stable-harness 0.0.3 → 0.0.5

package/packages/cli/package.json

@@ -15,7 +15,9 @@
   "dependencies": {
     "@langchain/langgraph-api": "^1.2.1",
     "@stable-harness/adapter-deepagents": "0.0.1",
+    "@stable-harness/adapter-langgraph": "0.0.1",
     "@stable-harness/core": "0.0.1",
+    "@stable-harness/memory": "0.0.1",
     "@stable-harness/protocols": "0.0.1",
     "@stable-harness/tool-gateway": "0.0.1",
     "@stable-harness/workspace-yaml": "0.0.1"

package/packages/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/packages/core/dist/evaluations/index.js

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

package/packages/core/dist/execution-contract.js

@@ -1 +1 @@
-export function assertExecutionContract(r){(function(e){const t=n(e?.executionContract)?e.executionContract:void 0;return!1===t?.enabled})(r.metadata)||(function(n){const r=e(n.agent);if(!r.requiresPlan)return;const i=o(r.planEvidenceTools);if(0===i.length)throw n.emit({type:"execution.contract.failed",requestId:n.requestId,sessionId:n.sessionId,agentId:n.agent.id,reason:"invalid_plan_contract"}),new Error("Execution contract requires plan evidence tools when requiresPlan is enabled");const s=n.store.getRun(n.requestId),c=new Set((s?.events??[]).flatMap(t));if(!i.some(e=>c.has(e)))throw n.emit({type:"execution.contract.failed",requestId:n.requestId,sessionId:n.sessionId,agentId:n.agent.id,reason:"missing_plan"}),new Error(`Execution contract requires a planning checkpoint from one of: ${i.join(", ")}`)}(r),function(n){const r=o(e(n.agent).requiredEvidenceTools);if(0===r.length)return;const i=n.store.getRun(n.requestId),s=new Set((i?.events??[]).flatMap(t)),c=r.filter(e=>!s.has(e));if(0!==c.length)throw n.emit({type:"execution.contract.failed",requestId:n.requestId,sessionId:n.sessionId,agentId:n.agent.id,reason:"missing_required_evidence",missingEvidenceTools:c}),new Error(`Execution contract missing required evidence tools: ${c.join(", ")}`)}(r))}function e(e){return n(e.config.executionContract)?e.config.executionContract:{}}function t(e){return"tool.completed"===e.type?[e.toolId]:"adapter.event"===e.type&&n(e.event)&&"tool.result"===e.event.phase&&"string"==typeof e.event.toolId?[e.event.toolId]:[]}function n(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function o(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}
+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 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?[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/packages/core/dist/index.d.ts

@@ -1,5 +1,6 @@
 export * from "./runtime/persistence/artifacts.js";
 export * from "./execution-contract.js";
+export * from "./recovery/tool-call.js";
 export * from "./runtime/persistence/inspection.js";
 export * from "./memory-plugins.js";
 export * from "./runtime/persistence/queue.js";
@@ -7,4 +8,6 @@ export * from "./runtime.js";
 export * from "./runtime/persistence/stores.js";
 export * from "./trace.js";
 export * from "./types.js";
+export * from "./evaluations/index.js";
+export * from "./spec-driven/index.js";
 export * from "./workflows/index.js";

package/packages/core/dist/index.js

@@ -1 +1 @@
-export*from"./runtime/persistence/artifacts.js";export*from"./execution-contract.js";export*from"./runtime/persistence/inspection.js";export*from"./memory-plugins.js";export*from"./runtime/persistence/queue.js";export*from"./runtime.js";export*from"./runtime/persistence/stores.js";export*from"./trace.js";export*from"./types.js";export*from"./workflows/index.js";
+export*from"./runtime/persistence/artifacts.js";export*from"./execution-contract.js";export*from"./recovery/tool-call.js";export*from"./runtime/persistence/inspection.js";export*from"./memory-plugins.js";export*from"./runtime/persistence/queue.js";export*from"./runtime.js";export*from"./runtime/persistence/stores.js";export*from"./trace.js";export*from"./types.js";export*from"./evaluations/index.js";export*from"./spec-driven/index.js";export*from"./workflows/index.js";

package/packages/core/dist/memory-plugins/maintenance.js

@@ -1 +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(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??u(e.workspace)},s=[];for(const t of e.targets)s.push(await o(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 c=s(t.workspace,t.agent,{input:"",sessionId:t.sessionId},o),d=i(await a.search({namespace:c,query:t.prompt,limit:e.limit??100}),await a.search({namespace:c,query:"",limit:e.limit??100}));n.push(...await a.consolidate({records:d}))}return{operations:n}}}}export function createSkillCandidateMinerTarget(e){return{name:"skillCandidateMiner",async run(t){const n=await async function(e,t){const n=[];for(const o of a(t.workspace,"all")){const a=r(e??[],o.provider);if(!a)continue;const c=s(t.workspace,t.agent,{input:"",sessionId:t.sessionId},o);n.push(...i(await a.search({namespace:c,query:t.prompt,limit:100}),await a.search({namespace:c,query:"",limit:100})))}return i(n,[])}(e.providers,t),o=function(e){const t=new Map;for(const n of e){const e=c(n);e&&t.set(e,[...t.get(e)??[],n])}return t}(n),m=[];for(const[n,a]of o){if(a.length<(e.minEvidenceCount??3))continue;const r=await e.store.upsert(d(t.workspace,n,a));m.push(r),t.emit({type:"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 ${r.name}`,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,subject:{candidate:r}}))}return{operations:[],metadata:{candidateCount:m.length}}}}}async function o(t,n){t.emit({type:"memory.maintenance.started",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,target:n.name});try{const e=await n.run(t);return t.emit({type:"memory.maintenance.completed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,target:n.name,operationCount:e.operations?.length??0}),e}catch(a){return t.emit({type:"memory.maintenance.failed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,target:n.name,error:e(a)}),{operations:[],metadata:{error:e(a)}}}}function i(e,t){const n=new Map;for(const a of[...e,...t])n.set(a.id,a);return[...n.values()]}function c(e){const n=t(e.metadata?.skillCandidate),a=t(t(e.metadata?.requestMetadata)?.skillCandidate),r=n?.name??a?.name;return"string"==typeof r&&r.trim()?function(e){return e.toLowerCase().replace(/[^a-z0-9]+/gu,"-").replace(/^-+|-+$/gu,"")}(r):void 0}function d(e,t,a){const r=function(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:m(r,a),evidence:a.map(e=>({evidenceType:"memory_record",evidenceRef:e.id,summary:e.summary??e.content.slice(0,160),weight:e.confidence}))}}function m(e,t){return[`# ${e}`,"","## Purpose","","Apply this stable workflow when the task matches the evidence.","","## Evidence","",t.map(e=>`- ${e.summary??e.content}`).join("\n")].join("\n")}function u(e){const n=t(e.runtime.memory?.maintenance);return"string"==typeof n?.prompt?n.prompt:"durable preferences workspace facts reusable procedures prior corrections stale duplicated contradicted memories"}
+import{formatError as e,readRecord as t,readWorkspaceId as n,resolveEnabledMemories as a,resolveMemoryProvider as r,resolvePluginNamespace as s}from"./shared.js";export function createMemoryMaintenanceDaemon(e){let t;const n=e.emit??(()=>{});return{async runOnce(t={}){const a=t.agent??function resolveMaintenanceAgent(e,t){const n=e.agents.get(t??e.runtime.defaultAgentId);if(!n)throw new Error("Memory maintenance requires a configured default agent");return n}(e.workspace,e.defaultAgentId),r={emit:n,requestId:t.requestId??`memory-maintenance-${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/packages/core/dist/memory-plugins/shared.js

