@rulemetric/cli 0.14.0 → 0.14.1

package/dist/base-command.js

@@ -1 +1 @@
-import{a}from"./chunk-N4RJ6CN7.js";import"./chunk-IZTIYR5Q.js";import"./chunk-E5YE25QG.js";import"./chunk-YUIRDBQQ.js";import"./chunk-HAYO6JYI.js";import"./chunk-3AV3GTMP.js";import"./chunk-ZZHIBOMT.js";import"./chunk-TWPG3TEQ.js";import"./chunk-74QIUVWY.js";import"./chunk-PZNWH4VQ.js";import"./chunk-ZLLER3HP.js";export{a as BaseCommand};
+import{a}from"./chunk-MU7KKU3O.js";import"./chunk-IZTIYR5Q.js";import"./chunk-FKTT7A4H.js";import"./chunk-VIFOLTZT.js";import"./chunk-4FJYVYCU.js";import"./chunk-3AV3GTMP.js";import"./chunk-76UKAFNM.js";import"./chunk-42A4PUW7.js";import"./chunk-XQYAVSUQ.js";import"./chunk-PZNWH4VQ.js";import"./chunk-ZLLER3HP.js";export{a as BaseCommand};

package/dist/{chunk-LMCICVE7.js → chunk-2UK4DZFW.js}

@@ -1,4 +1,4 @@
-import{a as U,b as F}from"./chunk-J7UZXRTH.js";import{a as h,b as I}from"./chunk-BVOXWAJO.js";import{a as w,b as k,e as R}from"./chunk-LJRE3LZ2.js";import{d as $}from"./chunk-GYNZ3OZF.js";import{g as C,i as D,j as P}from"./chunk-UPFH72PF.js";import{e as v,f as x,i as j}from"./chunk-ZZHIBOMT.js";import{D as S,F as b,G as T,o as A}from"./chunk-TWPG3TEQ.js";import{a as l}from"./chunk-ZLLER3HP.js";import*as u from"node:fs";import*as m from"node:path";var O=new Set(["mcp","subagent","feature"]);function ie(e,t,n){return L(m.join(e,"CLAUDE.md"),t,n)}l(ie,"appendToCLAUDEmd");function L(e,t,n){let i=w(t),s=k(t),a=u.existsSync(e)?u.readFileSync(e,"utf8"):"";if(a.includes(i))return!1;let p=`${a.length>0&&!a.endsWith(`
+import{a as U,b as F}from"./chunk-3FNTV2TW.js";import{a as h,b as I}from"./chunk-B5X6HCBK.js";import{a as w,b as k,e as R}from"./chunk-LJRE3LZ2.js";import{d as $}from"./chunk-GYNZ3OZF.js";import{g as C,i as D,j as P}from"./chunk-UPFH72PF.js";import{e as v,f as x,i as j}from"./chunk-76UKAFNM.js";import{D as S,F as b,G as T,o as A}from"./chunk-42A4PUW7.js";import{a as l}from"./chunk-ZLLER3HP.js";import*as u from"node:fs";import*as m from"node:path";var O=new Set(["mcp","subagent","feature"]);function ie(e,t,n){return L(m.join(e,"CLAUDE.md"),t,n)}l(ie,"appendToCLAUDEmd");function L(e,t,n){let i=w(t),s=k(t),a=u.existsSync(e)?u.readFileSync(e,"utf8"):"";if(a.includes(i))return!1;let p=`${a.length>0&&!a.endsWith(`
 `)?`
 
 `:a.length>0?`

package/dist/{chunk-FK55FXKQ.js → chunk-35YVHBLL.js}

@@ -1,2 +1,2 @@
-import{c as E,n as m}from"./chunk-SE3UKA3M.js";import{a as d}from"./chunk-ZZHIBOMT.js";import{a as c}from"./chunk-ZLLER3HP.js";import*as p from"node:fs";import*as u from"node:path";import*as _ from"node:os";function T(){if(process.env.RULEMETRIC_API_KEY||process.env.RULEMETRIC_ACCESS_TOKEN)return;let s=u.join(_.homedir(),".config","rulemetric","env");C(s)||C(u.join(process.cwd(),".env.local"))}c(T,"ensureEnv");function C(s){try{let t=p.readFileSync(s,"utf-8");for(let n of t.split(`
+import{c as E,n as m}from"./chunk-T75KAEJO.js";import{a as d}from"./chunk-76UKAFNM.js";import{a as c}from"./chunk-ZLLER3HP.js";import*as p from"node:fs";import*as u from"node:path";import*as _ from"node:os";function T(){if(process.env.RULEMETRIC_API_KEY||process.env.RULEMETRIC_ACCESS_TOKEN)return;let s=u.join(_.homedir(),".config","rulemetric","env");C(s)||C(u.join(process.cwd(),".env.local"))}c(T,"ensureEnv");function C(s){try{let t=p.readFileSync(s,"utf-8");for(let n of t.split(`
 `)){let e=n.trim();if(!e||e.startsWith("#"))continue;let o=e.indexOf("=");if(o<0)continue;let r=e.slice(0,o).trim(),i=e.slice(o+1).trim();process.env[r]||(process.env[r]=i)}return!!(process.env.RULEMETRIC_API_KEY||process.env.RULEMETRIC_ACCESS_TOKEN)}catch{return!1}}c(C,"loadEnvFile");T();async function I(s){let t={totalSessions:0,totalCalls:0,totalInputTokens:0,totalOutputTokens:0,totalCacheReadTokens:0,totalCacheWriteTokens:0,totalCostUSD:0,providers:{}},n=await m(s);t.providers={};for(let a of n)t.providers[a.provider]=(t.providers[a.provider]||0)+1;let e=new E,o=n.filter(a=>a.provider==="claude");for(let a of o)try{let h=p.readdirSync(a.path).filter(f=>f.endsWith(".jsonl"));t.totalSessions+=h.length;for(let f of h){let v=p.readFileSync(u.join(a.path,f),"utf-8"),R=e.parse(v);for(let l of R.events)l.tokenUsage&&(t.totalCalls++,t.totalInputTokens+=l.tokenUsage.input,t.totalOutputTokens+=l.tokenUsage.output,t.totalCacheReadTokens+=l.tokenUsage.cacheRead??0,t.totalCacheWriteTokens+=l.tokenUsage.cacheWrite??0)}}catch{}let r=15,i=75;return t.totalCostUSD=Math.round((t.totalInputTokens/1e6*r+t.totalOutputTokens/1e6*i+t.totalCacheReadTokens/1e6*(r*.1)+t.totalCacheWriteTokens/1e6*(r*1.25))*100)/100,t}c(I,"loadLocalStats");async function U(s){let t=process.env.RULEMETRIC_API_URL??d,n=process.env.RULEMETRIC_API_KEY,e=process.env.RULEMETRIC_ACCESS_TOKEN,o=n??e;if(!o)return null;let r=new URLSearchParams;s&&r.set("projectPath",s);try{let i=await fetch(`${t}/api/insights/compute?${r}`,{headers:{Authorization:`Bearer ${o}`},signal:AbortSignal.timeout(5e3)});return i.ok?await i.json():null}catch{return null}}c(U,"loadApiInsights");async function g(s){let t=process.env.RULEMETRIC_API_URL??d,n=process.env.RULEMETRIC_API_KEY??process.env.RULEMETRIC_ACCESS_TOKEN;if(!n)return null;let e=new URLSearchParams;s&&e.set("projectPath",s);try{let o=await fetch(`${t}/api/insights/linked-instructions?${e}`,{headers:{Authorization:`Bearer ${n}`},signal:AbortSignal.timeout(5e3)});return o.ok?await o.json():null}catch{return null}}c(g,"loadLinkedInstructions");export{I as a,U as b,g as c};

package/dist/{chunk-J7UZXRTH.js → chunk-3FNTV2TW.js}

@@ -1 +1 @@
-import{b as A,c as l,d as b}from"./chunk-N7SZWKCU.js";import{a as I}from"./chunk-BVOXWAJO.js";import{f as k}from"./chunk-ZZHIBOMT.js";import{n as f,o as g,p as u,v,w as $,x as y}from"./chunk-TWPG3TEQ.js";import{a as c}from"./chunk-ZLLER3HP.js";import{randomUUID as P}from"node:crypto";async function H(s){let{applyMode:e="human",...r}=s;await k("/api/recommendation-applications",{...r,harness:"claude_code",applyMode:e,status:"applied",applyBatchId:P(),applySeq:0})}c(H,"recordApplication");async function M(s,e){let r=c(t=>({action:"manual-required",instructionId:e.id,name:e.name,fileModified:!1,messages:t}),"refuse");if(s.actor!=="human")return r([`Not installed: ${e.name} is a hook, and a hook is a command this machine runs unattended.`,"Installing one is always a human decision \u2014 the same refusal the API makes server-side.",`  Review it:  rulemetric suggestions accept ${s.suggestionId}`]);let n=y(e);if(!n.spec)return r([`Not installed: ${e.name} does not contain an installable hook spec.`,...n.problems.map(t=>`  ${t.field}: ${t.message}`)]);if(n.problems.length>0)return r([`Not installed: ${e.name}'s hook spec is not fit to act on.`,...n.problems.map(t=>`  ${t.field}: ${t.message}`)]);let w=u(e.frontmatter),i=v(n.spec,w??void 0),p=f(i);if(p.length>0)return r([`Not installed: ${e.name}'s artifact spec is not fit to act on.`,...p.map(t=>`  ${t.field}: ${t.message}`)]);let o=b(s.dir,i.requires);if(o.missing.length>0)return r([`Not installed: ${e.name} needs ${o.missing.join(", ")}, which is not present here.`,"It is left unapplied rather than installed and broken \u2014 a hook whose command is missing still runs, and fails, on every matching event.","Accept it again once the dependency exists."]);let a=A({dir:s.dir,instructionId:e.id,spec:n.spec});if(a.status==="refused")return r([`Not installed: ${e.name}.`,`  ${a.reason}`]);let m=g(i,a.inverse);if(m.length>0)return l({dir:s.dir,inverse:a.inverse}),r([`Not installed: ${e.name} was written and then rolled back \u2014 the actuator could not produce a valid inverse.`,...m.map(t=>`  ${t.field}: ${t.message}`)]);try{await H({instructionId:e.id,projectPath:s.projectPath,targetPath:a.target,beforeHash:a.beforeHash,appliedHash:a.inverse.hashAtWrite,boundaryClass:i.boundaryClass,targets:i.targets,inverse:a.inverse,requires:i.requires,provides:i.provides})}catch(t){let h=l({dir:s.dir,inverse:a.inverse});return r([`Not installed: ${e.name} \u2014 the effect could not be recorded, so it was undone.`,`  ${t instanceof Error?t.message:String(t)}`,h.removed?"  The hook was removed; your settings are as they were.":`  WARNING: the rollback did not complete (${h.reason??"unknown reason"}). Check ${a.target} for the entry tagged with ${e.id}.`])}await I(s.suggestionId,s.projectPath,e.id,s.actor);let d=a.status==="installed"?[`Accepted: installed a ${n.spec.event} hook into ${a.target}.`]:[`Already present: this hook is already installed in ${a.target}.`];for(let t of $(n.spec))d.push(`  ${t}`);return o.unverifiable.length>0&&d.push(`  unchecked: ${o.unverifiable.join(", ")} \u2014 this command cannot verify that namespace.`),d.push(`Undo with: rulemetric suggestions undo ${e.id}`),{action:a.status==="installed"?"added":"already-present",instructionId:e.id,name:e.name,fileModified:a.status==="installed",messages:d}}c(M,"acceptHook");export{H as a,M as b};
+import{b as A,c as l,d as b}from"./chunk-JTYVBR5H.js";import{a as I}from"./chunk-B5X6HCBK.js";import{f as k}from"./chunk-76UKAFNM.js";import{n as f,o as g,p as u,v,w as $,x as y}from"./chunk-42A4PUW7.js";import{a as c}from"./chunk-ZLLER3HP.js";import{randomUUID as P}from"node:crypto";async function H(s){let{applyMode:e="human",...r}=s;await k("/api/recommendation-applications",{...r,harness:"claude_code",applyMode:e,status:"applied",applyBatchId:P(),applySeq:0})}c(H,"recordApplication");async function M(s,e){let r=c(t=>({action:"manual-required",instructionId:e.id,name:e.name,fileModified:!1,messages:t}),"refuse");if(s.actor!=="human")return r([`Not installed: ${e.name} is a hook, and a hook is a command this machine runs unattended.`,"Installing one is always a human decision \u2014 the same refusal the API makes server-side.",`  Review it:  rulemetric suggestions accept ${s.suggestionId}`]);let n=y(e);if(!n.spec)return r([`Not installed: ${e.name} does not contain an installable hook spec.`,...n.problems.map(t=>`  ${t.field}: ${t.message}`)]);if(n.problems.length>0)return r([`Not installed: ${e.name}'s hook spec is not fit to act on.`,...n.problems.map(t=>`  ${t.field}: ${t.message}`)]);let w=u(e.frontmatter),i=v(n.spec,w??void 0),p=f(i);if(p.length>0)return r([`Not installed: ${e.name}'s artifact spec is not fit to act on.`,...p.map(t=>`  ${t.field}: ${t.message}`)]);let o=b(s.dir,i.requires);if(o.missing.length>0)return r([`Not installed: ${e.name} needs ${o.missing.join(", ")}, which is not present here.`,"It is left unapplied rather than installed and broken \u2014 a hook whose command is missing still runs, and fails, on every matching event.","Accept it again once the dependency exists."]);let a=A({dir:s.dir,instructionId:e.id,spec:n.spec});if(a.status==="refused")return r([`Not installed: ${e.name}.`,`  ${a.reason}`]);let m=g(i,a.inverse);if(m.length>0)return l({dir:s.dir,inverse:a.inverse}),r([`Not installed: ${e.name} was written and then rolled back \u2014 the actuator could not produce a valid inverse.`,...m.map(t=>`  ${t.field}: ${t.message}`)]);try{await H({instructionId:e.id,projectPath:s.projectPath,targetPath:a.target,beforeHash:a.beforeHash,appliedHash:a.inverse.hashAtWrite,boundaryClass:i.boundaryClass,targets:i.targets,inverse:a.inverse,requires:i.requires,provides:i.provides})}catch(t){let h=l({dir:s.dir,inverse:a.inverse});return r([`Not installed: ${e.name} \u2014 the effect could not be recorded, so it was undone.`,`  ${t instanceof Error?t.message:String(t)}`,h.removed?"  The hook was removed; your settings are as they were.":`  WARNING: the rollback did not complete (${h.reason??"unknown reason"}). Check ${a.target} for the entry tagged with ${e.id}.`])}await I(s.suggestionId,s.projectPath,e.id,s.actor);let d=a.status==="installed"?[`Accepted: installed a ${n.spec.event} hook into ${a.target}.`]:[`Already present: this hook is already installed in ${a.target}.`];for(let t of $(n.spec))d.push(`  ${t}`);return o.unverifiable.length>0&&d.push(`  unchecked: ${o.unverifiable.join(", ")} \u2014 this command cannot verify that namespace.`),d.push(`Undo with: rulemetric suggestions undo ${e.id}`),{action:a.status==="installed"?"added":"already-present",instructionId:e.id,name:e.name,fileModified:a.status==="installed",messages:d}}c(M,"acceptHook");export{H as a,M as b};

package/dist/{chunk-U7BPWFOP.js → chunk-3IMRLHJJ.js}

@@ -1 +1 @@
-import{b as g}from"./chunk-3ZHVP3KE.js";import{d as k,e as h}from"./chunk-O62OLKQL.js";import{e as m,i as f}from"./chunk-ZZHIBOMT.js";import{a as c}from"./chunk-ZLLER3HP.js";import{existsSync as b}from"node:fs";async function w($,y){let r=c(t=>y?.logger?.info?.(t)??console.log(t),"log"),{projectPath:s}=$,n=`[cron_retire_unused_skills] ${s}`;if(!s||!b(s))return{skipped:"project-missing",reason:`${s} does not exist on this machine \u2014 refusing to delete files for a project that is not here.`};let o=await m("/api/insights/groom-candidates"),e=g({captureHealthy:o.skillCaptureHealthy===!0,captureReason:o.skillCaptureHealthy===void 0?"API did not report skill capture health (older deployment)":o.skillCaptureReason,candidates:o.unusedSkills??[]});if(e.refused)return r(`${n} ${e.refused}: ${e.summary}`),{skipped:e.refused,reason:e.summary,held:e.held.length};r(`${n} ${e.summary}`);let a=[],d=[],l=0;for(let t of e.retire){let u=k({kind:"skill_file",target:t.name,cwd:s});if(!u.ok){l+=1,r(`${n} could not plan removal of ${t.name}: ${u.refusal??"unresolved"}`);continue}let R=new Date().toISOString().replace(/[:.]/g,"-"),i=h(u,R);if(!i.applied){l+=1,r(`${n} failed to remove ${t.name}: ${i.error??"unknown error"}`);continue}a.push(t.name),i.backupPath&&d.push(i.backupPath),r(`${n} retired ${t.name} (backup: ${i.backupPath??"none"})`);try{await f(`/api/instructions/${t.instructionId}`)}catch(p){r(`${n} removed ${t.name} but could not archive its instruction row: ${p instanceof Error?p.message:String(p)}`)}}return{retired:a.length,failed:l,held:e.held.length,names:a,backups:d,reason:e.summary}}c(w,"cronRetireUnusedSkills");export{w as a};
+import{b as g}from"./chunk-3ZHVP3KE.js";import{d as k,e as h}from"./chunk-RXZLBSDO.js";import{e as m,i as f}from"./chunk-76UKAFNM.js";import{a as c}from"./chunk-ZLLER3HP.js";import{existsSync as b}from"node:fs";async function w($,y){let r=c(t=>y?.logger?.info?.(t)??console.log(t),"log"),{projectPath:s}=$,n=`[cron_retire_unused_skills] ${s}`;if(!s||!b(s))return{skipped:"project-missing",reason:`${s} does not exist on this machine \u2014 refusing to delete files for a project that is not here.`};let o=await m("/api/insights/groom-candidates"),e=g({captureHealthy:o.skillCaptureHealthy===!0,captureReason:o.skillCaptureHealthy===void 0?"API did not report skill capture health (older deployment)":o.skillCaptureReason,candidates:o.unusedSkills??[]});if(e.refused)return r(`${n} ${e.refused}: ${e.summary}`),{skipped:e.refused,reason:e.summary,held:e.held.length};r(`${n} ${e.summary}`);let a=[],d=[],l=0;for(let t of e.retire){let u=k({kind:"skill_file",target:t.name,cwd:s});if(!u.ok){l+=1,r(`${n} could not plan removal of ${t.name}: ${u.refusal??"unresolved"}`);continue}let R=new Date().toISOString().replace(/[:.]/g,"-"),i=h(u,R);if(!i.applied){l+=1,r(`${n} failed to remove ${t.name}: ${i.error??"unknown error"}`);continue}a.push(t.name),i.backupPath&&d.push(i.backupPath),r(`${n} retired ${t.name} (backup: ${i.backupPath??"none"})`);try{await f(`/api/instructions/${t.instructionId}`)}catch(p){r(`${n} removed ${t.name} but could not archive its instruction row: ${p instanceof Error?p.message:String(p)}`)}}return{retired:a.length,failed:l,held:e.held.length,names:a,backups:d,reason:e.summary}}c(w,"cronRetireUnusedSkills");export{w as a};

package/dist/{chunk-TWPG3TEQ.js → chunk-42A4PUW7.js}

@@ -1,4 +1,4 @@
-import{a as I,b as N}from"./chunk-74QIUVWY.js";import{a as i}from"./chunk-ZLLER3HP.js";function ne(e){return(e.trim().split(`
+import{f as I,g as N}from"./chunk-XQYAVSUQ.js";import{a as i}from"./chunk-ZLLER3HP.js";function ne(e){return(e.trim().split(`
 `)[0]?.replace(/^#+\s*/,"").trim()||"Suggested instruction").slice(0,120)}i(ne,"recommendationTitle");function oe(e){if(!e||typeof e!="object")return e;let t=e;if(Array.isArray(t.recommendation_candidates))return e;let n=t.suggestions;if(!n||typeof n!="object")return e;let o=n.claude_md_additions;return Array.isArray(o)&&(t.recommendation_candidates=o.flatMap(r=>{if(!r||typeof r!="object")return[];let s=r;if(typeof s.addition!="string"||s.addition.trim()==="")return[];let a=s.addition.trim();return[{title:ne(a),artifact:"instruction",canonicalBody:a,rationale:typeof s.why=="string"?s.why:"",evidence:[],outcomeMetric:"session_outcome",renderings:[{harness:"claude_code",target:"CLAUDE.md",body:a},{harness:"codex",target:"AGENTS.md",body:a}]}]})),e}i(oe,"normalizeRecommendationCandidates");var re=new Set(["instruction"]);function se(e){if(!e||typeof e!="object")return e;let t=e,n=t.recommendation_candidates;if(!Array.isArray(n)||n.length===0)return e;let o=n.filter(s=>!!s&&typeof s=="object"&&typeof s.canonicalBody=="string"&&s.canonicalBody.trim().length>0);for(let s of o){if(Array.isArray(s.renderings)&&s.renderings.length>0)continue;let a=s.canonicalBody.trim(),c=re.has(String(s.artifact));s.renderings=c?[{harness:"claude_code",target:"CLAUDE.md",body:a},{harness:"codex",target:"AGENTS.md",body:a}]:[{harness:"claude_code",target:"manual",body:a},{harness:"codex",target:"manual",body:a}]}let r=t.suggestions&&typeof t.suggestions=="object"?t.suggestions:{};return(!Array.isArray(r.claude_md_additions)||r.claude_md_additions.length===0)&&(r.claude_md_additions=o.filter(s=>s.artifact==="instruction").map(s=>({addition:s.canonicalBody.trim(),why:typeof s.rationale=="string"?s.rationale:"",prompt_scaffold:"Add to CLAUDE.md (and AGENTS.md for Codex)."}))),(!Array.isArray(r.features_to_try)||r.features_to_try.length===0)&&(r.features_to_try=o.filter(s=>s.artifact!=="instruction").map(s=>({feature:typeof s.title=="string"?s.title:"Proposed change",one_liner:typeof s.rationale=="string"?s.rationale:"",why_for_you:typeof s.outcomeMetric=="string"?`Aimed at ${s.outcomeMetric}.`:"",example_code:s.canonicalBody.trim()}))),Array.isArray(r.usage_patterns)||(r.usage_patterns=[]),t.suggestions=r,e}i(se,"deriveLegacyPresentation");function ie(e){if(!e||typeof e!="object")return e;let t=e,n=t.project_areas;if(n!==void 0&&!Array.isArray(n)){let o=n&&typeof n=="object"?n.areas:void 0;t.project_areas=Array.isArray(o)?o:void 0}return ae(e),oe(e),e}i(ie,"normalizeRecommendations");function ae(e){if(!e||typeof e!="object")return;let t=e;if(Array.isArray(t.recommendation_candidates)&&t.recommendation_candidates.length>0)return;let n=t.suggestions;if(!n||typeof n!="object")return;let o=n.recommendation_candidates;!Array.isArray(o)||o.length===0||(t.recommendation_candidates=o)}i(ae,"liftNestedCandidates");import{z as d}from"zod";var P=["org","team","project","personal"],D=d.object({name:d.string().min(1),description:d.string().min(1),type:d.literal("instruction").default("instruction"),scope:d.enum(P),version:d.string().min(1),tags:d.array(d.string()).optional(),compatibility:d.array(d.string()).optional(),slots:d.record(d.string(),d.string()).optional(),metadata:d.record(d.string(),d.unknown()).optional()});import{z as p}from"zod";var j=p.object({name:p.string().min(1),description:p.string().min(1),type:p.literal("prompt"),version:p.string().min(1),tags:p.array(p.string()).optional(),variables:p.array(p.string()).optional(),metadata:p.record(p.string(),p.unknown()).optional()});function _(e,t){let n=t.map(o=>`- ID: ${o.id}
   Name: ${o.name}
   Description: ${o.description}
@@ -43,6 +43,6 @@ Only include instructions with a relevance score >= ${.5}. Order by score descen
 
