@rulemetric/cli 0.17.3 → 0.18.1

package/dist/chunk-A6JUIZ3H.js

@@ -0,0 +1 @@
+import{d as a}from"./chunk-5VWVQCNR.js";import{a as s}from"./chunk-ZLLER3HP.js";async function u(){let[e,t]=await Promise.allSettled([a("/api/setup/status",{timeoutMs:5e3}),a("/api/workers/health",{timeoutMs:5e3})]),i=s((n,r)=>({label:n,severity:"warn",message:`Status unavailable: ${r instanceof Error?r.message:String(r)}`,hints:[]}),"unavailable");return[e.status==="fulfilled"?{label:"Capture",severity:e.value.anyCapturedSession?"ok":"warn",message:e.value.anyCapturedSession?"Recorded sessions found":"No captured sessions received yet",hints:e.value.anyCapturedSession?[]:["Check capture setup with: rulemetric setup"]}:i("Capture",e.reason),t.status==="fulfilled"?{label:"Worker",severity:t.value.verdict.severity,message:t.value.verdict.headline,hints:t.value.verdict.hints}:i("Worker",t.reason)]}s(u,"loadTelemetryChecks");export{u as a};

package/dist/{chunk-TG4C4ENM.js → chunk-CXENLSMR.js}

@@ -1 +1 @@
-import{b as u,c as d}from"./chunk-EQ2O6DSN.js";import{b as g}from"./chunk-UQJNUESP.js";import{e as m,f as p,h as l}from"./chunk-5VWVQCNR.js";import{a as s}from"./chunk-ZLLER3HP.js";async function x(f,R){let r=s(e=>R?.logger?.info?.(e)??console.log(e),"log"),{evalTargetId:n}=f,i=`[cron_refine_cases] ${n}`,t=await m(`/api/eval-targets/${n}`);if(!(t.metadata?.autoRefineCases===!0))return r(`${i} not opted in (metadata.autoRefineCases)`),{skipped:"not-opted-in",reason:`"${t.name}" has not opted into automatic case refinement. Set metadata.autoRefineCases=true to let the loop cull cases that carry no treatment signal, or run \`rulemetric evals refine-cases ${n}\` by hand.`,targetName:t.name};let C=await m(`/api/eval-targets/${n}/judge-calibration`),a=u(C.caseDiscrimination);if(a.refused)return r(`${i} ${a.refused}: ${a.summary}`),{skipped:a.refused,reason:a.summary,targetName:t.name};r(`${i} ${a.summary}`);let o=await d(a,{generate:s(e=>g({content:t.content,name:t.name,feedback:e}),"generate"),createCase:s(e=>p(`/api/eval-targets/${n}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations}),"createCase"),retireCase:s(async(e,y)=>{await l(`/api/eval-targets/${n}/evals/${e}`,{retire:!0,retiredReason:y})},"retireCase"),log:r});r(`${i} ${o.summary}`);let{skipped:c,...$}=o;return{...$,targetName:t.name,...c?{skipped:c,reason:o.summary}:{}}}s(x,"cronRefineCases");export{x as a};
+import{b as u,c as d}from"./chunk-YZQZD2E6.js";import{b as g}from"./chunk-UQJNUESP.js";import{e as m,f as p,h as l}from"./chunk-5VWVQCNR.js";import{a as s}from"./chunk-ZLLER3HP.js";async function x(f,R){let r=s(e=>R?.logger?.info?.(e)??console.log(e),"log"),{evalTargetId:n}=f,i=`[cron_refine_cases] ${n}`,t=await m(`/api/eval-targets/${n}`);if(!(t.metadata?.autoRefineCases===!0))return r(`${i} not opted in (metadata.autoRefineCases)`),{skipped:"not-opted-in",reason:`"${t.name}" has not opted into automatic case refinement. Set metadata.autoRefineCases=true to let the loop cull cases that carry no treatment signal, or run \`rulemetric evals refine-cases ${n}\` by hand.`,targetName:t.name};let C=await m(`/api/eval-targets/${n}/judge-calibration`),a=u(C.caseDiscrimination);if(a.refused)return r(`${i} ${a.refused}: ${a.summary}`),{skipped:a.refused,reason:a.summary,targetName:t.name};r(`${i} ${a.summary}`);let o=await d(a,{generate:s(e=>g({content:t.content,name:t.name,feedback:e}),"generate"),createCase:s(e=>p(`/api/eval-targets/${n}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations}),"createCase"),retireCase:s(async(e,y)=>{await l(`/api/eval-targets/${n}/evals/${e}`,{retire:!0,retiredReason:y})},"retireCase"),log:r});r(`${i} ${o.summary}`);let{skipped:c,...$}=o;return{...$,targetName:t.name,...c?{skipped:c,reason:o.summary}:{}}}s(x,"cronRefineCases");export{x as a};

package/dist/{chunk-BTKVLLV2.js → chunk-FLJTAD2R.js}

@@ -1 +1 @@
-import{b as t,e as E}from"./chunk-MGT5FZJ4.js";import{a as r}from"./chunk-ZLLER3HP.js";function s(o=10,_=t){for(let e=1;e<=o;e++)if(E(e,o,_).feasible)return e;return o}r(s,"minFeasibleLiveCases");var S=s(),A=3;export{s as a,S as b,A as c};
+import{b as t,e as E}from"./chunk-L5ILEWHH.js";import{a as r}from"./chunk-ZLLER3HP.js";function s(o=10,_=t){for(let e=1;e<=o;e++)if(E(e,o,_).feasible)return e;return o}r(s,"minFeasibleLiveCases");var S=s(),A=3;export{s as a,S as b,A as c};

package/dist/chunk-I2FPXRA6.js

@@ -0,0 +1,2 @@
+import{c as C,o as E}from"./chunk-RQBN7WNS.js";import{a as h}from"./chunk-5VWVQCNR.js";import{a as l}from"./chunk-ZLLER3HP.js";import*as p from"node:fs";import*as u from"node:path";import*as m from"node:os";function k(){if(process.env.RULEMETRIC_API_KEY||process.env.RULEMETRIC_ACCESS_TOKEN)return;let s=u.join(m.homedir(),".config","rulemetric","env");v(s)||v(u.join(process.cwd(),".env.local"))}l(k,"ensureEnv");function v(s){try{let t=p.readFileSync(s,"utf-8");for(let a of t.split(`
+`)){let o=a.trim();if(!o||o.startsWith("#"))continue;let n=o.indexOf("=");if(n<0)continue;let e=o.slice(0,n).trim(),i=o.slice(n+1).trim();process.env[e]||(process.env[e]=i)}return!!(process.env.RULEMETRIC_API_KEY||process.env.RULEMETRIC_ACCESS_TOKEN)}catch{return!1}}l(v,"loadEnvFile");k();async function I(s){let t={totalSessions:0,totalCalls:0,totalInputTokens:0,totalOutputTokens:0,totalCacheReadTokens:0,totalCacheWriteTokens:0,totalCostUSD:0,providers:{}},a=await E(s);t.providers={};for(let r of a)t.providers[r.provider]=(t.providers[r.provider]||0)+1;let o=new C,n=a.filter(r=>r.provider==="claude");for(let r of n)try{let d=p.readdirSync(r.path).filter(f=>f.endsWith(".jsonl"));t.totalSessions+=d.length;for(let f of d){let T=p.readFileSync(u.join(r.path,f),"utf-8"),_=o.parse(T);for(let c of _.events)c.tokenUsage&&(t.totalCalls++,t.totalInputTokens+=c.tokenUsage.input,t.totalOutputTokens+=c.tokenUsage.output,t.totalCacheReadTokens+=c.tokenUsage.cacheRead??0,t.totalCacheWriteTokens+=c.tokenUsage.cacheWrite??0)}}catch{}let e=15,i=75;return t.totalCostUSD=Math.round((t.totalInputTokens/1e6*e+t.totalOutputTokens/1e6*i+t.totalCacheReadTokens/1e6*(e*.1)+t.totalCacheWriteTokens/1e6*(e*1.25))*100)/100,t}l(I,"loadLocalStats");async function U(s){let t=process.env.RULEMETRIC_API_URL??h,a=process.env.RULEMETRIC_API_KEY,o=process.env.RULEMETRIC_ACCESS_TOKEN,n=a??o;if(!n)return null;let e=new URLSearchParams;s&&e.set("projectPath",s);try{let i=await fetch(`${t}/api/insights/compute?${e}`,{headers:{Authorization:`Bearer ${n}`},signal:AbortSignal.timeout(5e3)});return i.ok?await i.json():null}catch{return null}}l(U,"loadApiInsights");export{I as a,U as b};