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

@@ -1 +1 @@
-import{formatError as e,readRecord as r,readWorkspaceId as t,resolveEnabledMemories as n,resolveMemoryProvider as o,resolvePluginNamespace as s}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 a(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=>async function(e,r){const t=o(e.providers??[],r.provider);if(!t)return;const n=s(e.workspace,e.agent,e.request,r),a=await t.search({namespace:n,query:i(e),limit:c(e.workspace)});return{namespace:n,records:a,context:a.map(e=>`- ${e.summary??e.content}\n  ${e.content}`).join("\n")}}(e,r)));return t.filter(e=>Boolean(e?.context))}async function a(r,t){const n=o(r.providers??[],t.provider);if(!n)return;const a=s(r.workspace,r.agent,r.request,t);r.emit({type:"memory.plugin.started",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:a});try{const e=await n.propose(function(e,r,t){return{namespace:t,content:[`User input:\n${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,a));r.emit({type:"memory.plugin.completed",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:a,candidateCount:e.length})}catch(o){r.emit({type:"memory.plugin.failed",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:a,error:e(o)})}}function i(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=(o=e.request.metadata,"string"==typeof o?.recentContext?o.recentContext.slice(0,3e3):Array.isArray(o?.recentTurns)?o.recentTurns.filter(e=>"string"==typeof e).slice(-6).join("\n").slice(0,3e3):void 0);var o;return n?[...r,`recent_context: ${n}`].join("\n"):r.join("\n")}function c(e){const t=r(e.runtime.memory?.LangMem),n=r(t?.recall),o=n?.topK??n?.limit;return"number"==typeof o&&Number.isFinite(o)?o:10}
+import{formatError as e,readRecord as r,readWorkspaceId as t,resolveEnabledMemories as n,resolveMemoryProvider as o,resolvePluginNamespace as a}from"./memory-plugins/shared.js";export{resolveEnabledMemories,resolvePluginNamespace}from"./memory-plugins/shared.js";export{createLangMemMaintenanceTarget,createMemoryMaintenanceDaemon,createSkillCandidateMinerTarget}from"./memory-plugins/maintenance.js";export async function runMemoryPlugins(e){const r=n(e.workspace,"write");if(e.providers?.length&&0!==r.length&&!1!==e.request.metadata?.memoryWrite)for(const t of r)await runMemoryPlugin(e,t)}export async function recallMemoryPlugins(e){const r=n(e.workspace,"recall");if(!e.providers?.length||0===r.length)return[];const t=await Promise.all(r.map(async r=>{try{return await async function recallMemoryPlugin(e,r){const t=o(e.providers??[],r.provider);if(!t)return;const n=a(e.workspace,e.agent,e.request,r),s=await t.search({namespace:n,query:buildRecallQuery(e),limit:readRecallLimit(e.workspace)});return{namespace:n,records:s,context:formatRecallContext(e.workspace,s)}}(e,r)}catch{return}}));return t.filter(e=>Boolean(e?.context))}async function runMemoryPlugin(r,t){const n=o(r.providers??[],t.provider);if(!n)return;const s=a(r.workspace,r.agent,r.request,t);r.emit({type:"runtime.memory.plugin.started",requestId:r.requestId,sessionId:r.sessionId,agentId:r.agent.id,memoryId:t.id,provider:n.name,namespace:s});try{const e=await n.propose(function createPluginProposeInput(e,r,t){return{namespace:t,content:[`User input:\n${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/packages/core/dist/recovery/tool-call.d.ts

@@ -6,8 +6,23 @@ 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/packages/core/dist/recovery/tool-call.js

@@ -1 +1 @@
-export function toolCallRecoveryEnabled(t){return!0===o(t).enabled}export function isRecoverableAdapterError(t,e){const n=o(e);if(!0!==n.enabled)return!1;const a=t instanceof Error?t.message:String(t);return r(n.adapterErrorPatterns,[/XML syntax error|tool.?call.*syntax|malformed.*(?:XML|tool)|Non string tool message content/iu]).some(t=>t.test(a))}export function buildAdapterErrorRecoveryPrompt(e,r,n){const a=r instanceof Error?r.message:String(r),l=o(n).instruction;return t(e,["Stable runtime recovery: the backend failed while parsing a tool call.",`Parser error: ${a}`,"string"==typeof l?l:"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(r){const a=o(r.policy);if(!0!==a.enabled)return;if(e(r.output,a))return t(r.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, then return a final human-readable answer. Do not print XML, JSON, or pseudo tool-call text in the final answer.","","Previous invalid final answer:",r.output]);const l=function(t,e){const o=(r=e.eventRecoveryHints,(Array.isArray(r)?r:[]).flatMap(t=>n(t)&&"string"==typeof t.instruction?[{..."string"==typeof t.toolId?{toolId:t.toolId}:{},..."string"==typeof t.phase?{phase:t.phase}:{},..."string"==typeof t.outputIncludes?{outputIncludes:t.outputIncludes}:{},instruction:t.instruction}]:[]));var r;if(0!==o.length)return t.flatMap(t=>function(t,e){const o="adapter.event"===t.type&&n(t.event)?t.event:void 0;return o?e.filter(t=>function(t,e){return(!e.toolId||t.toolId===e.toolId)&&(!e.phase||t.phase===e.phase)&&(!e.outputIncludes||"string"==typeof t.output&&t.output.includes(e.outputIncludes))}(o,t)).map(t=>({output:"string"==typeof o.output?o.output:"Adapter event matched configured recovery hint.",instruction:t.instruction})):[]}(t,o)).at(-1)}(r.events,a);return l?t(r.request,["Stable runtime recovery: a previous adapter event matched a configured recovery hint.",l.output,l.instruction]):void 0}export function assertNoRawToolCallOutput(t,e){if(containsRawToolCallOutput(t,e))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.")}export function containsRawToolCallOutput(t,r){const n=o(r);return!0===n.enabled&&e(t,n)}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(" ")}function t(t,e){return{...t,input:[t.input,"",...e].join("\n"),metadata:{...t.metadata,stableHarnessRecovery:"tool_call"}}}function e(t,e){return r(e.rawOutputPatterns,[/<\s*(?:tool_call|task)\b[^>]*>/iu,/<\s*\/\s*(?:tool_call|task)\s*>/iu]).some(e=>e.test(t))}function o(t){if(!n(t))return{};const e=n(t.recovery)?t.recovery:{};return n(e.toolCall)?e.toolCall:{}}function r(t,e){const o=(Array.isArray(t)?t:[]).filter(t=>"string"==typeof t&&t.length>0).map(t=>new RegExp(t,"iu"));return o.length>0?o:e}function n(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}
+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: ${previewOutput(e)}`)}export function rawToolCallFailureMessage(){return["The model attempted to call a tool but returned the tool call as text instead of executing it.","Please retry the request or use a model/backend configuration with reliable tool calling for this workspace."].join(" ")}export function buildEvidenceSynthesisOutput(e){const t=readToolCallRecovery(e.policy);if(!0!==t.enabled||!1===t.synthesizeFromEvidenceOnFailure||!containsRecoverableResultFailureText(e.output,t))return;const o=recentToolEvidence(e.events,6e3);return 0!==o.length?"zh"==(/\p{Script=Han}/u.test(e.request.input)?"zh":"en")?function buildChineseEvidenceSynthesis(e,t){return["上游模型在已有工具证据后仍输出了伪工具调用；runtime 已拒绝该 raw 输出，并直接交付已执行工具返回的证据结果。","","已执行的工具证据：",...t,"",`被拒绝的最终输出预览：${previewRejectedOutput(e)}`].join("\n")}(e.output,o):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: ${previewRejectedOutput(e)}`].join("\n")}(e.output,o):void 0}function previewOutput(e){const t=e.replace(/\s+/gu," ").trim();return t.length>300?`${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)}

