isRecord(e.args)?`completed with args: ${previewOutput(JSON.stringify(e.args))}`:"completed"}function readToolCallRecovery(e){if(!isRecord(e))return{};const t=isRecord(e.recovery)?e.recovery:{};return isRecord(t.toolCall)?t.toolCall:{}}function readRegexps(e,t){const o=(Array.isArray(e)?e:[]).filter(e=>"string"==typeof e&&e.length>0).map(e=>new RegExp(e,"iu"));return o.length>0?o:t}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)}