stable-harness 0.0.38 → 0.0.39

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

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

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

@@ -34,12 +34,14 @@ export type QualityPolicy = {
         enabled: boolean;
         mode: "evidence_only";
         maxEvidenceItems: number;
+        fieldKinds?: Partial<Record<FactKind, string[]>>;
     };
     recovery: {
         enabled: boolean;
         maxLoops: number;
     };
 };
+export type FactKind = "context" | "data" | "timeBound" | "limit" | "other";
 export type QualityReviewInput = {
     workspace: CompiledWorkspace;
     agent: WorkspaceAgent;

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

@@ -5,6 +5,7 @@ import type { CompiledWorkspace, WorkspaceAgent } from "../workspace/types.js";
 export type RuntimeCapabilityState = {
     memory?: RuntimeMemoryContext;
     pluginMemories?: RuntimeMemoryContext[];
+    modules: Record<string, Record<string, unknown>>;
 };
 export type RuntimeCapabilityContext = {
     workspace: CompiledWorkspace;

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

@@ -1 +1 @@
-export function createRuntimeCapabilityRegistry(e){const t=e.filter(e=>Boolean(e)),o={};return{modules:t,state:o,emitSideEffects(e,o){for(const r of t)try{r.onEvent?.(e,o)}catch{continue}},async beforeAdapterRun(e){for(const r of t){const t=await(r.beforeAdapterRun?.({...e,state:o}));mergeCapabilityState(o,t)}return o},async beforeAdapterResultContract(e){for(const r of t)await(r.beforeAdapterResultContract?.({...e,state:o}))},async afterAdapterResponse(e){for(const r of t)await(r.afterAdapterResponse?.({...e,state:o}))},async stop(){await Promise.all(t.map(e=>e.stop?.()))}}}export function normalizeAdapterResult(e){return"string"==typeof e?{text:e}:e}function mergeCapabilityState(e,t){t&&(void 0!==t.memory&&(e.memory=t.memory),t.pluginMemories&&(e.pluginMemories=[...e.pluginMemories??[],...t.pluginMemories]))}
+export function createRuntimeCapabilityRegistry(e){const t=e.filter(e=>Boolean(e)),o={modules:{}};return{modules:t,state:o,emitSideEffects(e,o){for(const r of t)try{r.onEvent?.(e,o)}catch{continue}},async beforeAdapterRun(e){for(const r of t){const t=await(r.beforeAdapterRun?.({...e,state:o}));mergeCapabilityState(o,t)}return o},async beforeAdapterResultContract(e){for(const r of t)await(r.beforeAdapterResultContract?.({...e,state:o}))},async afterAdapterResponse(e){for(const r of t)await(r.afterAdapterResponse?.({...e,state:o}))},async stop(){await Promise.all(t.map(e=>e.stop?.()))}}}export function normalizeAdapterResult(e){return"string"==typeof e?{text:e}:e}function mergeCapabilityState(e,t){t&&(void 0!==t.memory&&(e.memory=t.memory),t.pluginMemories&&(e.pluginMemories=[...e.pluginMemories??[],...t.pluginMemories]),t.modules&&(e.modules=function mergeModuleState(e,t){const o={...e};for(const[e,r]of Object.entries(t))o[e]={...o[e]??{},...r};return o}(e.modules,t.modules)))}

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

@@ -0,0 +1,11 @@
+import type { CompiledWorkspace, RuntimeRequest, WorkspaceAgent } from "../../types.js";
+export type RuntimePolicyProjectionInput = {
+    workspace: CompiledWorkspace;
+    agent?: WorkspaceAgent;
+    request: RuntimeRequest;
+};
+export type RuntimePolicyProjection = {
+    systemPromptSections: string[];
+};
+export declare function projectRuntimePolicies(input: RuntimePolicyProjectionInput): RuntimePolicyProjection;
+export declare function buildRuntimeSystemPrompt(input: RuntimePolicyProjectionInput, basePrompt?: string): string | undefined;

