stable-harness 0.0.19 → 0.0.21

package/docs/guides/operator-runbook.md

@@ -20,6 +20,12 @@ npm run release:pack
 npm run release:smoke
 ```
 
+Stable Harness publishes compiled npm artifacts only. The package must not
+include TypeScript sources, source maps, source-map content markers, tests,
+scripts, examples, benchmarks, or development-only source directories. The
+release package may include compiled JavaScript, declaration files, package
+metadata, docs, README, NOTICE, and LICENSE.
+
 `release:smoke` installs the packed package into an isolated temporary project,
 initializes a workspace, and runs a CLI tool call. It catches missing published
 dependencies that monorepo tests can hide.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stable-harness",
-  "version": "0.0.19",
+  "version": "0.0.21",
   "type": "module",
   "description": "Stable application runtime and operator control plane for agent workspaces.",
   "license": "Apache-2.0",

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,observedToolEvidence as l,stringifyDeepAgentResult as u,toolControlProjection as c}from"./gateway-tools.js";import{validateSkillFileBuiltinCall as d}from"./skill-file-policy.js";import{filterRepeatLimitedTools as g}from"./tool-repeat-visibility.js";import{traceProjectionForBuiltinTool as p}from"./trace-projection.js";const f=new Set(["ls","read_file","write_file","edit_file","glob","grep"]),m=new Set(["write_todos","read_todos","task","execute",...f]);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)&&f.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,c={}){const g=c.repeatState??s(e.workspace.runtime.toolGateway);return{name:"StableHarnessObserver",async wrapToolCall(s,p){const f=s.toolCall?.name;if(!f||!m.has(f))return p(s);const h=await r({toolId:f,request:s,workspaceRoot:e.workspace.root});emitToolEvent(e,f,"agent.tool.start",h.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){const e={...a,...l.args,subagent_type:l.candidateId};return{request:{...r,toolCall:{...r.toolCall,args:e}}}}const u=s?`: ${s}`:"",c=i.length>0?i.join(", "):"none";return{request:r,blocked:new o({tool_call_id:r.toolCall?.id??"stable-harness-task-policy",name:"task",status:"error",content:[`Task delegation target is not in the workspace inventory${u}. Allowed task targets: ${c}.`,"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,h):{request:h},b=y.request,v=y.blocked;if(v)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:v.content}),v;const k=validateFilesystemBuiltinCall(e,f,b);if(k)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:k.content}),k;const w=d(e,f,b);if(w)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:w.content}),w;try{const t=g?a(f,b.toolCall?.args,g):void 0;if(t)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:t.eventOutput}),builtinToolMessage(s,f,t.modelOutput);const o=await p(b),n=function observedToolOutput(e,t,o){return"write_todos"===e?JSON.stringify({status:"recorded",args:t.toolCall?.args}):u(o)}(f,b,o),r=g?i(f,b.toolCall?.args,n,o,g):{};return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:r.eventOutput??n}),c.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=l(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-${t}-argument-error`,name:t,status:"error",content:JSON.stringify({status:"tool_argument_error",toolId:t,instruction:"The upstream builtin tool rejected these arguments. Fix the tool arguments according to the tool schema, or choose a more appropriate available tool.",error:previewError(i)})})}(e,f,s,t);if(n)return emitToolEvent(e,f,"agent.tool.result",h.toolCall?.args,{output:n.content}),n;throw emitToolEvent(e,f,"agent.tool.result",h.toolCall?.args,{error:formatError(t)}),t}}}}function builtinToolMessage(e,t,n){return new o({tool_call_id:e.toolCall?.id??`stable-harness-${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,`${n}/**`]:[]]}(e)):[]}(e.workspace.root,e.workspace.skills.get(t)?.path)).filter(e=>e.startsWith("/")))];return o.length>0?[{operations:["read"],paths:o,mode:"allow"}]:[]}(e,t)),!1!