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

@@ -1 +1 @@
-export function completeRun(e){const s=e.store.getRun(e.requestId);if("cancelled"===s?.state)return t(e,"cancelled");s&&e.store.updateRun(e.requestId,{state:"completed",output:e.result.text,metadata:{...s.metadata,...e.result.metadata},artifacts:[...s.artifacts,...e.result.artifacts??[]],completedAt:(new Date).toISOString()});for(const t of e.result.artifacts??[])e.emit({type:"artifact.created",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,artifact:t});return e.emit({type:"request.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,output:e.result.text}),t(e,"completed")}export function failRun(t){const e=t.error instanceof Error?t.error.message:String(t.error);return t.store.getRun(t.requestId)&&t.store.updateRun(t.requestId,{state:"failed",error:e,completedAt:(new Date).toISOString()}),t.emit({type:"request.failed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,error:e}),{requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,state:"failed",output:e}}function t(t,e){return{requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,state:e,output:t.result.text,metadata:t.result.metadata,artifacts:t.result.artifacts}}
+export function completeRun(e){const t=e.store.getRun(e.requestId);if("cancelled"===t?.state)return response(e,"cancelled",t.artifacts);if(t){const s=function mergeArtifacts(e,t){const s=new Map;for(const r of[...e??[],...t??[]])s.set(r.id,r);return[...s.values()]}(t.artifacts,e.result.artifacts??[]);e.store.updateRun(e.requestId,{state:"completed",output:e.result.text,metadata:{...t.metadata,...e.result.metadata},artifacts:s,completedAt:(new Date).toISOString()})}for(const t of e.result.artifacts??[])e.emit({type:"runtime.artifact.created",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,artifact:t});return e.emit({type:"runtime.request.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,output:e.result.text}),response(e,"completed",e.store.getRun(e.requestId)?.artifacts)}export function failRun(e){const t=e.error instanceof Error?e.error.message:String(e.error);return e.store.getRun(e.requestId)&&e.store.updateRun(e.requestId,{state:"failed",error:t,completedAt:(new Date).toISOString()}),e.emit({type:"runtime.request.failed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,error:t}),{requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,state:"failed",output:t}}function response(e,t,s){return{requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,state:t,output:e.result.text,metadata:e.result.metadata,artifacts:s??e.result.artifacts}}

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

@@ -1 +1 @@
-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(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}}));if(t&&o.agent.tools.includes(t.toolId)&&o.gateway.get(t.toolId))return t;if(!o.agent.tools.includes(o.toolId))throw new Error(`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`);throw new Error(`Tool is not registered: ${o.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,toolId:t.toolId,args:t.args});o.emit({type:"tool.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:e.toolId});const s=await o.gateway.invoke({toolId:e.toolId,args:e.args,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id}});return o.emit({type:"tool.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:s.toolId,output:s.output}),{text:(a=s.output,"string"==typeof a?a:JSON.stringify(a)),metadata:{toolCall:{toolId:s.toolId}}};var a}
+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}}));if(t&&o.agent.tools.includes(t.toolId)&&o.gateway.get(t.toolId))return t;if(!o.agent.tools.includes(o.toolId))throw new Error(`Tool ${o.toolId} is not assigned to agent ${o.agent.id}`);throw new Error(`Tool is not registered: ${o.toolId}`)}({gateway:o.gateway,workspace:o.workspace,requestId:o.requestId,sessionId:o.sessionId,agent:o.agent,toolId:t.toolId,args:t.args});o.emit({type:"runtime.tool.direct.started",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:e.toolId});const s=await o.gateway.invoke({toolId:e.toolId,args:e.args,context:{workspaceRoot:o.workspace.root,requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id}});return o.emit({type:"runtime.tool.direct.completed",requestId:o.requestId,sessionId:o.sessionId,agentId:o.agent.id,toolId:s.toolId,output:s.output}),{text:(r=s.output,"string"==typeof r?r:JSON.stringify(r)),metadata:{toolCall:{toolId:s.toolId}}};var r}

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

@@ -2,62 +2,108 @@ import type { MemoryCandidate, MemoryDecision, MemoryRecord } from "@stable-harn
 import type { ApprovalRequest } from "@stable-harness/governance";
 import type { RuntimeArtifact } from "../types.js";
 export type RuntimeEvent = {
-    type: "request.started";
+    type: "runtime.request.started";
     requestId: string;
     sessionId: string;
     agentId: string;
+    input?: string;
 } | {
-    type: "request.completed";
+    type: "runtime.request.completed";
     requestId: string;
     sessionId: string;
     agentId: string;
     output: string;
 } | {
-    type: "request.failed";
+    type: "runtime.request.failed";
     requestId: string;
     sessionId: string;
     agentId: string;
     error: string;
 } | {
-    type: "request.cancelled";
+    type: "runtime.request.cancelled";
     requestId: string;
     sessionId: string;
     agentId: string;
     reason?: string;
 } | {
-    type: "artifact.created";
+    type: "runtime.artifact.created";
     requestId: string;
     sessionId: string;
     agentId: string;
     artifact: RuntimeArtifact;
 } | {
-    type: "execution.contract.failed";
+    type: "runtime.execution.contract.failed";
     requestId: string;
     sessionId: string;
     agentId: string;
     reason: string;
     missingEvidenceTools?: string[];
 } | {
-    type: "tool.started";
+    type: "runtime.tool.direct.started";
     requestId: string;
     sessionId: string;
     agentId: string;
     toolId: string;
 } | {
-    type: "tool.completed";
+    type: "runtime.tool.direct.completed";
     requestId: string;
     sessionId: string;
     agentId: string;
     toolId: string;
     output: unknown;
 } | {
-    type: "memory.lifecycle";
+    type: "runtime.workflow.started";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    workflowId: string;
+    adapter: string;
+} | {
+    type: "runtime.workflow.completed";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    workflowId: string;
+    adapter: string;
+} | {
+    type: "runtime.specDriven.phase.started";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    phaseId: string;
+    workflowId?: string;
+} | {
+    type: "runtime.specDriven.phase.blocked";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    phaseId: string;
+    workflowId?: string;
+    reason: string;
+} | {
+    type: "runtime.specDriven.phase.completed";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    phaseId: string;
+    workflowId?: string;
+    artifact?: RuntimeArtifact;
+} | {
+    type: "runtime.specDriven.phase.verified";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    phaseId: string;
+    workflowId?: string;
+    artifact?: RuntimeArtifact;
+} | {
+    type: "runtime.memory.lifecycle";
     requestId: string;
     sessionId: string;
     agentId: string;
     hook: RuntimeMemoryHook;
 } | {
-    type: "memory.recall.completed";
+    type: "runtime.memory.recall.completed";
     requestId: string;
     sessionId: string;
     agentId: string;
@@ -65,7 +111,7 @@ export type RuntimeEvent = {
     recordIds: string[];
     context: string;
 } | {
-    type: "memory.candidate.submitted";
+    type: "runtime.memory.candidate.submitted";
     requestId: string;
     sessionId: string;
     agentId: string;
@@ -73,13 +119,13 @@ export type RuntimeEvent = {
     decision: MemoryDecision;
     record?: MemoryRecord;
 } | {
-    type: "memory.approval.requested";
+    type: "runtime.memory.approval.requested";
     requestId: string;
     sessionId: string;
     agentId: string;
     approval: ApprovalRequest;
 } | {
-    type: "memory.plugin.started";
+    type: "runtime.memory.plugin.started";
     requestId: string;
     sessionId: string;
     agentId: string;
@@ -87,7 +133,7 @@ export type RuntimeEvent = {
     provider: string;
     namespace: string;
 } | {
-    type: "memory.plugin.completed";
+    type: "runtime.memory.plugin.completed";
     requestId: string;
     sessionId: string;
     agentId: string;
@@ -96,7 +142,7 @@ export type RuntimeEvent = {
     namespace: string;
     candidateCount: number;
 } | {
-    type: "memory.plugin.failed";
+    type: "runtime.memory.plugin.failed";
     requestId: string;
     sessionId: string;
     agentId: string;
@@ -105,27 +151,27 @@ export type RuntimeEvent = {
     namespace: string;
     error: string;
 } | {
-    type: "memory.maintenance.started";
+    type: "runtime.memory.maintenance.started";
     requestId: string;
     sessionId: string;
     agentId: string;
     target: string;
 } | {
-    type: "memory.maintenance.completed";
+    type: "runtime.memory.maintenance.completed";
     requestId: string;
     sessionId: string;
     agentId: string;
     target: string;
     operationCount: number;
 } | {
-    type: "memory.maintenance.failed";
+    type: "runtime.memory.maintenance.failed";
     requestId: string;
     sessionId: string;
     agentId: string;
     target: string;
     error: string;
 } | {
-    type: "skill.candidate.created";
+    type: "runtime.skill.candidate.created";
     requestId: string;
     sessionId: string;
     agentId: string;
@@ -136,7 +182,18 @@ export type RuntimeEvent = {
     status: string;
     proposedPath?: string;
 } | {
-    type: "adapter.event";
+    type: "runtime.progress.narration";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    message: string;
+    provider: string;
+    sourceEventTypes: string[];
+    sourceEventIds?: string[];
+    model?: string;
+    style?: string;
+} | {
+    type: "runtime.adapter.event";
     requestId: string;
     sessionId: string;
     agentId: string;

