agentic-qe 3.12.1 → 3.12.3

package/dist/cli/chunks/{witness-chain-XBEPT4XD.js → witness-chain-GTITDIA2.js}

@@ -1,2 +1,2 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b,c,d,e,f,g}from"./chunk-BKKJEKNT.js";import"./chunk-FNMEVLSN.js";import"./chunk-NOCPO55P.js";import"./chunk-USTUWUCE.js";import"./chunk-U5ITD7ZD.js";export{c as PersistentWitnessChain,a as SPRTAccumulator,b as WitnessChain,f as createPersistentWitnessChain,e as createWitnessChain,d as createWitnessChainSQLitePersistence,g as isWitnessChainFeatureEnabled};
+import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.3");process.exit(0)}
+import{a,b,c,d,e,f,g}from"./chunk-GV2BS5NC.js";import"./chunk-G5OGKVLO.js";import"./chunk-4EHW4TNA.js";import"./chunk-XUUVAAHV.js";import"./chunk-SGMJ277D.js";export{c as PersistentWitnessChain,a as SPRTAccumulator,b as WitnessChain,f as createPersistentWitnessChain,e as createWitnessChain,d as createWitnessChainSQLitePersistence,g as isWitnessChainFeatureEnabled};

package/dist/cli/chunks/witness-chain-QNFPV4UL.js

@@ -0,0 +1,2 @@
+import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.3");process.exit(0)}
+import{a,b,c,d,e,f,g,h,i}from"./chunk-DZ775UGW.js";import"./chunk-SPI4LUJ6.js";import"./chunk-NOSQI2OW.js";import"./chunk-FJACKKRL.js";import"./chunk-L2X3BWMN.js";import"./chunk-G5OGKVLO.js";import"./chunk-4EHW4TNA.js";import"./chunk-NJDLJ3KD.js";import"./chunk-OBL6W4Z5.js";import"./chunk-2LNDU7AW.js";import"./chunk-ULQCEM7K.js";import"./chunk-UAPUA6OY.js";import"./chunk-GS6G7AVS.js";import"./chunk-YUDGQZJN.js";import"./chunk-HJ4U64BL.js";import"./chunk-CYSBM52D.js";import"./chunk-NLDLUHOF.js";import"./chunk-PEHEANAM.js";import"./chunk-XUUVAAHV.js";import"./chunk-SGMJ277D.js";export{a as GENESIS_PREV_HASH,f as WitnessChain,i as _resetWitnessChainForTests,h as createWitnessChain,g as getWitnessChain,d as hashWith,e as serializeEntry,b as sha256,c as shake256};

package/dist/cli/chunks/{witness-key-manager-DAAMNYOW.js → witness-key-manager-AHRM2M52.js}

@@ -1,2 +1,2 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b,c}from"./chunk-OXVAGR64.js";import"./chunk-IE6FDICX.js";import"./chunk-U5ITD7ZD.js";export{a as WitnessKeyManager,c as _resetDefaultWitnessKeyManagerForTests,b as getDefaultWitnessKeyManager};
+import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.3");process.exit(0)}
+import{a,b,c}from"./chunk-SPI4LUJ6.js";import"./chunk-NJDLJ3KD.js";import"./chunk-SGMJ277D.js";export{a as WitnessKeyManager,c as _resetDefaultWitnessKeyManagerForTests,b as getDefaultWitnessKeyManager};