==o?.filesystem){if(deepagentsMemoryWritable(e))return n.length>0?n:void 0}else n.push({operations:["read"],paths:["/memories/**"],mode:"allow"}),n.push({operations:["read","write"],paths:["/**"],mode:"deny"});return deepagentsMemoryWritable(e)||n.push({operations:["write"],paths:["/memories/**"],mode:"deny"}),n.length>0?n:void 0}function emitToolEvent(e,t,o,n,r={}){e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:o,toolId:t,..."agent.tool.start"===o?{args:n}:{},...r,..."string"==typeof r.output?c(r.output):{},...p(t,o,n)}})}function backendSkillPath(e,t){const o=function pathRelative(e,t){const o=e.split("/").filter(Boolean),n=t.split("/").filter(Boolean);if(!(n.length<o.length)){for(let e=0;e<o.length;e+=1)if(o[e]!==n[e])return;return n.slice(o.length).join("/")}}(e,t);return void 0===o?t:`/${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 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&&f.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?`${t.slice(0,797)}...`:t}
+import{realpathSync as e}from"node:fs";import{repairCallSelection as t}from"@botbotgo/better-call";import{ToolMessage as o}from"@langchain/core/messages";import{normalizeArgsRecord as n}from"./builtin-args.js";import{repairBuiltinToolRequest as r}from"./builtin-call-repair.js";import{afterToolInvoke as i,beforeToolInvoke as a,createToolRepeatState as s,stringifyDeepAgentResult as l,toolControlProjection as u}from"./gateway-tools.js";import{observedToolEvidence as c,recordObservedToolEvidence as d}from"./gateway/tool-evidence.js";import{validateSkillFileBuiltinCall as g}from"./skill-file-policy.js";import{filterRepeatLimitedTools as p}from"./tool-repeat-visibility.js";import{traceProjectionForBuiltinTool as f}from"./trace-projection.js";const m=new Set(["ls","read_file","write_file","edit_file","glob","grep"]),h=new Set(["write_todos","read_todos","task","execute",...m]);export function createBuiltinToolPolicyMiddleware(e,t={}){return{name:"StableHarnessBuiltinToolPolicy",async wrapModelCall(o,n){const r=Array.isArray(o.tools)?p(o.tools.filter(t=>!function hasHiddenBuiltins(e){return isFilesystemDisabled(e)||!isTaskVisible(e)}(e)||isModelVisibleBuiltin(e,t.name)),t.repeatState):o.tools,i=function normalizeToolChoice(e,t,o){return"required"===t?o&&o.length>0?t:"auto":function isForcedHiddenTool(e,t){return"string"==typeof t?.function?.name&&!isModelVisibleBuiltin(e,t.function.name)}(e,t)?"auto":t}(e,o.toolChoice,r);return n({...o,tools:r,toolChoice:i})}}}export function validateFilesystemBuiltinCall(e,t,n){if(isFilesystemDisabled(e)&&m.has(t))return new o({tool_call_id:n.toolCall?.id??`stable-harness-${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 d=u.repeatState??s(e.workspace.runtime.toolGateway);return{name:"StableHarnessObserver",async wrapToolCall(s,p){const f=s.toolCall?.name;if(!f||!h.has(f))return p(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){const e={...a,...l.args,subagent_type:l.candidateId};return{request:{...r,toolCall:{...r.toolCall,args:e}}}}const u=s?`: ${s}`:"",c=i.length>0?i.join(", "):"none";return{request:r,blocked:new o({tool_call_id:r.toolCall?.id??"stable-harness-task-policy",name:"task",status:"error",content:[`Task delegation target is not in the workspace inventory${u}. Allowed task targets: ${c}.`,"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},b=y.request,v=y.blocked;if(v)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:v.content}),v;const w=validateFilesystemBuiltinCall(e,f,b);if(w)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:w.content}),w;const k=g(e,f,b);if(k)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:k.content}),k;try{const t=d?a(f,b.toolCall?.args,d):void 0;if(t)return emitToolEvent(e,f,"agent.tool.result",b.toolCall?.args,{output:t.eventOutput}),builtinToolMessage(s,f,t.modelOutput);const o=await p(b),n=function observedToolOutput(e,t,o){return"write_todos"===e?JSON.stringify({status:"recorded",args:t.toolCall?.args}):l(o)}(f,b,o),r=d?i(f,b.toolCall?.args,n,o,d):{};return emitToolEvent(e,f,"agent.tool.result",b.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=c(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,`${n}/**`]:[]]}(e)):[]}(e.workspace.root,e.workspace.skills.get(t)?.path)).filter(e=>e.startsWith("/")))];return o.length>0?[{operations:["read"],paths:o,mode:"allow"}]:[]}(e,t)),!1!