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

@@ -1 +1 @@
-export function emitMemoryLifecycle(e,t,r,s,n,o){e&&t({type:"memory.lifecycle",requestId:r,sessionId:s,agentId:n,hook:o})}export async function runMemoryRecall(t){if(!t.memory)return;if(emitMemoryLifecycle(t.memory,t.emit,t.requestId,t.sessionId,t.agent.id,"read-before-plan"),!1===t.request.memory?.recall)return;const r=e(t.workspace,t.agent,t.request),s=t.request.memory?.recall?.query??t.request.input,n=await t.memory.recall({namespace:r,query:s,limit:t.request.memory?.recall?.limit});return t.emit({type:"memory.recall.completed",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,namespace:r,recordIds:n.records.map(e=>e.id),context:n.context}),{namespace:r,records:n.records,context:n.context}}export async function submitMemoryCandidates(t){if(!t.memory||!t.request.memory?.candidates?.length)return;emitMemoryLifecycle(t.memory,t.emit,t.requestId,t.sessionId,t.agent.id,"write-after-run");const r=e(t.workspace,t.agent,t.request);for(const e of t.request.memory.candidates){const s={...e,namespace:e.namespace??r},n=await t.memory.submitCandidate(s);if(t.emit({type:"memory.candidate.submitted",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,candidate:n.candidate,decision:n.decision,record:n.record}),"review"===n.decision.action&&t.approvals){const e=await t.approvals.create({kind:"memory_write",reason:n.decision.reason,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,subject:{candidate:n.candidate,decision:n.decision}});t.emit({type:"memory.approval.requested",requestId:t.requestId,sessionId:t.sessionId,agentId:t.agent.id,approval:e})}}}function e(e,t,r){return r.memory?.namespace??`${e.root}:${t.id}`}
+export function emitMemoryLifecycle(e,r,t,s,o,n){e&&r({type:"runtime.memory.lifecycle",requestId:t,sessionId:s,agentId:o,hook:n})}export async function runMemoryRecall(e){if(!e.memory)return;if(emitMemoryLifecycle(e.memory,e.emit,e.requestId,e.sessionId,e.agent.id,"read-before-plan"),!1===e.request.memory?.recall)return;const r=resolveMemoryNamespace(e.workspace,e.agent,e.request),t=e.request.memory?.recall?.query??e.request.input,s=await e.memory.recall({namespace:r,query:t,limit:e.request.memory?.recall?.limit});return e.emit({type:"runtime.memory.recall.completed",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,namespace:r,recordIds:s.records.map(e=>e.id),context:s.context}),{namespace:r,records:s.records,context:s.context}}export async function submitMemoryCandidates(e){if(!e.memory||!e.request.memory?.candidates?.length)return;emitMemoryLifecycle(e.memory,e.emit,e.requestId,e.sessionId,e.agent.id,"write-after-run");const r=resolveMemoryNamespace(e.workspace,e.agent,e.request);for(const t of e.request.memory.candidates){const s={...t,namespace:t.namespace??r},o=await e.memory.submitCandidate(s);if(e.emit({type:"runtime.memory.candidate.submitted",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,candidate:o.candidate,decision:o.decision,record:o.record}),"review"===o.decision.action&&e.approvals){const r=await e.approvals.create({kind:"memory_write",reason:o.decision.reason,requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,subject:{candidate:o.candidate,decision:o.decision}});e.emit({type:"runtime.memory.approval.requested",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,approval:r})}}}function resolveMemoryNamespace(e,r,t){return t.memory?.namespace??`${e.root}:${r.id}`}

package/packages/core/dist/runtime/persistence/artifacts.js

@@ -1 +1 @@
-import{existsSync as t,mkdirSync as e,readFileSync as r,writeFileSync as n}from"node:fs";import i from"node:path";export function createInMemoryArtifactStore(t){const e=new Map((t?.records??[]).map(t=>[t.id,s(t)])),r=new Map(Object.entries(t?.contents??{}).map(([t,e])=>[t,structuredClone(e)]));return{createArtifact(t){const n=function(t){return{id:t.id,kind:t.kind,uri:t.uri,metadata:t.metadata,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agentId,contentType:t.contentType,createdAt:t.createdAt??(new Date).toISOString(),size:(e=t.content,Buffer.byteLength(JSON.stringify(e??null),"utf8"))};var e}(t);return e.set(n.id,n),r.set(n.id,structuredClone(t.content??null)),s(n)},listArtifacts:t=>[...e.values()].filter(e=>function(t,e){return!e||!(e.requestId&&e.requestId!==t.requestId||e.sessionId&&e.sessionId!==t.sessionId||e.agentId&&e.agentId!==t.agentId)}(e,t)).map(s),readArtifact(t){if(!r.has(t))throw new Error(`Artifact not found: ${t}`);return structuredClone(r.get(t))}}}export function createJsonFileArtifactStore(s){const o=i.resolve(s),c=createInMemoryArtifactStore(function(e){if(!t(e))return{records:[],contents:{}};const n=JSON.parse(r(e,"utf8"));if("object"!=typeof(i=n)||null===i||!Array.isArray(i.records)||"object"!=typeof i.contents)throw new Error(`Invalid artifact store file: ${e}`);var i;return n}(o));return{createArtifact(t){const r=c.createArtifact(t);return function(t,r){e(i.dirname(t),{recursive:!0}),n(t,`${JSON.stringify({version:1,...r},null,2)}\n`)}(o,function(t){const e=t.listArtifacts(),r=Object.fromEntries(e.map(e=>[e.id,t.readArtifact(e.id)]));return{records:e,contents:r}}(c)),r},listArtifacts:t=>c.listArtifacts(t),readArtifact:t=>c.readArtifact(t)}}function s(t){return structuredClone(t)}
+import{existsSync as t,mkdirSync as e,readFileSync as r,writeFileSync as n}from"node:fs";import i from"node:path";export function createInMemoryArtifactStore(t){const e=new Map((t?.records??[]).map(t=>[t.id,cloneRecord(t)])),r=new Map(Object.entries(t?.contents??{}).map(([t,e])=>[t,structuredClone(e)]));return{createArtifact(t){const n=function toRecord(t){return{id:t.id,kind:t.kind,uri:t.uri,metadata:t.metadata,requestId:t.requestId,sessionId:t.sessionId,agentId:t.agentId,contentType:t.contentType,createdAt:t.createdAt??(new Date).toISOString(),size:(e=t.content,Buffer.byteLength(JSON.stringify(e??null),"utf8"))};var e}(t);return e.set(n.id,n),r.set(n.id,structuredClone(t.content??null)),cloneRecord(n)},listArtifacts:t=>[...e.values()].filter(e=>function matchesFilter(t,e){return!e||!(e.requestId&&e.requestId!==t.requestId||e.sessionId&&e.sessionId!==t.sessionId||e.agentId&&e.agentId!==t.agentId)}(e,t)).map(cloneRecord),readArtifact(t){if(!r.has(t))throw new Error(`Artifact not found: ${t}`);return structuredClone(r.get(t))}}}export function createJsonFileArtifactStore(c){const o=i.resolve(c),s=createInMemoryArtifactStore(function readArtifactFile(e){if(!t(e))return{records:[],contents:{}};const n=JSON.parse(r(e,"utf8"));if(!function isState(t){return"object"==typeof t&&null!==t&&Array.isArray(t.records)&&"object"==typeof t.contents}(n))throw new Error(`Invalid artifact store file: ${e}`);return n}(o));return{createArtifact(t){const r=s.createArtifact(t);return function writeArtifactFile(t,r){e(i.dirname(t),{recursive:!0}),n(t,`${JSON.stringify({version:1,...r},null,2)}\n`)}(o,function snapshot(t){const e=t.listArtifacts(),r=Object.fromEntries(e.map(e=>[e.id,t.readArtifact(e.id)]));return{records:e,contents:r}}(s)),r},listArtifacts:t=>s.listArtifacts(t),readArtifact:t=>s.readArtifact(t)}}function cloneRecord(t){return structuredClone(t)}