package/dist/{chunk-MGT5FZJ4.js → chunk-L5ILEWHH.js}

@@ -1 +1 @@
-import{a as _e}from"./chunk-M7QSC64D.js";import{i as Me}from"./chunk-DZRBB34J.js";import{a as Pe}from"./chunk-KR32Y5Q5.js";import{a as xe,b as Se,c as De,d as Ie,e as se,f as Oe}from"./chunk-RITL4NU3.js";import{b as pe,i as me,j as ge,k as fe,l as he,m as be,n as $e}from"./chunk-GYKYJJNF.js";import{c as Re}from"./chunk-G7V4ZZZ3.js";import{a as re}from"./chunk-AHLVY52I.js";import{d as ke,f as Ce,g as Ee,i as Ae}from"./chunk-PPQFN62M.js";import{a as we}from"./chunk-V6L25LFD.js";import{c as ye}from"./chunk-NAS7FIGM.js";import{b as ve}from"./chunk-VZYNZS5U.js";import{d as ue,e as ne,f as de}from"./chunk-KFNFLMUR.js";import{e as y,f as te}from"./chunk-5VWVQCNR.js";import{a as g}from"./chunk-ZLLER3HP.js";import{createHash as Ge}from"node:crypto";import{existsSync as Te,readFileSync as Ne,writeFileSync as Xe}from"node:fs";var H=g(t=>Ge("sha256").update(t).digest("hex"),"sha256"),Le=1.3,ae=4,ze=20;function Be(t,o){if(!t.configured||o<=0)return ae;let e=Math.ceil(t.minPairs*Le/o);return Math.min(ze,Math.max(ae,e))}g(Be,"passCapFor");function Qe(t,o=10,e=ae){if(t<=0)return{passes:0,expectedPairs:0,feasible:!1,reason:"no live cases"};let i=Math.ceil(o*Le),l=Math.min(e,Math.ceil(i/t)),r=l*t;return r<o?{passes:l,expectedPairs:r,feasible:!1,reason:`${t} live case(s) x ${e} passes = ${t*e} pairs at best, below the ${o} a controlled verdict requires \u2014 add cases rather than replaying these`}:{passes:l,expectedPairs:r,feasible:!0}}g(Qe,"planPasses");var Ve=60,Ye=5,We=g(t=>Array.isArray(t)?t.filter(o=>typeof o=="string"):[],"stringList"),J=g(t=>typeof t=="number"&&Number.isFinite(t)?t:null,"finite");function Ze(t,o,e=Ye){let i=[...t].filter(s=>s.fromContentHash===o).sort((s,c)=>Date.parse(c.createdAt)-Date.parse(s.createdAt)),l=[],r=new Set;for(let s of i){let c=s.toContentHash??s.id;if(r.has(c))continue;r.add(c);let d=s.verdict??null,k=J(d?.delta),M=J(d?.ci?.lower),x=J(d?.ci?.upper),S=J(d?.nWith),L=d?.sufficient===!0&&k!=null&&M!=null&&x!=null&&S!=null?{delta:k,ciLower:M,ciUpper:x,nPairs:S}:null;if(l.push({decision:typeof s.decision=="string"?s.decision:"unknown",applied:s.applied===!0,at:typeof s.createdAt=="string"?s.createdAt:"",rationale:typeof s.proposalRationale=="string"?s.proposalRationale:"",touched:We(s.touchedSections),refusals:We(s.refusals),measured:L}),l.length>=e)break}return l}g(Ze,"summarisePriorDecisions");function et(t,o){let e=o.accumulate??(t.autoEvolveAccumulate===!0||t.autoEvolveAccumulate==="true"),i=o.accumulateTargetPairs??t.autoEvolveAccumulateTarget,l=typeof i=="string"?Number.parseInt(i,10):i,r=typeof l=="number"&&Number.isInteger(l)&&l>0?l:Ve;return{enabled:e,targetPairs:r}}g(et,"readAccumulateConfig");function tt(t){let o=t.accumulateEnabled&&t.pairCount<t.targetPairs;return{apply:t.consented&&!o,peeking:o}}g(tt,"resolveApply");async function Pt(t,o){if(t.restoration)return Pe(t.evalTargetId,t.restoration);let e=g(n=>o?.logger?.info?.(n)??console.log(n),"log"),{evalTargetId:i}=t,l=t.measurement?ye(t):{},r=`[cron_instruction_evolution] ${i}`,s=await y(`/api/eval-targets/${i}`),c=s.metadata??{},d=t.measurement?JSON.stringify([l.engine,l.model,l.graderEngine,l.graderModel,s.type,ve(t.measurement.engine)]):void 0,k=null,M=!1;if(s.name.startsWith("/")&&Te(s.name))try{if(k=Ne(s.name,"utf8"),k!==s.content){let n=await te(`/api/eval-targets/${i}/sync-content`,{content:k});M=n.synced,e(`${r}: target content drifted from disk by ${Math.abs(k.length-s.content.length)} chars \u2014 synced to disk truth (v${n.version}). Prior verdicts against the old text dissolve via content-hash keying.`),s={...s,content:k,version:n.version}}}catch(n){e(`${r}: disk sync failed (${n instanceof Error?n.message:String(n)}) \u2014 measuring the stored snapshot`)}let x=!1;try{x=(await y("/api/settings/evolution-account")).paused}catch(n){e(`${r}: could not read the account switch (${n instanceof Error?n.message:String(n)}) \u2014 pausing to be safe`),x=!0}if(x)return e(`${r}: PAUSED for the whole account (profiles.evolution_paused) \u2014 no model budget spent.`),{skipped:"paused-account"};if(ge(c))return e(`${r}: PAUSED (metadata.autoEvolvePaused) \u2014 no model budget spent. Opt-in is intact.`),{skipped:"paused"};let S=_e(s);if(S)return e(`${r}: ${S.reason}`),S;let L=t.apply??(c.autoEvolveApply===!0||c.autoEvolveApply==="true"),F=nt(c);if(F==null){if(L)return e(`${r}: refusing \u2014 metadata.immutableSections is unset. Set it (use [] to explicitly protect nothing) before letting the loop edit this target unattended.`),{skipped:"immutable-sections-unset"};e(`${r}: metadata.immutableSections unset \u2014 measure-only cycle proceeds with [] (nothing protected). Writing still requires setting it explicitly.`),F=[]}let v=(await y(`/api/eval-targets/${i}/evals`)).filter(n=>!n.retiredAt);if(v.length===0)return{skipped:"no-live-cases"};let j=me(c);j.note&&e(`${r}: ${j.note}`);let C=Qe(v.length,j.minPairs,Be(j,v.length));if(!C.feasible)return e(`${r}: cannot reach a sufficient verdict \u2014 ${C.reason}`),{skipped:"cannot-reach-min-n",reason:C.reason};let Fe=fe({liveCases:v.length,passes:C.passes});e(`${r}: budget for this run \u2014 ${$e(Fe)}`);let G=await y(`/api/eval-targets/${i}/judge-calibration`);if(!G.powerCheck.adequate){let n=(G.powerCheck.reasons??[]).join("; ");return e(`${r}: not adequately powered, skipping before spending budget \u2014 ${n}`),{skipped:"not-adequate",reason:n}}let oe=de,je=Date.now()-oe*24*60*60*1e3,X=await y(`/api/eval-targets/${i}/promotions?limit=200`),ie=X.promotions.filter(n=>n.applied&&Date.parse(n.createdAt)>=je).length;if(ie>=ne)return e(`${r}: change budget looks exhausted (${ie}/${ne} applied in ${oe}d), skipping before spending budget`),{skipped:"budget-exhausted"};try{let n=await y("/api/usage/current-limits"),u=Re({limits:n,now:new Date});if(!u.ok)return e(`${r}: deferring before spending budget \u2014 ${u.reason}`),{skipped:"quota-exhausted",reason:u.reason}}catch(n){e(`${r}: quota unknown (${n instanceof Error?n.message:String(n)}) \u2014 proceeding`)}let U=ue();if(!U.green)return e(`${r}: A/A canary is RED \u2014 halting. ${U.failures.join(" | ")}`),{skipped:"canary-red",reason:U.failures.join(" | ")};let z=s.content,D=H(z),f=et(c,t),w=f.enabled?re(X.promotions,D,f.targetPairs,d):null,b,q,$=[],_=0,B=pe,p=w?null:Se(c,D,d);if(p){b=p.content,q="",B=0;try{$=await Ie({evalTargetId:i,candidateContentHash:p.hash,executionKey:d,observedKey:p.observedKey})}catch(n){return e(`${r}: could not rebuild banked pairs (${n instanceof Error?n.message:String(n)})`),{skipped:"resume-read-failed",reason:p.hash}}e(`${r}: resuming interrupted cycle \u2014 ${$.length}/${p.plannedPairs} pair(s) rebuilt from eval_runs, no proposal call spent`)}else if(w){let n=await y(`/api/promotions/${w.promotionId}`);if(H(n.toContent)!==w.candidateContentHash)return e(`${r}: stored candidate does not match its recorded hash \u2014 refusing to continue`),{skipped:"accumulation-hash-mismatch",reason:w.promotionId};b=n.toContent,q=n.proposalRationale??"",$=w.pairs,_=w.runs,B=0,e(`${r}: continuing accumulation ${w.promotionId} \u2014 ${$.length}/${f.targetPairs} pair(s) carried from ${_} prior run(s)`)}else{let n=G.caseDiscrimination?.cases??[],u=[];if(s.name.startsWith("/")){let a=s.name.slice(0,s.name.lastIndexOf("/"));try{let m=await y(`/api/insights/evidence?projectPath=${encodeURIComponent(a)}`);m.analyzedAt&&(u=[...m.friction.map(ee=>`Real-session friction (insights, ${m.analyzedAt.slice(0,10)}): ${ee}`),...m.whatWorks.map(ee=>`Working well in real sessions \u2014 do not break: ${ee}`)],u.length>0&&e(`${r}: proposer sees ${m.friction.length} friction + ${m.whatWorks.length} what-works item(s) from insights`))}catch(m){e(`${r}: insights evidence unavailable (${m instanceof Error?m.message:String(m)}) \u2014 proposing without it`)}}let N=[];try{N=Ze(X.promotions,D)}catch(a){e(`${r}: could not summarise prior decisions (${a instanceof Error?a.message:String(a)}) \u2014 proposing without them`)}N.length>0&&e(`${r}: proposer sees ${N.length} prior decision(s) against this exact body (${N.map(a=>`${a.decision}${a.measured?"/measured":""}`).join(", ")})`);let W=await Me({currentContent:s.content,targetType:s.type,targetName:s.name,immutableSections:F,evidence:{nonDiscriminating:n.filter(a=>!a.discriminating).map(a=>`${a.name??a.assertion??"unnamed"} \u2014 ${a.note}`),weakest:n.filter(a=>a.discriminating&&a.withPassRate<a.withoutPassRate).map(a=>({name:a.name??a.assertion??"unnamed",withPassRate:a.withPassRate,withoutPassRate:a.withoutPassRate})),priorDecisions:N,notes:u}},void 0,t.measurement?(a,m)=>we(t.measurement,a,m):void 0);if(!W.ok)return e(`${r}: no usable proposal \u2014 ${W.reason}`),{skipped:"no-proposal",reason:W.reason};b=W.proposal.candidateContent,q=W.proposal.rationale}let I=p?De({plannedPairs:p.plannedPairs||C.expectedPairs,carriedPairs:$.length,liveCases:v.length}):C.passes;e(`${r}: measuring ${v.length} case(s) x ${I} pass(es) = up to ${v.length*I} pairs, candidate vs prior version`+(p?` (resumed; ${$.length} pair(s) already banked)`:""));let E=[],K=[],Ue=Date.now(),le=0,Q=0;for(let n=1;n<=I;n++){let u=await Ae({cases:v,candidateContent:b,priorVersionContent:z,concurrency:t.concurrency??2,runNumber:n,opts:{targetId:i,targetVersion:s.version??1,iteration:1,timeout:t.timeout??24e4,model:t.model,graderModel:t.graderModel,graderEngine:t.graderEngine,...l,executionKey:d,skipGrading:!1,targetType:s.type,targetName:s.name,candidateContentHash:H(b)},onLog:e});if(E.push(...u.pairs),K.push(...u.results),le+=u.completed,Q+=u.failed,e(`${r}: pass ${n}/${I} \u2014 ${u.pairs.length} pair(s), running total ${E.length}`),u.abortedForQuota){e(`${r}: stopping after pass ${n}/${I} \u2014 ${u.abortedForQuota}. Remaining passes would spend against an exhausted account.`);break}}let O=ke(K),qe=p?.observedKey??w?.observedKey;if(t.measurement&&E.length>0&&(!O||$.length>0&&qe!==O))return{skipped:"incomparable-execution",reason:"Actual actor/judge models or delivery changed; recorded runs are retained, but incompatible evidence cannot be pooled or promoted."};let A=[...$,...E],T=Ce(K),V=E.length>0,h=he({runs:K,plannedPairs:C.expectedPairs,usablePairs:E.length,verdictPairs:V?A.length:E.length,proposalCalls:B,passes:I,liveCases:v.length,wallMs:Date.now()-Ue});if(e(`${r}: ${le} completed, ${Q} failed, ${E.length} usable pair(s) \u2014 spent ~${h.modelCalls} model calls (${h.proposalCalls} proposal + ${h.executorCalls} executor + >=${h.graderCalls} grader) in ${(h.wallMs/6e4).toFixed(1)} min`+(V?`; the ${h.verdictPairs} pair(s) behind this verdict resolve +/-${be(h.mde)} at 80% power`:"")),!V)return{skipped:T.infrastructureFailed?"infrastructure-failure":"no-pairs",reason:T.infrastructureFailed?`${T.infrastructureFailed} run(s) never executed (${T.summary.describe()}) \u2014 the environment failed, not the target`:`${Q} run(s) failed or were ungradeable`,cost:h};let Y=Ee({pairs:A.length,minPairs:10,accounting:T});if(Y.kind==="infrastructure-incomplete")return e(`${r}: not posting a verdict \u2014 ${Y.reason}`),await se({evalTargetId:i,meta:c,blob:Oe({executionKey:d,...O?{observedKey:O}:{},content:b,hash:H(b),priorHash:D,plannedPairs:p?.plannedPairs||C.expectedPairs,now:new Date}),log:e,tag:r}),{skipped:"infrastructure-incomplete",reason:Y.reason,cost:h};f.enabled&&e(`${r}: accumulated ${A.length}/${f.targetPairs} pair(s) over ${_+1} run(s) on one candidate`);let ce=L,{apply:Ke,peeking:Je}=tt({consented:ce,accumulateEnabled:f.enabled,pairCount:A.length,targetPairs:f.targetPairs});ce&&Je&&e(`${r}: interim readout at ${A.length}/${f.targetPairs} pairs \u2014 recording the decision but not acting until the target is reached`);let P=await te(`/api/eval-targets/${i}/promotions`,{candidateContent:b,proposalSource:"llm",proposalRationale:q,pairs:A,canaryGreen:U.green,measuredAgainstContentHash:D,immutableSections:F,apply:Ke,measurementCost:h,...f.enabled?{accumulation:{executionKey:d,...O?{observedKey:O}:{},candidateContentHash:H(b),priorContentHash:D,targetPairs:f.targetPairs,carriedPairs:$.length,runs:_+1}}:{}});e(`${r}: ${P.decision}${P.applied?" (APPLIED)":""} \u2014 ${P.rationale}`+(P.refusals.length>0?` | refusals: ${P.refusals.join("; ")}`:"")),(p||c[xe])&&await se({evalTargetId:i,meta:c,blob:null,log:e,tag:r});let Z=!1,R;if(P.applied&&s.name.startsWith("/"))try{Te(s.name)?Ne(s.name,"utf8")!==z?R="file changed since measurement started \u2014 refusing to clobber; next cycle re-syncs":(Xe(s.name,b,"utf8"),Z=!0):R="file no longer exists on disk",e(Z?`${r}: measured improvement written to ${s.name} (undo: git checkout -- ${s.name})`:`${r}: NOT written to disk \u2014 ${R}`)}catch(n){R=n instanceof Error?n.message:String(n),e(`${r}: disk write-back failed \u2014 ${R}. The eval target is updated; the file is not.`)}return{decision:P.decision,applied:P.applied,nPairs:A.length,promotionId:P.id,cost:h,diskSynced:M,...P.applied?{fileWritten:Z,...R?{fileWriteSkipped:R}:{}}:{},...f.enabled?{carriedPairs:$.length,accumulationRuns:_+1}:{}}}g(Pt,"cronInstructionEvolution");function nt(t){let o=t.immutableSections;return o==null||!Array.isArray(o)?null:o.filter(e=>typeof e=="string")}g(nt,"readImmutableSections");export{Le as a,ae as b,ze as c,Be as d,Qe as e,Ve as f,Ye as g,Ze as h,et as i,tt as j,Pt as k,nt as l};
+import{a as _e}from"./chunk-M7QSC64D.js";import{i as Me}from"./chunk-DZRBB34J.js";import{a as Pe}from"./chunk-KR32Y5Q5.js";import{c as Re}from"./chunk-G7V4ZZZ3.js";import{a as xe,b as Se,c as De,d as Ie,e as se,f as Oe}from"./chunk-RITL4NU3.js";import{b as pe,i as me,j as ge,k as fe,l as he,m as be,n as $e}from"./chunk-GYKYJJNF.js";import{a as we}from"./chunk-V6L25LFD.js";import{a as re}from"./chunk-AHLVY52I.js";import{d as ke,f as Ce,g as Ee,i as Ae}from"./chunk-PPQFN62M.js";import{c as ye}from"./chunk-NAS7FIGM.js";import{b as ve}from"./chunk-VZYNZS5U.js";import{d as ue,e as ne,f as de}from"./chunk-KFNFLMUR.js";import{e as y,f as te}from"./chunk-5VWVQCNR.js";import{a as g}from"./chunk-ZLLER3HP.js";import{createHash as Ge}from"node:crypto";import{existsSync as Te,readFileSync as Ne,writeFileSync as Xe}from"node:fs";var H=g(t=>Ge("sha256").update(t).digest("hex"),"sha256"),Le=1.3,ae=4,ze=20;function Be(t,o){if(!t.configured||o<=0)return ae;let e=Math.ceil(t.minPairs*Le/o);return Math.min(ze,Math.max(ae,e))}g(Be,"passCapFor");function Qe(t,o=10,e=ae){if(t<=0)return{passes:0,expectedPairs:0,feasible:!1,reason:"no live cases"};let i=Math.ceil(o*Le),l=Math.min(e,Math.ceil(i/t)),r=l*t;return r<o?{passes:l,expectedPairs:r,feasible:!1,reason:`${t} live case(s) x ${e} passes = ${t*e} pairs at best, below the ${o} a controlled verdict requires \u2014 add cases rather than replaying these`}:{passes:l,expectedPairs:r,feasible:!0}}g(Qe,"planPasses");var Ve=60,Ye=5,We=g(t=>Array.isArray(t)?t.filter(o=>typeof o=="string"):[],"stringList"),J=g(t=>typeof t=="number"&&Number.isFinite(t)?t:null,"finite");function Ze(t,o,e=Ye){let i=[...t].filter(s=>s.fromContentHash===o).sort((s,c)=>Date.parse(c.createdAt)-Date.parse(s.createdAt)),l=[],r=new Set;for(let s of i){let c=s.toContentHash??s.id;if(r.has(c))continue;r.add(c);let d=s.verdict??null,k=J(d?.delta),M=J(d?.ci?.lower),x=J(d?.ci?.upper),S=J(d?.nWith),L=d?.sufficient===!0&&k!=null&&M!=null&&x!=null&&S!=null?{delta:k,ciLower:M,ciUpper:x,nPairs:S}:null;if(l.push({decision:typeof s.decision=="string"?s.decision:"unknown",applied:s.applied===!0,at:typeof s.createdAt=="string"?s.createdAt:"",rationale:typeof s.proposalRationale=="string"?s.proposalRationale:"",touched:We(s.touchedSections),refusals:We(s.refusals),measured:L}),l.length>=e)break}return l}g(Ze,"summarisePriorDecisions");function et(t,o){let e=o.accumulate??(t.autoEvolveAccumulate===!0||t.autoEvolveAccumulate==="true"),i=o.accumulateTargetPairs??t.autoEvolveAccumulateTarget,l=typeof i=="string"?Number.parseInt(i,10):i,r=typeof l=="number"&&Number.isInteger(l)&&l>0?l:Ve;return{enabled:e,targetPairs:r}}g(et,"readAccumulateConfig");function tt(t){let o=t.accumulateEnabled&&t.pairCount<t.targetPairs;return{apply:t.consented&&!o,peeking:o}}g(tt,"resolveApply");async function Pt(t,o){if(t.restoration)return Pe(t.evalTargetId,t.restoration);let e=g(n=>o?.logger?.info?.(n)??console.log(n),"log"),{evalTargetId:i}=t,l=t.measurement?ye(t):{},r=`[cron_instruction_evolution] ${i}`,s=await y(`/api/eval-targets/${i}`),c=s.metadata??{},d=t.measurement?JSON.stringify([l.engine,l.model,l.graderEngine,l.graderModel,s.type,ve(t.measurement.engine)]):void 0,k=null,M=!1;if(s.name.startsWith("/")&&Te(s.name))try{if(k=Ne(s.name,"utf8"),k!==s.content){let n=await te(`/api/eval-targets/${i}/sync-content`,{content:k});M=n.synced,e(`${r}: target content drifted from disk by ${Math.abs(k.length-s.content.length)} chars \u2014 synced to disk truth (v${n.version}). Prior verdicts against the old text dissolve via content-hash keying.`),s={...s,content:k,version:n.version}}}catch(n){e(`${r}: disk sync failed (${n instanceof Error?n.message:String(n)}) \u2014 measuring the stored snapshot`)}let x=!1;try{x=(await y("/api/settings/evolution-account")).paused}catch(n){e(`${r}: could not read the account switch (${n instanceof Error?n.message:String(n)}) \u2014 pausing to be safe`),x=!0}if(x)return e(`${r}: PAUSED for the whole account (profiles.evolution_paused) \u2014 no model budget spent.`),{skipped:"paused-account"};if(ge(c))return e(`${r}: PAUSED (metadata.autoEvolvePaused) \u2014 no model budget spent. Opt-in is intact.`),{skipped:"paused"};let S=_e(s);if(S)return e(`${r}: ${S.reason}`),S;let L=t.apply??(c.autoEvolveApply===!0||c.autoEvolveApply==="true"),F=nt(c);if(F==null){if(L)return e(`${r}: refusing \u2014 metadata.immutableSections is unset. Set it (use [] to explicitly protect nothing) before letting the loop edit this target unattended.`),{skipped:"immutable-sections-unset"};e(`${r}: metadata.immutableSections unset \u2014 measure-only cycle proceeds with [] (nothing protected). Writing still requires setting it explicitly.`),F=[]}let v=(await y(`/api/eval-targets/${i}/evals`)).filter(n=>!n.retiredAt);if(v.length===0)return{skipped:"no-live-cases"};let j=me(c);j.note&&e(`${r}: ${j.note}`);let C=Qe(v.length,j.minPairs,Be(j,v.length));if(!C.feasible)return e(`${r}: cannot reach a sufficient verdict \u2014 ${C.reason}`),{skipped:"cannot-reach-min-n",reason:C.reason};let Fe=fe({liveCases:v.length,passes:C.passes});e(`${r}: budget for this run \u2014 ${$e(Fe)}`);let G=await y(`/api/eval-targets/${i}/judge-calibration`);if(!G.powerCheck.adequate){let n=(G.powerCheck.reasons??[]).join("; ");return e(`${r}: not adequately powered, skipping before spending budget \u2014 ${n}`),{skipped:"not-adequate",reason:n}}let oe=de,je=Date.now()-oe*24*60*60*1e3,X=await y(`/api/eval-targets/${i}/promotions?limit=200`),ie=X.promotions.filter(n=>n.applied&&Date.parse(n.createdAt)>=je).length;if(ie>=ne)return e(`${r}: change budget looks exhausted (${ie}/${ne} applied in ${oe}d), skipping before spending budget`),{skipped:"budget-exhausted"};try{let n=await y("/api/usage/current-limits"),u=Re({limits:n,now:new Date});if(!u.ok)return e(`${r}: deferring before spending budget \u2014 ${u.reason}`),{skipped:"quota-exhausted",reason:u.reason}}catch(n){e(`${r}: quota unknown (${n instanceof Error?n.message:String(n)}) \u2014 proceeding`)}let U=ue();if(!U.green)return e(`${r}: A/A canary is RED \u2014 halting. ${U.failures.join(" | ")}`),{skipped:"canary-red",reason:U.failures.join(" | ")};let z=s.content,D=H(z),f=et(c,t),w=f.enabled?re(X.promotions,D,f.targetPairs,d):null,b,q,$=[],_=0,B=pe,p=w?null:Se(c,D,d);if(p){b=p.content,q="",B=0;try{$=await Ie({evalTargetId:i,candidateContentHash:p.hash,executionKey:d,observedKey:p.observedKey})}catch(n){return e(`${r}: could not rebuild banked pairs (${n instanceof Error?n.message:String(n)})`),{skipped:"resume-read-failed",reason:p.hash}}e(`${r}: resuming interrupted cycle \u2014 ${$.length}/${p.plannedPairs} pair(s) rebuilt from eval_runs, no proposal call spent`)}else if(w){let n=await y(`/api/promotions/${w.promotionId}`);if(H(n.toContent)!==w.candidateContentHash)return e(`${r}: stored candidate does not match its recorded hash \u2014 refusing to continue`),{skipped:"accumulation-hash-mismatch",reason:w.promotionId};b=n.toContent,q=n.proposalRationale??"",$=w.pairs,_=w.runs,B=0,e(`${r}: continuing accumulation ${w.promotionId} \u2014 ${$.length}/${f.targetPairs} pair(s) carried from ${_} prior run(s)`)}else{let n=G.caseDiscrimination?.cases??[],u=[];if(s.name.startsWith("/")){let a=s.name.slice(0,s.name.lastIndexOf("/"));try{let m=await y(`/api/insights/evidence?projectPath=${encodeURIComponent(a)}`);m.analyzedAt&&(u=[...m.friction.map(ee=>`Real-session friction (insights, ${m.analyzedAt.slice(0,10)}): ${ee}`),...m.whatWorks.map(ee=>`Working well in real sessions \u2014 do not break: ${ee}`)],u.length>0&&e(`${r}: proposer sees ${m.friction.length} friction + ${m.whatWorks.length} what-works item(s) from insights`))}catch(m){e(`${r}: insights evidence unavailable (${m instanceof Error?m.message:String(m)}) \u2014 proposing without it`)}}let N=[];try{N=Ze(X.promotions,D)}catch(a){e(`${r}: could not summarise prior decisions (${a instanceof Error?a.message:String(a)}) \u2014 proposing without them`)}N.length>0&&e(`${r}: proposer sees ${N.length} prior decision(s) against this exact body (${N.map(a=>`${a.decision}${a.measured?"/measured":""}`).join(", ")})`);let W=await Me({currentContent:s.content,targetType:s.type,targetName:s.name,immutableSections:F,evidence:{nonDiscriminating:n.filter(a=>!a.discriminating).map(a=>`${a.name??a.assertion??"unnamed"} \u2014 ${a.note}`),weakest:n.filter(a=>a.discriminating&&a.withPassRate<a.withoutPassRate).map(a=>({name:a.name??a.assertion??"unnamed",withPassRate:a.withPassRate,withoutPassRate:a.withoutPassRate})),priorDecisions:N,notes:u}},void 0,t.measurement?(a,m)=>we(t.measurement,a,m):void 0);if(!W.ok)return e(`${r}: no usable proposal \u2014 ${W.reason}`),{skipped:"no-proposal",reason:W.reason};b=W.proposal.candidateContent,q=W.proposal.rationale}let I=p?De({plannedPairs:p.plannedPairs||C.expectedPairs,carriedPairs:$.length,liveCases:v.length}):C.passes;e(`${r}: measuring ${v.length} case(s) x ${I} pass(es) = up to ${v.length*I} pairs, candidate vs prior version`+(p?` (resumed; ${$.length} pair(s) already banked)`:""));let E=[],K=[],Ue=Date.now(),le=0,Q=0;for(let n=1;n<=I;n++){let u=await Ae({cases:v,candidateContent:b,priorVersionContent:z,concurrency:t.concurrency??2,runNumber:n,opts:{targetId:i,targetVersion:s.version??1,iteration:1,timeout:t.timeout??24e4,model:t.model,graderModel:t.graderModel,graderEngine:t.graderEngine,...l,executionKey:d,skipGrading:!1,targetType:s.type,targetName:s.name,candidateContentHash:H(b)},onLog:e});if(E.push(...u.pairs),K.push(...u.results),le+=u.completed,Q+=u.failed,e(`${r}: pass ${n}/${I} \u2014 ${u.pairs.length} pair(s), running total ${E.length}`),u.abortedForQuota){e(`${r}: stopping after pass ${n}/${I} \u2014 ${u.abortedForQuota}. Remaining passes would spend against an exhausted account.`);break}}let O=ke(K),qe=p?.observedKey??w?.observedKey;if(t.measurement&&E.length>0&&(!O||$.length>0&&qe!==O))return{skipped:"incomparable-execution",reason:"Actual actor/judge models or delivery changed; recorded runs are retained, but incompatible evidence cannot be pooled or promoted."};let A=[...$,...E],T=Ce(K),V=E.length>0,h=he({runs:K,plannedPairs:C.expectedPairs,usablePairs:E.length,verdictPairs:V?A.length:E.length,proposalCalls:B,passes:I,liveCases:v.length,wallMs:Date.now()-Ue});if(e(`${r}: ${le} completed, ${Q} failed, ${E.length} usable pair(s) \u2014 spent ~${h.modelCalls} model calls (${h.proposalCalls} proposal + ${h.executorCalls} executor + >=${h.graderCalls} grader) in ${(h.wallMs/6e4).toFixed(1)} min`+(V?`; the ${h.verdictPairs} pair(s) behind this verdict resolve +/-${be(h.mde)} at 80% power`:"")),!V)return{skipped:T.infrastructureFailed?"infrastructure-failure":"no-pairs",reason:T.infrastructureFailed?`${T.infrastructureFailed} run(s) never executed (${T.summary.describe()}) \u2014 the environment failed, not the target`:`${Q} run(s) failed or were ungradeable`,cost:h};let Y=Ee({pairs:A.length,minPairs:10,accounting:T});if(Y.kind==="infrastructure-incomplete")return e(`${r}: not posting a verdict \u2014 ${Y.reason}`),await se({evalTargetId:i,meta:c,blob:Oe({executionKey:d,...O?{observedKey:O}:{},content:b,hash:H(b),priorHash:D,plannedPairs:p?.plannedPairs||C.expectedPairs,now:new Date}),log:e,tag:r}),{skipped:"infrastructure-incomplete",reason:Y.reason,cost:h};f.enabled&&e(`${r}: accumulated ${A.length}/${f.targetPairs} pair(s) over ${_+1} run(s) on one candidate`);let ce=L,{apply:Ke,peeking:Je}=tt({consented:ce,accumulateEnabled:f.enabled,pairCount:A.length,targetPairs:f.targetPairs});ce&&Je&&e(`${r}: interim readout at ${A.length}/${f.targetPairs} pairs \u2014 recording the decision but not acting until the target is reached`);let P=await te(`/api/eval-targets/${i}/promotions`,{candidateContent:b,proposalSource:"llm",proposalRationale:q,pairs:A,canaryGreen:U.green,measuredAgainstContentHash:D,immutableSections:F,apply:Ke,measurementCost:h,...f.enabled?{accumulation:{executionKey:d,...O?{observedKey:O}:{},candidateContentHash:H(b),priorContentHash:D,targetPairs:f.targetPairs,carriedPairs:$.length,runs:_+1}}:{}});e(`${r}: ${P.decision}${P.applied?" (APPLIED)":""} \u2014 ${P.rationale}`+(P.refusals.length>0?` | refusals: ${P.refusals.join("; ")}`:"")),(p||c[xe])&&await se({evalTargetId:i,meta:c,blob:null,log:e,tag:r});let Z=!1,R;if(P.applied&&s.name.startsWith("/"))try{Te(s.name)?Ne(s.name,"utf8")!==z?R="file changed since measurement started \u2014 refusing to clobber; next cycle re-syncs":(Xe(s.name,b,"utf8"),Z=!0):R="file no longer exists on disk",e(Z?`${r}: measured improvement written to ${s.name} (undo: git checkout -- ${s.name})`:`${r}: NOT written to disk \u2014 ${R}`)}catch(n){R=n instanceof Error?n.message:String(n),e(`${r}: disk write-back failed \u2014 ${R}. The eval target is updated; the file is not.`)}return{decision:P.decision,applied:P.applied,nPairs:A.length,promotionId:P.id,cost:h,diskSynced:M,...P.applied?{fileWritten:Z,...R?{fileWriteSkipped:R}:{}}:{},...f.enabled?{carriedPairs:$.length,accumulationRuns:_+1}:{}}}g(Pt,"cronInstructionEvolution");function nt(t){let o=t.immutableSections;return o==null||!Array.isArray(o)?null:o.filter(e=>typeof e=="string")}g(nt,"readImmutableSections");export{Le as a,ae as b,ze as c,Be as d,Qe as e,Ve as f,Ye as g,Ze as h,et as i,tt as j,Pt as k,nt as l};

