stable-harness 0.0.29 → 0.0.30

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stable-harness",
-  "version": "0.0.29",
+  "version": "0.0.30",
   "type": "module",
   "description": "Stable application runtime and operator control plane for agent workspaces.",
   "license": "Apache-2.0",
@@ -62,7 +62,7 @@
     "benchmark:tool-guard:matrix": "node scripts/benchmarks/tool-argument-guard-matrix.mjs",
     "test:langmem:sqlite:e2e": "node scripts/run-langmem-sqlite-e2e.mjs",
     "validate:workspace": "node scripts/validate-workspace.mjs",
-    "test": "rm -rf dist/test && tsc -b test/tsconfig.json && node --test dist/test/*.test.js dist/test/adapter/*.test.js dist/test/adapter/*/*.test.js dist/test/compat/*.test.js dist/test/memory/*.test.js dist/test/protocol/*.test.js dist/test/retry/*.test.js dist/test/runtime/*.test.js dist/test/runtime/*/*.test.js dist/test/sdk/*.test.js dist/test/workspace/*.test.js",
+    "test": "rm -rf dist/test && tsc -b test/tsconfig.json && node --test dist/test/*.test.js dist/test/adapter/*.test.js dist/test/adapter/*/*.test.js dist/test/compat/*.test.js dist/test/evaluation/*.test.js dist/test/memory/*.test.js dist/test/protocol/*.test.js dist/test/retry/*.test.js dist/test/runtime/*.test.js dist/test/runtime/*/*.test.js dist/test/sdk/*.test.js dist/test/workspace/*.test.js",
     "test:langmem:maintenance:e2e": "node scripts/run-langmem-maintenance-e2e.mjs",
     "test:skill-mining:e2e": "node scripts/run-skill-candidate-mining-e2e.mjs",
     "prepublishOnly": "npm run build && npm run release:check-package",

package/packages/adapter-deepagents/dist/src/internal/builtin-tool-policy.js

@@ -1 +1 @@
-import{realpathSync as e}from"node:fs";import{repairCallSelection as t}from"@botbotgo/better-call";import{ToolMessage as o}from"@langchain/core/messages";import{normalizeArgsRecord as n}from"./builtin-args.js";import{repairBuiltinToolRequest as r}from"./builtin-call-repair.js";import{afterToolInvoke as i,beforeToolInvoke as a,createToolRepeatState as s,stringifyDeepAgentResult as l,toolControlProjection as d}from"./gateway-tools.js";import{observedToolEvidence as u,recordObservedToolEvidence as c}from"./gateway/tool-evidence.js";import{validateSkillFileBuiltinCall as p}from"./skill-file-policy.js";import{filterRepeatLimitedTools as g}from"./tool-repeat-visibility.js";import{traceProjectionForBuiltinTool as f}from"./trace-projection.js";const m=new Set(["ls","read_file","write_file","edit_file","glob","grep"]),h=new Set(["write_todos","read_todos","task","execute",...m]);export function createBuiltinToolPolicyMiddleware(e,t={}){return{name:"StableHarnessBuiltinToolPolicy",async wrapModelCall(o,n){const r=Array.isArray(o.tools)?g(o.tools.filter(t=>!function hasHiddenBuiltins(e){return isFilesystemDisabled(e)||!isTaskVisible(e)}(e)||isModelVisibleBuiltin(e,t.name)),t.repeatState):o.tools,i=function normalizeToolChoice(e,t,o){return"required"===t?o&&o.length>0?t:"auto":function isForcedHiddenTool(e,t){return"string"==typeof t?.function?.name&&!isModelVisibleBuiltin(e,t.function.name)}(e,t)?"auto":t}(e,o.toolChoice,r);return n({...o,tools:r,toolChoice:i})}}}export function validateFilesystemBuiltinCall(e,t,n){if(isFilesystemDisabled(e)&&m.has(t))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-policy`,name:t,content:`Filesystem builtin tool '${t}' is disabled for this agent. Do not retry filesystem tools; use the agent's registered workspace tools and already collected evidence instead.`})}export function createObserverMiddleware(e,d={}){const c=d.repeatState??s(e.workspace.runtime.toolGateway);return{name:"StableHarnessObserver",async wrapToolCall(s,g){const f=s.toolCall?.name;if(!f||!h.has(f))return g(s);const m=await r({toolId:f,request:s,workspaceRoot:e.workspace.root});emitToolEvent(e,f,"agent.tool.start",m.toolCall?.args);const y="task"===f?await async function repairTaskCall(e,r){const i=function allowedTaskTypes(e){const t=readConfigRecord(e.agent.config,"deepagents");if(!0===t?.generalPurposeAgent)return;const o=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1===o?[]:Array.isArray(o)?o.includes("task")?e.agent.subagents:[]:0===e.agent.subagents.length?[]:void 0}(e);if(void 0===i)return{request:r};const a=n(r.toolCall?.args),s=function readTaskSubagentType(e){const t=isRecord(e)?e:{};return readString(t.subagent_type)??readString(t.subagentType)}(a);if(s&&i.includes(s))return{request:r};const l=await t({call:{id:s,args:a},candidates:taskCallCandidates(e,i),mode:"repair"});if(l.ok){emitInventoryRepair(e,"repaired",s,l.candidateId,i);const t={...a,...l.args,subagent_type:l.candidateId};return{request:{...r,toolCall:{...r.toolCall,args:t}}}}emitInventoryRepair(e,"blocked",s,void 0,i,l.reason);const d=s?`: ${s}`:"",u=i.length>0?i.join(", "):"none";return{request:r,blocked:new o({tool_call_id:r.toolCall?.id??"stable-harness-task-policy",name:"task",status:"error",content:[`Task delegation target is not in the workspace inventory${d}. Allowed task targets: ${u}.`,"Retry with an allowed target only when that target semantically owns the original user request.","Do not substitute another specialist just to continue the same evidence need; synthesize from already collected evidence when no allowed target is a semantic match."].join(" ")})}}(e,m):{request:m},v=y.request,b=y.blocked;if(b)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:b.content}),b;const k=validateFilesystemBuiltinCall(e,f,v);if(k)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:k.content}),k;const w=p(e,f,v);if(w)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:w.content}),w;try{const t=c?a(f,v.toolCall?.args,c):void 0;if(t)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:t.eventOutput}),builtinToolMessage(s,f,t.modelOutput);const o=await g(v),n=function observedToolOutput(e,t,o){return"write_todos"===e?JSON.stringify({status:"recorded",args:t.toolCall?.args}):l(o)}(f,v,o),r=c?i(f,v.toolCall?.args,n,o,c):{};return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:r.eventOutput??n}),d.observedToolIds?.add(f),void 0===r.modelOutput?o:builtinToolMessage(s,f,r.modelOutput)}catch(t){const n=function recoverableBuiltinToolError(e,t,n,r){const i=formatError(r);if("task"===t&&/repeat limit reached for tool/iu.test(i)){const t=function formatObservedEvidence(e){const t=u(e);if(0===t.length)return"";const o=t.map(e=>[`Agent: ${e.agentId}`,`Tool: ${e.toolId}`,e.output].join("\n")).join("\n\n---\n\n");return o.length>12e3?`${o.slice(0,12e3)}\n[truncated]`:o}(e);return new o({tool_call_id:n.toolCall?.id??"stable-harness-task-repeat-limit",name:"task",content:JSON.stringify({status:"delegated_task_repeat_limit",finalizationRequired:!0,instruction:"The delegated agent reached a configured tool repeat limit. Stop delegating this evidence need and do not send a synthesis task to another subagent for the same need. Finalize only from observedEvidence and other evidence that is already visible in this run. If the visible evidence is insufficient for a requested claim, report an explicit blocker or evidence gap instead of estimating, inventing, or using generic knowledge.",...t?{observedEvidence:t}:{},error:previewError(i)})})}if(/Received tool input did not match expected schema|Invalid input:/iu.test(i))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-argument-error`,name:t,status:"error",content:JSON.stringify({status:"tool_argument_error",toolId:t,instruction:"The upstream builtin tool rejected these arguments. Fix the tool arguments according to the tool schema, or choose a more appropriate available tool.",error:previewError(i)})})}(e,f,s,t);if(n)return emitToolEvent(e,f,"agent.tool.result",m.toolCall?.args,{output:n.content}),n;throw emitToolEvent(e,f,"agent.tool.result",m.toolCall?.args,{error:formatError(t)}),t}}}}function builtinToolMessage(e,t,n){return new o({tool_call_id:e.toolCall?.id??`stable-harness-${t}-repeat-guard`,name:t,content:n})}export function resolveFilesystemPermissions(e,t){const o=readConfigRecord(t?.config,"builtinTools"),n=[];if(n.push(...function skillReadPermissions(e,t){const o=[...new Set((t?.skills??[]).flatMap(t=>function skillReadPaths(e,t){return t?[...new Set([t,canonicalPath(t),backendSkillPath(e,t)])].flatMap(e=>function skillReadPathCandidates(e){const t=e.replace(/\/+$/u,""),o=t.endsWith("/SKILL.md")?t.slice(0,-9):t,n=function parentPath(e){const t=e.lastIndexOf("/");return t>0?e.slice(0,t):void 0}(o);return[t,o,`${o}/**`,...n?[n,`${n}/**`]:[]]}(e)):[]}(e.workspace.root,e.workspace.skills.get(t)?.path)).filter(e=>e.startsWith("/")))];return o.length>0?[{operations:["read"],paths:o,mode:"allow"}]:[]}(e,t)),!1!==o?.filesystem){if(deepagentsMemoryWritable(e))return n.length>0?n:void 0}else n.push({operations:["read"],paths:["/memories/**"],mode:"allow"}),n.push({operations:["read","write"],paths:["/**"],mode:"deny"});return deepagentsMemoryWritable(e)||n.push({operations:["write"],paths:["/memories/**"],mode:"deny"}),n.length>0?n:void 0}function emitToolEvent(e,t,o,n,r={}){"string"==typeof r.output&&c(e,e.agent.id,t,r.output),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:o,toolId:t,..."agent.tool.start"===o?{args:n}:{},...r,..."string"==typeof r.output?d(r.output):{},...f(t,o,n)}})}function backendSkillPath(e,t){const o=function pathRelative(e,t){const o=e.split("/").filter(Boolean),n=t.split("/").filter(Boolean);if(!(n.length<o.length)){for(let e=0;e<o.length;e+=1)if(o[e]!==n[e])return;return n.slice(o.length).join("/")}}(e,t);return void 0===o?t:`/${o.split("/").filter(Boolean).join("/")}`}function canonicalPath(t){try{return e.native(t)}catch{return t}}function deepagentsMemoryWritable(e){const t=readConfigRecord(e.workspace.runtime.memory,"deepagentsMem");return!1!==t?.write}function emitInventoryRepair(e,t,o,n,r,i){e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:"inventory.repair",status:t,diagnostic:{layer:"task",owner:"stable_runtime_policy",originalId:o,repairedId:n,candidateIds:r,reason:i}}})}function taskCallCandidates(e,t){return t.map(t=>({id:t,description:e.workspace.agents.get(t)?.description,schema:{type:"object",properties:{subagent_type:{type:"string"},subagentType:{type:"string"},description:{type:"string"}},required:["description"],additionalProperties:!0}}))}function isFilesystemDisabled(e){const t=readConfigRecord(e.agent.config,"builtinTools");return!1===t?.filesystem}function isTaskVisible(e){const t=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1!==t&&(!Array.isArray(t)||t.includes("task"))}function isModelVisibleBuiltin(e,t){return(!isFilesystemDisabled(e)||!function isFilesystemTool(e){return"string"==typeof e&&m.has(e)}(t))&&("task"!==t||isTaskVisible(e))}function readConfigRecord(e,t){const o=isRecord(e)?e:{};return isRecord(o[t])?o[t]:void 0}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function formatError(e){return e instanceof Error?e.message:String(e)}function previewError(e){const t=e.replace(/\s+/gu," ").trim();return t.length>800?`${t.slice(0,797)}...`:t}
+import{realpathSync as e}from"node:fs";import{repairCallSelection as t}from"@botbotgo/better-call";import{ToolMessage as o}from"@langchain/core/messages";import{normalizeArgsRecord as n}from"./builtin-args.js";import{repairBuiltinToolRequest as r}from"./builtin-call-repair.js";import{afterToolInvoke as i,beforeToolInvoke as a,createToolRepeatState as s,stringifyDeepAgentResult as l,toolControlProjection as u}from"./gateway-tools.js";import{observedToolEvidence as d,recordObservedToolEvidence as c}from"./gateway/tool-evidence.js";import{validateSkillFileBuiltinCall as p}from"./skill-file-policy.js";import{filterRepeatLimitedTools as g}from"./tool-repeat-visibility.js";import{traceProjectionForBuiltinTool as f}from"./trace-projection.js";const m=new Set(["ls","read_file","write_file","edit_file","glob","grep"]),h=new Set(["write_todos","read_todos","task","execute",...m]);export function createBuiltinToolPolicyMiddleware(e,t={}){return{name:"StableHarnessBuiltinToolPolicy",async wrapModelCall(o,n){const r=Array.isArray(o.tools)?g(o.tools.filter(t=>!function hasHiddenBuiltins(e){return isFilesystemDisabled(e)||!isTaskVisible(e)}(e)||isModelVisibleBuiltin(e,t.name)),t.repeatState):o.tools,i=function normalizeToolChoice(e,t,o){return"required"===t?o&&o.length>0?t:"auto":function isForcedHiddenTool(e,t){return"string"==typeof t?.function?.name&&!isModelVisibleBuiltin(e,t.function.name)}(e,t)?"auto":t}(e,o.toolChoice,r);return n({...o,tools:r,toolChoice:i})}}}export function validateFilesystemBuiltinCall(e,t,n){if(isFilesystemDisabled(e)&&m.has(t))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-policy`,name:t,content:`Filesystem builtin tool '${t}' is disabled for this agent. Do not retry filesystem tools; use the agent's registered workspace tools and already collected evidence instead.`})}export function createObserverMiddleware(e,u={}){const c=u.repeatState??s(e.workspace.runtime.toolGateway);return{name:"StableHarnessObserver",async wrapToolCall(s,g){const f=s.toolCall?.name;if(!f||!h.has(f))return g(s);const m=await r({toolId:f,request:s,workspaceRoot:e.workspace.root});emitToolEvent(e,f,"agent.tool.start",m.toolCall?.args);const y="task"===f?await async function repairTaskCall(e,r){const i=function allowedTaskTypes(e){const t=readConfigRecord(e.agent.config,"deepagents");if(!0===t?.generalPurposeAgent)return;const o=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1===o?[]:Array.isArray(o)?o.includes("task")?e.agent.subagents:[]:0===e.agent.subagents.length?[]:void 0}(e);if(void 0===i)return{request:r};const a=n(r.toolCall?.args),s=function readTaskSubagentType(e){const t=isRecord(e)?e:{};return readString(t.subagent_type)??readString(t.subagentType)}(a);if(s&&i.includes(s))return{request:r};const l=await t({call:{id:s,args:a},candidates:taskCallCandidates(e,i),mode:"repair"});if(l.ok){emitInventoryRepair(e,"repaired",s,l.candidateId,i);const t={...a,...l.args,subagent_type:l.candidateId};return{request:{...r,toolCall:{...r.toolCall,args:t}}}}emitInventoryRepair(e,"blocked",s,void 0,i,l.reason);const u=s?`: ${s}`:"",d=i.length>0?i.join(", "):"none";return{request:r,blocked:new o({tool_call_id:r.toolCall?.id??"stable-harness-task-policy",name:"task",status:"error",content:[`Task delegation target is not in the workspace inventory${u}. Allowed task targets: ${d}.`,"Retry with an allowed target only when that target semantically owns the original user request.","Do not substitute another specialist just to continue the same evidence need; synthesize from already collected evidence when no allowed target is a semantic match."].join(" ")})}}(e,m):{request:m},v=y.request,b=y.blocked;if(b)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:b.content}),b;const k=validateFilesystemBuiltinCall(e,f,v);if(k)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:k.content}),k;const w=p(e,f,v);if(w)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:w.content}),w;try{const t=c?a(f,v.toolCall?.args,c):void 0;if(t)return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:t.eventOutput}),builtinToolMessage(s,f,function modelOutputForRepeatedBuiltin(e,t){return"task"!==e?t:function readPreviousOutput(e){try{const t=JSON.parse(e);if(!isRecord(t)||"repeated_tool_call_limit"!==t.status||"string"!=typeof t.previousOutput)return;const o=t.previousOutput.trim();return o.length>0?o:void 0}catch{return}}(t)??t}(f,t.modelOutput));const o=await g(v),n=function observedToolOutput(e,t,o){return"write_todos"===e?JSON.stringify({status:"recorded",args:t.toolCall?.args}):l(o)}(f,v,o),r=c?i(f,v.toolCall?.args,n,o,c):{};return emitToolEvent(e,f,"agent.tool.result",v.toolCall?.args,{output:r.eventOutput??n}),u.observedToolIds?.add(f),void 0===r.modelOutput?o:builtinToolMessage(s,f,r.modelOutput)}catch(t){const n=function recoverableBuiltinToolError(e,t,n,r){const i=formatError(r);if("task"===t&&/repeat limit reached for tool/iu.test(i)){const t=function formatObservedEvidence(e){const t=d(e);if(0===t.length)return"";const o=t.map(e=>[`Agent: ${e.agentId}`,`Tool: ${e.toolId}`,e.output].join("\n")).join("\n\n---\n\n");return o.length>12e3?`${o.slice(0,12e3)}\n[truncated]`:o}(e);return new o({tool_call_id:n.toolCall?.id??"stable-harness-task-repeat-limit",name:"task",content:JSON.stringify({status:"delegated_task_repeat_limit",finalizationRequired:!0,instruction:"The delegated agent reached a configured tool repeat limit. Stop delegating this evidence need and do not send a synthesis task to another subagent for the same need. Finalize only from observedEvidence and other evidence that is already visible in this run. If the visible evidence is insufficient for a requested claim, report an explicit blocker or evidence gap instead of estimating, inventing, or using generic knowledge.",...t?{observedEvidence:t}:{},error:previewError(i)})})}if(/Received tool input did not match expected schema|Invalid input:/iu.test(i))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${t}-argument-error`,name:t,status:"error",content:JSON.stringify({status:"tool_argument_error",toolId:t,instruction:"The upstream builtin tool rejected these arguments. Fix the tool arguments according to the tool schema, or choose a more appropriate available tool.",error:previewError(i)})})}(e,f,s,t);if(n)return emitToolEvent(e,f,"agent.tool.result",m.toolCall?.args,{output:n.content}),n;throw emitToolEvent(e,f,"agent.tool.result",m.toolCall?.args,{error:formatError(t)}),t}}}}function builtinToolMessage(e,t,n){return new o({tool_call_id:e.toolCall?.id??`stable-harness-${t}-repeat-guard`,name:t,content:n})}export function resolveFilesystemPermissions(e,t){const o=readConfigRecord(t?.config,"builtinTools"),n=[];if(n.push(...function skillReadPermissions(e,t){const o=[...new Set((t?.skills??[]).flatMap(t=>function skillReadPaths(e,t){return t?[...new Set([t,canonicalPath(t),backendSkillPath(e,t)])].flatMap(e=>function skillReadPathCandidates(e){const t=e.replace(/\/+$/u,""),o=t.endsWith("/SKILL.md")?t.slice(0,-9):t,n=function parentPath(e){const t=e.lastIndexOf("/");return t>0?e.slice(0,t):void 0}(o);return[t,o,`${o}/**`,...n?[n,`${n}/**`]:[]]}(e)):[]}(e.workspace.root,e.workspace.skills.get(t)?.path)).filter(e=>e.startsWith("/")))];return o.length>0?[{operations:["read"],paths:o,mode:"allow"}]:[]}(e,t)),!1!==o?.filesystem){if(deepagentsMemoryWritable(e))return n.length>0?n:void 0}else n.push({operations:["read"],paths:["/memories/**"],mode:"allow"}),n.push({operations:["read","write"],paths:["/**"],mode:"deny"});return deepagentsMemoryWritable(e)||n.push({operations:["write"],paths:["/memories/**"],mode:"deny"}),n.length>0?n:void 0}function emitToolEvent(e,t,o,n,r={}){"string"==typeof r.output&&c(e,e.agent.id,t,r.output),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:o,toolId:t,..."agent.tool.start"===o?{args:n}:{},...r,..."string"==typeof r.output?u(r.output):{},...f(t,o,n)}})}function backendSkillPath(e,t){const o=function pathRelative(e,t){const o=e.split("/").filter(Boolean),n=t.split("/").filter(Boolean);if(!(n.length<o.length)){for(let e=0;e<o.length;e+=1)if(o[e]!==n[e])return;return n.slice(o.length).join("/")}}(e,t);return void 0===o?t:`/${o.split("/").filter(Boolean).join("/")}`}function canonicalPath(t){try{return e.native(t)}catch{return t}}function deepagentsMemoryWritable(e){const t=readConfigRecord(e.workspace.runtime.memory,"deepagentsMem");return!1!==t?.write}function emitInventoryRepair(e,t,o,n,r,i){e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:"inventory.repair",status:t,diagnostic:{layer:"task",owner:"stable_runtime_policy",originalId:o,repairedId:n,candidateIds:r,reason:i}}})}function taskCallCandidates(e,t){return t.map(t=>({id:t,description:e.workspace.agents.get(t)?.description,schema:{type:"object",properties:{subagent_type:{type:"string"},subagentType:{type:"string"},description:{type:"string"}},required:["description"],additionalProperties:!0}}))}function isFilesystemDisabled(e){const t=readConfigRecord(e.agent.config,"builtinTools");return!1===t?.filesystem}function isTaskVisible(e){const t=readConfigRecord(e.agent.config,"builtinTools")?.modelExposed;return!1!==t&&(!Array.isArray(t)||t.includes("task"))}function isModelVisibleBuiltin(e,t){return(!isFilesystemDisabled(e)||!function isFilesystemTool(e){return"string"==typeof e&&m.has(e)}(t))&&("task"!==t||isTaskVisible(e))}function readConfigRecord(e,t){const o=isRecord(e)?e:{};return isRecord(o[t])?o[t]:void 0}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function formatError(e){return e instanceof Error?e.message:String(e)}function previewError(e){const t=e.replace(/\s+/gu," ").trim();return t.length>800?`${t.slice(0,797)}...`:t}