package/packages/core/dist/runtime/persistence/inspection.js

@@ -1 +1 @@
-export function projectRequestSummary(t){return{requestId:t.requestId,sessionId:t.sessionId,agentId:t.agentId,parentRunId:t.parentRunId,state:t.state,startedAt:t.startedAt,completedAt:t.completedAt,eventCount:t.events.length,artifactCount:t.artifacts.length}}export function projectSessionSummaries(t){const e=new Map;for(const n of t)e.set(n.sessionId,[...e.get(n.sessionId)??[],n]);return[...e.entries()].map(([t,e])=>function(t,e){const r=[...e].sort((t,e)=>t.startedAt.localeCompare(e.startedAt));return{sessionId:t,requestCount:e.length,agentIds:[...new Set(e.map(t=>t.agentId))].sort(),states:n(e),startedAt:r[0]?.startedAt??"",updatedAt:s(e)}}(t,e)).sort((t,e)=>e.updatedAt.localeCompare(t.updatedAt))}export function projectRequestInspection(n,s){return{summary:projectRequestSummary(s),input:s.input,output:s.output,error:s.error,metadata:s.metadata,artifacts:s.artifacts,runtimeSnapshot:t(n,s.agentId),timeline:e(s)}}function t(t,e){const n=t.agents.get(e);return{workspaceRoot:t.root,agentId:e,backend:n?.backend??"unknown",modelRef:n?.modelRef,tools:[...n?.tools??[]],skills:[...n?.skills??[]],subagents:[...n?.subagents??[]]}}function e(t){return t.events.map((t,e)=>({index:e,type:t.type,event:t}))}function n(t){return t.reduce((t,e)=>(t[e.state]=(t[e.state]??0)+1,t),{})}function s(t){return t.map(t=>t.completedAt??t.startedAt).sort((t,e)=>e.localeCompare(t))[0]??""}
+import{applySpecDrivenPhaseTransition as t,createSpecDrivenWorkflowState as e}from"../../spec-driven/index.js";export function projectRequestSummary(t){return{requestId:t.requestId,sessionId:t.sessionId,agentId:t.agentId,parentRunId:t.parentRunId,state:t.state,startedAt:t.startedAt,completedAt:t.completedAt,eventCount:t.events.length,artifactCount:t.artifacts.length}}export function projectSessionSummaries(t){const e=new Map;for(const n of t)e.set(n.sessionId,[...e.get(n.sessionId)??[],n]);return[...e.entries()].map(([t,e])=>function projectSessionSummary(t,e){const n=[...e].sort((t,e)=>t.startedAt.localeCompare(e.startedAt));return{sessionId:t,requestCount:e.length,agentIds:[...new Set(e.map(t=>t.agentId))].sort(),states:countStates(e),startedAt:n[0]?.startedAt??"",updatedAt:latestTimestamp(e)}}(t,e)).sort((t,e)=>e.updatedAt.localeCompare(t.updatedAt))}export function projectRequestInspection(t,e){return{summary:projectRequestSummary(e),input:e.input,output:e.output,error:e.error,metadata:e.metadata,artifacts:e.artifacts,runtimeSnapshot:projectBindingSnapshot(t,e.agentId),...t.runtime.specDrivenWorkflow?{specDrivenWorkflow:projectSpecDrivenState(t.runtime.specDrivenWorkflow,e)}:{},timeline:projectTimeline(e)}}function projectSpecDrivenState(n,r){const s=new Set(n.phases.map(t=>t.id));return r.events.map(specDrivenTransition).filter(t=>void 0!==t&&s.has(t.phaseId)).reduce(t,e(n))}function specDrivenTransition(t){if(!function isSpecDrivenPhaseEvent(t){return t.type.startsWith("runtime.specDriven.phase.")}(t))return;const e=t.type.slice(25);return["started","blocked","completed","verified"].includes(e)?{phaseId:t.phaseId,status:e,..."artifact"in t&&t.artifact?{artifact:t.artifact}:{},..."reason"in t?{reason:t.reason}:{}}:void 0}function projectBindingSnapshot(t,e){const n=t.agents.get(e);return{workspaceRoot:t.root,agentId:e,backend:n?.backend??"unknown",modelRef:n?.modelRef,tools:[...n?.tools??[]],skills:[...n?.skills??[]],subagents:[...n?.subagents??[]]}}function projectTimeline(t){return t.events.map((t,e)=>({index:e,type:t.type,event:t}))}function countStates(t){return t.reduce((t,e)=>(t[e.state]=(t[e.state]??0)+1,t),{})}function latestTimestamp(t){return t.map(t=>t.completedAt??t.startedAt).sort((t,e)=>e.localeCompare(t))[0]??""}