 `),removed:!0,removedText:l}}i(ge,"removeMarkdownSection");function ye(e,t){let n=t.trim();if(n.length===0)return{content:e,removed:!1,removedText:null,reason:"not-found"};let o=e.indexOf(n);if(o===-1)return{content:e,removed:!1,removedText:null,reason:"not-found"};if(e.indexOf(n,o+1)!==-1)return{content:e,removed:!1,removedText:null,reason:"ambiguous"};let r=(e.slice(0,o)+e.slice(o+n.length)).replace(/\n{3,}/g,`
 
-`);return r.trim().length===0?{content:e,removed:!1,removedText:null,reason:"would-empty-file"}:{content:r,removed:!0,removedText:n}}i(ye,"removeExactBlock");var U=["HTTPS_PROXY","HTTP_PROXY","NODE_EXTRA_CA_CERTS"];function xe(e,t){if(!e||typeof e!="object")return{value:e,removed:!1,removedDescriptions:[],reason:"not-found"};let n=structuredClone(e);if(!n.hooks)return{value:n,removed:!1,removedDescriptions:[],reason:"not-found"};let o=[];for(let[r,s]of Object.entries(n.hooks)){if(!Array.isArray(s))continue;let a=s.filter(c=>{let l=(c.hooks??[]).some(u=>typeof u.command=="string"&&u.command.includes(t));return l&&o.push(`${r}: ${c.matcher??"*"} \u2192 ${(c.hooks??[]).map(u=>u.command).join(", ")}`),!l});a.length===0?delete n.hooks[r]:n.hooks[r]=a}return Object.keys(n.hooks).length===0&&delete n.hooks,{value:n,removed:o.length>0,removedDescriptions:o,...o.length===0?{reason:"not-found"}:{}}}i(xe,"removeHookByCommand");function _e(e,t){if(!e||typeof e!="object")return{value:e,removed:!1,removedDescriptions:[],reason:"not-found"};let n=structuredClone(e);return!n.mcpServers||!(t in n.mcpServers)?{value:n,removed:!1,removedDescriptions:[],reason:"not-found"}:(delete n.mcpServers[t],Object.keys(n.mcpServers).length===0&&delete n.mcpServers,{value:n,removed:!0,removedDescriptions:[`mcpServers.${t}`]})}i(_e,"removeMcpServer");function v(e){if(!e||typeof e!="object")return[];let t=e.env;return t?U.filter(n=>n in t):[]}i(v,"protectedEnvKeysPresent");function Ee(e,t){let n=v(e),o=new Set(v(t));return n.filter(r=>!o.has(r))}i(Ee,"lostProtectedEnvKeys");function Se(e,t){let n=`HTTP ${e}`;if(!t||typeof t!="object")return{message:n};let o=t,r=typeof o.error=="string"&&o.error?o.error:o.error?JSON.stringify(o.error):n,s=typeof o.code=="string"&&o.code?o.code:void 0;return s?{message:r,code:s}:{message:r}}i(Se,"parseApiError");var T=["inside","outside"];function y(e){return e.trim().toLowerCase()}i(y,"normalizeDependencyKey");function ve(e){let t=[];if(T.includes(e.boundaryClass)||t.push({field:"boundaryClass",message:`must be one of ${T.join(" | ")}`}),!Array.isArray(e.targets)||e.targets.length===0)t.push({field:"targets",message:"a proposal that cannot name the paths it writes cannot be applied"});else{for(let r of e.targets)if(typeof r!="string"||r.trim().length===0){t.push({field:"targets",message:"target paths must be non-empty"});break}let n=e.targets.filter(r=>typeof r=="string"&&(r.startsWith("/")||/^[A-Za-z]:[\\/]/.test(r)));n.length>0&&t.push({field:"targets",message:`targets must be project-relative (absolute: ${n.join(", ")})`});let o=e.targets.filter(r=>typeof r=="string"&&r.split(/[\\/]/).includes(".."));o.length>0&&t.push({field:"targets",message:`targets must not traverse outside the project (${o.join(", ")})`})}for(let n of["requires","provides"]){let o=e[n];if(!Array.isArray(o)){t.push({field:n,message:"must be an array of dependency keys"});continue}o.some(r=>typeof r!="string"||r.trim().length===0)&&t.push({field:n,message:"dependency keys must be non-empty strings"})}return t}i(ve,"validateArtifactSpec");function Ae(e,t){let n=[],o=new Set(e.targets??[]);if(t.kind==="delete-paths"){e.boundaryClass!=="inside"&&n.push({field:"inverse",message:"deleting paths is only an inverse for `inside` artifacts; an `outside` path may hold text we did not write"}),t.paths.length===0&&n.push({field:"inverse.paths",message:"nothing to delete \u2014 this is not an inverse"});let r=t.paths.filter(s=>!o.has(s));return r.length>0&&n.push({field:"inverse.paths",message:`would delete paths the spec never declared: ${r.join(", ")}`}),n}return e.boundaryClass!=="outside"&&n.push({field:"inverse",message:"a compensation is only an inverse for `outside` artifacts; an `inside` path owes an exact undo"}),(!t.backupPath||t.backupPath.trim().length===0)&&n.push({field:"inverse.backupPath",message:"an `outside` write owes a backup"}),(!t.hashAtWrite||t.hashAtWrite.trim().length===0)&&n.push({field:"inverse.hashAtWrite",message:"without the post-write hash the undo cannot tell a drifted file from an untouched one"}),o.has(t.file)||n.push({field:"inverse.file",message:`wrote ${t.file}, which the spec did not declare as a target`}),t.kind==="remove-marked-region"?(!t.startMarker||!t.endMarker)&&n.push({field:"inverse",message:"an `outside` region needs both markers, or the undo cannot tell where our text ends"}):t.command.trim().length===0&&n.push({field:"inverse.command",message:"the exact command written is the identity of the entry; without it the undo is a search"}),n}i(Ae,"validateArtifactInverse");var z="artifact";function A(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&t.trim().length>0):[]}i(A,"stringArray");function Te(e){if(!e||typeof e!="object")return null;let t=e[z];if(!t||typeof t!="object")return null;let n=t;return{boundaryClass:n.boundaryClass,targets:A(n.targets),requires:A(n.requires),provides:A(n.provides)}}i(Te,"parseArtifactSpec");var be=["applied","measuring","improved","inconclusive","harmful","rollback_ready"];function K(e){return e.appliedAt?e.appliedAt instanceof Date?e.appliedAt.getTime():new Date(e.appliedAt).getTime():0}i(K,"appliedAtMs");function ke(e){return[...e].sort((t,n)=>{let o=K(n)-K(t);return o!==0?o:(n.applySeq??0)-(t.applySeq??0)})}i(ke,"orderForReversal");var Y=["SessionStart","SessionEnd","UserPromptSubmit","PreToolUse","PostToolUse","PreCompact","Notification","Stop","SubagentStop"],b=".claude/settings.local.json";function G(e){return`# rm-artifact:${e}`}i(G,"hookTag");function Ce(e,t){return`${e.trim()} ${G(t)}`}i(Ce,"taggedHookCommand");var Oe=new Set(["if","then","else","elif","fi","for","while","until","do","done","case","esac","function","return","exit","cd","echo","test","true","false"]);function k(e){let t=e.trim().split(/\s+/)[0]??"";return!/^[a-z][a-z0-9_-]*$/.test(t)||Oe.has(t)?null:y(`cmd:${t}`)}i(k,"inferredCommandRequirement");function we(e,t){let n=new Set((t?.requires??[]).map(y).filter(Boolean)),o=k(e.command);return o&&n.add(o),{boundaryClass:"outside",targets:[b],requires:[...n],provides:(t?.provides??[]).map(y).filter(Boolean)}}i(we,"hookArtifactSpec");function x(e){let t=[];Y.includes(e.event)||t.push({field:"event",message:`unknown hook event ${JSON.stringify(e.event)} \u2014 a hook on an event that never fires is indistinguishable from one that was never installed`});let n=(e.command??"").trim();return n.length===0?t.push({field:"command",message:"a hook with no command installs nothing"}):(/[\r\n]/.test(n)&&t.push({field:"command",message:"multi-line commands are refused: the identity tag is a trailing shell comment, which would silence only the final line, and a human reviewing an install should read one line"}),n.includes("\0")&&t.push({field:"command",message:"command contains a NUL byte"}),n.includes("rulemetric")&&t.push({field:"command",message:'refused: `rulemetric hooks uninstall` deletes every hook whose command contains "rulemetric", so installing one here would make this hook collateral damage of an unrelated command'}),n.includes("rm-artifact:")&&t.push({field:"command",message:"command already carries an artifact tag \u2014 installing it would produce two identities for one entry"})),e.matcher!==void 0&&(typeof e.matcher!="string"||e.matcher.trim().length===0)&&t.push({field:"matcher",message:"matcher must be a non-empty string when present"}),t}i(x,"validateHookSpec");function Re(e){let t=[`runs on: ${e.event}${e.matcher?` (tools matching ${e.matcher})`:""}`,`spawns:  ${e.command.trim()}`,`writes:  ${b} (one entry, tagged with this artifact's id)`],n=k(e.command);return n&&t.push(`needs:   ${n} on PATH`),t}i(Re,"describeHookCapability");var q="hook";function h(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}i(h,"asRecord");function F(e){let t=e.event,n=e.command;if(typeof t!="string"||typeof n!="string")return null;let o=typeof e.matcher=="string"?e.matcher:void 0;return{event:t,matcher:o,command:n}}i(F,"readBareSpec");function W(e){let t=h(e.hooks)??e,n=[];for(let[o,r]of Object.entries(t))if(Array.isArray(r))for(let s of r){let a=h(s);if(!a)continue;let c=typeof a.matcher=="string"&&a.matcher.length>0?a.matcher:void 0,l=Array.isArray(a.hooks)?a.hooks:[];for(let u of l){let f=h(u)?.command;typeof f=="string"&&n.push({event:o,matcher:c,command:f})}}return n.length===0?{spec:null,problems:[]}:n.length>1?{spec:null,problems:[{field:"hook",message:`proposal describes ${n.length} hook entries; one proposal installs exactly one hook, so its inverse stays exact and its review stays readable`}]}:{spec:n[0],problems:[]}}i(W,"readSettingsFragment");function Le(e){let t=[],n=/```[a-zA-Z]*\n([\s\S]*?)```/g,o;for(;(o=n.exec(e))!==null;)t.push(o[1]);return t}i(Le,"fencedBlocks");function Ie(e){let t=h(h(e.frontmatter)?.artifact),n=t?h(t[q]):null;if(n){let o=F(n);if(o)return{spec:o,problems:x(o),source:"envelope"};let r=W(n);if(r.spec)return{spec:r.spec,problems:x(r.spec),source:"envelope"};if(r.problems.length>0)return{spec:null,problems:r.problems,source:"envelope"}}for(let o of Le(e.content??"")){let r;try{r=JSON.parse(o)}catch{continue}let s=h(r);if(!s)continue;let a=F(s);if(a)return{spec:a,problems:x(a),source:"content-fragment"};let c=W(s);if(c.spec)return{spec:c.spec,problems:x(c.spec),source:"content-fragment"};if(c.problems.length>0)return{spec:null,problems:c.problems,source:"content-fragment"}}return{spec:null,problems:[{field:"hook",message:"no hook spec found \u2014 a hook proposal must carry {event, command} in its artifact envelope or a JSON settings fragment in its body"}],source:"none"}}i(Ie,"parseHookSpec");var V=25,C=N,X=[{task:"rm-batch-task",instructions:"rm-batch-task/inject.md",class:"proven",axis:"reward",measured:"2026-09-02 n=25/arm Haiku 4.5: 0/25 \u2192 25/25, delta +1.000, 95% CI [+0.734, +1.000]"},{task:"rm-locality-task",instructions:"rm-locality-task/inject.md",class:"proven",axis:"cost",measured:"2026-09-02 n=25/arm Haiku 4.5: 25/25 both arms, cost \u221225.6%/trial, 95% CI [\u22120.0171, \u22120.0107]"},{task:"rm-migrations-task",instructions:"rm-migrations-task/inject.md",class:"negative",axis:"reward",measured:"2026-08-31 Haiku 4.5: flat after two reworks \u2014 redundant with the model's priors on this shape"},{task:"rm-pagination-task",instructions:"rm-pagination-task/inject.md",class:"negative",axis:"reward",measured:"2026-09-02 n=25/arm Haiku 4.5: 25/25 baseline under --allowedTools Bash; the rule costs +8.0%/trial and buys nothing"}];function Ne(e){return X.filter(t=>t.class===e)}i(Ne,"harborTasksOfClass");function J(e,t){return t?.engine==="codex"&&e.axis==="cost"?`${e.task}: dollar-cost telemetry is unavailable for Codex; reward calibration remains supported`:null}i(J,"harborTaskSupport");function Pe(e,t,n=null){let o=n?I.parse(n):null,r=J(e,o);if(r)throw new Error(r);let s=!o||o.engine==="claude"&&o.model===C;return{task:e.task,instructions:e.instructions,class:s?e.class:"unvalidated",axis:e.axis,model:o?.model??C,...o?{measurement:o}:{},nTrials:V,lane:t}}i(Pe,"harborPayloadFor");var De=/[\p{L}\p{N}]+/gu;function O(e){let t=new Set;for(let n of e.toLowerCase().matchAll(De)){let o=`  ${n[0]} `;for(let r=0;r+3<=o.length;r+=1)t.add(o.slice(r,r+3))}return t}i(O,"trigrams");function g(e,t){let n=O(e),o=O(t);if(n.size===0||o.size===0)return 0;let r=0,[s,a]=n.size<=o.size?[n,o]:[o,n];for(let l of s)a.has(l)&&(r+=1);let c=n.size+o.size-r;return c===0?0:r/c}i(g,"trigramSimilarity");var R=.3,je=15;function w(e){let n=e.replace(/<!--[\s\S]*?-->/g," ").replace(/```[\s\S]*?```/g," "),o=[];for(let r of n.split(`
+`);return r.trim().length===0?{content:e,removed:!1,removedText:null,reason:"would-empty-file"}:{content:r,removed:!0,removedText:n}}i(ye,"removeExactBlock");var U=["HTTPS_PROXY","HTTP_PROXY","NODE_EXTRA_CA_CERTS"];function xe(e,t){if(!e||typeof e!="object")return{value:e,removed:!1,removedDescriptions:[],reason:"not-found"};let n=structuredClone(e);if(!n.hooks)return{value:n,removed:!1,removedDescriptions:[],reason:"not-found"};let o=[];for(let[r,s]of Object.entries(n.hooks)){if(!Array.isArray(s))continue;let a=s.filter(c=>{let l=(c.hooks??[]).some(u=>typeof u.command=="string"&&u.command.includes(t));return l&&o.push(`${r}: ${c.matcher??"*"} \u2192 ${(c.hooks??[]).map(u=>u.command).join(", ")}`),!l});a.length===0?delete n.hooks[r]:n.hooks[r]=a}return Object.keys(n.hooks).length===0&&delete n.hooks,{value:n,removed:o.length>0,removedDescriptions:o,...o.length===0?{reason:"not-found"}:{}}}i(xe,"removeHookByCommand");function _e(e,t){if(!e||typeof e!="object")return{value:e,removed:!1,removedDescriptions:[],reason:"not-found"};let n=structuredClone(e);return!n.mcpServers||!(t in n.mcpServers)?{value:n,removed:!1,removedDescriptions:[],reason:"not-found"}:(delete n.mcpServers[t],Object.keys(n.mcpServers).length===0&&delete n.mcpServers,{value:n,removed:!0,removedDescriptions:[`mcpServers.${t}`]})}i(_e,"removeMcpServer");function v(e){if(!e||typeof e!="object")return[];let t=e.env;return t?U.filter(n=>n in t):[]}i(v,"protectedEnvKeysPresent");function Ee(e,t){let n=v(e),o=new Set(v(t));return n.filter(r=>!o.has(r))}i(Ee,"lostProtectedEnvKeys");function Se(e,t){let n=`HTTP ${e}`;if(!t||typeof t!="object")return{message:n};let o=t,r=typeof o.error=="string"&&o.error?o.error:o.error?JSON.stringify(o.error):n,s=typeof o.code=="string"&&o.code?o.code:void 0;return s?{message:r,code:s}:{message:r}}i(Se,"parseApiError");var T=["inside","outside"];function y(e){return e.trim().toLowerCase()}i(y,"normalizeDependencyKey");function ve(e){let t=[];if(T.includes(e.boundaryClass)||t.push({field:"boundaryClass",message:`must be one of ${T.join(" | ")}`}),!Array.isArray(e.targets)||e.targets.length===0)t.push({field:"targets",message:"a proposal that cannot name the paths it writes cannot be applied"});else{for(let r of e.targets)if(typeof r!="string"||r.trim().length===0){t.push({field:"targets",message:"target paths must be non-empty"});break}let n=e.targets.filter(r=>typeof r=="string"&&(r.startsWith("/")||/^[A-Za-z]:[\\/]/.test(r)));n.length>0&&t.push({field:"targets",message:`targets must be project-relative (absolute: ${n.join(", ")})`});let o=e.targets.filter(r=>typeof r=="string"&&r.split(/[\\/]/).includes(".."));o.length>0&&t.push({field:"targets",message:`targets must not traverse outside the project (${o.join(", ")})`})}for(let n of["requires","provides"]){let o=e[n];if(!Array.isArray(o)){t.push({field:n,message:"must be an array of dependency keys"});continue}o.some(r=>typeof r!="string"||r.trim().length===0)&&t.push({field:n,message:"dependency keys must be non-empty strings"})}return t}i(ve,"validateArtifactSpec");function Ae(e,t){let n=[],o=new Set(e.targets??[]);if(t.kind==="delete-paths"){e.boundaryClass!=="inside"&&n.push({field:"inverse",message:"deleting paths is only an inverse for `inside` artifacts; an `outside` path may hold text we did not write"}),t.paths.length===0&&n.push({field:"inverse.paths",message:"nothing to delete \u2014 this is not an inverse"});let r=t.paths.filter(s=>!o.has(s));return r.length>0&&n.push({field:"inverse.paths",message:`would delete paths the spec never declared: ${r.join(", ")}`}),n}return e.boundaryClass!=="outside"&&n.push({field:"inverse",message:"a compensation is only an inverse for `outside` artifacts; an `inside` path owes an exact undo"}),(!t.backupPath||t.backupPath.trim().length===0)&&n.push({field:"inverse.backupPath",message:"an `outside` write owes a backup"}),(!t.hashAtWrite||t.hashAtWrite.trim().length===0)&&n.push({field:"inverse.hashAtWrite",message:"without the post-write hash the undo cannot tell a drifted file from an untouched one"}),o.has(t.file)||n.push({field:"inverse.file",message:`wrote ${t.file}, which the spec did not declare as a target`}),t.kind==="remove-marked-region"?(!t.startMarker||!t.endMarker)&&n.push({field:"inverse",message:"an `outside` region needs both markers, or the undo cannot tell where our text ends"}):t.command.trim().length===0&&n.push({field:"inverse.command",message:"the exact command written is the identity of the entry; without it the undo is a search"}),n}i(Ae,"validateArtifactInverse");var z="artifact";function A(e){return Array.isArray(e)?e.filter(t=>typeof t=="string"&&t.trim().length>0):[]}i(A,"stringArray");function Te(e){if(!e||typeof e!="object")return null;let t=e[z];if(!t||typeof t!="object")return null;let n=t;return{boundaryClass:n.boundaryClass,targets:A(n.targets),requires:A(n.requires),provides:A(n.provides)}}i(Te,"parseArtifactSpec");var be=["applied","measuring","improved","inconclusive","harmful","rollback_ready"];function K(e){return e.appliedAt?e.appliedAt instanceof Date?e.appliedAt.getTime():new Date(e.appliedAt).getTime():0}i(K,"appliedAtMs");function ke(e){return[...e].sort((t,n)=>{let o=K(n)-K(t);return o!==0?o:(n.applySeq??0)-(t.applySeq??0)})}i(ke,"orderForReversal");var Y=["SessionStart","SessionEnd","UserPromptSubmit","PreToolUse","PostToolUse","PreCompact","Notification","Stop","SubagentStop"],b=".claude/settings.local.json";function G(e){return`# rm-artifact:${e}`}i(G,"hookTag");function Ce(e,t){return`${e.trim()} ${G(t)}`}i(Ce,"taggedHookCommand");var Oe=new Set(["if","then","else","elif","fi","for","while","until","do","done","case","esac","function","return","exit","cd","echo","test","true","false"]);function k(e){let t=e.trim().split(/\s+/)[0]??"";return!/^[a-z][a-z0-9_-]*$/.test(t)||Oe.has(t)?null:y(`cmd:${t}`)}i(k,"inferredCommandRequirement");function we(e,t){let n=new Set((t?.requires??[]).map(y).filter(Boolean)),o=k(e.command);return o&&n.add(o),{boundaryClass:"outside",targets:[b],requires:[...n],provides:(t?.provides??[]).map(y).filter(Boolean)}}i(we,"hookArtifactSpec");function x(e){let t=[];Y.includes(e.event)||t.push({field:"event",message:`unknown hook event ${JSON.stringify(e.event)} \u2014 a hook on an event that never fires is indistinguishable from one that was never installed`});let n=(e.command??"").trim();return n.length===0?t.push({field:"command",message:"a hook with no command installs nothing"}):(/[\r\n]/.test(n)&&t.push({field:"command",message:"multi-line commands are refused: the identity tag is a trailing shell comment, which would silence only the final line, and a human reviewing an install should read one line"}),n.includes("\0")&&t.push({field:"command",message:"command contains a NUL byte"}),n.includes("rulemetric")&&t.push({field:"command",message:'refused: `rulemetric hooks uninstall` deletes every hook whose command contains "rulemetric", so installing one here would make this hook collateral damage of an unrelated command'}),n.includes("rm-artifact:")&&t.push({field:"command",message:"command already carries an artifact tag \u2014 installing it would produce two identities for one entry"})),e.matcher!==void 0&&(typeof e.matcher!="string"||e.matcher.trim().length===0)&&t.push({field:"matcher",message:"matcher must be a non-empty string when present"}),t}i(x,"validateHookSpec");function Re(e){let t=[`runs on: ${e.event}${e.matcher?` (tools matching ${e.matcher})`:""}`,`spawns:  ${e.command.trim()}`,`writes:  ${b} (one entry, tagged with this artifact's id)`],n=k(e.command);return n&&t.push(`needs:   ${n} on PATH`),t}i(Re,"describeHookCapability");var q="hook";function h(e){return e&&typeof e=="object"&&!Array.isArray(e)?e:null}i(h,"asRecord");function F(e){let t=e.event,n=e.command;if(typeof t!="string"||typeof n!="string")return null;let o=typeof e.matcher=="string"?e.matcher:void 0;return{event:t,matcher:o,command:n}}i(F,"readBareSpec");function W(e){let t=h(e.hooks)??e,n=[];for(let[o,r]of Object.entries(t))if(Array.isArray(r))for(let s of r){let a=h(s);if(!a)continue;let c=typeof a.matcher=="string"&&a.matcher.length>0?a.matcher:void 0,l=Array.isArray(a.hooks)?a.hooks:[];for(let u of l){let f=h(u)?.command;typeof f=="string"&&n.push({event:o,matcher:c,command:f})}}return n.length===0?{spec:null,problems:[]}:n.length>1?{spec:null,problems:[{field:"hook",message:`proposal describes ${n.length} hook entries; one proposal installs exactly one hook, so its inverse stays exact and its review stays readable`}]}:{spec:n[0],problems:[]}}i(W,"readSettingsFragment");function Le(e){let t=[],n=/```[a-zA-Z]*\n([\s\S]*?)```/g,o;for(;(o=n.exec(e))!==null;)t.push(o[1]);return t}i(Le,"fencedBlocks");function Ie(e){let t=h(h(e.frontmatter)?.artifact),n=t?h(t[q]):null;if(n){let o=F(n);if(o)return{spec:o,problems:x(o),source:"envelope"};let r=W(n);if(r.spec)return{spec:r.spec,problems:x(r.spec),source:"envelope"};if(r.problems.length>0)return{spec:null,problems:r.problems,source:"envelope"}}for(let o of Le(e.content??"")){let r;try{r=JSON.parse(o)}catch{continue}let s=h(r);if(!s)continue;let a=F(s);if(a)return{spec:a,problems:x(a),source:"content-fragment"};let c=W(s);if(c.spec)return{spec:c.spec,problems:x(c.spec),source:"content-fragment"};if(c.problems.length>0)return{spec:null,problems:c.problems,source:"content-fragment"}}return{spec:null,problems:[{field:"hook",message:"no hook spec found \u2014 a hook proposal must carry {event, command} in its artifact envelope or a JSON settings fragment in its body"}],source:"none"}}i(Ie,"parseHookSpec");var V=25,C=N,X=[{task:"rm-batch-task",instructions:"rm-batch-task/inject.md",class:"proven",axis:"reward",measured:"2026-09-02 n=25/arm Haiku 4.5: 0/25 \u2192 25/25, delta +1.000, 95% CI [+0.734, +1.000]"},{task:"rm-locality-task",instructions:"rm-locality-task/inject.md",class:"proven",axis:"cost",measured:"2026-09-02 n=25/arm Haiku 4.5: 25/25 both arms, cost \u221225.6%/trial, 95% CI [\u22120.0171, \u22120.0107]"},{task:"rm-migrations-task",instructions:"rm-migrations-task/inject.md",class:"negative",axis:"reward",measured:"2026-08-31 Haiku 4.5: flat after two reworks \u2014 redundant with the model's priors on this shape"},{task:"rm-pagination-task",instructions:"rm-pagination-task/inject.md",class:"negative",axis:"reward",measured:"2026-09-02 n=25/arm Haiku 4.5: 25/25 baseline under --allowedTools Bash; the rule costs +8.0%/trial and buys nothing"}];function Ne(e){return X.filter(t=>t.class===e)}i(Ne,"harborTasksOfClass");function J(e,t){return t?.engine==="pi"?"Pi is supported for eval trials, not Harbor tasks":t?.engine==="codex"&&e.axis==="cost"?`${e.task}: dollar-cost telemetry is unavailable for Codex; reward calibration remains supported`:null}i(J,"harborTaskSupport");function Pe(e,t,n=null){let o=n?I.parse(n):null,r=J(e,o);if(r)throw new Error(r);let s=!o||o.engine==="claude"&&o.model===C;return{task:e.task,instructions:e.instructions,class:s?e.class:"unvalidated",axis:e.axis,model:o?.model??C,...o?{measurement:o}:{},nTrials:V,lane:t}}i(Pe,"harborPayloadFor");var De=/[\p{L}\p{N}]+/gu;function O(e){let t=new Set;for(let n of e.toLowerCase().matchAll(De)){let o=`  ${n[0]} `;for(let r=0;r+3<=o.length;r+=1)t.add(o.slice(r,r+3))}return t}i(O,"trigrams");function g(e,t){let n=O(e),o=O(t);if(n.size===0||o.size===0)return 0;let r=0,[s,a]=n.size<=o.size?[n,o]:[o,n];for(let l of s)a.has(l)&&(r+=1);let c=n.size+o.size-r;return c===0?0:r/c}i(g,"trigramSimilarity");var R=.3,je=15;function w(e){let n=e.replace(/<!--[\s\S]*?-->/g," ").replace(/```[\s\S]*?```/g," "),o=[];for(let r of n.split(`
 `)){let s=r.trim();if(s.length===0||s.startsWith("#"))continue;let a=s.replace(/^(?:[-*+]|\d+[.)])\s+/,"").trim();if(a.length!==0)for(let c of a.split(/(?<=[.!?])\s+(?=[A-Z(`"'*])/)){let l=c.trim();l.length>=je&&o.push(l)}}return o}i(w,"splitDirectives");function Z(e,t,n=R){let o=w(e);if(o.length===0)return{redundant:!1,coverage:0,matches:[],novel:[]};let r=t.flatMap(c=>w(c.content).map(l=>({text:l,id:c.id,source:c.source}))),s=[],a=[];for(let c of o){let l=null;for(let u of r){let f=g(c,u.text);f<n||(l===null||f>l.similarity)&&(l={directive:c,matchedDirective:u.text,matchedId:u.id,matchedSource:u.source,similarity:f})}l?s.push(l):a.push(c)}return{redundant:a.length===0,coverage:s.length/o.length,matches:s,novel:a}}i(Z,"assessRedundancy");function $e(e,t=R){let n=[];for(let o=0;o<e.length;o+=1){let r=Z(e[o].content,e.slice(0,o),t);r.matches.length>0&&n.push({instruction:e[o],verdict:r})}return n}i($e,"findInternalRedundancy");var Q=.5,Me=/-(?=[0-9a-f]{6}$)[0-9a-f]*[0-9][0-9a-f]*$/;function L(e){return e.trim().toLowerCase().replace(/^skills--/,"").replace(/[^a-z0-9]+/g,"-").replace(/^-+|-+$/g,"").replace(Me,"").replace(/-+$/,"")}i(L,"skillNameStem");function He(e,t,n=Q){let o=L(e.name);for(let s of t)if(o.length>0&&L(s.name)===o)return{name:s.name,similarity:g(e.content,s.content),reason:"name-stem"};let r=null;for(let s of t){let a=g(e.content,s.content);a<n||(r===null||a>r.similarity)&&(r={name:s.name,similarity:a,reason:"body"})}return r}i(He,"assessSkillDuplication");var Be=new Set(["sh","bash","zsh","fish"]),Ue=new Set(["mjs","cjs","js","ts","py","rb","pl"]),Ke=/(?:^|[\s(`'"])(?:sudo\s+)?(?:node|npx|pnpm|npm|yarn|bun|deno|tsx|ts-node|vitest|jest|python3?|python|ruby|perl|bash|sh|zsh|exec|run|dlx)\s+(?:[-\w=@./]+\s+)*$/,ze=/(?:\.\/)?(?:[\w.@-]+\/)+[\w.@-]+\.\w+/g;function Fe(e){return e.replace(/```[\w-]*\n?/g," ").replace(/`/g," ")}i(Fe,"stripCodeMarkers");function We(e){let t=e.trim();return t=t.replace(/[.,;:)\]]+$/,""),t.startsWith("./")&&(t=t.slice(2)),t.length===0||t.startsWith("/")||t.startsWith("~")||t.startsWith("..")||/[<>{}*?]/.test(t)?null:t}i(We,"normalise");function ee(e){let t=Fe(e),n=new Set,o=[];for(let r of t.matchAll(ze)){let s=r[0],a=r.index??0,c=t.slice(Math.max(0,a-1),a);if(c.length>0&&!/[\s(['"=]/.test(c))continue;let l=We(s);if(!l)continue;let u=l.slice(l.lastIndexOf(".")+1).toLowerCase();(Be.has(u)||Ue.has(u)&&(s.startsWith("./")||Ke.test(t.slice(t.lastIndexOf(`
 `,Math.max(0,a-1))+1,a))))&&(n.has(l)||(n.add(l),o.push({reference:s,path:l})))}return o}i(ee,"extractCommandReferences");function Ye(e,t){return ee(e).filter(n=>!t(n.path))}i(Ye,"findUnresolvableReferences");import Yt from"gray-matter";import{z as te}from"zod";var Xt=te.object({type:te.enum(["instruction","prompt"])});export{se as a,ie as b,_ as c,M as d,fe as e,he as f,ge as g,ye as h,xe as i,_e as j,Ee as k,Se as l,y as m,ve as n,Ae as o,Te as p,be as q,ke as r,b as s,G as t,Ce as u,we as v,Re as w,Ie as x,C as y,Ne as z,J as A,Pe as B,w as C,Z as D,$e as E,He as F,Ye as G};

