@rulemetric/cli 0.16.0 → 0.17.1

package/dist/chunk-XFOG2PDJ.js

@@ -0,0 +1 @@
+import{a as r}from"./chunk-ZLLER3HP.js";import{createHash as I,randomUUID as T}from"node:crypto";import{closeSync as w,fsyncSync as _,linkSync as b,openSync as L,existsSync as u,mkdirSync as a,readdirSync as x,readFileSync as d,renameSync as P,rmSync as R,statSync as N,writeFileSync as C}from"node:fs";import{homedir as k}from"node:os";import{join as s}from"node:path";var E=1,z=["session","reimport","verify","settle"];function O(){let e=process.env.RULEMETRIC_CONFIG_DIR??s(k(),".config","rulemetric");return s(e,"scoped-outcomes")}r(O,"scopedOutcomeStoreRoot");var A=r(e=>I("sha256").update(e).digest("hex"),"sha256");function v(e,t){if(!/^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$/.test(e))throw new Error(`Unsafe ${t} for the outcome store: ${JSON.stringify(e)}`);return e}r(v,"safeSegment");function g(e,t,o=!1){let n=`${e}.tmp-${T()}`,c=L(n,"wx",384);try{C(c,t),_(c)}finally{w(c)}try{return o?b(n,e):P(n,e),!0}catch(i){if(o&&i.code==="EEXIST")return!1;throw i}finally{R(n,{force:!0})}}r(g,"writeAtomic");function p(e){return u(e)?JSON.parse(d(e,"utf8")):null}r(p,"readJson");function f(e){return s(O(),v(e,"confirmation id"))}r(f,"confirmationDir");function M(e,t){return s(f(e),v(t,"session id"))}r(M,"outcomeDir");var h=r(e=>s(f(e),"protocol.json"),"protocolPath"),S=r(e=>s(e,"transcript.jsonl"),"transcriptPath"),l=r(e=>s(e,"receipt.json"),"receiptPath"),D=r(e=>s(e,"delivery.json"),"ledgerPath");function V(e,t){let o=h(e);return a(f(e),{recursive:!0,mode:448}),g(o,JSON.stringify(t)),o}r(V,"stageProtocol");function Z(e){return p(h(e))}r(Z,"readStagedProtocol");function G(e,t){a(e,{recursive:!0,mode:448});let o=S(e);u(o)||g(o,d(t),!0);let n=d(o);return{path:o,sha256:A(n),bytes:n.length}}r(G,"preserveTranscript");function H(e,t){let o=p(l(e));if(o)return{record:o,created:!1};a(e,{recursive:!0,mode:448});let n={version:E,...t},c=g(l(e),JSON.stringify(n,null,2),!0);return{record:c?n:p(l(e)),created:c}}r(H,"writeReceiptOnce");function J(e){return p(l(e))}r(J,"readReceipt");function U(e){return p(D(e))??{version:E,sessionId:null,completed:[],attempts:0,lastError:null,completedAt:null}}r(U,"readLedger");function X(e,t){return a(e,{recursive:!0,mode:448}),g(D(e),JSON.stringify(t,null,2)),t}r(X,"writeLedger");function Y(e){return d(S(e))}r(Y,"transcriptBytes");function q(e){try{return N(e).size}catch{return null}}r(q,"transcriptSize");function K(){let e=O();if(!u(e))return[];let t=[];for(let o of x(e,{withFileTypes:!0}).filter(n=>n.isDirectory()).map(n=>n.name)){let n=s(e,o);for(let c of x(n,{withFileTypes:!0}).filter(i=>i.isDirectory()).map(i=>i.name)){let i=s(n,c),m=null,y=null;try{m=J(i),y=U(i)}catch{}t.push({dir:i,confirmationId:o,externalSessionId:c,receipt:m,ledger:y,hasTranscript:u(S(i))})}}return t}r(K,"listScopedOutcomeEntries");function Q(e){return!!e.ledger?.completedAt}r(Q,"isDelivered");function W(e){R(e,{recursive:!0,force:!0})}r(W,"removeScopedOutcomeEntry");export{E as a,z as b,O as c,f as d,M as e,h as f,S as g,l as h,D as i,V as j,Z as k,G as l,H as m,J as n,U as o,X as p,Y as q,q as r,K as s,Q as t,W as u};

package/dist/{chunk-KGU2IW54.js → chunk-YSY25OGG.js}

@@ -1 +1 @@
-import{a as _e}from"./chunk-M7QSC64D.js";import{i as Me}from"./chunk-ZKPURXUS.js";import{a as Pe}from"./chunk-ISEUEYSC.js";import{b as pe,i as me,j as ge,k as fe,l as he,m as be,n as $e}from"./chunk-GYKYJJNF.js";import{c as Re}from"./chunk-G7V4ZZZ3.js";import{a as xe,b as Se,c as De,d as Ie,e as se,f as Oe}from"./chunk-LJE5J3EW.js";import{a as re}from"./chunk-AHLVY52I.js";import{d as ke,f as Ce,g as Ee,i as Ae}from"./chunk-YXMAGC53.js";import{a as we}from"./chunk-T7HZZ4DJ.js";import{c as ye}from"./chunk-FADINV4H.js";import{b as ve}from"./chunk-SWSOEYC7.js";import{d as ue,e as ne,f as de}from"./chunk-KFNFLMUR.js";import{e as y,f as te}from"./chunk-GIXHG5FK.js";import{a as g}from"./chunk-ZLLER3HP.js";import{createHash as Ge}from"node:crypto";import{existsSync as Te,readFileSync as Ne,writeFileSync as Xe}from"node:fs";var H=g(t=>Ge("sha256").update(t).digest("hex"),"sha256"),Le=1.3,ae=4,ze=20;function Be(t,o){if(!t.configured||o<=0)return ae;let e=Math.ceil(t.minPairs*Le/o);return Math.min(ze,Math.max(ae,e))}g(Be,"passCapFor");function Qe(t,o=10,e=ae){if(t<=0)return{passes:0,expectedPairs:0,feasible:!1,reason:"no live cases"};let i=Math.ceil(o*Le),l=Math.min(e,Math.ceil(i/t)),r=l*t;return r<o?{passes:l,expectedPairs:r,feasible:!1,reason:`${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-CEM7JMUW.js";import{a as Pe}from"./chunk-L3AXNCWC.js";import{a as xe,b as Se,c as De,d as Ie,e as se,f as Oe}from"./chunk-ANHFE6US.js";import{b as pe,i as me,j as ge,k as fe,l as he,m as be,n as $e}from"./chunk-GYKYJJNF.js";import{c as Re}from"./chunk-G7V4ZZZ3.js";import{a as re}from"./chunk-AHLVY52I.js";import{d as ke,f as Ce,g as Ee,i as Ae}from"./chunk-YX2IGS72.js";import{a as we}from"./chunk-55HFAX5Y.js";import{c as ye}from"./chunk-AXTEV56C.js";import{b as ve}from"./chunk-VZYNZS5U.js";import{d as ue,e as ne,f as de}from"./chunk-KFNFLMUR.js";import{e as y,f as te}from"./chunk-Z7UN4R7T.js";import{a as g}from"./chunk-ZLLER3HP.js";import{createHash as Ge}from"node:crypto";import{existsSync as Te,readFileSync as Ne,writeFileSync as Xe}from"node:fs";var H=g(t=>Ge("sha256").update(t).digest("hex"),"sha256"),Le=1.3,ae=4,ze=20;function Be(t,o){if(!t.configured||o<=0)return ae;let e=Math.ceil(t.minPairs*Le/o);return Math.min(ze,Math.max(ae,e))}g(Be,"passCapFor");function Qe(t,o=10,e=ae){if(t<=0)return{passes:0,expectedPairs:0,feasible:!1,reason:"no live cases"};let i=Math.ceil(o*Le),l=Math.min(e,Math.ceil(i/t)),r=l*t;return r<o?{passes:l,expectedPairs:r,feasible:!1,reason:`${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-WL34G6IR.js → chunk-YVEI7XR6.js}

@@ -1,4 +1,4 @@
-import{b as $}from"./chunk-HBDKOU2N.js";import{a as C}from"./chunk-PDXHRF4E.js";import{a as y}from"./chunk-FG343JM3.js";import{f}from"./chunk-4GOXEEXZ.js";import{e as h,f as w}from"./chunk-GIXHG5FK.js";import{a as v}from"./chunk-ZLLER3HP.js";async function N(n,i,p){let{conversationId:a,messageId:g,messageContent:r}=n,o=await h(`/api/eval-targets/${i}`),c=await h(`/api/eval-targets/${i}/evals`),e=await h(`/api/eval-conversations/${a}/messages`),t=e.findIndex(l=>l.id===g);if(t>=0&&e.slice(t+1).some(u=>u.role==="assistant"&&(u.type==="results"||u.type==="proposal"||u.type==="generated"||u.type==="text"))){p(`Message ${g} already processed \u2014 skipping (idempotency guard)`);return}let s=v(async(l,u,m)=>{await w(`/api/eval-conversations/${a}/messages/assistant`,{content:l,type:u,data:m})},"postMessage"),d=await _(o,c,e,r,p);p(`Action: ${d.action}`);try{switch(d.action){case"run_evals":await x(o,c,a,s,p);break;case"generate_cases":await k(o,i,a,s,p);break;case"analyze_results":case"propose_changes":await E(o,c,e,r,a,s,p);break;case"respond":await s(d.response,"text");break}}catch(l){let u=l instanceof Error?l.message:String(l);throw await s(`Something went wrong: ${u}`,"text").catch(()=>{}),l}}v(N,"processConversationJob");async function _(n,i,p,a,g){let r=p.slice(-10).map(e=>`${e.role}: ${e.content.slice(0,200)}`).join(`
+import{b as $}from"./chunk-AZB34RT6.js";import{a as C}from"./chunk-TL57YUJV.js";import{a as y}from"./chunk-U6AASXY3.js";import{g as f}from"./chunk-LARY6ZFT.js";import{e as h,f as w}from"./chunk-Z7UN4R7T.js";import{a as v}from"./chunk-ZLLER3HP.js";async function N(n,i,p){let{conversationId:a,messageId:g,messageContent:r}=n,o=await h(`/api/eval-targets/${i}`),c=await h(`/api/eval-targets/${i}/evals`),e=await h(`/api/eval-conversations/${a}/messages`),t=e.findIndex(l=>l.id===g);if(t>=0&&e.slice(t+1).some(u=>u.role==="assistant"&&(u.type==="results"||u.type==="proposal"||u.type==="generated"||u.type==="text"))){p(`Message ${g} already processed \u2014 skipping (idempotency guard)`);return}let s=v(async(l,u,m)=>{await w(`/api/eval-conversations/${a}/messages/assistant`,{content:l,type:u,data:m})},"postMessage"),d=await _(o,c,e,r,p);p(`Action: ${d.action}`);try{switch(d.action){case"run_evals":await x(o,c,a,s,p);break;case"generate_cases":await k(o,i,a,s,p);break;case"analyze_results":case"propose_changes":await E(o,c,e,r,a,s,p);break;case"respond":await s(d.response,"text");break}}catch(l){let u=l instanceof Error?l.message:String(l);throw await s(`Something went wrong: ${u}`,"text").catch(()=>{}),l}}v(N,"processConversationJob");async function _(n,i,p,a,g){let r=p.slice(-10).map(e=>`${e.role}: ${e.content.slice(0,200)}`).join(`
 `),o=`You are an eval assistant. Decide what action to take based on the user's message.
 