package/packages/adapter-deepagents/dist/src/internal/gateway/tool-failure-events.d.ts

@@ -0,0 +1,2 @@
+import type { RuntimeAdapter } from "@stable-harness/core";
+export declare function emitStructuredToolFailure(input: Parameters<RuntimeAdapter["run"]>[0], agentId: string, toolId: string, error: unknown): void;

package/packages/adapter-deepagents/dist/src/internal/gateway/tool-failure-events.js

@@ -0,0 +1 @@
+import{toolCircuitOpenEvent as e,toolFailureEvent as o}from"@stable-harness/core";export function emitStructuredToolFailure(r,t,s,n){const i=o({requestId:r.requestId,sessionId:r.sessionId,agentId:t,toolId:s,error:n});r.emit(i),r.toolFailureTracker?.recordFailure(s)&&r.emit(e({requestId:r.requestId,sessionId:r.sessionId,agentId:t,toolId:s,reason:"runtime.tool.failure"===i.type?i.failure.reason:"unknown"}))}

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

@@ -12,6 +12,7 @@ export type ToolRepeatState = {
     maxSuccessfulCallsPerTool?: number;
     maxCallsByTool: Map<string, number>;
     maxSuccessfulCallsByTool: Map<string, number>;
+    returnPreviousOutputOnRepeatLimit: boolean;
 };
 export declare function buildGatewayTools(input: Parameters<RuntimeAdapter["run"]>[0], agentId: string, toolIds: string[], repairModel?: RuntimeToolRepairModel, repeatState?: ToolRepeatState | undefined): import("@langchain/core/tools").DynamicTool<string | ToolMessage<import("@langchain/core/messages").MessageStructure<import("@langchain/core/messages").MessageToolSet>>, unknown>[];
 export declare function createToolRepeatState(config: unknown): ToolRepeatState | undefined;