package/dist/{chunk-SV73ASP2.js → chunk-43VKTVM3.js}

@@ -1 +1 @@
-import{d as h,e as g}from"./chunk-RRM4HQG4.js";import{e as p,g as f}from"./chunk-ZZHIBOMT.js";import{a as o}from"./chunk-ZLLER3HP.js";var u="openCandidate";function K(n,s,a){let t=(n??{})[u];if(!t||typeof t!="object"||Array.isArray(t))return null;let e=t;return e.executionKey!==a||typeof e.content!="string"||!e.content||typeof e.hash!="string"||!e.hash||typeof e.priorHash!="string"||e.priorHash!==s?null:{...a?{executionKey:a}:{},...typeof e.observedKey=="string"?{observedKey:e.observedKey}:{},content:e.content,hash:e.hash,priorHash:e.priorHash,plannedPairs:typeof e.plannedPairs=="number"?e.plannedPairs:0,updatedAt:typeof e.updatedAt=="string"?e.updatedAt:""}}o(K,"resolveOpenCandidate");function C(n){let{plannedPairs:s,carriedPairs:a,liveCases:t}=n;if(t<=0)return 0;let e=s-a;return e<=0?0:Math.ceil(e/t)}o(C,"owedPasses");var d=100;async function x(n){let{evalTargetId:s,candidateContentHash:a}=n,t=[];for(let e=0;;e+=d){let i=await p(`/api/eval-targets/${s}/runs?candidateContentHash=${encodeURIComponent(a)}&limit=${d}&offset=${e}`);if(!Array.isArray(i)||i.length===0||(t.push(...i.filter(r=>{if(r.outputs?.executionKey!==n.executionKey)return!1;if(!n.executionKey)return!0;if(!n.observedKey)return!1;let l=h([r]);if(!l)return!1;let[y,c]=JSON.parse(l),[b,m]=JSON.parse(n.observedKey);return y===b&&(c===null||c===m)})),i.length<d))break}return g(t)}o(x,"fetchCarriedPairs");async function A(n){let{evalTargetId:s,meta:a,blob:t,log:e,tag:i}=n;try{let r={...a};t?r[u]=t:delete r[u],await f(`/api/eval-targets/${s}`,{metadata:r}),e(t?`${i}: banked candidate ${t.hash.slice(0,8)} \u2014 the next run resumes instead of restarting`:`${i}: cleared the banked candidate (this cycle reached a decision)`)}catch(r){e(`${i}: could not update the candidate bank (${r instanceof Error?r.message:String(r)})`)}}o(A,"bankOpenCandidate");function O(n){return{...n.executionKey?{executionKey:n.executionKey}:{},...n.observedKey?{observedKey:n.observedKey}:{},content:n.content,hash:n.hash,priorHash:n.priorHash,plannedPairs:n.plannedPairs,updatedAt:n.now.toISOString()}}o(O,"buildOpenCandidate");export{u as a,K as b,C as c,x as d,A as e,O as f};
+import{d as h,e as g}from"./chunk-SCQK52SL.js";import{e as p,g as f}from"./chunk-76UKAFNM.js";import{a as o}from"./chunk-ZLLER3HP.js";var u="openCandidate";function K(n,s,a){let t=(n??{})[u];if(!t||typeof t!="object"||Array.isArray(t))return null;let e=t;return e.executionKey!==a||typeof e.content!="string"||!e.content||typeof e.hash!="string"||!e.hash||typeof e.priorHash!="string"||e.priorHash!==s?null:{...a?{executionKey:a}:{},...typeof e.observedKey=="string"?{observedKey:e.observedKey}:{},content:e.content,hash:e.hash,priorHash:e.priorHash,plannedPairs:typeof e.plannedPairs=="number"?e.plannedPairs:0,updatedAt:typeof e.updatedAt=="string"?e.updatedAt:""}}o(K,"resolveOpenCandidate");function C(n){let{plannedPairs:s,carriedPairs:a,liveCases:t}=n;if(t<=0)return 0;let e=s-a;return e<=0?0:Math.ceil(e/t)}o(C,"owedPasses");var d=100;async function x(n){let{evalTargetId:s,candidateContentHash:a}=n,t=[];for(let e=0;;e+=d){let i=await p(`/api/eval-targets/${s}/runs?candidateContentHash=${encodeURIComponent(a)}&limit=${d}&offset=${e}`);if(!Array.isArray(i)||i.length===0||(t.push(...i.filter(r=>{if(r.outputs?.executionKey!==n.executionKey)return!1;if(!n.executionKey)return!0;if(!n.observedKey)return!1;let l=h([r]);if(!l)return!1;let[y,c]=JSON.parse(l),[b,m]=JSON.parse(n.observedKey);return y===b&&(c===null||c===m)})),i.length<d))break}return g(t)}o(x,"fetchCarriedPairs");async function A(n){let{evalTargetId:s,meta:a,blob:t,log:e,tag:i}=n;try{let r={...a};t?r[u]=t:delete r[u],await f(`/api/eval-targets/${s}`,{metadata:r}),e(t?`${i}: banked candidate ${t.hash.slice(0,8)} \u2014 the next run resumes instead of restarting`:`${i}: cleared the banked candidate (this cycle reached a decision)`)}catch(r){e(`${i}: could not update the candidate bank (${r instanceof Error?r.message:String(r)})`)}}o(A,"bankOpenCandidate");function O(n){return{...n.executionKey?{executionKey:n.executionKey}:{},...n.observedKey?{observedKey:n.observedKey}:{},content:n.content,hash:n.hash,priorHash:n.priorHash,plannedPairs:n.plannedPairs,updatedAt:n.now.toISOString()}}o(O,"buildOpenCandidate");export{u as a,K as b,C as c,x as d,A as e,O as f};

package/dist/{chunk-MX34N7Y5.js → chunk-44TYRMIS.js}

@@ -1 +1 @@
-import{a}from"./chunk-ZZHIBOMT.js";import{a as e}from"./chunk-ZLLER3HP.js";function i(){return process.env.RULEMETRIC_API_URL??a}e(i,"apiUrl");function u(){let n=process.env.RULEMETRIC_API_KEY??process.env.RULEMETRIC_ACCESS_TOKEN;if(!n)throw new Error("not signed in \u2014 run `rulemetric auth login`");return n}e(u,"authToken");async function t(n,o){let r=await fetch(`${i()}${n}`,{...o,headers:{Authorization:`Bearer ${u()}`,"Content-Type":"application/json",...o?.headers??{}},signal:AbortSignal.timeout(1e4)});if(!r.ok)throw new Error(`${r.status} ${r.statusText}`);return await r.json()}e(t,"request");function c({limit:n}){return t(`/api/launch-jobs?limit=${n}`)}e(c,"fetchLaunchJobs");function h(n){return t("/api/launch-jobs",{method:"POST",body:JSON.stringify(n)})}e(h,"createLaunchJob");function l(n){return t(`/api/launch-jobs/${n}/cancel`,{method:"POST"})}e(l,"cancelLaunchJob");function p(){return t("/api/launchers")}e(p,"listLaunchers");export{c as a,h as b,l as c,p as d};
+import{a}from"./chunk-76UKAFNM.js";import{a as e}from"./chunk-ZLLER3HP.js";function i(){return process.env.RULEMETRIC_API_URL??a}e(i,"apiUrl");function u(){let n=process.env.RULEMETRIC_API_KEY??process.env.RULEMETRIC_ACCESS_TOKEN;if(!n)throw new Error("not signed in \u2014 run `rulemetric auth login`");return n}e(u,"authToken");async function t(n,o){let r=await fetch(`${i()}${n}`,{...o,headers:{Authorization:`Bearer ${u()}`,"Content-Type":"application/json",...o?.headers??{}},signal:AbortSignal.timeout(1e4)});if(!r.ok)throw new Error(`${r.status} ${r.statusText}`);return await r.json()}e(t,"request");function c({limit:n}){return t(`/api/launch-jobs?limit=${n}`)}e(c,"fetchLaunchJobs");function h(n){return t("/api/launch-jobs",{method:"POST",body:JSON.stringify(n)})}e(h,"createLaunchJob");function l(n){return t(`/api/launch-jobs/${n}/cancel`,{method:"POST"})}e(l,"cancelLaunchJob");function p(){return t("/api/launchers")}e(p,"listLaunchers");export{c as a,h as b,l as c,p as d};