package/packages/core/dist/runtime/policy/projection.js

@@ -0,0 +1 @@
+export function projectRuntimePolicies(e){return{systemPromptSections:[skillInventoryPolicy(e),responseLanguagePolicy(e),responsePresentationPolicy(e)].filter(e=>Boolean(e))}}export function buildRuntimeSystemPrompt(e,t){const n=[t,...projectRuntimePolicies(e).systemPromptSections].filter(e=>Boolean(e));return n.length>0?n.join("\n\n"):void 0}function skillInventoryPolicy(e){const t=(e.agent?.skills??[]).map(t=>e.workspace.skills.get(t)).filter(e=>Boolean(e));if(0!==t.length)return["## Stable Harness Skill Inventory","Use only the skills listed in this workspace inventory. Do not infer or invent skill names, skill directories, or SKILL.md paths from generic examples.","Skill files are already registered with the backend. Do not read local SKILL.md files or skill directories through filesystem tools.","If none of these skills match the task, continue with the configured tools and collected evidence instead of reading an unlisted skill path.",t.map(e=>{const t=e.allowedTools.length>0?`; allowed tools: ${e.allowedTools.join(", ")}`:"",n=e.description?`: ${e.description}${t}`:t;return`- ${e.id}${n}`}).join("\n")].join("\n")}function responseLanguagePolicy(e){const t=readRecord(e.workspace.runtime.responseLanguage);if("matchUser"!==(readString(t.mode)??readString(t.strategy)))return;const n=function detectRequestLanguage(e){return/\p{Script=Han}/u.test(e)?"Chinese":/[\p{Script=Hiragana}\p{Script=Katakana}]/u.test(e)?"Japanese":/\p{Script=Hangul}/u.test(e)?"Korean":/\p{Script=Arabic}/u.test(e)?"Arabic":/\p{Script=Hebrew}/u.test(e)?"Hebrew":/\p{Script=Thai}/u.test(e)?"Thai":/\p{Script=Cyrillic}/u.test(e)?"Cyrillic-script language":void 0}(e.request.input);return["Stable runtime response language policy:",...n?[`- Detected request language: ${n}.`]:[],"- Match the final answer language to the original user request unless the user explicitly asks for another language.","- If tool or subagent evidence is in a different language, translate the final user-facing answer into the detected request language.","- When delegating to subagents, include the same response-language requirement in delegated instructions.","- Do not call another tool or subagent only to translate, rewrite, format, or synthesize a completed answer.",`Original user request:\n${e.request.input}`].join("\n")}function responsePresentationPolicy(e){const t=readRecord(e.workspace.runtime.responsePresentation);if(!0===t.enabled&&!function hasStructuredResponseFormat(e){return void 0!==e?.config.responseFormat||isRecord(e?.config.deepagents)&&void 0!==e.config.deepagents.responseFormat}(e.agent))return"markdown"===(readString(t.style)??"markdown")?["Stable runtime final-answer presentation policy:","- For user-facing natural-language final answers, use GitHub-flavored Markdown.","- Prefer clear section headings, short paragraphs, and concise bullets over dense prose.","- Use tables only when they make comparison or planning details easier to scan.","- For detailed investigations, plans, or reports, include assumptions, findings, recommendations, and concrete next steps.","- Preserve exact plain text, JSON, code, or other structured output when the user or response format asks for it.","- Do not end with generic follow-up offers; deliver the requested answer directly."].join("\n"):void 0}function readRecord(e){return isRecord(e)?e:{}}function readString(e){return"string"==typeof e&&e.trim()?e:void 0}function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}

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

@@ -0,0 +1,45 @@
+import type { RuntimeEvent, RuntimeRequest, WorkspaceAgent } from "../../types.js";
+export type ToolRepeatState = {
+    successfulCalls: Map<string, string>;
+    duplicateCallCounts: Map<string, number>;
+    latestSuccessfulOutputByTool: Map<string, string>;
+    successfulToolCounts: Map<string, number>;
+    toolCallCounts: Map<string, number>;
+    maxDuplicateCallsPerTool?: number;
+    maxCallsPerTool?: number;
+    maxSuccessfulCallsPerTool?: number;
+    maxCallsByTool: Map<string, number>;
+    maxSuccessfulCallsByTool: Map<string, number>;
+    returnPreviousOutputOnRepeatLimit: boolean;
+};
+export type ToolRepeatDecision = {
+    eventOutput: string;
+    modelOutput: string;
+};
+export declare function createToolRepeatState(config: unknown): ToolRepeatState | undefined;
+export declare function beforeToolInvoke(toolId: string, args: unknown, state: ToolRepeatState): ToolRepeatDecision | undefined;
+export declare function afterToolInvoke(input: {
+    toolId: string;
+    args: unknown;
+    output: string;
+    successful: boolean;
+    state: ToolRepeatState;
+}): {
+    eventOutput?: string;
+    modelOutput?: string;
+};
+export declare function isToolRepeatLimitReached(toolId: string, state: ToolRepeatState | undefined): boolean;
+export declare function missingRequiredPlanContent(input: {
+    agent: WorkspaceAgent;
+    events: RuntimeEvent[];
+    toolId: string;
+}): string;
+export declare function missingToolDependencyContent(input: {
+    agent: WorkspaceAgent;
+    events: RuntimeEvent[];
+    toolId: string;
+}): string;
+export declare function toolInvocationEvents(input: {
+    request: RuntimeRequest;
+    getEvents?: () => RuntimeEvent[];
+}): RuntimeEvent[];