package/dist/cli/chunks/{workflow-SZOOU6D3.js → workflow-XCGHEL3I.js}

@@ -1,5 +1,5 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import"./chunk-X2K3BTI4.js";import{b as $}from"./chunk-45QNHMTP.js";import{d as R}from"./chunk-NEXJUTHF.js";import{b as k,c as C,d as b}from"./chunk-NK2Z6FGF.js";import{c as W}from"./chunk-O65OPAZH.js";import"./chunk-4LKQZW2H.js";import"./chunk-UBPJQKYS.js";import"./chunk-U5ITD7ZD.js";import{Command as O}from"commander";import o from"chalk";function P(d,a){let m=d.indexOf("=");return m>0&&(a[d.substring(0,m)]=d.substring(m+1)),a}function x(d,a,m){let y=new O("workflow").description("Manage QE workflows and pipelines (ADR-041)");return y.command("run <file>").description("Execute a QE pipeline from YAML file").option("-w, --watch","Watch execution progress").option("-v, --verbose","Show detailed output").option("--params <json>","Additional parameters as JSON","{}").action(async(t,l)=>{if(!await m())return;let e=await import("fs"),c=(await import("path")).resolve(t);try{console.log(o.blue(`
+import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.3");process.exit(0)}
+import"./chunk-DMD6WVJ2.js";import{b as $}from"./chunk-FTM562PF.js";import{d as R}from"./chunk-S47YGWXK.js";import{b as k,c as C,d as b}from"./chunk-BEPP57C2.js";import{c as W}from"./chunk-OBL6W4Z5.js";import"./chunk-NKTLKMUB.js";import"./chunk-2LNDU7AW.js";import"./chunk-SGMJ277D.js";import{Command as O}from"commander";import o from"chalk";function P(d,a){let m=d.indexOf("=");return m>0&&(a[d.substring(0,m)]=d.substring(m+1)),a}function x(d,a,m){let y=new O("workflow").description("Manage QE workflows and pipelines (ADR-041)");return y.command("run <file>").description("Execute a QE pipeline from YAML file").option("-w, --watch","Watch execution progress").option("-v, --verbose","Show detailed output").option("--params <json>","Additional parameters as JSON","{}").action(async(t,l)=>{if(!await m())return;let e=await import("fs"),c=(await import("path")).resolve(t);try{console.log(o.blue(`
  Running workflow from: ${t}
 `));let s=k(c);if(!s.success||!s.workflow){console.log(o.red("Failed to parse pipeline:"));for(let w of s.errors)console.log(o.red(`   ${w}`));await a(1)}let r={...W(l.params,"params")};if(s.pipeline){for(let w of s.pipeline.stages)if(w.params)for(let[S,h]of Object.entries(w.params))r[S]=h}if(!d.workflowOrchestrator.getWorkflow(s.workflow.id)){let w=d.workflowOrchestrator.registerWorkflow(s.workflow);w.success||(console.log(o.red(`Failed to register workflow: ${w.error.message}`)),await a(1))}let n=await d.workflowOrchestrator.executeWorkflow(s.workflow.id,r);if(!n.success){console.log(o.red(`Failed to start workflow: ${n.error.message}`)),await a(1);return}let p=n.value;if(console.log(o.cyan(`  Execution ID: ${p}`)),console.log(o.gray(`  Workflow: ${s.workflow.name}`)),console.log(o.gray(`  Stages: ${s.workflow.steps.length}`)),console.log(""),l.watch){console.log(o.blue(`Workflow Progress:
 `));let w,S=Date.now();for(;;){let g=d.workflowOrchestrator.getWorkflowStatus(p);if(!g)break;if(!w||w.progress!==g.progress||w.status!==g.status||JSON.stringify(w.currentSteps)!==JSON.stringify(g.currentSteps)){process.stdout.write("\r\x1B[K");let v="\u2588".repeat(Math.floor(g.progress/5))+"\u2591".repeat(20-Math.floor(g.progress/5)),D=g.status==="completed"?o.green:g.status==="failed"?o.red:g.status==="running"?o.yellow:o.gray;console.log(`  [${v}] ${g.progress}% - ${D(g.status)}`),g.currentSteps.length>0&&l.verbose&&console.log(o.gray(`    Running: ${g.currentSteps.join(", ")}`)),w=g}if(g.status==="completed"||g.status==="failed"||g.status==="cancelled")break;await new Promise(v=>setTimeout(v,500))}let h=d.workflowOrchestrator.getWorkflowStatus(p);if(h){console.log("");let g=h.duration||Date.now()-S;h.status==="completed"?(console.log(o.green("Workflow completed successfully")),console.log(o.gray(`   Duration: ${$(g)}`)),console.log(o.gray(`   Completed: ${h.completedSteps.length} stages`)),h.skippedSteps.length>0&&console.log(o.yellow(`   Skipped: ${h.skippedSteps.length} stages`))):h.status==="failed"?(console.log(o.red("Workflow failed")),console.log(o.red(`   Error: ${h.error}`)),console.log(o.gray(`   Failed stages: ${h.failedSteps.join(", ")}`))):console.log(o.yellow(`Workflow ${h.status}`))}}else console.log(o.green("Workflow execution started")),console.log(o.gray(`   Use 'aqe workflow status ${p}' to check progress`));console.log(""),await a(0)}catch(s){console.error(o.red(`
@@ -35,10 +35,10 @@ Workflow cancelled: ${t}
 `)):console.log(o.red(`
 Failed to cancel workflow: ${l.error.message}
 `)),await a(l.success?0:1)}catch(l){console.error(o.red(`
-Failed to cancel workflow:`),l),await a(1)}}),y.command("browser-list").description("List available browser workflow templates").action(async()=>{try{let{BrowserWorkflowTool:t}=await import("./browser-workflow-A77OFBVY.js"),e=await new t().invoke({});if(e.success&&e.data){console.log(o.blue(`
+Failed to cancel workflow:`),l),await a(1)}}),y.command("browser-list").description("List available browser workflow templates").action(async()=>{try{let{BrowserWorkflowTool:t}=await import("./browser-workflow-ZMMKCGYW.js"),e=await new t().invoke({});if(e.success&&e.data){console.log(o.blue(`
  Browser Workflow Templates:
 `));for(let u of e.data.availableTemplates)console.log(`  ${o.cyan(u)}`);console.log("")}else console.log(o.red(`Failed: ${e.error||"Unknown error"}`));await a(0)}catch(t){console.error(o.red(`
-Failed:`),t),await a(1)}}),y.command("browser-load [template]").description("Load and validate a browser workflow template or inline YAML").option("--yaml <yaml>","Inline YAML workflow definition").option("-v, --var <key=value>","Variable override (repeatable)",P,{}).option("-F, --format <format>","Output format (text|json)","text").option("-o, --output <path>","Write output to file").action(async(t,l)=>{try{let{BrowserWorkflowTool:e}=await import("./browser-workflow-A77OFBVY.js"),{writeOutput:u,toJSON:c}=await import("./ci-output-CLI4A43F.js"),s=new e,f={variables:l.var||{}};if(l.yaml)f.workflowYaml=l.yaml;else if(t)if(t.endsWith(".yaml")||t.endsWith(".yml")){let i=await import("fs"),p=(await import("path")).resolve(t);i.existsSync(p)||(console.log(o.red(`
+Failed:`),t),await a(1)}}),y.command("browser-load [template]").description("Load and validate a browser workflow template or inline YAML").option("--yaml <yaml>","Inline YAML workflow definition").option("-v, --var <key=value>","Variable override (repeatable)",P,{}).option("-F, --format <format>","Output format (text|json)","text").option("-o, --output <path>","Write output to file").action(async(t,l)=>{try{let{BrowserWorkflowTool:e}=await import("./browser-workflow-ZMMKCGYW.js"),{writeOutput:u,toJSON:c}=await import("./ci-output-25GS6RXR.js"),s=new e,f={variables:l.var||{}};if(l.yaml)f.workflowYaml=l.yaml;else if(t)if(t.endsWith(".yaml")||t.endsWith(".yml")){let i=await import("fs"),p=(await import("path")).resolve(t);i.existsSync(p)||(console.log(o.red(`
 File not found: ${p}
 `)),await a(1)),f.workflowYaml=i.readFileSync(p,"utf-8")}else f.templateName=t;else console.log(o.red(`
 Provide a template name or --yaml. Use "workflow browser-list" to see templates.