package/dist/{chunk-HAYO6JYI.js → chunk-4FJYVYCU.js}

@@ -1 +1 @@
-import{a as s,b as n,c as a,d as g}from"./chunk-3AV3GTMP.js";import{e as o}from"./chunk-ZZHIBOMT.js";import{a as e}from"./chunk-ZLLER3HP.js";var i=!1;async function c(){if(process.env.RULEMETRIC_ORG_ID)return;let r=s();if(!(r&&!g(r)))try{let t=await o("/api/profile");if(!t.activeOrgId){a();return}if(r&&t.activeOrgId===r.orgId){n({orgId:r.orgId,orgSlug:r.orgSlug,syncedAt:Date.now()});return}let d=(await o("/api/orgs")).find(l=>l.id===t.activeOrgId);n({orgId:t.activeOrgId,orgSlug:d?.slug,syncedAt:Date.now()})}catch{}}e(c,"syncActiveOrgFromServer");async function p(){i||(i=!0,await c())}e(p,"maybeRefreshActiveOrg");async function v(){i=!0,await c()}e(v,"bootstrapActiveOrg");function O(){i=!1}e(O,"_resetRefreshAttempt");export{p as a,v as b,O as c};
+import{a as s,b as n,c as a,d as g}from"./chunk-3AV3GTMP.js";import{e as o}from"./chunk-76UKAFNM.js";import{a as e}from"./chunk-ZLLER3HP.js";var i=!1;async function c(){if(process.env.RULEMETRIC_ORG_ID)return;let r=s();if(!(r&&!g(r)))try{let t=await o("/api/profile");if(!t.activeOrgId){a();return}if(r&&t.activeOrgId===r.orgId){n({orgId:r.orgId,orgSlug:r.orgSlug,syncedAt:Date.now()});return}let d=(await o("/api/orgs")).find(l=>l.id===t.activeOrgId);n({orgId:t.activeOrgId,orgSlug:d?.slug,syncedAt:Date.now()})}catch{}}e(c,"syncActiveOrgFromServer");async function p(){i||(i=!0,await c())}e(p,"maybeRefreshActiveOrg");async function v(){i=!0,await c()}e(v,"bootstrapActiveOrg");function O(){i=!1}e(O,"_resetRefreshAttempt");export{p as a,v as b,O as c};

package/dist/{chunk-TTSDPDRH.js → chunk-5BT375C7.js}

@@ -1,4 +1,4 @@
-import{e as d}from"./chunk-QWCEZEDO.js";import{a as R,g as v,h as T,j as I,m as C}from"./chunk-OZQNSELV.js";import{d as g}from"./chunk-SBGXIPRN.js";import{e as _}from"./chunk-ZZHIBOMT.js";import{g as m,l as k}from"./chunk-PZNWH4VQ.js";import{a as i}from"./chunk-ZLLER3HP.js";import{execSync as j}from"node:child_process";import{existsSync as p,readFileSync as y}from"node:fs";import{homedir as f}from"node:os";import{join as a}from"node:path";var P=["install","auth","hooks","worker"],O=["judge"],M=[...P,...O];function $(){return{projectPath:process.cwd(),configDir:k()}}i($,"defaultContext");var A=a(f(),".mitmproxy","mitmproxy-ca-cert.pem");function S(e){if(!e)return"";if(e.length<=8)return"*".repeat(e.length);let t=e.slice(0,4),n=e.slice(-4);return`${t}${"*".repeat(Math.max(4,e.length-8))}${n}`}i(S,"maskToken");function te(e,t){let n=e===null?[]:e.replace(/\n$/,"").split(`
+import{e as d}from"./chunk-QWCEZEDO.js";import{a as R,g as v,h as T,j as I,m as C}from"./chunk-OZQNSELV.js";import{d as g}from"./chunk-GPBN4ROE.js";import{e as _}from"./chunk-76UKAFNM.js";import{g as m,l as k}from"./chunk-PZNWH4VQ.js";import{a as i}from"./chunk-ZLLER3HP.js";import{execSync as j}from"node:child_process";import{existsSync as p,readFileSync as y}from"node:fs";import{homedir as f}from"node:os";import{join as a}from"node:path";var P=["install","auth","hooks","worker"],O=["judge"],M=[...P,...O];function $(){return{projectPath:process.cwd(),configDir:k()}}i($,"defaultContext");var A=a(f(),".mitmproxy","mitmproxy-ca-cert.pem");function S(e){if(!e)return"";if(e.length<=8)return"*".repeat(e.length);let t=e.slice(0,4),n=e.slice(-4);return`${t}${"*".repeat(Math.max(4,e.length-8))}${n}`}i(S,"maskToken");function te(e,t){let n=e===null?[]:e.replace(/\n$/,"").split(`
 `),o=t===null?[]:t.replace(/\n$/,"").split(`
 `),s=[],u=Math.max(n.length,o.length);for(let c=0;c<u;c++){let l=n[c],r=o[c];l===r?s.push(` ${l}`):(l!==void 0&&s.push(`-${l}`),r!==void 0&&s.push(`+${r}`))}return s.join(`
 `)}i(te,"renderDiff");function x(e,t){let n=a(e,"env");if(!p(n))return null;try{let s=y(n,"utf-8").match(new RegExp(`^${t}=(.+)$`,"m"));return s?s[1]:null}catch{return null}}i(x,"readEnvValue");async function b(){if(!m())return!1;try{return(await Promise.race([_("/api/setup/status"),new Promise((t,n)=>setTimeout(()=>n(new Error("setup/status timeout")),2500))]))?.workerActive===!0}catch{return!1}}i(b,"workerActive");function E(e,t){let n=a(e,".claude",t),o={},s=null;if(p(n))try{s=y(n,"utf-8").replace(/\n$/,""),o=JSON.parse(s)}catch{o={}}return{settings:o,before:s}}i(E,"readProjectSettings");function D(e){let{settings:t,before:n}=E(e,"settings.local.json");return v(t),{before:n,after:JSON.stringify(t,null,2)}}i(D,"projectHooksSettings");function N(e){let{settings:t,before:n}=E(e,"settings.local.json"),o={...t.env??{}};return o.HTTPS_PROXY=`http://localhost:${d}`,o.NO_PROXY=I,o.NODE_EXTRA_CA_CERTS=A,{before:n,after:JSON.stringify({...t,env:o},null,2)}}i(N,"projectProxyEnvSettings");function U(e){let{settings:t,before:n}=E(e,"settings.json");if(n===null)return null;let{removed:o}=C(t,d);return o?{before:n,after:JSON.stringify(t,null,2)}:null}i(U,"projectSharedSettingsMigration");var Y=a(f(),"Library","LaunchAgents","com.rulemetric.worker.plist"),G=a(f(),".config","systemd","user","rulemetric-worker.service");function h(e){j(e,{stdio:"inherit"})}i(h,"runCommand");var L=[{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",async detect(){return m()||await b()},async plan(){return{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",artifacts:[{path:"npm install -g @rulemetric/cli",kind:"command",before:null,after:"rulemetric --version   # verifies the binary is on PATH"}]}},async apply(){h("npm install -g @rulemetric/cli")}},{id:"auth",title:"Authenticate",command:"rulemetric auth login",async detect(){return m()},async plan(e){let t=a(e.configDir,"env"),n=null;if(p(t))try{n=X(y(t,"utf-8"))}catch{n="(unreadable)"}let o=x(e.configDir,"RULEMETRIC_ACCESS_TOKEN"),u=[`RULEMETRIC_ACCESS_TOKEN=${S(o||"<token-from-login>")}`],c=x(e.configDir,"RULEMETRIC_API_KEY");c&&u.push(`RULEMETRIC_API_KEY=${S(c)}`);let l=process.env.RULEMETRIC_API_URL;l&&u.push(`RULEMETRIC_API_URL=${l}`);let r=u.join(`

package/dist/{chunk-N26UEHI3.js → chunk-5Z3H5JHK.js}

@@ -1 +1 @@
-import{a as i}from"./chunk-MX34N7Y5.js";import{a as f}from"./chunk-ZLLER3HP.js";import{useEffect as J,useState as u,useCallback as m,useRef as s}from"react";var y=5e3;function T(a=20){let[d,b]=u([]),[g,l]=u(null),[h,p]=u(!0),r=s(!1),e=s(null),c=s(0),t=m(async()=>{let o=++c.current;try{let n=await i({limit:a});if(r.current||o!==c.current)return;b(n),l(null)}catch(n){if(r.current||o!==c.current)return;l(n instanceof Error?n.message:String(n))}finally{if(r.current||o!==c.current)return;p(!1),e.current=setTimeout(()=>{t()},y)}},[a]),L=m(()=>{e.current&&clearTimeout(e.current),t()},[t]);return J(()=>(r.current=!1,t(),()=>{r.current=!0,e.current&&clearTimeout(e.current)}),[t]),{jobs:d,error:g,loading:h,refetch:L}}f(T,"useLaunchJobs");export{T as a};
+import{a as i}from"./chunk-44TYRMIS.js";import{a as f}from"./chunk-ZLLER3HP.js";import{useEffect as J,useState as u,useCallback as m,useRef as s}from"react";var y=5e3;function T(a=20){let[d,b]=u([]),[g,l]=u(null),[h,p]=u(!0),r=s(!1),e=s(null),c=s(0),t=m(async()=>{let o=++c.current;try{let n=await i({limit:a});if(r.current||o!==c.current)return;b(n),l(null)}catch(n){if(r.current||o!==c.current)return;l(n instanceof Error?n.message:String(n))}finally{if(r.current||o!==c.current)return;p(!1),e.current=setTimeout(()=>{t()},y)}},[a]),L=m(()=>{e.current&&clearTimeout(e.current),t()},[t]);return J(()=>(r.current=!1,t(),()=>{r.current=!0,e.current&&clearTimeout(e.current)}),[t]),{jobs:d,error:g,loading:h,refetch:L}}f(T,"useLaunchJobs");export{T as a};

package/dist/{chunk-EWGE7YTQ.js → chunk-6ANODRDT.js}

@@ -1 +1 @@
-import{b as u,c as d}from"./chunk-HFSISGCJ.js";import{b as g}from"./chunk-E7RH43OS.js";import{e as m,f as p,h as l}from"./chunk-ZZHIBOMT.js";import{a as s}from"./chunk-ZLLER3HP.js";async function x(f,R){let r=s(e=>R?.logger?.info?.(e)??console.log(e),"log"),{evalTargetId:n}=f,i=`[cron_refine_cases] ${n}`,t=await m(`/api/eval-targets/${n}`);if(!(t.metadata?.autoRefineCases===!0))return r(`${i} not opted in (metadata.autoRefineCases)`),{skipped:"not-opted-in",reason:`"${t.name}" has not opted into automatic case refinement. Set metadata.autoRefineCases=true to let the loop cull cases that carry no treatment signal, or run \`rulemetric evals refine-cases ${n}\` by hand.`,targetName:t.name};let C=await m(`/api/eval-targets/${n}/judge-calibration`),a=u(C.caseDiscrimination);if(a.refused)return r(`${i} ${a.refused}: ${a.summary}`),{skipped:a.refused,reason:a.summary,targetName:t.name};r(`${i} ${a.summary}`);let o=await d(a,{generate:s(e=>g({content:t.content,name:t.name,feedback:e}),"generate"),createCase:s(e=>p(`/api/eval-targets/${n}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations}),"createCase"),retireCase:s(async(e,y)=>{await l(`/api/eval-targets/${n}/evals/${e}`,{retire:!0,retiredReason:y})},"retireCase"),log:r});r(`${i} ${o.summary}`);let{skipped:c,...$}=o;return{...$,targetName:t.name,...c?{skipped:c,reason:o.summary}:{}}}s(x,"cronRefineCases");export{x as a};
+import{b as u,c as d}from"./chunk-MXHAXY66.js";import{b as g}from"./chunk-OP4U3XBZ.js";import{e as m,f as p,h as l}from"./chunk-76UKAFNM.js";import{a as s}from"./chunk-ZLLER3HP.js";async function x(f,R){let r=s(e=>R?.logger?.info?.(e)??console.log(e),"log"),{evalTargetId:n}=f,i=`[cron_refine_cases] ${n}`,t=await m(`/api/eval-targets/${n}`);if(!(t.metadata?.autoRefineCases===!0))return r(`${i} not opted in (metadata.autoRefineCases)`),{skipped:"not-opted-in",reason:`"${t.name}" has not opted into automatic case refinement. Set metadata.autoRefineCases=true to let the loop cull cases that carry no treatment signal, or run \`rulemetric evals refine-cases ${n}\` by hand.`,targetName:t.name};let C=await m(`/api/eval-targets/${n}/judge-calibration`),a=u(C.caseDiscrimination);if(a.refused)return r(`${i} ${a.refused}: ${a.summary}`),{skipped:a.refused,reason:a.summary,targetName:t.name};r(`${i} ${a.summary}`);let o=await d(a,{generate:s(e=>g({content:t.content,name:t.name,feedback:e}),"generate"),createCase:s(e=>p(`/api/eval-targets/${n}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations}),"createCase"),retireCase:s(async(e,y)=>{await l(`/api/eval-targets/${n}/evals/${e}`,{retire:!0,retiredReason:y})},"retireCase"),log:r});r(`${i} ${o.summary}`);let{skipped:c,...$}=o;return{...$,targetName:t.name,...c?{skipped:c,reason:o.summary}:{}}}s(x,"cronRefineCases");export{x as a};

package/dist/{chunk-ZZHIBOMT.js → chunk-76UKAFNM.js}

@@ -1 +1 @@
-import{l as E}from"./chunk-TWPG3TEQ.js";import{a as p,e as w}from"./chunk-PZNWH4VQ.js";import{a as o}from"./chunk-ZLLER3HP.js";import{fileURLToPath as v}from"node:url";import{existsSync as A,readFileSync as b}from"node:fs";import{dirname as L,join as x}from"node:path";function C(){let t=L(v(import.meta.url));for(let e=0;e<8;e++){let s=x(t,"package.json");if(A(s))try{let i=JSON.parse(b(s,"utf8"));if(i.name==="@rulemetric/cli"&&i.version)return i.version}catch{}let n=L(t);if(n===t)break;t=n}return"0.0.0"}o(C,"readCliVersion");var O=C(),I="https://rulemetric.com",l=class extends Error{constructor(s,n){super(n);this.statusCode=s;this.name="ApiError"}static{o(this,"ApiError")}};function S(){let t=p(),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=I}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: ${e}`);return e}o(S,"getBaseUrl");async function a(t,e={}){let n=`${S()}${t}`,i={"Content-Type":"application/json","User-Agent":`rulemetric-cli/${O}`,...e.headers},m=p(),d=process.env.RULEMETRIC_API_KEY||m.apiKey||w()?.accessToken||m.accessToken;d&&(i.Authorization=`Bearer ${d}`);let h=new URL(n),f=h.hostname==="localhost"||h.hostname==="127.0.0.1",P=process.env.HTTPS_PROXY,R=process.env.HTTP_PROXY;f&&(delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY);let u=e.timeoutMs?new AbortController:void 0,g=!1,y=u?setTimeout(()=>{g=!0,u.abort()},e.timeoutMs):void 0,r;try{r=await fetch(n,{method:e.method??"GET",headers:i,body:e.body?JSON.stringify(e.body):void 0,signal:u?.signal})}catch(T){if(g){let U=new URL(n).host,_=Math.round(e.timeoutMs/1e3);throw new l(408,`Request to ${U} timed out after ${_}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 T}finally{y&&clearTimeout(y),f&&(P!==void 0&&(process.env.HTTPS_PROXY=P),R!==void 0&&(process.env.HTTP_PROXY=R))}let c;if(r.headers.get("content-type")?.includes("application/json")?c=await r.json():c=await r.text(),!r.ok){let{message:T}=E(r.status,c);throw new l(r.status,T)}return c}o(a,"api");var j=o(t=>a(t),"apiGet"),F=o((t,e)=>a(t,{method:"POST",body:e}),"apiPost"),B=o((t,e)=>a(t,{method:"PUT",body:e}),"apiPut"),D=o((t,e)=>a(t,{method:"PATCH",body:e}),"apiPatch"),G=o(t=>a(t,{method:"DELETE"}),"apiDelete");export{I as a,l as b,S as c,a as d,j as e,F as f,B as g,D as h,G as i};
+import{l as E}from"./chunk-42A4PUW7.js";import{a as p,e as w}from"./chunk-PZNWH4VQ.js";import{a as o}from"./chunk-ZLLER3HP.js";import{fileURLToPath as v}from"node:url";import{existsSync as A,readFileSync as b}from"node:fs";import{dirname as L,join as x}from"node:path";function C(){let t=L(v(import.meta.url));for(let e=0;e<8;e++){let s=x(t,"package.json");if(A(s))try{let i=JSON.parse(b(s,"utf8"));if(i.name==="@rulemetric/cli"&&i.version)return i.version}catch{}let n=L(t);if(n===t)break;t=n}return"0.0.0"}o(C,"readCliVersion");var O=C(),I="https://rulemetric.com",l=class extends Error{constructor(s,n){super(n);this.statusCode=s;this.name="ApiError"}static{o(this,"ApiError")}};function S(){let t=p(),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=I}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: ${e}`);return e}o(S,"getBaseUrl");async function a(t,e={}){let n=`${S()}${t}`,i={"Content-Type":"application/json","User-Agent":`rulemetric-cli/${O}`,...e.headers},m=p(),d=process.env.RULEMETRIC_API_KEY||m.apiKey||w()?.accessToken||m.accessToken;d&&(i.Authorization=`Bearer ${d}`);let h=new URL(n),f=h.hostname==="localhost"||h.hostname==="127.0.0.1",P=process.env.HTTPS_PROXY,R=process.env.HTTP_PROXY;f&&(delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY);let u=e.timeoutMs?new AbortController:void 0,g=!1,y=u?setTimeout(()=>{g=!0,u.abort()},e.timeoutMs):void 0,r;try{r=await fetch(n,{method:e.method??"GET",headers:i,body:e.body?JSON.stringify(e.body):void 0,signal:u?.signal})}catch(T){if(g){let U=new URL(n).host,_=Math.round(e.timeoutMs/1e3);throw new l(408,`Request to ${U} timed out after ${_}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 T}finally{y&&clearTimeout(y),f&&(P!==void 0&&(process.env.HTTPS_PROXY=P),R!==void 0&&(process.env.HTTP_PROXY=R))}let c;if(r.headers.get("content-type")?.includes("application/json")?c=await r.json():c=await r.text(),!r.ok){let{message:T}=E(r.status,c);throw new l(r.status,T)}return c}o(a,"api");var j=o(t=>a(t),"apiGet"),F=o((t,e)=>a(t,{method:"POST",body:e}),"apiPost"),B=o((t,e)=>a(t,{method:"PUT",body:e}),"apiPut"),D=o((t,e)=>a(t,{method:"PATCH",body:e}),"apiPatch"),G=o(t=>a(t,{method:"DELETE"}),"apiDelete");export{I as a,l as b,S as c,a as d,j as e,F as f,B as g,D as h,G as i};

package/dist/{chunk-IOX4VJPR.js → chunk-7MTYUTUN.js}

@@ -1,4 +1,4 @@
-import{f as u}from"./chunk-UJ7EPLKC.js";import{a as o}from"./chunk-ZLLER3HP.js";var l=`You are a meta-judge evaluating the quality of an eval run \u2014 not the code output, but whether the test cases and grading themselves are rigorous and meaningful.
+import{f as u}from"./chunk-FBSZ6KCM.js";import{a as o}from"./chunk-ZLLER3HP.js";var l=`You are a meta-judge evaluating the quality of an eval run \u2014 not the code output, but whether the test cases and grading themselves are rigorous and meaningful.
 