package/dist/{chunk-YLEXR62G.js → chunk-LSZ32DUR.js}

@@ -1,11 +1,11 @@
 import{e as d}from"./chunk-QWCEZEDO.js";import{a as R,g as v,h as T,j as I,m as C}from"./chunk-OZQNSELV.js";import{d as g}from"./chunk-GPBN4ROE.js";import{e as _}from"./chunk-5VWVQCNR.js";import{g as m,l as k}from"./chunk-PZNWH4VQ.js";import{a as i}from"./chunk-ZLLER3HP.js";import{execSync as j}from"node:child_process";import{existsSync as p,readFileSync as y}from"node:fs";import{homedir as f}from"node:os";import{join as a}from"node:path";var P=["install","auth","hooks","worker"],O=["judge"],M=[...P,...O];function $(){return{projectPath:process.cwd(),configDir:k()}}i($,"defaultContext");var A=a(f(),".mitmproxy","mitmproxy-ca-cert.pem");function S(e){if(!e)return"";if(e.length<=8)return"*".repeat(e.length);let t=e.slice(0,4),n=e.slice(-4);return`${t}${"*".repeat(Math.max(4,e.length-8))}${n}`}i(S,"maskToken");function te(e,t){let n=e===null?[]:e.replace(/\n$/,"").split(`
 `),o=t===null?[]:t.replace(/\n$/,"").split(`
-`),s=[],u=Math.max(n.length,o.length);for(let c=0;c<u;c++){let l=n[c],r=o[c];l===r?s.push(` ${l}`):(l!==void 0&&s.push(`-${l}`),r!==void 0&&s.push(`+${r}`))}return s.join(`
-`)}i(te,"renderDiff");function x(e,t){let n=a(e,"env");if(!p(n))return null;try{let s=y(n,"utf-8").match(new RegExp(`^${t}=(.+)$`,"m"));return s?s[1]:null}catch{return null}}i(x,"readEnvValue");async function b(){if(!m())return!1;try{return(await Promise.race([_("/api/setup/status"),new Promise((t,n)=>setTimeout(()=>n(new Error("setup/status timeout")),2500))]))?.workerActive===!0}catch{return!1}}i(b,"workerActive");function E(e,t){let n=a(e,".claude",t),o={},s=null;if(p(n))try{s=y(n,"utf-8").replace(/\n$/,""),o=JSON.parse(s)}catch{o={}}return{settings:o,before:s}}i(E,"readProjectSettings");function D(e){let{settings:t,before:n}=E(e,"settings.local.json");return v(t),{before:n,after:JSON.stringify(t,null,2)}}i(D,"projectHooksSettings");function N(e){let{settings:t,before:n}=E(e,"settings.local.json"),o={...t.env??{}};return o.HTTPS_PROXY=`http://localhost:${d}`,o.NO_PROXY=I,o.NODE_EXTRA_CA_CERTS=A,{before:n,after:JSON.stringify({...t,env:o},null,2)}}i(N,"projectProxyEnvSettings");function U(e){let{settings:t,before:n}=E(e,"settings.json");if(n===null)return null;let{removed:o}=C(t,d);return o?{before:n,after:JSON.stringify(t,null,2)}:null}i(U,"projectSharedSettingsMigration");var Y=a(f(),"Library","LaunchAgents","com.rulemetric.worker.plist"),G=a(f(),".config","systemd","user","rulemetric-worker.service");function h(e){j(e,{stdio:"inherit"})}i(h,"runCommand");var L=[{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",async detect(){return m()||await b()},async plan(){return{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",artifacts:[{path:"npm install -g @rulemetric/cli",kind:"command",before:null,after:"rulemetric --version   # verifies the binary is on PATH"}]}},async apply(){h("npm install -g @rulemetric/cli")}},{id:"auth",title:"Authenticate",command:"rulemetric auth login",async detect(){return m()},async plan(e){let t=a(e.configDir,"env"),n=null;if(p(t))try{n=X(y(t,"utf-8"))}catch{n="(unreadable)"}let o=x(e.configDir,"RULEMETRIC_ACCESS_TOKEN"),u=[`RULEMETRIC_ACCESS_TOKEN=${S(o||"<token-from-login>")}`],c=x(e.configDir,"RULEMETRIC_API_KEY");c&&u.push(`RULEMETRIC_API_KEY=${S(c)}`);let l=process.env.RULEMETRIC_API_URL;l&&u.push(`RULEMETRIC_API_URL=${l}`);let r=u.join(`
-`);return{id:"auth",title:"Authenticate",command:"rulemetric auth login",artifacts:[{path:t,kind:"file",before:n,after:r,masked:!0},{path:a(e.configDir,"auth.json"),kind:"file",before:p(a(e.configDir,"auth.json"))?"(existing session tokens)":null,after:"{ accessToken, refreshToken, expiresAt, email }  \u2014 OAuth session tokens (secret)",masked:!0}]}},async apply(){h("rulemetric auth login")}},{id:"hooks",title:"Install hooks + capture proxy",command:process.env.RULEMETRIC_SETUP_NO_PROXY==="1"||process.env.RULEMETRIC_SETUP_NO_PROXY==="true"?"rulemetric hooks install --no-proxy":"rulemetric hooks install",async detect(e){for(let t of["settings.local.json","settings.json"]){let n=a(e.projectPath,".claude",t);if(p(n))try{let o=JSON.parse(y(n,"utf-8"));if(T(o))return!0}catch{}}return!1},async plan(e){let t=a(e.projectPath,".claude","settings.local.json"),{before:n,after:o}=D(e.projectPath),s=[{path:t,kind:"file",before:n,after:o},{path:"what these hooks capture",kind:"service",before:null,after:"session-end reimports the FULL transcript to the API; post-tool-use records each tool call and snapshots the working tree to a LOCAL shadow git ref (paper-trail, never pushed); user-prompt posts prompt text. All scoped to sessions in this project."},{path:`${t} (proxy env block)`,kind:"file",...N(e.projectPath)},{path:"system keychain \u2014 mitmproxy root CA",kind:"service",before:p(A)?"(CA generated; trust state unchanged)":"(no CA yet)",after:"Installs + trusts the mitmproxy CA (requires sudo). This lets the proxy DECRYPT Claude Code HTTPS request/response bodies \u2014 the capability that powers rendered-system-prompt / instruction linking. Decline the sudo prompt and setup stays hooks-only (still fully onboarded)."},{path:"launchctl setenv HTTPS_PROXY (machine-wide, all GUI apps)",kind:"command",before:null,after:`http://localhost:${d}  \u2014 routes every GUI app's HTTPS through the local gateway (only when the CA is trusted AND :${d} is live)`},{path:"VS Code + Cursor user settings (http.proxy)",kind:"file",before:null,after:`http.proxy \u2192 http://localhost:${d} (+ http.proxyStrictSSL=false)`},{path:`gateway process on :${d}`,kind:"service",before:null,after:"started (routes Claude Code \u2192 mitmproxy \u2192 Anthropic; falls back to direct if mitmproxy is down)"}],u=U(e.projectPath);u&&s.push({path:`${a(e.projectPath,".claude","settings.json")} (removes RuleMetric machine-local state left by an older install \u2014 see the local-file artifacts above for where it moves)`,kind:"file",...u});for(let c of R(e.projectPath).filter(l=>l==="cursor"||l==="copilot_agent"))s.push({path:`${e.projectPath} (${c} workspace proxy config)`,kind:"file",before:null,after:`proxy \u2192 http://localhost:${d}`});for(let[c,l,r]of[["Cursor",a(f(),".cursor"),a(f(),".cursor","hooks.json")],["Antigravity/Gemini",a(f(),".gemini"),a(f(),".gemini","config","hooks.json")]])p(l)&&s.push({path:`${r} (${c} user hooks)`,kind:"file",before:p(r)?"(existing config)":null,after:"adds RuleMetric session-capture hooks for this tool"});return{id:"hooks",title:"Install hooks + capture proxy",command:"rulemetric hooks install",artifacts:s}},async apply(){let e=process.env.RULEMETRIC_SETUP_NO_PROXY==="1"||process.env.RULEMETRIC_SETUP_NO_PROXY==="true";h(e?"rulemetric hooks install --no-proxy":"rulemetric hooks install")}},{id:"worker",title:"Run the worker",command:"rulemetric service install --worker-only",async detect(){return b()},async plan(){let e=process.platform==="darwin",t=e?Y:G,n=p(t)?`(service already installed at ${t})`:null,o=e?["Label: com.rulemetric.worker","ProgramArguments: <node> <cli>/bin/run.js evals agent","RunAtLoad: true, KeepAlive: true"].join(`
+`),r=[],u=Math.max(n.length,o.length);for(let c=0;c<u;c++){let l=n[c],s=o[c];l===s?r.push(` ${l}`):(l!==void 0&&r.push(`-${l}`),s!==void 0&&r.push(`+${s}`))}return r.join(`
+`)}i(te,"renderDiff");function x(e,t){let n=a(e,"env");if(!p(n))return null;try{let r=y(n,"utf-8").match(new RegExp(`^${t}=(.+)$`,"m"));return r?r[1]:null}catch{return null}}i(x,"readEnvValue");async function b(){if(!m())return!1;try{return(await Promise.race([_("/api/setup/status"),new Promise((t,n)=>setTimeout(()=>n(new Error("setup/status timeout")),2500))]))?.workerActive===!0}catch{return!1}}i(b,"workerActive");function E(e,t){let n=a(e,".claude",t),o={},r=null;if(p(n))try{r=y(n,"utf-8").replace(/\n$/,""),o=JSON.parse(r)}catch{o={}}return{settings:o,before:r}}i(E,"readProjectSettings");function D(e){let{settings:t,before:n}=E(e,"settings.local.json");return v(t),{before:n,after:JSON.stringify(t,null,2)}}i(D,"projectHooksSettings");function N(e){let{settings:t,before:n}=E(e,"settings.local.json"),o={...t.env??{}};return o.HTTPS_PROXY=`http://localhost:${d}`,o.NO_PROXY=I,o.NODE_EXTRA_CA_CERTS=A,{before:n,after:JSON.stringify({...t,env:o},null,2)}}i(N,"projectProxyEnvSettings");function U(e){let{settings:t,before:n}=E(e,"settings.json");if(n===null)return null;let{removed:o}=C(t,d);return o?{before:n,after:JSON.stringify(t,null,2)}:null}i(U,"projectSharedSettingsMigration");var Y=a(f(),"Library","LaunchAgents","com.rulemetric.worker.plist"),G=a(f(),".config","systemd","user","rulemetric-worker.service");function h(e){j(e,{stdio:"inherit"})}i(h,"runCommand");var L=[{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",async detect(){return m()||await b()},async plan(){return{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",artifacts:[{path:"npm install -g @rulemetric/cli",kind:"command",before:null,after:"rulemetric --version   # verifies the binary is on PATH"}]}},async apply(){h("npm install -g @rulemetric/cli")}},{id:"auth",title:"Authenticate",command:"rulemetric auth login",async detect(){return m()},async plan(e){let t=a(e.configDir,"env"),n=null;if(p(t))try{n=X(y(t,"utf-8"))}catch{n="(unreadable)"}let o=x(e.configDir,"RULEMETRIC_ACCESS_TOKEN"),u=[`RULEMETRIC_ACCESS_TOKEN=${S(o||"<token-from-login>")}`],c=x(e.configDir,"RULEMETRIC_API_KEY");c&&u.push(`RULEMETRIC_API_KEY=${S(c)}`);let l=process.env.RULEMETRIC_API_URL;l&&u.push(`RULEMETRIC_API_URL=${l}`);let s=u.join(`
+`);return{id:"auth",title:"Authenticate",command:"rulemetric auth login",artifacts:[{path:t,kind:"file",before:n,after:s,masked:!0},{path:a(e.configDir,"auth.json"),kind:"file",before:p(a(e.configDir,"auth.json"))?"(existing session tokens)":null,after:"{ accessToken, refreshToken, expiresAt, email }  \u2014 OAuth session tokens (secret)",masked:!0}]}},async apply(){h("rulemetric auth login")}},{id:"hooks",title:"Install hooks + capture proxy",command:process.env.RULEMETRIC_SETUP_NO_PROXY==="1"||process.env.RULEMETRIC_SETUP_NO_PROXY==="true"?"rulemetric hooks install --no-proxy":"rulemetric hooks install",async detect(e){for(let t of new Set([e.projectPath,f()]))for(let n of["settings.local.json","settings.json"]){let o=a(t,".claude",n);if(p(o))try{let r=JSON.parse(y(o,"utf-8"));if(T(r))return!0}catch{}}return!1},async plan(e){let t=a(e.projectPath,".claude","settings.local.json"),{before:n,after:o}=D(e.projectPath),r=[{path:t,kind:"file",before:n,after:o},{path:"what these hooks capture",kind:"service",before:null,after:"session-end reimports the FULL transcript to the API; post-tool-use records each tool call and snapshots the working tree to a LOCAL shadow git ref (paper-trail, never pushed); user-prompt posts prompt text. All scoped to sessions in this project."},{path:`${t} (proxy env block)`,kind:"file",...N(e.projectPath)},{path:"system keychain \u2014 mitmproxy root CA",kind:"service",before:p(A)?"(CA generated; trust state unchanged)":"(no CA yet)",after:"Installs + trusts the mitmproxy CA (requires sudo). This lets the proxy DECRYPT Claude Code HTTPS request/response bodies \u2014 the capability that powers rendered-system-prompt / instruction linking. Decline the sudo prompt and setup stays hooks-only (still fully onboarded)."},{path:"launchctl setenv HTTPS_PROXY (machine-wide, all GUI apps)",kind:"command",before:null,after:`http://localhost:${d}  \u2014 routes every GUI app's HTTPS through the local gateway (only when the CA is trusted AND :${d} is live)`},{path:"VS Code + Cursor user settings (http.proxy)",kind:"file",before:null,after:`http.proxy \u2192 http://localhost:${d} (+ http.proxyStrictSSL=false)`},{path:`gateway process on :${d}`,kind:"service",before:null,after:"started (routes Claude Code \u2192 mitmproxy \u2192 Anthropic; falls back to direct if mitmproxy is down)"}],u=U(e.projectPath);u&&r.push({path:`${a(e.projectPath,".claude","settings.json")} (removes RuleMetric machine-local state left by an older install \u2014 see the local-file artifacts above for where it moves)`,kind:"file",...u});for(let c of R(e.projectPath).filter(l=>l==="cursor"||l==="copilot_agent"))r.push({path:`${e.projectPath} (${c} workspace proxy config)`,kind:"file",before:null,after:`proxy \u2192 http://localhost:${d}`});for(let[c,l,s]of[["Cursor",a(f(),".cursor"),a(f(),".cursor","hooks.json")],["Antigravity/Gemini",a(f(),".gemini"),a(f(),".gemini","config","hooks.json")]])p(l)&&r.push({path:`${s} (${c} user hooks)`,kind:"file",before:p(s)?"(existing config)":null,after:"adds RuleMetric session-capture hooks for this tool"});return{id:"hooks",title:"Install hooks + capture proxy",command:"rulemetric hooks install",artifacts:r}},async apply(){let e=process.env.RULEMETRIC_SETUP_NO_PROXY==="1"||process.env.RULEMETRIC_SETUP_NO_PROXY==="true";h(e?"rulemetric hooks install --no-proxy":"rulemetric hooks install")}},{id:"worker",title:"Run the worker",command:"rulemetric service install --worker-only",async detect(){return b()},async plan(){let e=process.platform==="darwin",t=e?Y:G,n=p(t)?`(service already installed at ${t})`:null,o=e?["Label: com.rulemetric.worker","ProgramArguments: <node> <cli>/bin/run.js evals agent","RunAtLoad: true, KeepAlive: true"].join(`
 `):["[Service]","ExecStart=<node> <cli>/bin/run.js evals agent","Restart=on-failure"].join(`
 `);return{id:"worker",title:"Run the worker",command:"rulemetric service install --worker-only",artifacts:[{path:t,kind:"service",before:n,after:o}]}},async apply(){h("rulemetric service install --worker-only")}},{id:"judge",title:"Second LLM engine for cross-vendor eval grading (optional)",command:"npm install -g @openai/codex",optional:!0,async detect(){let e=g();return e.claudeBin!=null&&e.codexBin!=null},async plan(){let e=g();return{id:"judge",title:"Second LLM engine for cross-vendor eval grading (optional)",command:"npm install -g @openai/codex",artifacts:[{path:"installed LLM engines",kind:"command",before:[`claude CLI: ${e.claudeBin??"not found"}`,`codex CLI: ${e.codexBin??"not found"}`].join(`
 `),after:["claude CLI + codex CLI both installed.","Hand labels + one judge already give calibrated verdicts; a second","engine additionally unlocks cross-vendor judge validation","(`rulemetric evals validate-judge`)."].join(`
 `)}]}},async apply(){let e=g();if(e.claudeBin!=null&&e.codexBin!=null){console.log("[OK] Both engines already installed \u2014 nothing to do.");return}let t=e.codexBin==null?"codex":"claude";console.log(`[..] Second GRADING engine not detected (missing: ${t} CLI).`),t==="codex"?(console.log("     This is about GRADING EVALS with a second vendor, not about capturing"),console.log("     your Codex sessions. Codex capture needs no setup: the worker imports"),console.log("     ~/.codex/sessions automatically. Nothing is missing for that."),console.log("     Optional install: npm install -g @openai/codex")):console.log("     Install it yourself when ready: npm install -g @anthropic-ai/claude-code"),console.log("     Then re-run: rulemetric setup judge --show")}}],H=new Map(L.map(e=>[e.id,e]));function ne(e){let t=H.get(e);if(!t)throw new Error(`Unknown setup step: ${e}`);return t}i(ne,"getStep");function oe(e){return M.includes(e)}i(oe,"isStepId");var K=new Set(["RULEMETRIC_API_URL","RULEMETRIC_USER_ID","RULEMETRIC_CLIENT_NAME","RESEND_FROM_EMAIL","SENTRY_DSN","PG_POOL_MAX","PG_IDLE_TIMEOUT","PG_STATEMENT_TIMEOUT_MS","NODE_ENV"]);function X(e){return e.replace(/\n$/,"").split(`
-`).map(t=>{let n=t.match(/^([A-Z0-9_]+)=(.+)$/);if(!n)return t;let[,o,s]=n;return K.has(o)?t:`${o}=${S(s)}`}).join(`
-`)}i(X,"maskEnvFile");async function re(e=$()){let t=[];for(let r of L){let w=await r.detect(e);t.push({id:r.id,done:w})}let n=new Set(P),o=t.find(r=>r.id==="install");o&&!o.done&&(o.done=t.some(r=>r.id!=="install"&&n.has(r.id)&&r.done));let s=new Set(P),u=t.filter(r=>s.has(r.id)),c=u.find(r=>!r.done)?.id??null,l=u.every(r=>r.done);return{steps:t,coreComplete:l,currentStepId:c}}i(re,"computeStatus");export{P as a,O as b,M as c,$ as d,S as e,te as f,L as g,ne as h,oe as i,re as j};
+`).map(t=>{let n=t.match(/^([A-Z0-9_]+)=(.+)$/);if(!n)return t;let[,o,r]=n;return K.has(o)?t:`${o}=${S(r)}`}).join(`
+`)}i(X,"maskEnvFile");async function re(e=$()){let t=[];for(let s of L){let w=await s.detect(e);t.push({id:s.id,done:w})}let n=new Set(P),o=t.find(s=>s.id==="install");o&&!o.done&&(o.done=t.some(s=>s.id!=="install"&&n.has(s.id)&&s.done));let r=new Set(P),u=t.filter(s=>r.has(s.id)),c=u.find(s=>!s.done)?.id??null,l=u.every(s=>s.done);return{steps:t,coreComplete:l,currentStepId:c}}i(re,"computeStatus");export{P as a,O as b,M as c,$ as d,S as e,te as f,L as g,ne as h,oe as i,re as j};