==o?.filesystem){if(deepagentsMemoryWritable(e))return n.length>0?n:void 0}else n.push({operations:["read"],paths:["/memories/**"],mode:"allow"}),n.push({operations:["read","write"],paths:["/**"],mode:"deny"});return deepagentsMemoryWritable(e)||n.push({operations:["write"],paths:["/memories/**"],mode:"deny"}),n.length>0?n:void 0}function emitToolEvent(e,t,o,n,r={}){"string"==typeof r.output&&d(e,e.agent.id,t,r.output),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:e.agent.id,event:{adapter:"deepagents",phase:o,toolId:t,..."agent.tool.start"===o?{args:n}:{},...r,..."string"==typeof r.output?u(r.output):{},...f(t,o,n)}})}function backendSkillPath(e,t){const o=function pathRelative(e,t){const o=e.split("/").filter(Boolean),n=t.split("/").filter(Boolean);if(!(n.length<o.length)){for(let e=0;e<o.length;e+=1)if(o[e]!==n[e])return;return n.slice(o.length).join("/")}}(e,t);return void 0===o?t:`/${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 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-evidence.d.ts

@@ -0,0 +1,10 @@
+import type { RuntimeAdapter } from "@stable-harness/core";
+export type ObservedToolEvidence = {
+    agentId: string;
+    toolId: string;
+    output: string;
+};
+export declare function observedToolEvidence(input: Parameters<RuntimeAdapter["run"]>[0]): ObservedToolEvidence[];
+export declare function recordObservedToolEvidence(input: Parameters<RuntimeAdapter["run"]>[0], agentId: string, toolId: string, output: string): void;
+export declare function isControlOnlyOutput(output: string): boolean;
+export declare function isSuccessfulEvidenceOutput(output: string): boolean;

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

@@ -0,0 +1 @@
+const e="deepagents.observedToolEvidence";export function observedToolEvidence(t){const r=t.requestState?.get(e);return Array.isArray(r)?r:[]}export function recordObservedToolEvidence(t,r,o,n){if(!t.requestState||!isSuccessfulEvidenceOutput(n)||isControlOnlyOutput(n))return;const s=observedToolEvidence(t);t.requestState.set(e,[...s,{agentId:r,toolId:o,output:n}].slice(-12))}export function isControlOnlyOutput(e){return/"status"\s*:\s*"(?:duplicate_tool_call|repeated_tool_call_limit|tool_argument_error|delegated_task_repeat_limit|invalid_input|blocked|failed|error|plan_required|dependency_required)"|Status:\s*(?:invalid_input|blocked|failed|error|plan_required|dependency_required)\b|Filesystem builtin tool .* is disabled/iu.test(e)}export function isSuccessfulEvidenceOutput(e){const t=function parseJsonRecord(e){try{const t=JSON.parse(e);return function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}(t)?t:void 0}catch{return}}(e),r="string"==typeof t?.status?t.status:function readTextStatus(e){return String(e).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}(e);return!r||/^(?:completed|success|ok|recorded)$/iu.test(r)}

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

@@ -13,12 +13,6 @@ export type ToolRepeatState = {
     maxCallsByTool: Map<string, number>;
     maxSuccessfulCallsByTool: Map<string, number>;
 };
-export type ObservedToolEvidence = {
-    agentId: string;
-    toolId: string;
-    output: string;
-};
-export declare function observedToolEvidence(input: Parameters<RuntimeAdapter["run"]>[0]): ObservedToolEvidence[];
 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;
 export declare function beforeToolInvoke(toolId: string, args: unknown, state: ToolRepeatState): {

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";const o="deepagents.observedToolEvidence";export function observedToolEvidence(e){const t=e.requestState?.get(o);return Array.isArray(t)?t:[]}export function buildGatewayTools(o,r,s,n,a=createToolRepeatState(o.workspace.runtime.toolGateway)){return o.toolGateway?s.flatMap(s=>{const l=o.toolGateway?.get(s);if(!l)return[];const i=o.workspace.tools.get(s),u=i?.schema??l.schema;return[t(async t=>async function invokeGuardedGatewayTool(t){emitToolResult(t.input,t.agentId,t.toolId,void 0);const o=t.repeatState?beforeToolInvoke(t.toolId,t.args,t.repeatState):void 0;if(o)return emitToolResult(t.input,t.agentId,t.toolId,o.eventOutput),o.modelOutput;const r=await async function invokeGatewayTool(t,o,r,s,n){try{return await t.toolGateway.invoke({toolId:r,args:s,repairModel:n,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),s=r instanceof e?String(r.content):stringifyDeepAgentResult(r.output),n=t.repeatState?afterToolInvoke(t.toolId,t.args,s,r,t.repeatState):{};return emitToolResult(t.input,t.agentId,t.toolId,n.eventOutput??s),void 0!