 ## Current State

package/dist/{chunk-YXMAGC53.js → chunk-YX2IGS72.js}

@@ -1 +1 @@
-import{d as E}from"./chunk-FADINV4H.js";import{b as F,c as x}from"./chunk-KGURGIZ2.js";import{f as R}from"./chunk-GIXHG5FK.js";import{a}from"./chunk-ZLLER3HP.js";var _="with_candidate",w="with_prior_version";function S(e){return JSON.stringify([e.outputs.engine??null,e.outputs.model??null,e.outputs.methodology??null,e.outputs.deliveryRegime??null,e.grading.judge_engine??null,e.grading.judge_model??null])}a(S,"observedExecutionKey");function I(e){let r=e.filter(n=>n.status==="completed"),i=new Set(r.map(n=>JSON.stringify([n.outputs.engine??null,n.outputs.model??null,n.outputs.methodology??null,n.outputs.deliveryRegime??null]))),t=new Set(r.filter(n=>n.grading.judge_model!=null).map(n=>JSON.stringify([n.grading.judge_engine??"claude",n.grading.judge_model])));return i.size===1&&t.size<=1?JSON.stringify([[...i][0],[...t][0]??null]):null}a(I,"measurementBatchKey");function A(e){let r=e.filter(t=>t.configuration===_),i=[];for(let t of r){let n=e.find(s=>s.configuration===w&&s.evalId===t.evalId&&s.runNumber===t.runNumber);if(!n||S(n)!==S(t))continue;let l=h(t),m=h(n);l==null||m==null||i.push({withScore:l,withoutScore:m})}return i}a(A,"pairCandidateRuns");function h(e){return e.status!=="completed"?null:e.grading?.summary?.pass_rate??null}a(h,"evidenceOf");function $(e){let r=[],i=0;for(let t of e){if(t.status==="failed"){r.push(t.failure??{kind:"unknown",class:"infrastructure"});continue}h(t)==null&&(i+=1)}return{infrastructureFailed:r.length,ungradeable:i,summary:F(r)}}a($,"accountForFailures");function Q(e){let{pairs:r,minPairs:i,accounting:t}=e;return r>=i?{kind:"proceed"}:t.infrastructureFailed<=0?{kind:"proceed"}:{kind:"infrastructure-incomplete",reason:`${r}/${i} pair(s) \u2014 ${t.infrastructureFailed} run(s) never executed (${t.summary.describe()}). Retry when the environment recovers; this is not a statement about the target.`}}a(Q,"attributeShortfall");function k(e,r,i,t){let n=[];for(let l of e)n.push({eval_:l,configuration:_,targetContent:r,runNumber:t}),n.push({eval_:l,configuration:w,targetContent:i,runNumber:t});return n}a(k,"buildCandidateQueue");async function j(e){let{cases:r,candidateContent:i,priorVersionContent:t,concurrency:n,runNumber:l,opts:m,onLog:s=a(()=>{},"onLog")}=e,g=k(r,i,t,l),o=[],y=Math.max(1,n),p=null;for(let u=0;u<g.length;u+=y){let d=g.slice(u,u+y);o.push(...await Promise.all(d.map(c=>E(c,m,s))));let C=o.find(c=>c.failure&&x(c.failure.kind));if(C?.failure){p=C.failure.kind;let c=g.length-o.length;s(`  \u26D4 aborting batch \u2014 ${p}: the account is out of quota. Skipped ${c} remaining run(s) that would certainly have failed.`);break}}for(let u of o)try{await R("/api/eval-runs",u)}catch(d){s(`  \u26A0 failed to post a run result: ${d instanceof Error?d.message:String(d)}`)}let b=A(o),f=$(o),v=r.length-b.length;if(v>0){let u=f.infrastructureFailed?` \u2014 ${f.infrastructureFailed} run(s) never executed (${f.summary.describe()})`:"";s(`  \u26A0 ${v}/${r.length} case(s) produced no usable pair (failed or ungradeable arm)${u}`)}return{completed:o.filter(u=>u.status==="completed").length,failed:o.filter(u=>u.status==="failed").length,results:o,pairs:b,accounting:f,abortedForQuota:p}}a(j,"runCandidateBatchAndPost");export{_ as a,w as b,S as c,I as d,A as e,$ as f,Q as g,k as h,j as i};
+import{d as E}from"./chunk-AXTEV56C.js";import{b as F,c as x}from"./chunk-KGURGIZ2.js";import{f as R}from"./chunk-Z7UN4R7T.js";import{a}from"./chunk-ZLLER3HP.js";var _="with_candidate",w="with_prior_version";function S(e){return JSON.stringify([e.outputs.engine??null,e.outputs.model??null,e.outputs.methodology??null,e.outputs.deliveryRegime??null,e.grading.judge_engine??null,e.grading.judge_model??null])}a(S,"observedExecutionKey");function I(e){let r=e.filter(n=>n.status==="completed"),i=new Set(r.map(n=>JSON.stringify([n.outputs.engine??null,n.outputs.model??null,n.outputs.methodology??null,n.outputs.deliveryRegime??null]))),t=new Set(r.filter(n=>n.grading.judge_model!=null).map(n=>JSON.stringify([n.grading.judge_engine??"claude",n.grading.judge_model])));return i.size===1&&t.size<=1?JSON.stringify([[...i][0],[...t][0]??null]):null}a(I,"measurementBatchKey");function A(e){let r=e.filter(t=>t.configuration===_),i=[];for(let t of r){let n=e.find(s=>s.configuration===w&&s.evalId===t.evalId&&s.runNumber===t.runNumber);if(!n||S(n)!==S(t))continue;let l=h(t),m=h(n);l==null||m==null||i.push({withScore:l,withoutScore:m})}return i}a(A,"pairCandidateRuns");function h(e){return e.status!=="completed"?null:e.grading?.summary?.pass_rate??null}a(h,"evidenceOf");function $(e){let r=[],i=0;for(let t of e){if(t.status==="failed"){r.push(t.failure??{kind:"unknown",class:"infrastructure"});continue}h(t)==null&&(i+=1)}return{infrastructureFailed:r.length,ungradeable:i,summary:F(r)}}a($,"accountForFailures");function Q(e){let{pairs:r,minPairs:i,accounting:t}=e;return r>=i?{kind:"proceed"}:t.infrastructureFailed<=0?{kind:"proceed"}:{kind:"infrastructure-incomplete",reason:`${r}/${i} pair(s) \u2014 ${t.infrastructureFailed} run(s) never executed (${t.summary.describe()}). Retry when the environment recovers; this is not a statement about the target.`}}a(Q,"attributeShortfall");function k(e,r,i,t){let n=[];for(let l of e)n.push({eval_:l,configuration:_,targetContent:r,runNumber:t}),n.push({eval_:l,configuration:w,targetContent:i,runNumber:t});return n}a(k,"buildCandidateQueue");async function j(e){let{cases:r,candidateContent:i,priorVersionContent:t,concurrency:n,runNumber:l,opts:m,onLog:s=a(()=>{},"onLog")}=e,g=k(r,i,t,l),o=[],y=Math.max(1,n),p=null;for(let u=0;u<g.length;u+=y){let d=g.slice(u,u+y);o.push(...await Promise.all(d.map(c=>E(c,m,s))));let C=o.find(c=>c.failure&&x(c.failure.kind));if(C?.failure){p=C.failure.kind;let c=g.length-o.length;s(`  \u26D4 aborting batch \u2014 ${p}: the account is out of quota. Skipped ${c} remaining run(s) that would certainly have failed.`);break}}for(let u of o)try{await R("/api/eval-runs",u)}catch(d){s(`  \u26A0 failed to post a run result: ${d instanceof Error?d.message:String(d)}`)}let b=A(o),f=$(o),v=r.length-b.length;if(v>0){let u=f.infrastructureFailed?` \u2014 ${f.infrastructureFailed} run(s) never executed (${f.summary.describe()})`:"";s(`  \u26A0 ${v}/${r.length} case(s) produced no usable pair (failed or ungradeable arm)${u}`)}return{completed:o.filter(u=>u.status==="completed").length,failed:o.filter(u=>u.status==="failed").length,results:o,pairs:b,accounting:f,abortedForQuota:p}}a(j,"runCandidateBatchAndPost");export{_ as a,w as b,S as c,I as d,A as e,$ as f,Q as g,k as h,j as i};

package/dist/chunk-Z7UN4R7T.js

@@ -0,0 +1 @@
+import{l as A}from"./chunk-X6CF5KTW.js";import{a as T,e as w}from"./chunk-PZNWH4VQ.js";import{a as n}from"./chunk-ZLLER3HP.js";import{fileURLToPath as _}from"node:url";import{existsSync as v,readFileSync as O}from"node:fs";import{dirname as E,join as b}from"node:path";function k(){let t=E(_(import.meta.url));for(let e=0;e<8;e++){let s=b(t,"package.json");if(v(s))try{let i=JSON.parse(O(s,"utf8"));if(i.name==="@rulemetric/cli"&&i.version)return i.version}catch{}let o=E(t);if(o===t)break;t=o}return"0.0.0"}n(k,"readCliVersion");var x=k(),C="https://rulemetric.com",u=class extends Error{constructor(s,o){super(o);this.statusCode=s;this.name="ApiError"}static{n(this,"ApiError")}};function M(){let t=T(),e=process.env.RULEMETRIC_API_URL??t.apiUrl;if(!e){if(process.env.RULEMETRIC_STRICT_LOCAL==="1"||t.strictLocal==="1")throw new Error("RULEMETRIC_STRICT_LOCAL is set but no RULEMETRIC_API_URL is configured \u2014 refusing the cloud default. Set RULEMETRIC_API_URL in the environment or ~/.config/rulemetric/env.");e=C}let s=new URL(e);if(s.protocol==="http:"&&s.hostname!=="localhost"&&s.hostname!=="127.0.0.1")throw new Error(`API URL must use HTTPS for non-localhost hosts: ${e}`);return e}n(M,"getBaseUrl");async function c(t,e={}){let o=`${M()}${t}`,i={"Content-Type":"application/json","User-Agent":`rulemetric-cli/${x}`,...e.headers},m=T(),d=process.env.RULEMETRIC_API_KEY||m.apiKey||w()?.accessToken||m.accessToken;d&&(i.Authorization=`Bearer ${d}`);let h=new URL(o),f=h.hostname==="localhost"||h.hostname==="127.0.0.1",g=process.env.HTTPS_PROXY,P=process.env.HTTP_PROXY;f&&(delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY);let a=e.timeoutMs?new AbortController:void 0,R=!1,y=a?setTimeout(()=>{R=!0,a.abort()},e.timeoutMs):void 0,r;try{r=await fetch(o,{method:e.method??"GET",headers:i,body:e.body?JSON.stringify(e.body):void 0,signal:a&&e.signal?AbortSignal.any([a.signal,e.signal]):e.signal??a?.signal})}catch(p){if(R){let L=new URL(o).host,U=Math.round(e.timeoutMs/1e3);throw new u(408,`Request to ${L} timed out after ${U}s. This is usually transient or a local proxy/gateway issue \u2014 retry, and if it persists check HTTPS_PROXY / the RuleMetric gateway (:8787).`)}throw p}finally{y&&clearTimeout(y),f&&(g!==void 0&&(process.env.HTTPS_PROXY=g),P!==void 0&&(process.env.HTTP_PROXY=P))}let l;if(r.headers.get("content-type")?.includes("application/json")?l=await r.json():l=await r.text(),!r.ok){let{message:p}=A(r.status,l);throw new u(r.status,p)}return l}n(c,"api");var j=n((t,e)=>c(t,e),"apiGet"),F=n((t,e,s={})=>c(t,{...s,method:"POST",body:e}),"apiPost"),B=n((t,e)=>c(t,{method:"PUT",body:e}),"apiPut"),D=n((t,e,s={})=>c(t,{...s,method:"PATCH",body:e}),"apiPatch"),G=n(t=>c(t,{method:"DELETE"}),"apiDelete");export{C as a,u as b,M as c,c as d,j as e,F as f,B as g,D as h,G as i};