package/packages/core/dist/runtime/persistence/queue.js

@@ -1 +1 @@
-export function createInMemoryRuntimeQueueStore(){const s=new Map,a=new Map,o=new Map;return{enqueue(e){const t=e.enqueuedAt??(new Date).toISOString(),n=r({...e,enqueuedAt:t,attemptCount:e.attemptCount??0});return s.set(n.requestId,n),r(n)},listQueue:()=>e([...s.values()]).map(r),getQueuedRequest:e=>t(s.get(e),r),claimNext(t){const n=function(t,r){const n=Date.parse(r.now??(new Date).toISOString());return e(t).find(e=>{const t=Date.parse(e.availableAt)<=n,u=!r.queueKey||e.queueKey===r.queueKey,s=!e.leaseExpiresAt||Date.parse(e.leaseExpiresAt)<=n;return t&&u&&s})}([...s.values()],t);if(!n)return;const u=t.now??(new Date).toISOString(),o=function(e,t,r,n){return{...e,claimedBy:t,claimedAt:r,leaseExpiresAt:new Date(Date.parse(r)+n).toISOString(),attemptCount:e.attemptCount+1}}(n,t.workerId,u,t.leaseMs);return s.set(o.requestId,o),a.set(o.requestId,{requestId:o.requestId,cancelRequested:a.get(o.requestId)?.cancelRequested??!1,workerId:t.workerId,workerStartedAt:a.get(o.requestId)?.workerStartedAt??u,heartbeatAt:u}),r(o)},heartbeat(e){const t=a.get(e.requestId);if(!t||t.workerId!==e.workerId)return;const r={...t,heartbeatAt:e.now??(new Date).toISOString()};return a.set(e.requestId,r),n(r)},requestCancel(e){const t={...a.get(e.requestId),requestId:e.requestId,cancelRequested:!0,cancelReason:e.reason,cancelRequestedAt:e.now??(new Date).toISOString()};return a.set(e.requestId,t),n(t)},getControl:e=>t(a.get(e),n),listStuck(e){const t=Date.parse(e.now??(new Date).toISOString());return[...s.values()].filter(r=>function(e,t,r,n){if(!e.claimedAt)return!1;const u=Date.parse(t?.heartbeatAt??e.claimedAt);return!!e.leaseExpiresAt&&Date.parse(e.leaseExpiresAt)<r||r-u>n}(r,a.get(r.requestId),t,e.staleAfterMs)).map(r)},saveRecoveryIntent:e=>(o.set(e.requestId,u(e)),u(e)),getRecoveryIntent:e=>t(o.get(e),u),clearRecoveryIntent:e=>o.delete(e)}}function e(e){return[...e].sort((e,t)=>{const r=t.priority-e.priority;return 0===r?e.enqueuedAt.localeCompare(t.enqueuedAt):r})}function t(e,t){return e?t(e):void 0}function r(e){return structuredClone(e)}function n(e){return structuredClone(e)}function u(e){return structuredClone(e)}
+export function createInMemoryRuntimeQueueStore(){const e=new Map,t=new Map,r=new Map;return{enqueue(t){const r=t.enqueuedAt??(new Date).toISOString(),n=cloneQueueRecord({...t,enqueuedAt:r,attemptCount:t.attemptCount??0});return e.set(n.requestId,n),cloneQueueRecord(n)},listQueue:()=>sortQueue([...e.values()]).map(cloneQueueRecord),getQueuedRequest:t=>cloneOptional(e.get(t),cloneQueueRecord),claimNext(r){const n=function findClaimCandidate(e,t){const r=Date.parse(t.now??(new Date).toISOString());return sortQueue(e).find(e=>{const n=Date.parse(e.availableAt)<=r,o=!t.queueKey||e.queueKey===t.queueKey,u=!e.leaseExpiresAt||Date.parse(e.leaseExpiresAt)<=r;return n&&o&&u})}([...e.values()],r);if(!n)return;const o=r.now??(new Date).toISOString(),u=function claimRecord(e,t,r,n){return{...e,claimedBy:t,claimedAt:r,leaseExpiresAt:new Date(Date.parse(r)+n).toISOString(),attemptCount:e.attemptCount+1}}(n,r.workerId,o,r.leaseMs);return e.set(u.requestId,u),t.set(u.requestId,{requestId:u.requestId,cancelRequested:t.get(u.requestId)?.cancelRequested??!1,workerId:r.workerId,workerStartedAt:t.get(u.requestId)?.workerStartedAt??o,heartbeatAt:o}),cloneQueueRecord(u)},heartbeat(e){const r=t.get(e.requestId);if(!r||r.workerId!==e.workerId)return;const n={...r,heartbeatAt:e.now??(new Date).toISOString()};return t.set(e.requestId,n),cloneControl(n)},requestCancel(e){const r={...t.get(e.requestId),requestId:e.requestId,cancelRequested:!0,cancelReason:e.reason,cancelRequestedAt:e.now??(new Date).toISOString()};return t.set(e.requestId,r),cloneControl(r)},getControl:e=>cloneOptional(t.get(e),cloneControl),listStuck(r){const n=Date.parse(r.now??(new Date).toISOString());return[...e.values()].filter(e=>function isStuck(e,t,r,n){if(!e.claimedAt)return!1;const o=Date.parse(t?.heartbeatAt??e.claimedAt);return!!e.leaseExpiresAt&&Date.parse(e.leaseExpiresAt)<r||r-o>n}(e,t.get(e.requestId),n,r.staleAfterMs)).map(cloneQueueRecord)},saveRecoveryIntent:e=>(r.set(e.requestId,cloneRecovery(e)),cloneRecovery(e)),getRecoveryIntent:e=>cloneOptional(r.get(e),cloneRecovery),clearRecoveryIntent:e=>r.delete(e)}}function sortQueue(e){return[...e].sort((e,t)=>{const r=t.priority-e.priority;return 0===r?e.enqueuedAt.localeCompare(t.enqueuedAt):r})}function cloneOptional(e,t){return e?t(e):void 0}function cloneQueueRecord(e){return structuredClone(e)}function cloneControl(e){return structuredClone(e)}function cloneRecovery(e){return structuredClone(e)}

package/packages/core/dist/runtime/persistence/stores.js

@@ -1 +1 @@
-import{existsSync as e,mkdirSync as n,readFileSync as t,writeFileSync as r}from"node:fs";import u from"node:path";export function createInMemoryRuntimeStore(e=[]){const n=new Map(e.map(e=>[e.requestId,o(e)]));return{createRun(e){n.set(e.requestId,o(e))},updateRun(e,t){const r=n.get(e);if(r)return Object.assign(r,function(e){return structuredClone(e)}(t)),o(r)},appendEvent(e){const t=n.get(e.requestId);if(t)return t.events.push(function(e){return structuredClone(e)}(e)),o(t)},getRun(e){return(t=n.get(e))?o(t):void 0;var t},listRuns:e=>[...n.values()].filter(n=>function(e,n){return!n||!(n.agentId&&n.agentId!==e.agentId||n.sessionId&&n.sessionId!==e.sessionId||n.state&&n.state!==e.state)}(n,e)).map(o)}}export function createJsonFileRuntimeStore(n){const r=u.resolve(n),o=createInMemoryRuntimeStore(function(n){if(!e(n))return[];const r=JSON.parse(t(n,"utf8"));if(!a(r)||!Array.isArray(r.runs))throw new Error(`Invalid runtime store file: ${n}`);return r.runs.map(i)}(r));return{createRun(e){o.createRun(e),s(r,o.listRuns())},updateRun(e,n){const t=o.updateRun(e,n);return s(r,o.listRuns()),t},appendEvent(e){const n=o.appendEvent(e);return s(r,o.listRuns()),n},getRun:e=>o.getRun(e),listRuns:e=>o.listRuns(e)}}function s(e,t){n(u.dirname(e),{recursive:!0}),r(e,`${JSON.stringify({version:1,runs:t},null,2)}\n`)}function i(e){if(!a(e)||"string"!=typeof e.requestId)throw new Error("Invalid runtime run record in store file");return o(e)}function o(e){return structuredClone(e)}function a(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}
+import{existsSync as e,mkdirSync as t,readFileSync as n,writeFileSync as r}from"node:fs";import u from"node:path";export function createInMemoryRuntimeStore(e=[]){const t=new Map(e.map(e=>[e.requestId,cloneRun(e)]));return{createRun(e){t.set(e.requestId,cloneRun(e))},updateRun(e,n){const r=t.get(e);if(r)return Object.assign(r,function clonePatch(e){return structuredClone(e)}(n)),cloneRun(r)},appendEvent(e){const n=t.get(e.requestId);if(!n)return;const r=function eventArtifact(e){return"artifact"in e&&e.artifact?e.artifact:void 0}(e);return r&&!n.artifacts.some(e=>e.id===r.id)&&n.artifacts.push(structuredClone(r)),n.events.push(function cloneEvent(e){return structuredClone(e)}(e)),cloneRun(n)},getRun:e=>function cloneOptionalRun(e){return e?cloneRun(e):void 0}(t.get(e)),listRuns:e=>[...t.values()].filter(t=>function matchesFilter(e,t){return!t||!(t.agentId&&t.agentId!==e.agentId||t.sessionId&&t.sessionId!==e.sessionId||t.state&&t.state!==e.state)}(t,e)).map(cloneRun)}}export function createJsonFileRuntimeStore(t){const r=u.resolve(t),i=createInMemoryRuntimeStore(function readStoreFile(t){if(!e(t))return[];const r=JSON.parse(n(t,"utf8"));if(!isRecord(r)||!Array.isArray(r.runs))throw new Error(`Invalid runtime store file: ${t}`);return r.runs.map(assertRunRecord)}(r));return{createRun(e){i.createRun(e),writeStoreFile(r,i.listRuns())},updateRun(e,t){const n=i.updateRun(e,t);return writeStoreFile(r,i.listRuns()),n},appendEvent(e){const t=i.appendEvent(e);return writeStoreFile(r,i.listRuns()),t},getRun:e=>i.getRun(e),listRuns:e=>i.listRuns(e)}}function writeStoreFile(e,n){t(u.dirname(e),{recursive:!0}),r(e,`${JSON.stringify({version:1,runs:n},null,2)}\n`)}function assertRunRecord(e){if(!isRecord(e)||"string"!=typeof e.requestId)throw new Error("Invalid runtime run record in store file");return cloneRun(e)}function cloneRun(e){return structuredClone(e)}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/packages/core/dist/runtime/progress-narration.d.ts