==n.modelOutput?n.modelOutput:r instanceof e?r:s}({input:o,agentId:r,toolId:s,args:t,repairModel:n,repeatState:a}),{name:s,description:buildToolDescription(i?.description??l.description??s,u,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 s=o.maxCallsByTool.get(e)??o.maxCallsPerTool;if(void 0!==s&&r>=s)throw new Error(`Stable Harness repeat limit reached for tool '${e}' after ${s} call(s).`);const n=o.maxSuccessfulCallsByTool.get(e)??o.maxSuccessfulCallsPerTool;if(void 0!==n&&(o.successfulToolCounts.get(e)??0)>=n){const t=repeatedToolCallLimitContent(e,o.latestSuccessfulOutputByTool.get(e));return{eventOutput:t,modelOutput:t}}const a=stableToolCallKey(e,t),l=o.successfulCalls.get(a);if(void 0!==l){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,l);return{eventOutput:r,modelOutput:l}}}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,o,r,s,n){return s instanceof e&&"error"===s.status?{}:isSuccessfulEvidenceOutput(r)?(n.successfulCalls.set(stableToolCallKey(t,o),r),n.latestSuccessfulOutputByTool.set(t,r),n.successfulToolCounts.set(t,(n.successfulToolCounts.get(t)??0)+1),{}):{}}function emitToolResult(e,t,r,s){void 0!==s&&function recordObservedToolEvidence(e,t,r,s){if(!e.requestState||!isSuccessfulEvidenceOutput(s)||function isControlOnlyOutput(e){return/"status"\s*:\s*"(?:duplicate_tool_call|repeated_tool_call_limit|tool_argument_error|delegated_task_repeat_limit|invalid_input|blocked|failed|error)"|Status:\s*(?:invalid_input|blocked|failed|error)\b|Filesystem builtin tool .* is disabled/iu.test(e)}(s))return;const n=observedToolEvidence(e);e.requestState.set(o,[...n,{agentId:t,toolId:r,output:s}].slice(-12))}(e,t,r,s),e.emit({type:"runtime.adapter.event",requestId:e.requestId,sessionId:e.sessionId,agentId:t,event:void 0===s?{adapter:"deepagents",phase:"agent.tool.start",toolId:r}:{adapter:"deepagents",phase:"agent.tool.result",toolId:r,output:previewToolOutput(s),...toolControlProjection(s)}})}function isSuccessfulEvidenceOutput(e){const t=parseJsonRecord(e),o="string"==typeof t?.status?t.status:readTextStatus(e);return!o||/^(?:completed|success|ok|recorded)$/iu.test(o)}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 s=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 ${r} successful time(s) for this request. If more detail is needed, include the dimensions in the first call and synthesize after the result returns.`}(o,r),n=s?`${e}\n\n${s}`:e;return t?`${n}\n\nStable tool input schema:\n${previewToolOutput(JSON.stringify(t))}`:n}function previewToolOutput(e){const t=e.replace(/\s+/gu," ").trim();return t.length>500?`${t.slice(0,497)}...`:t}export function toolControlProjection(e){const t=parseJsonRecord(e);if("string"==typeof t?.status)return{controlStatus:t.status};const o=readTextStatus(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 readTextStatus(e){return String(e).match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function parseJsonRecord(e){try{const t=JSON.parse(e);return isRecord(t)?t:void 0}catch{return}}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=observedToolEvidence(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 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";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)}

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

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

package/packages/core/dist/quality/event-evidence.d.ts

@@ -1,4 +1,5 @@
 import type { RuntimeEvent } from "../types.js";
 export declare function hasPlanningEvidence(events: RuntimeEvent[]): boolean;
 export declare function successfulEvidenceToolIds(events: RuntimeEvent[]): string[];
+export declare function successfulEvidenceOutputs(events: RuntimeEvent[]): string[];
 export declare function controlBlockers(events: RuntimeEvent[]): string[];

package/packages/core/dist/quality/event-evidence.js

@@ -1 +1 @@