package/dist/{chunk-TQM7MF5C.js → chunk-ZZZ7ECPT.js}

@@ -1 +1 @@
-import{a as G,c as I}from"./chunk-FADINV4H.js";import{f as j}from"./chunk-SWSOEYC7.js";import{a as R}from"./chunk-FG343JM3.js";import{b as U}from"./chunk-KFNFLMUR.js";import{f as A}from"./chunk-4GOXEEXZ.js";import{b as P,e as C,h as _}from"./chunk-GIXHG5FK.js";import{a as x}from"./chunk-ZLLER3HP.js";var J=x(l=>l.summary.pass_rate===1?1:0,"fullPassScore");async function z(l,d={}){let $=d.executeEval??A,M=d.grade??R,q=d.getExperiment??(async o=>{try{return await C(`/api/memory-experiments/${o}`)}catch(e){if(e instanceof P&&e.statusCode===404)return null;throw e}}),B=d.getMemoryBody??(async o=>{try{let e=await C(`/api/memories/${o}`);return t?.memoryVersion!=null&&e.version!==t.memoryVersion?null:e.body}catch(e){if(e instanceof P&&e.statusCode===404)return null;throw e}}),p=d.updateExperiment??(async(o,e)=>{await _(`/api/memory-experiments/${o}`,e)}),t=await q(l.experimentId);if(!t)return{skipped:"not_found"};if(t.status==="completed"||t.status==="failed")return{skipped:"already_terminal",status:t.status};let E=t.memoryBody??await B(t.memoryId);if(E==null)return await p(t.id,{status:"failed",error:"memory no longer exists or its queued version is no longer available"}),{skipped:"memory_gone",status:"failed"};await p(t.id,{status:"running"});let v=t.expectations??[],c=t.model??void 0,i=t.engine??"claude",g=G({engine:i,model:c},{engine:l.graderEngine,model:l.graderModel}),f=x(async o=>{let e=(d.createWorkspace??j)({engine:i,targetType:"prompt",targetContent:o});try{if(e.leakedFiles.length)throw new Error("Contaminated memory trial workspace");return await $({prompt:t.taskPrompt,targetContent:e.systemPrompt,appendTarget:!0,ignoreContextFiles:!0,workDir:e.workDir,model:c,engine:i})}finally{e.cleanup()}},"executeArm"),s=[],u=new Set,b,k,m=[],w=0;try{if(l.measurement){let e=I(l);if(e.engine!==i||e.model!==c)throw new Error("Memory experiment differs from its queued measurement snapshot");g={engine:e.graderEngine,model:e.graderModel}}for(let e=0;e<t.trials;e++){let[r,n]=e%2===0?[await f(E),await f()]:await(async()=>{let y=await f();return[await f(E),y]})();if(r.error||n.error||r.exitCode!==0||n.exitCode!==0||!r.output.trim()||!n.output.trim()){w++,m.push({trial:e+1,reason:(r.error??n.error??"Nonzero exit or empty output").slice(0,500)});continue}if(r.engine!==i||n.engine!==i||!r.model||!n.model||r.model!==n.model||r.provider!==n.provider||r.model&&u.size>0&&!u.has(r.model)||i==="pi"&&c&&r.model!==`pi:${c}`){w++,m.push({trial:e+1,reason:"Harness, provider or model changed within the experiment"});continue}r.model&&u.add(r.model),n.model&&u.add(n.model);let[a,h]=await Promise.all([M({prompt:t.taskPrompt,output:r.output,expectations:v,...g}),M({prompt:t.taskPrompt,output:n.output,expectations:v,...g})]),T=JSON.stringify([a.judge_engine??null,a.judge_model??null]);if([a,h].some(y=>y.summary.pass_rate==null||y.summary.errored||y.summary.total!==v.length)||a.judge_model!==h.judge_model||a.judge_engine!==h.judge_engine||b!==void 0&&b!==T){w++,m.push({trial:e+1,reason:"Grading incomplete or judge changed; not a model failure"});continue}b=T,k={engine:a.judge_engine??null,model:a.judge_model??null},s.push({withScore:J(a),withoutScore:J(h),withLatency:r.durationMs/1e3,withoutLatency:n.durationMs/1e3,withTokens:r.totalTokens,withoutTokens:n.totalTokens,...r.costUsd!==void 0?{withCostUsd:r.costUsd}:{},...n.costUsd!==void 0?{withoutCostUsd:n.costUsd}:{}})}if(s.length===0){let e=`no trials completed (${w}/${t.trials} runs errored \u2014 is the ${i} CLI available?)`;return await p(t.id,{status:"failed",error:e,result:{pairs:s,excluded:m,completedTrials:0,engine:i}}),{status:"failed",requestedTrials:t.trials,completedTrials:0}}let o=U(s.map(e=>({...e,withCost:e.withCostUsd,withoutCost:e.withoutCostUsd})));return await p(t.id,{status:"completed",result:{verdict:o,pairs:s,excluded:m,judge:g,actualJudge:k,methodology:"v2-neutral-appended-memory",completedTrials:s.length,engine:i,model:[...u][0]}}),{status:"completed",requestedTrials:t.trials,completedTrials:s.length,significant:o.significant,delta:o.delta}}catch(o){let e=o instanceof Error?o.message:String(o);return await p(t.id,{status:"failed",error:e.slice(0,500),result:{pairs:s,excluded:m,completedTrials:s.length,engine:i,model:[...u][0]??null,judge:g,actualJudge:k,methodology:"v2-neutral-appended-memory"}}).catch(()=>{}),{status:"failed",requestedTrials:t.trials,completedTrials:s.length}}}x(z,"processMemoryAb");export{z as a};
+import{a as G,c as I}from"./chunk-AXTEV56C.js";import{f as j}from"./chunk-VZYNZS5U.js";import{a as R}from"./chunk-U6AASXY3.js";import{b as U}from"./chunk-KFNFLMUR.js";import{g as A}from"./chunk-LARY6ZFT.js";import{b as P,e as C,h as _}from"./chunk-Z7UN4R7T.js";import{a as x}from"./chunk-ZLLER3HP.js";var J=x(l=>l.summary.pass_rate===1?1:0,"fullPassScore");async function z(l,d={}){let $=d.executeEval??A,M=d.grade??R,q=d.getExperiment??(async o=>{try{return await C(`/api/memory-experiments/${o}`)}catch(e){if(e instanceof P&&e.statusCode===404)return null;throw e}}),B=d.getMemoryBody??(async o=>{try{let e=await C(`/api/memories/${o}`);return t?.memoryVersion!=null&&e.version!==t.memoryVersion?null:e.body}catch(e){if(e instanceof P&&e.statusCode===404)return null;throw e}}),p=d.updateExperiment??(async(o,e)=>{await _(`/api/memory-experiments/${o}`,e)}),t=await q(l.experimentId);if(!t)return{skipped:"not_found"};if(t.status==="completed"||t.status==="failed")return{skipped:"already_terminal",status:t.status};let E=t.memoryBody??await B(t.memoryId);if(E==null)return await p(t.id,{status:"failed",error:"memory no longer exists or its queued version is no longer available"}),{skipped:"memory_gone",status:"failed"};await p(t.id,{status:"running"});let v=t.expectations??[],c=t.model??void 0,i=t.engine??"claude",g=G({engine:i,model:c},{engine:l.graderEngine,model:l.graderModel}),f=x(async o=>{let e=(d.createWorkspace??j)({engine:i,targetType:"prompt",targetContent:o});try{if(e.leakedFiles.length)throw new Error("Contaminated memory trial workspace");return await $({prompt:t.taskPrompt,targetContent:e.systemPrompt,appendTarget:!0,ignoreContextFiles:!0,workDir:e.workDir,model:c,engine:i})}finally{e.cleanup()}},"executeArm"),s=[],u=new Set,b,k,m=[],w=0;try{if(l.measurement){let e=I(l);if(e.engine!==i||e.model!==c)throw new Error("Memory experiment differs from its queued measurement snapshot");g={engine:e.graderEngine,model:e.graderModel}}for(let e=0;e<t.trials;e++){let[r,n]=e%2===0?[await f(E),await f()]:await(async()=>{let y=await f();return[await f(E),y]})();if(r.error||n.error||r.exitCode!==0||n.exitCode!==0||!r.output.trim()||!n.output.trim()){w++,m.push({trial:e+1,reason:(r.error??n.error??"Nonzero exit or empty output").slice(0,500)});continue}if(r.engine!==i||n.engine!==i||!r.model||!n.model||r.model!==n.model||r.provider!==n.provider||r.model&&u.size>0&&!u.has(r.model)||i==="pi"&&c&&r.model!==`pi:${c}`){w++,m.push({trial:e+1,reason:"Harness, provider or model changed within the experiment"});continue}r.model&&u.add(r.model),n.model&&u.add(n.model);let[a,h]=await Promise.all([M({prompt:t.taskPrompt,output:r.output,expectations:v,...g}),M({prompt:t.taskPrompt,output:n.output,expectations:v,...g})]),T=JSON.stringify([a.judge_engine??null,a.judge_model??null]);if([a,h].some(y=>y.summary.pass_rate==null||y.summary.errored||y.summary.total!==v.length)||a.judge_model!==h.judge_model||a.judge_engine!==h.judge_engine||b!==void 0&&b!==T){w++,m.push({trial:e+1,reason:"Grading incomplete or judge changed; not a model failure"});continue}b=T,k={engine:a.judge_engine??null,model:a.judge_model??null},s.push({withScore:J(a),withoutScore:J(h),withLatency:r.durationMs/1e3,withoutLatency:n.durationMs/1e3,withTokens:r.totalTokens,withoutTokens:n.totalTokens,...r.costUsd!==void 0?{withCostUsd:r.costUsd}:{},...n.costUsd!==void 0?{withoutCostUsd:n.costUsd}:{}})}if(s.length===0){let e=`no trials completed (${w}/${t.trials} runs errored \u2014 is the ${i} CLI available?)`;return await p(t.id,{status:"failed",error:e,result:{pairs:s,excluded:m,completedTrials:0,engine:i}}),{status:"failed",requestedTrials:t.trials,completedTrials:0}}let o=U(s.map(e=>({...e,withCost:e.withCostUsd,withoutCost:e.withoutCostUsd})));return await p(t.id,{status:"completed",result:{verdict:o,pairs:s,excluded:m,judge:g,actualJudge:k,methodology:"v2-neutral-appended-memory",completedTrials:s.length,engine:i,model:[...u][0]}}),{status:"completed",requestedTrials:t.trials,completedTrials:s.length,significant:o.significant,delta:o.delta}}catch(o){let e=o instanceof Error?o.message:String(o);return await p(t.id,{status:"failed",error:e.slice(0,500),result:{pairs:s,excluded:m,completedTrials:s.length,engine:i,model:[...u][0]??null,judge:g,actualJudge:k,methodology:"v2-neutral-appended-memory"}}).catch(()=>{}),{status:"failed",requestedTrials:t.trials,completedTrials:s.length}}}x(z,"processMemoryAb");export{z as a};

package/dist/commands/antigravity/backfill.js

@@ -1,2 +1,2 @@
-import{a as v,b as u,e as y,g as d}from"../../chunk-BDQTVYYF.js";import{a as h}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as p}from"../../chunk-ZLLER3HP.js";import{Args as k,Flags as o}from"@oclif/core";import{existsSync as T}from"node:fs";import{homedir as P}from"node:os";import{join as I}from"node:path";var A=I(P(),".gemini","antigravity-ide","brain"),g=class x extends h{static{p(this,"AntigravityBackfill")}static description="Backfill historical events from an Antigravity conversation transcript into a RuleMetric session. Useful for seeding a conversation that existed before hooks were installed \u2014 the live hook intentionally only captures forward.";static examples=["<%= config.bin %> antigravity backfill 26fd215d-f43b-4ea5-b177-dc2c0e696f97 --session 158db9aa-68ee-43fc-961b-f3a180578694","<%= config.bin %> antigravity backfill 26fd215d-f43b-4ea5-b177-dc2c0e696f97 --dry-run"];static args={conversationId:k.string({description:"Antigravity conversationId (UUID) \u2014 the directory name under ~/.gemini/antigravity-ide/brain/",required:!0})};static flags={session:o.string({char:"s",description:"RuleMetric session UUID to POST events into. If omitted, resolved from the hook tmp file ($TMPDIR/rulemetric/<conversationId>) \u2014 fails if neither is present."}),transcript:o.string({description:"Path to the transcript.jsonl file. Defaults to ~/.gemini/antigravity-ide/brain/<conversationId>/.system_generated/logs/transcript.jsonl"}),"after-step":o.integer({description:"Only emit records with step_index > this value. Default: read from existing state file, else -1 (everything)."}),limit:o.integer({description:"Cap on the number of records to process this run."}),"dry-run":o.boolean({default:!1,description:"Classify records without POSTing. Reports counts only."}),"save-state":o.boolean({default:!0,allowNo:!0,description:"After backfill, persist the highest step_index so the live hook resumes from there."})};async run(){let{args:b,flags:e}=await this.parse(x),t=b.conversationId;d.loadEnvFile(),process.env.RULEMETRIC_API_URL||this.error("RULEMETRIC_API_URL is not set. Configure it via env or ~/.config/rulemetric/env.");let l=e.transcript??I(A,t,".system_generated","logs","transcript.jsonl");T(l)||this.error(`Transcript not found: ${l}`);let m=e.session??d.resolveRwSessionId(t);m||this.error(`No RuleMetric session resolved. Pass --session <uuid>, or run the Antigravity hook at least once so $TMPDIR/rulemetric/${t} is populated.`);let n=e["after-step"];n===void 0&&(n=d.load(t)?.lastStepIdx??-1);let i=v(l,n);e.limit!==void 0&&i.length>e.limit&&(i=i.slice(0,e.limit));let R=u(l);if(this.log(`Conversation ${t}: transcript max step = ${R}, after-step = ${n}, records to process = ${i.length}`),i.length===0){this.log("Nothing to backfill.");return}if(e["dry-run"]){let r=0,a=0,c={};for(let S of i){let f=(await import("../../lib/antigravity-transcript.js")).classifyRecord(S,0);f?(r++,c[f.eventType]=(c[f.eventType]??0)+1):a++}this.log(`Dry run: would emit ${r}, skip ${a}. By eventType: ${JSON.stringify(c)}`);return}let w=Date.now(),s=await y(i,m,t,n,{onProgress:p(({attempted:r,total:a})=>{(r%100===0||r===a)&&process.stderr.write(`  \u2026${r}/${a}
+import{a as v,b as u,e as y,g as d}from"../../chunk-UDNEOJWY.js";import{a as h}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as p}from"../../chunk-ZLLER3HP.js";import{Args as k,Flags as o}from"@oclif/core";import{existsSync as T}from"node:fs";import{homedir as P}from"node:os";import{join as I}from"node:path";var A=I(P(),".gemini","antigravity-ide","brain"),g=class x extends h{static{p(this,"AntigravityBackfill")}static description="Backfill historical events from an Antigravity conversation transcript into a RuleMetric session. Useful for seeding a conversation that existed before hooks were installed \u2014 the live hook intentionally only captures forward.";static examples=["<%= config.bin %> antigravity backfill 26fd215d-f43b-4ea5-b177-dc2c0e696f97 --session 158db9aa-68ee-43fc-961b-f3a180578694","<%= config.bin %> antigravity backfill 26fd215d-f43b-4ea5-b177-dc2c0e696f97 --dry-run"];static args={conversationId:k.string({description:"Antigravity conversationId (UUID) \u2014 the directory name under ~/.gemini/antigravity-ide/brain/",required:!0})};static flags={session:o.string({char:"s",description:"RuleMetric session UUID to POST events into. If omitted, resolved from the hook tmp file ($TMPDIR/rulemetric/<conversationId>) \u2014 fails if neither is present."}),transcript:o.string({description:"Path to the transcript.jsonl file. Defaults to ~/.gemini/antigravity-ide/brain/<conversationId>/.system_generated/logs/transcript.jsonl"}),"after-step":o.integer({description:"Only emit records with step_index > this value. Default: read from existing state file, else -1 (everything)."}),limit:o.integer({description:"Cap on the number of records to process this run."}),"dry-run":o.boolean({default:!1,description:"Classify records without POSTing. Reports counts only."}),"save-state":o.boolean({default:!0,allowNo:!0,description:"After backfill, persist the highest step_index so the live hook resumes from there."})};async run(){let{args:b,flags:e}=await this.parse(x),t=b.conversationId;d.loadEnvFile(),process.env.RULEMETRIC_API_URL||this.error("RULEMETRIC_API_URL is not set. Configure it via env or ~/.config/rulemetric/env.");let l=e.transcript??I(A,t,".system_generated","logs","transcript.jsonl");T(l)||this.error(`Transcript not found: ${l}`);let m=e.session??d.resolveRwSessionId(t);m||this.error(`No RuleMetric session resolved. Pass --session <uuid>, or run the Antigravity hook at least once so $TMPDIR/rulemetric/${t} is populated.`);let n=e["after-step"];n===void 0&&(n=d.load(t)?.lastStepIdx??-1);let i=v(l,n);e.limit!==void 0&&i.length>e.limit&&(i=i.slice(0,e.limit));let R=u(l);if(this.log(`Conversation ${t}: transcript max step = ${R}, after-step = ${n}, records to process = ${i.length}`),i.length===0){this.log("Nothing to backfill.");return}if(e["dry-run"]){let r=0,a=0,c={};for(let S of i){let f=(await import("../../lib/antigravity-transcript.js")).classifyRecord(S,0);f?(r++,c[f.eventType]=(c[f.eventType]??0)+1):a++}this.log(`Dry run: would emit ${r}, skip ${a}. By eventType: ${JSON.stringify(c)}`);return}let w=Date.now(),s=await y(i,m,t,n,{onProgress:p(({attempted:r,total:a})=>{(r%100===0||r===a)&&process.stderr.write(`  \u2026${r}/${a}
 `)},"onProgress")}),$=((Date.now()-w)/1e3).toFixed(1);this.log(`Backfill complete in ${$}s. attempted=${s.attempted} emitted=${s.emitted} skipped=${s.skipped} failed=${s.failed} highestStepIdx=${s.highestStepIdx}`),e["save-state"]?(d.save(t,{lastStepIdx:s.highestStepIdx}),this.log(`State saved \u2192 live hook will resume from step_index > ${s.highestStepIdx}.`)):this.log("State NOT saved (--no-save-state). Live hook will re-emit any new records from after-step.")}};export{g as default};