 You will receive a summary of eval cases, their expectations, and grading results.
 

package/dist/{chunk-JLBQKV4V.js → chunk-AGEZAKCC.js}

@@ -1,3 +1,3 @@
-import{h as R}from"./chunk-FJX7WDPG.js";import{a as S,b as v,c as $,f as w,g as A}from"./chunk-SIGZEQ3K.js";import{e as y,f}from"./chunk-ZZHIBOMT.js";import{a as c}from"./chunk-ZLLER3HP.js";import{execFileSync as x}from"node:child_process";import{openSync as P,readSync as C,closeSync as j,readFileSync as E}from"node:fs";function I(l){let t;try{t=P(l,"r");let o=Buffer.alloc(64*1024),s=C(t,o,0,o.length,0),i=o.subarray(0,s).toString("utf-8");for(let r of i.split(`
+import{h as R}from"./chunk-I6ZQOMZS.js";import{a as S,b as v,c as $,f as w,g as A}from"./chunk-SIGZEQ3K.js";import{e as y,f}from"./chunk-76UKAFNM.js";import{a as c}from"./chunk-ZLLER3HP.js";import{execFileSync as x}from"node:child_process";import{openSync as P,readSync as C,closeSync as j,readFileSync as E}from"node:fs";function I(l){let t;try{t=P(l,"r");let o=Buffer.alloc(64*1024),s=C(t,o,0,o.length,0),i=o.subarray(0,s).toString("utf-8");for(let r of i.split(`
 `))if(r.startsWith("{"))try{let a=JSON.parse(r).cwd;if(typeof a=="string"&&a.startsWith("/"))return a}catch{}return null}catch{return null}finally{t!==void 0&&j(t)}}c(I,"readTranscriptCwd");function W(l){let t=c(r=>{try{return x("git",["-C",l,...r],{encoding:"utf-8",stdio:["ignore","pipe","ignore"],timeout:5e3}).trim()||null}catch{return null}},"git");if(!t(["rev-parse","--show-toplevel"]))return null;let s=t(["rev-parse","--git-dir"]),i=t(["rev-parse","--git-common-dir"]);return{gitRemote:t(["remote","get-url","origin"]),gitRootCommit:t(["rev-list","--max-parents=0","HEAD"])?.split(`
 `).pop()?.trim()??null,gitBranch:t(["branch","--show-current"]),isWorktree:!!s&&!!i&&s!==i}}c(W,"gitAnchors");async function O(l,t,o=v(W,I)){let s=Math.min(l.limit??200,1e3),i=await R(()=>y(`/api/sessions/unattributed?limit=${s}`),{log:c(e=>t.logger.info(e),"log"),label:"sessions/unattributed",idempotent:!0}),r=i.sessions??[],a=i.total??r.length;if(r.length===0)return t.logger.info("no unattributed sessions"),{scanned:0,repaired:0,reimported:0,unrepairable:0,remaining:0};let m=0,g=0,u=0;for(let e of r){if(!e.externalSessionId){u++;continue}let n=A(e,o);if(!n){u++;continue}try{await f("/api/sessions",{tool:e.tool,externalSessionId:e.externalSessionId,projectPath:n.projectPath,...n.gitBranch?{gitBranch:n.gitBranch}:{},metadata:{...n.gitRemote?{gitRemote:n.gitRemote}:{},...n.gitRootCommit?{gitRootCommit:n.gitRootCommit}:{},isWorktree:n.isWorktree,reconciledAttribution:!0}}),m++}catch(d){u++,t.logger.info(`reconcile failed for ${e.externalSessionId}: ${d.message}`);continue}let p=e.tool==="claude_code"?$(e.externalSessionId,o):null,b=w({tool:e.tool,pathWasJunk:!!e.projectPath&&S(e.projectPath),endedAt:e.endedAt??null,transcript:p?o.stat(p):null});if(b.reimport&&p)try{let d=E(p,"utf-8");await f(`/api/sessions/${e.id}/reimport`,{content:d,tool:"claude_code"}),g++}catch(d){t.logger.info(`reimport failed for ${e.externalSessionId} (${b.reason}): ${d.message}`)}}let h=a-m;return t.logger.info(`reconciled attribution: scanned=${r.length} repaired=${m} reimported=${g} unrepairable=${u} remaining=${h} (of ${a} total)`),{scanned:r.length,repaired:m,reimported:g,unrepairable:u,remaining:h}}c(O,"cronReconcileAttribution");export{O as a};

package/dist/{chunk-XAOLOR4Q.js → chunk-AIYFT4P6.js}

@@ -1,4 +1,4 @@
-import{e as b,g as y}from"./chunk-GN46S7T4.js";import{e as M,f as w}from"./chunk-ZZHIBOMT.js";import{a as u}from"./chunk-ZLLER3HP.js";var $={anthropic:["https://platform.claude.com/docs/en/about-claude/pricing.md"],openai:["https://platform.openai.com/docs/pricing"]},R=.7,v=16e3,x=["You extract LLM API pricing from a provider pricing page.","You will be given the page text in <data> tags and, in this system prompt,","a list of model ids to find. For EACH requested model id, return its","published USD price per 1,000,000 tokens.","","Rules:","- Only report a model if you can find its rate ON THE PAGE. If it is not on","  the page, set found=false and omit the numbers \u2014 never guess from memory.","- inputPerM and outputPerM are the standard (non-batch, non-priority) rates.","- cacheReadPerM: the cached/discounted-input rate if published, else omit.","- For Anthropic models ONLY, if cache-write rates are not explicitly listed,","  compute them as cacheWrite5mPerM = inputPerM * 1.25 and","  cacheWrite1hPerM = inputPerM * 2.0. For all other providers, omit them.","- confidence is 0..1: how sure you are the numbers are correct for THIS id.","- contextWindow in tokens if stated, else omit.","","Respond with ONLY a JSON object of the form:",'{"rates":[{"modelId":"...","found":true,"confidence":0.9,"inputPerM":5,"outputPerM":25,"cacheReadPerM":0.5,"contextWindow":1000000}]}'].join(`
+import{e as b,g as y}from"./chunk-N4YPDZJF.js";import{e as M,f as w}from"./chunk-76UKAFNM.js";import{a as u}from"./chunk-ZLLER3HP.js";var $={anthropic:["https://platform.claude.com/docs/en/about-claude/pricing.md"],openai:["https://platform.openai.com/docs/pricing"]},R=.7,v=16e3,x=["You extract LLM API pricing from a provider pricing page.","You will be given the page text in <data> tags and, in this system prompt,","a list of model ids to find. For EACH requested model id, return its","published USD price per 1,000,000 tokens.","","Rules:","- Only report a model if you can find its rate ON THE PAGE. If it is not on","  the page, set found=false and omit the numbers \u2014 never guess from memory.","- inputPerM and outputPerM are the standard (non-batch, non-priority) rates.","- cacheReadPerM: the cached/discounted-input rate if published, else omit.","- For Anthropic models ONLY, if cache-write rates are not explicitly listed,","  compute them as cacheWrite5mPerM = inputPerM * 1.25 and","  cacheWrite1hPerM = inputPerM * 2.0. For all other providers, omit them.","- confidence is 0..1: how sure you are the numbers are correct for THIS id.","- contextWindow in tokens if stated, else omit.","","Respond with ONLY a JSON object of the form:",'{"rates":[{"modelId":"...","found":true,"confidence":0.9,"inputPerM":5,"outputPerM":25,"cacheReadPerM":0.5,"contextWindow":1000000}]}'].join(`
 `);async function C(f){try{let n=await fetch(f,{headers:{accept:"text/plain, text/markdown, text/html"}});if(!n.ok)return null;let i=await n.text();return i.length>v?i.slice(0,v):i}catch{return null}}u(C,"fetchPage");var L=u(async(f,n)=>{let i=await M("/api/pricing/coverage"),I=[...i.unpriced,...i.stale],m=new Map;for(let r of I)m.set(`${r.provider}::${r.modelId}`,r);let a=[...m.values()];if(a.length===0)return n.logger.info("[cron-refresh-pricing] coverage clean \u2014 nothing unpriced or stale"),{status:"clean",checked:i.unpriced.length+i.stale.length};if(b()==="none")return n.logger.warn(`[cron-refresh-pricing] ${a.length} model(s) need pricing but no LLM backend (install the claude CLI or set ANTHROPIC_API_KEY) \u2014 skipping`),{status:"no-llm",pending:a.length};let d=new Map;for(let r of a){let o=d.get(r.provider)??[];o.push(r.modelId),d.set(r.provider,o)}let t={written:0,unchanged:0,flagged:0,rejected:0,skipped:0},c=[];for(let[r,o]of d){let h=$[r];if(!h){n.logger.warn(`[cron-refresh-pricing] no pricing source configured for provider '${r}' (${o.length} model(s)) \u2014 skipping`),t.skipped+=o.length;continue}let p="",l="";for(let e of h){let s=await C(e);s&&(p+=`
 
 === ${e} ===

package/dist/{chunk-BVOXWAJO.js → chunk-B5X6HCBK.js}

@@ -1 +1 @@
-import{f as r}from"./chunk-ZZHIBOMT.js";import{a as i}from"./chunk-ZLLER3HP.js";async function p(e,t,n,o){await r(`/api/instruction-suggestions/${encodeURIComponent(e)}/events`,{eventType:"accepted",projectPath:t,instructionId:n,actor:o})}i(p,"logAccepted");async function a(e,t,n,o,s="redundant"){await r(`/api/instruction-suggestions/${encodeURIComponent(e)}/events`,{eventType:s,projectPath:t,instructionId:n,actor:o})}i(a,"logRefused");export{p as a,a as b};
+import{f as r}from"./chunk-76UKAFNM.js";import{a as i}from"./chunk-ZLLER3HP.js";async function p(e,t,n,o){await r(`/api/instruction-suggestions/${encodeURIComponent(e)}/events`,{eventType:"accepted",projectPath:t,instructionId:n,actor:o})}i(p,"logAccepted");async function a(e,t,n,o,s="redundant"){await r(`/api/instruction-suggestions/${encodeURIComponent(e)}/events`,{eventType:s,projectPath:t,instructionId:n,actor:o})}i(a,"logRefused");export{p as a,a as b};

package/dist/chunk-BAITMD76.js

@@ -0,0 +1 @@
+import{b as R,c as L,f as s,g as I,i as h}from"./chunk-W23OYQXI.js";import{c as ne}from"./chunk-EMLLGASZ.js";import{b as re}from"./chunk-IEQDLURA.js";import{d as te,e as oe,f as se}from"./chunk-KINIFUZQ.js";import{a as U}from"./chunk-V55ROUNP.js";import{a as F}from"./chunk-XMA6SWOT.js";import{a as W}from"./chunk-RAQLVO2X.js";import{a as $}from"./chunk-DHX4U6EC.js";import{a as q}from"./chunk-NYNHX7OX.js";import{b as J}from"./chunk-VXL5H3VZ.js";import{a as K}from"./chunk-QAMHZY2P.js";import{a as Y}from"./chunk-RMRVUI7B.js";import{a as O}from"./chunk-ZYGBABMZ.js";import{b as V}from"./chunk-5RBLIVBF.js";import{a as Q}from"./chunk-AIYFT4P6.js";import{a as D}from"./chunk-PZS55PW2.js";import{a as ee}from"./chunk-3IMRLHJJ.js";import{e as P}from"./chunk-EUHRTEVJ.js";import{b as G}from"./chunk-GKVHFXUQ.js";import{c as N}from"./chunk-LJRJ2RAC.js";import{a as z}from"./chunk-QRIJ3AHP.js";import{a as X}from"./chunk-CT6GJQIR.js";import{a as Z}from"./chunk-OVS7ZDJ4.js";import{a as x}from"./chunk-AGEZAKCC.js";import{a as j}from"./chunk-6ANODRDT.js";import{b as C}from"./chunk-PIVEJTNF.js";import{k as B}from"./chunk-POEMCITL.js";import{d as S}from"./chunk-GPBN4ROE.js";import{b as M,e as T,f as E}from"./chunk-76UKAFNM.js";import{a as l}from"./chunk-ZLLER3HP.js";var H=null,w=null,ce={process_eval:W,process_insights:$,process_launch:J,process_send_message:O,process_conversation:F,process_session_goal:U,distill_memories:G,memory_ab:K,process_run_cleanup:Y,cron_refine_cases:j,cron_retire_unused_skills:ee,cron_reconcile_attribution:x,cron_suggest_instructions:P,cron_refresh_skills:D,cron_refresh_pricing:Q,cron_eval_autorun:X,cron_auto_accept_suggestions:z,cron_instruction_training:Z,cron_instruction_evolution:B,process_judge_review:q,harbor_ab:N},ae=Object.keys(ce),me=new Set([]);function pe(e=process.env){return e.DATABASE_URL?ae:ae.filter(t=>!me.has(t))}l(pe,"availableCapabilities");var ge=3e4,_e=3e4,fe=3e4,ie=500,le=1e4,ve=`thin-agent-${process.pid}`,A=null;function ke(e,t,r,p){let d;try{let i=S();d={claude:i.claudeBin!=null,codex:i.codexBin!=null,pi:i.piBin!=null,codexModel:i.codexConfiguredModel}}catch{}let m;try{m=C()}catch{}let g=l(i=>E("/api/workers/heartbeat",{workerType:"cli-worker",clientName:ve,clientVersion:p,workerVersion:"0.1.0",currentJobId:null,inFlightCount:0,lastError:null,status:i,metadata:{capabilities:e,mode:"thin-agent",measurementProtocol:1,memorySnapshots:1,boundedAutorun:1,llmEngines:d,harborReady:m,machine:H?.summary()}}).then(c=>{let u=c?.advisory;u&&u.message!==A?(A=u.message,s(r,{event:"worker_advisory",code:u.code,message:u.message})):u||(A=null)}).catch(c=>{s(r,{event:"presence_heartbeat_error",error:c instanceof Error?c.message:String(c)})}),"send");g("alive");let f=setInterval(()=>g("alive"),fe),a=null,n=l(()=>a||(clearInterval(f),a=Promise.race([g("shutting_down"),new Promise(i=>setTimeout(i,2e3).unref())]),a),"sayGoodbye");return t.addEventListener("abort",()=>{n()},{once:!0}),{shutdown:n}}l(ke,"startPresenceHeartbeat");async function be(e,t,r=S()){let p=[...r.claudeBin?["claude"]:[],...r.codexBin?["codex"]:[],...r.piBin?["pi"]:[]],d=`capabilities=${encodeURIComponent(e.join(","))}&waitMs=${t}&measurementProtocol=1&memorySnapshots=1&boundedAutorun=1&measurementEngines=${encodeURIComponent(p.join(","))}`,m=await T(`/api/work/next?${d}`);return m==null||typeof m=="string"?null:m}l(be,"pollForWork");function he(e,t){return{logger:{info:l(r=>t(r),"info"),warn:l(r=>t(r),"warn"),error:l(r=>t(r),"error")},job:{attempts:e}}}l(he,"makeHelpers");async function we(e,t,r,p){let d=ce[e.kind],m=Date.now();if(s(r,{event:"claim",jobId:e.id,kind:e.kind,attempts:e.attempts}),!d){let n=`unknown kind: ${e.kind}`;s(r,{event:"fail",jobId:e.id,kind:e.kind,terminal:!0,error:n}),await E(`/api/work/${e.id}/fail`,{leaseToken:e.leaseToken,error:n,terminal:!0}).catch(()=>{});return}let g=l(()=>{s(r,{event:"heartbeat",jobId:e.id}),h(`/api/work/${e.id}/heartbeat`,{leaseToken:e.leaseToken},{budgetMs:Math.min(L,Math.max(0,t-1e3)),signal:p,log:r,event:"heartbeat",jobId:e.id}).then(n=>{n.ok||s(r,{event:"heartbeat_error",jobId:e.id,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})})},"beat"),f=setInterval(g,t);w={jobId:e.id,beat:g};let a;try{let n=he(e.attempts,r);a={kind:"complete",result:await d(e.payload,n)??{}}}catch(n){let i=H?.pausesSince(m)??[],c=se(i),u=n instanceof Error?n.message:String(n),v=V(n),o=!v&&oe(i);a={kind:"fail",message:c?`${c}; ${u}`:u,pausedDuringJob:!!c,terminal:v,hostSuspended:o}}try{if(a.kind==="complete"){s(r,{event:"complete",jobId:e.id,kind:e.kind,durationMs:Date.now()-m});let n=await h(`/api/work/${e.id}/complete`,{leaseToken:e.leaseToken,result:a.result},{budgetMs:R,signal:p,log:r,event:"complete_post",jobId:e.id});n.ok||s(r,{event:"complete_post_error",jobId:e.id,kind:e.kind,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}else{s(r,{event:"fail",jobId:e.id,kind:e.kind,durationMs:Date.now()-m,terminal:a.terminal,error:a.message,...a.pausedDuringJob?{pausedDuringJob:!0}:{},...a.hostSuspended?{hostSuspended:!0}:{}});let n=await h(`/api/work/${e.id}/fail`,{leaseToken:e.leaseToken,error:a.message,terminal:a.terminal,hostSuspended:a.hostSuspended},{budgetMs:R,signal:p,log:r,event:"fail_post",jobId:e.id});n.ok||s(r,{event:"fail_post_error",jobId:e.id,attempts:n.attempts,error:n.error instanceof Error?n.error.message:String(n.error)})}}finally{clearInterval(f),w=null}}l(we,"runOne");var de=new Set(["process_eval","process_conversation","process_insights","process_session_goal","distill_memories","memory_ab","cron_suggest_instructions","cron_eval_autorun","cron_instruction_evolution","process_judge_review","harbor_ab"]);async function Ze(e){let t=e.log??(()=>{}),r=e.pollWaitMs??ge,p=e.heartbeatIntervalMs??_e,d=e.capabilities??pe();s(t,{event:"agent_start",capabilities:d,pollWaitMs:r,heartbeatIntervalMs:p});try{let o=re();o.removed.length>0&&s(t,{event:"eval_worktrees_swept",removed:o.removed.length})}catch(o){s(t,{event:"eval_worktree_sweep_failed",error:o instanceof Error?o.message:String(o)})}H=te({signal:e.signal,onPause:l(o=>{s(t,{event:"machine_pause",...o,jobId:w?.jobId??null}),w?.beat()},"onPause")});let m=ke(d,e.signal,t,e.version);ne({signal:e.signal,log:t}).catch(o=>{s(t,{event:"tt_heartbeat_fatal",error:o instanceof Error?o.message:String(o)})});let g=5*6e4,f=0,a=l(()=>{if(!e.readInstalledVersion||!e.version)return!1;let o=Date.now();if(o-f<g)return!1;f=o;let b=e.readInstalledVersion();return!!(b&&b!==e.version)},"versionDrifted"),n=new AbortController,i=n.signal;e.signal.aborted?n.abort():e.signal.addEventListener("abort",()=>n.abort(),{once:!0});let c=l(async(o,b)=>{let _=ie;for(;!i.aborted;){if(a()){s(t,{event:"agent_version_drift",bootVersion:e.version,installedVersion:e.readInstalledVersion?.()??null}),n.abort(),e.onVersionDrift?.();break}let y=null;try{y=await be([...b],r)}catch(k){let ue=k instanceof M?k.statusCode:void 0;if(s(t,{event:"poll_error",lane:o,status:ue,error:k instanceof Error?k.message:String(k)}),!await I(_,i))break;_=Math.min(_*2,le);continue}if(!y){if(s(t,{event:"poll_204",lane:o}),!await I(_,i))break;_=Math.min(_*2,le);continue}_=ie,await we(y,p,t,i)}},"runLane"),u=d.filter(o=>de.has(o)),v=d.filter(o=>!de.has(o));u.length>0&&v.length>0?(s(t,{event:"agent_lanes",heavy:u.length,fast:v.length}),await Promise.all([c("heavy",u),c("fast",v)])):await c("single",d),await m.shutdown(),s(t,{event:"agent_stop"})}l(Ze,"runAgentLoop");export{ce as a,ae as b,pe as c,be as d,Ze as e};

package/dist/{chunk-AXJAGWZB.js → chunk-BFQVAU6N.js}

@@ -1 +1 @@
-import{b as o,f as a}from"./chunk-ZZHIBOMT.js";import{a as n}from"./chunk-ZLLER3HP.js";function c(e){return e==null?e??null:e instanceof Date?e.toISOString():e}n(c,"serializeDate");async function d(e){let r={provider:e.provider,plan:e.plan,limitType:e.limitType,value:e.value,unit:e.unit,comparator:e.comparator,resetsAt:c(e.resetsAt),sourceType:e.sourceType,sourceUrl:e.sourceUrl,sourceEvidence:e.sourceEvidence,confidence:e.confidence,observedAt:c(e.observedAt)};return a("/api/internal/research/observe",r)}n(d,"postObservation");async function f(e){let r=0,i=0,s=[];for(let u of e)try{await d(u),r++}catch(t){i++;let p=t instanceof o?`HTTP ${t.statusCode}: ${t.message}`:String(t);s.length<5&&s.push(p)}return{posted:r,failed:i,errors:s}}n(f,"postObservations");export{d as a,f as b};
+import{b as o,f as a}from"./chunk-76UKAFNM.js";import{a as n}from"./chunk-ZLLER3HP.js";function c(e){return e==null?e??null:e instanceof Date?e.toISOString():e}n(c,"serializeDate");async function d(e){let r={provider:e.provider,plan:e.plan,limitType:e.limitType,value:e.value,unit:e.unit,comparator:e.comparator,resetsAt:c(e.resetsAt),sourceType:e.sourceType,sourceUrl:e.sourceUrl,sourceEvidence:e.sourceEvidence,confidence:e.confidence,observedAt:c(e.observedAt)};return a("/api/internal/research/observe",r)}n(d,"postObservation");async function f(e){let r=0,i=0,s=[];for(let u of e)try{await d(u),r++}catch(t){i++;let p=t instanceof o?`HTTP ${t.statusCode}: ${t.message}`:String(t);s.length<5&&s.push(p)}return{posted:r,failed:i,errors:s}}n(f,"postObservations");export{d as a,f as b};

package/dist/{chunk-2WQUOC6S.js → chunk-BMT5WMEV.js}

@@ -1,4 +1,4 @@
-import{a as f,b as M}from"./chunk-HRJDKAJ4.js";import{e as _,f as S}from"./chunk-TWPG3TEQ.js";import{a as o}from"./chunk-ZLLER3HP.js";import*as m from"node:fs/promises";import*as x from"node:path";import{eq as j}from"drizzle-orm";import{randomBytes as Y}from"node:crypto";var A="#0d9488",h="#1a1a1a",$="#666666",C="#f5f5f7",N="#ffffff",D="#e5e5e9",I=["Added","Changed","Fixed","Removed","Deprecated","Security"];function U(e){let t=["January","February","March","April","May","June","July","August","September","October","November","December"],[n,r,a]=e.split("-").map(d=>parseInt(d,10));return!n||!r||!a||r<1||r>12?e:`${t[r-1]} ${a}, ${n}`}o(U,"formatIsoDate");function l(e){return e.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;").replace(/"/g,"&quot;").replace(/'/g,"&#39;")}o(l,"escapeHtml");function z(e){let t=[];for(let n of I){let r=e[n];r&&r.length>0&&t.push({name:n,items:r})}return t}o(z,"orderedSections");function v(e){let{version:t,date:n,sections:r,unsubscribeUrl:a,productUrl:d,recipientName:u}=e,c=U(n),s=`What's new in RuleMetric \u2014 ${c}`,y=u?`Hi ${u},`:"Hi there,",i=z(r),p=new Date().getUTCFullYear(),b=i.map(({name:R,items:E})=>{let w=E.map(O=>`              <li style="margin:0 0 6px 0;font-size:15px;line-height:1.5;color:${h};">${l(O)}</li>`).join(`
+import{a as f,b as M}from"./chunk-Q776YUBD.js";import{e as _,f as S}from"./chunk-42A4PUW7.js";import{a as o}from"./chunk-ZLLER3HP.js";import*as m from"node:fs/promises";import*as x from"node:path";import{eq as j}from"drizzle-orm";import{randomBytes as Y}from"node:crypto";var A="#0d9488",h="#1a1a1a",$="#666666",C="#f5f5f7",N="#ffffff",D="#e5e5e9",I=["Added","Changed","Fixed","Removed","Deprecated","Security"];function U(e){let t=["January","February","March","April","May","June","July","August","September","October","November","December"],[n,r,a]=e.split("-").map(d=>parseInt(d,10));return!n||!r||!a||r<1||r>12?e:`${t[r-1]} ${a}, ${n}`}o(U,"formatIsoDate");function l(e){return e.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;").replace(/"/g,"&quot;").replace(/'/g,"&#39;")}o(l,"escapeHtml");function z(e){let t=[];for(let n of I){let r=e[n];r&&r.length>0&&t.push({name:n,items:r})}return t}o(z,"orderedSections");function v(e){let{version:t,date:n,sections:r,unsubscribeUrl:a,productUrl:d,recipientName:u}=e,c=U(n),s=`What's new in RuleMetric \u2014 ${c}`,y=u?`Hi ${u},`:"Hi there,",i=z(r),p=new Date().getUTCFullYear(),b=i.map(({name:R,items:E})=>{let w=E.map(O=>`              <li style="margin:0 0 6px 0;font-size:15px;line-height:1.5;color:${h};">${l(O)}</li>`).join(`
 `);return`            <h3 style="margin:24px 0 8px 0;font-size:16px;line-height:1.3;font-weight:600;color:${h};">${R}</h3>
             <ul style="margin:0 0 8px 0;padding-left:20px;">
 ${w}

