stable-harness 0.0.22 → 0.0.23

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "stable-harness",
-  "version": "0.0.22",
+  "version": "0.0.23",
   "type": "module",
   "description": "Stable application runtime and operator control plane for agent workspaces.",
   "license": "Apache-2.0",
@@ -98,7 +98,7 @@
     "zod": "^4.1.13"
   },
   "devDependencies": {
-    "@types/node": "^24.10.1",
+    "@types/node": "^25.6.2",
     "terser": "^5.47.1",
     "typescript": "^5.9.3"
   }

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

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

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

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

@@ -1 +1 @@
-import{controlBlockers as e,controlGaps as n,successfulEvidenceItems as s}from"./event-evidence.js";export function synthesizeEvidenceOnlyReport(t,o,i){if(!i.enabled||!i.synthesis.enabled||"evidence_only"!==i.synthesis.mode)return;if("pass"===o.verdict||!function hasRecoverableSynthesisIssue(e){return e.issues.some(e=>"control_blocker"!==e.code)}(o))return;const r=s(t.events).slice(-i.synthesis.maxEvidenceItems),c=e(t.events),a=n(t.events);return 0!==r.length||0!==c.length||0!==a.length?["# Grounded report","","Stable Harness rejected the previous final answer because it was not fully supported by observed runtime evidence. The report below is synthesized only from completed tool or delegated-task evidence.","","## Evidence",...evidenceLines(r),"","## Evidence gaps and blockers",...gapLines(c,a)].join("\n"):void 0}function evidenceLines(e){return 0===e.length?["- Evidence gap: no successful tool or delegated-task evidence was available."]:e.map(e=>`- ${e.source}: ${function formatEvidence(e){const n=e.replace(/\s+/gu," ").trim();return n.length>1200?`${n.slice(0,1197)}...`:n}(e.output)}`)}function gapLines(e,n){return 0===e.length&&0===n.length?["- Evidence gap: any requested category not present in the evidence above remains unsupported."]:[...e.map(e=>`- Blocked: ${e}`),...n.map(e=>`- Evidence gap: ${e}`)]}
+import{controlBlockers as e,controlGaps as n,successfulEvidenceItems as t}from"./event-evidence.js";export function synthesizeEvidenceOnlyReport(s,r,i){if(!i.enabled||!i.synthesis.enabled||"evidence_only"!==i.synthesis.mode)return;if("pass"===r.verdict||!function hasRecoverableSynthesisIssue(e){return e.issues.some(e=>"control_blocker"!==e.code)}(r))return;const o=t(s.events).slice(-i.synthesis.maxEvidenceItems),c=e(s.events),u=n(s.events);if(0===o.length&&0===c.length&&0===u.length)return;const a=function detectSynthesisLanguage(e){const n=e.workspace.runtime.responseLanguage;if(function isRecord(e){return"object"==typeof e&&null!==e&&!Array.isArray(e)}(n)){const e=readString(n.language)??readString(n.locale)??readString(n.target);if(e&&/^(?:zh|zh-|chinese|中文|汉语|漢語)/iu.test(e))return"zh";if(e&&/^(?:en|en-|english)$/iu.test(e))return"en"}return/\p{Script=Han}/u.test(e.request.input)?"zh":"en"}(s);return"zh"===a?function buildChineseReport(e,n,t){return["# 有依据的报告","","Stable Harness 已拒绝上一版最终回答，因为它未被已观察到的运行时证据完全支持。以下报告只根据已完成的工具或委托任务证据生成。","","## 已完成的证据来源",...sourceLines(e,"zh"),"","## 有证据支持的事实",...evidenceLines(e,"zh"),"","## 证据缺口与阻塞",...gapLines(n,t,"zh")].join("\n")}(o,c,u):function buildEnglishReport(e,n,t){return["# Grounded report","","Stable Harness rejected the previous final answer because it was not fully supported by observed runtime evidence. This report is synthesized only from completed tool or delegated-task evidence.","","## Completed evidence sources",...sourceLines(e,"en"),"","## Evidence-backed facts",...evidenceLines(e,"en"),"","## Evidence gaps and blockers",...gapLines(n,t,"en")].join("\n")}(o,c,u)}function sourceLines(e,n){return 0===e.length?["zh"===n?"- 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."]:[...e.reduce((e,n)=>e.set(n.source,(e.get(n.source)??0)+1),new Map)].map(([e,n])=>`- ${e}${n>1?` (${n})`:""}`)}function evidenceLines(e,n){return 0===e.length?["zh"===n?"- 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."]:e.map(e=>`- ${e.source}: ${function formatEvidence(e){const n=e.replace(/\s+/gu," ").trim();return n.length>1200?`${n.slice(0,1197)}...`:n}(e.output)}`)}function gapLines(e,n,t){return 0===e.length&&0===n.length?["zh"===t?"- 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"===t?[...e.map(e=>`- 阻塞：${e}`),...n.map(e=>`- 证据缺口：${e}`)]:[...e.map(e=>`- Blocked: ${e}`),...n.map(e=>`- Evidence gap: ${e}`)]}function readString(e){return"string"==typeof e&&e.trim()?e.trim():void 0}

package/packages/protocols/dist/src/http-server.d.ts

@@ -1,3 +1,3 @@
 import { type IncomingMessage, type ServerResponse } from "node:http";
 import { type StableHarnessRuntime } from "@stable-harness/core";
-export declare function createHttpServer(runtime: StableHarnessRuntime): import("http").Server<typeof IncomingMessage, typeof ServerResponse>;
+export declare function createHttpServer(runtime: StableHarnessRuntime): import("node:http").Server<typeof IncomingMessage, typeof ServerResponse>;

package/packages/protocols/dist/src/openai-compatible.d.ts

@@ -6,4 +6,4 @@ export type OpenAiCompatibleServerOptions = {
     modelAgentMap?: Record<string, string>;
     defaultModel?: string;
 };
-export declare function createOpenAiCompatibleHttpServer(runtime: StableHarnessRuntime, options?: OpenAiCompatibleServerOptions): import("http").Server<typeof IncomingMessage, typeof ServerResponse>;
+export declare function createOpenAiCompatibleHttpServer(runtime: StableHarnessRuntime, options?: OpenAiCompatibleServerOptions): import("node:http").Server<typeof IncomingMessage, typeof ServerResponse>;