package/dist/{chunk-JLX2D3QL.js → chunk-NVOYP4HX.js}

@@ -1 +1 @@
-import{c as te}from"./chunk-5U5SHJJA.js";import{b as oe}from"./chunk-IEQDLURA.js";import{d as ie,e as le,f as de}from"./chunk-KINIFUZQ.js";import{a as Y}from"./chunk-QRKMRG6W.js";import{a as B}from"./chunk-D7L7IMJQ.js";import{a as x}from"./chunk-MRQIYIQE.js";import{a as W}from"./chunk-PMMHMRFS.js";import{a as K}from"./chunk-IYX65N6T.js";import{a as V}from"./chunk-GHUW7P55.js";import{a as U}from"./chunk-MBP5OJKE.js";import{a as z}from"./chunk-BRYURAUN.js";import{b as F}from"./chunk-NLUHMHPM.js";import{a as Q}from"./chunk-DTTR4EUJ.js";import{a as X}from"./chunk-WLLIFRRE.js";import{a as G}from"./chunk-ZYGBABMZ.js";import{b as $}from"./chunk-5RBLIVBF.js";import{a as se}from"./chunk-NSO2EBYE.js";import{e as T}from"./chunk-2C7CFCMM.js";import{b as q}from"./chunk-WSUKOU3X.js";import{c as C}from"./chunk-A6XQSO62.js";import{a as ee}from"./chunk-IPNBW7ZW.js";import{a as j}from"./chunk-P577DWWL.js";import{a as ne}from"./chunk-7MQZPDDZ.js";import{a as L}from"./chunk-ANLPPIMS.js";import{a as re}from"./chunk-TG4C4ENM.js";import{a as Z}from"./chunk-MJLW5KN2.js";import{d as P}from"./chunk-SW24KZ5C.js";import{a as ae}from"./chunk-RWQG25JU.js";import{b as y,c as O,f as o,g as A,i as J,j as b}from"./chunk-3ZHBOILE.js";import{k as N}from"./chunk-MGT5FZJ4.js";import{d as M}from"./chunk-GPBN4ROE.js";import{b as R,e as H,f as D}from"./chunk-5VWVQCNR.js";import{a as l}from"./chunk-ZLLER3HP.js";var E=null,S=new Map,_e={process_eval:V,process_suggestion_eval:W,process_suggestion_discovery:x,process_insights:U,process_launch:F,process_send_message:G,process_conversation:K,process_session_goal:Y,distill_memories:q,memory_ab:Q,process_run_cleanup:X,cron_refine_cases:re,cron_retire_unused_skills:se,cron_reconcile_attribution:L,cron_suggest_instructions:T,cron_refresh_skills:P,cron_refresh_pricing:Z,cron_eval_autorun:j,process_suggestion_adoption:B,cron_auto_accept_suggestions:ee,cron_instruction_training:ne,cron_instruction_evolution:N,process_judge_review:z,harbor_ab:C},ce=500,ue=1e4,pe=Object.keys(_e),ve=new Set([]);function we(e=process.env){return e.DATABASE_URL?pe:pe.filter(s=>!ve.has(s))}l(we,"availableCapabilities");var he=3e4,be=3e4;async function Se(e,s,t=M(),i){let d=[...t.claudeBin?["claude"]:[],...t.codexBin?["codex"]:[],...t.piBin?["pi"]:[]],c=`capabilities=${encodeURIComponent(e.join(","))}&waitMs=${s}&measurementProtocol=1&studyProtocol=1&scopedCollector=2&memorySnapshots=1&boundedAutorun=1&measurementEngines=${encodeURIComponent(d.join(","))}`,u=await H(`/api/work/next?${c}${i?`&launchMode=${i}`:""}`);return u==null||typeof u=="string"?null:u}l(Se,"pollForWork");function Ie(e,s,t){return{logger:{info:l(i=>s(i),"info"),warn:l(i=>s(i),"warn"),error:l(i=>s(i),"error")},job:{attempts:e},leaseSignal:t}}l(Ie,"makeHelpers");async function ye(e,s,t,i){let d=_e[e.kind],c=Date.now();if(o(t,{event:"claim",jobId:e.id,kind:e.kind,attempts:e.attempts}),!d){let n=`unknown kind: ${e.kind}`;o(t,{event:"fail",jobId:e.id,kind:e.kind,terminal:!0,error:n}),await D(`/api/work/${e.id}/fail`,{leaseToken:e.leaseToken,error:n,terminal:!0}).catch(()=>{});return}let u=new AbortController,m=!1,k,g=l(()=>{m||(o(t,{event:"heartbeat",jobId:e.id}),b(`/api/work/${e.id}/heartbeat`,{leaseToken:e.leaseToken},{budgetMs:Math.min(O,Math.max(0,s-1e3)),signal:i,log:t,event:"heartbeat",jobId:e.id}).then(n=>{if(!(n.ok||m)){if(J(n.error)){m=!0,k!==void 0&&clearInterval(k),o(t,{event:"lease_lost",jobId:e.id,kind:e.kind,elapsedMs:Date.now()-c,error:n.error instanceof Error?n.error.message:String(n.error)}),u.abort(new Error(`lease lost for job ${e.id} (${e.kind}); stop work`));return}o(t,{event:"heartbeat_error",jobId:e.id,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}}))},"beat");k=setInterval(g,s),S.set(e.id,g);let a;try{let n=Ie(e.attempts,t,u.signal);a={kind:"complete",result:await d(e.payload,n)??{}}}catch(n){let v=E?.pausesSince(c)??[],_=de(v),w=n instanceof Error?n.message:String(n),r=$(n),p=!r&&le(v);a={kind:"fail",message:_?`${_}; ${w}`:w,pausedDuringJob:!!_,terminal:r,hostSuspended:p}}try{if(m)o(t,{event:"lease_lost_stop",jobId:e.id,kind:e.kind,durationMs:Date.now()-c,handlerOutcome:a.kind,...a.kind==="fail"?{error:a.message}:{}});else if(a.kind==="complete"){o(t,{event:"complete",jobId:e.id,kind:e.kind,durationMs:Date.now()-c});let n=await b(`/api/work/${e.id}/complete`,{leaseToken:e.leaseToken,result:a.result},{budgetMs:y,signal:i,log:t,event:"complete_post",jobId:e.id});n.ok||o(t,{event:"complete_post_error",jobId:e.id,kind:e.kind,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}else{o(t,{event:"fail",jobId:e.id,kind:e.kind,durationMs:Date.now()-c,terminal:a.terminal,error:a.message,...a.pausedDuringJob?{pausedDuringJob:!0}:{},...a.hostSuspended?{hostSuspended:!0}:{}});let n=await b(`/api/work/${e.id}/fail`,{leaseToken:e.leaseToken,error:a.message,terminal:a.terminal,hostSuspended:a.hostSuspended},{budgetMs:y,signal:i,log:t,event:"fail_post",jobId:e.id});n.ok||o(t,{event:"fail_post_error",jobId:e.id,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}}finally{clearInterval(k),S.delete(e.id)}}l(ye,"runOne");var me=new Set(["process_launch","process_eval","process_suggestion_eval","process_suggestion_discovery","process_conversation","process_insights","process_session_goal","distill_memories","memory_ab","cron_suggest_instructions","cron_eval_autorun","cron_instruction_evolution","process_judge_review","harbor_ab"]),ge=new Set(["cron_refresh_skills"]);async function tn(e){let s=e.log??(()=>{}),t=e.pollWaitMs??he,i=e.heartbeatIntervalMs??be,d=e.capabilities??we();o(s,{event:"agent_start",capabilities:d,pollWaitMs:t,heartbeatIntervalMs:i});try{let r=oe();r.removed.length>0&&o(s,{event:"eval_worktrees_swept",removed:r.removed.length})}catch(r){o(s,{event:"eval_worktree_sweep_failed",error:r instanceof Error?r.message:String(r)})}E=ie({signal:e.signal,onPause:l(r=>{o(s,{event:"machine_pause",...r,jobIds:[...S.keys()]});for(let p of S.values())p()},"onPause")});let c=ae(d,e.signal,s,e.version,()=>E);te({signal:e.signal,log:s}).catch(r=>{o(s,{event:"tt_heartbeat_fatal",error:r instanceof Error?r.message:String(r)})});let u=5*6e4,m=0,k=l(()=>{if(!e.readInstalledVersion||!e.version)return!1;let r=Date.now();if(r-m<u)return!1;m=r;let p=e.readInstalledVersion();return!!(p&&p!==e.version)},"versionDrifted"),g=new AbortController,a=g.signal;e.signal.aborted?g.abort():e.signal.addEventListener("abort",()=>g.abort(),{once:!0});let n=l(async(r,p,fe)=>{let f=ce;for(;!a.aborted;){if(k()){o(s,{event:"agent_version_drift",bootVersion:e.version,installedVersion:e.readInstalledVersion?.()??null}),g.abort(),e.onVersionDrift?.();break}let I=null;try{I=await Se([...p],t,void 0,fe)}catch(h){let ke=h instanceof R?h.statusCode:void 0;if(o(s,{event:"poll_error",lane:r,status:ke,error:h instanceof Error?h.message:String(h)}),!await A(f,a))break;f=Math.min(f*2,ue);continue}if(!I){if(o(s,{event:"poll_204",lane:r}),!await A(f,a))break;f=Math.min(f*2,ue);continue}f=ce,await ye(I,i,s,a)}},"runLane"),v=d.filter(r=>me.has(r)),_=d.filter(r=>(r==="process_launch"||!me.has(r))&&!ge.has(r)),w=d.filter(r=>ge.has(r));o(s,{event:"agent_lanes",heavy:v.length,fast:_.length,background:w.length}),await Promise.all([...v.length?[n("heavy",v,"headless")]:[],..._.length?[n("fast",_,"interactive")]:[],...w.length?[n("background",w)]:[]]),await c.shutdown(),o(s,{event:"agent_stop"})}l(tn,"runAgentLoop");export{_e as a,pe as b,we as c,Se as d,tn as e};
+import{c as te}from"./chunk-5U5SHJJA.js";import{b as oe}from"./chunk-IEQDLURA.js";import{d as ie,e as le,f as de}from"./chunk-KINIFUZQ.js";import{a as G}from"./chunk-ZYGBABMZ.js";import{a as Y}from"./chunk-QRKMRG6W.js";import{a as B}from"./chunk-D7L7IMJQ.js";import{a as x}from"./chunk-MRQIYIQE.js";import{a as W}from"./chunk-PMMHMRFS.js";import{a as K}from"./chunk-IYX65N6T.js";import{a as V}from"./chunk-GHUW7P55.js";import{a as U}from"./chunk-RAMQVEJB.js";import{a as z}from"./chunk-BRYURAUN.js";import{b as F}from"./chunk-NLUHMHPM.js";import{a as Q}from"./chunk-DTTR4EUJ.js";import{a as X}from"./chunk-WLLIFRRE.js";import{b as $}from"./chunk-5RBLIVBF.js";import{d as P}from"./chunk-SW24KZ5C.js";import{a as se}from"./chunk-NSO2EBYE.js";import{e as T}from"./chunk-2C7CFCMM.js";import{b as q}from"./chunk-WSUKOU3X.js";import{c as C}from"./chunk-A6XQSO62.js";import{a as ee}from"./chunk-IPNBW7ZW.js";import{a as j}from"./chunk-P577DWWL.js";import{a as ne}from"./chunk-7MQZPDDZ.js";import{a as L}from"./chunk-ANLPPIMS.js";import{a as re}from"./chunk-CXENLSMR.js";import{a as Z}from"./chunk-MJLW5KN2.js";import{a as ae}from"./chunk-RWQG25JU.js";import{b as y,c as O,f as o,g as A,i as J,j as b}from"./chunk-3ZHBOILE.js";import{k as N}from"./chunk-L5ILEWHH.js";import{d as M}from"./chunk-GPBN4ROE.js";import{b as R,e as H,f as D}from"./chunk-5VWVQCNR.js";import{a as l}from"./chunk-ZLLER3HP.js";var E=null,S=new Map,_e={process_eval:V,process_suggestion_eval:W,process_suggestion_discovery:x,process_insights:U,process_launch:F,process_send_message:G,process_conversation:K,process_session_goal:Y,distill_memories:q,memory_ab:Q,process_run_cleanup:X,cron_refine_cases:re,cron_retire_unused_skills:se,cron_reconcile_attribution:L,cron_suggest_instructions:T,cron_refresh_skills:P,cron_refresh_pricing:Z,cron_eval_autorun:j,process_suggestion_adoption:B,cron_auto_accept_suggestions:ee,cron_instruction_training:ne,cron_instruction_evolution:N,process_judge_review:z,harbor_ab:C},ce=500,ue=1e4,pe=Object.keys(_e),ve=new Set([]);function we(e=process.env){return e.DATABASE_URL?pe:pe.filter(s=>!ve.has(s))}l(we,"availableCapabilities");var he=3e4,be=3e4;async function Se(e,s,t=M(),i){let d=[...t.claudeBin?["claude"]:[],...t.codexBin?["codex"]:[],...t.piBin?["pi"]:[]],c=`capabilities=${encodeURIComponent(e.join(","))}&waitMs=${s}&measurementProtocol=1&studyProtocol=1&scopedCollector=2&memorySnapshots=1&boundedAutorun=1&measurementEngines=${encodeURIComponent(d.join(","))}`,u=await H(`/api/work/next?${c}${i?`&launchMode=${i}`:""}`);return u==null||typeof u=="string"?null:u}l(Se,"pollForWork");function Ie(e,s,t){return{logger:{info:l(i=>s(i),"info"),warn:l(i=>s(i),"warn"),error:l(i=>s(i),"error")},job:{attempts:e},leaseSignal:t}}l(Ie,"makeHelpers");async function ye(e,s,t,i){let d=_e[e.kind],c=Date.now();if(o(t,{event:"claim",jobId:e.id,kind:e.kind,attempts:e.attempts}),!d){let n=`unknown kind: ${e.kind}`;o(t,{event:"fail",jobId:e.id,kind:e.kind,terminal:!0,error:n}),await D(`/api/work/${e.id}/fail`,{leaseToken:e.leaseToken,error:n,terminal:!0}).catch(()=>{});return}let u=new AbortController,m=!1,k,g=l(()=>{m||(o(t,{event:"heartbeat",jobId:e.id}),b(`/api/work/${e.id}/heartbeat`,{leaseToken:e.leaseToken},{budgetMs:Math.min(O,Math.max(0,s-1e3)),signal:i,log:t,event:"heartbeat",jobId:e.id}).then(n=>{if(!(n.ok||m)){if(J(n.error)){m=!0,k!==void 0&&clearInterval(k),o(t,{event:"lease_lost",jobId:e.id,kind:e.kind,elapsedMs:Date.now()-c,error:n.error instanceof Error?n.error.message:String(n.error)}),u.abort(new Error(`lease lost for job ${e.id} (${e.kind}); stop work`));return}o(t,{event:"heartbeat_error",jobId:e.id,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}}))},"beat");k=setInterval(g,s),S.set(e.id,g);let a;try{let n=Ie(e.attempts,t,u.signal);a={kind:"complete",result:await d(e.payload,n)??{}}}catch(n){let v=E?.pausesSince(c)??[],_=de(v),w=n instanceof Error?n.message:String(n),r=$(n),p=!r&&le(v);a={kind:"fail",message:_?`${_}; ${w}`:w,pausedDuringJob:!!_,terminal:r,hostSuspended:p}}try{if(m)o(t,{event:"lease_lost_stop",jobId:e.id,kind:e.kind,durationMs:Date.now()-c,handlerOutcome:a.kind,...a.kind==="fail"?{error:a.message}:{}});else if(a.kind==="complete"){o(t,{event:"complete",jobId:e.id,kind:e.kind,durationMs:Date.now()-c});let n=await b(`/api/work/${e.id}/complete`,{leaseToken:e.leaseToken,result:a.result},{budgetMs:y,signal:i,log:t,event:"complete_post",jobId:e.id});n.ok||o(t,{event:"complete_post_error",jobId:e.id,kind:e.kind,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}else{o(t,{event:"fail",jobId:e.id,kind:e.kind,durationMs:Date.now()-c,terminal:a.terminal,error:a.message,...a.pausedDuringJob?{pausedDuringJob:!0}:{},...a.hostSuspended?{hostSuspended:!0}:{}});let n=await b(`/api/work/${e.id}/fail`,{leaseToken:e.leaseToken,error:a.message,terminal:a.terminal,hostSuspended:a.hostSuspended},{budgetMs:y,signal:i,log:t,event:"fail_post",jobId:e.id});n.ok||o(t,{event:"fail_post_error",jobId:e.id,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}}finally{clearInterval(k),S.delete(e.id)}}l(ye,"runOne");var me=new Set(["process_launch","process_eval","process_suggestion_eval","process_suggestion_discovery","process_conversation","process_insights","process_session_goal","distill_memories","memory_ab","cron_suggest_instructions","cron_eval_autorun","cron_instruction_evolution","process_judge_review","harbor_ab"]),ge=new Set(["cron_refresh_skills"]);async function tn(e){let s=e.log??(()=>{}),t=e.pollWaitMs??he,i=e.heartbeatIntervalMs??be,d=e.capabilities??we();o(s,{event:"agent_start",capabilities:d,pollWaitMs:t,heartbeatIntervalMs:i});try{let r=oe();r.removed.length>0&&o(s,{event:"eval_worktrees_swept",removed:r.removed.length})}catch(r){o(s,{event:"eval_worktree_sweep_failed",error:r instanceof Error?r.message:String(r)})}E=ie({signal:e.signal,onPause:l(r=>{o(s,{event:"machine_pause",...r,jobIds:[...S.keys()]});for(let p of S.values())p()},"onPause")});let c=ae(d,e.signal,s,e.version,()=>E);te({signal:e.signal,log:s}).catch(r=>{o(s,{event:"tt_heartbeat_fatal",error:r instanceof Error?r.message:String(r)})});let u=5*6e4,m=0,k=l(()=>{if(!e.readInstalledVersion||!e.version)return!1;let r=Date.now();if(r-m<u)return!1;m=r;let p=e.readInstalledVersion();return!!(p&&p!==e.version)},"versionDrifted"),g=new AbortController,a=g.signal;e.signal.aborted?g.abort():e.signal.addEventListener("abort",()=>g.abort(),{once:!0});let n=l(async(r,p,fe)=>{let f=ce;for(;!a.aborted;){if(k()){o(s,{event:"agent_version_drift",bootVersion:e.version,installedVersion:e.readInstalledVersion?.()??null}),g.abort(),e.onVersionDrift?.();break}let I=null;try{I=await Se([...p],t,void 0,fe)}catch(h){let ke=h instanceof R?h.statusCode:void 0;if(o(s,{event:"poll_error",lane:r,status:ke,error:h instanceof Error?h.message:String(h)}),!await A(f,a))break;f=Math.min(f*2,ue);continue}if(!I){if(o(s,{event:"poll_204",lane:r}),!await A(f,a))break;f=Math.min(f*2,ue);continue}f=ce,await ye(I,i,s,a)}},"runLane"),v=d.filter(r=>me.has(r)),_=d.filter(r=>(r==="process_launch"||!me.has(r))&&!ge.has(r)),w=d.filter(r=>ge.has(r));o(s,{event:"agent_lanes",heavy:v.length,fast:_.length,background:w.length}),await Promise.all([...v.length?[n("heavy",v,"headless")]:[],..._.length?[n("fast",_,"interactive")]:[],...w.length?[n("background",w)]:[]]),await c.shutdown(),o(s,{event:"agent_stop"})}l(tn,"runAgentLoop");export{_e as a,pe as b,we as c,Se as d,tn as e};

package/dist/{chunk-GBSIOACS.js → chunk-O7ZQEKND.js}

@@ -1,35 +1,35 @@
-import{a as f,b as M}from"./chunk-BDWQ6J4V.js";import{e as _,f as S}from"./chunk-MHLHLJ5T.js";import{a as o}from"./chunk-ZLLER3HP.js";import*as m from"node:fs/promises";import*as x from"node:path";import{eq as j}from"drizzle-orm";import{randomBytes as Y}from"node:crypto";var A="#0d9488",h="#1a1a1a",$="#666666",C="#f5f5f7",N="#ffffff",D="#e5e5e9",I=["Added","Changed","Fixed","Removed","Deprecated","Security"];function U(e){let t=["January","February","March","April","May","June","July","August","September","October","November","December"],[n,r,a]=e.split("-").map(d=>parseInt(d,10));return!n||!r||!a||r<1||r>12?e:`${t[r-1]} ${a}, ${n}`}o(U,"formatIsoDate");function l(e){return e.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;").replace(/"/g,"&quot;").replace(/'/g,"&#39;")}o(l,"escapeHtml");function z(e){let t=[];for(let n of I){let r=e[n];r&&r.length>0&&t.push({name:n,items:r})}return t}o(z,"orderedSections");function v(e){let{version:t,date:n,sections:r,unsubscribeUrl:a,productUrl:d,recipientName:u}=e,c=U(n),s=`What's new in RuleMetric \u2014 ${c}`,y=u?`Hi ${u},`:"Hi there,",i=z(r),p=new Date().getUTCFullYear(),b=i.map(({name:R,items:E})=>{let w=E.map(O=>`              <li style="margin:0 0 6px 0;font-size:15px;line-height:1.5;color:${h};">${l(O)}</li>`).join(`
-`);return`            <h3 style="margin:24px 0 8px 0;font-size:16px;line-height:1.3;font-weight:600;color:${h};">${R}</h3>
+import{a as f,b as k}from"./chunk-BDWQ6J4V.js";import{e as _,f as S}from"./chunk-MHLHLJ5T.js";import{a as o}from"./chunk-ZLLER3HP.js";import*as m from"node:fs/promises";import*as $ from"node:path";import{eq as j}from"drizzle-orm";import{randomBytes as Y}from"node:crypto";var C="#0d9488",h="#1a1a1a",x="#666666",A="#f5f5f7",N="#ffffff",D="#e5e5e9",I=["Added","Changed","Fixed","Removed","Deprecated","Security"];function U(e){let t=["January","February","March","April","May","June","July","August","September","October","November","December"],[n,r,a]=e.split("-").map(d=>parseInt(d,10));return!n||!r||!a||r<1||r>12?e:`${t[r-1]} ${a}, ${n}`}o(U,"formatIsoDate");function c(e){return e.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;").replace(/"/g,"&quot;").replace(/'/g,"&#39;")}o(c,"escapeHtml");function z(e){let t=[];for(let n of I){let r=e[n];r&&r.length>0&&t.push({name:n,items:r})}return t}o(z,"orderedSections");function v(e){let{version:t,date:n,sections:r,unsubscribeUrl:a,productUrl:d,recipientName:u}=e,l=U(n),s=`What's new in RuleMetric \u2014 ${l}`,y=u?`Hi ${u},`:"Hi there,",i=z(r),p=new Date().getUTCFullYear(),b=i.map(({name:w,items:R})=>{let E=R.map(O=>`              <li style="margin:0 0 6px 0;font-size:15px;line-height:1.5;color:${h};">${c(O)}</li>`).join(`
+`);return`            <h3 style="margin:24px 0 8px 0;font-size:16px;line-height:1.3;font-weight:600;color:${h};">${w}</h3>
             <ul style="margin:0 0 8px 0;padding-left:20px;">
-${w}
+${E}
             </ul>`}).join(`
-`),k=`<!doctype html>
+`),M=`<!doctype html>
 <html lang="en">
   <head>
     <meta charset="utf-8" />
     <meta name="viewport" content="width=device-width, initial-scale=1" />
-    <title>${l(s)}</title>
+    <title>${c(s)}</title>
   </head>
-  <body style="margin:0;padding:0;background:${C};font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,Helvetica,Arial,sans-serif;color:${h};">
-    <table role="presentation" width="100%" cellspacing="0" cellpadding="0" border="0" style="background:${C};padding:32px 16px;">
+  <body style="margin:0;padding:0;background:${A};font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,Helvetica,Arial,sans-serif;color:${h};">
+    <table role="presentation" width="100%" cellspacing="0" cellpadding="0" border="0" style="background:${A};padding:32px 16px;">
       <tr>
         <td align="center">
           <table role="presentation" width="560" cellspacing="0" cellpadding="0" border="0" style="max-width:560px;background:${N};border:1px solid ${D};border-radius:8px;overflow:hidden;">
             <tr>
               <td style="padding:24px 32px 8px 32px;">
-                <a href="${l(d)}" style="text-decoration:none;color:${A};font-size:14px;font-weight:600;">RuleMetric</a>
+                <a href="${c(d)}" style="text-decoration:none;color:${C};font-size:14px;font-weight:600;">RuleMetric</a>
               </td>
             </tr>
             <tr>
               <td style="padding:8px 32px 8px 32px;">
                 <h1 style="margin:0 0 8px 0;font-size:22px;line-height:1.3;font-weight:600;color:${h};">
-                  What's new \u2014 ${l(c)}
+                  What's new \u2014 ${c(l)}
                 </h1>
-                <p style="margin:0 0 8px 0;font-size:13px;line-height:1.5;color:${$};">
-                  Release ${l(t)}
+                <p style="margin:0 0 8px 0;font-size:13px;line-height:1.5;color:${x};">
+                  Release ${c(t)}
                 </p>
                 <p style="margin:16px 0 0 0;font-size:15px;line-height:1.5;color:${h};">
-                  ${l(y)}
+                  ${c(y)}
                 </p>
                 <p style="margin:8px 0 0 0;font-size:15px;line-height:1.5;color:${h};">
                   Here's what shipped in the latest RuleMetric release.
@@ -43,16 +43,16 @@ ${b}
             </tr>
             <tr>
               <td style="padding:16px 32px 24px 32px;border-top:1px solid ${D};">
-                <p style="margin:0 0 8px 0;font-size:12px;line-height:1.5;color:${$};">
+                <p style="margin:0 0 8px 0;font-size:12px;line-height:1.5;color:${x};">
                   You're receiving this because you signed up for RuleMetric product updates. If you'd rather not hear about new releases, you can unsubscribe at any time.
                 </p>
                 <p style="margin:0 0 4px 0;font-size:12px;line-height:1.5;">
-                  <a href="${l(a)}" style="color:${A};text-decoration:underline;">Unsubscribe</a>
+                  <a href="${c(a)}" style="color:${C};text-decoration:underline;">Unsubscribe</a>
                 </p>
-                <p style="margin:0 0 12px 0;font-size:12px;line-height:1.5;color:${$};word-break:break-all;">
-                  ${l(a)}
+                <p style="margin:0 0 12px 0;font-size:12px;line-height:1.5;color:${x};word-break:break-all;">
+                  ${c(a)}
                 </p>
-                <p style="margin:0;font-size:11px;line-height:1.5;color:${$};">
+                <p style="margin:0;font-size:11px;line-height:1.5;color:${x};">
                   RuleMetric \u2022 ${p}
                 </p>
               </td>
@@ -62,5 +62,5 @@ ${b}
       </tr>
     </table>
   </body>
-</html>`,g=[`What's new in RuleMetric \u2014 ${c}`,`Release ${t}`,"",y,"","Here's what shipped in the latest RuleMetric release.",""];for(let{name:R,items:E}of i){g.push(`${R}:`);for(let w of E)g.push(`  - ${w}`);g.push("")}g.push("--","You're receiving this because you signed up for RuleMetric product updates.",`Unsubscribe: ${a}`,"",`RuleMetric \u2022 ${p}`);let T=g.join(`
-`);return{subject:s,html:k,text:T}}o(v,"renderChangelogAnnouncement");async function P(e){let t=e;for(let n=0;n<50;n++){try{return await m.stat(x.join(t,"pnpm-workspace.yaml")),t}catch{}let r=x.dirname(t);if(r===t)return null;t=r}return null}o(P,"findRepoRoot");async function L(e){if(e)try{return await m.readFile(e,"utf-8")}catch{return null}let t=await P(process.cwd())??process.cwd(),n=x.join(t,"CHANGELOG.md");try{return await m.readFile(n,"utf-8")}catch{return null}}o(L,"readChangelog");var B="https://rulemetric.com";function F(e){let t=e.date??new Date().toISOString().slice(0,10),n=v({version:e.version,date:t,sections:e.sections,unsubscribeUrl:"{{unsubscribe_url}}",recipientName:"{{recipient_name}}",productUrl:B});return{version:e.version,releaseDate:t,subject:n.subject,htmlBody:n.html,textBody:n.text,sections:e.sections}}o(F,"renderRow");async function Re(e={}){let t=e.db??M(),n=e.logger??{info:o(s=>console.log(s),"info"),warn:o(s=>console.warn(s),"warn")},r=await L(e.changelogPath);if(r===null)return n.info("[cron_check_changelog] CHANGELOG.md not found; nothing to do"),{createdCount:0,skippedCount:0};let a=_(r),d=S(a);if(d.length===0)return n.info("[cron_check_changelog] No announceable entries; nothing to do"),{createdCount:0,skippedCount:0};let u=0,c=0;for(let s of d){if((await t.select({id:f.id}).from(f).where(j(f.version,s.version)).limit(1)).length>0){c++;continue}let i=F(s);try{await t.insert(f).values({version:i.version,releaseDate:i.releaseDate,subject:i.subject,htmlBody:i.htmlBody,textBody:i.textBody,sections:i.sections,status:"draft"}),u++,n.info(`[cron_check_changelog] Created draft announcement for ${i.version}`)}catch(p){let b=p instanceof Error?p.message:String(p);if(/duplicate key|unique/i.test(b)){c++;continue}throw p}}return n.info(`[cron_check_changelog] Done: created ${u}, skipped ${c}`),{createdCount:u,skippedCount:c}}o(Re,"processChangelog");export{Re as a};
+</html>`,g=[`What's new in RuleMetric \u2014 ${l}`,`Release ${t}`,"",y,"","Here's what shipped in the latest RuleMetric release.",""];for(let{name:w,items:R}of i){g.push(`${w}:`);for(let E of R)g.push(`  - ${E}`);g.push("")}g.push("--","You're receiving this because you signed up for RuleMetric product updates.",`Unsubscribe: ${a}`,"",`RuleMetric \u2022 ${p}`);let T=g.join(`
+`);return{subject:s,html:M,text:T}}o(v,"renderChangelogAnnouncement");async function L(e){let t=e;for(let n=0;n<50;n++){try{return await m.stat($.join(t,"pnpm-workspace.yaml")),t}catch{}let r=$.dirname(t);if(r===t)return null;t=r}return null}o(L,"findRepoRoot");async function P(e){if(e)try{return await m.readFile(e,"utf-8")}catch{return null}let t=await L(process.cwd())??process.cwd(),n=$.join(t,"CHANGELOG.md");try{return await m.readFile(n,"utf-8")}catch{return null}}o(P,"readChangelog");var B="https://rulemetric.com";function F(e){let t=e.date??new Date().toISOString().slice(0,10),n=v({version:e.version,date:t,sections:e.sections,unsubscribeUrl:"{{unsubscribe_url}}",recipientName:"{{recipient_name}}",productUrl:B});return{version:e.version,releaseDate:t,subject:n.subject,htmlBody:n.html,textBody:n.text,sections:e.sections}}o(F,"renderRow");async function we(e={}){let t=e.db??k(),n=e.logger??{info:o(s=>console.log(s),"info"),warn:o(s=>console.warn(s),"warn")},r=await P(e.changelogPath);if(r===null)return n.info("[cron_check_changelog] CHANGELOG.md not found; nothing to do"),{createdCount:0,skippedCount:0};let a=_(r),d=S(a);if(d.length===0)return n.info("[cron_check_changelog] No announceable entries; nothing to do"),{createdCount:0,skippedCount:0};let u=0,l=0;for(let s of d){if((await t.select({id:f.id}).from(f).where(j(f.version,s.version)).limit(1)).length>0){l++;continue}let i=F(s);try{await t.insert(f).values({version:i.version,releaseDate:i.releaseDate,subject:i.subject,htmlBody:i.htmlBody,textBody:i.textBody,sections:i.sections,status:"draft"}),u++,n.info(`[cron_check_changelog] Created draft announcement for ${i.version}`)}catch(p){let b=p instanceof Error?p.message:String(p);if(/duplicate key|unique/i.test(b)){l++;continue}throw p}}return n.info(`[cron_check_changelog] Done: created ${u}, skipped ${l}`),{createdCount:u,skippedCount:l}}o(we,"processChangelog");export{we as a};