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

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

package/packages/core/dist/runtime.js

@@ -1 +1 @@
-import{randomUUID as e}from"node:crypto";import{assertExecutionContract as t}from"./execution-contract.js";import{assertNoRawToolCallOutput as r,assertNoToolExecutionErrorOutput as a,buildAdapterErrorRecoveryPrompt as o,buildEvidenceSynthesisOutput as s,buildExecutionContractRecoveryRequest as n,buildResultRecoveryRequest as i,containsRawToolCallOutput as u,isRecoverableAdapterError as c,rawToolCallFailureMessage as l,toolCallRecoveryEnabled as p}from"./recovery/tool-call.js";import{recoverQualityReview as d,resolveQualityPolicy as m}from"./quality/index.js";import{completeRun as w,failRun as y}from"./runtime/completion.js";import{runDirectToolCall as f}from"./runtime/direct-tool-call.js";import{createApprovalGatedToolGateway as g}from"./runtime/governance/approval-gate.js";import{createSandboxedToolGateway as I}from"./runtime/governance/sandbox.js";import{createRuntimeInspectionMethods as q}from"./runtime/inspection/methods.js";import{createRuntimeCapabilityRegistry as R,normalizeAdapterResult as v}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as b}from"./runtime/memory.js";import{createInMemoryRuntimeStore as k}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 h}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const r=new Set,a=t.store??k(),s=R([b(t),A({options:t.progressNarration,policy:t.workspace.runtime}),...t.capabilities??[]]),emitBase=e=>{a.appendEvent(e);for(const t of r)t(e)},emit=e=>{emitBase(e),s.emitSideEffects(e,emitBase)},i=I({gateway:g({gateway:t.toolGateway,approvals:t.approvals,workspace:t.workspace,emit:emit}),workspace:t.workspace,sandbox:t.sandbox,emit:emit}),u={...t,toolGateway:i},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 t=>async function runRuntimeRequest(t){const r=t.request.requestId??e(),a=t.request.sessionId??e(),s=[],{agent:i,adapter:u}=await async function resolveExecution(e,t,r){const a=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const a=await 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 a.ok?a.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(a);if(!o)throw new Error(`Agent ${a} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:o,adapter:void 0};const s=e.adapters.find(e=>e.canRun(o));if(!s)throw new Error(`No runtime adapter can run backend ${o.backend} for agent ${o.id}`);return{agent:o,adapter:s}}(t.input,t.request,{requestId:r,sessionId:a,emit:e=>s.push(e)});t.store.createRun(function createRunRecord(e,t,r,a){return{requestId:t,sessionId:r,agentId:a.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,r,a,i)),s.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:r,sessionId:a,agentId:i.id,input:t.request.input});try{if(t.request.workflow){const e=await h({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:r,sessionId:a,agentId:i.id,emit:t.emit});return w({store:t.store,emit:t.emit,requestId:r,sessionId:a,agent:i,result:e,artifacts:t.input.artifacts})}if(t.request.toolCall){const e=await f({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:r,sessionId:a,agent:i,toolFailureTracker:t.toolFailureTracker});return w({store:t.store,emit:t.emit,requestId:r,sessionId:a,agent:i,result:e,artifacts:t.input.artifacts})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,r=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),a=r.memory,s=r.pluginMemories??[],i=e.input.workspace.runtime,u=m(e.input.workspace.runtime,e.agent),l=new Map;let p;try{p=await runAdapterOnce(e,t,e.request,a,s,l)}catch(r){if(!c(r,i))throw r;p=await runAdapterOnce(e,t,o(e.request,r,i),a,s,l)}p=await recoverAdapterResultOutput(e,t,e.request,p,a,s,i,l),p=await d(createQualityRuntimeInput(e,a,s,l),e.request,p,u),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:p});try{assertRequestExecutionContract(e)}catch(r){const o=n({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:i});if(!o)throw r;p=await runAdapterOnce(e,t,o,a,s,l),p=await recoverAdapterResultOutput(e,t,o,p,a,s,i,l),p=await d(createQualityRuntimeInput(e,a,s,l),o,p,u),assertRequestExecutionContract(e)}const y=w({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:p,artifacts:e.input.artifacts});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:p,response:y}),y}({...t,adapter:u,requestId:r,sessionId:a,agent:i})}catch(e){return y({store:t.store,emit:t.emit,requestId:r,sessionId:a,agent:i,error:e})}}({input:u,capabilities:s,store:a,emit:emit,request:t,toolFailureTracker:l}),subscribe:e=>(r.add(e),()=>r.delete(e)),...q({workspace:t.workspace,store:a,artifacts:t.artifacts,approvals:t.approvals,emit:emit}),cancel(e,t){const r=a.getRun(e);r&&"running"===r.state&&(a.