package/dist/commands/auth/create-key.js

@@ -1 +1 @@
-import{a as n}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as r}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{i as o,l as s}from"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import*as c from"node:path";import{Flags as l}from"@oclif/core";var t=class g extends n{static{a(this,"AuthCreateKey")}static description="Create a long-lived API key for CLI and proxy authentication";static examples=['<%= config.bin %> auth create-key --name "laptop"','<%= config.bin %> auth create-key --name "ci token" --no-save'];static flags={name:l.string({char:"n",description:"Name for this key",required:!0}),save:l.boolean({char:"s",description:"Save key to ~/.config/rulemetric/env for hooks/proxy",default:!0,allowNo:!0})};async run(){let{flags:i}=await this.parse(g),e=await r("/api/auth/api-keys",{name:i.name});this.log(""),this.log("API key created successfully:"),this.log(""),this.log(`  Name:   ${e.name}`),this.log(`  Key:    ${e.key}`),this.log(`  Prefix: ${e.keyPrefix}`),this.log(""),this.warn("Save this key now \u2014 it will not be shown again."),i.save&&(o(e.key,process.env.RULEMETRIC_API_URL),this.log(""),this.log(`Key saved to ${c.join(s(),"env")} (RULEMETRIC_API_KEY)`),this.log("Hooks and proxy will use this key automatically."))}};export{t as default};
+import{a as n}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{f as r}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{i as o,l as s}from"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import*as c from"node:path";import{Flags as l}from"@oclif/core";var t=class g extends n{static{a(this,"AuthCreateKey")}static description="Create a long-lived API key for CLI and proxy authentication";static examples=['<%= config.bin %> auth create-key --name "laptop"','<%= config.bin %> auth create-key --name "ci token" --no-save'];static flags={name:l.string({char:"n",description:"Name for this key",required:!0}),save:l.boolean({char:"s",description:"Save key to ~/.config/rulemetric/env for hooks/proxy",default:!0,allowNo:!0})};async run(){let{flags:i}=await this.parse(g),e=await r("/api/auth/api-keys",{name:i.name});this.log(""),this.log("API key created successfully:"),this.log(""),this.log(`  Name:   ${e.name}`),this.log(`  Key:    ${e.key}`),this.log(`  Prefix: ${e.keyPrefix}`),this.log(""),this.warn("Save this key now \u2014 it will not be shown again."),i.save&&(o(e.key,process.env.RULEMETRIC_API_URL),this.log(""),this.log(`Key saved to ${c.join(s(),"env")} (RULEMETRIC_API_KEY)`),this.log("Hooks and proxy will use this key automatically."))}};export{t as default};

package/dist/commands/auth/login.js

@@ -1,4 +1,4 @@
-import{a as f}from"../../chunk-QPVUGTDP.js";import{a as w}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import{c}from"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import{b as n}from"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{b as r,d as a}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{b as m,d,h as u,i as g}from"../../chunk-PZNWH4VQ.js";import{a as l}from"../../chunk-ZLLER3HP.js";import{Flags as p}from"@oclif/core";var h=class y extends w{static{l(this,"AuthLogin")}static description="Log in to RuleMetric";static examples=["<%= config.bin %> auth login","<%= config.bin %> auth login -e user@example.com"];static flags={email:p.string({char:"e",description:"Email address"}),password:p.string({char:"p",description:"Password (for non-interactive use)"}),"api-key":p.string({description:"Authenticate with an API key from https://rulemetric.com/connect (required for Google/GitHub sign-ins, which have no password)",exclusive:["password"]})};async run(){let{flags:s}=await this.parse(y);if(s["api-key"]){await this.loginWithApiKey(s["api-key"]);return}let e=s.email,t=s.password;if(!e||!t){process.stdin.isTTY||this.error("Missing credentials in a non-interactive shell. Pass --email/-e and --password/-p, or set RULEMETRIC_API_KEY (create a key at https://rulemetric.com/connect).");let{input:o,password:k}=await import("@inquirer/prompts");this.log("Signed up with Google or GitHub? You have no password \u2014 press Ctrl-C and run:"),this.log("  rulemetric auth login --api-key <key>   (create one at https://rulemetric.com/connect)"),this.log(""),e||(e=await o({message:"Email:"})),t||(t=await k({message:"Password:"}))}let i;try{i=await a("/auth/login",{method:"POST",body:{email:e,password:t},timeoutMs:2e4})}catch(o){throw o instanceof r&&o.statusCode===401&&this.error(`Invalid email or password.
+import{a as f}from"../../chunk-EOTSCWFS.js";import{a as w}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import{c}from"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import{b as n}from"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{b as r,d as a}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{b as m,d,h as u,i as g}from"../../chunk-PZNWH4VQ.js";import{a as l}from"../../chunk-ZLLER3HP.js";import{Flags as p}from"@oclif/core";var h=class y extends w{static{l(this,"AuthLogin")}static description="Log in to RuleMetric";static examples=["<%= config.bin %> auth login","<%= config.bin %> auth login -e user@example.com"];static flags={email:p.string({char:"e",description:"Email address"}),password:p.string({char:"p",description:"Password (for non-interactive use)"}),"api-key":p.string({description:"Authenticate with an API key from https://rulemetric.com/connect (required for Google/GitHub sign-ins, which have no password)",exclusive:["password"]})};async run(){let{flags:s}=await this.parse(y);if(s["api-key"]){await this.loginWithApiKey(s["api-key"]);return}let e=s.email,t=s.password;if(!e||!t){process.stdin.isTTY||this.error("Missing credentials in a non-interactive shell. Pass --email/-e and --password/-p, or set RULEMETRIC_API_KEY (create a key at https://rulemetric.com/connect).");let{input:o,password:k}=await import("@inquirer/prompts");this.log("Signed up with Google or GitHub? You have no password \u2014 press Ctrl-C and run:"),this.log("  rulemetric auth login --api-key <key>   (create one at https://rulemetric.com/connect)"),this.log(""),e||(e=await o({message:"Email:"})),t||(t=await k({message:"Password:"}))}let i;try{i=await a("/auth/login",{method:"POST",body:{email:e,password:t},timeoutMs:2e4})}catch(o){throw o instanceof r&&o.statusCode===401&&this.error(`Invalid email or password.
 
 If you signed up with Google or GitHub, you have no password \u2014 that is why this cannot work. Create an API key at https://rulemetric.com/connect, then run:
   rulemetric auth login --api-key <key>`),o}let v=i.expiresAt??new Date(Date.now()+3600*1e3).toISOString();d({accessToken:i.accessToken,refreshToken:i.refreshToken,expiresAt:v,email:i.email}),u(i.accessToken,process.env.RULEMETRIC_API_URL),await n(),await f(this.log.bind(this),this.warn.bind(this),this.config.version),c("auth_login",{},this.config.version),this.log(`Logged in as ${i.email}`)}async loginWithApiKey(s){let e=s.trim();e||this.error("Empty --api-key. Create one at https://rulemetric.com/connect"),process.env.RULEMETRIC_API_KEY=e;try{await a("/api/setup/status",{timeoutMs:2e4})}catch(t){throw delete process.env.RULEMETRIC_API_KEY,t instanceof r&&(t.statusCode===401||t.statusCode===403)&&this.error("That API key was rejected. Check you copied it whole (it is shown only once at creation) or mint a new one at https://rulemetric.com/connect"),t}g(e,process.env.RULEMETRIC_API_URL),m(),await n(),c("auth_login",{method:"api_key"},this.config.version),this.log(`Authenticated with API key ${e.slice(0,11)}...`),this.log("Verify any time with: rulemetric auth status")}};export{h as default};

package/dist/commands/auth/logout.js

@@ -1 +1 @@
-import{a as s}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{f as i,j as r}from"../../chunk-PZNWH4VQ.js";import{a as o}from"../../chunk-ZLLER3HP.js";var e=class t extends s{static{o(this,"AuthLogout")}static description="Log out of RuleMetric";static examples=["<%= config.bin %> auth logout"];async run(){await this.parse(t),i(),r(),this.log("Logged out successfully.")}};export{e as default};
+import{a as s}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{f as i,j as r}from"../../chunk-PZNWH4VQ.js";import{a as o}from"../../chunk-ZLLER3HP.js";var e=class t extends s{static{o(this,"AuthLogout")}static description="Log out of RuleMetric";static examples=["<%= config.bin %> auth logout"];async run(){await this.parse(t),i(),r(),this.log("Logged out successfully.")}};export{e as default};

package/dist/commands/auth/revoke-key.js

@@ -1,4 +1,4 @@
-import{a as k}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as n,i as c}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as s}from"../../chunk-ZLLER3HP.js";import{Args as y}from"@oclif/core";var a=class m extends k{static{s(this,"AuthRevokeKey")}static description="Revoke an API key (run with no argument to list active keys)";static examples=["<%= config.bin %> auth revoke-key","<%= config.bin %> auth revoke-key rm_abc12345",'<%= config.bin %> auth revoke-key "my-laptop (auto)"'];static args={id:y.string({description:"API key ID, prefix (rm_xxxxxxxx), or name"})};async run(){let{args:o}=await this.parse(m),r=(await n("/api/auth/api-keys")).filter(e=>!e.revokedAt);if(!o.id){if(r.length===0){this.log("No active API keys.");return}this.log(`Active API keys \u2014 pass an id, prefix, or name to revoke one:
+import{a as k}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as n,i as c}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as s}from"../../chunk-ZLLER3HP.js";import{Args as y}from"@oclif/core";var a=class m extends k{static{s(this,"AuthRevokeKey")}static description="Revoke an API key (run with no argument to list active keys)";static examples=["<%= config.bin %> auth revoke-key","<%= config.bin %> auth revoke-key rm_abc12345",'<%= config.bin %> auth revoke-key "my-laptop (auto)"'];static args={id:y.string({description:"API key ID, prefix (rm_xxxxxxxx), or name"})};async run(){let{args:o}=await this.parse(m),r=(await n("/api/auth/api-keys")).filter(e=>!e.revokedAt);if(!o.id){if(r.length===0){this.log("No active API keys.");return}this.log(`Active API keys \u2014 pass an id, prefix, or name to revoke one:
 `);for(let e of r)this.log(`  ${e.id}  ${e.keyPrefix}...  ${e.name}`);return}let i=o.id,t=r.filter(e=>e.id===i||e.name===i||i.startsWith("rm_")&&i.startsWith(e.keyPrefix));t.length===0&&this.error(`No active API key matches "${i}". Run \`rulemetric auth revoke-key\` with no argument to list keys.`),t.length>1&&this.error(`"${i}" matches ${t.length} keys \u2014 use the exact key ID:
 `+t.map(e=>`  ${e.id}  ${e.keyPrefix}...  ${e.name}`).join(`
 `)),await c(`/api/auth/api-keys/${t[0].id}`),this.log(`API key ${t[0].keyPrefix}... ("${t[0].name}") revoked.`)}};export{a as default};