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

@@ -1 +1 @@
-import{ToolMessage as e}from"@langchain/core/messages";import{tool as t}from"@langchain/core/tools";import{isSuccessfulEvidenceOutput as o,observedToolEvidence as r,recordObservedToolEvidence as n}from"./gateway/tool-evidence.js";export function buildGatewayTools(o,n,s,a,i=createToolRepeatState(o.workspace.runtime.toolGateway)){return o.toolGateway?s.flatMap(s=>{const l=o.toolGateway?.get(s);if(!l)return[];const u=o.workspace.tools.get(s),c=u?.schema??l.schema;return[t(async t=>async function invokeGuardedGatewayTool(t){emitToolResult(t.input,t.agentId,t.toolId,void 0);const o=function missingRequiredPlanContent(e,t){const o=readRecord(e.agent.config.executionContract);if(!0!==o?.requiresPlan)return"";const n=readStringArray(o.planEvidenceTools);if(0===n.length||n.includes(t))return"";const s=new Set(r(e).map(e=>e.toolId));return n.some(e=>s.has(e))?"":["Status: plan_required",`Evidence tool: ${t}`,`Blocker: execution contract requires a planning checkpoint from one of: ${n.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}(t.input,t.toolId);if(o)return emitToolResult(t.input,t.agentId,t.toolId,o),o;const n=function missingToolDependencyContent(e,t){const o=readRecord(e.agent.config.executionContract),n=readStringArray(readRecord(o.toolDependencies)[t]);if(0===n.length)return"";const s=new Set(r(e).map(e=>e.toolId)),a=n.filter(e=>!s.has(e));return 0===a.length?"":["Status: dependency_required",`Evidence tool: ${t}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${a.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}(t.input,t.toolId);if(n)return emitToolResult(t.input,t.agentId,t.toolId,n),n;const s=t.repeatState?beforeToolInvoke(t.toolId,t.args,t.repeatState):void 0;if(s)return emitToolResult(t.input,t.agentId,t.toolId,s.eventOutput),s.modelOutput;const a=await async function invokeGatewayTool(t,o,r,n,s){try{return await t.toolGateway.invoke({toolId:r,args:n,repairModel:s,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:o,requestInput:t.request.input,observedEvidence:formatObservedEvidenceForToolContext(t)}})}catch(o){if(function isToolArgumentValidationError(e){return e instanceof Error&&"ToolArgumentValidationError"===e.name&&"string"==typeof e.toolId&&Array.isArray(e.issues)}(o))return new e({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(o)});if(t.workspace.runtime.retry?.tools?.enabled)throw o;return new e({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(o),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(t.input,t.agentId,t.toolId,t.args,t.repairModel),i=a instanceof e?String(a.content):stringifyDeepAgentResult(a.output),l=t.repeatState?afterToolInvoke(t.toolId,t.args,i,a,t.repeatState):{};return emitToolResult(t.input,t.agentId,t.toolId,l.eventOutput??i),void 0!==l.modelOutput?l.modelOutput:a instanceof e?a:i}({input:o,agentId:n,toolId:s,args:t,repairModel:a,repeatState:i}),{name:s,description:buildToolDescription(u?.description??l.description??s,c,o.workspace.runtime.toolGateway,s),schema:{type:"object",additionalProperties:!0}})]}):[]}export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool)}}function repeatGuardConfig(e){return isRecord(e)&&isRecord(e.repeatGuard)?e.repeatGuard:{}}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){return isRecord(e)?new Map(Object.entries(e).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1])):new Map}export function beforeToolInvoke(e,t,o){const r=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,r+1);const n=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==n&&r>=n){const t=repeatedToolCallLimitContent(e,o.latestSuccessfulOutputByTool.get(e));return{eventOutput:t,modelOutput:t}}const s=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==s&&(o.successfulToolCounts.get(e)??0)>=s){const t=repeatedToolCallLimitContent(e,o.latestSuccessfulOutputByTool.get(e));return{eventOutput:t,modelOutput:t}}const a=stableToolCallKey(e,t),i=o.successfulCalls.get(a);if(void 0!==i){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return{eventOutput:t,modelOutput:t}}const r=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,i);return{eventOutput:r,modelOutput:i}}}export function isToolRepeatLimitReached(e,t){if(!t)return!1;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const r=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==r&&(t.successfulToolCounts.get(e)??0)>=r}export function afterToolInvoke(t,r,n,s,a){return s instanceof e&&"error"===s.status?{}:o(n)?(a.successfulCalls.set(stableToolCallKey(t,r),n),a.latestSuccessfulOutputByTool.set(t,n),a.successfulToolCounts.set(t,(a.successfulToolCounts.get(t)??0)+1),{}):{}}function emitToolResult(e,t,o,r){void 0!==r&&n(e,t,o,r),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:t,event:void 0===r?{adapter:"deepagents",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),...toolControlProjection(r)}})}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}export function stringifyDeepAgentResult(t){if(t instanceof e)return function stringifyToolMessageContent(e){return"string"==typeof e?e:JSON.stringify(e)}(t.content);if("string"==typeof t)return t;if(isRecord(t)){const e=t.structuredResponse??t.structured_response;if(void 0!==e)return"string"==typeof e?e:JSON.stringify(e);const o=(Array.isArray(t.messages)?t.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(t.update)&&Array.isArray(t.update.messages)?t.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(t)}function buildToolDescription(e,t,o,r){const n=function toolRepeatPolicyDescription(e,t){const o=repeatGuardConfig(e),r=readPositiveIntegerMap(o.maxSuccessfulCallsByTool).get(t)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${e}\n\n${n}`:e;return t?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(t))}`:s}function previewToolOutput(e){const t=e.replace(/\s+/gu," ").trim();return t.length>500?`${t.slice(0,497)}...`:t}export function toolControlProjection(e){const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);if("string"==typeof t?.status)return{controlStatus:t.status};const o=function readTextStatus(e){return String(e).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e);return o?{controlStatus:o}:"string"==typeof t?.error?{controlStatus:t.error}:e.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function formatObservedEvidenceForToolContext(e){const t=r(e).map(e=>`Tool: ${e.toolId}\n${e.output}`).join("\n\n---\n\n");return t.length>12e3?`${t.slice(0,12e3)}\n[truncated]`:t}function formatToolArgumentError(e){return JSON.stringify({error:"tool_argument_validation_failed",toolId:e.toolId,issues:e.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function formatError(e){return e instanceof Error?e.message:String(e)}
+import{ToolMessage as e}from"@langchain/core/messages";import{tool as t}from"@langchain/core/tools";import{isSuccessfulEvidenceOutput as o,observedToolEvidence as r,recordObservedToolEvidence as n}from"./gateway/tool-evidence.js";import{emitStructuredToolFailure as s}from"./gateway/tool-failure-events.js";export function buildGatewayTools(o,n,a,i,l=createToolRepeatState(o.workspace.runtime.toolGateway)){return o.toolGateway?a.flatMap(a=>{const u=o.toolGateway?.get(a);if(!u)return[];const c=o.workspace.tools.get(a),d=c?.schema??u.schema;return[t(async t=>async function invokeGuardedGatewayTool(t){emitToolResult(t.input,t.agentId,t.toolId,void 0);const o=function missingRequiredPlanContent(e,t){const o=readRecord(e.agent.config.executionContract);if(!0!==o?.requiresPlan)return"";const n=readStringArray(o.planEvidenceTools);if(0===n.length||n.includes(t))return"";const s=new Set(r(e).map(e=>e.toolId));return n.some(e=>s.has(e))?"":["Status: plan_required",`Evidence tool: ${t}`,`Blocker: execution contract requires a planning checkpoint from one of: ${n.join(", ")} before evidence tools run.`,"Instruction: call the planning tool first, then retry this atomic evidence tool with repaired arguments."].join("\n")}(t.input,t.toolId);if(o)return emitToolResult(t.input,t.agentId,t.toolId,o),o;const n=function missingToolDependencyContent(e,t){const o=readRecord(e.agent.config.executionContract),n=readStringArray(readRecord(o.toolDependencies)[t]);if(0===n.length)return"";const s=new Set(r(e).map(e=>e.toolId)),a=n.filter(e=>!s.has(e));return 0===a.length?"":["Status: dependency_required",`Evidence tool: ${t}`,`Blocker: this atomic evidence tool requires completed dependency evidence from: ${a.join(", ")}.`,"Instruction: complete the dependency tool first, evaluate it, then retry this atomic evidence tool."].join("\n")}(t.input,t.toolId);if(n)return emitToolResult(t.input,t.agentId,t.toolId,n),n;const a=t.repeatState?beforeToolInvoke(t.toolId,t.args,t.repeatState):void 0;if(a)return emitToolResult(t.input,t.agentId,t.toolId,a.eventOutput),a.modelOutput;const i=await async function invokeGatewayTool(t,o,r,n,a){try{if(t.toolFailureTracker?.isCircuitOpen(r))throw new Error(`Tool circuit is open: ${r}`);const e=await t.toolGateway.invoke({toolId:r,args:n,repairModel:a,context:{workspaceRoot:t.workspace.root,requestId:t.requestId,sessionId:t.sessionId,agentId:o,requestInput:t.request.input,observedEvidence:formatObservedEvidenceForToolContext(t)}});return t.toolFailureTracker?.recordSuccess(r),e}catch(n){if(s(t,o,r,n),function isToolArgumentValidationError(e){return e instanceof Error&&"ToolArgumentValidationError"===e.name&&"string"==typeof e.toolId&&Array.isArray(e.issues)}(n))return new e({tool_call_id:`stable-harness-${r}-argument-guard`,name:r,status:"error",content:formatToolArgumentError(n)});if(t.workspace.runtime.retry?.tools?.enabled)throw n;return new e({tool_call_id:`stable-harness-${r}-execution-error`,name:r,status:"error",content:JSON.stringify({error:"tool_execution_failed",toolId:r,message:formatError(n),retry:"Use the error as evidence, adjust the tool arguments if possible, or return a final answer with the blocker."})})}}(t.input,t.agentId,t.toolId,t.args,t.repairModel),l=i instanceof e?String(i.content):stringifyDeepAgentResult(i.output),u=t.repeatState?afterToolInvoke(t.toolId,t.args,l,i,t.repeatState):{};return emitToolResult(t.input,t.agentId,t.toolId,u.eventOutput??l),void 0!==u.modelOutput?u.modelOutput:i instanceof e?i:l}({input:o,agentId:n,toolId:a,args:t,repairModel:i,repeatState:l}),{name:a,description:buildToolDescription(c?.description??u.description??a,d,o.workspace.runtime.toolGateway,a),schema:{type:"object",additionalProperties:!0}})]}):[]}export function createToolRepeatState(e){if(function repeatGuardEnabled(e){return!0===repeatGuardConfig(e).enabled}(e))return{successfulCalls:new Map,duplicateCallCounts:new Map,latestSuccessfulOutputByTool:new Map,successfulToolCounts:new Map,toolCallCounts:new Map,maxDuplicateCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxDuplicateCallsPerTool)??3,maxCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxCallsPerTool),maxSuccessfulCallsPerTool:readPositiveInteger(repeatGuardConfig(e).maxSuccessfulCallsPerTool),maxCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxCallsByTool),maxSuccessfulCallsByTool:readPositiveIntegerMap(repeatGuardConfig(e).maxSuccessfulCallsByTool),returnPreviousOutputOnRepeatLimit:!0===repeatGuardConfig(e).returnPreviousOutputOnRepeatLimit}}function repeatGuardConfig(e){return isRecord(e)&&isRecord(e.repeatGuard)?e.repeatGuard:{}}function readPositiveInteger(e){return"number"==typeof e&&Number.isInteger(e)&&e>0?e:void 0}function readPositiveIntegerMap(e){return isRecord(e)?new Map(Object.entries(e).map(([e,t])=>[e,readPositiveInteger(t)]).filter(e=>void 0!==e[1])):new Map}export function beforeToolInvoke(e,t,o){const r=o.toolCallCounts.get(e)??0;o.toolCallCounts.set(e,r+1);const n=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==n&&r>=n){const t=o.latestSuccessfulOutputByTool.get(e),r=repeatedToolCallLimitContent(e,t);return{eventOutput:r,modelOutput:repeatLimitModelOutput(r,t,o)}}const s=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==s&&(o.successfulToolCounts.get(e)??0)>=s){const t=o.latestSuccessfulOutputByTool.get(e),r=repeatedToolCallLimitContent(e,t);return{eventOutput:r,modelOutput:repeatLimitModelOutput(r,t,o)}}const a=stableToolCallKey(e,t),i=o.successfulCalls.get(a);if(void 0!==i){const t=o.duplicateCallCounts.get(a)??0;if(o.duplicateCallCounts.set(a,t+1),void 0!==o.maxDuplicateCallsPerTool&&t>=o.maxDuplicateCallsPerTool){const t=repeatedToolCallLimitContent(e);return{eventOutput:t,modelOutput:t}}const r=function duplicateToolCallContent(e,t){return JSON.stringify({status:"duplicate_tool_call",toolId:e,instruction:"This agent already completed an equivalent tool call. Use the prior evidence instead of calling the tool again.",previousOutput:t})}(e,i);return{eventOutput:r,modelOutput:i}}}function repeatLimitModelOutput(e,t,o){return o.returnPreviousOutputOnRepeatLimit&&void 0!==t&&0!==t.trim().length?t:e}export function isToolRepeatLimitReached(e,t){if(!t)return!1;const o=t.maxCallsByTool.get(e)??t.maxCallsPerTool;if(void 0!==o&&(t.toolCallCounts.get(e)??0)>=o)return!0;const r=t.maxSuccessfulCallsByTool.get(e)??t.maxSuccessfulCallsPerTool;return void 0!==r&&(t.successfulToolCounts.get(e)??0)>=r}export function afterToolInvoke(t,r,n,s,a){return s instanceof e&&"error"===s.status?{}:o(n)?(a.successfulCalls.set(stableToolCallKey(t,r),n),a.latestSuccessfulOutputByTool.set(t,n),a.successfulToolCounts.set(t,(a.successfulToolCounts.get(t)??0)+1),{}):{}}function emitToolResult(e,t,o,r){void 0!==r&&n(e,t,o,r),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:t,event:void 0===r?{adapter:"deepagents",phase:"agent.tool.start",toolId:o}:{adapter:"deepagents",phase:"agent.tool.result",toolId:o,output:previewToolOutput(r),...toolControlProjection(r)}})}function repeatedToolCallLimitContent(e,t){return JSON.stringify({status:"repeated_tool_call_limit",toolId:e,instruction:"This tool reached the configured repeat limit for this request. Do not call this tool or a substitute tool for the same evidence need again. Use previousOutput and the collected evidence to produce the final answer now, or report the remaining gap explicitly.",...void 0!==t?{previousOutput:t}:{}})}export function stringifyDeepAgentResult(t){if(t instanceof e)return function stringifyToolMessageContent(e){return"string"==typeof e?e:JSON.stringify(e)}(t.content);if("string"==typeof t)return t;if(isRecord(t)){const e=t.structuredResponse??t.structured_response;if(void 0!==e)return"string"==typeof e?e:JSON.stringify(e);const o=(Array.isArray(t.messages)?t.messages:[]).at(-1);if(isRecord(o)&&"string"==typeof o.content)return o.content;const r=(isRecord(t.update)&&Array.isArray(t.update.messages)?t.update.messages:[]).at(-1);if(isRecord(r)&&isRecord(r.kwargs)&&"string"==typeof r.kwargs.content)return r.kwargs.content;if(isRecord(r)&&"string"==typeof r.content)return r.content}return JSON.stringify(t)}function buildToolDescription(e,t,o,r){const n=function toolRepeatPolicyDescription(e,t){const o=repeatGuardConfig(e),r=readPositiveIntegerMap(o.maxSuccessfulCallsByTool).get(t)??readPositiveInteger(o.maxSuccessfulCallsPerTool);return void 0===r?"":`Stable runtime repeat policy: call this tool at most ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),s=n?`${e}\n\n${n}`:e;return t?`${s}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(t))}`:s}function previewToolOutput(e){const t=e.replace(/\s+/gu," ").trim();return t.length>500?`${t.slice(0,497)}...`:t}export function toolControlProjection(e){const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}(e);if("string"==typeof t?.status)return{controlStatus:t.status};const o=function readTextStatus(e){return String(e).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e);return o?{controlStatus:o}:"string"==typeof t?.error?{controlStatus:t.error}:e.startsWith("Task delegation target is not in the workspace inventory")?{controlStatus:"task_inventory_blocked"}:{}}function stableToolCallKey(e,t){return`${e}:${stableJson(t)}`}function stableJson(e){return Array.isArray(e)?`[${e.map(stableJson).join(",")}]`:isRecord(e)?`{${Object.keys(e).sort().map(t=>`${JSON.stringify(t)}:${stableJson(e[t])}`).join(",")}}`:JSON.stringify(e)}function formatObservedEvidenceForToolContext(e){const t=r(e).map(e=>`Tool: ${e.toolId}\n${e.output}`).join("\n\n---\n\n");return t.length>12e3?`${t.slice(0,12e3)}\n[truncated]`:t}function formatToolArgumentError(e){return JSON.stringify({error:"tool_argument_validation_failed",toolId:e.toolId,issues:e.issues,retry:"Call the same tool again with arguments that satisfy the reported schema and semantic issues."})}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readRecord(e){return isRecord(e)?e:{}}function readStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.length>0):[]}function formatError(e){return e instanceof Error?e.message:String(e)}

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

@@ -7,9 +7,11 @@ 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";

package/packages/core/dist/index.js

@@ -1 +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*from"./memory-plugins.js";export*from"./runtime/persistence/queue.js";export*from"./runtime.js";export*from"./runtime/selection-repair.js";export*from"./runtime/persistence/stores.js";export*from"./trace.js";export*from"./types.js";export*from"./evaluations/index.js";export*from"./quality/index.js";export*from"./spec-driven/index.js";export*from"./workflows/index.js";
+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*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/packages/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/packages/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/packages/core/dist/quality/synthesis.js