package/dist/cli/chunks/workflow-orchestrator-MIVBCALG.js

@@ -0,0 +1,2 @@
+import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.3");process.exit(0)}
+import{a,b,c,d}from"./chunk-PDX7IYTJ.js";import"./chunk-Y6XNISHP.js";import"./chunk-NOSQI2OW.js";import"./chunk-FJACKKRL.js";import"./chunk-L2X3BWMN.js";import"./chunk-G5OGKVLO.js";import"./chunk-4EHW4TNA.js";import"./chunk-NJDLJ3KD.js";import"./chunk-OBL6W4Z5.js";import"./chunk-NKTLKMUB.js";import"./chunk-2LNDU7AW.js";import"./chunk-EOUJ6X5H.js";import"./chunk-ULQCEM7K.js";import"./chunk-UAPUA6OY.js";import"./chunk-GS6G7AVS.js";import"./chunk-YUDGQZJN.js";import"./chunk-HJ4U64BL.js";import"./chunk-CYSBM52D.js";import"./chunk-NLDLUHOF.js";import"./chunk-PEHEANAM.js";import"./chunk-XUUVAAHV.js";import"./chunk-SGMJ277D.js";export{b as DEFAULT_WORKFLOW_CONFIG,a as WorkflowEvents,c as WorkflowOrchestrator,d as createWorkflowOrchestrator};

package/dist/cli/chunks/{wrappers-I2CBKWGF.js → wrappers-VGO2GO4J.js}

@@ -1,2 +1,2 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.1");process.exit(0)}
-import{a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v as L,w as M,x as N}from"./chunk-WPG5LXR7.js";import{a as v,b as w,c as x,d as y,e as z,f as A,g as B,h as C,i as D,j as E,k as F,l as G,m as H,n as I,o as J,p as K}from"./chunk-6ONUCDVL.js";import"./chunk-XDRAU3W6.js";import"./chunk-PQ4NSXZM.js";import"./chunk-FNMEVLSN.js";import"./chunk-NOCPO55P.js";import"./chunk-IE6FDICX.js";import"./chunk-O65OPAZH.js";import"./chunk-BLU7OLQZ.js";import"./chunk-UBPJQKYS.js";import"./chunk-TMPRR3FZ.js";import"./chunk-R4A6PY5D.js";import"./chunk-JRNDFC23.js";import"./chunk-6BCVKCG5.js";import"./chunk-UR655D4K.js";import"./chunk-PZ3Q34OT.js";import"./chunk-CXUHD2O5.js";import"./chunk-IC6SOVGO.js";import"./chunk-USTUWUCE.js";import"./chunk-U5ITD7ZD.js";export{z as GNNLayerFactory,i as QEFlashAttention,x as QEGNNEmbeddingIndex,y as QEGNNIndexFactory,a as QESONA,e as QE_FLASH_ATTENTION_CONFIG,g as QE_PERFORMANCE_TARGETS,f as QE_SONA_CONFIG,A as TensorCompressionFactory,u as batchComputeAttention,K as batchDifferentiableSearch,M as checkRuvectorPackagesAvailable,c as createDomainQESONA,j as createQEFlashAttention,b as createQESONA,D as getDifferentiableSearch,F as getGetCompressionLevel,E as getHierarchicalForward,G as getInit,h as getOptimalBlockConfig,k as getQEFlashAttentionConfig,n as getRuvectorDotProductAttention,m as getRuvectorFlashAttention,p as getRuvectorHyperbolicAttention,B as getRuvectorLayer,q as getRuvectorLinearAttention,r as getRuvectorMoEAttention,o as getRuvectorMultiHeadAttention,L as getRuvectorPackageVersions,C as getTensorCompress,l as getWorkloadTypes,N as initAllRuvectorPackages,w as initGNN,d as isAttentionAvailable,v as isGNNAvailable,s as toFloat32Array,H as toFloat32ArrayGNN,J as toIEmbedding,t as toNumberArrayAttn,I as toNumberArrayGNN};
+import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.12.3");process.exit(0)}
+import{a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v as L,w as M,x as N}from"./chunk-QILF6BZH.js";import{a as v,b as w,c as x,d as y,e as z,f as A,g as B,h as C,i as D,j as E,k as F,l as G,m as H,n as I,o as J,p as K}from"./chunk-YOG5KQZF.js";import"./chunk-NOSQI2OW.js";import"./chunk-FJACKKRL.js";import"./chunk-L2X3BWMN.js";import"./chunk-G5OGKVLO.js";import"./chunk-4EHW4TNA.js";import"./chunk-NJDLJ3KD.js";import"./chunk-OBL6W4Z5.js";import"./chunk-2LNDU7AW.js";import"./chunk-ULQCEM7K.js";import"./chunk-UAPUA6OY.js";import"./chunk-GS6G7AVS.js";import"./chunk-YUDGQZJN.js";import"./chunk-HJ4U64BL.js";import"./chunk-CYSBM52D.js";import"./chunk-NLDLUHOF.js";import"./chunk-PEHEANAM.js";import"./chunk-XUUVAAHV.js";import"./chunk-SGMJ277D.js";export{z as GNNLayerFactory,i as QEFlashAttention,x as QEGNNEmbeddingIndex,y as QEGNNIndexFactory,a as QESONA,e as QE_FLASH_ATTENTION_CONFIG,g as QE_PERFORMANCE_TARGETS,f as QE_SONA_CONFIG,A as TensorCompressionFactory,u as batchComputeAttention,K as batchDifferentiableSearch,M as checkRuvectorPackagesAvailable,c as createDomainQESONA,j as createQEFlashAttention,b as createQESONA,D as getDifferentiableSearch,F as getGetCompressionLevel,E as getHierarchicalForward,G as getInit,h as getOptimalBlockConfig,k as getQEFlashAttentionConfig,n as getRuvectorDotProductAttention,m as getRuvectorFlashAttention,p as getRuvectorHyperbolicAttention,B as getRuvectorLayer,q as getRuvectorLinearAttention,r as getRuvectorMoEAttention,o as getRuvectorMultiHeadAttention,L as getRuvectorPackageVersions,C as getTensorCompress,l as getWorkloadTypes,N as initAllRuvectorPackages,w as initGNN,d as isAttentionAvailable,v as isGNNAvailable,s as toFloat32Array,H as toFloat32ArrayGNN,J as toIEmbedding,t as toNumberArrayAttn,I as toNumberArrayGNN};