package/dist/commands/auth/signup.js

@@ -1 +1 @@
-import{a as g}from"../../chunk-QPVUGTDP.js";import{a as h}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import{c as d}from"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import{b as l}from"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{b as m,d as p}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{d as n,h as c}from"../../chunk-PZNWH4VQ.js";import{a as o}from"../../chunk-ZLLER3HP.js";import{Flags as u}from"@oclif/core";var t=class f extends h{static{o(this,"AuthSignup")}static description="Create a RuleMetric account from the CLI";static examples=["<%= config.bin %> auth signup","<%= config.bin %> auth signup -e user@example.com"];static flags={email:u.string({char:"e",description:"Email address"}),password:u.string({char:"p",description:"Password (min 8 chars; for non-interactive use)"})};async run(){let{flags:a}=await this.parse(f),i=a.email,r=a.password;if(!i||!r){process.stdin.isTTY||this.error("Missing credentials in a non-interactive shell. Pass --email/-e and --password/-p.");let{input:s,password:v}=await import("@inquirer/prompts");i||(i=await s({message:"Email:"})),r||(r=await v({message:"Password (min 8 chars):"}))}let e;try{e=await p("/auth/signup",{method:"POST",body:{email:i,password:r},timeoutMs:2e4})}catch(s){throw s instanceof m&&s.statusCode===409&&this.error(`An account with ${i} already exists. Run: rulemetric auth login -e ${i}`),s}if(e.confirmationRequired||!e.accessToken||!e.refreshToken){this.log(`Account created for ${e.email}.`),this.log("Check your email and click the confirmation link, then run:"),this.log(`  rulemetric auth login -e ${e.email}`);return}let w=e.expiresAt??new Date(Date.now()+3600*1e3).toISOString();n({accessToken:e.accessToken,refreshToken:e.refreshToken,expiresAt:w,email:e.email}),c(e.accessToken,process.env.RULEMETRIC_API_URL),await l(),await g(this.log.bind(this),this.warn.bind(this),this.config.version),d("auth_signup",{},this.config.version),this.log(`Signed up and logged in as ${e.email}`)}};export{t as default};
+import{a as g}from"../../chunk-EOTSCWFS.js";import{a as h}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import{c as d}from"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import{b as l}from"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{b as m,d as p}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{d as n,h as c}from"../../chunk-PZNWH4VQ.js";import{a as o}from"../../chunk-ZLLER3HP.js";import{Flags as u}from"@oclif/core";var t=class f extends h{static{o(this,"AuthSignup")}static description="Create a RuleMetric account from the CLI";static examples=["<%= config.bin %> auth signup","<%= config.bin %> auth signup -e user@example.com"];static flags={email:u.string({char:"e",description:"Email address"}),password:u.string({char:"p",description:"Password (min 8 chars; for non-interactive use)"})};async run(){let{flags:a}=await this.parse(f),i=a.email,r=a.password;if(!i||!r){process.stdin.isTTY||this.error("Missing credentials in a non-interactive shell. Pass --email/-e and --password/-p.");let{input:s,password:v}=await import("@inquirer/prompts");i||(i=await s({message:"Email:"})),r||(r=await v({message:"Password (min 8 chars):"}))}let e;try{e=await p("/auth/signup",{method:"POST",body:{email:i,password:r},timeoutMs:2e4})}catch(s){throw s instanceof m&&s.statusCode===409&&this.error(`An account with ${i} already exists. Run: rulemetric auth login -e ${i}`),s}if(e.confirmationRequired||!e.accessToken||!e.refreshToken){this.log(`Account created for ${e.email}.`),this.log("Check your email and click the confirmation link, then run:"),this.log(`  rulemetric auth login -e ${e.email}`);return}let w=e.expiresAt??new Date(Date.now()+3600*1e3).toISOString();n({accessToken:e.accessToken,refreshToken:e.refreshToken,expiresAt:w,email:e.email}),c(e.accessToken,process.env.RULEMETRIC_API_URL),await l(),await g(this.log.bind(this),this.warn.bind(this),this.config.version),d("auth_signup",{},this.config.version),this.log(`Signed up and logged in as ${e.email}`)}};export{t as default};

package/dist/commands/auth/status.js

@@ -1 +1 @@
-import{a as c}from"../../chunk-W2MV25SB.js";import{a as g}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{a,e as l,l as t}from"../../chunk-PZNWH4VQ.js";import{a as r}from"../../chunk-ZLLER3HP.js";import*as u from"node:path";var o=class h extends g{static{r(this,"AuthStatus")}static description="Show current authentication status";static examples=["<%= config.bin %> auth status"];async run(){await this.parse(h);let e=l(),i=process.env.RULEMETRIC_API_KEY,s=a().apiKey;if(i||s){let m=i||s,p=i?"RULEMETRIC_API_KEY env var":u.join(t(),"env");this.log(`Authenticated via API key ${m.slice(0,11)}... (${p})`),e&&this.log(`Logged in as: ${e.email}`),this.log(`Config: ${t()}`);return}if(!e){this.log("Not logged in. Run `rulemetric auth login` to authenticate.");return}let n=new Date(e.expiresAt).getTime();if(Number.isNaN(n)||Date.now()>=n){this.log("Session expired. Run `rulemetric auth login` to re-authenticate.");return}this.log(`Logged in as: ${e.email}`),this.log(`Session expires: ${c(e.expiresAt)}`),this.log(`Config: ${t()}`)}};export{o as default};
+import{a as c}from"../../chunk-W2MV25SB.js";import{a as g}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{a,e as l,l as t}from"../../chunk-PZNWH4VQ.js";import{a as r}from"../../chunk-ZLLER3HP.js";import*as u from"node:path";var o=class h extends g{static{r(this,"AuthStatus")}static description="Show current authentication status";static examples=["<%= config.bin %> auth status"];async run(){await this.parse(h);let e=l(),i=process.env.RULEMETRIC_API_KEY,s=a().apiKey;if(i||s){let m=i||s,p=i?"RULEMETRIC_API_KEY env var":u.join(t(),"env");this.log(`Authenticated via API key ${m.slice(0,11)}... (${p})`),e&&this.log(`Logged in as: ${e.email}`),this.log(`Config: ${t()}`);return}if(!e){this.log("Not logged in. Run `rulemetric auth login` to authenticate.");return}let n=new Date(e.expiresAt).getTime();if(Number.isNaN(n)||Date.now()>=n){this.log("Session expired. Run `rulemetric auth login` to re-authenticate.");return}this.log(`Logged in as: ${e.email}`),this.log(`Session expires: ${c(e.expiresAt)}`),this.log(`Config: ${t()}`)}};export{o as default};

package/dist/commands/capture/scope.js

@@ -1,2 +1,2 @@
-import{a as f,b as k,c as S,d as x,e as R,f as n,g as y}from"../../chunk-VALRPZXH.js";import{b as C,c as w,g as v,h as b}from"../../chunk-QI3FJKGK.js";import{a as u}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{l as m}from"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{existsSync as O}from"node:fs";import{Flags as c}from"@oclif/core";var g=class E extends u{static{h(this,"CaptureScope")}static description="Show or set which projects RuleMetric captures \u2014 the Codex sweep and hook-driven sessions each have their own scope";static examples=["<%= config.bin %> <%= command.id %>","<%= config.bin %> <%= command.id %> --add ~/work/client-a --add ~/work/client-b","<%= config.bin %> <%= command.id %> --tool hooks --add ~/work/client-a","<%= config.bin %> <%= command.id %> --tool hooks --off","<%= config.bin %> <%= command.id %> --remove ~/work/client-a","<%= config.bin %> <%= command.id %> --off","<%= config.bin %> <%= command.id %> --all"];static flags={tool:c.string({description:'Which capture path to edit: "codex" (the worker transcript sweep, the default) or "hooks" (Claude Code / Cursor / Copilot / Antigravity hook sessions).',options:["codex","hooks"],default:"codex"}),add:c.string({description:"Capture sessions under this path (repeatable). Implies an allowlist.",multiple:!0}),remove:c.string({description:"Stop capturing under this path (repeatable).",multiple:!0}),off:c.boolean({description:"Turn this capture path off entirely.",exclusive:["all"]}),all:c.boolean({description:"Capture every project (the default).",exclusive:["off"]}),clear:c.boolean({description:"Remove this scope, restoring the default."})};async run(){let{flags:o}=await this.parse(E),s=m(),d=f(s);if(!((o.add?.length??0)>0||(o.remove?.length??0)>0||o.off||o.all||o.clear)){await this.show(s,d);return}for(let t of[...o.add??[],...o.remove??[]])R(t)||this.error(`Not an absolute path: ${t}
+import{a as f,b as k,c as S,d as x,e as R,f as n,g as y}from"../../chunk-VALRPZXH.js";import{b as C,c as w,g as v,h as b}from"../../chunk-QI3FJKGK.js";import{a as u}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{l as m}from"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{existsSync as O}from"node:fs";import{Flags as c}from"@oclif/core";var g=class E extends u{static{h(this,"CaptureScope")}static description="Show or set which projects RuleMetric captures \u2014 the Codex sweep and hook-driven sessions each have their own scope";static examples=["<%= config.bin %> <%= command.id %>","<%= config.bin %> <%= command.id %> --add ~/work/client-a --add ~/work/client-b","<%= config.bin %> <%= command.id %> --tool hooks --add ~/work/client-a","<%= config.bin %> <%= command.id %> --tool hooks --off","<%= config.bin %> <%= command.id %> --remove ~/work/client-a","<%= config.bin %> <%= command.id %> --off","<%= config.bin %> <%= command.id %> --all"];static flags={tool:c.string({description:'Which capture path to edit: "codex" (the worker transcript sweep, the default) or "hooks" (Claude Code / Cursor / Copilot / Antigravity hook sessions).',options:["codex","hooks"],default:"codex"}),add:c.string({description:"Capture sessions under this path (repeatable). Implies an allowlist.",multiple:!0}),remove:c.string({description:"Stop capturing under this path (repeatable).",multiple:!0}),off:c.boolean({description:"Turn this capture path off entirely.",exclusive:["all"]}),all:c.boolean({description:"Capture every project (the default).",exclusive:["off"]}),clear:c.boolean({description:"Remove this scope, restoring the default."})};async run(){let{flags:o}=await this.parse(E),s=m(),d=f(s);if(!((o.add?.length??0)>0||(o.remove?.length??0)>0||o.off||o.all||o.clear)){await this.show(s,d);return}for(let t of[...o.add??[],...o.remove??[]])R(t)||this.error(`Not an absolute path: ${t}
 Use an absolute path or ~/\u2026 \u2014 relative paths cannot be matched against a session's cwd.`);for(let t of o.add??[]){let a=x(t);O(a)||this.warn(`${a} does not exist yet \u2014 added anyway, but nothing will match until it does.`)}let e;try{e=k(s)}catch(t){o.clear||this.error(t instanceof Error?t.message:String(t));let a=y(s);this.warn(`${f(s)} was not valid JSON \u2014 saved it to ${a} and starting fresh.`),e={}}let i=o.tool??"codex",r=void 0,l=!1;if(o.clear&&({config:e}=n(e,{op:"clear"},r,i)),o.off&&({config:e}=n(e,{op:"mode",mode:"off"},r,i)),o.all&&({config:e}=n(e,{op:"mode",mode:"all"},r,i)),o.remove?.length){let t=n(e,{op:"remove",roots:o.remove},r,i);e=t.config,l=t.emptiedToOff}o.add?.length&&({config:e}=n(e,{op:"add",roots:o.add},r,i)),S(s,e),l&&(this.log(`Removed the last root \u2014 ${i==="hooks"?"hook":"Codex"} capture is now OFF rather than silently capturing everything.`),this.log("Use --all to go back to capturing every project.")),await this.show(s,d),this.log(""),this.log(i==="hooks"?"Takes effect on the next hook fire. No restart needed.":"Takes effect on the next sweep (within ~30s). No restart needed.")}async show(o,s){let d=await C({configDir:o}),p=await w({configDir:o}),e=h((i,r,l,t)=>{if(this.log(`${i}: ${l}`),r.roots.length)for(let a of r.roots)this.log(`  \u2022 ${a}`);r.error&&(this.log("  This scope is FAILING CLOSED \u2014 nothing is being captured on this path."),this.log(`  Fix the file, or run \`rulemetric capture scope ${i==="Hook capture"?"--tool hooks ":""}--clear\` to start over.`)),r.source==="env"&&this.log(`  Set by environment variable, which outranks the config file (${t}).`)},"print");e("Codex capture",d,v(d),"RULEMETRIC_CODEX_CAPTURE / RULEMETRIC_CODEX_ROOTS"),e("Hook capture",p,b(p),"RULEMETRIC_HOOKS_CAPTURE / RULEMETRIC_HOOKS_ROOTS"),this.log(`Config: ${s}`)}};export{g as default};

package/dist/commands/catalog/index.js

@@ -1,2 +1,2 @@
-import{a as g,b as d,c as p}from"../../chunk-W2MV25SB.js";import{a as c}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Flags as i}from"@oclif/core";var o=class m extends c{static{a(this,"Catalog")}static description="Browse the org instruction catalog";static flags={badge:i.string({description:"Filter by badge",options:["recommended","required"]}),scope:i.string({description:"Filter by scope",options:["org","team"]}),search:i.string({description:"Search by name or description"})};async run(){let{flags:e}=await this.parse(m);this.requireAuth();let s=new URLSearchParams;e.badge&&s.set("badge",e.badge),e.scope&&s.set("scope",e.scope),e.search&&s.set("search",e.search);let r=await n(`/api/catalog?${s.toString()}`);if(r.length===0){this.log("No instructions found in catalog.");return}for(let t of r){let l=t.badge?` [${t.badge}]`:"",f=t.tags?.length?` (${t.tags.join(", ")})`:"";this.log(`  ${d(t.name,40)} \u2014 ${p(t.scope)}${l} v${t.version}${f}  ${g(t.updatedAt)}`)}this.log(`
+import{a as g,b as d,c as p}from"../../chunk-W2MV25SB.js";import{a as c}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Flags as i}from"@oclif/core";var o=class m extends c{static{a(this,"Catalog")}static description="Browse the org instruction catalog";static flags={badge:i.string({description:"Filter by badge",options:["recommended","required"]}),scope:i.string({description:"Filter by scope",options:["org","team"]}),search:i.string({description:"Search by name or description"})};async run(){let{flags:e}=await this.parse(m);this.requireAuth();let s=new URLSearchParams;e.badge&&s.set("badge",e.badge),e.scope&&s.set("scope",e.scope),e.search&&s.set("search",e.search);let r=await n(`/api/catalog?${s.toString()}`);if(r.length===0){this.log("No instructions found in catalog.");return}for(let t of r){let l=t.badge?` [${t.badge}]`:"",f=t.tags?.length?` (${t.tags.join(", ")})`:"";this.log(`  ${d(t.name,40)} \u2014 ${p(t.scope)}${l} v${t.version}${f}  ${g(t.updatedAt)}`)}this.log(`
 ${r.length} instruction(s) found.`)}};export{o as default};