@@ -1 +1 @@
-import{controlBlockers as e,controlGaps as t,successfulEvidenceItems as n}from"./event-evidence.js";export function synthesizeEvidenceOnlyReport(s,r,i){if(!i.enabled||!i.synthesis.enabled||"evidence_only"!==i.synthesis.mode)return;if("pass"===r.verdict||!function hasRecoverableSynthesisIssue(e){return e.issues.some(e=>"control_blocker"!==e.code)}(r))return;const o=n(s.events).slice(-i.synthesis.maxEvidenceItems),c=e(s.events),l=t(s.events);if(0===o.length&&0===c.length&&0===l.length)return;const a=function detectSynthesisLanguage(e){const t=e.workspace.runtime.responseLanguage;if(isRecord(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)$/iu.test(e))return"en"}return/\p{Script=Han}/u.test(e.request.input)?"zh":"en"}(s);return"zh"===a?function buildChineseReport(e,t,n){const s=evidenceFacts(e);return["# 调查结果","","## 结论",...resultSummaryLines(e,s,t,n,"zh"),"","## 证据摘要",...summaryLines(e,t,n,"zh"),"",...structuredFactSections(s,"zh"),"","## 证据缺口与阻塞",...gapLines(t,n,"zh"),"","## 使用的证据来源",...sourceLines(sourceSummary(e),"zh")].join("\n")}(o,c,l):function buildEnglishReport(e,t,n){const s=evidenceFacts(e);return["# Result","","## Conclusion",...resultSummaryLines(e,s,t,n,"en"),"","## Evidence summary",...summaryLines(e,t,n,"en"),"",...structuredFactSections(s,"en"),"","## Evidence gaps and blockers",...gapLines(t,n,"en"),"","## Sources used",...sourceLines(sourceSummary(e),"en")].join("\n")}(o,c,l)}function resultSummaryLines(e,t,n,s,r){if(0===e.length)return["zh"===r?"- 没有成功证据可用于形成调查结论。":"- No successful evidence was available for a result."];const i=[...new Set(e.map(e=>humanSourceLabel(e.source)))],o=function observedResultLines(e,t){const value=(...t)=>function fieldValue(e,t){const n=new Set(t.map(normalizeFieldLabel));return e.find(e=>n.has(normalizeFieldLabel(e.label)))?.value}(e,t),n=value("Resolved Ticker","Symbol","Ticker"),s=value("Company Name","Company"),r=value("Close"),i=value("Open"),o=value("High"),c=value("Low"),l=value("Volume"),a=value("Date"),u=value("Inc Revenue","Revenue"),d=value("Filed 2026 03 06 Net Income","Net Income"),m=value("Filed 2026 03 06 Operating Income","Operating Income"),p=value("Filed 2026 03 06 Diluted Eps","Diluted Eps"),f=value("Workday Wday Stock Dips While Market Gains"),h=value("Summary"),g=[];if((n||s)&&g.push("zh"===t?`- 标的已确认：${s??"公司未命名"}${n?`（${n}）`:""}。`:`- Confirmed target: ${s??"unnamed company"}${n?` (${n})`:""}.`),r){const e=o&&c?"zh"===t?`，盘中区间 ${c}-${o}`:`, intraday range ${c}-${o}`:"",n=i?"zh"===t?`，开盘 ${i}`:`, open ${i}`:"",s=l?"zh"===t?`，成交量 ${l}`:`, volume ${l}`:"";g.push("zh"===t?`- 报价观察：${a?`${a} `:""}收盘 ${r}${n}${e}${s}。`:`- Quote observed: ${a?`${a} `:""}close ${r}${n}${e}${s}.`)}return(u||d||m||p)&&g.push("zh"===t?`- 财务观察：${compactParts([u&&`营收 ${u}`,d&&`净利润 ${d}`,m&&`经营利润 ${m}`,p&&`摊薄 EPS ${p}`]).join("；")}。`:`- Financials observed: ${compactParts([u&&`revenue ${u}`,d&&`net income ${d}`,m&&`operating income ${m}`,p&&`diluted EPS ${p}`]).join("; ")}.`),(f||h)&&g.push("zh"===t?`- 新闻观察：${compactParts([f,h]).join("；")}。`:`- News observed: ${compactParts([f,h]).join("; ")}.`),g.length>0?g:["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."]}(evidenceFields(t),r),c=0===n.length&&0===s.length?"zh"===r?"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"===r?`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"===r?`- 已基于 ${i.join("、")} 形成可交付结果。`:`- Result formed from ${i.join(", ")}.`,`- ${c}`,...o]}function compactParts(e){return e.filter(e=>"string"==typeof e&&e.trim().length>0)}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:classifyFact(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,i=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).`,`- ${i}`]}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=evidenceFields(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=n.filter(e=>e.kind===s.kind).slice(0,12),i=e.filter(e=>e.kind===s.kind&&!function hasExtractableFields(e){return extractKeyValueFields(e).some(e=>!isBoilerplateField(e.label)&&e.value.length>0)}(e.text)).slice(0,5);return 0===r.length&&0===i.length?[]:[s.title,...fieldTableLines(r,t),...i.map(e=>function factLine(e,t){const n=humanSourceLabel(e.source);return"zh"===t?`- ${e.text}（来源：${n}）`:`- ${e.text} (source: ${n})`}(e,t)),""]}).filter((e,t,n)=>""!==e||t<n.length-1)}function evidenceFields(e){const t=new Set,n=[];for(const s of e)for(const e of parseEvidenceFields(s)){const s=`${normalizeFieldLabel(e.label)}\0${e.value.toLowerCase()}`;t.has(s)||(t.add(s),n.push(e))}return n}function parseEvidenceFields(e){return extractKeyValueFields(e.text).filter(e=>e.value.length>0&&!isBoilerplateField(e.label)).map(t=>{return{source:e.source,label:(n=t.label,/\p{Script=Han}/u.test(n)?n:normalizeFieldLabel(n).split(" ").filter(Boolean).map(e=>/^[A-Z0-9]+$/u.test(e)?e:e.charAt(0).toUpperCase()+e.slice(1)).join(" ")),value:formatEvidence(t.value),kind:"other"===classifyFact(`${t.label}: ${t.value}`)?e.kind:classifyFact(`${t.label}: ${t.value}`)};var n})}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(humanSourceLabel(e.source))} |`)]}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 isRecord(t)?Object.entries(t).filter(([e])=>!/^(?:status|controlStatus)$/iu.test(e)).slice(0,8).map(([e,t])=>`${e}: ${formatEvidence(function stringifyJsonValue(e){return"string"==typeof e?e:JSON.stringify(e)}(t))}`):[]}catch{return[]}}(e);return t.length>0?t:function splitPlainTextFacts(e){const t=function stripControlPreamble(e){return e.replace(/^(?:Status:\s*(?:completed|success|ok|recorded)\b\.?\s*)+/iu,"").replace(/^(?:Evidence tool:\s*[A-Za-z0-9_.-]+\b\.?\s*)+/iu,"").trim()}(e);return t.split(/\r?\n|;\s*/u).map(e=>e.trim()).map(stripControlFragments).filter(e=>e&&!function isControlFact(e){return/^Status:\s*(?:completed|success|ok|recorded)$/iu.test(e)||/^Evidence tool:\s*[A-Za-z0-9_.-]+$/iu.test(e)}(e))}(e).slice(0,8).map(formatEvidence)}function extractKeyValueFields(e){const t=[...e.matchAll(/(^|[\s,;|])([\p{L}][\p{L}\p{N} _./()%&+\-]{0,48})[:：]\s*/gu)].filter(e=>void 0!==e.index);return t.map((n,s)=>{const r=(n.index??0)+n[0].length,i=t[s+1]?.index??e.length,o=t[s-1]?.[2]??"";return{label:cleanFieldLabel(n[2]??"",o),value:(c=e.slice(r,i),c.replace(/^[\s,;|.-]+/u,"").replace(/[\s,;|.-]+$/u,"").replace(/\s+/gu," ").trim())};var c}).filter(e=>e.label.length>0)}function cleanFieldLabel(e,t=""){const n=e.replace(/^(?:[-*]\s*)+/u,"").replace(/\s+/gu," ").trim(),s=n.split(" ").filter(Boolean);return s.length>1&&isBoilerplateField(t)||s.length>1&&/^[A-Z]$/u.test(s[0]??"")?s.slice(1).join(" "):n}function isBoilerplateField(e){return/^(?:status|control status|evidence tool|tool|source|来源)$/iu.test(e.trim())}function normalizeFieldLabel(e){return e.toLowerCase().replace(/[^a-z0-9\p{L}\p{N}]+/giu," ").trim()}function escapeTableCell(e){return e.replace(/\|/gu,"\\|").replace(/\r?\n/gu," ")}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 evidenceTextForSource(e,t){return e.find(e=>e.source===t)?.output??""}function fallbackEvidenceText(e){return formatEvidence(e)}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 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(" ")}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}
+import{controlBlockers as e,controlGaps as t,successfulEvidenceItems as n}from"./event-evidence.js";import{detectSynthesisLanguage as s}from"./synthesis/language.js";export function synthesizeEvidenceOnlyReport(r,i,o){if(!o.enabled||!o.synthesis.enabled||"evidence_only"!==o.synthesis.mode)return;if("pass"===i.verdict||!function hasRecoverableSynthesisIssue(e){return e.issues.some(e=>"control_blocker"!==e.code)}(i))return;const l=n(r.events).slice(-o.synthesis.maxEvidenceItems),c=e(r.events),a=t(r.events);if(0===l.length&&0===c.length&&0===a.length)return;const u=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)).at(-1)?.output.trim();return t||void 0}(l);return u||("zh"===s(r)?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")}(l,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")}(l,c,a))}function resultSummaryLines(e,t,n,s,r){if(0===e.length)return["zh"===r?"- 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)))],o=function observedResultLines(e,t){const value=(...t)=>function fieldValue(e,t){const n=new Set(t.map(normalizeFieldLabel));return e.find(e=>n.has(normalizeFieldLabel(e.label)))?.value}(e,t),n=value("Resolved Ticker","Symbol","Ticker"),s=value("Company Name","Company"),r=value("Close"),i=value("Open"),o=value("High"),l=value("Low"),c=value("Volume"),a=value("Date"),u=value("Inc Revenue","Revenue"),d=value("Filed 2026 03 06 Net Income","Net Income"),m=value("Filed 2026 03 06 Operating Income","Operating Income"),p=value("Filed 2026 03 06 Diluted Eps","Diluted Eps"),f=value("Workday Wday Stock Dips While Market Gains"),h=value("Summary"),v=[];if((n||s)&&v.push("zh"===t?`- 标的已确认：${s??"公司未命名"}${n?`（${n}）`:""}。`:`- Confirmed target: ${s??"unnamed company"}${n?` (${n})`:""}.`),r){const e=o&&l?"zh"===t?`，盘中区间 ${l}-${o}`:`, intraday range ${l}-${o}`:"",n=i?"zh"===t?`，开盘 ${i}`:`, open ${i}`:"",s=c?"zh"===t?`，成交量 ${c}`:`, volume ${c}`:"";v.push("zh"===t?`- 报价观察：${a?`${a} `:""}收盘 ${r}${n}${e}${s}。`:`- Quote observed: ${a?`${a} `:""}close ${r}${n}${e}${s}.`)}return(u||d||m||p)&&v.push("zh"===t?`- 财务观察：${compactParts([u&&`营收 ${u}`,d&&`净利润 ${d}`,m&&`经营利润 ${m}`,p&&`摊薄 EPS ${p}`]).join("；")}。`:`- Financials observed: ${compactParts([u&&`revenue ${u}`,d&&`net income ${d}`,m&&`operating income ${m}`,p&&`diluted EPS ${p}`]).join("; ")}.`),(f||h)&&v.push("zh"===t?`- 新闻观察：${compactParts([f,h]).join("；")}。`:`- News observed: ${compactParts([f,h]).join("; ")}.`),v.length>0?v:["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."]}(evidenceFields(t),r),l=0===n.length&&0===s.length?"zh"===r?"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"===r?`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"===r?`- 已基于 ${i.join("、")} 形成可交付结果。`:`- Result formed from ${i.join(", ")}.`,`- ${l}`,...o]}function compactParts(e){return e.filter(e=>"string"==typeof e&&e.trim().length>0)}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:classifyFact(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,i=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).`,`- ${i}`]}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=evidenceFields(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=n.filter(e=>e.kind===s.kind).slice(0,12),i=e.filter(e=>e.kind===s.kind&&!function hasExtractableFields(e){return extractKeyValueFields(e).some(e=>!isBoilerplateField(e.label)&&e.value.length>0)}(e.text)).slice(0,5);return 0===r.length&&0===i.length?[]:[s.title,...fieldTableLines(r,t),...i.map(e=>function factLine(e,t){const n=humanSourceLabel(e.source);return"zh"===t?`- ${e.text}（来源：${n}）`:`- ${e.text} (source: ${n})`}(e,t)),""]}).filter((e,t,n)=>""!==e||t<n.length-1)}function evidenceFields(e){const t=new Set,n=[];for(const s of e)for(const e of parseEvidenceFields(s)){const s=`${normalizeFieldLabel(e.label)}\0${e.value.toLowerCase()}`;t.has(s)||(t.add(s),n.push(e))}return n}function parseEvidenceFields(e){return extractKeyValueFields(e.text).filter(e=>e.value.length>0&&!isBoilerplateField(e.label)).map(t=>{return{source:e.source,label:(n=t.label,/\p{Script=Han}/u.test(n)?n:normalizeFieldLabel(n).split(" ").filter(Boolean).map(e=>/^[A-Z0-9]+$/u.test(e)?e:e.charAt(0).toUpperCase()+e.slice(1)).join(" ")),value:formatEvidence(t.value),kind:"other"===classifyFact(`${t.label}: ${t.value}`)?e.kind:classifyFact(`${t.label}: ${t.value}`)};var n})}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(humanSourceLabel(e.source))} |`)]}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}: ${formatEvidence(function stringifyJsonValue(e){return"string"==typeof e?e:JSON.stringify(e)}(t))}`):[]}catch{return[]}}(e);return t.length>0?t:function splitPlainTextFacts(e){const t=function stripControlPreamble(e){return e.replace(/^(?:Status:\s*(?:completed|success|ok|recorded)\b\.?\s*)+/iu,"").replace(/^(?:Evidence tool:\s*[A-Za-z0-9_.-]+\b\.?\s*)+/iu,"").trim()}(e);return t.split(/\r?\n|;\s*/u).map(e=>e.trim()).map(stripControlFragments).filter(e=>e&&!function isControlFact(e){return/^Status:\s*(?:completed|success|ok|recorded)$/iu.test(e)||/^Evidence tool:\s*[A-Za-z0-9_.-]+$/iu.test(e)}(e))}(e).slice(0,8).map(formatEvidence)}function extractKeyValueFields(e){const t=[...e.matchAll(/(^|[\s,;|])([\p{L}][\p{L}\p{N} _./()%&+\-]{0,48})[:：]\s*/gu)].filter(e=>void 0!==e.index);return t.map((n,s)=>{const r=(n.index??0)+n[0].length,i=t[s+1]?.index??e.length,o=t[s-1]?.[2]??"";return{label:cleanFieldLabel(n[2]??"",o),value:(l=e.slice(r,i),l.replace(/^[\s,;|.-]+/u,"").replace(/[\s,;|.-]+$/u,"").replace(/\s+/gu," ").trim())};var l}).filter(e=>e.label.length>0)}function cleanFieldLabel(e,t=""){const n=e.replace(/^(?:[-*]\s*)+/u,"").replace(/\s+/gu," ").trim(),s=n.split(" ").filter(Boolean);return s.length>1&&isBoilerplateField(t)||s.length>1&&/^[A-Z]$/u.test(s[0]??"")?s.slice(1).join(" "):n}function isBoilerplateField(e){return/^(?:status|control status|evidence tool|tool|source|来源)$/iu.test(e.trim())}function normalizeFieldLabel(e){return e.toLowerCase().replace(/[^a-z0-9\p{L}\p{N}]+/giu," ").trim()}function escapeTableCell(e){return e.replace(/\|/gu,"\\|").replace(/\r?\n/gu," ")}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 evidenceTextForSource(e,t){return e.find(e=>e.source===t)?.output??""}function fallbackEvidenceText(e){return formatEvidence(e)}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 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/packages/core/dist/recovery/tool-call.js

@@ -1 +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: ${previewOutput(e)}`)}export function rawToolCallFailureMessage(){return["The model attempted to call a tool but returned the tool call as text instead of executing it.","Please retry the request or use a model/backend configuration with reliable tool calling for this workspace."].join(" ")}export function buildEvidenceSynthesisOutput(e){const t=readToolCallRecovery(e.policy);if(!0!==t.enabled||!1===t.synthesizeFromEvidenceOnFailure||!containsRecoverableResultFailureText(e.output,t))return;const o=recentToolEvidence(e.events,6e3);return 0!==o.length?"zh"==(/\p{Script=Han}/u.test(e.request.input)?"zh":"en")?function buildChineseEvidenceSynthesis(e,t){return["上游模型在已有工具证据后仍输出了伪工具调用；runtime 已拒绝该 raw 输出，并直接交付已执行工具返回的证据结果。","","已执行的工具证据：",...t,"",`被拒绝的最终输出预览：${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)}
+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)}

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

@@ -1,4 +1,4 @@
-import type { CompiledWorkspace, RuntimeEvent, RuntimeOutput, RuntimeRequest, RuntimeToolGateway, WorkspaceAgent } from "../types.js";
+import type { CompiledWorkspace, RuntimeEvent, RuntimeOutput, RuntimeRequest, RuntimeToolFailureTracker, RuntimeToolGateway, WorkspaceAgent } from "../types.js";
 export declare function runDirectToolCall(input: {
     gateway: RuntimeToolGateway | undefined;
     workspace: CompiledWorkspace;
@@ -7,4 +7,5 @@ export declare function runDirectToolCall(input: {
     requestId: string;
     sessionId: string;
     agent: WorkspaceAgent;
+    toolFailureTracker?: RuntimeToolFailureTracker;
 }): Promise<RuntimeOutput>;

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

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

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

@@ -1,6 +1,7 @@
 import type { MemoryCandidate, MemoryDecision, MemoryRecord } from "@stable-harness/memory";
 import type { ApprovalRequest } from "@stable-harness/governance";
 import type { RuntimeArtifact } from "../types.js";
+import type { RuntimeToolFailureClassification, RuntimeToolFailureReason } from "./tool-failure.js";
 export type RuntimeInventoryRepairLayer = "agent" | "workflow_route" | "workflow" | "tool" | "task" | "skill";
 export type RuntimeInventoryRepairStatus = "repaired" | "blocked";
 export type RuntimeInventoryRepairDiagnostic = {
@@ -70,6 +71,21 @@ export type RuntimeEvent = {
     agentId: string;
     toolId: string;
     output: unknown;
+} | {
+    type: "runtime.tool.failure";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    toolId: string;
+    error: unknown;
+    failure: RuntimeToolFailureClassification;
+} | {
+    type: "runtime.tool.circuit.opened";
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    toolId: string;
+    reason: RuntimeToolFailureReason;
 } | {
     type: "runtime.workflow.started";
     requestId: string;

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

@@ -0,0 +1,41 @@
+import type { RuntimeEvent } from "./events.js";
+export type RuntimeToolFailureKind = "transient" | "business" | "control" | "unknown";
+export type RuntimeToolFailureReason = "timeout" | "network" | "rate_limit" | "server_error" | "invalid_input" | "not_found" | "permission_denied" | "approval_required" | "schema_repair_failed" | "repeat_limit" | "tool_execution_failed" | "unknown";
+export type RuntimeToolFailureClassification = {
+    kind: RuntimeToolFailureKind;
+    reason: RuntimeToolFailureReason;
+    retryable: boolean;
+    message: string;
+};
+export type RuntimeToolFailurePolicy = {
+    enabled?: boolean;
+    circuitBreaker?: {
+        enabled?: boolean;
+        failureThreshold?: number;
+    };
+};
+export type RuntimeToolFailureTracker = {
+    isCircuitOpen(toolId: string): boolean;
+    recordFailure(toolId: string): boolean;
+    recordSuccess(toolId: string): void;
+};
+export declare function createToolFailureTracker(policy: RuntimeToolFailurePolicy | undefined): RuntimeToolFailureTracker;
+export declare function classifyToolFailure(error: unknown): RuntimeToolFailureClassification;
+export declare function toolFailureEvent(input: {
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    toolId: string;
+    error: unknown;
+}): Extract<RuntimeEvent, {
+    type: "runtime.tool.failure";
+}>;
+export declare function toolCircuitOpenEvent(input: {
+    requestId: string;
+    sessionId: string;
+    agentId: string;
+    toolId: string;
+    reason: RuntimeToolFailureReason;
+}): Extract<RuntimeEvent, {
+    type: "runtime.tool.circuit.opened";
+}>;

package/packages/core/dist/runtime/tool-failure.js

@@ -0,0 +1 @@
+export function createToolFailureTracker(e){const t=function readCircuitThreshold(e){const t=e?.circuitBreaker,i=t?.failureThreshold;return!1!==e?.enabled&&!0===t?.enabled&&"number"==typeof i&&Number.isInteger(i)&&i>0?i:void 0}(e),i=new Map,r=new Set;return{isCircuitOpen:e=>r.has(e),recordFailure(e){if(void 0===t)return!1;const n=(i.get(e)??0)+1;return i.set(e,n),n>=t&&(r.add(e),!0)},recordSuccess(e){i.delete(e),r.delete(e)}}}export function classifyToolFailure(e){const t=function formatToolFailureMessage(e){return e instanceof Error?e.message:String(e)}(e),i=t.toLowerCase();return/\b(timeout|timed out|etimedout)\b/u.test(i)?classification("transient","timeout",!0,t):/\b(econnreset|econnrefused|enotfound|network|fetch failed)\b/u.test(i)?classification("transient","network",!0,t):/\b(rate.?limit|too many requests|429)\b/u.test(i)?classification("transient","rate_limit",!0,t):/\b(500|502|503|504|server error|bad gateway)\b/u.test(i)?classification("transient","server_error",!0,t):/\b(approval_required|requires approval|permission denied|forbidden|401|403)\b/u.test(i)?classification("business",i.includes("approval")?"approval_required":"permission_denied",!1,t):/\b(schema repair failed|tool_argument_validation_failed|invalid input|expected .* received|validation)\b/u.test(i)?classification("control",i.includes("schema repair")?"schema_repair_failed":"invalid_input",!1,t):/\b(repeated_tool_call_limit|repeat limit)\b/u.test(i)?classification("control","repeat_limit",!1,t):/\b(not found|404)\b/u.test(i)?classification("business","not_found",!1,t):classification("unknown","tool_execution_failed",!1,t)}export function toolFailureEvent(e){const t=classifyToolFailure(e.error);return{type:"runtime.tool.failure",...e,failure:t}}export function toolCircuitOpenEvent(e){return{type:"runtime.tool.circuit.opened",...e}}function classification(e,t,i,r){return{kind:e,reason:t,retryable:i,message:r}}