package/dist/cli/commands/eval.js

@@ -15,7 +15,8 @@ import chalk from 'chalk';
 import * as fs from 'fs';
 import * as path from 'path';
 import * as yaml from 'yaml';
-import { createParallelEvalRunner, DEFAULT_PARALLEL_EVAL_CONFIG, } from '../../validation/parallel-eval-runner.js';
+import { createParallelEvalRunner, DEFAULT_PARALLEL_EVAL_CONFIG, MockLLMExecutor, } from '../../validation/parallel-eval-runner.js';
+import { resolveEvalExecutor } from '../../validation/provider-llm-executor.js';
 import { createCommandEvalRunner, isCommandEvalSuite, } from '../../validation/command-eval-runner.js';
 import { createSkillValidationLearner, } from '../../learning/skill-validation-learner.js';
 import { createRealQEReasoningBank, } from '../../learning/real-qe-reasoning-bank.js';
@@ -54,6 +55,16 @@ const P2_SKILLS = [
 /**
  * Get skills for a specific tier
  */
+/**
+ * ADR-123: propagate a `--max-budget-usd` flag to the provider layer via the
+ * AQE_MAX_BUDGET_USD env var, which ProviderManager reads as a per-run cap.
+ */
+function applyMaxBudget(maxBudgetUsd) {
+    if (typeof maxBudgetUsd === 'number' && Number.isFinite(maxBudgetUsd) && maxBudgetUsd > 0) {
+        process.env.AQE_MAX_BUDGET_USD = String(maxBudgetUsd);
+        console.log(chalk.gray(`  Budget cap: $${maxBudgetUsd.toFixed(2)} for this run`));
+    }
+}
 function getSkillsByTier(tier) {
     switch (tier) {
         case 3:
@@ -295,6 +306,36 @@ async function handleRunCommand(options) {
         }
         process.exit(result.passed ? 0 : 1);
     }
+    // Issue #557 follow-up: resolve a REAL LLM executor by default. `--mock`
+    // opts into canned responses (offline testing only). If neither a provider
+    // is configured nor --mock is passed, fail loudly rather than fabricate.
+    let executor;
+    let evalManager;
+    if (options.mock) {
+        console.log(chalk.yellow.bold('  ⚠️  MOCK MODE: canned responses, NOT real model output. Results are not valid eval evidence.'));
+        executor = new MockLLMExecutor();
+    }
+    else {
+        const resolved = await resolveEvalExecutor();
+        if (!resolved.executor) {
+            console.error(chalk.red('\n  Cannot run real eval: ') + (resolved.reason ?? 'no provider'));
+            process.exit(1);
+        }
+        executor = resolved.executor;
+        evalManager = resolved.manager;
+        const modeLabel = resolved.billingMode === 'subscription'
+            ? chalk.green('subscription (no per-token charge)')
+            : resolved.billingMode === 'local'
+                ? chalk.green('local (no cost)')
+                : resolved.billingMode === 'metered-capped'
+                    ? chalk.yellow('metered (server-side cap)')
+                    : chalk.red('pay-per-token API');
+        console.log(`  LLM provider: ${chalk.cyan(resolved.providerType)}  billing: ${modeLabel}`);
+        if (process.env.AQE_MAX_BUDGET_USD) {
+            console.log(chalk.gray(`  Budget cap: $${process.env.AQE_MAX_BUDGET_USD} for this run`));
+        }
+        console.log('');
+    }
     const { runner, reasoningBank } = await initializeRunner();
     try {
         // Configure runner
@@ -305,7 +346,7 @@ async function handleRunCommand(options) {
             timeout: options.timeout,
             retryFailedTests: options.retry,
         };
-        const customRunner = createParallelEvalRunner(createSkillValidationLearner(reasoningBank), config);
+        const customRunner = createParallelEvalRunner(createSkillValidationLearner(reasoningBank), config, executor);
         // Set up progress reporting
         if (options.verbose) {
             customRunner.onProgress(displayProgress);
@@ -333,11 +374,30 @@ async function handleRunCommand(options) {
     }
     finally {
         await reasoningBank.dispose();
+        if (evalManager)
+            await evalManager.dispose();
     }
 }
 async function handleRunAllCommand(options) {
     console.log(chalk.bold('\nAQE Parallel Eval Runner - Multi-Skill'));
     console.log(chalk.dim('ADR-056 Phase 5: Worker Pool Pattern\n'));
+    // Issue #557 follow-up: real executor by default; --mock is offline-only.