package/dist/chunk-CT6GJQIR.js

@@ -0,0 +1 @@
+import{a as E}from"./chunk-DQYYHX6K.js";import{a as I}from"./chunk-RBSEE2Z6.js";import{a as w}from"./chunk-M7QSC64D.js";import{c as v,e as h}from"./chunk-H6FVXVGG.js";import{b as R}from"./chunk-SWSOEYC7.js";import{e as m}from"./chunk-76UKAFNM.js";import{a as d}from"./chunk-ZLLER3HP.js";var T=6e5;async function P(e,$){let i=d(n=>$?.logger?.info?.(n)??console.log(n),"log"),{evalTargetId:r}=e,k=e.measurement?v(e):{},o=await m(`/api/eval-targets/${r}/judge-calibration`),l=!e.measurement||o.methodology?.measurement?.engine===e.measurement.engine&&o.methodology.measurement.model===e.measurement.model&&o.methodology.current===R(e.measurement.engine),a=I({powered:l&&o.powerCheck.powered,mde:l?o.powerCheck.mde??Number.POSITIVE_INFINITY:Number.POSITIVE_INFINITY},(l?o.caseDiscrimination:void 0)??{total:0,discriminating:0,nonDiscriminating:0,unpaired:0},{runsCompleted:0,maxRuns:e.maxRuns??24,batchesRun:0,maxBatches:1});if(a.action==="stop")return i(`[cron_eval_autorun] ${r}: ${a.reason}`),{skipped:a.code,reason:a.reason};let t=await m(`/api/eval-targets/${r}`);if(t.metadata?.autoRunEnabled===!1)return{skipped:"paused",reason:"Nightly measurement is off for this target."};let c=w(t);if(c)return i(`[cron_eval_autorun] ${r}: ${c.reason}`),c;let p=(await m(`/api/eval-targets/${r}/evals`)).filter(n=>!n.retiredAt);if(p.length===0)return{skipped:"no-live-cases"};let{completed:f,failed:b,results:N}=await h({cases:E(p,e.maxRuns??12),targetContent:t.content,concurrency:2,runNumber:Math.floor(Date.now()/1e3),opts:{targetId:r,targetVersion:t.version??1,iteration:1,timeout:e.timeout??T,model:e.model,graderModel:e.graderModel,graderEngine:e.graderEngine,...k,skipGrading:!1,targetType:t.type,targetName:t.name},onLog:i}),g={};for(let n of N){if(n.status!=="failed")continue;let s=n.failure?.kind??"unknown";g[s]=(g[s]??0)+1}let u=Object.entries(g).sort(([,n],[,s])=>s-n).map(([n,s])=>`${n}x${s}`).join(", ");return i(`[cron_eval_autorun] ${t.name}: ${f} completed, ${b} failed`+(u===""?"":` (${u})`)),{ranRuns:f,failedRuns:b,...u===""?{}:{failures:u},reason:a.reason}}d(P,"cronEvalAutorun");export{P as a};

package/dist/{chunk-FCTFHY3X.js → chunk-DHX4U6EC.js}

@@ -1,4 +1,4 @@
-import{a as A}from"./chunk-5RBLIVBF.js";import{c as W}from"./chunk-NSPDDSRR.js";import{a as T,b as k,c as U,d as O}from"./chunk-SJKS3PYX.js";import{a as B}from"./chunk-ET7753IY.js";import{a as I,d as M}from"./chunk-3UKZJAFS.js";import{b as x,e as H,g as R}from"./chunk-GN46S7T4.js";import{e as _,f as L,h as d}from"./chunk-ZZHIBOMT.js";import{a as F,b as J}from"./chunk-TWPG3TEQ.js";import{a as f}from"./chunk-ZLLER3HP.js";var V=f((l,{limit:i,sort:n,eventTypes:s})=>_(`/api/sessions/${l}/events?limit=${i}&sort=${n}&eventType=${encodeURIComponent(s.join(","))}&hasContent=true`),"fetchTranscriptTurns");async function is(l,i){let{mirrorId:n,projectPath:s,projectId:r,config:g}=l,t=f(e=>i.logger.info(e),"log");if(g?.sessionId)return Y({mirrorId:n,sessionId:g.sessionId},t);t(`Processing insights job ${n}...`),await d(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:e}=await L("/api/sessions/repair-analyzed-at",{});e>0&&t(`  Repaired analyzed_at on ${e} orphan session(s).`)}catch(e){t(`  Warning: analyzed_at repair failed: ${e instanceof Error?e.message:String(e)}`)}let p=20,w=200,D=15*6e4,N=3,E=Date.now()+D,C=H()==="none"?"no-llm-backend":null;if(C&&i.logger.warn?.("[process-insights] no LLM backend (no claude/codex CLI and no ANTHROPIC_API_KEY) \u2014 facet grading SKIPPED for this run; the summary below is computed from already-graded sessions only"),!C){let e=0,o=0;try{for(;e<w&&Date.now()<E;){let c=await _(`/api/sessions?needsFacets=true&graderVersion=${2}&limit=${p}&offset=${o}`);if(c.length===0)break;t(`  Generating facets for ${c.length} un-analyzed sessions (drained ${e} so far)...`);for(let $=0;$<c.length&&!(e>=w||Date.now()>=E);$+=N){let S=c.slice($,$+N);await Promise.all(S.map(async m=>{if(!(e>=w||Date.now()>=E))try{let u=await M(K=>V(m.id,K));if(u.turnCount<2){e++,o++;return}let j=await R(O,u.text,4096);j&&typeof j.outcome=="string"?(await d(`/api/sessions/${m.id}/enrich`,{analysis:{source:"generated_facets",...j,graderVersion:2},...m.eventCount!=null?{analyzedEventCount:m.eventCount}:{},analyzedTurnCount:u.turnCount}),t(`    Generated facets for session ${m.id} (${m.tool}, ${u.turnCount} turns)`)):o++,e++}catch(u){t(`    Failed to generate facets for session ${m.id}: ${u instanceof Error?u.message:String(u)}`),e++,o++}}))}if(c.length<p)break}}catch(c){t(`  Warning: facet generation failed after ${e} sessions: ${c instanceof Error?c.message:String(c)}`)}if(e>0){let c=Date.now()>=E?`time budget (${Math.round(D/6e4)}m)`:e>=w?`hard cap (${w})`:"queue empty";t(`  Facet drain stopped on ${c} after ${e} session(s); the rest resume next run.`)}}let b=await Q(s,r);if(!b)throw new A("No session data found \u2014 ensure sessions are captured and enriched.");t(`  Found ${b.stats.sessions} sessions, generating recommendations...`);let P=await Z(s,t);P&&t(`  Generator memory: ${P.split(`
+import{a as A}from"./chunk-5RBLIVBF.js";import{a as I,d as M}from"./chunk-3UKZJAFS.js";import{c as W}from"./chunk-XJ25HFIT.js";import{a as T,b as k,c as U,d as O}from"./chunk-SJKS3PYX.js";import{a as B}from"./chunk-YK2IEHLB.js";import{b as x,e as H,g as R}from"./chunk-N4YPDZJF.js";import{e as _,f as L,h as d}from"./chunk-76UKAFNM.js";import{a as F,b as J}from"./chunk-42A4PUW7.js";import{a as f}from"./chunk-ZLLER3HP.js";var V=f((l,{limit:i,sort:n,eventTypes:s})=>_(`/api/sessions/${l}/events?limit=${i}&sort=${n}&eventType=${encodeURIComponent(s.join(","))}&hasContent=true`),"fetchTranscriptTurns");async function is(l,i){let{mirrorId:n,projectPath:s,projectId:r,config:g}=l,t=f(e=>i.logger.info(e),"log");if(g?.sessionId)return Y({mirrorId:n,sessionId:g.sessionId},t);t(`Processing insights job ${n}...`),await d(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:e}=await L("/api/sessions/repair-analyzed-at",{});e>0&&t(`  Repaired analyzed_at on ${e} orphan session(s).`)}catch(e){t(`  Warning: analyzed_at repair failed: ${e instanceof Error?e.message:String(e)}`)}let p=20,w=200,D=15*6e4,N=3,E=Date.now()+D,C=H()==="none"?"no-llm-backend":null;if(C&&i.logger.warn?.("[process-insights] no LLM backend (no claude/codex CLI and no ANTHROPIC_API_KEY) \u2014 facet grading SKIPPED for this run; the summary below is computed from already-graded sessions only"),!C){let e=0,o=0;try{for(;e<w&&Date.now()<E;){let c=await _(`/api/sessions?needsFacets=true&graderVersion=${2}&limit=${p}&offset=${o}`);if(c.length===0)break;t(`  Generating facets for ${c.length} un-analyzed sessions (drained ${e} so far)...`);for(let $=0;$<c.length&&!(e>=w||Date.now()>=E);$+=N){let S=c.slice($,$+N);await Promise.all(S.map(async m=>{if(!(e>=w||Date.now()>=E))try{let u=await M(K=>V(m.id,K));if(u.turnCount<2){e++,o++;return}let j=await R(O,u.text,4096);j&&typeof j.outcome=="string"?(await d(`/api/sessions/${m.id}/enrich`,{analysis:{source:"generated_facets",...j,graderVersion:2},...m.eventCount!=null?{analyzedEventCount:m.eventCount}:{},analyzedTurnCount:u.turnCount}),t(`    Generated facets for session ${m.id} (${m.tool}, ${u.turnCount} turns)`)):o++,e++}catch(u){t(`    Failed to generate facets for session ${m.id}: ${u instanceof Error?u.message:String(u)}`),e++,o++}}))}if(c.length<p)break}}catch(c){t(`  Warning: facet generation failed after ${e} sessions: ${c instanceof Error?c.message:String(c)}`)}if(e>0){let c=Date.now()>=E?`time budget (${Math.round(D/6e4)}m)`:e>=w?`hard cap (${w})`:"queue empty";t(`  Facet drain stopped on ${c} after ${e} session(s); the rest resume next run.`)}}let b=await Q(s,r);if(!b)throw new A("No session data found \u2014 ensure sessions are captured and enriched.");t(`  Found ${b.stats.sessions} sessions, generating recommendations...`);let P=await Z(s,t);P&&t(`  Generator memory: ${P.split(`
 `).length} lines of prior-advice context`);let a={},y=[],h=k.length+1;for(let e=0;e<k.length;e++){let o=k[e];t(`  [${e+1}/${h}] Generating ${o.name}...`);try{await d(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:o.name,current:e+1,total:h,partial:{...a}}})}catch(u){t(`  Warning: progress update failed for ${o.name}: ${u instanceof Error?u.message:String(u)}`)}t(`  Running LLM prompt for ${o.name}...`);let c=o.usesGeneratorMemory&&P?`${b.summary}
 ${P}`:b.summary,$=Date.now(),S=await R(o.prompt,c,o.maxTokens,{timeoutMs:o.timeoutMs}),m=Date.now()-$;S?(a[o.name]=S,t(`  Got result for ${o.name} (${Math.round(m/1e3)}s)`)):(y.push(o.name),t(`  No result for ${o.name} after ${Math.round(m/1e3)}s (budget ${Math.round((o.timeoutMs??12e4)/1e3)}s) \u2014 section DROPPED`))}t(`  [${h}/${h}] Generating at_a_glance...`),await d(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"at_a_glance",current:h,total:h,partial:{...a}}});let q=[b.summary,a.what_works?`
 ## What Works

package/dist/chunk-DQYYHX6K.js

@@ -0,0 +1 @@
+import{a as i}from"./chunk-ZLLER3HP.js";function g(n,e=12,l=new Date){if(!Number.isFinite(e)||e<2)return[];let h=Math.min(6,Math.floor(e/2),n.length),t=[...n].sort((o,r)=>o.id.localeCompare(r.id));if(!t.length)return[];let a=Math.floor(l.getTime()/864e5)%t.length;return Array.from({length:h},(o,r)=>t[(a+r)%t.length])}i(g,"selectNightlyCases");export{g as a};

package/dist/{chunk-OO2ZRKZY.js → chunk-DVU5PHPM.js}

@@ -1,4 +1,4 @@
-import{a as w,b as S}from"./chunk-CDY7TQFZ.js";import{f as A}from"./chunk-UJ7EPLKC.js";import{d as G}from"./chunk-SG4UDDHB.js";import{a as p}from"./chunk-ZLLER3HP.js";var j=`You are a rigorous eval grader. Your job is to evaluate whether an AI assistant's output meets a set of expectations.
+import{a as w,b as S}from"./chunk-CDY7TQFZ.js";import{f as A}from"./chunk-FBSZ6KCM.js";import{d as G}from"./chunk-SG4UDDHB.js";import{a as p}from"./chunk-ZLLER3HP.js";var j=`You are a rigorous eval grader. Your job is to evaluate whether an AI assistant's output meets a set of expectations.
 
 You will receive:
 1. The original prompt that was given to the AI

package/dist/{chunk-ZA6NXNPR.js → chunk-EMLLGASZ.js}

@@ -1 +1 @@
-import{b as g,c as m}from"./chunk-N7ODX6PO.js";import{e as s,h as c}from"./chunk-ZZHIBOMT.js";import{a as i}from"./chunk-ZLLER3HP.js";var u=6e4;function o(t,e){try{t(JSON.stringify({ts:new Date().toISOString(),...e}))}catch{t(`[tt-heartbeat] log_error event=${e.event}`)}}i(o,"structuredLog");function d(t,e){return e.aborted?Promise.resolve(!1):t<=0?Promise.resolve(!0):new Promise(r=>{let a=setTimeout(()=>{e.removeEventListener("abort",n),r(!0)},t),n=i(()=>{clearTimeout(a),e.removeEventListener("abort",n),r(!1)},"onAbort");e.addEventListener("abort",n,{once:!0})})}i(d,"abortableSleep");async function l(t,e){if(t.kind!=="tmux")return;let a=(await g(m(t.target))).code===0?"active":"missing";try{await c(`/api/terminal-targets/${t.id}/status`,{status:a,lastSeenAt:new Date().toISOString()}),o(e,{event:"tt_heartbeat",targetId:t.id,target:t.target,status:a})}catch(n){o(e,{event:"tt_heartbeat_error",targetId:t.id,target:t.target,error:n instanceof Error?n.message:String(n)})}}i(l,"probeAndReport");async function b(t){let e;try{e=await s("/api/terminal-targets?status=active")}catch(r){o(t,{event:"tt_heartbeat_list_error",error:r instanceof Error?r.message:String(r)});return}for(let r of e.targets??[])try{await l(r,t)}catch(a){o(t,{event:"tt_heartbeat_probe_error",targetId:r.id,error:a instanceof Error?a.message:String(a)})}}i(b,"runOneHeartbeatCycle");async function p(t){let e=t.log??(()=>{}),r=t.intervalMs??u;for(o(e,{event:"tt_heartbeat_start",intervalMs:r});!t.signal.aborted&&(await b(e),!!await d(r,t.signal)););o(e,{event:"tt_heartbeat_stop"})}i(p,"startTerminalTargetHeartbeat");export{l as a,b,p as c};
+import{b as g,c as m}from"./chunk-N7ODX6PO.js";import{e as s,h as c}from"./chunk-76UKAFNM.js";import{a as i}from"./chunk-ZLLER3HP.js";var u=6e4;function o(t,e){try{t(JSON.stringify({ts:new Date().toISOString(),...e}))}catch{t(`[tt-heartbeat] log_error event=${e.event}`)}}i(o,"structuredLog");function d(t,e){return e.aborted?Promise.resolve(!1):t<=0?Promise.resolve(!0):new Promise(r=>{let a=setTimeout(()=>{e.removeEventListener("abort",n),r(!0)},t),n=i(()=>{clearTimeout(a),e.removeEventListener("abort",n),r(!1)},"onAbort");e.addEventListener("abort",n,{once:!0})})}i(d,"abortableSleep");async function l(t,e){if(t.kind!=="tmux")return;let a=(await g(m(t.target))).code===0?"active":"missing";try{await c(`/api/terminal-targets/${t.id}/status`,{status:a,lastSeenAt:new Date().toISOString()}),o(e,{event:"tt_heartbeat",targetId:t.id,target:t.target,status:a})}catch(n){o(e,{event:"tt_heartbeat_error",targetId:t.id,target:t.target,error:n instanceof Error?n.message:String(n)})}}i(l,"probeAndReport");async function b(t){let e;try{e=await s("/api/terminal-targets?status=active")}catch(r){o(t,{event:"tt_heartbeat_list_error",error:r instanceof Error?r.message:String(r)});return}for(let r of e.targets??[])try{await l(r,t)}catch(a){o(t,{event:"tt_heartbeat_probe_error",targetId:r.id,error:a instanceof Error?a.message:String(a)})}}i(b,"runOneHeartbeatCycle");async function p(t){let e=t.log??(()=>{}),r=t.intervalMs??u;for(o(e,{event:"tt_heartbeat_start",intervalMs:r});!t.signal.aborted&&(await b(e),!!await d(r,t.signal)););o(e,{event:"tt_heartbeat_stop"})}i(p,"startTerminalTargetHeartbeat");export{l as a,b,p as c};