package/packages/core/dist/runtime.js

@@ -1 +1 @@
-import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{projectRequestInspection as r,projectRequestSummary as s,projectSessionSummaries as o}from"./runtime/persistence/inspection.js";import{assertNoRawToolCallOutput as a,assertNoToolExecutionErrorOutput as n,buildAdapterErrorRecoveryPrompt as i,buildEvidenceSynthesisOutput as u,buildExecutionContractRecoveryRequest as c,buildResultRecoveryRequest as l,containsRawToolCallOutput as p,isRecoverableAdapterError as d,rawToolCallFailureMessage as m,toolCallRecoveryEnabled as w}from"./recovery/tool-call.js";import{recoverQualityReview as f,resolveQualityPolicy as y}from"./quality/index.js";import{completeRun as R,failRun as g}from"./runtime/completion.js";import{runDirectToolCall as I}from"./runtime/direct-tool-call.js";import{createRuntimeCapabilityRegistry as q,normalizeAdapterResult as k}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as v}from"./runtime/memory.js";import{createInMemoryRuntimeStore as b}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as A}from"./runtime/progress-narration.js";import{repairRuntimeSelection as C}from"./runtime/selection-repair.js";import{runWorkflowRequest as x}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const a=new Set,n=t.store??b(),u=q([v(t),A({options:t.progressNarration,policy:t.workspace.runtime}),...t.capabilities??[]]),emitBase=e=>{n.appendEvent(e);for(const t of a)t(e)},emit=e=>{emitBase(e),u.emitSideEffects(e,emitBase)};return{request:async r=>async function runRuntimeRequest(t){const r=t.request.requestId??e(),s=t.request.sessionId??e(),o=[],{agent:a,adapter:n}=await async function resolveExecution(e,t,r){const s=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const s=await C({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description})),trace:{...r,agentId:t,layer:"agent",owner:"stable_runtime_policy"}});return s.ok?s.id:t}(e.workspace,t.agentId,r):function resolveRoutedAgentId(e,t){for(const r of e.runtime.routes??[])if(routeMatches(r,t))return r.agentId;return e.runtime.defaultAgentId}(e.workspace,t.input),o=e.workspace.agents.get(s);if(!o)throw new Error(`Agent ${s} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:o,adapter:void 0};const a=e.adapters.find(e=>e.canRun(o));if(!a)throw new Error(`No runtime adapter can run backend ${o.backend} for agent ${o.id}`);return{agent:o,adapter:a}}(t.input,t.request,{requestId:r,sessionId:s,emit:e=>o.push(e)});t.store.createRun(function createRunRecord(e,t,r,s){return{requestId:t,sessionId:r,agentId:s.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,r,s,a)),o.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:r,sessionId:s,agentId:a.id,input:t.request.input});try{if(t.request.workflow){const e=await x({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:r,sessionId:s,agentId:a.id,emit:t.emit});return R({store:t.store,emit:t.emit,requestId:r,sessionId:s,agent:a,result:e})}if(t.request.toolCall){const e=await I({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:r,sessionId:s,agent:a});return R({store:t.store,emit:t.emit,requestId:r,sessionId:s,agent:a,result:e})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,r=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),s=r.memory,o=r.pluginMemories??[],a=e.input.workspace.runtime,n=y(e.input.workspace.runtime,e.agent),u=new Map;let l;try{l=await runAdapterOnce(e,t,e.request,s,o,u)}catch(r){if(!d(r,a))throw r;l=await runAdapterOnce(e,t,i(e.request,r,a),s,o,u)}l=await recoverAdapterResultOutput(e,t,e.request,l,s,o,a,u),l=await f(createQualityRuntimeInput(e,s,o,u),e.request,l,n),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:l});try{assertRequestExecutionContract(e)}catch(r){const i=c({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:a});if(!i)throw r;l=await runAdapterOnce(e,t,i,s,o,u),l=await recoverAdapterResultOutput(e,t,i,l,s,o,a,u),l=await f(createQualityRuntimeInput(e,s,o,u),i,l,n),assertRequestExecutionContract(e)}const p=R({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:l});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:l,response:p}),p}({...t,adapter:n,requestId:r,sessionId:s,agent:a})}catch(e){return g({store:t.store,emit:t.emit,requestId:r,sessionId:s,agent:a,error:e})}}({input:t,capabilities:u,store:n,emit:emit,request:r}),subscribe:e=>(a.add(e),()=>a.delete(e)),inspect(){return{workspaceRoot:t.workspace.root,agents:[...t.workspace.agents.keys()].sort(),workflows:[...t.workspace.workflows.keys()].sort(),...t.workspace.runtime.specDrivenWorkflow?{specDrivenWorkflow:(e=t.workspace.runtime.specDrivenWorkflow,{enabled:e.enabled,artifactsDir:e.artifactsDir,...e.constitution?{constitution:e.constitution}:{},phases:e.phases.map(e=>e.id)})}:{},evaluations:[...t.workspace.evaluations?.keys()??[]].sort(),...t.workspace.runtime.workflowRouting?.defaultWorkflowId?{defaultWorkflowId:t.workspace.runtime.workflowRouting.defaultWorkflowId}:{},workflowRoutes:(t.workspace.runtime.workflowRouting?.routes??[]).map(e=>e.id).sort(),models:[...t.workspace.models.keys()].sort(),tools:[...t.workspace.tools.keys()].sort(),runs:n.listRuns()};var e},getRuntimePolicy:()=>t.workspace.runtime,getWorkflow:e=>t.workspace.workflows.get(e),getRun:e=>n.getRun(e),listRequests:e=>n.listRuns(e).map(s),listSessions:()=>o(n.listRuns()),inspectRequest(e){const s=n.getRun(e);return s?r(t.workspace,s):void 0},cancel(e,t){const r=n.getRun(e);r&&"running"===r.state&&(n.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await u.stop(),a.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:o=>runAdapterOnce(e,e.adapter,o,t,r,s),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}async function recoverAdapterResultOutput(e,t,r,s,o,i,c,d){let f=r;const y=function resultRecoveryAttempts(e){const t="object"!=typeof e||null===e||Array.isArray(e)?void 0:e.recovery,r="object"!=typeof t||null===t||Array.isArray(t)?void 0:t.toolCall,s="object"!=typeof r||null===r||Array.isArray(r)?void 0:r.maxResultRecoveryAttempts;return"number"==typeof s&&Number.isInteger(s)&&s>0?s:3}(c);let R=0;for(let r=0;r<y;r+=1){const r=e.store.getRun(e.requestId)?.events??[],a=l({request:f,output:s.text,events:r.slice(R),availableToolIds:e.agent.tools,policy:c});if(!a)break;f=a,R=e.store.getRun(e.requestId)?.events.length??0,s=await runAdapterOnce(e,t,a,o,i,d)}if(w(c)){let t=!1;p(s.text,c)&&function rawToolCallFailureReturnsMessage(e){return"message"===("object"!=typeof e?.toolCallRecovery||null===e.toolCallRecovery||Array.isArray(e.toolCallRecovery)?{}:e.toolCallRecovery).onFailure}(r.metadata)&&(s={...s,text:m(),metadata:{...s.metadata,toolCallRecovery:{failed:!0,reason:"raw_tool_call_output"}}});const o=u({request:r,output:s.text,events:e.store.getRun(e.requestId)?.events??[],policy:c});o&&(t=!0,s={...s,text:o,metadata:{...s.metadata,toolCallRecovery:{synthesized:!0,reason:"raw_tool_call_output_with_evidence"}}}),t||(a(s.text,c),n(s.text,c))}return s}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,s,o,a){return k(await t.run({workspace:e.input.workspace,agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:s,pluginMemories:o,toolGateway:e.input.toolGateway,requestState:a,emit:e.emit}))}function routeMatches(e,t){if(e.pattern)try{if(new RegExp(e.pattern,"iu").test(t))return!0}catch{return!1}const r=t.toLowerCase();return(e.keywords??[]).some(e=>r.includes(e.toLowerCase()))}
+import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{projectRequestInspection as r,projectRequestSummary as s,projectSessionSummaries as o}from"./runtime/persistence/inspection.js";import{assertNoRawToolCallOutput as a,assertNoToolExecutionErrorOutput as n,buildAdapterErrorRecoveryPrompt as i,buildEvidenceSynthesisOutput as u,buildExecutionContractRecoveryRequest as c,buildResultRecoveryRequest as l,containsRawToolCallOutput as p,isRecoverableAdapterError as d,rawToolCallFailureMessage as m,toolCallRecoveryEnabled as w}from"./recovery/tool-call.js";import{recoverQualityReview as f,resolveQualityPolicy as y}from"./quality/index.js";import{completeRun as R,failRun as g}from"./runtime/completion.js";import{runDirectToolCall as I}from"./runtime/direct-tool-call.js";import{createRuntimeCapabilityRegistry as q,normalizeAdapterResult as k}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as v}from"./runtime/memory.js";import{createInMemoryRuntimeStore as b}from"./runtime/persistence/stores.js";import{createProgressNarrationCapability as A}from"./runtime/progress-narration.js";import{repairRuntimeSelection as C}from"./runtime/selection-repair.js";import{createToolFailureTracker as x}from"./runtime/tool-failure.js";import{runWorkflowRequest as j}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const a=new Set,n=t.store??b(),u=q([v(t),A({options:t.progressNarration,policy:t.workspace.runtime}),...t.capabilities??[]]),emitBase=e=>{n.appendEvent(e);for(const t of a)t(e)},emit=e=>{emitBase(e),u.emitSideEffects(e,emitBase)},l=x(function readToolFailurePolicy(e){if("object"!=typeof e||null===e||Array.isArray(e))return;const t=e.failurePolicy;return"object"!=typeof t||null===t||Array.isArray(t)?void 0:t}(t.workspace.runtime.toolGateway));return{request:async r=>async function runRuntimeRequest(t){const r=t.request.requestId??e(),s=t.request.sessionId??e(),o=[],{agent:a,adapter:n}=await async function resolveExecution(e,t,r){const s=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const s=await C({id:t,candidates:[...e.agents.values()].map(e=>({id:e.id,description:e.description})),trace:{...r,agentId:t,layer:"agent",owner:"stable_runtime_policy"}});return s.ok?s.id:t}(e.workspace,t.agentId,r):function resolveRoutedAgentId(e,t){for(const r of e.runtime.routes??[])if(routeMatches(r,t))return r.agentId;return e.runtime.defaultAgentId}(e.workspace,t.input),o=e.workspace.agents.get(s);if(!o)throw new Error(`Agent ${s} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:o,adapter:void 0};const a=e.adapters.find(e=>e.canRun(o));if(!a)throw new Error(`No runtime adapter can run backend ${o.backend} for agent ${o.id}`);return{agent:o,adapter:a}}(t.input,t.request,{requestId:r,sessionId:s,emit:e=>o.push(e)});t.store.createRun(function createRunRecord(e,t,r,s){return{requestId:t,sessionId:r,agentId:s.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,r,s,a)),o.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:r,sessionId:s,agentId:a.id,input:t.request.input});try{if(t.request.workflow){const e=await j({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:r,sessionId:s,agentId:a.id,emit:t.emit});return R({store:t.store,emit:t.emit,requestId:r,sessionId:s,agent:a,result:e})}if(t.request.toolCall){const e=await I({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:r,sessionId:s,agent:a,toolFailureTracker:t.toolFailureTracker});return R({store:t.store,emit:t.emit,requestId:r,sessionId:s,agent:a,result:e})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,r=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),s=r.memory,o=r.pluginMemories??[],a=e.input.workspace.runtime,n=y(e.input.workspace.runtime,e.agent),u=new Map;let l;try{l=await runAdapterOnce(e,t,e.request,s,o,u)}catch(r){if(!d(r,a))throw r;l=await runAdapterOnce(e,t,i(e.request,r,a),s,o,u)}l=await recoverAdapterResultOutput(e,t,e.request,l,s,o,a,u),l=await f(createQualityRuntimeInput(e,s,o,u),e.request,l,n),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:l});try{assertRequestExecutionContract(e)}catch(r){const i=c({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:a});if(!i)throw r;l=await runAdapterOnce(e,t,i,s,o,u),l=await recoverAdapterResultOutput(e,t,i,l,s,o,a,u),l=await f(createQualityRuntimeInput(e,s,o,u),i,l,n),assertRequestExecutionContract(e)}const p=R({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:l});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:l,response:p}),p}({...t,adapter:n,requestId:r,sessionId:s,agent:a})}catch(e){return g({store:t.store,emit:t.emit,requestId:r,sessionId:s,agent:a,error:e})}}({input:t,capabilities:u,store:n,emit:emit,request:r,toolFailureTracker:l}),subscribe:e=>(a.add(e),()=>a.delete(e)),inspect(){return{workspaceRoot:t.workspace.root,agents:[...t.workspace.agents.keys()].sort(),workflows:[...t.workspace.workflows.keys()].sort(),...t.workspace.runtime.specDrivenWorkflow?{specDrivenWorkflow:(e=t.workspace.runtime.specDrivenWorkflow,{enabled:e.enabled,artifactsDir:e.artifactsDir,...e.constitution?{constitution:e.constitution}:{},phases:e.phases.map(e=>e.id)})}:{},evaluations:[...t.workspace.evaluations?.keys()??[]].sort(),...t.workspace.runtime.workflowRouting?.defaultWorkflowId?{defaultWorkflowId:t.workspace.runtime.workflowRouting.defaultWorkflowId}:{},workflowRoutes:(t.workspace.runtime.workflowRouting?.routes??[]).map(e=>e.id).sort(),models:[...t.workspace.models.keys()].sort(),tools:[...t.workspace.tools.keys()].sort(),runs:n.listRuns()};var e},getRuntimePolicy:()=>t.workspace.runtime,getWorkflow:e=>t.workspace.workflows.get(e),getRun:e=>n.getRun(e),listRequests:e=>n.listRuns(e).map(s),listSessions:()=>o(n.listRuns()),inspectRequest(e){const s=n.getRun(e);return s?r(t.workspace,s):void 0},cancel(e,t){const r=n.getRun(e);r&&"running"===r.state&&(n.updateRun(e,{state:"cancelled",completedAt:(new Date).toISOString()}),emit({type:"runtime.request.cancelled",requestId:e,sessionId:r.sessionId,agentId:r.agentId,reason:t}))},async stop(){await u.stop(),a.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,s){return{workspace:e.input.workspace,agent:e.agent,request:e.request,requestId:e.requestId,sessionId:e.sessionId,events:e.store.getRun(e.requestId)?.events??[],emit:e.emit,getEvents:()=>e.store.getRun(e.requestId)?.events??[],runAdapter:o=>runAdapterOnce(e,e.adapter,o,t,r,s),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}async function recoverAdapterResultOutput(e,t,r,s,o,i,c,d){let f=r;const y=function resultRecoveryAttempts(e){const t="object"!=typeof e||null===e||Array.isArray(e)?void 0:e.recovery,r="object"!=typeof t||null===t||Array.isArray(t)?void 0:t.toolCall,s="object"!=typeof r||null===r||Array.isArray(r)?void 0:r.maxResultRecoveryAttempts;return"number"==typeof s&&Number.isInteger(s)&&s>0?s:3}(c);let R=0;for(let r=0;r<y;r+=1){const r=e.store.getRun(e.requestId)?.events??[],a=l({request:f,output:s.text,events:r.slice(R),availableToolIds:e.agent.tools,policy:c});if(!a)break;f=a,R=e.store.getRun(e.requestId)?.events.length??0,s=await runAdapterOnce(e,t,a,o,i,d)}if(w(c)){let t=!1;p(s.text,c)&&function rawToolCallFailureReturnsMessage(e){return"message"===("object"!=typeof e?.toolCallRecovery||null===e.toolCallRecovery||Array.isArray(e.toolCallRecovery)?{}:e.toolCallRecovery).onFailure}(r.metadata)&&(s={...s,text:m(),metadata:{...s.metadata,toolCallRecovery:{failed:!0,reason:"raw_tool_call_output"}}});const o=u({request:r,output:s.text,events:e.store.getRun(e.requestId)?.events??[],policy:c});o&&(t=!0,s={...s,text:o,metadata:{...s.metadata,toolCallRecovery:{synthesized:!0,reason:"raw_tool_call_output_with_evidence"}}}),t||(a(s.text,c),n(s.text,c))}return s}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,s,o,a){return k(await t.run({workspace:e.input.workspace,agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:s,pluginMemories:o,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,requestState:a,emit:e.emit}))}function routeMatches(e,t){if(e.pattern)try{if(new RegExp(e.pattern,"iu").test(t))return!0}catch{return!1}const r=t.toLowerCase();return(e.keywords??[]).some(e=>r.includes(e.toLowerCase()))}