+    let executor;
+    let evalManager;
+    if (options.mock) {
+        console.log(chalk.yellow.bold('  ⚠️  MOCK MODE: canned responses, NOT real model output. Results are not valid eval evidence.'));
+        executor = new MockLLMExecutor();
+    }
+    else {
+        const resolved = await resolveEvalExecutor();
+        if (!resolved.executor) {
+            console.error(chalk.red('\n  Cannot run real eval: ') + (resolved.reason ?? 'no provider'));
+            process.exit(1);
+        }
+        executor = resolved.executor;
+        evalManager = resolved.manager;
+        console.log(`  LLM provider: ${chalk.cyan(resolved.providerType)}  billing: ${resolved.billingMode}`);
+    }
     const { runner, learner, reasoningBank } = await initializeRunner();
     try {
         // Get skills for tier
@@ -353,7 +413,7 @@ async function handleRunAllCommand(options) {
             ...DEFAULT_PARALLEL_EVAL_CONFIG,
             maxWorkers: options.parallel ? options.workers : 1,
         };
-        const customRunner = createParallelEvalRunner(learner, config);
+        const customRunner = createParallelEvalRunner(learner, config, executor);
         // Run all evals
         const results = await customRunner.runMultipleEvalsParallel(skills, models);
         // Flatten results for display
@@ -394,6 +454,8 @@ async function handleRunAllCommand(options) {
     }
     finally {
         await reasoningBank.dispose();
+        if (evalManager)
+            await evalManager.dispose();
     }
 }
 async function handleStatusCommand(skill) {
@@ -531,7 +593,10 @@ export function createEvalCommand() {
         .option('--no-retry', 'Disable retry of failed tests')
         .option('-o, --output <path>', 'Output file path for results')
         .option('-v, --verbose', 'Show progress during execution', false)
+        .option('--max-budget-usd <n>', 'ADR-123: cap total LLM spend for this run in USD (aborts with COST_LIMIT_EXCEEDED past the cap)', parseFloat)
+        .option('--mock', 'Use canned responses instead of real model calls (offline testing only — results are NOT valid eval evidence)', false)
         .action(async (opts) => {
+        applyMaxBudget(opts.maxBudgetUsd);
         await handleRunCommand({
             skill: opts.skill,
             model: opts.model,
@@ -542,6 +607,7 @@ export function createEvalCommand() {
             retry: opts.retry !== false,
             output: opts.output,
             verbose: opts.verbose,
+            mock: opts.mock === true,
         });
     });
     // Run all evals
@@ -554,7 +620,10 @@ export function createEvalCommand() {
         .option('-w, --workers <n>', 'Number of parallel workers', parseInt, 5)
         .option('-o, --output <path>', 'Output file path for report')
         .option('-v, --verbose', 'Show progress during execution', false)
+        .option('--max-budget-usd <n>', 'ADR-123: cap total LLM spend for this run in USD (aborts with COST_LIMIT_EXCEEDED past the cap)', parseFloat)
+        .option('--mock', 'Use canned responses instead of real model calls (offline testing only — results are NOT valid eval evidence)', false)
         .action(async (opts) => {
+        applyMaxBudget(opts.maxBudgetUsd);
         await handleRunAllCommand({
             skillsTier: opts.skillsTier,
             models: opts.models,
@@ -562,6 +631,7 @@ export function createEvalCommand() {
             workers: opts.workers,
             output: opts.output,
             verbose: opts.verbose,
+            mock: opts.mock === true,
         });
     });
     // Status command

package/dist/cli/commands/llm-router.js

@@ -399,12 +399,14 @@ function formatCost(cost) {
 function getDefaultModelForProvider(provider) {
     const defaults = {
         claude: 'claude-sonnet-4-6',
+        'claude-code': 'sonnet',
         openai: 'gpt-4o',
         ollama: 'llama3.1',
         openrouter: 'anthropic/claude-sonnet-4',
         gemini: 'gemini-2.0-pro',
         'azure-openai': 'gpt-4o',
         bedrock: 'anthropic.claude-sonnet-4-v1:0',
+        cognitum: 'cognitum-auto',
         onnx: 'phi-4',
     };
     return defaults[provider] || '';