package/dist/commands/catalog/recommend.js

@@ -1,3 +1,3 @@
-import{a as i}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as n}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as t}from"../../chunk-ZLLER3HP.js";import{Args as c}from"@oclif/core";var o=class s extends i{static{t(this,"CatalogRecommend")}static description="Get AI-powered instruction recommendations for a project";static args={projectId:c.string({description:"ID of the project",required:!0})};async run(){let{args:a}=await this.parse(s);this.requireAuth(),this.log(`Analyzing project and generating recommendations...
+import{a as i}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{f as n}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as t}from"../../chunk-ZLLER3HP.js";import{Args as c}from"@oclif/core";var o=class s extends i{static{t(this,"CatalogRecommend")}static description="Get AI-powered instruction recommendations for a project";static args={projectId:c.string({description:"ID of the project",required:!0})};async run(){let{args:a}=await this.parse(s);this.requireAuth(),this.log(`Analyzing project and generating recommendations...
 `);let r=await n(`/api/projects/${a.projectId}/recommendations`,{});if(r.error&&this.warn(`Note: ${r.error}`),r.recommendations.length===0){this.log("No recommendations found.");return}for(let e of r.recommendations)this.log(`  ${(e.score*100).toFixed(0)}% \u2014 ${e.instructionId}`),this.log(`       ${e.reason}
 `)}};export{o as default};

package/dist/commands/crons/list.js

@@ -1 +1 @@
-import{a as o}from"../../chunk-4VN5PG2S.js";import{a as n}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as i}from"../../chunk-ZLLER3HP.js";import{printTable as t}from"@oclif/table";var a=class s extends n{static{i(this,"CronsList")}static description="List manually-runnable maintenance tasks";static examples=["<%= config.bin %> crons list"];async run(){await this.parse(s),t({data:o.map(e=>({Name:e.name,"Scheduled by":e.scheduledBy,Description:e.description})),columns:["Name","Scheduled by","Description"]}),this.log(""),this.log("Scheduled crons run in Supabase pg_cron \u2014 audit them with:"),this.log(`  psql "$DATABASE_URL" -c 'select jobname, schedule, active from cron.job;'`),this.log("Run any task above on demand with: rulemetric crons run <name>")}};export{a as default};
+import{a as o}from"../../chunk-4VN5PG2S.js";import{a as n}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as i}from"../../chunk-ZLLER3HP.js";import{printTable as t}from"@oclif/table";var a=class s extends n{static{i(this,"CronsList")}static description="List manually-runnable maintenance tasks";static examples=["<%= config.bin %> crons list"];async run(){await this.parse(s),t({data:o.map(e=>({Name:e.name,"Scheduled by":e.scheduledBy,Description:e.description})),columns:["Name","Scheduled by","Description"]}),this.log(""),this.log("Scheduled crons run in Supabase pg_cron \u2014 audit them with:"),this.log(`  psql "$DATABASE_URL" -c 'select jobname, schedule, active from cron.job;'`),this.log("Run any task above on demand with: rulemetric crons run <name>")}};export{a as default};

package/dist/commands/crons/run.js

@@ -1,2 +1,2 @@
-import{a as $}from"../../chunk-4VN5PG2S.js";import{a as k}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as b}from"../../chunk-GIXHG5FK.js";import{A as f,B as g,z as h}from"../../chunk-X6CF5KTW.js";import{f as p}from"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as d}from"../../chunk-ZLLER3HP.js";import{Args as _,Flags as y}from"@oclif/core";var l=class w extends k{static{d(this,"CronsRun")}static description="Run a maintenance task synchronously, right now";static examples=["<%= config.bin %> crons run cron_check_changelog","<%= config.bin %> crons run cron_suggest_instructions --quiet"];static args={name:_.string({description:"Task name (run `rulemetric crons list` to see all)",required:!0})};static flags={"dry-run":y.boolean({description:"For Harbor, print the saved engine/model and supported tasks without running them"}),quiet:y.boolean({char:"q",description:"Suppress per-line log output; print summary only"})};async run(){let{args:s,flags:a}=await this.parse(w);this.requireAuth();let o=$.find(e=>e.name===s.name);o||this.error(`Unknown task: ${s.name}
+import{a as $}from"../../chunk-4VN5PG2S.js";import{a as k}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as b}from"../../chunk-Z7UN4R7T.js";import{A as f,B as g,z as h}from"../../chunk-X6CF5KTW.js";import{f as p}from"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as d}from"../../chunk-ZLLER3HP.js";import{Args as _,Flags as y}from"@oclif/core";var l=class w extends k{static{d(this,"CronsRun")}static description="Run a maintenance task synchronously, right now";static examples=["<%= config.bin %> crons run cron_check_changelog","<%= config.bin %> crons run cron_suggest_instructions --quiet"];static args={name:_.string({description:"Task name (run `rulemetric crons list` to see all)",required:!0})};static flags={"dry-run":y.boolean({description:"For Harbor, print the saved engine/model and supported tasks without running them"}),quiet:y.boolean({char:"q",description:"Suppress per-line log output; print summary only"})};async run(){let{args:s,flags:a}=await this.parse(w);this.requireAuth();let o=$.find(e=>e.name===s.name);o||this.error(`Unknown task: ${s.name}
 Run \`rulemetric crons list\` to see available tasks.`);let t=null;if(s.name==="harbor_ab"){let e=await b("/api/settings/measurement");t=e.execution.harbor?p.parse(e.execution.harbor):null;for(let r of h("proven")){let m=f(r,t);if(m){this.log(`Unsupported: ${m}`);continue}let i=g(r,"manual",t);this.log(`${r.task}: ${t?.engine??"claude"} / ${i.model}, ${i.class}, ${i.nTrials} trials per arm`)}if(a["dry-run"])return}else a["dry-run"]&&this.error("--dry-run is currently supported only for harbor_ab");let u;try{({runManualTask:u}=await import("../../lib/manual-tasks.js"))}catch(e){let r=e?.code;throw(r==="ERR_MODULE_NOT_FOUND"||r==="MODULE_NOT_FOUND")&&this.error("`crons run` needs the worker runtime, which isn't available in this install.\nReinstall the CLI and verify that its runtime dependencies are present. A database credential is not required for end-user maintenance tasks."),e}this.log(`\u25B6 Running ${o.name} \u2026`);let v=Date.now(),n=await u(s.name,{measurement:t});if(!a.quiet)for(let e of n.logs){let r=e.level==="error"||e.level==="warn"?`[${e.level}]`:"";this.log(`  ${r} ${e.message}`)}let c=((n.durationMs??Date.now()-v)/1e3).toFixed(1);n.ok?this.log(`\u2713 ${o.name} completed in ${c}s (${n.logs.length} log line${n.logs.length===1?"":"s"})`):this.error(`\u2717 ${o.name} failed after ${c}s: ${n.error}`)}};export{l as default};

package/dist/commands/dashboard.js

@@ -1 +1 @@
-import{a as o}from"../chunk-WUCACPGC.js";import"../chunk-IZTIYR5Q.js";import"../chunk-WL5ZWQAS.js";import"../chunk-KEVDLLHT.js";import"../chunk-DX6IPDED.js";import"../chunk-3AV3GTMP.js";import"../chunk-GIXHG5FK.js";import"../chunk-X6CF5KTW.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-MOQTSJNZ.js";import"../chunk-IZTIYR5Q.js";import"../chunk-TDDJOSFQ.js";import"../chunk-433HK3QQ.js";import"../chunk-WDLAPZPN.js";import"../chunk-3AV3GTMP.js";import"../chunk-Z7UN4R7T.js";import"../chunk-X6CF5KTW.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};

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-KGU2IW54.js";import"../chunk-M7QSC64D.js";import"../chunk-ZKPURXUS.js";import"../chunk-ISEUEYSC.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-LJE5J3EW.js";import"../chunk-AHLVY52I.js";import"../chunk-YXMAGC53.js";import"../chunk-T7HZZ4DJ.js";import"../chunk-DHIEGG4T.js";import"../chunk-PPAHIVG5.js";import"../chunk-5MXZJWQS.js";import"../chunk-CT7FWSIU.js";import"../chunk-N7ODX6PO.js";import{a as V}from"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-FADINV4H.js";import"../chunk-SWSOEYC7.js";import"../chunk-U2NT5OSQ.js";import"../chunk-KGURGIZ2.js";import"../chunk-GYNZ3OZF.js";import"../chunk-UPFH72PF.js";import"../chunk-FG343JM3.js";import"../chunk-CDY7TQFZ.js";import"../chunk-KFNFLMUR.js";import"../chunk-4GOXEEXZ.js";import"../chunk-JP6ULOEX.js";import"../chunk-JNLDISBA.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-WUCACPGC.js";import"../chunk-IZTIYR5Q.js";import"../chunk-WL5ZWQAS.js";import"../chunk-KEVDLLHT.js";import"../chunk-DX6IPDED.js";import"../chunk-3AV3GTMP.js";import{a as U,e as y}from"../chunk-GIXHG5FK.js";import"../chunk-X6CF5KTW.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-YSY25OGG.js";import"../chunk-M7QSC64D.js";import"../chunk-CEM7JMUW.js";import"../chunk-L3AXNCWC.js";import"../chunk-ANHFE6US.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-YX2IGS72.js";import"../chunk-55HFAX5Y.js";import"../chunk-26XE3JCL.js";import"../chunk-PPAHIVG5.js";import"../chunk-5MXZJWQS.js";import"../chunk-CT7FWSIU.js";import"../chunk-N7ODX6PO.js";import{a as V}from"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-AXTEV56C.js";import"../chunk-VZYNZS5U.js";import"../chunk-U2NT5OSQ.js";import"../chunk-KGURGIZ2.js";import"../chunk-GYNZ3OZF.js";import"../chunk-UPFH72PF.js";import"../chunk-U6AASXY3.js";import"../chunk-CDY7TQFZ.js";import"../chunk-KFNFLMUR.js";import"../chunk-LARY6ZFT.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-MOQTSJNZ.js";import"../chunk-IZTIYR5Q.js";import"../chunk-TDDJOSFQ.js";import"../chunk-433HK3QQ.js";import"../chunk-WDLAPZPN.js";import"../chunk-3AV3GTMP.js";import{a as U,e as y}from"../chunk-Z7UN4R7T.js";import"../chunk-X6CF5KTW.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/eval-targets/create.js

@@ -1 +1 @@
-import{a}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as s}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as o}from"../../chunk-ZLLER3HP.js";import{Flags as e}from"@oclif/core";import{readFileSync as p}from"node:fs";var n=class c extends a{static{o(this,"EvalTargetCreate")}static description="Create an eval target (instruction, skill, prompt, or plan to test)";static examples=["<%= config.bin %> eval-targets create --name my-rules --type instruction --file ./CLAUDE.md",'<%= config.bin %> eval-targets create --name my-prompt --type prompt --content "Always use TypeScript"'];static flags={name:e.string({char:"n",description:"Name for the eval target",required:!0}),type:e.string({char:"t",description:"Target type",options:["instruction","skill","prompt","plan"],required:!0}),file:e.string({char:"f",description:"Path to content file"}),content:e.string({char:"c",description:"Content string (alternative to --file)"}),description:e.string({char:"d",description:"Description of the target",default:""})};async run(){let{flags:t}=await this.parse(c);this.requireAuth();let i;t.file?i=p(t.file,"utf-8"):t.content?i=t.content:this.error("Provide either --file or --content");let r=await s("/api/eval-targets",{name:t.name,type:t.type,content:i,description:t.description});this.log(`Created eval target: ${r.name} (${r.id}) v${r.version}`)}};export{n as default};
+import{a}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{f as s}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as o}from"../../chunk-ZLLER3HP.js";import{Flags as e}from"@oclif/core";import{readFileSync as p}from"node:fs";var n=class c extends a{static{o(this,"EvalTargetCreate")}static description="Create an eval target (instruction, skill, prompt, or plan to test)";static examples=["<%= config.bin %> eval-targets create --name my-rules --type instruction --file ./CLAUDE.md",'<%= config.bin %> eval-targets create --name my-prompt --type prompt --content "Always use TypeScript"'];static flags={name:e.string({char:"n",description:"Name for the eval target",required:!0}),type:e.string({char:"t",description:"Target type",options:["instruction","skill","prompt","plan"],required:!0}),file:e.string({char:"f",description:"Path to content file"}),content:e.string({char:"c",description:"Content string (alternative to --file)"}),description:e.string({char:"d",description:"Description of the target",default:""})};async run(){let{flags:t}=await this.parse(c);this.requireAuth();let i;t.file?i=p(t.file,"utf-8"):t.content?i=t.content:this.error("Provide either --file or --content");let r=await s("/api/eval-targets",{name:t.name,type:t.type,content:i,description:t.description});this.log(`Created eval target: ${r.name} (${r.id}) v${r.version}`)}};export{n as default};