package/dist/{chunk-MBP5OJKE.js → chunk-RAMQVEJB.js}

@@ -1,4 +1,4 @@
-import{a as N}from"./chunk-5RBLIVBF.js";import{b as Z,d as X}from"./chunk-K65DDLTZ.js";import{a as se}from"./chunk-LELSVLNK.js";import{c as ee}from"./chunk-A2JFQUDW.js";import{a as K,b as Y}from"./chunk-LZPE6424.js";import{a as A,d as F}from"./chunk-3UKZJAFS.js";import{a as O,b as M,c as Q,d as G}from"./chunk-77FH6UZP.js";import{b as B,e as V,g as R}from"./chunk-GWQ24FQP.js";import{e as $,f as U,h as f}from"./chunk-5VWVQCNR.js";import{a as q,b as W}from"./chunk-MHLHLJ5T.js";import{a as _}from"./chunk-ZLLER3HP.js";var te=_((c,{limit:i,sort:n,eventTypes:e})=>$(`/api/sessions/${c}/events?limit=${i}&sort=${n}&eventType=${encodeURIComponent(e.join(","))}&hasContent=true`),"fetchTranscriptTurns");async function _e(c,i){let{mirrorId:n,projectPath:e,projectId:r,config:u}=c,s=_(t=>i.logger.info(t),"log");if(u?.sessionId)return re({mirrorId:n,sessionId:u.sessionId},s);s(`Processing insights job ${n}...`),await f(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:t}=await U("/api/sessions/repair-analyzed-at",{});t>0&&s(`  Repaired analyzed_at on ${t} orphan session(s).`)}catch(t){s(`  Warning: analyzed_at repair failed: ${t instanceof Error?t.message:String(t)}`)}let y=20,w=200,x=15*6e4,J=3,v=Date.now()+x,C=V()==="none"?"no-llm-backend":null,g={attempted:0,generated:0,failed:0,ungradeable:0,stopReason:C??"queue-exhausted"};if(C&&i.logger.warn?.("[process-insights] no LLM backend (no claude/codex CLI and no ANTHROPIC_API_KEY) \u2014 facet grading SKIPPED for this run; the summary below is computed from already-graded sessions only"),!C){let t=0,a=0;try{for(;t<w&&Date.now()<v;){let d=await $(`/api/sessions?needsFacets=true&graderVersion=${2}&limit=${y}&offset=${a}`);if(d.length===0)break;s(`  Generating facets for ${d.length} un-analyzed sessions (drained ${t} so far)...`);for(let I=0;I<d.length&&!(t>=w||Date.now()>=v);I+=J){let j=d.slice(I,I+J);await Promise.all(j.map(async p=>{if(!(t>=w||Date.now()>=v)){g.attempted++;try{let l=await F(oe=>te(p.id,oe));if(l.turnCount<2){g.ungradeable++,t++,a++;return}let D=await R(G,l.text,4096);D&&typeof D.outcome=="string"?(await f(`/api/sessions/${p.id}/enrich`,{analysis:{source:"generated_facets",...D,graderVersion:2},...p.eventCount!=null?{analyzedEventCount:p.eventCount}:{},analyzedTurnCount:l.turnCount}),s(`    Generated facets for session ${p.id} (${p.tool}, ${l.turnCount} turns)`),g.generated++):(a++,g.failed++),t++}catch(l){g.failed++,s(`    Failed to generate facets for session ${p.id}: ${l instanceof Error?l.message:String(l)}`),t++,a++}}}))}if(d.length<y)break}}catch(d){g.stopReason="read-failed",s(`  Warning: facet generation failed after ${t} sessions: ${d instanceof Error?d.message:String(d)}`)}if(g.stopReason!=="read-failed"&&(g.stopReason=Date.now()>=v?"time-budget":t>=w?"hard-cap":"queue-exhausted"),t>0||g.stopReason==="read-failed"){let d=g.stopReason==="read-failed"?"database read failure":Date.now()>=v?`time budget (${Math.round(x/6e4)}m)`:t>=w?`hard cap (${w})`:"end of queue scan";s(`  Facet drain stopped on ${d} after ${t} session(s); the rest resume next run.`)}}let b=await ie(e,r);if(!b)throw new N("No session data found \u2014 ensure sessions are captured and enriched.");s(`  Found ${b.stats.sessions} sessions, generating recommendations...`);let m=await Y({projectId:r,projectPath:e},$);b.summary+=`
+import{a as N}from"./chunk-5RBLIVBF.js";import{a as O,b as M,c as Q,d as G}from"./chunk-77FH6UZP.js";import{b as Z,d as X}from"./chunk-K65DDLTZ.js";import{a as se}from"./chunk-LELSVLNK.js";import{c as ee}from"./chunk-A2JFQUDW.js";import{a as K,b as Y}from"./chunk-LZPE6424.js";import{a as A,d as F}from"./chunk-3UKZJAFS.js";import{b as B,e as V,g as R}from"./chunk-GWQ24FQP.js";import{e as $,f as U,h as f}from"./chunk-5VWVQCNR.js";import{a as q,b as W}from"./chunk-MHLHLJ5T.js";import{a as _}from"./chunk-ZLLER3HP.js";var te=_((c,{limit:i,sort:n,eventTypes:e})=>$(`/api/sessions/${c}/events?limit=${i}&sort=${n}&eventType=${encodeURIComponent(e.join(","))}&hasContent=true`),"fetchTranscriptTurns");async function _e(c,i){let{mirrorId:n,projectPath:e,projectId:r,config:u}=c,s=_(t=>i.logger.info(t),"log");if(u?.sessionId)return re({mirrorId:n,sessionId:u.sessionId},s);s(`Processing insights job ${n}...`),await f(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:t}=await U("/api/sessions/repair-analyzed-at",{});t>0&&s(`  Repaired analyzed_at on ${t} orphan session(s).`)}catch(t){s(`  Warning: analyzed_at repair failed: ${t instanceof Error?t.message:String(t)}`)}let y=20,w=200,x=15*6e4,J=3,v=Date.now()+x,C=V()==="none"?"no-llm-backend":null,g={attempted:0,generated:0,failed:0,ungradeable:0,stopReason:C??"queue-exhausted"};if(C&&i.logger.warn?.("[process-insights] no LLM backend (no claude/codex CLI and no ANTHROPIC_API_KEY) \u2014 facet grading SKIPPED for this run; the summary below is computed from already-graded sessions only"),!C){let t=0,a=0;try{for(;t<w&&Date.now()<v;){let d=await $(`/api/sessions?needsFacets=true&graderVersion=${2}&limit=${y}&offset=${a}`);if(d.length===0)break;s(`  Generating facets for ${d.length} un-analyzed sessions (drained ${t} so far)...`);for(let I=0;I<d.length&&!(t>=w||Date.now()>=v);I+=J){let j=d.slice(I,I+J);await Promise.all(j.map(async p=>{if(!(t>=w||Date.now()>=v)){g.attempted++;try{let l=await F(oe=>te(p.id,oe));if(l.turnCount<2){g.ungradeable++,t++,a++;return}let D=await R(G,l.text,4096);D&&typeof D.outcome=="string"?(await f(`/api/sessions/${p.id}/enrich`,{analysis:{source:"generated_facets",...D,graderVersion:2},...p.eventCount!=null?{analyzedEventCount:p.eventCount}:{},analyzedTurnCount:l.turnCount}),s(`    Generated facets for session ${p.id} (${p.tool}, ${l.turnCount} turns)`),g.generated++):(a++,g.failed++),t++}catch(l){g.failed++,s(`    Failed to generate facets for session ${p.id}: ${l instanceof Error?l.message:String(l)}`),t++,a++}}}))}if(d.length<y)break}}catch(d){g.stopReason="read-failed",s(`  Warning: facet generation failed after ${t} sessions: ${d instanceof Error?d.message:String(d)}`)}if(g.stopReason!=="read-failed"&&(g.stopReason=Date.now()>=v?"time-budget":t>=w?"hard-cap":"queue-exhausted"),t>0||g.stopReason==="read-failed"){let d=g.stopReason==="read-failed"?"database read failure":Date.now()>=v?`time budget (${Math.round(x/6e4)}m)`:t>=w?`hard cap (${w})`:"end of queue scan";s(`  Facet drain stopped on ${d} after ${t} session(s); the rest resume next run.`)}}let b=await ie(e,r);if(!b)throw new N("No session data found \u2014 ensure sessions are captured and enriched.");s(`  Found ${b.stats.sessions} sessions, generating recommendations...`);let m=await Y({projectId:r,projectPath:e},$);b.summary+=`
 
 ${m.text}`,s(`  Session evidence: ${m.sessionIds.length} usable, ${m.failedSessionIds.length} unavailable`);let P=await ae(e,s);P&&s(`  Generator memory: ${P.split(`
 `).length} lines of prior-advice context`);let o={},E=[],h=M.length+1;for(let t=0;t<M.length;t++){let a=M[t];s(`  [${t+1}/${h}] Generating ${a.name}...`);try{await f(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:a.name,current:t+1,total:h,partial:{...o}}})}catch(l){s(`  Warning: progress update failed for ${a.name}: ${l instanceof Error?l.message:String(l)}`)}s(`  Running LLM prompt for ${a.name}...`);let d=a.usesGeneratorMemory&&P?`${b.summary}

package/dist/chunk-YALBJNHO.js

@@ -0,0 +1 @@
+import{e as y,f as o,h as f}from"./chunk-NJ5UED3H.js";import{a as n}from"./chunk-ZLLER3HP.js";import{Box as p,Text as r}from"ink";import{jsx as x,jsxs as l}from"react/jsx-runtime";function T({checks:t,width:a}){let m=t?.filter(e=>e.severity!=="ok")??[];return!t||m.length===0?null:l(p,{flexDirection:"column",borderStyle:"round",borderColor:o,paddingX:1,width:a,children:[x(r,{bold:!0,color:o,children:"Telemetry"}),m.map(e=>l(p,{flexDirection:"column",children:[l(r,{color:e.severity==="fail"?f:o,children:[e.label,": ",e.message]}),e.hints.map(i=>x(r,{color:y,children:i},i))]},e.label))]})}n(T,"TelemetryPanel");export{T as a};

package/dist/{chunk-EQ2O6DSN.js → chunk-YZQZD2E6.js}

@@ -1 +1 @@
-import{b as m}from"./chunk-BTKVLLV2.js";import{a as g}from"./chunk-ZLLER3HP.js";var p=5;function b(t){let r=g((e,f)=>({retire:[],survivingCases:0,replacementsRequired:!1,refused:e,summary:f}),"none");if(!t||t.cases.length===0&&t.unpairedCases.length===0)return r("no-discrimination-data","No case-discrimination data yet \u2014 the target needs runs in BOTH cohorts before any case can be judged dead.");let c=[...t.cases.filter(e=>!e.discriminating).map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate,withoutPassRate:e.withoutPassRate,reason:e.note}})),...t.unpairedCases.map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate??0,withoutPassRate:e.withoutPassRate,reason:"runs in only one cohort (no with-vs-without comparison possible)"}}))];if(c.length===0)return r("all-discriminating","Every case discriminates \u2014 nothing to retire.");let s=new Map(t.cases.map(e=>[e.evalId,e.separation])),o=[...c].sort((e,f)=>(s.get(e.evalId)??-1)-(s.get(f.evalId)??-1)).slice(0,p),u=c.length-o.length,d=t.total+t.unpaired,l=d-o.length,i=l<m,n=`${o.length} of ${d} live case(s) carry no treatment signal`+(u>0?` (${u} more over the ${p}-per-run cap)`:"")+(i?`; retiring them leaves ${l}, below the ${m}-case evolution floor, so replacements must be generated first`:"");return{retire:o,survivingCases:l,replacementsRequired:i,refused:null,summary:n}}g(b,"planCaseRefinement");async function v(t,r){let h=r.log??(()=>{}),c=g((n,e)=>({retired:0,generated:0,skipped:n,summary:e,retiredNames:[],generatedNames:[]}),"empty");if(t.refused)return c(t.refused,t.summary);let s=await r.generate(t.retire.map(n=>n.feedback)),a=[];for(let n of s??[]){let e=await r.createCase(n);a.push(e.name),h(`  + ${e.name} (${e.id})`)}let o=t.survivingCases+t.retire.length+a.length,u=Math.max(0,o-m),d=t.retire.slice(0,Math.min(t.retire.length,u)),l=t.retire.length-d.length;if(d.length===0)return{retired:0,generated:a.length,skipped:s===null?"no-llm-backend":"no-replacements-generated",summary:`${t.summary}. `+(s===null?"No local LLM backend available, so nothing was retired":`The generator produced ${a.length} usable replacement(s), not enough to retire anything`)+` \u2014 dropping below the ${m}-case floor would silently disenrol the target from the nightly loop.`,retiredNames:[],generatedNames:a};let i=[];for(let n of d)await r.retireCase(n.evalId,n.feedback.reason),i.push(n.name),h(`  retired ${n.name}`);return{retired:i.length,generated:a.length,skipped:null,summary:`Retired ${i.length} non-discriminating case(s), added ${a.length} replacement(s)`+(l>0?`; held ${l} back to stay at the ${m}-case floor (they remain condemned and will be retired once replacements exist)`:"")+". The replacements have no runs yet \u2014 the next autorun batch builds their depth.",retiredNames:i,generatedNames:a}}g(v,"executeCaseRefinement");export{p as a,b,v as c};
+import{b as m}from"./chunk-FLJTAD2R.js";import{a as g}from"./chunk-ZLLER3HP.js";var p=5;function b(t){let r=g((e,f)=>({retire:[],survivingCases:0,replacementsRequired:!1,refused:e,summary:f}),"none");if(!t||t.cases.length===0&&t.unpairedCases.length===0)return r("no-discrimination-data","No case-discrimination data yet \u2014 the target needs runs in BOTH cohorts before any case can be judged dead.");let c=[...t.cases.filter(e=>!e.discriminating).map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate,withoutPassRate:e.withoutPassRate,reason:e.note}})),...t.unpairedCases.map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate??0,withoutPassRate:e.withoutPassRate,reason:"runs in only one cohort (no with-vs-without comparison possible)"}}))];if(c.length===0)return r("all-discriminating","Every case discriminates \u2014 nothing to retire.");let s=new Map(t.cases.map(e=>[e.evalId,e.separation])),o=[...c].sort((e,f)=>(s.get(e.evalId)??-1)-(s.get(f.evalId)??-1)).slice(0,p),u=c.length-o.length,d=t.total+t.unpaired,l=d-o.length,i=l<m,n=`${o.length} of ${d} live case(s) carry no treatment signal`+(u>0?` (${u} more over the ${p}-per-run cap)`:"")+(i?`; retiring them leaves ${l}, below the ${m}-case evolution floor, so replacements must be generated first`:"");return{retire:o,survivingCases:l,replacementsRequired:i,refused:null,summary:n}}g(b,"planCaseRefinement");async function v(t,r){let h=r.log??(()=>{}),c=g((n,e)=>({retired:0,generated:0,skipped:n,summary:e,retiredNames:[],generatedNames:[]}),"empty");if(t.refused)return c(t.refused,t.summary);let s=await r.generate(t.retire.map(n=>n.feedback)),a=[];for(let n of s??[]){let e=await r.createCase(n);a.push(e.name),h(`  + ${e.name} (${e.id})`)}let o=t.survivingCases+t.retire.length+a.length,u=Math.max(0,o-m),d=t.retire.slice(0,Math.min(t.retire.length,u)),l=t.retire.length-d.length;if(d.length===0)return{retired:0,generated:a.length,skipped:s===null?"no-llm-backend":"no-replacements-generated",summary:`${t.summary}. `+(s===null?"No local LLM backend available, so nothing was retired":`The generator produced ${a.length} usable replacement(s), not enough to retire anything`)+` \u2014 dropping below the ${m}-case floor would silently disenrol the target from the nightly loop.`,retiredNames:[],generatedNames:a};let i=[];for(let n of d)await r.retireCase(n.evalId,n.feedback.reason),i.push(n.name),h(`  retired ${n.name}`);return{retired:i.length,generated:a.length,skipped:null,summary:`Retired ${i.length} non-discriminating case(s), added ${a.length} replacement(s)`+(l>0?`; held ${l} back to stay at the ${m}-case floor (they remain condemned and will be retired once replacements exist)`:"")+". The replacements have no runs yet \u2014 the next autorun batch builds their depth.",retiredNames:i,generatedNames:a}}g(v,"executeCaseRefinement");export{p as a,b,v as c};