package/dist/cli/handlers/status-handler.d.ts

@@ -22,6 +22,11 @@ export interface DaemonLivenessStatus {
  * tracks a *different*, in-process `QualityDaemon` instance that can't see a
  * truly detached process across separate CLI invocations anyway.
  */
+/**
+ * ADR-123 (issue #557): print how the active LLM provider bills, plus any
+ * configured per-run budget cap. Best-effort — never throws into `aqe health`.
+ */
+export declare function printLlmBilling(): void;
 export declare function checkDaemonLiveness(): DaemonLivenessStatus;
 export declare class StatusHandler implements ICommandHandler {
     readonly name = "status";

package/dist/cli/handlers/status-handler.js

@@ -9,6 +9,8 @@ import { join } from 'node:path';
 import { getStatusColor, formatUptime, } from './interfaces.js';
 import { writeOutput, toJSON } from '../utils/ci-output.js';
 import { findProjectRoot } from '../../kernel/unified-memory.js';
+import { loadRouterConfig } from '../../shared/llm/router/config-store.js';
+import { billingModeForType } from '../../shared/llm/billing-modes.js';
 /**
  * Check whether the detached background-worker daemon (`.agentic-qe/workers/
  * start-daemon.cjs`, PID recorded in `daemon.pid` — see `10-workers.ts`) is
@@ -19,6 +21,36 @@ import { findProjectRoot } from '../../kernel/unified-memory.js';
  * tracks a *different*, in-process `QualityDaemon` instance that can't see a
  * truly detached process across separate CLI invocations anyway.
  */
+/**
+ * ADR-123 (issue #557): print how the active LLM provider bills, plus any
+ * configured per-run budget cap. Best-effort — never throws into `aqe health`.
+ */
+export function printLlmBilling() {
+    try {
+        const config = loadRouterConfig();
+        const primary = config.defaultProvider;
+        const mode = billingModeForType(primary);
+        const modeLabel = {
+            'metered-api': `${chalk.red('●')} pay-per-token API key (no cap)`,
+            'metered-capped': `${chalk.yellow('●')} pay-per-token with server-side cap`,
+            subscription: `${chalk.green('●')} Claude subscription (no per-token charge)`,
+            local: `${chalk.green('●')} local (no cost)`,
+        };
+        console.log(chalk.blue('\n  LLM Billing:'));
+        console.log(`  Provider: ${chalk.cyan(primary)}  ${modeLabel[mode] ?? mode}`);
+        const cap = Number.parseFloat(process.env.AQE_MAX_BUDGET_USD ?? '');
+        if (Number.isFinite(cap) && cap > 0) {
+            console.log(`  Per-run budget cap: ${chalk.cyan('$' + cap.toFixed(2))}`);
+        }
+        else if (mode === 'metered-api') {
+            console.log(chalk.gray('  No budget cap set — use AQE_MAX_BUDGET_USD or --max-budget-usd.'));
+            console.log(chalk.gray('  Tip: AQE_LLM_PROVIDER=claude-code runs on your Claude subscription.'));
+        }
+    }
+    catch {
+        // Billing display is informational; failures must not break health.
+    }
+}
 export function checkDaemonLiveness() {
     try {
         const pidFile = join(findProjectRoot(), '.agentic-qe', 'workers', 'daemon.pid');
@@ -291,6 +323,9 @@ export class HealthHandler {
                 else {
                     console.log(`  ${chalk.red('\u25CF')} Not running (stale PID: ${daemon.pid ?? 'unknown'}) \u2014 learning/consolidation snapshots will not advance`);
                 }
+                // ADR-123 (issue #557): make LLM billing visible so a paid API key is
+                // never a silent surprise.
+                printLlmBilling();
             }
             console.log('');
             await this.cleanupAndExit(0);

package/dist/coordination/consensus/providers/index.js

@@ -184,6 +184,17 @@ export function registerProvidersFromEnv(enableLogging = false) {
         config.claude = {
             apiKey: process.env.ANTHROPIC_API_KEY,
         };
+        // ADR-123 (issue #557): this consensus/multi-model-verification path bills
+        // the pay-per-token Anthropic API, not a Claude subscription, and has no
+        // budget cap of its own. Surface it so the spend isn't silent. (Full
+        // subscription + budget support for the consensus ModelProvider layer is
+        // tracked as ADR-123 follow-up — it uses a different interface than the
+        // shared LLM ProviderManager.)
+        if ((process.env.AQE_LLM_NO_BILLING_NOTICE ?? '') !== '1') {
+            // eslint-disable-next-line no-console
+            console.error('⚠️  Consensus verification will use the Anthropic API key (pay-per-token, ' +
+                'no budget cap). Unset ANTHROPIC_API_KEY to disable this path.');
+        }
     }
     // Detect OpenAI
     if (process.env.OPENAI_API_KEY) {