package/dist/commands/eval-targets/list.js

@@ -1 +1 @@
-import{a as p,b as e}from"../../chunk-W2MV25SB.js";import{a as o}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as s}from"../../chunk-ZLLER3HP.js";import{Flags as m}from"@oclif/core";import{printTable as d}from"@oclif/table";var r=class l extends o{static{s(this,"EvalTargetList")}static description="List eval targets";static examples=["<%= config.bin %> eval-targets list","<%= config.bin %> eval-targets list --type instruction"];static flags={type:m.string({char:"t",description:"Filter by type",options:["instruction","skill","prompt","plan"]})};async run(){let{flags:i}=await this.parse(l);this.requireAuth();let c=i.type?`?type=${i.type}`:"",a=await n(`/api/eval-targets${c}`);if(a.length===0){this.log("No eval targets found. Create one with `rulemetric eval-targets create`.");return}d({data:a.map(t=>({Name:e(t.name,30),Type:t.type,Version:`v${t.version}`,Description:e(t.description,40),Updated:p(t.updatedAt)})),columns:["Name","Type","Version","Description","Updated"]})}};export{r as default};
+import{a as p,b as e}from"../../chunk-W2MV25SB.js";import{a as o}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as s}from"../../chunk-ZLLER3HP.js";import{Flags as m}from"@oclif/core";import{printTable as d}from"@oclif/table";var r=class l extends o{static{s(this,"EvalTargetList")}static description="List eval targets";static examples=["<%= config.bin %> eval-targets list","<%= config.bin %> eval-targets list --type instruction"];static flags={type:m.string({char:"t",description:"Filter by type",options:["instruction","skill","prompt","plan"]})};async run(){let{flags:i}=await this.parse(l);this.requireAuth();let c=i.type?`?type=${i.type}`:"",a=await n(`/api/eval-targets${c}`);if(a.length===0){this.log("No eval targets found. Create one with `rulemetric eval-targets create`.");return}d({data:a.map(t=>({Name:e(t.name,30),Type:t.type,Version:`v${t.version}`,Description:e(t.description,40),Updated:p(t.updatedAt)})),columns:["Name","Type","Version","Description","Updated"]})}};export{r 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-J4W6GNCF.js";import{b as M}from"../../chunk-GH4EZK2O.js";import{c as g}from"../../chunk-HTAHNQZJ.js";import{e as D}from"../../chunk-OCCJEWOF.js";import"../../chunk-XXPWPKTD.js";import"../../chunk-5TM7AOH2.js";import"../../chunk-IEQDLURA.js";import"../../chunk-KINIFUZQ.js";import"../../chunk-U3YPVLM3.js";import"../../chunk-TQM7MF5C.js";import"../../chunk-WOANSBMJ.js";import"../../chunk-IG4JMWFK.js";import"../../chunk-ZYGBABMZ.js";import"../../chunk-GIQ2HK3W.js";import"../../chunk-TTZWVZIW.js";import"../../chunk-FZMO7XSN.js";import"../../chunk-MOJY6S5J.js";import"../../chunk-I6W2KC6W.js";import"../../chunk-5A6TMSRO.js";import"../../chunk-HYQ5YKBC.js";import"../../chunk-JU5ZNXPK.js";import"../../chunk-3WCX2SLU.js";import"../../chunk-KOORMTZO.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-Q7IDLOSP.js";import"../../chunk-CZBO3A6B.js";import"../../chunk-USKRCVX3.js";import"../../chunk-7YZURVZD.js";import"../../chunk-PZS55PW2.js";import"../../chunk-QPAOX7P2.js";import"../../chunk-3ZHVP3KE.js";import"../../chunk-EFEPD4AZ.js";import"../../chunk-HOQPVOMF.js";import"../../chunk-E4CIJ67Z.js";import"../../chunk-EHW2KS6R.js";import"../../chunk-D4FP5M2B.js";import"../../chunk-CVPAKEEZ.js";import"../../chunk-VUNJGZPP.js";import"../../chunk-GGUXEZC3.js";import"../../chunk-GLBWNB2N.js";import"../../chunk-CQD2R3BA.js";import"../../chunk-3UKZJAFS.js";import"../../chunk-77FH6UZP.js";import"../../chunk-KE5K3F7J.js";import"../../chunk-PIVEJTNF.js";import"../../chunk-KR36ZDGN.js";import"../../chunk-WL34G6IR.js";import"../../chunk-GIRPEMTW.js";import"../../chunk-SIGZEQ3K.js";import"../../chunk-DQYYHX6K.js";import"../../chunk-RIYVZDOQ.js";import"../../chunk-6T6YFPBW.js";import"../../chunk-ISYU5BMU.js";import"../../chunk-DANMUHNW.js";import"../../chunk-MMMT34KO.js";import"../../chunk-FXQS7AEZ.js";import"../../chunk-LJRE3LZ2.js";import"../../chunk-TSLBCZ5P.js";import"../../chunk-65LFPUDB.js";import"../../chunk-HBDKOU2N.js";import"../../chunk-3H44B6PB.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-MHVHGKDK.js";import"../../chunk-RBSEE2Z6.js";import"../../chunk-PDXHRF4E.js";import"../../chunk-KGU2IW54.js";import"../../chunk-M7QSC64D.js";import"../../chunk-ZKPURXUS.js";import"../../chunk-ISEUEYSC.js";import"../../chunk-GYKYJJNF.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-LJE5J3EW.js";import"../../chunk-AHLVY52I.js";import"../../chunk-YXMAGC53.js";import"../../chunk-T7HZZ4DJ.js";import"../../chunk-DHIEGG4T.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-FADINV4H.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-KFNFLMUR.js";import"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.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 $}from"../../chunk-QI3FJKGK.js";import{a as O}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as m}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{e as I,l as p}from"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as de,context as me}from"@opentelemetry/api";import{NodeSDK as W}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as Y}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as J,ConsoleSpanExporter as z,SimpleSpanProcessor as Q}from"@opentelemetry/sdk-trace-node";import{Resource as Z}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as ee,ATTR_SERVICE_VERSION as te}from"@opentelemetry/semantic-conventions";function F(r){if(!r)return{};let n={};for(let o of r.split(",")){let s=o.indexOf("=");if(s===-1)continue;let i=o.slice(0,s).trim(),c=o.slice(s+1).trim();i&&(n[i]=c)}return n}e(F,"parseOtlpHeaders");var u=null;function T(){return process.env.TELEMETRY_ENABLED!=="false"}e(T,"isTelemetryEnabled");function x(r){if(!T()||u!==null)return"off";let n=new Z({[ee]:r.serviceName,[te]:r.serviceVersion??"0.1.0"}),o=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(o){let c=F(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new Y({url:`${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-J4W6GNCF.js";import{b as M}from"../../chunk-GH4EZK2O.js";import{c as g}from"../../chunk-HTAHNQZJ.js";import{e as D}from"../../chunk-2IQBXE6F.js";import"../../chunk-GPWLS64K.js";import"../../chunk-IEQDLURA.js";import"../../chunk-KINIFUZQ.js";import"../../chunk-HB4RUEEE.js";import"../../chunk-TFJAT263.js";import"../../chunk-JU5ZNXPK.js";import"../../chunk-GLZBTGN5.js";import"../../chunk-V2G6E53W.js";import"../../chunk-G2226CG6.js";import"../../chunk-ZZZ7ECPT.js";import"../../chunk-AZK4Y5DB.js";import"../../chunk-2K4TF5RM.js";import"../../chunk-ZYGBABMZ.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-BLGICIV4.js";import"../../chunk-RAIXY6YA.js";import"../../chunk-HOQPVOMF.js";import"../../chunk-LEUXIXIS.js";import"../../chunk-FZMO7XSN.js";import"../../chunk-MOJY6S5J.js";import"../../chunk-RCJTCLZO.js";import"../../chunk-VBGHPAUT.js";import"../../chunk-4XRPSKIJ.js";import"../../chunk-M2CPTQZL.js";import"../../chunk-LMNUO32P.js";import"../../chunk-N7P3HUPD.js";import"../../chunk-ACJ2DBOE.js";import"../../chunk-FPEC6DJV.js";import"../../chunk-SW24KZ5C.js";import"../../chunk-57YXM7AI.js";import"../../chunk-3ZHVP3KE.js";import"../../chunk-WM437XV6.js";import"../../chunk-R7URROQD.js";import"../../chunk-VFWQ2W4B.js";import"../../chunk-GLBWNB2N.js";import"../../chunk-CQD2R3BA.js";import"../../chunk-3UKZJAFS.js";import"../../chunk-77FH6UZP.js";import"../../chunk-YVEI7XR6.js";import"../../chunk-4GM4OEQY.js";import"../../chunk-GIRPEMTW.js";import"../../chunk-DQYYHX6K.js";import"../../chunk-DB4ZM5N3.js";import"../../chunk-3WNQ52TW.js";import"../../chunk-PIVEJTNF.js";import"../../chunk-KSKC5SEQ.js";import"../../chunk-LMQDS47I.js";import"../../chunk-DANMUHNW.js";import"../../chunk-JQ2ZCUEC.js";import"../../chunk-WWL4HNXM.js";import"../../chunk-3IAZMJA4.js";import"../../chunk-XFOG2PDJ.js";import"../../chunk-SIGZEQ3K.js";import"../../chunk-FXQS7AEZ.js";import"../../chunk-LJRE3LZ2.js";import"../../chunk-QH7JB4UQ.js";import"../../chunk-52OHAULM.js";import"../../chunk-AZB34RT6.js";import"../../chunk-FLXWL4F4.js";import"../../chunk-7V6LLVAY.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-MHVHGKDK.js";import"../../chunk-RBSEE2Z6.js";import"../../chunk-TL57YUJV.js";import"../../chunk-YSY25OGG.js";import"../../chunk-M7QSC64D.js";import"../../chunk-CEM7JMUW.js";import"../../chunk-L3AXNCWC.js";import"../../chunk-ANHFE6US.js";import"../../chunk-GYKYJJNF.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-AHLVY52I.js";import"../../chunk-YX2IGS72.js";import"../../chunk-55HFAX5Y.js";import"../../chunk-26XE3JCL.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-AXTEV56C.js";import"../../chunk-VZYNZS5U.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-U6AASXY3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-KFNFLMUR.js";import"../../chunk-LARY6ZFT.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 $}from"../../chunk-QI3FJKGK.js";import{a as O}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{f as m}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{e as I,l as p}from"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as de,context as me}from"@opentelemetry/api";import{NodeSDK as W}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as Y}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as J,ConsoleSpanExporter as z,SimpleSpanProcessor as Q}from"@opentelemetry/sdk-trace-node";import{Resource as Z}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as ee,ATTR_SERVICE_VERSION as te}from"@opentelemetry/semantic-conventions";function F(r){if(!r)return{};let n={};for(let o of r.split(",")){let s=o.indexOf("=");if(s===-1)continue;let i=o.slice(0,s).trim(),c=o.slice(s+1).trim();i&&(n[i]=c)}return n}e(F,"parseOtlpHeaders");var u=null;function T(){return process.env.TELEMETRY_ENABLED!=="false"}e(T,"isTelemetryEnabled");function x(r){if(!T()||u!==null)return"off";let n=new Z({[ee]:r.serviceName,[te]:r.serviceVersion??"0.1.0"}),o=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(o){let c=F(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new Y({url:`${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/auto-run.js