package/dist/chunk-ZFEDOPRU.js

@@ -0,0 +1,7 @@
+import{c as a,d as y}from"./chunk-5VWVQCNR.js";import{l as m}from"./chunk-PZNWH4VQ.js";import{a as s}from"./chunk-ZLLER3HP.js";import{createHash as h}from"node:crypto";import{mkdirSync as O,readFileSync as x,writeFileSync as U,renameSync as $}from"node:fs";import{join as S}from"node:path";function _(n){if(process.env.RULEMETRIC_HOOK_TOOL&&process.env.RULEMETRIC_HOOK_TOOL!=="claude_code")return null;try{let e=JSON.parse(n);return e.hook_event_name==="SessionStart"&&typeof e.session_id=="string"&&typeof e.cwd=="string"?{cwd:e.cwd}:null}catch{return null}}s(_,"claudeSessionStart");function j(n,e){return JSON.stringify({...e?{systemMessage:e}:{},hookSpecificOutput:{hookEventName:"SessionStart",additionalContext:n}})+`
+`}s(j,"sessionStartOutput");function w(n,e,o){let p=o.toISOString().slice(0,10);return[...n.blockers.map(t=>({item:t,key:`${t.id}:${p}`})).filter(({key:t})=>!e[t]).slice(0,1),...n.events.map(t=>({item:t,key:t.id}))].filter(({key:t})=>!e[t]).slice(0,3)}s(w,"selectSessionUpdates");async function A(n){try{let e=await y(`/api/experiment-updates?projectPath=${encodeURIComponent(n)}`,{timeoutMs:2500});if(!e.userId||!Array.isArray(e.events)||!Array.isArray(e.blockers))return null;let o=new Date,p=h("sha256").update(`${a()}
+${e.userId}
+${n}`).digest("hex"),d=S(m(),"experiment-notices"),t=S(d,`${p}.json`),i={};try{i=JSON.parse(x(t,"utf8"))}catch{}(!i||typeof i!="object"||Array.isArray(i))&&(i={});let l=w(e,i,o);if(!l.length)return null;let g=s(r=>r.replace(/[\x00-\x1f\x7f]/g," ").slice(0,650),"text"),f=l.map(({item:r})=>{let c=new URL(r.url,a());if(c.origin!==new URL(a()).origin)throw new Error("Unexpected evidence origin");return`\u2022 ${g(r.title)}: ${g(r.body)}
+  ${c.href}`});return e.total-l.filter(({key:r})=>!r.startsWith("waiting:")).length>0&&f.push(`More experiment history: ${a()}/loop`),{message:`RuleMetric \u2014 experiment updates
+${f.join(`
+`)}`,emitted:s(()=>{try{O(d,{recursive:!0,mode:448});let r=Object.fromEntries(Object.entries(i).filter(([,u])=>Date.parse(u)>o.getTime()-14*864e5));for(let{key:u}of l)r[u]=o.toISOString();let c=`${t}.${process.pid}.tmp`;U(c,JSON.stringify(r),{mode:384}),$(c,t)}catch{}},"emitted")}}catch{return null}}s(A,"sessionExperimentNotice");export{_ as a,j as b,w as c,A as d};

package/dist/commands/dashboard.js

@@ -1 +1 @@
-import{a as o}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as i}from"../chunk-ZLLER3HP.js";import{Flags as r}from"@oclif/core";var a=class s extends o{static{i(this,"Dashboard")}static description="Interactive TUI dashboard showing session analytics, outcomes, friction, and insights";static examples=["<%= config.bin %> dashboard","<%= config.bin %> dashboard --period 30days","<%= config.bin %> dashboard --launcher"];static flags={period:r.string({char:"p",default:"week",options:["today","week","30days","month","all"],description:"Initial time period"}),launcher:r.boolean({char:"l",default:!1,description:"Open the launcher overlay immediately"})};async run(){let{flags:e}=await this.parse(s),{renderDashboard:t}=await import("../dashboard/Dashboard.js");await t(e.period,{openLauncher:e.launcher})}};export{a as default};
+import{a as o}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as i}from"../chunk-ZLLER3HP.js";import{Flags as r}from"@oclif/core";var a=class s extends o{static{i(this,"Dashboard")}static description="Interactive TUI dashboard showing session analytics, outcomes, friction, and insights";static examples=["<%= config.bin %> dashboard","<%= config.bin %> dashboard --period 30days"];static flags={period:r.string({char:"p",default:"week",options:["today","week","30days","month","all"],description:"Initial time period"})};async run(){let{flags:e}=await this.parse(s),{renderDashboard:t}=await import("../dashboard/Dashboard.js");await t(e.period)}};export{a as default};

package/dist/commands/diagnostics.js

@@ -1 +1 @@
-import{c as q,d as K}from"../chunk-TYFEWLYZ.js";import{b as L,c as W}from"../chunk-V5U3E4BZ.js";import{a as J}from"../chunk-2TJ7Q2OU.js";import{e as O,f as _,g as M}from"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import{d as oe,e as ne}from"../chunk-MGT5FZJ4.js";import"../chunk-M7QSC64D.js";import"../chunk-DZRBB34J.js";import"../chunk-KR32Y5Q5.js";import"../chunk-RITL4NU3.js";import{i as z,j as Q,k as Z,m as ee,n as te}from"../chunk-GYKYJJNF.js";import"../chunk-G7V4ZZZ3.js";import"../chunk-AHLVY52I.js";import"../chunk-PPQFN62M.js";import"../chunk-V6L25LFD.js";import"../chunk-GWQ24FQP.js";import"../chunk-PPAHIVG5.js";import"../chunk-CT7FWSIU.js";import"../chunk-5MXZJWQS.js";import"../chunk-N7ODX6PO.js";import{a as V}from"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-NAS7FIGM.js";import"../chunk-VZYNZS5U.js";import"../chunk-U2NT5OSQ.js";import"../chunk-KGURGIZ2.js";import"../chunk-GYNZ3OZF.js";import"../chunk-UPFH72PF.js";import"../chunk-3HJT572X.js";import"../chunk-CDY7TQFZ.js";import"../chunk-KFNFLMUR.js";import"../chunk-FLP6WIJV.js";import"../chunk-JP6ULOEX.js";import"../chunk-SVZYRGO5.js";import"../chunk-WULEBQPI.js";import{d as F}from"../chunk-GPBN4ROE.js";import"../chunk-6XTLP2G7.js";import"../chunk-PJ3WHQ67.js";import{a as X,b as D,c as Y,g as j,h as G}from"../chunk-QI3FJKGK.js";import{a as B}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import{a as U,e as y}from"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import{a as N,g as v,l as x}from"../chunk-PZNWH4VQ.js";import{a as S}from"../chunk-ZLLER3HP.js";import{spawnSync as le}from"node:child_process";import{existsSync as ce,readFileSync as ue}from"node:fs";import{join as se}from"node:path";var I=class h extends B{static{S(this,"Diagnostics")}static description='Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';static examples=["<%= config.bin %> diagnostics"];baseUrl(){return process.env.RULEMETRIC_API_URL??N().apiUrl??U}async checkDirect(){let r=process.env.HTTPS_PROXY,i=process.env.HTTP_PROXY;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let s=new AbortController,l=setTimeout(()=>s.abort(),8e3);try{return(await globalThis.fetch(`${this.baseUrl()}/health`,{signal:s.signal})).ok}catch{return!1}finally{clearTimeout(l),r!==void 0&&(process.env.HTTPS_PROXY=r),i!==void 0&&(process.env.HTTP_PROXY=i)}}configuredProxy(){if(process.env.HTTPS_PROXY)return process.env.HTTPS_PROXY;for(let r of["settings.local.json","settings.json"]){let i=se(process.cwd(),".claude",r);if(ce(i))try{let s=JSON.parse(ue(i,"utf8"));if(s.env?.HTTPS_PROXY)return s.env.HTTPS_PROXY}catch{}}return null}checkThroughProxy(r){let i=le("curl",["-sS","-x",r,"--max-time","8","-o","/dev/null","-w","%{http_code}",`${this.baseUrl()}/health`],{encoding:"utf8",stdio:["ignore","pipe","ignore"]});return i.error||typeof i.stdout!="string"?null:i.stdout.trim()==="200"}async run(){await this.parse(h);let r=process.version,i=this.configuredProxy(),s=i!=null,[l,m,d]=await Promise.all([this.checkDirect(),M(O),M(_)]),p=s&&i?this.checkThroughProxy(i):null,t=S((e,c,T)=>{let u=c===null?"\u2013":c?"[OK]":"[!!]";this.log(`  ${u} ${e.padEnd(22)} ${T}`)},"line");this.log("RuleMetric diagnostics"),this.log(""),t("Node",!L(r),`${r}${L(r)?" (< 20.11 \u2014 breaks the bundled proxy)":""}`),t("API (direct)",l,`${this.baseUrl()}/health ${l?"reachable":"UNREACHABLE"}`),t("Authenticated",v(),v()?"yes":"no \u2014 run `rulemetric auth login`"),t("Proxy configured",s?!0:null,s?i:"no HTTPS_PROXY set"),s&&(t("Gateway :"+O,m,m?"listening":"NOT listening"),t("mitmproxy :"+_,d,d?"listening":"not listening (gateway should route direct)"),t("API (via proxy)",p,p===null?"inconclusive (curl unavailable)":p?"reachable":"FAILED \u2014 502/blocked through your proxy"));let a=F();t("claude CLI",a.claudeBin!=null,a.claudeBin??"not found \u2014 LLM features (grading, distill) will no-op"),t("codex CLI",a.codexBin!=null?!0:null,a.codexBin?`${a.codexBin}${a.codexConfiguredModel?` (model ${a.codexConfiguredModel})`:""}`:"not found \u2014 optional; enables the cross-vendor reviewer (evals validate-judge)");let o=null;if(v()&&l)try{o=await y("/api/workers/health")}catch{o=null}if(o){let e=o.worker;t("Worker",!!e.active,e.present?e.active?`online (last heartbeat ${e.ageSeconds??"?"}s ago${e.clientVersion?`, v${e.clientVersion}`:""})`:`NOT RUNNING \u2014 last heartbeat ${A(e.ageSeconds)} ago`:"never connected \u2014 no worker has ever reported in");let c=o.worker.engines;if(c!=null){let u=[c.claude&&"claude",c.codex&&"codex"].filter(Boolean);t("LLM engine",u.length>0,u.length>0?`${u.join(" + ")} available to the worker`:"NONE \u2014 the worker found no `claude` and no `codex`. LLM jobs will complete with empty output rather than failing.")}let T=o.pending.reduce((u,E)=>u+E.count,0);if(T>0){let u=o.verdict.stuck.reduce(($,k)=>$+k.count,0),E=o.verdict.unclaimable.reduce(($,k)=>$+k.count,0),R=(o.verdict.engineBlocked??[]).reduce(($,k)=>$+k.count,0);t("Job queue",u===0&&R===0,`${T} pending`+(u>0?` \u2014 ${u} waiting on this worker`:"")+(R>0?` \u2014 ${R} need a local LLM engine`:"")+(E>0?` \u2014 ${E} need a worker with DATABASE_URL`:""))}else t("Job queue",!0,"empty")}else v()&&l&&t("Worker",null,"could not read /workers/health (older API?)");let n=o?.worker.machine;if(n){let e=n.suspends+n.stalls,c=n.lastPause;t("Worker pauses",e===0?!0:null,e===0?`none since ${n.since} \u2014 no sleep or stall seen by the worker`:`${n.suspends} suspend${n.suspends===1?"":"s"}`+(n.stalls>0?` + ${n.stalls} event-loop stall${n.stalls===1?"":"s"}`:"")+` since ${n.since}, ${A(Math.round(n.pausedMs/1e3))} paused total`+(c?`; last: ${c.kind} of ${A(Math.round(c.wallGapMs/1e3))} ending ${c.detectedAt}`:"")+" \u2014 jobs in flight during a suspend lose their lease and are redone on wake")}let f=await q();if(f){t("Machine sleep (pmset)",f.sleeps===0?!0:null,K(f));for(let e of f.recent.slice(-3)){let c=e.seconds!=null?` (${A(e.seconds)})`:"";this.log(`        \u21B3 ${e.at.toISOString()} ${e.kind}${c}: ${e.reason.slice(0,80)}`)}}let C=V(process.cwd()),g=await D({configDir:x()}),re=await Y({configDir:x()}),w=` \u2014 ${G(re)}`,b=J(process.cwd()),H=[b.user&&`user hooks (${b.user})`,b.workspace&&`workspace hooks (${b.workspace})`].filter(Boolean),ie={claude_code:`session hooks (+ proxy for instruction linking)${w}`,codex:`worker transcript sync of ~/.codex/sessions \u2014 ${j(g)}`,cursor:`user hooks (~/.cursor/hooks.json)${w}`,copilot_agent:`workspace hooks${w}`,antigravity:H.length>0?`${H.join(" + ")}${w}`:"NOT CAPTURED \u2014 no rulemetric block in ~/.gemini/config/hooks.json or .agents/hooks.json; run `rulemetric hooks install` (user hooks are on by default)"};this.log(""),this.log("  Capture coverage (agents found on this machine):");for(let e of C)this.log(`    \u2022 ${e.padEnd(15)} ${ie[e]??"captured"}`);if(C.includes("codex"))if(g.error)this.log(`    ! ${g.error}`);else if(g.mode==="allowlist")for(let e of g.roots)this.log(`        \u21B3 ${e}`);else g.mode==="all"&&this.log(`        \u21B3 scope it with ${se(x(),X)} ({"codex":{"roots":["/path/to/repo"]}}) or RULEMETRIC_CODEX_CAPTURE=off`);o&&!o.worker.active&&C.includes("codex")&&g.mode!=="off"&&this.log("    ! Codex capture runs INSIDE the worker \u2014 it is stopped while the worker is down."),v()&&l&&await this.reportEvolutionSpend();let P=W({nodeVersion:r,directApiOk:l,httpsProxySet:s,gatewayAlive:m,mitmproxyAlive:d,proxyPathOk:p});this.log("");let ae=P.severity==="ok"?"[OK]":P.severity==="warn"?"[!!]":"[XX]";this.log(`${ae} ${P.headline}`);for(let e of P.hints)this.log(`   \u2022 ${e}`);if(o&&o.verdict.severity!=="ok"){let e=o.verdict.severity==="warn"?"[!!]":"[XX]";this.log(""),this.log(`${e} ${o.verdict.headline}`);for(let c of o.verdict.hints)this.log(`   \u2022 ${c}`)}}async reportEvolutionSpend(){let r;try{r=await y("/api/eval-targets?type=instruction&limit=200")}catch{return}let i=r.filter(s=>{let l=s.metadata??{};return l.autoEvolveEnabled===!0||l.autoEvolveEnabled==="true"});if(i.length!==0){this.log(""),this.log("  Nightly instruction evolution (cron_instruction_evolution, 07:30 UTC):");for(let s of i){let l=s.metadata??{},m=Q(l),d=z(l);if(m){this.log(`    \u2022 ${s.name} \u2014 PAUSED (metadata.autoEvolvePaused). Opt-in preserved.`),this.log("        \u21B3 resume: set autoEvolvePaused=false");continue}let p=null;try{p=(await y(`/api/eval-targets/${s.id}/evals`)).filter(a=>!a.retiredAt).length}catch{}if(this.log(`    \u2022 ${s.name}`),p!=null){let t=ne(p,d.minPairs,oe(d,p));if(!t.feasible)this.log(`        \u21B3 cannot reach a verdict \u2014 ${t.reason}`);else{let a=Z({liveCases:p,passes:t.passes});this.log(`        \u21B3 each night: ${te(a)}`)}}this.log(`        \u21B3 pairs budget ${d.minPairs}`+(d.configured?" (metadata.autoEvolvePairs)":" (default)")+` \xB7 writes ${l.autoEvolveApply===!0||l.autoEvolveApply==="true"?"ENABLED":"off (dry run)"}`),d.note&&this.log(`        ! ${d.note}`);try{let t=await y(`/api/eval-targets/${s.id}/promotions?limit=1`),a=t.promotions[0];if(a){let n=a.gates?.cost,f=n?.modelCalls?`${n.modelCalls} model calls, ${n.usablePairs??"?"} usable pair(s)`+(n.mde!=null?`, resolved +/-${ee(n.mde)}`:"")+(n.wallMs?`, ${(n.wallMs/6e4).toFixed(0)} min`:""):"cost not recorded (decision predates cost tracking)";this.log(`        \u21B3 last run ${new Date(a.createdAt).toISOString().slice(0,10)}: ${a.decision}${a.applied?" (APPLIED)":""} \u2014 ${f}`)}let o=t.counts??{};this.log(`        \u21B3 lifetime: ${o.promote??0} promote / ${o.revert??0} revert / ${o.refuse??0} refuse`)}catch{}}this.log("        \u21B3 pause the spend: metadata.autoEvolvePaused=true (keeps the opt-in)")}}};function A(h){if(h===null)return"never";if(h<90)return`${Math.round(h)}s`;let r=Math.round(h/60);return r<90?`${r}m`:`${Math.round(r/60)}h`}S(A,"formatAge");export{I as default};
+import{c as q,d as K}from"../chunk-TYFEWLYZ.js";import{b as L,c as W}from"../chunk-V5U3E4BZ.js";import{a as J}from"../chunk-2TJ7Q2OU.js";import{e as O,f as _,g as M}from"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import{d as oe,e as ne}from"../chunk-L5ILEWHH.js";import"../chunk-M7QSC64D.js";import"../chunk-DZRBB34J.js";import"../chunk-KR32Y5Q5.js";import"../chunk-G7V4ZZZ3.js";import"../chunk-RITL4NU3.js";import{i as z,j as Q,k as Z,m as ee,n as te}from"../chunk-GYKYJJNF.js";import"../chunk-V6L25LFD.js";import"../chunk-GWQ24FQP.js";import"../chunk-PPAHIVG5.js";import"../chunk-CT7FWSIU.js";import"../chunk-5MXZJWQS.js";import"../chunk-AHLVY52I.js";import"../chunk-PPQFN62M.js";import"../chunk-N7ODX6PO.js";import{a as V}from"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-NAS7FIGM.js";import"../chunk-VZYNZS5U.js";import"../chunk-U2NT5OSQ.js";import"../chunk-KGURGIZ2.js";import"../chunk-GYNZ3OZF.js";import"../chunk-UPFH72PF.js";import"../chunk-3HJT572X.js";import"../chunk-CDY7TQFZ.js";import"../chunk-KFNFLMUR.js";import"../chunk-FLP6WIJV.js";import"../chunk-JP6ULOEX.js";import"../chunk-SVZYRGO5.js";import"../chunk-WULEBQPI.js";import{d as F}from"../chunk-GPBN4ROE.js";import"../chunk-6XTLP2G7.js";import"../chunk-PJ3WHQ67.js";import{a as X,b as D,c as Y,g as j,h as G}from"../chunk-QI3FJKGK.js";import{a as B}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import{a as U,e as y}from"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import{a as N,g as v,l as x}from"../chunk-PZNWH4VQ.js";import{a as S}from"../chunk-ZLLER3HP.js";import{spawnSync as le}from"node:child_process";import{existsSync as ce,readFileSync as ue}from"node:fs";import{join as se}from"node:path";var I=class h extends B{static{S(this,"Diagnostics")}static description='Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';static examples=["<%= config.bin %> diagnostics"];baseUrl(){return process.env.RULEMETRIC_API_URL??N().apiUrl??U}async checkDirect(){let r=process.env.HTTPS_PROXY,i=process.env.HTTP_PROXY;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let s=new AbortController,l=setTimeout(()=>s.abort(),8e3);try{return(await globalThis.fetch(`${this.baseUrl()}/health`,{signal:s.signal})).ok}catch{return!1}finally{clearTimeout(l),r!==void 0&&(process.env.HTTPS_PROXY=r),i!==void 0&&(process.env.HTTP_PROXY=i)}}configuredProxy(){if(process.env.HTTPS_PROXY)return process.env.HTTPS_PROXY;for(let r of["settings.local.json","settings.json"]){let i=se(process.cwd(),".claude",r);if(ce(i))try{let s=JSON.parse(ue(i,"utf8"));if(s.env?.HTTPS_PROXY)return s.env.HTTPS_PROXY}catch{}}return null}checkThroughProxy(r){let i=le("curl",["-sS","-x",r,"--max-time","8","-o","/dev/null","-w","%{http_code}",`${this.baseUrl()}/health`],{encoding:"utf8",stdio:["ignore","pipe","ignore"]});return i.error||typeof i.stdout!="string"?null:i.stdout.trim()==="200"}async run(){await this.parse(h);let r=process.version,i=this.configuredProxy(),s=i!=null,[l,m,d]=await Promise.all([this.checkDirect(),M(O),M(_)]),p=s&&i?this.checkThroughProxy(i):null,t=S((e,c,T)=>{let u=c===null?"\u2013":c?"[OK]":"[!!]";this.log(`  ${u} ${e.padEnd(22)} ${T}`)},"line");this.log("RuleMetric diagnostics"),this.log(""),t("Node",!L(r),`${r}${L(r)?" (< 20.11 \u2014 breaks the bundled proxy)":""}`),t("API (direct)",l,`${this.baseUrl()}/health ${l?"reachable":"UNREACHABLE"}`),t("Authenticated",v(),v()?"yes":"no \u2014 run `rulemetric auth login`"),t("Proxy configured",s?!0:null,s?i:"no HTTPS_PROXY set"),s&&(t("Gateway :"+O,m,m?"listening":"NOT listening"),t("mitmproxy :"+_,d,d?"listening":"not listening (gateway should route direct)"),t("API (via proxy)",p,p===null?"inconclusive (curl unavailable)":p?"reachable":"FAILED \u2014 502/blocked through your proxy"));let a=F();t("claude CLI",a.claudeBin!=null,a.claudeBin??"not found \u2014 LLM features (grading, distill) will no-op"),t("codex CLI",a.codexBin!=null?!0:null,a.codexBin?`${a.codexBin}${a.codexConfiguredModel?` (model ${a.codexConfiguredModel})`:""}`:"not found \u2014 optional; enables the cross-vendor reviewer (evals validate-judge)");let o=null;if(v()&&l)try{o=await y("/api/workers/health")}catch{o=null}if(o){let e=o.worker;t("Worker",!!e.active,e.present?e.active?`online (last heartbeat ${e.ageSeconds??"?"}s ago${e.clientVersion?`, v${e.clientVersion}`:""})`:`NOT RUNNING \u2014 last heartbeat ${A(e.ageSeconds)} ago`:"never connected \u2014 no worker has ever reported in");let c=o.worker.engines;if(c!=null){let u=[c.claude&&"claude",c.codex&&"codex"].filter(Boolean);t("LLM engine",u.length>0,u.length>0?`${u.join(" + ")} available to the worker`:"NONE \u2014 the worker found no `claude` and no `codex`. LLM jobs will complete with empty output rather than failing.")}let T=o.pending.reduce((u,E)=>u+E.count,0);if(T>0){let u=o.verdict.stuck.reduce(($,k)=>$+k.count,0),E=o.verdict.unclaimable.reduce(($,k)=>$+k.count,0),R=(o.verdict.engineBlocked??[]).reduce(($,k)=>$+k.count,0);t("Job queue",u===0&&R===0,`${T} pending`+(u>0?` \u2014 ${u} waiting on this worker`:"")+(R>0?` \u2014 ${R} need a local LLM engine`:"")+(E>0?` \u2014 ${E} need a worker with DATABASE_URL`:""))}else t("Job queue",!0,"empty")}else v()&&l&&t("Worker",null,"could not read /workers/health (older API?)");let n=o?.worker.machine;if(n){let e=n.suspends+n.stalls,c=n.lastPause;t("Worker pauses",e===0?!0:null,e===0?`none since ${n.since} \u2014 no sleep or stall seen by the worker`:`${n.suspends} suspend${n.suspends===1?"":"s"}`+(n.stalls>0?` + ${n.stalls} event-loop stall${n.stalls===1?"":"s"}`:"")+` since ${n.since}, ${A(Math.round(n.pausedMs/1e3))} paused total`+(c?`; last: ${c.kind} of ${A(Math.round(c.wallGapMs/1e3))} ending ${c.detectedAt}`:"")+" \u2014 jobs in flight during a suspend lose their lease and are redone on wake")}let f=await q();if(f){t("Machine sleep (pmset)",f.sleeps===0?!0:null,K(f));for(let e of f.recent.slice(-3)){let c=e.seconds!=null?` (${A(e.seconds)})`:"";this.log(`        \u21B3 ${e.at.toISOString()} ${e.kind}${c}: ${e.reason.slice(0,80)}`)}}let C=V(process.cwd()),g=await D({configDir:x()}),re=await Y({configDir:x()}),w=` \u2014 ${G(re)}`,b=J(process.cwd()),H=[b.user&&`user hooks (${b.user})`,b.workspace&&`workspace hooks (${b.workspace})`].filter(Boolean),ie={claude_code:`session hooks (+ proxy for instruction linking)${w}`,codex:`worker transcript sync of ~/.codex/sessions \u2014 ${j(g)}`,cursor:`user hooks (~/.cursor/hooks.json)${w}`,copilot_agent:`workspace hooks${w}`,antigravity:H.length>0?`${H.join(" + ")}${w}`:"NOT CAPTURED \u2014 no rulemetric block in ~/.gemini/config/hooks.json or .agents/hooks.json; run `rulemetric hooks install` (user hooks are on by default)"};this.log(""),this.log("  Capture coverage (agents found on this machine):");for(let e of C)this.log(`    \u2022 ${e.padEnd(15)} ${ie[e]??"captured"}`);if(C.includes("codex"))if(g.error)this.log(`    ! ${g.error}`);else if(g.mode==="allowlist")for(let e of g.roots)this.log(`        \u21B3 ${e}`);else g.mode==="all"&&this.log(`        \u21B3 scope it with ${se(x(),X)} ({"codex":{"roots":["/path/to/repo"]}}) or RULEMETRIC_CODEX_CAPTURE=off`);o&&!o.worker.active&&C.includes("codex")&&g.mode!=="off"&&this.log("    ! Codex capture runs INSIDE the worker \u2014 it is stopped while the worker is down."),v()&&l&&await this.reportEvolutionSpend();let P=W({nodeVersion:r,directApiOk:l,httpsProxySet:s,gatewayAlive:m,mitmproxyAlive:d,proxyPathOk:p});this.log("");let ae=P.severity==="ok"?"[OK]":P.severity==="warn"?"[!!]":"[XX]";this.log(`${ae} ${P.headline}`);for(let e of P.hints)this.log(`   \u2022 ${e}`);if(o&&o.verdict.severity!=="ok"){let e=o.verdict.severity==="warn"?"[!!]":"[XX]";this.log(""),this.log(`${e} ${o.verdict.headline}`);for(let c of o.verdict.hints)this.log(`   \u2022 ${c}`)}}async reportEvolutionSpend(){let r;try{r=await y("/api/eval-targets?type=instruction&limit=200")}catch{return}let i=r.filter(s=>{let l=s.metadata??{};return l.autoEvolveEnabled===!0||l.autoEvolveEnabled==="true"});if(i.length!==0){this.log(""),this.log("  Nightly instruction evolution (cron_instruction_evolution, 07:30 UTC):");for(let s of i){let l=s.metadata??{},m=Q(l),d=z(l);if(m){this.log(`    \u2022 ${s.name} \u2014 PAUSED (metadata.autoEvolvePaused). Opt-in preserved.`),this.log("        \u21B3 resume: set autoEvolvePaused=false");continue}let p=null;try{p=(await y(`/api/eval-targets/${s.id}/evals`)).filter(a=>!a.retiredAt).length}catch{}if(this.log(`    \u2022 ${s.name}`),p!=null){let t=ne(p,d.minPairs,oe(d,p));if(!t.feasible)this.log(`        \u21B3 cannot reach a verdict \u2014 ${t.reason}`);else{let a=Z({liveCases:p,passes:t.passes});this.log(`        \u21B3 each night: ${te(a)}`)}}this.log(`        \u21B3 pairs budget ${d.minPairs}`+(d.configured?" (metadata.autoEvolvePairs)":" (default)")+` \xB7 writes ${l.autoEvolveApply===!0||l.autoEvolveApply==="true"?"ENABLED":"off (dry run)"}`),d.note&&this.log(`        ! ${d.note}`);try{let t=await y(`/api/eval-targets/${s.id}/promotions?limit=1`),a=t.promotions[0];if(a){let n=a.gates?.cost,f=n?.modelCalls?`${n.modelCalls} model calls, ${n.usablePairs??"?"} usable pair(s)`+(n.mde!=null?`, resolved +/-${ee(n.mde)}`:"")+(n.wallMs?`, ${(n.wallMs/6e4).toFixed(0)} min`:""):"cost not recorded (decision predates cost tracking)";this.log(`        \u21B3 last run ${new Date(a.createdAt).toISOString().slice(0,10)}: ${a.decision}${a.applied?" (APPLIED)":""} \u2014 ${f}`)}let o=t.counts??{};this.log(`        \u21B3 lifetime: ${o.promote??0} promote / ${o.revert??0} revert / ${o.refuse??0} refuse`)}catch{}}this.log("        \u21B3 pause the spend: metadata.autoEvolvePaused=true (keeps the opt-in)")}}};function A(h){if(h===null)return"never";if(h<90)return`${Math.round(h)}s`;let r=Math.round(h/60);return r<90?`${r}m`:`${Math.round(r/60)}h`}S(A,"formatAge");export{I as default};