package/dist/integrations/ruvector/brain-rvf-exporter.js

@@ -3,11 +3,82 @@
  * `.rvf` file using the @ruvector/rvf-node native binding via rvf-native-adapter.
  * Falls back to JSONL directory format when the native binding is unavailable.
  */
-import { existsSync, statSync, unlinkSync, writeFileSync } from 'fs';
-import { resolve } from 'path';
+import { existsSync, readdirSync, renameSync, statSync, unlinkSync, writeFileSync } from 'fs';
+import { basename, dirname, join, resolve } from 'path';
 import { createRvfStore, openRvfStore, openRvfStoreReadonly, isRvfNativeAvailable, } from './rvf-native-adapter.js';
 import { sha256, tableExists, queryAll, domainFilterForColumn, ensureTargetTables, mergeGenericRow, mergeAppendOnlyRow, TABLE_CONFIGS, TABLE_BLOB_COLUMNS, PK_COLUMNS, CONFIDENCE_COLUMNS, TIMESTAMP_COLUMNS, serializeRowBlobs, deserializeRowBlobs, } from './brain-shared.js';
 // --- Export ---
+/**
+ * Sidecar suffixes that make up a complete store on disk. `''` is the store
+ * itself; the adapter writes `.idmap.json` on close, and we write the manifest.
+ */
+const RVF_STORE_SUFFIXES = ['', '.idmap.json', '.manifest.json'];
+/**
+ * Everything an abandoned tmp export can leave behind. `.lock` is cleaned up
+ * but deliberately absent from RVF_STORE_SUFFIXES — renaming a lock into the
+ * final path is the very corruption this fix exists to prevent.
+ */
+const RVF_CLEANUP_SUFFIXES = [...RVF_STORE_SUFFIXES, '.lock'];
+/** Best-effort removal of a store and its sidecars at `base`. */
+function removeStoreFiles(base) {
+    for (const suffix of RVF_CLEANUP_SUFFIXES) {
+        try {
+            const p = `${base}${suffix}`;
+            if (existsSync(p))
+                unlinkSync(p);
+        }
+        catch {
+            // best-effort — a leftover tmp is inert, unlike a corrupt final store
+        }
+    }
+}
+/**
+ * Remove tmp stores orphaned by killed exports (#563). Each interrupted export
+ * strands a full-size tmp store, so on a project where the checkpoint hook is
+ * killed regularly these would accumulate unbounded.
+ *
+ * Only tmps whose owning pid is gone are removed: a live pid means a concurrent
+ * export is mid-write, and deleting its tmp would corrupt exactly what this
+ * change exists to protect.
+ */
+function sweepStaleTmpFiles(outPath) {
+    const dir = dirname(outPath);
+    const prefix = `${basename(outPath)}.tmp.`;
+    let entries;
+    try {
+        entries = readdirSync(dir);
+    }
+    catch {
+        return;
+    }
+    for (const entry of entries) {
+        if (!entry.startsWith(prefix))
+            continue;
+        // `<name>.rvf.tmp.<pid>` and its `<...>.tmp.<pid>.idmap.json` siblings.
+        const pid = Number.parseInt(entry.slice(prefix.length).split('.')[0], 10);
+        if (!Number.isInteger(pid) || pid <= 0)
+            continue;
+        if (pid !== process.pid && isPidAlive(pid))
+            continue;
+        try {
+            unlinkSync(join(dir, entry));
+        }
+        catch {
+            // best-effort
+        }
+    }
+}
+/** True when a process with `pid` exists (signal 0 probes without delivering). */
+function isPidAlive(pid) {
+    try {
+        process.kill(pid, 0);
+        return true;
+    }
+    catch (err) {
+        // EPERM means it exists but belongs to another user — still alive.
+        return err.code === 'EPERM';
+    }
+}
 /** Export brain state to a single `.rvf` file with vector embeddings + kernel. */
 export function exportBrainToRvf(db, options, sourceDbLabel = 'memory.db') {
     if (!isRvfNativeAvailable()) {
@@ -15,14 +86,20 @@ export function exportBrainToRvf(db, options, sourceDbLabel = 'memory.db') {
     }
     const outPath = resolve(options.outputPath);
     const dimension = options.dimension ?? 384;
-    // Clean up any existing file
-    if (existsSync(outPath))
-        unlinkSync(outPath);
-    const idmapPath = `${outPath}.idmap.json`;
-    if (existsSync(idmapPath))
-        unlinkSync(idmapPath);
+    // Issue #563: build the store at a private tmp path and rename it over the
+    // final path only once it is closed and complete. The previous code deleted
+    // the old store and then built the new one in place, so any interruption
+    // (hook timeout, session teardown, SIGKILL) left a created-but-incomplete
+    // store that could neither be opened (ManifestNotFound) nor created over
+    // (FsyncFailed) — permanently disabling the RVF backend. With tmp+rename an
+    // interrupted export leaves the *previous* good store untouched and orphans
+    // only an inert tmp file, which the next export sweeps.
+    const tmpPath = `${outPath}.tmp.${process.pid}`;
+    sweepStaleTmpFiles(outPath);
+    removeStoreFiles(tmpPath);
     // Create RVF container via adapter (handles idmap automatically)
-    const rvf = createRvfStore(outPath, dimension);
+    const rvf = createRvfStore(tmpPath, dimension);
+    let closed = false;
     try {
         // --- Export all tables using TABLE_CONFIGS ---
         const allTableData = {};
@@ -153,12 +230,37 @@ export function exportBrainToRvf(db, options, sourceDbLabel = 'memory.db') {
             catch { /* best-effort */ }
         }
         // Persist manifest as sidecar JSON alongside the .rvf file
-        const manifestPath = `${outPath}.manifest.json`;
-        writeFileSync(manifestPath, JSON.stringify(manifest, null, 2), 'utf-8');
+        writeFileSync(`${tmpPath}.manifest.json`, JSON.stringify(manifest, null, 2), 'utf-8');
+        // Close before promoting: close() flushes the store and writes the idmap
+        // sidecar, so renaming any earlier would publish an incomplete store.
+        rvf.close();
+        closed = true;
+        // Promote tmp → final. The store itself goes last: until it lands, the
+        // previous store (if any) is still the one on disk, and a crash mid-promote
+        // leaves the old store readable rather than a half-written new one.
+        for (const suffix of [...RVF_STORE_SUFFIXES].reverse()) {
+            const from = `${tmpPath}${suffix}`;
+            const to = `${outPath}${suffix}`;
+            if (existsSync(from)) {
+                renameSync(from, to);
+            }
+            else if (existsSync(to)) {
+                // No new sidecar produced — drop the stale one so it can't be read
+                // against the new store.
+                unlinkSync(to);
+            }
+        }
         return manifest;
     }
-    finally {
-        rvf.close();
+    catch (err) {
+        if (!closed) {
+            try {
+                rvf.close();
+            }
+            catch { /* best-effort */ }
+        }
+        removeStoreFiles(tmpPath);
+        throw err;
     }
 }
 /** Legacy field-to-table mapping for older RVF exports. */