@@ -1,4 +1,4 @@
-import{a as k}from"../../chunk-RBSEE2Z6.js";import{e as $}from"../../chunk-FADINV4H.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as w}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as A,Flags as r}from"@oclif/core";var m=class y extends w{static{h(this,"EvalAutoRun")}static description="Run evals in a loop until the effect verdict is statistically powered (\xA76f auto-run-to-power)";static examples=["<%= config.bin %> evals auto-run <target-id>            # preview (cost + plan)","<%= config.bin %> evals auto-run <target-id> --yes      # run until powered / budget","<%= config.bin %> evals auto-run <target-id> --yes --max-runs 120 --batch-parallel 4"];static args={targetId:A.string({description:"Eval target ID",required:!0})};static flags={yes:r.boolean({char:"y",description:"Actually run. Without this, only previews the plan + cost."}),"max-runs":r.integer({description:"Hard ceiling on completed runs",default:200}),"batch-parallel":r.integer({description:"Concurrent executions within a batch",default:3}),timeout:r.integer({description:"Timeout per eval in ms",default:12e4}),model:r.string({description:"Model for eval execution"}),"grader-model":r.string({description:"Model for grading"}),"skip-baseline":r.boolean({description:"Skip without_target runs (no delta \u2014 not recommended)",default:!1})};async run(){let{args:C,flags:e}=await this.parse(y);this.requireAuth();let i=C.targetId,c=20,a=await n(`/api/eval-targets/${i}`),l=await n(`/api/eval-targets/${i}/evals`);l.length===0&&this.error(`No active cases on "${a.name}". Generate some (\`evals create\` / the web Generate button) or run \`evals refine-cases\` first.`);let b=l.length*(e["skip-baseline"]?1:2),f=Math.ceil(e["max-runs"]/b);if(this.log(`Auto-run "${a.name}" v${a.version} toward a powered \xB1${c}pp verdict.`),this.log(`Plan: up to ${e["max-runs"]} runs (${l.length} active case(s) \xD7 ${e["skip-baseline"]?1:2}/batch = ${b}/batch, \u2264${f} batches). Each run is a synthetic \`claude -p\` execution + grade.`),this.log('Note: this is a controlled synthetic A/B. For a DEPLOYED instruction with real usage, the Insights observational verdict is cheaper and uses real sessions \u2014 auto-run is for pre-deploy, catalog skills, or targets with no real "without" cohort.'),!e.yes){this.log(`
+import{a as k}from"../../chunk-RBSEE2Z6.js";import{e as $}from"../../chunk-AXTEV56C.js";import"../../chunk-VZYNZS5U.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-U6AASXY3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-LARY6ZFT.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 w}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as A,Flags as r}from"@oclif/core";var m=class y extends w{static{h(this,"EvalAutoRun")}static description="Run evals in a loop until the effect verdict is statistically powered (\xA76f auto-run-to-power)";static examples=["<%= config.bin %> evals auto-run <target-id>            # preview (cost + plan)","<%= config.bin %> evals auto-run <target-id> --yes      # run until powered / budget","<%= config.bin %> evals auto-run <target-id> --yes --max-runs 120 --batch-parallel 4"];static args={targetId:A.string({description:"Eval target ID",required:!0})};static flags={yes:r.boolean({char:"y",description:"Actually run. Without this, only previews the plan + cost."}),"max-runs":r.integer({description:"Hard ceiling on completed runs",default:200}),"batch-parallel":r.integer({description:"Concurrent executions within a batch",default:3}),timeout:r.integer({description:"Timeout per eval in ms",default:12e4}),model:r.string({description:"Model for eval execution"}),"grader-model":r.string({description:"Model for grading"}),"skip-baseline":r.boolean({description:"Skip without_target runs (no delta \u2014 not recommended)",default:!1})};async run(){let{args:C,flags:e}=await this.parse(y);this.requireAuth();let i=C.targetId,c=20,a=await n(`/api/eval-targets/${i}`),l=await n(`/api/eval-targets/${i}/evals`);l.length===0&&this.error(`No active cases on "${a.name}". Generate some (\`evals create\` / the web Generate button) or run \`evals refine-cases\` first.`);let b=l.length*(e["skip-baseline"]?1:2),f=Math.ceil(e["max-runs"]/b);if(this.log(`Auto-run "${a.name}" v${a.version} toward a powered \xB1${c}pp verdict.`),this.log(`Plan: up to ${e["max-runs"]} runs (${l.length} active case(s) \xD7 ${e["skip-baseline"]?1:2}/batch = ${b}/batch, \u2264${f} batches). Each run is a synthetic \`claude -p\` execution + grade.`),this.log('Note: this is a controlled synthetic A/B. For a DEPLOYED instruction with real usage, the Insights observational verdict is cheaper and uses real sessions \u2014 auto-run is for pre-deploy, catalog skills, or targets with no real "without" cohort.'),!e.yes){this.log(`
 Preview only. Re-run with --yes to start.`);return}let t={runsCompleted:0,maxRuns:e["max-runs"],batchesRun:0,maxBatches:f},x={targetId:i,targetVersion:a.version,iteration:1,timeout:e.timeout,model:e.model,graderModel:e["grader-model"],skipGrading:!1,targetType:a.type,targetName:a.name};for(;;){let s;try{s=await n(`/api/eval-targets/${i}/judge-calibration`)}catch(o){this.log(`
 \u25A0 stopping \u2014 could not read the power check: ${o instanceof Error?o.message:String(o)}`);break}let E={powered:s.powerCheck?.powered??!1,mde:s.powerCheck?.mde==null?1/0:s.powerCheck.mde},B=s.caseDiscrimination??{total:0,discriminating:0,nonDiscriminating:0,unpaired:0},p=k(E,B,t,{targetEffectPp:c});if(this.log(`
 ${p.action==="stop"?"\u25A0":"\u25B8"} ${p.reason}`),p.action==="stop")break;let v=await n(`/api/eval-targets/${i}/evals`);if(v.length===0){this.log("\u25A0 no active cases left \u2014 stopping.");break}let{completed:g,failed:D}=await $({cases:v,targetContent:a.content,concurrency:e["batch-parallel"],runNumber:t.batchesRun+1,opts:x,skipBaseline:e["skip-baseline"],onLog:h(o=>this.log(o),"onLog")});if(t.runsCompleted+=g,t.batchesRun+=1,this.log(`Batch ${t.batchesRun}: ${g} completed, ${D} failed \u2014 ${t.runsCompleted}/${t.maxRuns} runs, ${t.batchesRun}/${t.maxBatches} batches.`),g===0){this.log("\u25A0 batch produced 0 completed runs \u2014 the executor appears broken (check the model / `claude -p`). Aborting.");break}}let d=await n(`/api/eval-targets/${i}/judge-calibration`),u=d.powerCheck?.mde,R=u==null||!Number.isFinite(u)?"unknown":`\xB1${Math.round(u*100)}pp`;this.log(`

package/dist/commands/evals/benchmark.js

@@ -1,4 +1,4 @@
-import{b as h,c as f}from"../../chunk-7LE2W3WP.js";import"../../chunk-KFNFLMUR.js";import{a as m}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as c}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as v}from"@oclif/core";var d=class b extends m{static{u(this,"EvalBenchmarkCommand")}static description="Aggregate eval runs into a benchmark for a target";static examples=["<%= config.bin %> evals benchmark <target-id>","<%= config.bin %> evals benchmark <target-id> --iteration 2"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={iteration:v.integer({description:"Filter runs by iteration",default:1})};async run(){let{args:o,flags:l}=await this.parse(b);this.requireAuth();let g=await c(`/api/eval-targets/${o.targetId}`);this.log(`Aggregating benchmark for "${g.name}" v${g.version}, iteration ${l.iteration}`);let a=await h(o.targetId,l.iteration);a.totalRuns===0&&this.error("No completed runs found. Run evals first with `rulemetric evals run`."),this.log(`
+import{b as h,c as f}from"../../chunk-PHBPHZSZ.js";import"../../chunk-KFNFLMUR.js";import{a as m}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as c}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as v}from"@oclif/core";var d=class b extends m{static{u(this,"EvalBenchmarkCommand")}static description="Aggregate eval runs into a benchmark for a target";static examples=["<%= config.bin %> evals benchmark <target-id>","<%= config.bin %> evals benchmark <target-id> --iteration 2"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={iteration:v.integer({description:"Filter runs by iteration",default:1})};async run(){let{args:o,flags:l}=await this.parse(b);this.requireAuth();let g=await c(`/api/eval-targets/${o.targetId}`);this.log(`Aggregating benchmark for "${g.name}" v${g.version}, iteration ${l.iteration}`);let a=await h(o.targetId,l.iteration);a.totalRuns===0&&this.error("No completed runs found. Run evals first with `rulemetric evals run`."),this.log(`
 Aggregated ${a.totalRuns} runs
 `);for(let[e,n]of Object.entries(a.runSummary)){this.log(`${e}:`);for(let[r,i]of Object.entries(n)){let s=i;this.log(`  ${r}: mean=${s.mean} stddev=${s.stddev} [${s.min}, ${s.max}]`)}}let t=a.verdict;if(t)if(this.log(`
 Verdict (${t.mode}, baseline=${t.baseline}):`),!t.sufficient)this.log(`  ${t.insufficientReason??"insufficient data"} \u2014 no effect claimed`);else{let e=t.delta>=0?"+":"";this.log(`  pass-rate delta: ${e}${(t.delta*100).toFixed(0)}% (n=${t.nWith} paired)`),this.log(`  95% CI: ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%`),t.pValue!=null&&this.log(`  McNemar p: ${t.pValue.toFixed(4)}`),this.log(`  significant: ${t.significant?"YES":"no"}`)}if(a.notes.length>0){this.log(`

package/dist/commands/evals/compare.js

@@ -1,4 +1,4 @@
-import{a as $,b as A}from"../../chunk-647M2I2Y.js";import{f as p}from"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as l,f as w}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as B,Flags as b}from"@oclif/core";var c=class I extends f{static{h(this,"EvalCompareCommand")}static description="Blind A/B comparison between two eval target versions";static examples=["<%= config.bin %> evals compare <target-a> <target-b>"];static args={targetA:B.string({description:"First eval target ID",required:!0}),targetB:B.string({description:"Second eval target ID",required:!0})};static flags={model:b.string({description:"Model for execution and comparison"}),timeout:b.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:g,flags:o}=await this.parse(I);this.requireAuth();let[e,r]=await Promise.all([l(`/api/eval-targets/${g.targetA}`),l(`/api/eval-targets/${g.targetB}`)]),m=await l(`/api/eval-targets/${g.targetA}/evals`);m.length===0&&this.error("No evals found for target A. Create some first."),this.log(`Comparing "${e.name}" v${e.version} vs "${r.name}" v${r.version}`),this.log(`Running ${m.length} eval(s)...
+import{a as $,b as A}from"../../chunk-647M2I2Y.js";import{g as p}from"../../chunk-LARY6ZFT.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SVZYRGO5.js";import"../../chunk-WULEBQPI.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{e as l,f as w}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as B,Flags as b}from"@oclif/core";var c=class I extends f{static{h(this,"EvalCompareCommand")}static description="Blind A/B comparison between two eval target versions";static examples=["<%= config.bin %> evals compare <target-a> <target-b>"];static args={targetA:B.string({description:"First eval target ID",required:!0}),targetB:B.string({description:"Second eval target ID",required:!0})};static flags={model:b.string({description:"Model for execution and comparison"}),timeout:b.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:g,flags:o}=await this.parse(I);this.requireAuth();let[e,r]=await Promise.all([l(`/api/eval-targets/${g.targetA}`),l(`/api/eval-targets/${g.targetB}`)]),m=await l(`/api/eval-targets/${g.targetA}/evals`);m.length===0&&this.error("No evals found for target A. Create some first."),this.log(`Comparing "${e.name}" v${e.version} vs "${r.name}" v${r.version}`),this.log(`Running ${m.length} eval(s)...
 `);let a=Math.random()>.5,E=a?r.content:e.content,T=a?e.content:r.content,C=a?r.name:e.name,x=a?e.name:r.name,v=[];for(let t of m){this.log(`[${t.name}] Running A...`);let s=await p({prompt:t.prompt,targetContent:E,timeout:o.timeout,model:o.model});this.log(`[${t.name}] Running B...`);let d=await p({prompt:t.prompt,targetContent:T,timeout:o.timeout,model:o.model});(s.error||d.error)&&this.error(`[${t.name}] execution failed (A: ${s.error??"ok"} | B: ${d.error??"ok"}) \u2014 aborting before the blind comparison`),v.push({evalName:t.name,outputA:s.output,outputB:d.output})}this.log(`
 Running blind comparison...`);let N=v.map(t=>`
 ## Eval: ${t.evalName}

package/dist/commands/evals/create.js

@@ -1 +1 @@
-import{a as c}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as i}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as m}from"../../chunk-ZLLER3HP.js";import{Args as f,Flags as r}from"@oclif/core";import{readFileSync as g}from"node:fs";var s=class d extends c{static{m(this,"EvalCreate")}static description="Create test cases (evals) for an eval target";static examples=['<%= config.bin %> evals create <target-id> --name "ts-check" --prompt "Write a function to add two numbers"',"<%= config.bin %> evals create <target-id> --from-file evals.json"];static args={targetId:f.string({description:"Eval target ID",required:!0})};static flags={name:r.string({char:"n",description:"Eval name"}),prompt:r.string({char:"p",description:"Test prompt"}),expectations:r.string({char:"e",description:"Assertion (repeatable)",multiple:!0}),"from-file":r.string({description:"Bulk create from JSON file"})};async run(){let{args:n,flags:e}=await this.parse(d);if(this.requireAuth(),e["from-file"]){let a=JSON.parse(g(e["from-file"],"utf-8")),p=Array.isArray(a)?a:a.evals;for(let t of p){let l=await i(`/api/eval-targets/${n.targetId}/evals`,{name:t.name,prompt:t.prompt,expectations:t.expectations??[],files:t.files??[]});this.log(`Created eval: ${l.name} (${l.id})`)}this.log(`Created ${p.length} evals.`);return}(!e.name||!e.prompt)&&this.error("Provide --name and --prompt, or use --from-file for bulk creation");let o=await i(`/api/eval-targets/${n.targetId}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations??[]});this.log(`Created eval: ${o.name} (${o.id})`)}};export{s as default};
+import{a as c}from"../../chunk-MOQTSJNZ.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-TDDJOSFQ.js";import"../../chunk-433HK3QQ.js";import"../../chunk-WDLAPZPN.js";import"../../chunk-3AV3GTMP.js";import{f as i}from"../../chunk-Z7UN4R7T.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as m}from"../../chunk-ZLLER3HP.js";import{Args as f,Flags as r}from"@oclif/core";import{readFileSync as g}from"node:fs";var s=class d extends c{static{m(this,"EvalCreate")}static description="Create test cases (evals) for an eval target";static examples=['<%= config.bin %> evals create <target-id> --name "ts-check" --prompt "Write a function to add two numbers"',"<%= config.bin %> evals create <target-id> --from-file evals.json"];static args={targetId:f.string({description:"Eval target ID",required:!0})};static flags={name:r.string({char:"n",description:"Eval name"}),prompt:r.string({char:"p",description:"Test prompt"}),expectations:r.string({char:"e",description:"Assertion (repeatable)",multiple:!0}),"from-file":r.string({description:"Bulk create from JSON file"})};async run(){let{args:n,flags:e}=await this.parse(d);if(this.requireAuth(),e["from-file"]){let a=JSON.parse(g(e["from-file"],"utf-8")),p=Array.isArray(a)?a:a.evals;for(let t of p){let l=await i(`/api/eval-targets/${n.targetId}/evals`,{name:t.name,prompt:t.prompt,expectations:t.expectations??[],files:t.files??[]});this.log(`Created eval: ${l.name} (${l.id})`)}this.log(`Created ${p.length} evals.`);return}(!e.name||!e.prompt)&&this.error("Provide --name and --prompt, or use --from-file for bulk creation");let o=await i(`/api/eval-targets/${n.targetId}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations??[]});this.log(`Created eval: ${o.name} (${o.id})`)}};export{s as default};