@@ -0,0 +1,33 @@
+import type { RuntimeEvent } from "./events.js";
+export type RuntimeProgressNarrationProvider = {
+    name: string;
+    narrate(event: RuntimeEvent, context: RuntimeProgressNarrationContext): RuntimeProgressNarrationResult;
+};
+export type RuntimeProgressNarrationResult = RuntimeProgressNarrationOutput | undefined | Promise<RuntimeProgressNarrationOutput | undefined>;
+export type RuntimeProgressNarrationContext = {
+    style?: string;
+    model?: string;
+};
+export type RuntimeProgressNarrationOutput = {
+    message: string;
+    sourceEventTypes?: string[];
+    sourceEventIds?: string[];
+    model?: string;
+    style?: string;
+};
+export type RuntimeProgressNarrationOptions = {
+    enabled?: boolean;
+    style?: string;
+    model?: string;
+    provider?: RuntimeProgressNarrationProvider;
+};
+export declare function resolveProgressNarrator(input: {
+    options?: RuntimeProgressNarrationOptions | false;
+    policy?: Record<string, unknown>;
+}): RuntimeProgressNarrationProvider | undefined;
+export declare function createProgressNarrationEvent(input: {
+    source: RuntimeEvent;
+    narrator: RuntimeProgressNarrationProvider | undefined;
+    options?: RuntimeProgressNarrationOptions | false;
+    policy?: Record<string, unknown>;
+}): RuntimeEvent | Promise<RuntimeEvent | undefined> | undefined;