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 s.stop(),r.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,a){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,a),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}async function recoverAdapterResultOutput(e,t,o,n,c,d,m,w){let y=o;const f=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,a="object"!=typeof r||null===r||Array.isArray(r)?void 0:r.maxResultRecoveryAttempts;return"number"==typeof a&&Number.isInteger(a)&&a>0?a:3}(m);let g=0;for(let r=0;r<f;r+=1){const r=e.store.getRun(e.requestId)?.events??[],a=i({request:y,output:n.text,events:r.slice(g),availableToolIds:e.agent.tools,policy:m});if(!a)break;y=a,g=e.store.getRun(e.requestId)?.events.length??0,n=await runAdapterOnce(e,t,a,c,d,w)}if(p(m)){let t=!1;u(n.text,m)&&function rawToolCallFailureReturnsMessage(e){return"message"===("object"!=typeof e?.toolCallRecovery||null===e.toolCallRecovery||Array.isArray(e.toolCallRecovery)?{}:e.toolCallRecovery).onFailure}(o.metadata)&&(n={...n,text:l(),metadata:{...n.metadata,toolCallRecovery:{failed:!0,reason:"raw_tool_call_output"}}});const i=s({request:o,output:n.text,events:e.store.getRun(e.requestId)?.events??[],policy:m});i&&(t=!0,n={...n,text:i,metadata:{...n.metadata,toolCallRecovery:{synthesized:!0,reason:"raw_tool_call_output_with_evidence"}}}),t||(r(n.text,m),a(n.text,m))}return n}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,a,o,s){return v(await t.run({workspace:e.input.workspace,agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:a,pluginMemories:o,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,requestState:s,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{assertNoRawToolCallOutput as r,assertNoToolExecutionErrorOutput as a,buildAdapterErrorRecoveryPrompt as s,buildEvidenceSynthesisOutput as o,buildExecutionContractRecoveryRequest as n,buildResultRecoveryRequest as i,containsRawToolCallOutput as u,isRecoverableAdapterError as c,rawToolCallFailureMessage as l,toolCallRecoveryEnabled as p}from"./recovery/tool-call.js";import{recoverQualityReview as d,resolveQualityPolicy as m}from"./quality/index.js";import{completeRun as w,failRun as y}from"./runtime/completion.js";import{runDirectToolCall as f}from"./runtime/direct-tool-call.js";import{createApprovalGatedToolGateway as g}from"./runtime/governance/approval-gate.js";import{createSandboxedToolGateway as I}from"./runtime/governance/sandbox.js";import{createRuntimeInspectionMethods as q}from"./runtime/inspection/methods.js";import{createRuntimeCapabilityRegistry as R,normalizeAdapterResult as v}from"./runtime/capabilities.js";import{createMemoryRuntimeCapability as b}from"./runtime/memory.js";import{createInMemoryRuntimeStore as k}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 h}from"./workflows/runtime.js";export function createStableHarnessRuntime(t){const r=new Set,a=t.store??k(),o=R([b(t),A({options:t.progressNarration,policy:t.workspace.runtime}),...t.capabilities??[]]),emitBase=e=>{a.appendEvent(e);for(const t of r)t(e)},emit=e=>{emitBase(e),o.emitSideEffects(e,emitBase)},i=I({gateway:g({gateway:t.toolGateway,approvals:t.approvals,workspace:t.workspace,emit:emit}),workspace:t.workspace,sandbox:t.sandbox,emit:emit}),u={...t,toolGateway:i},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 t=>async function runRuntimeRequest(t){const r=t.request.requestId??e(),a=t.request.sessionId??e(),o=[],{agent:i,adapter:u}=await async function resolveExecution(e,t,r){const a=t.agentId?await async function resolveRequestedAgentId(e,t,r){if(e.agents.has(t))return t;const a=await 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 a.ok?a.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),s=e.workspace.agents.get(a);if(!s)throw new Error(`Agent ${a} is not defined in the workspace`);if(t.toolCall||t.workflow)return{agent:s,adapter:void 0};const o=e.adapters.find(e=>e.canRun(s));if(!o)throw new Error(`No runtime adapter can run backend ${s.backend} for agent ${s.id}`);return{agent:s,adapter:o}}(t.input,t.request,{requestId:r,sessionId:a,emit:e=>o.push(e)});t.store.createRun(function createRunRecord(e,t,r,a){return{requestId:t,sessionId:r,agentId:a.id,input:e.input,state:"running",parentRunId:e.parentRunId,metadata:e.metadata,artifacts:[],startedAt:(new Date).toISOString(),events:[]}}(t.request,r,a,i)),o.forEach(t.emit),t.emit({type:"runtime.request.started",requestId:r,sessionId:a,agentId:i.id,input:t.request.input});try{if(t.request.workflow){const e=await h({workspace:t.input.workspace,adapters:t.input.workflowAdapters??