package/packages/core/dist/trace.js

@@ -1 +1 @@
-export function projectRuntimeTrace(e){return e.events.map(projectEvent).filter(isTraceEntry)}export function projectEvent(e){return"runtime.request.started"===e.type||"runtime.request.completed"===e.type||"runtime.request.failed"===e.type?base(e,"request",e.type):"runtime.request.cancelled"===e.type?base(e,"request",e.type,{reason:e.reason}):"runtime.execution.contract.failed"===e.type?base(e,"request",e.type,{reason:e.reason,missingEvidenceTools:e.missingEvidenceTools}):"runtime.inventory.repair"===e.type?base(e,"adapter","runtime.inventory.repair",{status:e.status,...e.diagnostic}):"runtime.tool.direct.started"===e.type?base(e,"tool","runtime.tool.direct.started",{toolId:e.toolId}):"runtime.tool.direct.completed"===e.type?base(e,"tool","runtime.tool.direct.completed",{toolId:e.toolId}):"runtime.workflow.started"===e.type||"runtime.workflow.completed"===e.type?base(e,"workflow",e.type,{workflowId:e.workflowId,adapter:e.adapter}):function isSpecDrivenPhaseEvent(e){return e.type.startsWith("runtime.specDriven.phase.")}(e)?base(e,"spec",e.type,{phaseId:e.phaseId,..."workflowId"in e&&e.workflowId?{workflowId:e.workflowId}:{},..."reason"in e?{reason:e.reason}:{},..."artifact"in e&&e.artifact?{artifact:e.artifact}:{}}):"runtime.adapter.event"===e.type?function adapterTrace(e){const t=e.event;if(isRecord(t)&&"string"==typeof t.phase){const r=function semanticAdapterTrace(e,t){if("inventory.repair"===t.phase)return base(e,"adapter","runtime.inventory.repair",{status:t.status,...isRecord(t.diagnostic)?t.diagnostic:{}});const r=function readTraceType(e){const t=readString(e);return"request"===t||"tool"===t||"workflow"===t||"spec"===t||"adapter"===t||"memory"===t||"artifact"===t||"progress"===t||"plan"===t||"delegation"===t?t:void 0}(t.traceType),o=readString(t.traceLabel);return r&&o?base(e,r,o,t):void 0}(e,t);return r||base(e,"adapter",t.phase.startsWith("agent.")?t.phase:`adapter.${t.phase}`,t)}return base(e,"adapter","runtime.adapter.event",{event:t})}(e):"runtime.artifact.created"===e.type?base(e,"artifact","runtime.artifact.created",{artifact:e.artifact}):e.type.startsWith("runtime.memory.")?base(e,"memory",e.type):"runtime.progress.narration"===e.type?base(e,"progress",e.type,{message:e.message,provider:e.provider,sourceEventTypes:e.sourceEventTypes}):void 0}export function readPlanTodos(e){const t=function readPlanRecord(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),r=function readTodosArray(e){const t=isRecord(e?.args)?e.args:void 0;return Array.isArray(e?.todos)?e.todos:Array.isArray(t?.todos)?t.todos:[]}(t);return r.map(readTodo).filter(isPlanTodoItem)}function readTodo(e){if(isRecord(e)&&"string"==typeof e.content)return{content:e.content,status:"string"==typeof e.status?e.status:"pending"}}function isPlanTodoItem(e){return void 0!==e}function base(e,t,r,o){return{type:t,label:r,agentId:e.agentId,requestId:e.requestId,detail:o}}function isTraceEntry(e){return void 0!==e}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}
+export function projectRuntimeTrace(e){return e.events.map(projectEvent).filter(isTraceEntry)}export function projectEvent(e){return"runtime.request.started"===e.type||"runtime.request.completed"===e.type||"runtime.request.failed"===e.type?base(e,"request",e.type):"runtime.request.cancelled"===e.type?base(e,"request",e.type,{reason:e.reason}):"runtime.execution.contract.failed"===e.type?base(e,"request",e.type,{reason:e.reason,missingEvidenceTools:e.missingEvidenceTools}):"runtime.inventory.repair"===e.type?base(e,"adapter","runtime.inventory.repair",{status:e.status,...e.diagnostic}):"runtime.tool.direct.started"===e.type?base(e,"tool","runtime.tool.direct.started",{toolId:e.toolId}):"runtime.tool.direct.completed"===e.type?base(e,"tool","runtime.tool.direct.completed",{toolId:e.toolId}):"runtime.tool.failure"===e.type?base(e,"tool","runtime.tool.failure",{toolId:e.toolId,...e.failure}):"runtime.tool.circuit.opened"===e.type?base(e,"tool","runtime.tool.circuit.opened",{toolId:e.toolId,reason:e.reason}):"runtime.workflow.started"===e.type||"runtime.workflow.completed"===e.type?base(e,"workflow",e.type,{workflowId:e.workflowId,adapter:e.adapter}):function isSpecDrivenPhaseEvent(e){return e.type.startsWith("runtime.specDriven.phase.")}(e)?base(e,"spec",e.type,{phaseId:e.phaseId,..."workflowId"in e&&e.workflowId?{workflowId:e.workflowId}:{},..."reason"in e?{reason:e.reason}:{},..."artifact"in e&&e.artifact?{artifact:e.artifact}:{}}):"runtime.adapter.event"===e.type?function adapterTrace(e){const t=e.event;if(isRecord(t)&&"string"==typeof t.phase){const r=function semanticAdapterTrace(e,t){if("inventory.repair"===t.phase)return base(e,"adapter","runtime.inventory.repair",{status:t.status,...isRecord(t.diagnostic)?t.diagnostic:{}});const r=function readTraceType(e){const t=readString(e);return"request"===t||"tool"===t||"workflow"===t||"spec"===t||"adapter"===t||"memory"===t||"artifact"===t||"progress"===t||"plan"===t||"delegation"===t?t:void 0}(t.traceType),o=readString(t.traceLabel);return r&&o?base(e,r,o,t):void 0}(e,t);return r||base(e,"adapter",t.phase.startsWith("agent.")?t.phase:`adapter.${t.phase}`,t)}return base(e,"adapter","runtime.adapter.event",{event:t})}(e):"runtime.artifact.created"===e.type?base(e,"artifact","runtime.artifact.created",{artifact:e.artifact}):e.type.startsWith("runtime.memory.")?base(e,"memory",e.type):"runtime.progress.narration"===e.type?base(e,"progress",e.type,{message:e.message,provider:e.provider,sourceEventTypes:e.sourceEventTypes}):void 0}export function readPlanTodos(e){const t=function readPlanRecord(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),r=function readTodosArray(e){const t=isRecord(e?.args)?e.args:void 0;return Array.isArray(e?.todos)?e.todos:Array.isArray(t?.todos)?t.todos:[]}(t);return r.map(readTodo).filter(isPlanTodoItem)}function readTodo(e){if(isRecord(e)&&"string"==typeof e.content)return{content:e.content,status:"string"==typeof e.status?e.status:"pending"}}function isPlanTodoItem(e){return void 0!==e}function base(e,t,r,o){return{type:t,label:r,agentId:e.agentId,requestId:e.requestId,detail:o}}function isTraceEntry(e){return void 0!==e}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}

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

@@ -3,6 +3,7 @@ import type { ApprovalQueue } from "@stable-harness/governance";
 import type { RuntimeWorkflowAdapter, RuntimeWorkflowRequest, WorkspaceWorkflow } from "./workflows/index.js";
 import type { SpecDrivenWorkflowState } from "./spec-driven/index.js";
 import type { RuntimeEvent, RuntimeEventListener, RuntimeEmit } from "./runtime/events.js";
+import type { RuntimeToolFailureTracker } from "./runtime/tool-failure.js";
 import type { RuntimeArtifact, RuntimeOutput, RuntimeRecordState, RuntimeRequest, RuntimeResponse, RuntimeRunFilter, RuntimeRunRecord } from "./runtime/types.js";
 import type { RuntimeToolGateway } from "./runtime/tool-gateway.js";
 import type { CompiledWorkspace, WorkspaceAgent, WorkspaceRuntimePolicy } from "./workspace/types.js";
@@ -12,9 +13,11 @@ export type { RuntimeCapabilityContext, RuntimeCapabilityModule, RuntimeCapabili
 export type { RuntimeProgressNarrationOptions, RuntimeProgressNarrationProvider } from "./runtime/progress-narration.js";
 export type { RuntimeArtifact, RuntimeArtifactFilter, RuntimeArtifactInput, RuntimeArtifactRecord, RuntimeArtifactStore, RuntimeCancelIntentInput, RuntimeHeartbeatInput, RuntimeMemoryCandidateInput, RuntimeOutput, RuntimeQueueClaimInput, RuntimeQueueRecord, RuntimeQueueStore, RuntimeRecoveryIntent, RuntimeRecordState, RuntimeRequest, RuntimeRequestControlRecord, RuntimeRequestMemory, RuntimeResponse, RuntimeRunFilter, RuntimeRunRecord, RuntimeStore, RuntimeStoreRunPatch, RuntimeStuckRequestInput, } from "./runtime/types.js";
 export type { CompiledWorkspace, WorkspaceAdapterPolicy, WorkspaceAgent, WorkspaceBoundaryDiagnostic, WorkspaceBoundaryDiagnosticCode, WorkspaceBoundaryScanPolicy, WorkspaceMemory, WorkspaceModel, WorkspaceRetryPolicy, WorkspaceRetryReason, WorkspaceRetryTargetPolicy, WorkspaceRuntimePolicy, WorkspaceSkill, WorkspaceTool, WorkspaceToolRetryPolicy, WorkspaceValidationPolicy, } from "./workspace/types.js";
+export type { WorkspaceToolQualityDiagnostic, WorkspaceToolQualityDiagnosticCode, WorkspaceToolQualityPolicy, } from "./workspace/tool-quality.js";
 export type { WorkspaceEvaluation, WorkspaceEvaluationCase, } from "./evaluations/index.js";
 export type { SpecDrivenPhaseId, SpecDrivenPhaseRecord, SpecDrivenPhaseStatus, SpecDrivenPhaseTransition, SpecDrivenWorkflowState, WorkspaceSpecDrivenPhase, WorkspaceSpecDrivenWorkflowPolicy, } from "./spec-driven/index.js";
 export type { RuntimeToolCallRequest, RuntimeToolGateway, RuntimeToolGatewayContext, RuntimeToolRepairModel, RuntimeToolGatewayTool, } from "./runtime/tool-gateway.js";
+export type { RuntimeToolFailureClassification, RuntimeToolFailureKind, RuntimeToolFailurePolicy, RuntimeToolFailureReason, RuntimeToolFailureTracker, } from "./runtime/tool-failure.js";
 export type RuntimeAdapterContext = {
     workspace: CompiledWorkspace;
     agent: WorkspaceAgent;
@@ -24,6 +27,7 @@ export type RuntimeAdapterContext = {
     memory?: RuntimeMemoryContext;
     pluginMemories?: RuntimeMemoryContext[];
     toolGateway?: RuntimeToolGateway;
+    toolFailureTracker?: RuntimeToolFailureTracker;
     requestState?: Map<string, unknown>;
     emit: RuntimeEmit;
 };

package/packages/core/dist/workspace/tool-quality.d.ts

@@ -0,0 +1,22 @@
+import type { CompiledWorkspace } from "./types.js";
+export type WorkspaceToolQualityDiagnosticCode = "missing-description" | "weak-description" | "missing-schema" | "missing-output-contract" | "missing-permission-metadata" | "overlapping-description";
+export type WorkspaceToolQualitySeverity = "error" | "warning";
+export type WorkspaceToolQualityDiagnostic = {
+    code: WorkspaceToolQualityDiagnosticCode;
+    layer: "tool";
+    owner: "stable_runtime_policy" | "workspace_config";
+    resources: string[];
+    severity: WorkspaceToolQualitySeverity;
+    message: string;
+    locations: string[];
+    details?: Record<string, unknown>;
+};
+export type WorkspaceToolQualityPolicy = {
+    enabled?: boolean;
+    requireSchema?: boolean;
+    requireOutputContract?: boolean;
+    requirePermissionMetadata?: boolean;
+    failOn?: Partial<Record<WorkspaceToolQualityDiagnosticCode, boolean>>;
+};
+export declare function scanWorkspaceToolQuality(workspace: CompiledWorkspace, policy?: WorkspaceToolQualityPolicy): WorkspaceToolQualityDiagnostic[];
+export declare function assertWorkspaceToolQualityDiagnostics(diagnostics: WorkspaceToolQualityDiagnostic[]): void;

package/packages/core/dist/workspace/tool-quality.js

@@ -0,0 +1 @@
+export function scanWorkspaceToolQuality(t,e={}){if(!1===e.enabled)return[];const o=[...t.tools.values()].flatMap(t=>function scanTool(t,e){const o=[],i=t.description?.trim()??"";return i?function isWeakDescription(t,e){const o=t.replace(/[_-]+/gu," ").toLowerCase(),i=e.replace(/[_-]+/gu," ").toLowerCase();return i.length<24||i===o}(t.id,i)&&o.push(diagnostic(t,"weak-description","Tool description is too short or repeats only the tool id.",e)):o.push(diagnostic(t,"missing-description","Tool is missing a description.",e)),!0===e.requireSchema&&void 0===t.schema&&o.push(diagnostic(t,"missing-schema","Tool is missing an input schema.",e)),!0!==e.requireOutputContract||void 0!==t.outputSchema||hasMetadata(t,"outputContract")||o.push(diagnostic(t,"missing-output-contract","Tool is missing output contract metadata.",e)),!0!==e.requirePermissionMetadata||hasMetadata(t,"permissions")||o.push(diagnostic(t,"missing-permission-metadata","Tool is missing permission metadata.",e)),o}(t,e));return[...o,...scanOverlaps([...t.tools.values()])]}export function assertWorkspaceToolQualityDiagnostics(t){const e=t.filter(t=>"error"===t.severity);if(0!==e.length)throw new Error(`Workspace tool quality gate failed:\n${e.map(formatDiagnostic).join("\n")}`)}function scanOverlaps(t){const e=[];for(let o=0;o<t.length;o+=1)for(let i=o+1;i<t.length;i+=1){const s=t[o],n=t[i],a=descriptionOverlap(s.description,n.description);a<.8||e.push({code:"overlapping-description",layer:"tool",owner:"workspace_config",resources:[s.id,n.id],severity:"warning",message:`Tools ${s.id} and ${n.id} have highly overlapping descriptions; add explicit selection boundaries.`,locations:[s.sourcePath,n.sourcePath].filter(isString),details:{score:a}})}return e}function diagnostic(t,e,o,i){return{code:e,layer:"tool",owner:"workspace_config",resources:[t.id],severity:!0===i.failOn?.[e]?"error":"warning",message:o,locations:t.sourcePath?[t.sourcePath]:[]}}function descriptionOverlap(t,e){const o=tokenSet(t),i=tokenSet(e);return o.size<4||i.size<4?0:[...o].filter(t=>i.has(t)).length/Math.min(o.size,i.size)}function tokenSet(t){return new Set((t??"").toLowerCase().split(/[^a-z0-9]+/u).filter(t=>t.length>2))}function hasMetadata(t,e){return"object"==typeof t.metadata&&null!==t.metadata&&!Array.isArray(t.metadata)&&void 0!==t.metadata[e]}function formatDiagnostic(t){return`- ${t.code}: ${t.message} (${t.resources.join(", ")})`}function isString(t){return"string"==typeof t&&t.length>0}