package/dist/{chunk-XFLKW6SJ.js → chunk-EUHRTEVJ.js}

@@ -1 +1 @@
-import{c as X}from"./chunk-HOQPVOMF.js";import{b as D}from"./chunk-GIRPEMTW.js";import{a as x}from"./chunk-MODXEGCU.js";import{e as J,g as B}from"./chunk-GN46S7T4.js";import{e as w,f as R}from"./chunk-ZZHIBOMT.js";import{c as H,d as U}from"./chunk-TWPG3TEQ.js";import{a as u}from"./chunk-ZLLER3HP.js";import{readFileSync as q,realpathSync as V}from"node:fs";import y from"node:path";import A from"node:os";function Q(s){let n=y.resolve(s);if(n===y.resolve(A.homedir()))return!1;let a=y.parse(n).root;if(n===a)return!1;try{let o=V(A.tmpdir());if(V(n)===o)return!1}catch{}return!0}u(Q,"isScorableProjectPath");var Z=5,ee=Number(process.env.RULEMETRIC_MAX_SUGGESTION_PROJECTS??10),te=48,ne=.5,O=400,oe=3,re=.6,se=1;function ae(s){let n=s.bloated??[],a=new Set([...s.harmful.map(o=>o.id),...s.candidates.map(o=>o.id)]);return n.filter(o=>!a.has(o.id)).slice(0,se).map(o=>({instructionId:o.id,estimatedTokens:o.estimatedTokens,topProjectId:o.topProjectId??null,score:.55,reason:`Costs ~${o.estimatedTokens} tokens every session and measures neutral across ${o.exposures} exposed session(s) \u2014 enough data to detect an effect, and none found. Consider shortening "${o.name}" rather than removing it.`+(o.containsSafetyLanguage?" CAUTION: it contains guardrail language (destructive-command or credential rules) \u2014 preserve those lines when shortening.":"")}))}u(ae,"buildTrimProposals");async function ie(s,n,a){let o=[],i=0;for(let d of s)try{await a("/api/instruction-experiments",{instructionId:d.instructionId,projectPath:n,source:"groom"}),i++}catch(g){let m=g instanceof Error?g.message:String(g);o.push({...d,reason:`${d.reason} (not measurable by withholding: ${m})`})}return{nominated:i,unmeasurable:o}}u(ie,"nominateRemovals");function ce(s){let n=[];for(let o of s.harmful){let i=o.verdict.delta;n.push({instructionId:o.id,score:Math.min(.95,.75+Math.abs(i)),topProjectId:o.topProjectId??null,reason:`Measured harmful: sessions using "${o.name}" scored ${(i*100).toFixed(1)}pp worse across ${o.exposures} exposed session(s). Observational \u2014 review before archiving.`})}let a=new Set(s.harmful.map(o=>o.id));for(let o of s.candidates)a.has(o.id)||n.push({instructionId:o.id,score:re,topProjectId:o.topProjectId??null,reason:`Unused: "${o.name}" is ${o.reason} (${o.totalExposures} lifetime exposure(s)) but still costs context budget every session.`});return n.slice(0,oe)}u(ce,"buildRemovalProposals");function le(){let s=y.join(A.homedir(),".config","rulemetric","skills-registry.json");try{let n=q(s,"utf-8"),a=JSON.parse(n),o=new Date(a.generatedAt).getTime();return Number.isNaN(o)?"unknown":Date.now()-o>te*60*60*1e3?"stale":"fresh"}catch{return"unknown"}}u(le,"observeSkillsRegistry");var Re=u(async(s,n)=>{let a=J(),o=a!=="none";o?n.logger.info(`Scoring backend: ${a}`):n.logger.warn("No LLM backend available (claude CLI missing and ANTHROPIC_API_KEY unset); skipping relevance scoring. Removal/trim proposals still run \u2014 they need no LLM.");let i;try{i=(await w("/api/sessions/project-paths?limit="+ee)).projectPaths}catch(e){n.logger.warn(`Failed to fetch project paths: ${e instanceof Error?e.message:String(e)}`);return}if(i.length===0){n.logger.info("No recent project paths; nothing to score");return}let d=new Map;try{let e=await w("/api/projects?limit=200");for(let t of e.projects??[])t.primaryCheckoutPath&&d.set(x(t.primaryCheckoutPath),t.id)}catch(e){n.logger.warn(`Failed to fetch projects for id resolution: ${e instanceof Error?e.message:String(e)}`)}let g=new Map;try{let e=await w(`/api/catalog?limit=200&proposable=1&contentPreview=${O}`);for(let t of e)g.set(t.id,{id:t.id,name:t.name,description:t.description??"",tags:t.tags??[],contentPreview:(t.contentPreview??"").slice(0,O)})}catch(e){n.logger.warn(`Failed to fetch catalog: ${e instanceof Error?e.message:String(e)}`)}try{let e=await w("/api/instructions?limit=100");for(let t of e.instructions??[])g.has(t.id)||g.set(t.id,{id:t.id,name:t.name,description:t.description??"",tags:t.tags??[],contentPreview:(t.content??"").slice(0,O)})}catch(e){n.logger.warn(`Failed to fetch personal instructions: ${e instanceof Error?e.message:String(e)}`)}let m=[],p=[];try{let e=await w("/api/insights/groom-candidates");if(!e.captureHealthy)n.logger.info(`Grooming paused: ${e.reason??"capture unhealthy"}`);else{m=ce(e),p=ae(e);let t=e.safetyGuarded+e.harmfulSafetyGuarded;n.logger.info(`Removal proposals: ${m.length} (${e.harmful.length} measured-harmful, ${e.candidates.length} never-used, ${t} withheld by the safety allowlist); trim proposals: ${p.length} (${(e.bloated??[]).length} measured-neutral and expensive, ${e.trimSafetyFlagged??0} flagged as containing guardrail language)`)}}catch(e){n.logger.warn(`Failed to fetch groom candidates: ${e instanceof Error?e.message:String(e)}`)}let b=[...g.values()];if(b.length===0&&m.length===0&&p.length===0){n.logger.info("No candidate instructions and nothing to retire; nothing to score");return}n.logger.info(`Scoring ${b.length} candidates across ${i.length} project(s)`);let M=0,l=null,I=new Set;for(let e of i)try{let t=x(e);if(!Q(t)){n.logger.info(`Skipping ${t}: not a project directory (home/root)`);continue}let f=D(t),k=o&&f.length>0;f.length===0&&n.logger.info(`${t}: insufficient language signals; removal/trim only`);let v=X(t),z={name:y.basename(t),description:v.description,languages:f,topics:v.topics,dependencies:v.dependencies},E=k?U(b):[];k&&n.logger.info(`Scoring ${b.length} candidates for ${t} in ${E.length} batch(es)`);let C=[],$=0;for(let r of E)try{let c=await B(H(z,r),"",4096);if(!c){$++;continue}let T=c.recommendations??[];for(let P of T)typeof P.instructionId=="string"&&typeof P.score=="number"&&typeof P.reason=="string"&&C.push({instructionId:P.instructionId,score:P.score,reason:P.reason})}catch(c){$++,n.logger.warn(`Batch failed for ${t}: ${c instanceof Error?c.message:String(c)}`)}$>0&&n.logger.warn(`${$}/${E.length} batch(es) failed for ${t} (backend ${a}); suggestions are partial`),k&&C.length===0&&n.logger.warn(`No usable LLM output for ${t} (backend ${a})`);let K=C.filter(r=>r.score>=ne).sort((r,c)=>c.score-r.score).slice(0,Z),_=d.get(t),W=l===null||l===t;l===null&&(l=t);let N=u(r=>r.topProjectId?r.topProjectId===_:W,"placeHere"),F=m.filter(N),{nominated:L,unmeasurable:G}=await ie(F,t,R);L>0&&n.logger.info(`${t}: ${L} removal candidate(s) nominated for randomised withholding instead of the feed`);let h=[...K.map(r=>({instructionId:r.instructionId,score:r.score,reason:r.reason,kind:"add"})),...G.map(({topProjectId:r,...c})=>({...c,kind:"remove"})),...p.filter(N).map(({estimatedTokens:r,topProjectId:c,...T})=>({...T,kind:"trim"}))];if(h.length===0){n.logger.info(`Nothing to suggest for ${t}`);continue}await R("/api/instruction-suggestions/batch-upsert",{projectPath:t,..._?{projectId:_}:{},suggestions:h}),n.logger.info(`${t}: ${h.filter(r=>r.kind==="add").length} add + ${h.filter(r=>r.kind==="remove").length} remove + ${h.filter(r=>r.kind==="trim").length} trim`);for(let r of h)r.kind!=="add"&&I.add(r.instructionId);let Y=new Set(G.map(r=>r.instructionId));for(let r of F)Y.has(r.instructionId)||I.add(r.instructionId);M++}catch(t){n.logger.warn(`Failed to score ${e}: ${t instanceof Error?t.message:String(t)}`)}let j=[...m.filter(e=>!I.has(e.instructionId)),...p.filter(e=>!I.has(e.instructionId)).map(({estimatedTokens:e,...t})=>t)];if(j.length>0&&l)try{let e=new Set(p.map(t=>t.instructionId));await R("/api/instruction-suggestions/batch-upsert",{projectPath:l,...d.get(l)?{projectId:d.get(l)}:{},suggestions:j.map(({topProjectId:t,...f})=>({...f,kind:e.has(f.instructionId)?"trim":"remove"}))}),n.logger.info(`Orphan sweep: ${j.length} proposal(s) whose project was not scored this run placed on ${l}`)}catch(e){n.logger.warn(`Orphan sweep failed: ${e instanceof Error?e.message:String(e)}`)}let S=le();if(S==="unknown")n.logger.warn("Skills registry unreadable or has no usable generatedAt; leaving suggestion staleness unchanged");else try{await R("/api/instruction-suggestions/mark-stale",{stale:S==="stale"}),S==="stale"?n.logger.warn("Skills registry is stale (>48h); marked catalog-skill suggestions stale"):n.logger.info("Skills registry is fresh; cleared staleness on catalog-skill suggestions")}catch{}n.logger.info(`Done: scored ${M}/${i.length} projects`)},"cronSuggestInstructions");export{Q as a,ae as b,ie as c,ce as d,Re as e};
+import{c as X}from"./chunk-HOQPVOMF.js";import{b as D}from"./chunk-GIRPEMTW.js";import{a as x}from"./chunk-MODXEGCU.js";import{e as J,g as B}from"./chunk-N4YPDZJF.js";import{e as w,f as R}from"./chunk-76UKAFNM.js";import{c as H,d as U}from"./chunk-42A4PUW7.js";import{a as u}from"./chunk-ZLLER3HP.js";import{readFileSync as q,realpathSync as V}from"node:fs";import y from"node:path";import A from"node:os";function Q(s){let n=y.resolve(s);if(n===y.resolve(A.homedir()))return!1;let a=y.parse(n).root;if(n===a)return!1;try{let o=V(A.tmpdir());if(V(n)===o)return!1}catch{}return!0}u(Q,"isScorableProjectPath");var Z=5,ee=Number(process.env.RULEMETRIC_MAX_SUGGESTION_PROJECTS??10),te=48,ne=.5,O=400,oe=3,re=.6,se=1;function ae(s){let n=s.bloated??[],a=new Set([...s.harmful.map(o=>o.id),...s.candidates.map(o=>o.id)]);return n.filter(o=>!a.has(o.id)).slice(0,se).map(o=>({instructionId:o.id,estimatedTokens:o.estimatedTokens,topProjectId:o.topProjectId??null,score:.55,reason:`Costs ~${o.estimatedTokens} tokens every session and measures neutral across ${o.exposures} exposed session(s) \u2014 enough data to detect an effect, and none found. Consider shortening "${o.name}" rather than removing it.`+(o.containsSafetyLanguage?" CAUTION: it contains guardrail language (destructive-command or credential rules) \u2014 preserve those lines when shortening.":"")}))}u(ae,"buildTrimProposals");async function ie(s,n,a){let o=[],i=0;for(let d of s)try{await a("/api/instruction-experiments",{instructionId:d.instructionId,projectPath:n,source:"groom"}),i++}catch(g){let m=g instanceof Error?g.message:String(g);o.push({...d,reason:`${d.reason} (not measurable by withholding: ${m})`})}return{nominated:i,unmeasurable:o}}u(ie,"nominateRemovals");function ce(s){let n=[];for(let o of s.harmful){let i=o.verdict.delta;n.push({instructionId:o.id,score:Math.min(.95,.75+Math.abs(i)),topProjectId:o.topProjectId??null,reason:`Measured harmful: sessions using "${o.name}" scored ${(i*100).toFixed(1)}pp worse across ${o.exposures} exposed session(s). Observational \u2014 review before archiving.`})}let a=new Set(s.harmful.map(o=>o.id));for(let o of s.candidates)a.has(o.id)||n.push({instructionId:o.id,score:re,topProjectId:o.topProjectId??null,reason:`Unused: "${o.name}" is ${o.reason} (${o.totalExposures} lifetime exposure(s)) but still costs context budget every session.`});return n.slice(0,oe)}u(ce,"buildRemovalProposals");function le(){let s=y.join(A.homedir(),".config","rulemetric","skills-registry.json");try{let n=q(s,"utf-8"),a=JSON.parse(n),o=new Date(a.generatedAt).getTime();return Number.isNaN(o)?"unknown":Date.now()-o>te*60*60*1e3?"stale":"fresh"}catch{return"unknown"}}u(le,"observeSkillsRegistry");var Re=u(async(s,n)=>{let a=J(),o=a!=="none";o?n.logger.info(`Scoring backend: ${a}`):n.logger.warn("No LLM backend available (claude CLI missing and ANTHROPIC_API_KEY unset); skipping relevance scoring. Removal/trim proposals still run \u2014 they need no LLM.");let i;try{i=(await w("/api/sessions/project-paths?limit="+ee)).projectPaths}catch(e){n.logger.warn(`Failed to fetch project paths: ${e instanceof Error?e.message:String(e)}`);return}if(i.length===0){n.logger.info("No recent project paths; nothing to score");return}let d=new Map;try{let e=await w("/api/projects?limit=200");for(let t of e.projects??[])t.primaryCheckoutPath&&d.set(x(t.primaryCheckoutPath),t.id)}catch(e){n.logger.warn(`Failed to fetch projects for id resolution: ${e instanceof Error?e.message:String(e)}`)}let g=new Map;try{let e=await w(`/api/catalog?limit=200&proposable=1&contentPreview=${O}`);for(let t of e)g.set(t.id,{id:t.id,name:t.name,description:t.description??"",tags:t.tags??[],contentPreview:(t.contentPreview??"").slice(0,O)})}catch(e){n.logger.warn(`Failed to fetch catalog: ${e instanceof Error?e.message:String(e)}`)}try{let e=await w("/api/instructions?limit=100");for(let t of e.instructions??[])g.has(t.id)||g.set(t.id,{id:t.id,name:t.name,description:t.description??"",tags:t.tags??[],contentPreview:(t.content??"").slice(0,O)})}catch(e){n.logger.warn(`Failed to fetch personal instructions: ${e instanceof Error?e.message:String(e)}`)}let m=[],p=[];try{let e=await w("/api/insights/groom-candidates");if(!e.captureHealthy)n.logger.info(`Grooming paused: ${e.reason??"capture unhealthy"}`);else{m=ce(e),p=ae(e);let t=e.safetyGuarded+e.harmfulSafetyGuarded;n.logger.info(`Removal proposals: ${m.length} (${e.harmful.length} measured-harmful, ${e.candidates.length} never-used, ${t} withheld by the safety allowlist); trim proposals: ${p.length} (${(e.bloated??[]).length} measured-neutral and expensive, ${e.trimSafetyFlagged??0} flagged as containing guardrail language)`)}}catch(e){n.logger.warn(`Failed to fetch groom candidates: ${e instanceof Error?e.message:String(e)}`)}let b=[...g.values()];if(b.length===0&&m.length===0&&p.length===0){n.logger.info("No candidate instructions and nothing to retire; nothing to score");return}n.logger.info(`Scoring ${b.length} candidates across ${i.length} project(s)`);let M=0,l=null,I=new Set;for(let e of i)try{let t=x(e);if(!Q(t)){n.logger.info(`Skipping ${t}: not a project directory (home/root)`);continue}let f=D(t),k=o&&f.length>0;f.length===0&&n.logger.info(`${t}: insufficient language signals; removal/trim only`);let v=X(t),z={name:y.basename(t),description:v.description,languages:f,topics:v.topics,dependencies:v.dependencies},E=k?U(b):[];k&&n.logger.info(`Scoring ${b.length} candidates for ${t} in ${E.length} batch(es)`);let C=[],$=0;for(let r of E)try{let c=await B(H(z,r),"",4096);if(!c){$++;continue}let T=c.recommendations??[];for(let P of T)typeof P.instructionId=="string"&&typeof P.score=="number"&&typeof P.reason=="string"&&C.push({instructionId:P.instructionId,score:P.score,reason:P.reason})}catch(c){$++,n.logger.warn(`Batch failed for ${t}: ${c instanceof Error?c.message:String(c)}`)}$>0&&n.logger.warn(`${$}/${E.length} batch(es) failed for ${t} (backend ${a}); suggestions are partial`),k&&C.length===0&&n.logger.warn(`No usable LLM output for ${t} (backend ${a})`);let K=C.filter(r=>r.score>=ne).sort((r,c)=>c.score-r.score).slice(0,Z),_=d.get(t),W=l===null||l===t;l===null&&(l=t);let N=u(r=>r.topProjectId?r.topProjectId===_:W,"placeHere"),F=m.filter(N),{nominated:L,unmeasurable:G}=await ie(F,t,R);L>0&&n.logger.info(`${t}: ${L} removal candidate(s) nominated for randomised withholding instead of the feed`);let h=[...K.map(r=>({instructionId:r.instructionId,score:r.score,reason:r.reason,kind:"add"})),...G.map(({topProjectId:r,...c})=>({...c,kind:"remove"})),...p.filter(N).map(({estimatedTokens:r,topProjectId:c,...T})=>({...T,kind:"trim"}))];if(h.length===0){n.logger.info(`Nothing to suggest for ${t}`);continue}await R("/api/instruction-suggestions/batch-upsert",{projectPath:t,..._?{projectId:_}:{},suggestions:h}),n.logger.info(`${t}: ${h.filter(r=>r.kind==="add").length} add + ${h.filter(r=>r.kind==="remove").length} remove + ${h.filter(r=>r.kind==="trim").length} trim`);for(let r of h)r.kind!=="add"&&I.add(r.instructionId);let Y=new Set(G.map(r=>r.instructionId));for(let r of F)Y.has(r.instructionId)||I.add(r.instructionId);M++}catch(t){n.logger.warn(`Failed to score ${e}: ${t instanceof Error?t.message:String(t)}`)}let j=[...m.filter(e=>!I.has(e.instructionId)),...p.filter(e=>!I.has(e.instructionId)).map(({estimatedTokens:e,...t})=>t)];if(j.length>0&&l)try{let e=new Set(p.map(t=>t.instructionId));await R("/api/instruction-suggestions/batch-upsert",{projectPath:l,...d.get(l)?{projectId:d.get(l)}:{},suggestions:j.map(({topProjectId:t,...f})=>({...f,kind:e.has(f.instructionId)?"trim":"remove"}))}),n.logger.info(`Orphan sweep: ${j.length} proposal(s) whose project was not scored this run placed on ${l}`)}catch(e){n.logger.warn(`Orphan sweep failed: ${e instanceof Error?e.message:String(e)}`)}let S=le();if(S==="unknown")n.logger.warn("Skills registry unreadable or has no usable generatedAt; leaving suggestion staleness unchanged");else try{await R("/api/instruction-suggestions/mark-stale",{stale:S==="stale"}),S==="stale"?n.logger.warn("Skills registry is stale (>48h); marked catalog-skill suggestions stale"):n.logger.info("Skills registry is fresh; cleared staleness on catalog-skill suggestions")}catch{}n.logger.info(`Done: scored ${M}/${i.length} projects`)},"cronSuggestInstructions");export{Q as a,ae as b,ie as c,ce as d,Re as e};

package/dist/chunk-FBSZ6KCM.js

@@ -0,0 +1 @@
+import{d as D}from"./chunk-JP6ULOEX.js";import{d as b}from"./chunk-SG4UDDHB.js";import{c as k,d as w}from"./chunk-GPBN4ROE.js";import{a}from"./chunk-ZLLER3HP.js";import{spawn as L}from"node:child_process";var g=["claude","codex","pi"];function A(e,t=process.env){if(e){if(g.includes(e))return e;throw new Error(`Eval engine "${String(e)}" is not known (expected one of: ${g.join(", ")})`)}let o=t.RULEMETRIC_EVAL_ENGINE?.trim().toLowerCase();if(!o)return"claude";if(g.includes(o))return o;throw new Error(`RULEMETRIC_EVAL_ENGINE="${t.RULEMETRIC_EVAL_ENGINE}" is not a known engine (expected one of: ${g.join(", ")})`)}a(A,"resolveEvalEngine");function N(e,t=process.env){return e?.trim()||t.RULEMETRIC_CODEX_MODEL?.trim()||w().codexConfiguredModel||k}a(N,"resolveCodexEvalModel");function S(e){try{let t=JSON.parse(e),o=t.result??t.content??e,u=t.usage?.total_tokens??(t.usage?.input_tokens??0)+(t.usage?.output_tokens??0),r;if(t.modelUsage&&typeof t.modelUsage=="object"){let n=-1;for(let[l,p]of Object.entries(t.modelUsage)){let s=typeof p?.outputTokens=="number"?p.outputTokens:0;s>n&&(n=s,r=l)}}return{output:o,totalTokens:u,model:r}}catch{return{output:e,totalTokens:0}}}a(S,"parseClaudeJsonOutput");function h(e){let{prompt:t,targetContent:o,systemPrompt:u,model:r}=e,n=["-p",t,"--output-format","json","--settings",JSON.stringify({disableAllHooks:!0})],l=o??u;return l&&n.push(e.appendTarget?"--append-system-prompt":"--system-prompt",l),r&&n.push("--model",r),n}a(h,"buildClaudeArgs");function U(e,t,o,u){let r=t.trim()||o.trim()||u.trim()||"(no output on stderr or stdout)";return`Exit code ${e}: ${r.slice(0,1e3)}`}a(U,"describeExitFailure");async function I(e){let{prompt:t,targetContent:o,systemPrompt:u,workDir:r,timeout:n,model:l}=e,p=o??u,s=await b({prompt:t,systemPrompt:p,model:N(l),timeout:n,workDir:r});return{output:s.output,exitCode:s.exitCode,durationMs:s.durationMs,totalTokens:s.usage?.totalTokens??0,model:s.model,engine:"codex",delivery:p?"prompt_prefix":"none",error:s.error}}a(I,"executeCodexArm");async function $(e){let t=e.targetContent??e.systemPrompt,o=await D({prompt:e.prompt,...e.appendTarget?{appendSystemPrompt:t}:{systemPrompt:t},ignoreContextFiles:e.ignoreContextFiles,model:e.model,timeout:e.timeout,workDir:e.workDir});return{output:o.output,exitCode:o.exitCode,durationMs:o.durationMs,totalTokens:o.usage?.totalTokens??0,model:o.model,provider:o.provider,costUsd:o.costUsd,engine:"pi",delivery:t?"system_prompt":"none",error:o.error}}a($,"executePiArm");async function B(e){let{prompt:t,targetContent:o,systemPrompt:u,workDir:r=process.cwd(),timeout:n=12e4,model:l}=e,p=A(e.engine);if(p==="codex")return I({...e,workDir:r,timeout:n});if(p==="pi")return $({...e,workDir:r,timeout:n});let s=h({prompt:t,targetContent:o,systemPrompt:u,model:l,appendTarget:e.appendTarget}),E=o??u?"system_prompt":"none",x=Date.now();return new Promise(f=>{let d={...process.env};delete d.CLAUDECODE,delete d.HTTPS_PROXY;let _=["/opt/homebrew/bin","/usr/local/bin"],v=(d.PATH??"").split(":").filter(Boolean),y=_.filter(i=>!v.includes(i));y.length&&(d.PATH=[...v,...y].join(":"));let m=L("claude",s,{cwd:r,env:d,stdio:["ignore","pipe","pipe"]}),c="",C="";m.stdout.on("data",i=>{c+=i.toString()}),m.stderr.on("data",i=>{C+=i.toString()});let T=setTimeout(()=>{m.kill("SIGTERM"),f({output:c,exitCode:null,durationMs:Date.now()-x,totalTokens:0,engine:"claude",delivery:E,error:`Timed out after ${n}ms`})},n);m.on("close",i=>{clearTimeout(T);let M=Date.now()-x,{output:P,totalTokens:O,model:R}=S(c);f({output:P,exitCode:i,durationMs:M,totalTokens:O,model:R,engine:"claude",delivery:E,error:i!==0?U(i,C,P,c):void 0})}),m.on("error",i=>{clearTimeout(T),f({output:"",exitCode:null,durationMs:Date.now()-x,totalTokens:0,engine:"claude",delivery:E,error:`Failed to spawn claude: ${i.message}`})})})}a(B,"executeEval");export{A as a,N as b,S as c,h as d,U as e,B as f};