[],toolGateway:t.input.toolGateway,request:{input:t.request.input,...t.request.workflow},requestId:r,sessionId:a,agentId:i.id,emit:t.emit});return w({store:t.store,emit:t.emit,requestId:r,sessionId:a,agent:i,result:e,artifacts:t.input.artifacts})}if(t.request.toolCall){const e=await f({gateway:t.input.toolGateway,workspace:t.input.workspace,emit:t.emit,request:t.request,requestId:r,sessionId:a,agent:i,toolFailureTracker:t.toolFailureTracker});return w({store:t.store,emit:t.emit,requestId:r,sessionId:a,agent:i,result:e,artifacts:t.input.artifacts})}return await async function runAdapterRequest(e){if(!e.adapter)throw new Error(`No runtime adapter can run backend ${e.agent.backend} for agent ${e.agent.id}`);const t=e.adapter,r=await e.capabilities.beforeAdapterRun(createCapabilityContext(e)),a=r.memory,o=r.pluginMemories??[],i=e.input.workspace.runtime,u=m(e.input.workspace.runtime,e.agent),l=new Map;let p;try{p=await runAdapterOnce(e,t,e.request,a,o,l)}catch(r){if(!c(r,i))throw r;p=await runAdapterOnce(e,t,s(e.request,r,i),a,o,l)}p=await recoverAdapterResultOutput(e,t,e.request,p,a,o,i,l),p=await d(createQualityRuntimeInput(e,a,o,l),e.request,p,u),await e.capabilities.beforeAdapterResultContract({...createCapabilityContext(e),result:p});try{assertRequestExecutionContract(e)}catch(r){const s=n({request:e.request,events:e.store.getRun(e.requestId)?.events??[],policy:i});if(!s)throw r;p=await runAdapterOnce(e,t,s,a,o,l),p=await recoverAdapterResultOutput(e,t,s,p,a,o,i,l),p=await d(createQualityRuntimeInput(e,a,o,l),s,p,u),assertRequestExecutionContract(e)}const y=w({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,result:p,artifacts:e.input.artifacts});return await e.capabilities.afterAdapterResponse({...createCapabilityContext(e),result:p,response:y}),y}({...t,adapter:u,requestId:r,sessionId:a,agent:i})}catch(e){return y({store:t.store,emit:t.emit,requestId:r,sessionId:a,agent:i,error:e})}}({input:u,capabilities:o,store:a,emit:emit,request:t,toolFailureTracker:l}),subscribe:e=>(r.add(e),()=>r.delete(e)),...q({workspace:t.workspace,store:a,artifacts:t.artifacts,approvals:t.approvals,emit:emit}),cancel(e,t){const r=a.getRun(e);r&&"running"===r.state&&(a.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 o.stop(),r.clear()}}}function createCapabilityContext(e){return{workspace:e.input.workspace,store:e.store,emit:e.emit,request:e.request,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent}}function createQualityRuntimeInput(e,t,r,a){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:s=>runAdapterOnce(e,e.adapter,s,t,r,a),reviewModel:e.input.qualityReviewModel,memory:t,pluginMemories:r}}async function recoverAdapterResultOutput(e,t,s,n,c,d,m,w){let y=s;const f=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,a="object"!=typeof r||null===r||Array.isArray(r)?void 0:r.maxResultRecoveryAttempts;return"number"==typeof a&&Number.isInteger(a)&&a>0?a:3}(m);let g=0;for(let r=0;r<f;r+=1){const r=e.store.getRun(e.requestId)?.events??[],a=i({request:y,output:n.text,events:r.slice(g),availableToolIds:e.agent.tools,policy:m});if(!a)break;y=a,g=e.store.getRun(e.requestId)?.events.length??0,n=await runAdapterOnce(e,t,a,c,d,w)}if(p(m)){let t=!1;u(n.text,m)&&function rawToolCallFailureReturnsMessage(e){return"message"===("object"!=typeof e?.toolCallRecovery||null===e.toolCallRecovery||Array.isArray(e.toolCallRecovery)?{}:e.toolCallRecovery).onFailure}(s.metadata)&&(n={...n,text:l(),metadata:{...n.metadata,toolCallRecovery:{failed:!0,reason:"raw_tool_call_output"}}});const i=o({request:s,output:n.text,events:e.store.getRun(e.requestId)?.events??[],policy:m});i&&(t=!0,n={...n,text:i,metadata:{...n.metadata,toolCallRecovery:{synthesized:!0,reason:"raw_tool_call_output_with_evidence"}}}),t||(r(n.text,m),a(n.text,m))}return n}function assertRequestExecutionContract(e){t({store:e.store,emit:e.emit,requestId:e.requestId,sessionId:e.sessionId,agent:e.agent,metadata:e.request.metadata})}async function runAdapterOnce(e,t,r,a,s,o){return v(await t.run({workspace:e.input.workspace,agent:e.agent,request:r,requestId:e.requestId,sessionId:e.sessionId,memory:a,pluginMemories:s,toolGateway:e.input.toolGateway,toolFailureTracker:e.input.toolFailureTracker,requestState:o,getEvents:()=>e.store.getRun(e.requestId)?.events??[],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.d.ts