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

@@ -1,6 +1,7 @@
 import type { WorkspaceEvaluation } from "../evaluations/index.js";
 import type { WorkspaceSpecDrivenWorkflowPolicy } from "../spec-driven/index.js";
 import type { WorkspaceWorkflow, WorkspaceWorkflowEdge, WorkspaceWorkflowRoutingPolicy } from "../workflows/index.js";
+import type { WorkspaceToolQualityDiagnostic, WorkspaceToolQualityPolicy } from "./tool-quality.js";
 export type WorkspaceModel = {
     id: string;
     provider: string;
@@ -11,6 +12,8 @@ export type WorkspaceTool = {
     id: string;
     description?: string;
     schema?: unknown;
+    outputSchema?: unknown;
+    metadata?: Record<string, unknown>;
     implementation?: string;
     sourcePath?: string;
 };
@@ -77,6 +80,7 @@ export type WorkspaceAgentRoute = {
 };
 export type WorkspaceValidationPolicy = {
     boundaryScan?: WorkspaceBoundaryScanPolicy;
+    toolQuality?: WorkspaceToolQualityPolicy;
 };
 export type WorkspaceBoundaryScanPolicy = {
     enabled?: boolean;
@@ -125,5 +129,5 @@ export type CompiledWorkspace = {
     memories: Map<string, WorkspaceMemory>;
     workflows: Map<string, WorkspaceWorkflow>;
     evaluations?: Map<string, WorkspaceEvaluation>;
-    diagnostics?: WorkspaceBoundaryDiagnostic[];
+    diagnostics?: Array<WorkspaceBoundaryDiagnostic | WorkspaceToolQualityDiagnostic>;
 };

package/packages/evaluation/dist/src/index.d.ts

@@ -1,2 +1,4 @@
 export { createTraceRecorder, createReplayManifest, createEvaluationBundle } from "./trace.js";
+export { computeToolCallMetrics } from "./tool-call-metrics.js";
 export type { EvaluationBundle, ReplayManifest } from "./types.js";
+export type { PerToolCallMetrics, ToolCallMetrics } from "./tool-call-metrics.js";

package/packages/evaluation/dist/src/index.js

@@ -1 +1 @@
-export{createTraceRecorder,createReplayManifest,createEvaluationBundle}from"./trace.js";
+export{createTraceRecorder,createReplayManifest,createEvaluationBundle}from"./trace.js";export{computeToolCallMetrics}from"./tool-call-metrics.js";

package/packages/evaluation/dist/src/tool-call-metrics.d.ts

@@ -0,0 +1,26 @@
+import type { RuntimeEvent } from "@stable-harness/core";
+export type ToolCallMetrics = {
+    toolStarted: number;
+    toolCompleted: number;
+    toolFailures: number;
+    inventoryRepairs: number;
+    inventoryRepairBlocks: number;
+    argumentFailureCount: number;
+    transientFailureCount: number;
+    controlFailureCount: number;
+    businessFailureCount: number;
+    circuitOpenCount: number;
+    firstAttemptArgumentValidityRate: number | undefined;
+    schemaRepairFailureRate: number | undefined;
+    repairRate: number | undefined;
+    failureRate: number | undefined;
+    perTool: Record<string, PerToolCallMetrics>;
+};
+export type PerToolCallMetrics = {
+    started: number;
+    completed: number;
+    failures: number;
+    inventoryRepairs: number;
+    circuitOpen: number;
+};
+export declare function computeToolCallMetrics(events: RuntimeEvent[]): ToolCallMetrics;

package/packages/evaluation/dist/src/tool-call-metrics.js

@@ -0,0 +1 @@
+export function computeToolCallMetrics(t){const e=function createCounters(){return{toolStarted:0,toolCompleted:0,toolFailures:0,inventoryRepairs:0,inventoryRepairBlocks:0,argumentFailureCount:0,transientFailureCount:0,controlFailureCount:0,businessFailureCount:0,circuitOpenCount:0,firstAttemptArgumentValidityRate:void 0,schemaRepairFailureRate:void 0,repairRate:void 0,failureRate:void 0,perTool:new Map}}();for(const o of t)applyEvent(e,o);return{...e,firstAttemptArgumentValidityRate:rate(e.toolStarted-e.argumentFailureCount,e.toolStarted),schemaRepairFailureRate:rate(e.argumentFailureCount,e.toolStarted),repairRate:rate(e.inventoryRepairs,e.toolStarted+e.inventoryRepairs),failureRate:rate(e.toolFailures,e.toolStarted),perTool:Object.fromEntries([...e.perTool.entries()])}}function applyEvent(t,e){"runtime.tool.direct.started"===e.type?(t.toolStarted+=1,perTool(t,e.toolId).started+=1):"runtime.tool.direct.completed"===e.type?(t.toolCompleted+=1,perTool(t,e.toolId).completed+=1):"runtime.tool.failure"===e.type?function applyFailure(t,e){t.toolFailures+=1,perTool(t,e.toolId).failures+=1,"transient"===e.failure.kind&&(t.transientFailureCount+=1),"control"===e.failure.kind&&(t.controlFailureCount+=1),"business"===e.failure.kind&&(t.businessFailureCount+=1),"invalid_input"!==e.failure.reason&&"schema_repair_failed"!==e.failure.reason||(t.argumentFailureCount+=1)}(t,e):"runtime.tool.circuit.opened"===e.type?(t.circuitOpenCount+=1,perTool(t,e.toolId).circuitOpen+=1):"runtime.inventory.repair"===e.type?function applyInventoryRepair(t,e){if("repaired"===e.status){t.inventoryRepairs+=1;for(const o of[e.diagnostic.originalId,e.diagnostic.repairedId].filter(isString))perTool(t,o).inventoryRepairs+=1}else t.inventoryRepairBlocks+=1}(t,e):"runtime.adapter.event"===e.type&&function applyAdapterEvent(t,e){(function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)})(e)&&("agent.tool.start"===e.phase&&"string"==typeof e.toolId?(t.toolStarted+=1,perTool(t,e.toolId).started+=1):"agent.tool.result"===e.phase&&"string"==typeof e.toolId&&(t.toolCompleted+=1,perTool(t,e.toolId).completed+=1))}(t,e.event)}function perTool(t,e){const o=t.perTool.get(e);if(o)return o;const r={started:0,completed:0,failures:0,inventoryRepairs:0,circuitOpen:0};return t.perTool.set(e,r),r}function rate(t,e){return e>0?t/e:void 0}function isString(t){return"string"==typeof t&&t.length>0}

package/packages/evaluation/dist/src/trace.js

@@ -1 +1 @@
-export function createTraceRecorder(){const e=[];return{listener(t){e.push(t)},events:()=>[...e]}}export function createReplayManifest(e){return{schemaVersion:1,createdAt:(new Date).toISOString(),runs:e.runs,events:e.events}}export function createEvaluationBundle(e){return{...createReplayManifest({runs:e.runs,events:e.events}),artifacts:e.artifacts??[],metadata:e.metadata}}
+import{computeToolCallMetrics as e}from"./tool-call-metrics.js";export function createTraceRecorder(){const e=[];return{listener(t){e.push(t)},events:()=>[...e]}}export function createReplayManifest(e){return{schemaVersion:1,createdAt:(new Date).toISOString(),runs:e.runs,events:e.events}}export function createEvaluationBundle(t){return{...createReplayManifest({runs:t.runs,events:t.events}),artifacts:t.artifacts??[],metadata:{...t.metadata??{},toolCallMetrics:e(t.events)}}}