package/dist/integrations/ruvector/rvf-dual-writer.js

@@ -87,14 +87,31 @@ export class RvfDualWriter {
                 nativeAdapter = adapter.openRvfStore(this.config.rvfPath);
             }
             catch {
-                nativeAdapter = adapter.createRvfStore(this.config.rvfPath, this.config.dimensions);
+                try {
+                    nativeAdapter = adapter.createRvfStore(this.config.rvfPath, this.config.dimensions);
+                }
+                catch (createErr) {
+                    // Issue #563: open failed *and* create failed, so the file on disk is
+                    // provably unusable — the signature of an export killed mid-write
+                    // (open → ManifestNotFound, create → FsyncFailed because the path
+                    // exists). Previously this fell to the catch below and disabled RVF
+                    // silently for the whole run. Quarantine and rebuild instead.
+                    const { quarantineUnusableStore } = await import('./rvf-store-integrity.js');
+                    const quarantined = quarantineUnusableStore(this.config.rvfPath, createErr instanceof Error ? createErr.message : String(createErr));
+                    if (!quarantined)
+                        throw createErr; // live peer, or nothing to move
+                    nativeAdapter = adapter.createRvfStore(this.config.rvfPath, this.config.dimensions);
+                }
             }
             this.rvfStore = wrapNativeAdapter(nativeAdapter, this.config.dimensions);
             this.rvfAvailable = true;
         }
-        catch {
-            // Native adapter module not available at all
+        catch (err) {
+            // Native adapter unavailable, or recovery itself failed. Degrade to
+            // sqlite-only — but say so, instead of vanishing silently (#563).
             this.rvfAvailable = false;
+            console.warn(`[RVF] Dual-writer disabled for ${this.config.rvfPath}; falling back to SQLite: ` +
+                `${err instanceof Error ? err.message : String(err)}`);
         }
     }
     /**

package/dist/integrations/ruvector/rvf-store-integrity.d.ts

@@ -0,0 +1,49 @@
+/**
+ * RVF store integrity helpers (issue #563).
+ *
+ * An export that is killed mid-write leaves a store that was created but never
+ * completed. Such a file cannot be opened (`RVF error 0x0106: ManifestNotFound`)
+ * and cannot be created over either, because the path already exists
+ * (`RVF error 0x0303: FsyncFailed`). Both paths fail forever, so the RVF
+ * backend stays disabled for the life of the project until someone deletes the
+ * file by hand.
+ *
+ * Recovery has to be evidence-based rather than fingerprint-based. Measured
+ * against @ruvector/rvf-node 0.1.8:
+ *
+ *   • A *valid, empty* store is 162 bytes with magic `SFVR` — byte-for-byte
+ *     the size the issue reports as "truncated: header only". Size and magic
+ *     therefore cannot distinguish a broken store from a healthy empty one.
+ *   • `<store>.lock` is a 104-byte record with magic `FLVR` and the owning pid
+ *     as a little-endian u32 at offset 4. `FLVR` in a lock is the *normal*
+ *     case — the issue reads it as "RVF store bytes leaked into the lock", but
+ *     the store magic is `SFVR`; a 104-byte `FLVR` lock is just a live-or-stale
+ *     lock record.
+ *
+ * So the only sound trigger is "open failed *and* create failed": at that point
+ * the file is provably unusable. The lock's pid is used as a *guard* — if it is
+ * held by a live process, the store is presumed healthy and left alone.
+ */
+/** The pid recorded in `<rvfPath>.lock`, or null if there is no readable lock. */
+export declare function readLockOwnerPid(rvfPath: string): number | null;
+/**
+ * True when the store's lock is held by a process that is still running.
+ *
+ * Deliberately conservative: an unreadable or pid-less lock reports false (the
+ * caller's open already failed, so there is nothing to protect), while a live
+ * pid reports true and blocks recovery. Pid reuse can only produce a false
+ * "alive", which fails safe — we leave the store alone.
+ */
+export declare function isLockHeldByLiveProcess(rvfPath: string): boolean;
+/**
+ * Move an unusable store and its sidecars aside so the next create succeeds.
+ * Returns the quarantine base path, or null if nothing was quarantined.
+ *
+ * Only call this once open *and* create have both failed — that is the proof
+ * the file is unusable. The store is a derived cache, rebuilt from the unified
+ * DB, so quarantining is recoverable; we move rather than delete because the
+ * bytes may hold writes that never reached SQLite, and they are the only
+ * evidence if this recurs.
+ */
+export declare function quarantineUnusableStore(rvfPath: string, reason: string): string | null;
+//# sourceMappingURL=rvf-store-integrity.d.ts.map