package/dist/commands/evals/agent.js

@@ -1,3 +1,3 @@
-import{a as G,b as V,c as j,d as q}from"../../chunk-4LVSD2XZ.js";import{b as M}from"../../chunk-GH4EZK2O.js";import{c as g}from"../../chunk-BDWQ6J4V.js";import{e as D}from"../../chunk-JLX2D3QL.js";import"../../chunk-5U5SHJJA.js";import"../../chunk-IEQDLURA.js";import"../../chunk-KINIFUZQ.js";import"../../chunk-QRKMRG6W.js";import"../../chunk-D7L7IMJQ.js";import"../../chunk-MRQIYIQE.js";import"../../chunk-R26SQ3BE.js";import"../../chunk-PMMHMRFS.js";import"../../chunk-IYX65N6T.js";import"../../chunk-GHUW7P55.js";import"../../chunk-JU5ZNXPK.js";import"../../chunk-MBP5OJKE.js";import"../../chunk-BRYURAUN.js";import"../../chunk-NLUHMHPM.js";import"../../chunk-DTTR4EUJ.js";import"../../chunk-WLLIFRRE.js";import"../../chunk-RRBFKVCI.js";import"../../chunk-ZYGBABMZ.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-NSO2EBYE.js";import"../../chunk-3ZHVP3KE.js";import"../../chunk-2C7CFCMM.js";import"../../chunk-HOQPVOMF.js";import"../../chunk-WSUKOU3X.js";import"../../chunk-A6XQSO62.js";import"../../chunk-MOJY6S5J.js";import"../../chunk-CJ2WSSHD.js";import"../../chunk-KOIFFSYW.js";import"../../chunk-IPNBW7ZW.js";import"../../chunk-P577DWWL.js";import"../../chunk-7MQZPDDZ.js";import"../../chunk-ANLPPIMS.js";import"../../chunk-AZRGURHR.js";import"../../chunk-TG4C4ENM.js";import"../../chunk-MJLW5KN2.js";import"../../chunk-SW24KZ5C.js";import"../../chunk-K65DDLTZ.js";import"../../chunk-LELSVLNK.js";import"../../chunk-A2JFQUDW.js";import"../../chunk-LZPE6424.js";import"../../chunk-3UKZJAFS.js";import"../../chunk-77FH6UZP.js";import"../../chunk-FGGESTXG.js";import"../../chunk-PX2E644Q.js";import"../../chunk-GIRPEMTW.js";import"../../chunk-DQYYHX6K.js";import"../../chunk-RWQG25JU.js";import"../../chunk-3ZHBOILE.js";import"../../chunk-PIVEJTNF.js";import"../../chunk-BOYWPUV2.js";import"../../chunk-ENW5SUQJ.js";import"../../chunk-C3XWDYZJ.js";import"../../chunk-CAQLSYOK.js";import"../../chunk-RQBN7WNS.js";import"../../chunk-ZJABRHIP.js";import"../../chunk-XFOG2PDJ.js";import"../../chunk-SIGZEQ3K.js";import"../../chunk-AZWY2DOT.js";import"../../chunk-LJRE3LZ2.js";import"../../chunk-EQ2O6DSN.js";import"../../chunk-BTKVLLV2.js";import"../../chunk-UQJNUESP.js";import"../../chunk-FLXWL4F4.js";import"../../chunk-KIQK7NVF.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-MHVHGKDK.js";import"../../chunk-RBSEE2Z6.js";import"../../chunk-NBTH6ZKD.js";import"../../chunk-MGT5FZJ4.js";import"../../chunk-M7QSC64D.js";import"../../chunk-DZRBB34J.js";import"../../chunk-KR32Y5Q5.js";import"../../chunk-RITL4NU3.js";import"../../chunk-GYKYJJNF.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-AHLVY52I.js";import"../../chunk-PPQFN62M.js";import"../../chunk-V6L25LFD.js";import"../../chunk-GWQ24FQP.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-NAS7FIGM.js";import"../../chunk-VZYNZS5U.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-3HJT572X.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-KFNFLMUR.js";import"../../chunk-FLP6WIJV.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SVZYRGO5.js";import"../../chunk-WULEBQPI.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{b as A,d as C,g as S,i as E,j as L,k as $}from"../../chunk-QI3FJKGK.js";import{a as O}from"../../chunk-TQFHRD5B.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5LTJNYCP.js";import"../../chunk-3EZZGA72.js";import"../../chunk-XAGNGQL3.js";import"../../chunk-3AV3GTMP.js";import{f as m}from"../../chunk-5VWVQCNR.js";import"../../chunk-MHLHLJ5T.js";import"../../chunk-XQYAVSUQ.js";import{e as I,l as p}from"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as de,context as me}from"@opentelemetry/api";import{NodeSDK as W}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as Y}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as J,ConsoleSpanExporter as z,SimpleSpanProcessor as Q}from"@opentelemetry/sdk-trace-node";import{Resource as Z}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as ee,ATTR_SERVICE_VERSION as te}from"@opentelemetry/semantic-conventions";function F(r){if(!r)return{};let n={};for(let o of r.split(",")){let s=o.indexOf("=");if(s===-1)continue;let i=o.slice(0,s).trim(),c=o.slice(s+1).trim();i&&(n[i]=c)}return n}e(F,"parseOtlpHeaders");var u=null;function T(){return process.env.TELEMETRY_ENABLED!=="false"}e(T,"isTelemetryEnabled");function x(r){if(!T()||u!==null)return"off";let n=new Z({[ee]:r.serviceName,[te]:r.serviceVersion??"0.1.0"}),o=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(o){let c=F(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new Y({url:`${o}/v1/traces`,headers:c});i=new J(l),s="otlp"}else if(process.env.OTEL_SPAN_EXPORT==="none")i=void 0,s="silent";else{let c=new z;i=new Q(c),s="console"}return u=new W({resource:n,spanProcessors:i?[i]:[]}),u.start(),s}e(x,"initTelemetry");async function h(){u&&(await u.shutdown(),u=null)}e(h,"shutdownTelemetry");import{SpanKind as Pe,SpanStatusCode as Re}from"@opentelemetry/api";import{trace as Ie,context as Oe}from"@opentelemetry/api";import*as f from"@sentry/node";var v=!1;function w(r){if(v)return!1;let n=process.env.SENTRY_DSN;return n?(f.init({dsn:n,serverName:r.serviceName,environment:process.env.NODE_ENV??"development",tracesSampleRate:0}),v=!0,!0):!1}e(w,"initSentry");async function y(r){v&&await f.flush(r)}e(y,"flushSentry");import{existsSync as ne,mkdirSync as oe,readFileSync as k,unlinkSync as X,writeFileSync as U}from"node:fs";import{dirname as se,join as R}from"node:path";import{execFileSync as ie}from"node:child_process";var a=R(p(),"evals-agent.lock");function ce(r){try{return process.kill(r,0),!0}catch(n){return n.code==="EPERM"}}e(ce,"pidAlive");function ae(r){if(!ce(r))return!1;try{let n=ie("ps",["-p",String(r),"-o","command="],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return/run\.js\s+evals\s+agent|evals\s+agent/.test(n)}catch{return!0}}e(ae,"isLiveAgent");function le(){try{oe(se(a),{recursive:!0})}catch{}for(let r=0;r<2;r++)try{return U(a,String(process.pid),{flag:"wx"}),{ok:!0}}catch(n){if(n.code!=="EEXIST")throw n;let o=0;try{o=Number(k(a,"utf-8").trim())}catch{}if(o&&o!==process.pid&&ae(o))return{ok:!1,holder:o};try{X(a)}catch{}}try{U(a,String(process.pid))}catch{}return{ok:!0}}e(le,"acquireAgentLock");function P(){try{ne(a)&&Number(k(a,"utf-8").trim())===process.pid&&X(a)}catch{}}e(P,"releaseAgentLock");var N=class r extends O{static{e(this,"EvalAgentCommand")}static description="The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.";static examples=["<%= config.bin %> evals agent"];async run(){await this.parse(r),this.requireAuth();let n=le();if(!n.ok){this.log(`Another rulemetric evals agent is already running (pid ${n.holder}). Refusing to start a second instance.`);return}process.env.PG_APPLICATION_NAME||="rulemetric-agent",w({serviceName:"rulemetric-agent"})&&this.log("[sentry] error capture enabled");let s=x({serviceName:"rulemetric-agent"});s==="otlp"?this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`):s==="console"?this.log("[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)"):this.log("[otel] disabled (TELEMETRY_ENABLED=false)"),!I()&&!process.env.RULEMETRIC_API_KEY&&this.error("No auth token found. Run `rulemetric auth login` first."),this.log("Starting thin agent (Phase 4b parallel-run)..."),this.log(`Long-polling /api/work/next. Press Ctrl+C to stop.
+import{a as G,b as V,c as j,d as q}from"../../chunk-4LVSD2XZ.js";import{b as M}from"../../chunk-GH4EZK2O.js";import{c as g}from"../../chunk-BDWQ6J4V.js";import{e as D}from"../../chunk-NVOYP4HX.js";import"../../chunk-5U5SHJJA.js";import"../../chunk-IEQDLURA.js";import"../../chunk-KINIFUZQ.js";import"../../chunk-ZYGBABMZ.js";import"../../chunk-QRKMRG6W.js";import"../../chunk-D7L7IMJQ.js";import"../../chunk-MRQIYIQE.js";import"../../chunk-R26SQ3BE.js";import"../../chunk-PMMHMRFS.js";import"../../chunk-IYX65N6T.js";import"../../chunk-GHUW7P55.js";import"../../chunk-JU5ZNXPK.js";import"../../chunk-RAMQVEJB.js";import"../../chunk-BRYURAUN.js";import"../../chunk-NLUHMHPM.js";import"../../chunk-DTTR4EUJ.js";import"../../chunk-WLLIFRRE.js";import"../../chunk-RRBFKVCI.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-SW24KZ5C.js";import"../../chunk-NSO2EBYE.js";import"../../chunk-3ZHVP3KE.js";import"../../chunk-2C7CFCMM.js";import"../../chunk-HOQPVOMF.js";import"../../chunk-WSUKOU3X.js";import"../../chunk-A6XQSO62.js";import"../../chunk-MOJY6S5J.js";import"../../chunk-CJ2WSSHD.js";import"../../chunk-KOIFFSYW.js";import"../../chunk-IPNBW7ZW.js";import"../../chunk-P577DWWL.js";import"../../chunk-7MQZPDDZ.js";import"../../chunk-ANLPPIMS.js";import"../../chunk-AZRGURHR.js";import"../../chunk-CXENLSMR.js";import"../../chunk-MJLW5KN2.js";import"../../chunk-77FH6UZP.js";import"../../chunk-K65DDLTZ.js";import"../../chunk-LELSVLNK.js";import"../../chunk-A2JFQUDW.js";import"../../chunk-LZPE6424.js";import"../../chunk-3UKZJAFS.js";import"../../chunk-FGGESTXG.js";import"../../chunk-PX2E644Q.js";import"../../chunk-DQYYHX6K.js";import"../../chunk-GIRPEMTW.js";import"../../chunk-RWQG25JU.js";import"../../chunk-3ZHBOILE.js";import"../../chunk-PIVEJTNF.js";import"../../chunk-BOYWPUV2.js";import"../../chunk-ENW5SUQJ.js";import"../../chunk-C3XWDYZJ.js";import"../../chunk-CAQLSYOK.js";import"../../chunk-RQBN7WNS.js";import"../../chunk-ZJABRHIP.js";import"../../chunk-XFOG2PDJ.js";import"../../chunk-SIGZEQ3K.js";import"../../chunk-AZWY2DOT.js";import"../../chunk-LJRE3LZ2.js";import"../../chunk-YZQZD2E6.js";import"../../chunk-FLJTAD2R.js";import"../../chunk-UQJNUESP.js";import"../../chunk-FLXWL4F4.js";import"../../chunk-KIQK7NVF.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-MHVHGKDK.js";import"../../chunk-RBSEE2Z6.js";import"../../chunk-NBTH6ZKD.js";import"../../chunk-L5ILEWHH.js";import"../../chunk-M7QSC64D.js";import"../../chunk-DZRBB34J.js";import"../../chunk-KR32Y5Q5.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-RITL4NU3.js";import"../../chunk-GYKYJJNF.js";import"../../chunk-V6L25LFD.js";import"../../chunk-GWQ24FQP.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-AHLVY52I.js";import"../../chunk-PPQFN62M.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-NAS7FIGM.js";import"../../chunk-VZYNZS5U.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-3HJT572X.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-KFNFLMUR.js";import"../../chunk-FLP6WIJV.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SVZYRGO5.js";import"../../chunk-WULEBQPI.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{b as A,d as C,g as S,i as E,j as L,k as $}from"../../chunk-QI3FJKGK.js";import{a as O}from"../../chunk-TQFHRD5B.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5LTJNYCP.js";import"../../chunk-3EZZGA72.js";import"../../chunk-XAGNGQL3.js";import"../../chunk-3AV3GTMP.js";import{f as m}from"../../chunk-5VWVQCNR.js";import"../../chunk-MHLHLJ5T.js";import"../../chunk-XQYAVSUQ.js";import{e as I,l as p}from"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as de,context as me}from"@opentelemetry/api";import{NodeSDK as W}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as Y}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as J,ConsoleSpanExporter as z,SimpleSpanProcessor as Q}from"@opentelemetry/sdk-trace-node";import{Resource as Z}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as ee,ATTR_SERVICE_VERSION as te}from"@opentelemetry/semantic-conventions";function F(r){if(!r)return{};let n={};for(let o of r.split(",")){let s=o.indexOf("=");if(s===-1)continue;let i=o.slice(0,s).trim(),c=o.slice(s+1).trim();i&&(n[i]=c)}return n}e(F,"parseOtlpHeaders");var u=null;function T(){return process.env.TELEMETRY_ENABLED!=="false"}e(T,"isTelemetryEnabled");function x(r){if(!T()||u!==null)return"off";let n=new Z({[ee]:r.serviceName,[te]:r.serviceVersion??"0.1.0"}),o=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(o){let c=F(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new Y({url:`${o}/v1/traces`,headers:c});i=new J(l),s="otlp"}else if(process.env.OTEL_SPAN_EXPORT==="none")i=void 0,s="silent";else{let c=new z;i=new Q(c),s="console"}return u=new W({resource:n,spanProcessors:i?[i]:[]}),u.start(),s}e(x,"initTelemetry");async function h(){u&&(await u.shutdown(),u=null)}e(h,"shutdownTelemetry");import{SpanKind as Pe,SpanStatusCode as Re}from"@opentelemetry/api";import{trace as Ie,context as Oe}from"@opentelemetry/api";import*as f from"@sentry/node";var v=!1;function w(r){if(v)return!1;let n=process.env.SENTRY_DSN;return n?(f.init({dsn:n,serverName:r.serviceName,environment:process.env.NODE_ENV??"development",tracesSampleRate:0}),v=!0,!0):!1}e(w,"initSentry");async function y(r){v&&await f.flush(r)}e(y,"flushSentry");import{existsSync as ne,mkdirSync as oe,readFileSync as k,unlinkSync as X,writeFileSync as U}from"node:fs";import{dirname as se,join as R}from"node:path";import{execFileSync as ie}from"node:child_process";var a=R(p(),"evals-agent.lock");function ce(r){try{return process.kill(r,0),!0}catch(n){return n.code==="EPERM"}}e(ce,"pidAlive");function ae(r){if(!ce(r))return!1;try{let n=ie("ps",["-p",String(r),"-o","command="],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return/run\.js\s+evals\s+agent|evals\s+agent/.test(n)}catch{return!0}}e(ae,"isLiveAgent");function le(){try{oe(se(a),{recursive:!0})}catch{}for(let r=0;r<2;r++)try{return U(a,String(process.pid),{flag:"wx"}),{ok:!0}}catch(n){if(n.code!=="EEXIST")throw n;let o=0;try{o=Number(k(a,"utf-8").trim())}catch{}if(o&&o!==process.pid&&ae(o))return{ok:!1,holder:o};try{X(a)}catch{}}try{U(a,String(process.pid))}catch{}return{ok:!0}}e(le,"acquireAgentLock");function P(){try{ne(a)&&Number(k(a,"utf-8").trim())===process.pid&&X(a)}catch{}}e(P,"releaseAgentLock");var N=class r extends O{static{e(this,"EvalAgentCommand")}static description="The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.";static examples=["<%= config.bin %> evals agent"];async run(){await this.parse(r),this.requireAuth();let n=le();if(!n.ok){this.log(`Another rulemetric evals agent is already running (pid ${n.holder}). Refusing to start a second instance.`);return}process.env.PG_APPLICATION_NAME||="rulemetric-agent",w({serviceName:"rulemetric-agent"})&&this.log("[sentry] error capture enabled");let s=x({serviceName:"rulemetric-agent"});s==="otlp"?this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`):s==="console"?this.log("[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)"):this.log("[otel] disabled (TELEMETRY_ENABLED=false)"),!I()&&!process.env.RULEMETRIC_API_KEY&&this.error("No auth token found. Run `rulemetric auth login` first."),this.log("Starting thin agent (Phase 4b parallel-run)..."),this.log(`Long-polling /api/work/next. Press Ctrl+C to stop.
 `);let c=new AbortController,l=await A({configDir:p()});if(this.log(`[transcript-sync] Codex capture: ${S(l)}`),l.error)this.log(`[transcript-sync] WARNING: ${l.error}`);else if(l.mode==="allowlist")for(let t of l.roots)this.log(`[transcript-sync]   \u2022 ${t}`);let d=await C({configDir:p()});if(this.log(`[transcript-sync] Pi capture: ${E(d)}`),d.error)this.log(`[transcript-sync] WARNING: ${d.error}`);else if(d.mode==="allowlist")for(let t of d.roots)this.log(`[transcript-sync]   \u2022 ${t}`);let B=new M({postSnapshot:e((t,K)=>m(`/api/sessions/${t}/context-snapshots`,K),"postSnapshot"),linkInstructions:e(t=>m("/api/sessions/link-instructions",{sessionId:t}),"linkInstructions"),log:e(t=>this.log(`[transcript-sync] ${t}`),"log")}),H=new G({providers:[V(void 0,void 0,L({configDir:p(),onChange:e(t=>this.log(`[transcript-sync] Codex capture changed: ${S(t)}`),"onChange")})),j(void 0,void 0,$({configDir:p(),onChange:e(t=>this.log(`[transcript-sync] Pi capture changed: ${E(t)}`),"onChange")}))],statePath:R(p(),"transcript-sync-state.json"),importTranscript:e(t=>m("/api/sessions/import",t),"importTranscript"),emitInstructions:e(t=>B.emit(t),"emitInstructions"),log:e(t=>this.log(`[transcript-sync] ${t}`),"log")});q(H,{signal:c.signal,log:e(t=>this.log(`[transcript-sync] ${t}`),"log")});let _=e(async t=>{this.log(`
 [${t}] shutting down agent...`),c.abort(),await Promise.allSettled([y(2e3),h(),g()]),P(),process.exit(0)},"shutdown");process.on("SIGINT",()=>{_("SIGINT")}),process.on("SIGTERM",()=>{_("SIGTERM")});let b=!1;try{await D({signal:c.signal,log:e(t=>this.log(t),"log"),version:this.config.version,readInstalledVersion:e(()=>{try{return JSON.parse(k(R(this.config.root,"package.json"),"utf-8")).version??null}catch{return null}},"readInstalledVersion"),onVersionDrift:e(()=>{b=!0},"onVersionDrift")}),b&&(this.log("[upgrade] Installed CLI version changed \u2014 restarting worker onto the new code."),await Promise.allSettled([y(2e3),h(),g()]),P(),process.exit(1))}finally{P(),await g()}}};export{N as default};