package/packages/workspace-yaml/dist/documents.js

@@ -1 +1 @@
-import{assertSpecDrivenWorkflowPolicy as e,createSpecDrivenWorkflowPolicy as r}from"@stable-harness/core";function assertRecord(e,r){if("object"!=typeof e||null===e||Array.isArray(e))throw new Error(`${r} must be an object`);return e}function readName(e,r){const t=e.metadata?.name;if("string"==typeof t&&t.trim())return t.trim();if(r)return r;throw new Error(`Document kind ${String(e.kind)} requires metadata.name`)}function readDescription(e){const r=e.metadata?.description;return"string"==typeof r&&r.trim()?r.trim():void 0}function readOptionalString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function toStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.trim().length>0):[]}function resolveValue(e){if("string"!=typeof e)return e;const r=e.match(/^\$\{env:([A-Za-z_][A-Za-z0-9_]*)(?::-(.*))?\}$/u);return r?process.env[r[1]]??r[2]??"":e}export function compileRuntime(e){const r=assertRecord(e.spec,"Runtime.spec"),t=assertRecord(r.routing??{},"Runtime.spec.routing");return{defaultAgentId:"string"==typeof t.defaultAgentId&&t.defaultAgentId.trim()?t.defaultAgentId.trim():"orchestra",...void 0!==t.routes?{routes:readAgentRoutes(t.routes)}:{},...readOptionalString(r.workspaceId)?{workspaceId:readOptionalString(r.workspaceId)}:{},...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...void 0!==r.adapters?{adapters:readAdapters(r.adapters)}:{},..."object"==typeof r.workflowRouting&&r.workflowRouting?{workflowRouting:readWorkflowRouting(r.workflowRouting)}:{},..."object"==typeof r.specDrivenWorkflow&&r.specDrivenWorkflow?{specDrivenWorkflow:readSpecDrivenWorkflow(r.specDrivenWorkflow)}:{},..."object"==typeof r.approvals&&r.approvals?{approvals:r.approvals}:{},..."object"==typeof r.recovery&&r.recovery?{recovery:r.recovery}:{},..."object"==typeof r.retry&&r.retry?{retry:r.retry}:{},..."object"==typeof r.toolGateway&&r.toolGateway?{toolGateway:r.toolGateway}:{},..."object"==typeof r.memory&&r.memory?{memory:r.memory}:{},..."object"==typeof r.protocols&&r.protocols?{protocols:r.protocols}:{},..."object"==typeof r.progress&&r.progress?{progress:r.progress}:{},..."object"==typeof r.cli&&r.cli?{cli:r.cli}:{},..."string"==typeof r.quality||"object"==typeof r.quality&&r.quality?{quality:r.quality}:{},..."object"==typeof r.workspaceValidation&&r.workspaceValidation?{workspaceValidation:r.workspaceValidation}:{},..."object"==typeof r.responseLanguage&&r.responseLanguage?{responseLanguage:r.responseLanguage}:{},..."object"==typeof r.responsePresentation&&r.responsePresentation?{responsePresentation:r.responsePresentation}:{}}}function readAgentRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.routing.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.routing.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.agentId);if(!t||!o)throw new Error("Runtime.spec.routing.routes[] requires id and agentId");const n=void 0===r.keywords?void 0:function assertStringArray(e,r){if(!Array.isArray(e))throw new Error(`${r} must be an array`);return e.map(e=>{if("string"!=typeof e||!e.trim())throw new Error(`${r} must contain non-empty strings`);return e.trim()})}(r.keywords,"Runtime.spec.routing.routes[].keywords"),i=readOptionalString(r.pattern);if(!(n&&0!==n.length||i))throw new Error("Runtime.spec.routing.routes[] requires keywords or pattern");return{id:t,agentId:o,...n&&n.length>0?{keywords:n}:{},...i?{pattern:i}:{},...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{}}})}function readSpecDrivenWorkflow(t){const o=assertRecord(t,"Runtime.spec.specDrivenWorkflow"),n=r({enabled:!0===o.enabled,constitution:readOptionalString(o.constitution),artifactsDir:readOptionalString(o.artifactsDir),phases:void 0===o.phases?void 0:readSpecDrivenPhases(o.phases),..."object"==typeof o.gates&&o.gates?{gates:o.gates}:{},..."object"==typeof o.config&&o.config?{config:o.config}:{}});return e(n),n}function readSpecDrivenPhases(e){if(!Array.isArray(e))throw new Error("Runtime.spec.specDrivenWorkflow.phases must be an array");return e.map(e=>{if("string"==typeof e&&e.trim())return{id:e.trim()};const r=assertRecord(e,"Runtime.spec.specDrivenWorkflow.phases[]"),t=readOptionalString(r.id);if(!t)throw new Error("Runtime.spec.specDrivenWorkflow.phases[] requires id");return{id:t,...readOptionalString(r.artifactKind)?{artifactKind:readOptionalString(r.artifactKind)}:{},..."boolean"==typeof r.required?{required:r.required}:{},...readOptionalString(r.gate)?{gate:readOptionalString(r.gate)}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}})}export function compileAgent(e,r){const t=assertRecord(e.spec,"Agent.spec"),o=readName(e),n=readOptionalString(t.backend);if(!n)throw new Error(`Agent ${o} requires spec.backend`);const i="object"==typeof t.config&&t.config?t.config:{},a="string"==typeof t.systemPrompt?t.systemPrompt:"string"==typeof i.systemPrompt?i.systemPrompt:void 0;return{id:o,...readDescription(e)?{description:readDescription(e)}:{},sourcePath:r,backend:n,..."string"==typeof t.modelRef&&t.modelRef.trim()?{modelRef:(s=t.modelRef,s.replace(/^[^/]+\//u,""))}:{},...void 0!==a?{systemPrompt:a}:{},tools:toStringArray(t.tools),skills:toStringArray(t.skills),memory:Array.isArray(t.memory)?t.memory:[],subagents:toStringArray(t.subagents),...void 0!==t.edges?{edges:readAgentEdges(t.edges,o)}:{},config:i};var s}export function compileModel(e){return compileModelSpec(assertRecord(e.spec,"Model.spec"),readName(e))}export function compileModelSpec(e,r){const t="string"==typeof e.name&&e.name.trim()?e.name.trim():r??"default",o=resolveValue(e.provider),n=resolveValue(e.model),i="string"==typeof o&&o.trim()?o.trim():"unknown",a="string"==typeof n&&n.trim()?n.trim():t,s={...e};return delete s.name,delete s.provider,delete s.model,{id:t,provider:i,model:a,config:Object.fromEntries(Object.entries(s).map(([e,r])=>[e,resolveValue(r)]))}}export function compileTool(e,r){const t=assertRecord(e.spec,"Tool.spec");return{id:readName(e),...r?{sourcePath:r}:{},..."string"==typeof t.description?{description:t.description}:{},...void 0!==t.schema?{schema:t.schema}:{},..."string"==typeof t.implementation?{implementation:t.implementation}:{}}}export function compileMemory(e){const r=assertRecord(e.spec,"Memory.spec"),t=readName(e),o={...r};return delete o.provider,delete o.profile,delete o.mode,delete o.enabled,delete o.prompts,{id:t,provider:readOptionalString(r.provider)??"langmem",...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...readOptionalString(r.mode)?{mode:readOptionalString(r.mode)}:{},enabled:!1!==r.enabled,..."object"==typeof r.prompts&&r.prompts?{prompts:readMemoryPrompts(r.prompts)}:{},...Object.keys(o).length>0?{config:o}:{}}}function readWorkflowRouting(e){const r=assertRecord(e,"Runtime.spec.workflowRouting"),t=void 0===r.routes?void 0:function readWorkflowRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.workflowRouting.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.workflowRouting.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.workflowId);if(!t||!o)throw new Error("Runtime.spec.workflowRouting.routes[] requires id and workflowId");return{id:t,workflowId:o,...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{},..."object"==typeof r.metadata&&r.metadata?{metadata:r.metadata}:{}}})}(r.routes);return{...readOptionalString(r.defaultWorkflowId)?{defaultWorkflowId:readOptionalString(r.defaultWorkflowId)}:{},...t?{routes:t}:{}}}function readAdapters(e){if(!Array.isArray(e))throw new Error("Runtime.spec.adapters must be an array");return e.map(readAdapter)}function readAgentEdges(e,r){if(!Array.isArray(e))throw new Error(`Agent ${r} spec.edges must be an array`);return e.map(e=>{const t=assertRecord(e,`Agent ${r} spec.edges[]`),o=readOptionalString(t.from),n=readOptionalString(t.to);if(!o||!n)throw new Error(`Agent ${r} spec.edges[] requires from and to`);return{from:o,to:n,...readOptionalString(t.condition)?{condition:readOptionalString(t.condition)}:{}}})}function readAdapter(e){if("string"==typeof e&&e.trim())return{name:e.trim()};const r=assertRecord(e,"Runtime.spec.adapters[]"),t=readOptionalString(r.name)??readOptionalString(r.id)??readOptionalString(r.backend);if(!t)throw new Error("Runtime.spec.adapters[] requires name");return{name:t,..."boolean"==typeof r.enabled?{enabled:r.enabled}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}}function readMemoryPrompts(e){const r=assertRecord(e,"Memory.spec.prompts");return{...readOptionalString(r.semantic)?{semantic:readOptionalString(r.semantic)}:{},...readOptionalString(r.episodic)?{episodic:readOptionalString(r.episodic)}:{},...readOptionalString(r.procedural)?{procedural:readOptionalString(r.procedural)}:{}}}
+import{assertSpecDrivenWorkflowPolicy as e,createSpecDrivenWorkflowPolicy as r}from"@stable-harness/core";function assertRecord(e,r){if("object"!=typeof e||null===e||Array.isArray(e))throw new Error(`${r} must be an object`);return e}function readName(e,r){const t=e.metadata?.name;if("string"==typeof t&&t.trim())return t.trim();if(r)return r;throw new Error(`Document kind ${String(e.kind)} requires metadata.name`)}function readDescription(e){const r=e.metadata?.description;return"string"==typeof r&&r.trim()?r.trim():void 0}function readOptionalString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}function toStringArray(e){return Array.isArray(e)?e.filter(e=>"string"==typeof e&&e.trim().length>0):[]}function resolveValue(e){if("string"!=typeof e)return e;const r=e.match(/^\$\{env:([A-Za-z_][A-Za-z0-9_]*)(?::-(.*))?\}$/u);return r?process.env[r[1]]??r[2]??"":e}export function compileRuntime(e){const r=assertRecord(e.spec,"Runtime.spec"),t=assertRecord(r.routing??{},"Runtime.spec.routing");return{defaultAgentId:"string"==typeof t.defaultAgentId&&t.defaultAgentId.trim()?t.defaultAgentId.trim():"orchestra",...void 0!==t.routes?{routes:readAgentRoutes(t.routes)}:{},...readOptionalString(r.workspaceId)?{workspaceId:readOptionalString(r.workspaceId)}:{},...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...void 0!==r.adapters?{adapters:readAdapters(r.adapters)}:{},..."object"==typeof r.workflowRouting&&r.workflowRouting?{workflowRouting:readWorkflowRouting(r.workflowRouting)}:{},..."object"==typeof r.specDrivenWorkflow&&r.specDrivenWorkflow?{specDrivenWorkflow:readSpecDrivenWorkflow(r.specDrivenWorkflow)}:{},..."object"==typeof r.approvals&&r.approvals?{approvals:r.approvals}:{},..."object"==typeof r.recovery&&r.recovery?{recovery:r.recovery}:{},..."object"==typeof r.retry&&r.retry?{retry:r.retry}:{},..."object"==typeof r.toolGateway&&r.toolGateway?{toolGateway:r.toolGateway}:{},..."object"==typeof r.memory&&r.memory?{memory:r.memory}:{},..."object"==typeof r.protocols&&r.protocols?{protocols:r.protocols}:{},..."object"==typeof r.progress&&r.progress?{progress:r.progress}:{},..."object"==typeof r.cli&&r.cli?{cli:r.cli}:{},..."string"==typeof r.quality||"object"==typeof r.quality&&r.quality?{quality:r.quality}:{},..."object"==typeof r.workspaceValidation&&r.workspaceValidation?{workspaceValidation:r.workspaceValidation}:{},..."object"==typeof r.responseLanguage&&r.responseLanguage?{responseLanguage:r.responseLanguage}:{},..."object"==typeof r.responsePresentation&&r.responsePresentation?{responsePresentation:r.responsePresentation}:{}}}function readAgentRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.routing.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.routing.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.agentId);if(!t||!o)throw new Error("Runtime.spec.routing.routes[] requires id and agentId");const n=void 0===r.keywords?void 0:function assertStringArray(e,r){if(!Array.isArray(e))throw new Error(`${r} must be an array`);return e.map(e=>{if("string"!=typeof e||!e.trim())throw new Error(`${r} must contain non-empty strings`);return e.trim()})}(r.keywords,"Runtime.spec.routing.routes[].keywords"),i=readOptionalString(r.pattern);if(!(n&&0!==n.length||i))throw new Error("Runtime.spec.routing.routes[] requires keywords or pattern");return{id:t,agentId:o,...n&&n.length>0?{keywords:n}:{},...i?{pattern:i}:{},...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{}}})}function readSpecDrivenWorkflow(t){const o=assertRecord(t,"Runtime.spec.specDrivenWorkflow"),n=r({enabled:!0===o.enabled,constitution:readOptionalString(o.constitution),artifactsDir:readOptionalString(o.artifactsDir),phases:void 0===o.phases?void 0:readSpecDrivenPhases(o.phases),..."object"==typeof o.gates&&o.gates?{gates:o.gates}:{},..."object"==typeof o.config&&o.config?{config:o.config}:{}});return e(n),n}function readSpecDrivenPhases(e){if(!Array.isArray(e))throw new Error("Runtime.spec.specDrivenWorkflow.phases must be an array");return e.map(e=>{if("string"==typeof e&&e.trim())return{id:e.trim()};const r=assertRecord(e,"Runtime.spec.specDrivenWorkflow.phases[]"),t=readOptionalString(r.id);if(!t)throw new Error("Runtime.spec.specDrivenWorkflow.phases[] requires id");return{id:t,...readOptionalString(r.artifactKind)?{artifactKind:readOptionalString(r.artifactKind)}:{},..."boolean"==typeof r.required?{required:r.required}:{},...readOptionalString(r.gate)?{gate:readOptionalString(r.gate)}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}})}export function compileAgent(e,r){const t=assertRecord(e.spec,"Agent.spec"),o=readName(e),n=readOptionalString(t.backend);if(!n)throw new Error(`Agent ${o} requires spec.backend`);const i="object"==typeof t.config&&t.config?t.config:{},a="string"==typeof t.systemPrompt?t.systemPrompt:"string"==typeof i.systemPrompt?i.systemPrompt:void 0;return{id:o,...readDescription(e)?{description:readDescription(e)}:{},sourcePath:r,backend:n,..."string"==typeof t.modelRef&&t.modelRef.trim()?{modelRef:(s=t.modelRef,s.replace(/^[^/]+\//u,""))}:{},...void 0!==a?{systemPrompt:a}:{},tools:toStringArray(t.tools),skills:toStringArray(t.skills),memory:Array.isArray(t.memory)?t.memory:[],subagents:toStringArray(t.subagents),...void 0!==t.edges?{edges:readAgentEdges(t.edges,o)}:{},config:i};var s}export function compileModel(e){return compileModelSpec(assertRecord(e.spec,"Model.spec"),readName(e))}export function compileModelSpec(e,r){const t="string"==typeof e.name&&e.name.trim()?e.name.trim():r??"default",o=resolveValue(e.provider),n=resolveValue(e.model),i="string"==typeof o&&o.trim()?o.trim():"unknown",a="string"==typeof n&&n.trim()?n.trim():t,s={...e};return delete s.name,delete s.provider,delete s.model,{id:t,provider:i,model:a,config:Object.fromEntries(Object.entries(s).map(([e,r])=>[e,resolveValue(r)]))}}export function compileTool(e,r){const t=assertRecord(e.spec,"Tool.spec");return{id:readName(e),...r?{sourcePath:r}:{},..."string"==typeof t.description?{description:t.description}:{},...void 0!==t.schema?{schema:t.schema}:{},...void 0!==t.outputSchema?{outputSchema:t.outputSchema}:{},..."object"==typeof t.metadata&&t.metadata?{metadata:t.metadata}:{},..."string"==typeof t.implementation?{implementation:t.implementation}:{}}}export function compileMemory(e){const r=assertRecord(e.spec,"Memory.spec"),t=readName(e),o={...r};return delete o.provider,delete o.profile,delete o.mode,delete o.enabled,delete o.prompts,{id:t,provider:readOptionalString(r.provider)??"langmem",...readOptionalString(r.profile)?{profile:readOptionalString(r.profile)}:{},...readOptionalString(r.mode)?{mode:readOptionalString(r.mode)}:{},enabled:!1!==r.enabled,..."object"==typeof r.prompts&&r.prompts?{prompts:readMemoryPrompts(r.prompts)}:{},...Object.keys(o).length>0?{config:o}:{}}}function readWorkflowRouting(e){const r=assertRecord(e,"Runtime.spec.workflowRouting"),t=void 0===r.routes?void 0:function readWorkflowRoutes(e){if(!Array.isArray(e))throw new Error("Runtime.spec.workflowRouting.routes must be an array");return e.map(e=>{const r=assertRecord(e,"Runtime.spec.workflowRouting.routes[]"),t=readOptionalString(r.id),o=readOptionalString(r.workflowId);if(!t||!o)throw new Error("Runtime.spec.workflowRouting.routes[] requires id and workflowId");return{id:t,workflowId:o,...readOptionalString(r.description)?{description:readOptionalString(r.description)}:{},..."object"==typeof r.metadata&&r.metadata?{metadata:r.metadata}:{}}})}(r.routes);return{...readOptionalString(r.defaultWorkflowId)?{defaultWorkflowId:readOptionalString(r.defaultWorkflowId)}:{},...t?{routes:t}:{}}}function readAdapters(e){if(!Array.isArray(e))throw new Error("Runtime.spec.adapters must be an array");return e.map(readAdapter)}function readAgentEdges(e,r){if(!Array.isArray(e))throw new Error(`Agent ${r} spec.edges must be an array`);return e.map(e=>{const t=assertRecord(e,`Agent ${r} spec.edges[]`),o=readOptionalString(t.from),n=readOptionalString(t.to);if(!o||!n)throw new Error(`Agent ${r} spec.edges[] requires from and to`);return{from:o,to:n,...readOptionalString(t.condition)?{condition:readOptionalString(t.condition)}:{}}})}function readAdapter(e){if("string"==typeof e&&e.trim())return{name:e.trim()};const r=assertRecord(e,"Runtime.spec.adapters[]"),t=readOptionalString(r.name)??readOptionalString(r.id)??readOptionalString(r.backend);if(!t)throw new Error("Runtime.spec.adapters[] requires name");return{name:t,..."boolean"==typeof r.enabled?{enabled:r.enabled}:{},..."object"==typeof r.config&&r.config?{config:r.config}:{}}}function readMemoryPrompts(e){const r=assertRecord(e,"Memory.spec.prompts");return{...readOptionalString(r.semantic)?{semantic:readOptionalString(r.semantic)}:{},...readOptionalString(r.episodic)?{episodic:readOptionalString(r.episodic)}:{},...readOptionalString(r.procedural)?{procedural:readOptionalString(r.procedural)}:{}}}

package/packages/workspace-yaml/dist/loader.js

@@ -1 +1 @@
-import{readFile as o}from"node:fs/promises";import e from"node:path";import{parseAllDocuments as t}from"yaml";import{discoverModuleTools as s,discoverSkills as r,listYamlFiles as n}from"./discovery.js";import{compileAgent as a,compileMemory as l,compileModel as i,compileModelSpec as c,compileRuntime as f,compileTool as u}from"./documents.js";import{compileWorkflow as w,validateWorkflows as m}from"./workflows.js";import{compileEvaluation as d,validateEvaluations as p}from"./evaluations.js";import{assertWorkspaceBoundaryDiagnostics as k,scanWorkspaceBoundaries as g}from"./boundary-scan.js";export async function loadWorkspaceFromYaml(a){const l=e.join(a,"config"),i=await n(l),c=[],f=new Map,u=new Map,w=new Map,d=new Map,v=new Map,M=new Map,y=new Map;for(const e of i){const s=await o(e,"utf8"),r=t(s).map(o=>o.toJSON()).filter(o=>null!==o);for(const o of r)collectWorkspaceDocument(o,e,{runtimeDocs:c,agents:f,models:u,tools:w,memories:v,workflows:M,evaluations:y})}for(const o of await s(a))w.has(o.id)||w.set(o.id,o);for(const o of await r(a))d.set(o.id,o);const W=c.at(-1)??{defaultAgentId:"orchestra"};m({workflows:M,agents:f,tools:w,skills:d}),p({evaluations:y,agents:f,tools:w,workflows:M}),function validateAgentRouting(o,e){for(const t of o.routes??[])if(!e.has(t.agentId))throw new Error(`Runtime routing route ${t.id} references unknown agent ${t.agentId}`)}(W,f),function validateWorkflowRouting(o,e){const t=o.workflowRouting;if(t){if(t.defaultWorkflowId&&!e.has(t.defaultWorkflowId))throw new Error(`Runtime workflowRouting.defaultWorkflowId references unknown workflow ${t.defaultWorkflowId}`);for(const o of t.routes??[])if(!e.has(o.workflowId))throw new Error(`Runtime workflowRouting route ${o.id} references unknown workflow ${o.workflowId}`)}}(W,M);const h={root:a,runtime:W,agents:f,models:u,tools:w,skills:d,memories:v,workflows:M,evaluations:y},R=g(h);return k(R),{...h,...R.length>0?{diagnostics:R}:{}}}function collectWorkspaceDocument(o,e,t){if("string"==typeof o.kind)switch(o.kind){case"Runtime":return void t.runtimeDocs.push(f(o));case"Agent":return collectOne(t.agents,a(o,e));case"Model":return collectOne(t.models,i(o));case"Models":return function collectModelSpecs(o,e){if(Array.isArray(o.spec))for(const t of o.spec)if("object"==typeof t&&null!==t&&!Array.isArray(t)){const o=c(t);e.set(o.id,o)}}(o,t.models);case"Tool":return collectOne(t.tools,u(o,e));case"Memory":return collectOne(t.memories,l(o));case"Workflow":return collectOne(t.workflows,w(o,e));case"Evaluation":return collectOne(t.evaluations,d(o,e));default:return}}function collectOne(o,e){o.set(e.id,e)}
+import{readFile as o}from"node:fs/promises";import e from"node:path";import{parseAllDocuments as s}from"yaml";import{discoverModuleTools as t,discoverSkills as r,listYamlFiles as a}from"./discovery.js";import{compileAgent as n,compileMemory as l,compileModel as i,compileModelSpec as c,compileRuntime as f,compileTool as u}from"./documents.js";import{compileWorkflow as m,validateWorkflows as w}from"./workflows.js";import{compileEvaluation as d,validateEvaluations as p}from"./evaluations.js";import{assertWorkspaceBoundaryDiagnostics as k,scanWorkspaceBoundaries as g}from"./boundary-scan.js";import{assertWorkspaceToolQualityDiagnostics as v,scanWorkspaceToolQuality as y}from"@stable-harness/core";export async function loadWorkspaceFromYaml(n){const l=e.join(n,"config"),i=await a(l),c=[],f=new Map,u=new Map,m=new Map,d=new Map,M=new Map,W=new Map,h=new Map;for(const e of i){const t=await o(e,"utf8"),r=s(t).map(o=>o.toJSON()).filter(o=>null!==o);for(const o of r)collectWorkspaceDocument(o,e,{runtimeDocs:c,agents:f,models:u,tools:m,memories:M,workflows:W,evaluations:h})}for(const o of await t(n))m.has(o.id)||m.set(o.id,o);for(const o of await r(n))d.set(o.id,o);const R=c.at(-1)??{defaultAgentId:"orchestra"};w({workflows:W,agents:f,tools:m,skills:d}),p({evaluations:h,agents:f,tools:m,workflows:W}),function validateAgentRouting(o,e){for(const s of o.routes??[])if(!e.has(s.agentId))throw new Error(`Runtime routing route ${s.id} references unknown agent ${s.agentId}`)}(R,f),function validateWorkflowRouting(o,e){const s=o.workflowRouting;if(s){if(s.defaultWorkflowId&&!e.has(s.defaultWorkflowId))throw new Error(`Runtime workflowRouting.defaultWorkflowId references unknown workflow ${s.defaultWorkflowId}`);for(const o of s.routes??[])if(!e.has(o.workflowId))throw new Error(`Runtime workflowRouting route ${o.id} references unknown workflow ${o.workflowId}`)}}(R,W);const A={root:n,runtime:R,agents:f,models:u,tools:m,skills:d,memories:M,workflows:W,evaluations:h},I=g(A);k(I);const O=y(A,R.workspaceValidation?.toolQuality);return v(O),{...A,...[...I??[],...O].length>0?{diagnostics:[...I,...O]}:{}}}function collectWorkspaceDocument(o,e,s){if("string"==typeof o.kind)switch(o.kind){case"Runtime":return void s.runtimeDocs.push(f(o));case"Agent":return collectOne(s.agents,n(o,e));case"Model":return collectOne(s.models,i(o));case"Models":return function collectModelSpecs(o,e){if(Array.isArray(o.spec))for(const s of o.spec)if("object"==typeof s&&null!==s&&!Array.isArray(s)){const o=c(s);e.set(o.id,o)}}(o,s.models);case"Tool":return collectOne(s.tools,u(o,e));case"Memory":return collectOne(s.memories,l(o));case"Workflow":return collectOne(s.workflows,m(o,e));case"Evaluation":return collectOne(s.evaluations,d(o,e));default:return}}function collectOne(o,e){o.set(e.id,e)}