-export function hasPlanningEvidence(t){return t.some(t=>{const e=readAdapterEvent(t);return!!e&&("plan"===e.traceType||String(e.traceLabel??"").startsWith("plan.")||"write_todos"===e.toolId&&String(e.phase??"").startsWith("agent.tool."))})}export function successfulEvidenceToolIds(t){const e=t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return[t.toolId];const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId&&function isSuccessfulEvidenceEvent(t){const e=readString(t.controlStatus)??readOutputStatus(t.output);return!e||/^(?:completed|success|ok|recorded)$/iu.test(e)}(e)?[e.toolId]:[]});return[...new Set(e)]}export function controlBlockers(t){return t.flatMap(t=>{const e=readAdapterEvent(t),r=readString(e?.controlStatus)??readOutputStatus(e?.output);return r&&function isBlockerStatus(t){return/^(?:blocked|approval_required|schema_repair_failed|tool_argument_error|invalid_input|repeated_tool_call_limit|duplicate_tool_call)$/iu.test(t)}(r)?[`${readString(e?.toolId)??"tool"}:${r}`]:[]})}function readAdapterEvent(t){return"runtime.adapter.event"===t.type&&isRecord(t.event)?t.event:void 0}function readOutputStatus(t){if("string"!=typeof t)return;const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);return"string"==typeof e?.status?e.status:t.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function readString(t){return"string"==typeof t&&t.trim()?t:void 0}
+export function hasPlanningEvidence(t){return t.some(t=>{const e=readAdapterEvent(t);return!!e&&("plan"===e.traceType||String(e.traceLabel??"").startsWith("plan.")||"write_todos"===e.toolId&&String(e.phase??"").startsWith("agent.tool."))})}export function successfulEvidenceToolIds(t){const e=t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return[t.toolId];const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId&&isSuccessfulEvidenceEvent(e)?[e.toolId]:[]});return[...new Set(e)]}export function successfulEvidenceOutputs(t){return t.flatMap(t=>{if("runtime.tool.direct.completed"===t.type)return stringifyEvidence(t.output);const e=readAdapterEvent(t);return e&&"agent.tool.result"===e.phase&&"string"==typeof e.toolId?!isSuccessfulEvidenceEvent(e)||function isPlanningTool(t){return/^(?:write_todos|read_todos)$/u.test(t)}(e.toolId)?[]:stringifyEvidence(e.output):[]})}export function controlBlockers(t){return t.flatMap(t=>{const e=readAdapterEvent(t),r=readString(e?.controlStatus)??readOutputStatus(e?.output);return r&&function isBlockerStatus(t){return/^(?:blocked|approval_required|schema_repair_failed|tool_argument_error|invalid_input)$/iu.test(t)}(r)?[`${readString(e?.toolId)??"tool"}:${r}`]:[]})}function stringifyEvidence(t){return"string"==typeof t?t.trim()?[t]:[]:null==t?[]:[JSON.stringify(t)]}function readAdapterEvent(t){return"runtime.adapter.event"===t.type&&isRecord(t.event)?t.event:void 0}function isSuccessfulEvidenceEvent(t){const e=readString(t.controlStatus)??readOutputStatus(t.output);return!e||/^(?:completed|success|ok|recorded)$/iu.test(e)}function readOutputStatus(t){if("string"!=typeof t)return;const e=function parseJsonRecord(t){try{const e=JSON.parse(t);return isRecord(e)?e:void 0}catch{return}}(t);return"string"==typeof e?.status?e.status:t.match(/^Status:\s*([A-Za-z0-9_-]+)/imu)?.[1]}function isRecord(t){return"object"==typeof t&&null!==t&&!Array.isArray(t)}function readString(t){return"string"==typeof t&&t.trim()?t:void 0}

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

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

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

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

package/packages/core/dist/quality/recovery-policy.d.ts

@@ -7,4 +7,5 @@ export declare function buildQualityRecoveryRequest(input: {
     policy: QualityPolicy;
     availableToolIds: string[];
     availableSubagentIds: string[];
+    observedEvidence?: string[];
 }): RuntimeRequest | undefined;

package/packages/core/dist/quality/recovery-policy.js

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

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

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

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

@@ -28,6 +28,7 @@ export type QualityPolicy = {
         requireToolEvidence: boolean;
         rejectEmptyFinal: boolean;
         stopOnBlocker: boolean;
+        rejectUngroundedNumbers: boolean;
     };
     recovery: {
         enabled: boolean;