package/dist/commands/evals/refine-cases.js

@@ -1,4 +1,4 @@
-import{b as w,c as $}from"../../chunk-EQ2O6DSN.js";import"../../chunk-BTKVLLV2.js";import{b as v}from"../../chunk-UQJNUESP.js";import"../../chunk-FLXWL4F4.js";import"../../chunk-MGT5FZJ4.js";import"../../chunk-M7QSC64D.js";import"../../chunk-DZRBB34J.js";import"../../chunk-KR32Y5Q5.js";import"../../chunk-RITL4NU3.js";import"../../chunk-GYKYJJNF.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-AHLVY52I.js";import"../../chunk-PPQFN62M.js";import"../../chunk-V6L25LFD.js";import"../../chunk-GWQ24FQP.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-NAS7FIGM.js";import"../../chunk-VZYNZS5U.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-3HJT572X.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-KFNFLMUR.js";import"../../chunk-FLP6WIJV.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SVZYRGO5.js";import"../../chunk-WULEBQPI.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-TQFHRD5B.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5LTJNYCP.js";import"../../chunk-3EZZGA72.js";import"../../chunk-XAGNGQL3.js";import"../../chunk-3AV3GTMP.js";import{e as d,f as p,h}from"../../chunk-5VWVQCNR.js";import"../../chunk-MHLHLJ5T.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Args as P,Flags as y}from"@oclif/core";var c=a(s=>s==null?"n/a":`${Math.round(s*100)}%`,"pct"),g=class s extends f{static{a(this,"EvalRefineCases")}static description="Retire non-discriminating eval cases and regenerate discriminating replacements (\xA76e)";static examples=["<%= config.bin %> evals refine-cases <target-id>            # preview","<%= config.bin %> evals refine-cases <target-id> --yes      # apply: retire + regenerate","<%= config.bin %> evals refine-cases <target-id> --yes --no-generate"];static args={targetId:P.string({description:"Eval target ID",required:!0})};static flags={yes:y.boolean({char:"y",description:"Apply (retire + regenerate). Without this, only previews."}),"no-generate":y.boolean({description:"Retire dead cases but do not regenerate replacements."})};async run(){let{args:b,flags:l}=await this.parse(s);this.requireAuth();let n=b.targetId,o=await d(`/api/eval-targets/${n}`),m=await d(`/api/eval-targets/${n}/judge-calibration`),i=m.caseDiscrimination,t=w(i);if(this.log(`Target "${o.name}" v${o.version} \u2014 judge ${m.judge.currentModel??"(unstamped)"}`),t.refused==="no-discrimination-data"){this.log("No case-discrimination data yet \u2014 run some evals with AND without the target first (`evals run` / the web Run Evals button), then refine.");return}if(this.log(`${i.total} diagnosable case(s): ${i.discriminating} discriminate, ${i.nonDiscriminating} flat/wrong-sign, ${i.unpaired} unpaired.`),t.refused==="all-discriminating"){this.log("Every case discriminates \u2014 nothing to retire. \u{1F389}");return}let C=i.cases.filter(e=>!e.discriminating),u=new Set(t.retire.map(e=>e.evalId));this.log(`
+import{b as w,c as $}from"../../chunk-YZQZD2E6.js";import"../../chunk-FLJTAD2R.js";import{b as v}from"../../chunk-UQJNUESP.js";import"../../chunk-FLXWL4F4.js";import"../../chunk-L5ILEWHH.js";import"../../chunk-M7QSC64D.js";import"../../chunk-DZRBB34J.js";import"../../chunk-KR32Y5Q5.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-RITL4NU3.js";import"../../chunk-GYKYJJNF.js";import"../../chunk-V6L25LFD.js";import"../../chunk-GWQ24FQP.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-AHLVY52I.js";import"../../chunk-PPQFN62M.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-NAS7FIGM.js";import"../../chunk-VZYNZS5U.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-3HJT572X.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-KFNFLMUR.js";import"../../chunk-FLP6WIJV.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SVZYRGO5.js";import"../../chunk-WULEBQPI.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-TQFHRD5B.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5LTJNYCP.js";import"../../chunk-3EZZGA72.js";import"../../chunk-XAGNGQL3.js";import"../../chunk-3AV3GTMP.js";import{e as d,f as p,h}from"../../chunk-5VWVQCNR.js";import"../../chunk-MHLHLJ5T.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Args as P,Flags as y}from"@oclif/core";var c=a(s=>s==null?"n/a":`${Math.round(s*100)}%`,"pct"),g=class s extends f{static{a(this,"EvalRefineCases")}static description="Retire non-discriminating eval cases and regenerate discriminating replacements (\xA76e)";static examples=["<%= config.bin %> evals refine-cases <target-id>            # preview","<%= config.bin %> evals refine-cases <target-id> --yes      # apply: retire + regenerate","<%= config.bin %> evals refine-cases <target-id> --yes --no-generate"];static args={targetId:P.string({description:"Eval target ID",required:!0})};static flags={yes:y.boolean({char:"y",description:"Apply (retire + regenerate). Without this, only previews."}),"no-generate":y.boolean({description:"Retire dead cases but do not regenerate replacements."})};async run(){let{args:b,flags:l}=await this.parse(s);this.requireAuth();let n=b.targetId,o=await d(`/api/eval-targets/${n}`),m=await d(`/api/eval-targets/${n}/judge-calibration`),i=m.caseDiscrimination,t=w(i);if(this.log(`Target "${o.name}" v${o.version} \u2014 judge ${m.judge.currentModel??"(unstamped)"}`),t.refused==="no-discrimination-data"){this.log("No case-discrimination data yet \u2014 run some evals with AND without the target first (`evals run` / the web Run Evals button), then refine.");return}if(this.log(`${i.total} diagnosable case(s): ${i.discriminating} discriminate, ${i.nonDiscriminating} flat/wrong-sign, ${i.unpaired} unpaired.`),t.refused==="all-discriminating"){this.log("Every case discriminates \u2014 nothing to retire. \u{1F389}");return}let C=i.cases.filter(e=>!e.discriminating),u=new Set(t.retire.map(e=>e.evalId));this.log(`
 Would retire ${t.retire.length} non-discriminating case(s):`);for(let e of C)u.has(e.evalId)&&this.log(`  \u2717 ${e.name} \u2014 ${c(e.withPassRate)} with vs ${c(e.withoutPassRate)} without (\u0394${(e.separation*100).toFixed(0)}pp) \u2014 ${e.note}`);for(let e of i.unpairedCases)u.has(e.evalId)&&this.log(`  \u2717 ${e.name} \u2014 ${c(e.withPassRate)} with vs ${c(e.withoutPassRate)} without \u2014 unpaired`);if(t.replacementsRequired&&this.log(`
 ${t.summary}. Replacements are generated and created BEFORE anything is retired \u2014 a target left under the floor drops out of the nightly evolution rotation with no error anywhere.`),!l.yes){this.log(`
 Preview only. Re-run with --yes to retire ${t.retire.length} case(s)${l["no-generate"]?"":" and generate replacements"}.`);return}let r=await $(t,{generate:a(e=>l["no-generate"]?Promise.resolve([]):v({content:o.content,name:o.name,feedback:e}),"generate"),createCase:a(e=>p(`/api/eval-targets/${n}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations}),"createCase"),retireCase:a(async(e,R)=>{await h(`/api/eval-targets/${n}/evals/${e}`,{retire:!0,retiredReason:R})},"retireCase"),log:a(e=>this.log(e),"log")});if(r.skipped==="no-llm-backend"){this.warn(r.summary+" Install the `claude` CLI (or set ANTHROPIC_API_KEY), then run `evals refine-cases` again.");return}if(r.skipped==="no-replacements-generated"){this.warn(r.summary);return}this.log(`

package/dist/commands/hooks/run.js

@@ -1,3 +1,3 @@
-import{a as p}from"../../chunk-TQFHRD5B.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5LTJNYCP.js";import"../../chunk-3EZZGA72.js";import"../../chunk-XAGNGQL3.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-5VWVQCNR.js";import"../../chunk-MHLHLJ5T.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as c}from"../../chunk-ZLLER3HP.js";import{Args as m}from"@oclif/core";import{execFileSync as f}from"node:child_process";import{existsSync as g}from"node:fs";import{join as h,delimiter as l}from"node:path";var v=["session-start","session-end","user-prompt","post-tool-use","assistant-response"],a=class d extends p{static{c(this,"HooksRun")}static description="Execute a hook script (called by AI tool hooks, not for direct use)";static hidden=!0;static args={hook:m.string({description:"Hook name to run",required:!0,options:[...v]})};async run(){let{args:e}=await this.parse(d),t=e.hook,{scriptsDir:o}=await import("@rulemetric/hooks/scripts-dir"),i=h(o,`${t}.sh`);g(i)||this.error(`Hook script not found: ${i}`);try{let s=await this.readStdin();try{let{evaluateHookCaptureGate:r}=await import("../../lib/hooks-capture-gate.js"),{getDefaultConfigDir:u}=await import("../../lib/auth.js");if(!(await r({stdin:s,configDir:u()})).allowed)return}catch{}let n=`${process.env.PATH??""}${l}${o}`;if(f("bash",[i],{input:s,stdio:["pipe","inherit","inherit"],timeout:t==="session-end"?3e4:1e4,env:{...process.env,PATH:n}}),process.env.RULEMETRIC_HOOK_TOOL==="antigravity"&&t==="post-tool-use")try{let{tailAntigravityTranscript:r}=await import("../../lib/antigravity-transcript.js");await r(s)}catch(r){process.stderr.write(`[antigravity-tailer] ${r.message}
-`)}}catch(s){let n=s.status;process.env.RULEMETRIC_HOOK_DEBUG&&process.stderr.write(`[rulemetric hooks run ${t}] non-fatal exit ${n??"?"}: ${s.message??""}
-`)}}readStdin(){return new Promise(e=>{let t=[];process.stdin.on("data",o=>t.push(o)),process.stdin.on("end",()=>e(Buffer.concat(t).toString("utf-8"))),process.stdin.readableEnded&&e(Buffer.concat(t).toString("utf-8"))})}};export{a as default};
+import{a as m,b as f,d as g}from"../../chunk-ZFEDOPRU.js";import{a as d}from"../../chunk-TQFHRD5B.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5LTJNYCP.js";import"../../chunk-3EZZGA72.js";import"../../chunk-XAGNGQL3.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-5VWVQCNR.js";import"../../chunk-MHLHLJ5T.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as v}from"@oclif/core";import{execFileSync as S}from"node:child_process";import{existsSync as y}from"node:fs";import{join as k,delimiter as E}from"node:path";var O=["session-start","session-end","user-prompt","post-tool-use","assistant-response"],p=class h extends d{static{u(this,"HooksRun")}static description="Execute a hook script (called by AI tool hooks, not for direct use)";static hidden=!0;static args={hook:v.string({description:"Hook name to run",required:!0,options:[...O]})};async run(){let{args:o}=await this.parse(h),s=o.hook,{scriptsDir:r}=await import("@rulemetric/hooks/scripts-dir"),n=k(r,`${s}.sh`);y(n)||this.error(`Hook script not found: ${n}`);try{let e=await this.readStdin();try{let{evaluateHookCaptureGate:t}=await import("../../lib/hooks-capture-gate.js"),{getDefaultConfigDir:w}=await import("../../lib/auth.js");if(!(await t({stdin:e,configDir:w()})).allowed)return}catch{}let a=`${process.env.PATH??""}${E}${r}`,l=s==="session-end"?3e4:1e4,i=s==="session-start"?m(e):null,c="";try{c=S("bash",[n],{input:e,stdio:["pipe",i?"pipe":"inherit","inherit"],timeout:l,env:{...process.env,PATH:a}})?.toString()??""}catch(t){if(!i)throw t;c=t.stdout?.toString()??""}if(i){let t=await g(i.cwd);process.stdout.write(f(c,t?.message??null)),t?.emitted()}if(process.env.RULEMETRIC_HOOK_TOOL==="antigravity"&&s==="post-tool-use")try{let{tailAntigravityTranscript:t}=await import("../../lib/antigravity-transcript.js");await t(e)}catch(t){process.stderr.write(`[antigravity-tailer] ${t.message}
+`)}}catch(e){let a=e.status;process.env.RULEMETRIC_HOOK_DEBUG&&process.stderr.write(`[rulemetric hooks run ${s}] non-fatal exit ${a??"?"}: ${e.message??""}
+`)}}readStdin(){return new Promise(o=>{let s=[];process.stdin.on("data",r=>s.push(r)),process.stdin.on("end",()=>o(Buffer.concat(s).toString("utf-8"))),process.stdin.readableEnded&&o(Buffer.concat(s).toString("utf-8"))})}};export{p as default};

package/dist/commands/launcher.js

@@ -1 +1 @@
-import{a as o}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as e}from"../chunk-ZLLER3HP.js";import{Flags as s}from"@oclif/core";var a=class r extends o{static{e(this,"Launcher")}static description="Open the launcher overlay (alias for `dashboard --launcher`)";static examples=["<%= config.bin %> launcher","<%= config.bin %> launcher --period today"];static flags={period:s.string({char:"p",default:"week",options:["today","week","30days","month","all"],description:"Initial time period for the dashboard panels behind the overlay"})};async run(){let{flags:i}=await this.parse(r),{renderDashboard:t}=await import("../dashboard/Dashboard.js");await t(i.period,{openLauncher:!0})}};export{a as default};
+import{a as r}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as i}from"../chunk-ZLLER3HP.js";var e=class extends r{static{i(this,"Launcher")}static hidden=!0;static description="Manual session launching is currently unavailable";async run(){this.error("Manual session launching has been shelved. Start your coding tool directly, or use `rulemetric dashboard` to view activity.")}};export{e as default};

package/dist/commands/setup.js

@@ -1,3 +1,3 @@
-import{c as l,d as h,f as m,g as p,h as f,i as y,j as r}from"../chunk-YLEXR62G.js";import"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-GPBN4ROE.js";import"../chunk-6XTLP2G7.js";import"../chunk-PJ3WHQ67.js";import{a as g}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as u}from"../chunk-ZLLER3HP.js";import{Args as b,Flags as c}from"@oclif/core";import*as w from"node:readline/promises";var d=class v extends g{static{u(this,"Setup")}static description="Agentic, transparent setup: plan / show diffs / apply the four core onboarding steps";static examples=["<%= config.bin %> setup                 # interactive walk (TTY)","<%= config.bin %> setup plan            # read-only: every remaining artifact + diff","<%= config.bin %> setup hooks --show    # print one step's diff","<%= config.bin %> setup hooks --apply   # apply one step","<%= config.bin %> setup --json          # machine-readable status (exit 0 iff complete)"];static args={target:b.string({description:`"plan" or a step id (${l.join(", ")})`,required:!1})};static flags={json:c.boolean({description:"Machine-readable output",default:!1}),show:c.boolean({description:"Print the step diff (read-only)",default:!1}),apply:c.boolean({description:"Apply the step",default:!1})};async run(){let{args:t,flags:e}=await this.parse(v),i=h();if(t.target==="plan"){await this.runPlan(i,e.json);return}if(t.target&&y(t.target)){if(e.apply){await f(t.target).apply(i);return}let n=await f(t.target).plan(i);this.printPlan(n);return}if(t.target&&this.error(`Unknown target "${t.target}". Use "plan" or one of: ${l.join(", ")}.`),e.json){let n=await r(i);this.log(JSON.stringify(n,null,2)),n.coreComplete||this.exit(1);return}await this.runInteractive(i)}async runPlan(t,e){let i=await r(t),n=p.filter(s=>!i.steps.find(a=>a.id===s.id)?.done),o=[];for(let s of n)o.push(await s.plan(t));if(e){this.log(JSON.stringify({...i,plans:o},null,2));return}if(o.length===0){this.log("[OK] All core steps complete \u2014 nothing to install.");return}this.log(`The following steps would be applied:
+import{c as l,d as h,f as m,g as p,h as f,i as y,j as r}from"../chunk-LSZ32DUR.js";import"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-GPBN4ROE.js";import"../chunk-6XTLP2G7.js";import"../chunk-PJ3WHQ67.js";import{a as g}from"../chunk-TQFHRD5B.js";import"../chunk-IZTIYR5Q.js";import"../chunk-5LTJNYCP.js";import"../chunk-3EZZGA72.js";import"../chunk-XAGNGQL3.js";import"../chunk-3AV3GTMP.js";import"../chunk-5VWVQCNR.js";import"../chunk-MHLHLJ5T.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as u}from"../chunk-ZLLER3HP.js";import{Args as b,Flags as c}from"@oclif/core";import*as w from"node:readline/promises";var d=class v extends g{static{u(this,"Setup")}static description="Agentic, transparent setup: plan / show diffs / apply the four core onboarding steps";static examples=["<%= config.bin %> setup                 # interactive walk (TTY)","<%= config.bin %> setup plan            # read-only: every remaining artifact + diff","<%= config.bin %> setup hooks --show    # print one step's diff","<%= config.bin %> setup hooks --apply   # apply one step","<%= config.bin %> setup --json          # machine-readable status (exit 0 iff complete)"];static args={target:b.string({description:`"plan" or a step id (${l.join(", ")})`,required:!1})};static flags={json:c.boolean({description:"Machine-readable output",default:!1}),show:c.boolean({description:"Print the step diff (read-only)",default:!1}),apply:c.boolean({description:"Apply the step",default:!1})};async run(){let{args:t,flags:e}=await this.parse(v),i=h();if(t.target==="plan"){await this.runPlan(i,e.json);return}if(t.target&&y(t.target)){if(e.apply){await f(t.target).apply(i);return}let n=await f(t.target).plan(i);this.printPlan(n);return}if(t.target&&this.error(`Unknown target "${t.target}". Use "plan" or one of: ${l.join(", ")}.`),e.json){let n=await r(i);this.log(JSON.stringify(n,null,2)),n.coreComplete||this.exit(1);return}await this.runInteractive(i)}async runPlan(t,e){let i=await r(t),n=p.filter(s=>!i.steps.find(a=>a.id===s.id)?.done),o=[];for(let s of n)o.push(await s.plan(t));if(e){this.log(JSON.stringify({...i,plans:o},null,2));return}if(o.length===0){this.log("[OK] All core steps complete \u2014 nothing to install.");return}this.log(`The following steps would be applied:
 `);for(let s of o)this.printPlan(s),this.log("")}printPlan(t){this.log(`[..] ${t.title} (${t.id})`),t.command&&this.log(`     command: ${t.command}`);for(let e of t.artifacts){this.log(`     ${e.kind}: ${e.path}${e.masked?"  (secret masked)":""}`);let i=m(e.before,e.after);for(let n of i.split(`
 `))this.log(`       ${n}`)}}async runInteractive(t){if(!process.stdin.isTTY){this.log("Not a TTY \u2014 interactive walk needs a terminal."),this.log("Use one of:"),this.log("  rulemetric setup plan          \u2014 see every remaining artifact + diff"),this.log("  rulemetric setup <step> --show \u2014 print one step diff"),this.log("  rulemetric setup <step> --apply \u2014 apply one step"),this.log("  rulemetric setup --json        \u2014 machine-readable status");return}let e=await r(t),i=p.filter(o=>!e.steps.find(s=>s.id===o.id)?.done);if(i.length===0){this.log("[OK] All core steps complete.");return}let n=w.createInterface({input:process.stdin,output:process.stdout});try{for(let o of i){let s=await o.plan(t);this.printPlan(s);let a=(await n.question(`Apply "${s.title}"? [y/N] `)).trim().toLowerCase();a==="y"||a==="yes"?(await o.apply(t),this.log(`[OK] Applied ${s.title}.`)):this.log(`[..] Skipped ${s.title}.`),this.log("")}}finally{n.close()}}};export{d as default};