package/packages/core/dist/runtime/progress-narration.js

@@ -0,0 +1 @@
+export function resolveProgressNarrator(e){if(!1===e.options)return;const t=readProgressPolicy(e.policy);return e.options?.enabled??t.enabled?e.options?.provider??function createTemplateProgressNarrator(){const e=new Map,t=new Map;return{name:"template",narrate(r,n){var o;"runtime.request.started"===r.type&&(e.set(r.requestId,{language:(o=r.input??"",/\p{Script=Han}/u.test(o)?"zh":void 0),input:r.input}),t.delete(r.requestId));const i=function templateMessage(e,t){const r="zh"===t?.language;return"runtime.request.started"===e.type?r?`我开始处理这个请求：${summarizeText(e.input)}`:`I'm starting on this request: ${summarizeText(e.input)}`:"runtime.request.completed"===e.type?r?"我已经完成执行链，正在交付最终结果。":"I've finished the execution chain and am returning the final result.":"runtime.request.failed"===e.type?r?"执行链失败了，我会保留具体错误方便定位。":"The execution chain failed; I'm keeping the concrete error visible for diagnosis.":"runtime.request.cancelled"===e.type?r?"这个请求已取消，后续执行会停止。":"The request was cancelled, so the remaining execution will stop.":"runtime.tool.direct.started"===e.type?r?`我正在运行 ${e.toolId}。`:`I'm running ${e.toolId}.`:"runtime.tool.direct.completed"===e.type?r?`${e.toolId} 已返回结果，我会继续使用这份证据。`:`${e.toolId} returned; I'll keep using that evidence.`:"runtime.workflow.started"===e.type?r?`我正在启动 workflow ${e.workflowId}。`:`I'm starting workflow ${e.workflowId}.`:"runtime.workflow.completed"===e.type?r?`workflow ${e.workflowId} 已完成。`:`Workflow ${e.workflowId} is complete.`:"runtime.specDriven.phase.started"===e.type?r?`我正在进入 spec-driven 阶段 ${e.phaseId}。`:`I'm starting spec-driven phase ${e.phaseId}.`:"runtime.specDriven.phase.blocked"===e.type?r?`spec-driven 阶段 ${e.phaseId} 被 gate 阻塞。`:`Spec-driven phase ${e.phaseId} is blocked by a gate.`:"runtime.specDriven.phase.completed"===e.type?r?`spec-driven 阶段 ${e.phaseId} 已完成。`:`Spec-driven phase ${e.phaseId} is complete.`:"runtime.specDriven.phase.verified"===e.type?r?`spec-driven 阶段 ${e.phaseId} 已验证。`:`Spec-driven phase ${e.phaseId} is verified.`:"runtime.artifact.created"===e.type?r?"我已经保存一份运行产物。":"I've saved a runtime artifact.":"runtime.execution.contract.failed"===e.type?r?"运行证据没有满足执行契约，我会进入恢复或失败路径。":"The run evidence did not satisfy the execution contract, so I'll recover or fail explicitly.":"runtime.skill.candidate.created"===e.type?r?`我发现一个可沉淀的 skill 候选：${e.name}。`:`I found a reusable skill candidate: ${e.name}.`:e.type.startsWith("runtime.memory.")?function memoryMessage(e,t){return"runtime.memory.lifecycle"===e.type?t?`我进入记忆阶段：${e.hook}。`:`I'm in the memory lifecycle phase: ${e.hook}.`:"runtime.memory.recall.completed"===e.type?t?`我召回了 ${e.recordIds.length} 条相关记忆。`:`I recalled ${e.recordIds.length} relevant memory records.`:"runtime.memory.candidate.submitted"===e.type?t?"我提交了一条候选记忆。":"I submitted a memory candidate.":"runtime.memory.approval.requested"===e.type?t?"我提交了记忆审批请求。":"I requested memory approval.":"runtime.memory.plugin.started"===e.type?t?`我正在运行记忆插件 ${e.provider}。`:`I'm running memory plugin ${e.provider}.`:"runtime.memory.plugin.completed"===e.type?t?`记忆插件 ${e.provider} 已完成。`:`Memory plugin ${e.provider} completed.`:"runtime.memory.plugin.failed"===e.type?t?`记忆插件 ${e.provider} 失败了；主任务会继续，不把它当成执行结果。`:`Memory plugin ${e.provider} failed; the main task will continue without treating it as the result.`:"runtime.memory.maintenance.started"===e.type?t?`我正在维护 ${e.target} 记忆。`:`I'm maintaining ${e.target} memory.`:"runtime.memory.maintenance.completed"===e.type?t?`${e.target} 记忆维护完成。`:`${e.target} memory maintenance completed.`:"runtime.memory.maintenance.failed"===e.type?t?`${e.target} 记忆维护失败。`:`${e.target} memory maintenance failed.`:void 0}(e,r):"runtime.adapter.event"===e.type?function adapterMessage(e,t){if(isRecord(e))return"agent.handoff"===e.phase?t?"我把请求交给上游 agent backend，让它负责规划和执行。":"I'm handing the request to the upstream agent backend for planning and execution.":"agent.tool.start"===e.phase&&"string"==typeof e.toolId?function toolStartMessage(e,t,r){if("task"===e){const e=function readTaskTarget(e){const t=isRecord(e)?e:{};return readString(t.subagent_type)??readString(t.subagentType)}(t),n=function readDescription(e){const t=isRecord(e)?e:{};return summarizeText(readString(t.description)??readString(t.task))}(t);if(e&&n)return r?`我正在请求 task 工具委派给 ${e}：${n}`:`I'm asking the task tool to delegate to ${e}: ${n}`;if(e)return r?`我正在请求 task 工具委派给 ${e}。`:`I'm asking the task tool to delegate to ${e}.`}return r?`我正在运行 ${e} 收集证据。`:`I'm running ${e} to gather evidence.`}(e.toolId,e.args,t):"agent.tool.result"===e.phase&&"string"==typeof e.toolId?function toolResultMessage(e,t,r,n,o){if(r)return o?`${e} 返回错误：${summarizeText(String(r))}`:`${e} returned an error: ${summarizeText(String(r))}`;const i="string"==typeof n?n:function readToolControlStatus(e){const t=function parseToolOutputRecord(e){if(isRecord(e))return e;if("string"==typeof e)try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);return"string"==typeof t?.status?t.status:"string"==typeof t?.error?t.error:"string"==typeof e&&e.startsWith("Task delegation target is not in the workspace inventory")?"task_inventory_blocked":void 0}(t);if(i)return function toolControlMessage(e,t,r){return"duplicate_tool_call"===t?r?`${e} 重复调用已复用已有证据。`:`${e} repeated an equivalent call, so the existing evidence was reused.`:"repeated_tool_call_limit"===t?r?`${e} 出现重复调用循环，我会基于已有证据继续收敛。`:`${e} entered a repeated-call loop, so I'll continue from the evidence already collected.`:"tool_argument_error"===t||"tool_argument_validation_failed"===t?r?`${e} 参数被工具 schema 拒绝，我会让 agent 修正参数或改用合适工具。`:`${e} arguments were rejected by the tool schema; I'll have the agent repair them or choose a suitable tool.`:"task_inventory_blocked"===t?r?"task 委派目标不在当前 workspace inventory 中，已被运行时策略阻止。":"The task delegation target is not in the current workspace inventory and was blocked by runtime policy.":r?`${e} 返回运行时控制状态：${t}。`:`${e} returned runtime control status: ${t}.`}(e,i,o);if("task"===e)return o?"委派任务已返回结果，我会基于这些证据继续推进。":"The delegated task returned; I'll keep going with that evidence.";const a=function summarizeToolOutput(e){if("string"==typeof e)return summarizeText(e);if(isRecord(e)){if("string"==typeof e.status)return summarizeText(e.status);if("string"==typeof e.summary)return summarizeText(e.summary)}}(t);return o?`${e} 已返回${a?`：${a}`:"，我会继续判断下一步。"}`:`${e} returned${a?`: ${a}`:"; I'll decide the next step from here."}`}(e.toolId,e.output,e.error,e.controlStatus,t):"agent.langgraph.invoke"===e.phase?t?"我正在调用 workflow backend。":"I'm invoking the workflow backend.":"agent.node.completed"===e.phase&&"string"==typeof e.nodeId?t?`节点 ${e.nodeId} 已完成。`:`Node ${e.nodeId} is complete.`:void 0}(e.event,r):void 0}(r,e.get(r.requestId));if(i&&i!==t.get(r.requestId))return t.set(r.requestId,i),"runtime.request.failed"!==r.type&&"runtime.request.cancelled"!==r.type||(e.delete(r.requestId),t.delete(r.requestId)),function pruneRequestState(e,t){for(;e.size>200;){const r=e.keys().next().value;if(!r)break;e.delete(r),t.delete(r)}}(e,t),{message:i,style:n.style,model:n.model}}}}():void 0}export function createProgressNarrationEvent(e){if(!e.narrator||"runtime.progress.narration"===e.source.type)return;const t=readProgressPolicy(e.policy),r=!1===e.options?void 0:e.options,n=e.narrator.narrate(e.source,{style:r?.style??t.style,model:r?.model??t.model});return function isPromiseLike(e){return isRecord(e)&&"function"==typeof e.then}(n)?n.then(t=>toNarrationEvent(e.source,e.narrator,t)):toNarrationEvent(e.source,e.narrator,n)}function toNarrationEvent(e,t,r){if(r?.message.trim())return{type:"runtime.progress.narration",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agentId,message:r.message,provider:t.name,sourceEventTypes:r.sourceEventTypes??[e.type],sourceEventIds:r.sourceEventIds,model:r.model,style:r.style}}function summarizeText(e){const t=e?.replace(/\s+/gu," ").trim();return t?t.length>120?`${t.slice(0,117)}...`:t:""}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}function readProgressPolicy(e){const t=isRecord(e?.progress)?e.progress:{},r=isRecord(t.narration)?t.narration:{};return{enabled:"boolean"==typeof r.enabled?r.enabled:void 0,style:"string"==typeof r.style?r.style:void 0,model:"string"==typeof r.model?r.model:void 0}}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

package/packages/core/dist/runtime/tool-gateway.d.ts

@@ -13,12 +13,16 @@ export type RuntimeToolGatewayTool = {
     description?: string;
     schema?: unknown;
 };
+export type RuntimeToolRepairModel = {
+    invoke(input: unknown): Promise<unknown> | unknown;
+};
 export type RuntimeToolGateway = {
     get(toolId: string): RuntimeToolGatewayTool | undefined;
     repairToolCall?(request: {
         toolId: string;
         args?: unknown;
         allowedToolIds?: string[];
+        repairModel?: RuntimeToolRepairModel;
         context: RuntimeToolGatewayContext;
     }): Promise<{
         toolId: string;
@@ -27,6 +31,7 @@ export type RuntimeToolGateway = {
     invoke(request: {
         toolId: string;
         args?: unknown;
+        repairModel?: RuntimeToolRepairModel;
         context: RuntimeToolGatewayContext;
     }): Promise<{
         toolId: string;

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

@@ -1,6 +1,6 @@
 import type { ApprovalQueue } from "@stable-harness/governance";
 import type { MemoryProvider, RuntimeMemoryStore } from "@stable-harness/memory";
-import type { CompiledWorkspace, RuntimeToolGateway, RuntimeAdapter, RuntimeStore, RuntimeWorkflowAdapter, StableHarnessRuntime } from "./types.js";
+import type { CompiledWorkspace, RuntimeToolGateway, RuntimeAdapter, RuntimeStore, RuntimeProgressNarrationOptions, RuntimeWorkflowAdapter, StableHarnessRuntime } from "./types.js";
 type RuntimeFactoryInput = {
     workspace: CompiledWorkspace;
     adapters: RuntimeAdapter[];
@@ -10,6 +10,7 @@ type RuntimeFactoryInput = {
     approvals?: ApprovalQueue;
     toolGateway?: RuntimeToolGateway;
     store?: RuntimeStore;
+    progressNarration?: RuntimeProgressNarrationOptions | false;
 };
 export declare function createStableHarnessRuntime(input: RuntimeFactoryInput): StableHarnessRuntime;
 export {};