@@ -10,6 +10,12 @@ export type RuntimeTraceEntry = {
     requestId: string;
     detail?: Record<string, unknown>;
 };
+export type RuntimeAdapterTraceProjection = {
+    traceType: RuntimeTraceEntry["type"];
+    traceLabel: string;
+} & Record<string, unknown>;
+export declare function createDelegationTraceProjection(label: "delegation.start" | "delegation.completed", detail?: Record<string, unknown>): RuntimeAdapterTraceProjection;
+export declare function createPlanTraceProjection(label: "plan.updated", detail?: Record<string, unknown>): RuntimeAdapterTraceProjection;
 export declare function projectRuntimeTrace(run: RuntimeRunRecord): RuntimeTraceEntry[];
 export declare function projectEvent(event: RuntimeEvent): RuntimeTraceEntry | undefined;
 export declare function readPlanTodos(output: unknown): PlanTodoItem[];

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.sandbox.decision"===e.type?base(e,"tool","runtime.sandbox.decision",{toolId:e.toolId,...e.decision}):"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}
+export function createDelegationTraceProjection(e,t={}){return{traceType:"delegation",traceLabel:e,...t}}export function createPlanTraceProjection(e,t={}){return{traceType:"plan",traceLabel:e,...t}}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.sandbox.decision"===e.type?base(e,"tool","runtime.sandbox.decision",{toolId:e.toolId,...e.decision}):"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

@@ -29,6 +29,7 @@ export type RuntimeAdapterContext = {
     toolGateway?: RuntimeToolGateway;
     toolFailureTracker?: RuntimeToolFailureTracker;
     requestState?: Map<string, unknown>;
+    getEvents?: () => RuntimeEvent[];
     emit: RuntimeEmit;
 };
 export type RuntimeMemoryContext = {