package/dist/{chunk-E5YE25QG.js → chunk-FKTT7A4H.js}

@@ -1 +1 @@
-import{a as d,d as m}from"./chunk-ZZHIBOMT.js";import{a as c,l as a}from"./chunk-PZNWH4VQ.js";import{a as o}from"./chunk-ZLLER3HP.js";import{createHash as g,randomUUID as v}from"node:crypto";import{hostname as y,platform as p}from"node:os";import{existsSync as P,mkdirSync as _,readFileSync as h,writeFileSync as R}from"node:fs";import{join as E}from"node:path";var O="device-id";function u(){let e=o(t=>t==="1"||t==="true","off");return e(process.env.RULEMETRIC_NO_TELEMETRY)||e(process.env.DO_NOT_TRACK)}o(u,"telemetryDisabled");function l(){let e=a(),t=E(e,O);try{if(P(t)){let r=h(t,"utf-8").trim();if(r)return r}}catch{}let n=g("sha256").update(`${v()}:${y()}`).digest("hex").slice(0,32);try{_(e,{recursive:!0}),R(t,n,{mode:384})}catch{}return n}o(l,"getDeviceId");async function U(e,t={},n){if(!u())try{await m("/api/telemetry/cli-events",{method:"POST",body:{event:e,deviceId:l(),cliVersion:n,os:p(),props:t},timeoutMs:5e3})}catch{}}o(U,"emitSetupEvent");async function w(e){if(u())return;let n=`${process.env.RULEMETRIC_API_URL??c().apiUrl??d}/api/telemetry/cli-error`,r=process.env.HTTPS_PROXY,s=process.env.HTTP_PROXY,T=r!==void 0||s!==void 0;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let i=new AbortController,f=setTimeout(()=>i.abort(),4e3);try{await globalThis.fetch(n,{method:"POST",headers:{"Content-Type":"application/json","User-Agent":`rulemetric-cli/${e.cliVersion??"unknown"}`},body:JSON.stringify({event:e.event??"onboarding_error",deviceId:l(),command:e.command,statusCode:e.statusCode,message:e.message?e.message.slice(0,500):void 0,cliVersion:e.cliVersion,os:p(),nodeVersion:process.version,proxied:T,props:e.props}),signal:i.signal})}catch{}finally{clearTimeout(f),r!==void 0&&(process.env.HTTPS_PROXY=r),s!==void 0&&(process.env.HTTP_PROXY=s)}}o(w,"emitErrorBeacon");export{u as a,l as b,U as c,w as d};
+import{a as d,d as m}from"./chunk-76UKAFNM.js";import{a as c,l as a}from"./chunk-PZNWH4VQ.js";import{a as o}from"./chunk-ZLLER3HP.js";import{createHash as g,randomUUID as v}from"node:crypto";import{hostname as y,platform as p}from"node:os";import{existsSync as P,mkdirSync as _,readFileSync as h,writeFileSync as R}from"node:fs";import{join as E}from"node:path";var O="device-id";function u(){let e=o(t=>t==="1"||t==="true","off");return e(process.env.RULEMETRIC_NO_TELEMETRY)||e(process.env.DO_NOT_TRACK)}o(u,"telemetryDisabled");function l(){let e=a(),t=E(e,O);try{if(P(t)){let r=h(t,"utf-8").trim();if(r)return r}}catch{}let n=g("sha256").update(`${v()}:${y()}`).digest("hex").slice(0,32);try{_(e,{recursive:!0}),R(t,n,{mode:384})}catch{}return n}o(l,"getDeviceId");async function U(e,t={},n){if(!u())try{await m("/api/telemetry/cli-events",{method:"POST",body:{event:e,deviceId:l(),cliVersion:n,os:p(),props:t},timeoutMs:5e3})}catch{}}o(U,"emitSetupEvent");async function w(e){if(u())return;let n=`${process.env.RULEMETRIC_API_URL??c().apiUrl??d}/api/telemetry/cli-error`,r=process.env.HTTPS_PROXY,s=process.env.HTTP_PROXY,T=r!==void 0||s!==void 0;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let i=new AbortController,f=setTimeout(()=>i.abort(),4e3);try{await globalThis.fetch(n,{method:"POST",headers:{"Content-Type":"application/json","User-Agent":`rulemetric-cli/${e.cliVersion??"unknown"}`},body:JSON.stringify({event:e.event??"onboarding_error",deviceId:l(),command:e.command,statusCode:e.statusCode,message:e.message?e.message.slice(0,500):void 0,cliVersion:e.cliVersion,os:p(),nodeVersion:process.version,proxied:T,props:e.props}),signal:i.signal})}catch{}finally{clearTimeout(f),r!==void 0&&(process.env.HTTPS_PROXY=r),s!==void 0&&(process.env.HTTP_PROXY=s)}}o(w,"emitErrorBeacon");export{u as a,l as b,U as c,w as d};

package/dist/{chunk-UQDOOQCM.js → chunk-FYZ6HHCV.js}

@@ -1 +1 @@
-import{b as t,e as E}from"./chunk-S5U76YUI.js";import{a as r}from"./chunk-ZLLER3HP.js";function s(o=10,_=t){for(let e=1;e<=o;e++)if(E(e,o,_).feasible)return e;return o}r(s,"minFeasibleLiveCases");var S=s(),A=3;export{s as a,S as b,A as c};
+import{b as t,e as E}from"./chunk-POEMCITL.js";import{a as r}from"./chunk-ZLLER3HP.js";function s(o=10,_=t){for(let e=1;e<=o;e++)if(E(e,o,_).feasible)return e;return o}r(s,"minFeasibleLiveCases");var S=s(),A=3;export{s as a,S as b,A as c};

package/dist/{chunk-5XEGRU2R.js → chunk-GKVHFXUQ.js}

@@ -1,4 +1,4 @@
-import{g as E}from"./chunk-GN46S7T4.js";import{b,e as y,f as k,h as M}from"./chunk-ZZHIBOMT.js";import{a as u}from"./chunk-ZLLER3HP.js";var v=50,I=24e3,P=3,R=3,C=40,N=4e3,x=new Set(["fact","decision","correction","convention"]),L=`You are distilling DURABLE, PROJECT-SPECIFIC knowledge from an AI coding session so a future agent working on the SAME project starts smarter.
+import{g as E}from"./chunk-N4YPDZJF.js";import{b,e as y,f as k,h as M}from"./chunk-76UKAFNM.js";import{a as u}from"./chunk-ZLLER3HP.js";var v=50,I=24e3,P=3,R=3,C=40,N=4e3,x=new Set(["fact","decision","correction","convention"]),L=`You are distilling DURABLE, PROJECT-SPECIFIC knowledge from an AI coding session so a future agent working on the SAME project starts smarter.
 
 Extract ONLY facts / decisions / corrections / conventions that will STILL be true and useful in FUTURE sessions \u2014 architecture facts, deliberate design decisions, corrections of a wrong approach, or project conventions.
 

package/dist/chunk-GPBN4ROE.js

@@ -0,0 +1,2 @@
+import{a as t}from"./chunk-6XTLP2G7.js";import{a as l}from"./chunk-ZLLER3HP.js";import{readFileSync as x}from"node:fs";import{homedir as m}from"node:os";import{join as a}from"node:path";function p(i){for(let n of i.split(`
+`)){let o=n.trim();if(o.startsWith("["))break;let e=/^model\s*=\s*"([^"]+)"/.exec(o);if(e)return e[1]}return null}l(p,"parseCodexConfigModel");function B(i){return/requires a newer version of Codex|is not supported when using Codex/i.test(i)}l(B,"isCodexModelUnsupported");var u="gpt-5.5";function M(i=a(m(),".codex")){let n=null;try{n=p(x(a(i,"config.toml"),"utf8"))}catch{}return{codexBin:t("codex"),claudeBin:t("claude"),piBin:t("pi"),codexConfiguredModel:n}}l(M,"detectEngines");function $(i){let{engineFlag:n,modelFlag:o,availability:e,judgeModels:d}=i,s=o??e.codexConfiguredModel??u,g=o?`--model ${o}`:e.codexConfiguredModel?`model ${e.codexConfiguredModel} from your ~/.codex/config.toml`:`fallback model ${u} (no ~/.codex/config.toml model found)`,r=o??(d.some(f=>/opus/i.test(f))?"sonnet":null),c=o?`--model ${o}`:r?`model ${r} (differs from the ${d.join(", ")} judge)`:"CLI default model";if(n==="codex"||n==="auto"&&e.codexBin){if(!e.codexBin)throw new Error("Codex CLI not found on PATH, well-known install dirs, or the node bin dir \u2014 install it or use --engine claude.");return{engine:"codex",model:s,rationale:`codex at ${e.codexBin} (cross-vendor reviewer), ${g}`}}if(!e.claudeBin)throw new Error("No reviewer CLI found \u2014 neither codex nor claude is installed on this machine.");return{engine:"claude",model:r,rationale:n==="auto"?`claude at ${e.claudeBin} (codex not installed \u2014 same-vendor reviewer, agreement will overstate independence), ${c}`:`claude at ${e.claudeBin}, ${c}`}}l($,"chooseReviewer");export{p as a,B as b,u as c,M as d,$ as e};

package/dist/chunk-H6FVXVGG.js

@@ -0,0 +1 @@
+import{f as w}from"./chunk-SWSOEYC7.js";import{a as C}from"./chunk-KGURGIZ2.js";import{a as M}from"./chunk-DVU5PHPM.js";import{a as x,f as b}from"./chunk-FBSZ6KCM.js";import{d as k}from"./chunk-GPBN4ROE.js";import{f as v}from"./chunk-76UKAFNM.js";import{f as y}from"./chunk-XQYAVSUQ.js";import{a as m}from"./chunk-ZLLER3HP.js";function $(t,e={}){let i=e.engine??(e.model||t.engine==="pi"?"claude":t.engine);return{engine:i,model:e.model??(i===t.engine?t.model:void 0)}}m($,"selectEvalJudge");function _(t){let e=$(t.actor,t.pin??{});if(!t.available.includes(e.engine))throw new Error(`configured grader unavailable: ${e.engine}`);if(e.engine==="codex"&&!e.model?.trim())throw new Error("Codex grader requires a model pin");return e}m(_,"resolveMeasurementJudge");function A(t,e){let i=t.measurement?y.parse(t.measurement):void 0,r=i?.engine??x(t.engine),o=i?i.model??void 0:t.model;if(!e){let s=k();e=[...s.claudeBin?["claude"]:[],...s.codexBin?["codex"]:[],...s.piBin?["pi"]:[]]}if(!e.includes(r))throw new Error(`configured actor unavailable: ${r}`);let u=t.skipGrading?$({engine:r,model:o}):_({actor:{engine:r,model:o},pin:t.graderEngine||t.graderModel?{engine:t.graderEngine,model:t.graderModel}:null,available:e});return{engine:r,model:o,graderEngine:u.engine,graderModel:u.model,...i?{measurement:i}:{}}}m(A,"resolveMeasurementJob");async function R(t,e,i=()=>{}){let{eval_:r,configuration:o,targetContent:u,runNumber:s}=t;i(`[${o}] ${r.name} run#${s} \u2014 executing...`);let l=x(e.engine),g=w({targetType:e.targetType,targetContent:u,targetName:e.targetName,engine:l}),c={...e.executionKey?{executionKey:e.executionKey}:{},deliveryRegime:`${e.targetType??"prompt"}:${l}:${g.methodology}`,...e.measurement?{measurement:e.measurement}:{},methodology:g.methodology,delivery:g.delivery,isolated:g.isWorktree,...e.candidateContentHash?{candidateContentHash:e.candidateContentHash}:{},...g.leakedFiles.length?{leakedFiles:g.leakedFiles}:{}};g.isWorktree||i(`[${o}] ${r.name} run#${s} \u2014 \u26A0 no git worktree; running without repo context`),g.leakedFiles.length&&i(`[${o}] ${r.name} run#${s} \u2014 \u26A0 CONTAMINATED: ${g.leakedFiles.join(", ")} survived the strip`);try{let n=await b({prompt:r.prompt,targetContent:g.systemPrompt,workDir:g.workDir,timeout:e.timeout,model:e.model,engine:l});if(n.error){let d=C(n.error);return i(`[${o}] ${r.name} run#${s} \u2014 FAILED (${d.kind}): ${n.error}`),{evalId:r.id,evalTargetId:e.targetId,evalTargetVersion:e.targetVersion,configuration:o,runNumber:s,iteration:e.iteration,status:"failed",outputs:{error:n.error,failureKind:d.kind,failureClass:d.class,...c,engine:n.engine??l,model:n.model??null,...n.provider?{provider:n.provider}:{},...n.costUsd!=null?{costUsd:n.costUsd}:{}},timing:{duration_ms:n.durationMs,total_tokens:n.totalTokens},grading:{},failure:d}}i(`[${o}] ${r.name} run#${s} \u2014 completed in ${(n.durationMs/1e3).toFixed(1)}s`);let a={};if(!e.skipGrading&&r.expectations.length>0){i(`[${o}] ${r.name} run#${s} \u2014 grading...`);let d=await M({prompt:r.prompt,output:n.output,expectations:r.expectations,...$({engine:l,model:e.model},{engine:e.graderEngine,model:e.graderModel}),timeout:e.timeout});a=d,i(`[${o}] ${r.name} run#${s} \u2014 grading done: pass_rate=${d.summary.pass_rate==null?"n/a (all assertions errored)":`${(d.summary.pass_rate*100).toFixed(0)}%`}`)}return{evalId:r.id,evalTargetId:e.targetId,evalTargetVersion:e.targetVersion,configuration:o,runNumber:s,iteration:e.iteration,status:"completed",outputs:{transcript:n.output,...c,engine:n.engine??l,...n.provider?{provider:n.provider}:{},model:n.model??null,...n.costUsd!=null?{costUsd:n.costUsd}:{}},timing:{duration_ms:n.durationMs,total_tokens:n.totalTokens,total_duration_seconds:n.durationMs/1e3},grading:a}}finally{g.cleanup()}}m(R,"executeAndGradeCase");async function B(t){let{cases:e,targetContent:i,concurrency:r,runNumber:o,opts:u,skipBaseline:s,onLog:l=m(()=>{},"onLog")}=t,g=s?["with_target"]:["with_target","without_target"],c=[];for(let a of e)for(let d of g)c.push({eval_:a,configuration:d,targetContent:d==="with_target"?i:void 0,runNumber:o});let n=[];for(let a=0;a<c.length;a+=Math.max(1,r)){let d=c.slice(a,a+Math.max(1,r)),h=await Promise.all(d.map(f=>R(f,u,l)));n.push(...h);let p=[];for(let f of h)try{await v("/api/eval-runs",f)}catch(E){p.push(E instanceof Error?E.message:String(E))}if(p.length)throw new Error(`Could not persist ${p.length} eval run(s); stopped further execution: ${p.join("; ")}`)}try{await v(`/api/eval-targets/${u.targetId}/label-state`,{})}catch(a){l(`  \u26A0 label-state notification skipped: ${a instanceof Error?a.message:String(a)}`)}return{completed:n.filter(a=>a.status==="completed").length,failed:n.filter(a=>a.status==="failed").length,results:n}}m(B,"runBatchAndPost");export{$ as a,_ as b,A as c,R as d,B as e};

package/dist/{chunk-JKWB6UXA.js → chunk-HNYHZO2J.js}

@@ -1 +1 @@
-import{b as D,c as j}from"./chunk-6UEFCMR2.js";import{a as G}from"./chunk-OO2ZRKZY.js";import{d as J}from"./chunk-SG4UDDHB.js";import{b as B,c as R,d as N,e as z}from"./chunk-SBGXIPRN.js";import{e as f,g as T}from"./chunk-ZZHIBOMT.js";import{a as p}from"./chunk-ZLLER3HP.js";function W(r,a,s){return s?r.reviewerModel===s||r.reviewerModel===`${a}:${s}`:r.reviewerEngine===a}p(W,"reviewMatchesReviewer");var C=100;function Z(r){let a=new Set;return r.forEach((s,d)=>{s.method!=="deterministic"&&a.add(d)}),a}p(Z,"llmExpectationIndexes");var w=class extends Error{constructor(s,d){super(d);this.reason=s}static{p(this,"JudgeReviewSkip")}};async function oe(r){let a=r.log??(()=>{}),s=r.warn??a,d=r.targetId,K=r.parallel??3,U=r.timeoutMs??12e4,v=await f(`/api/eval-targets/${d}`),u=(await f(`/api/eval-targets/${d}/annotations?limit=500`)).filter(e=>e.judgeMethod==="llm");if(u.length===0)throw new w("no-labels",`No human labels on "${v.name}" yet \u2014 label some runs in the Review tab first.`);let n;try{n=z({engineFlag:r.engineFlag??"auto",modelFlag:r.modelFlag,availability:N(),judgeModels:[...new Set(u.map(e=>e.judgeModel).filter(e=>e!=null))]})}catch(e){throw new w("no-engine",e instanceof Error?e.message:String(e))}a(`Reviewer: ${n.rationale}`);let c=0,b=u;if(r.skipReviewed){let e=await f(`/api/eval-targets/${d}/judge-reviews`),t=new Set(e.filter(i=>W(i,n.engine,r.modelFlag)).map(i=>`${i.evalRunId}:${i.expectationIndex}`));if(b=u.filter(i=>!t.has(`${i.evalRunId}:${i.expectationIndex}`)),c=u.length-b.length,b.length===0)return{targetName:v.name,reviewer:n,judgeModels:[],reviewerModel:n.model??`${n.engine} (CLI default)`,report:j([],0),persisted:0,alreadyReviewed:c,deferredPartialRuns:0}}let y=new Map;for(let e=0;;e+=C){let t=await f(`/api/eval-targets/${d}/runs?limit=${C}&offset=${e}`);for(let i of t)y.set(i.id,i);if(t.length<C)break}let q=await f(`/api/eval-targets/${d}/evals`),X=new Map(q.map(e=>[e.id,e.prompt])),o=new Map;for(let e of b){let t=o.get(e.evalRunId)??[];t.push(e),o.set(e.evalRunId,t)}let P=new Map;for(let e of u){let t=P.get(e.evalRunId)??new Set;t.add(e.expectationIndex),P.set(e.evalRunId,t)}let x=0;for(let[e]of[...o]){let t=y.get(e)?.grading?.expectations;if(!t)continue;let i=Z(t),g=P.get(e)??new Set;[...i].every(h=>g.has(h))||(o.delete(e),x+=1)}if(x>0&&a(`${x} partially-labeled run(s) deferred \u2014 a run is re-graded once, after all its assertions are labeled.`),o.size===0)return{targetName:v.name,reviewer:n,judgeModels:[],reviewerModel:n.model??`${n.engine} (CLI default)`,report:j([],0),persisted:0,alreadyReviewed:c,deferredPartialRuns:x};if(n.engine==="codex"&&n.model){let e=await J({prompt:"Reply with exactly: OK",model:n.model,timeout:6e4});if(e.error&&B(e.error)){if(n.model===R)throw new w("preflight-failed",`Codex fallback model ${R} is also unrunnable: ${e.error}`);s(`Your codex model "${n.model}" can't run on the installed CLI/account \u2014 falling back to ${R}.`),n={...n,model:R}}else if(e.error)throw new w("preflight-failed",`Codex preflight failed: ${e.error}`)}let S=[...o.values()].flat(),Y=[...new Set(S.map(e=>e.judgeModel).filter(e=>e!=null))];a(`Re-grading ${o.size} annotated run(s) / ${S.length} assertion(s) on "${v.name}"`+(c?` (${c} already reviewed \u2014 skipped)`:"")+".");let $=[],I=0,k=new Set,m=r.limit!=null?[...o.entries()].slice(0,Math.max(1,r.limit)):[...o.entries()];m.length<o.size&&a(`Sampling ${m.length} of ${o.size} annotated runs (limit).`);let H=0,Q=p(async()=>{for(;;){let e=H++;if(e>=m.length)return;let[t,i]=m[e],g=y.get(t),h=g?.grading?.expectations,O=g?.outputs?.transcript,F=g?X.get(g.evalId):void 0;if(!g||!h?.length||!O||!F){I+=i.length,s(`run ${t.slice(0,8)}: missing output/expectations/prompt \u2014 skipped`);continue}let M=await G({prompt:F,output:O,expectations:h.map(l=>l.text),model:n.model??void 0,timeout:U,engine:n.engine});M.judge_model&&k.add(M.judge_model);for(let l of i){let L=h[l.expectationIndex],A=M.expectations[l.expectationIndex];if(!L||!A||L.text!==l.expectationText){I+=1;continue}if(M.judge_model==null&&A.method==="llm"){I+=1;continue}$.push({runId:t,expectationIndex:l.expectationIndex,text:l.expectationText,humanPassed:D(l.verdict,l.judgePassed),judgeAPassed:l.judgePassed,judgeBPassed:A.passed})}a(`  ${e+1}/${m.length} re-graded`)}},"workerFn");await Promise.all(Array.from({length:Math.max(1,Math.min(K,m.length))},Q));let V=j($,I),_=[...k][0]??n.model??`${n.engine} (CLI default)`,E=0;if($.length>0)try{E=(await T(`/api/eval-targets/${d}/judge-reviews`,{reviewerEngine:n.engine,reviewerModel:_,verdicts:$.map(t=>({evalRunId:t.runId,expectationIndex:t.expectationIndex,expectationText:t.text,judgePassed:t.judgeAPassed,humanPassed:t.humanPassed,reviewerPassed:t.judgeBPassed}))})).upserted,a(`Persisted ${E} reviewer verdict(s) \u2014 disagreements appear on the Review tab.`)}catch(e){s(`Could not persist reviewer verdicts: ${e instanceof Error?e.message:e}`)}return{targetName:v.name,reviewer:n,judgeModels:Y,reviewerModel:_,report:V,persisted:E,alreadyReviewed:c,deferredPartialRuns:x}}p(oe,"runJudgeReviewPass");export{W as a,Z as b,w as c,oe as d};
+import{b as D,c as j}from"./chunk-6UEFCMR2.js";import{a as G}from"./chunk-DVU5PHPM.js";import{d as J}from"./chunk-SG4UDDHB.js";import{b as B,c as R,d as N,e as z}from"./chunk-GPBN4ROE.js";import{e as f,g as T}from"./chunk-76UKAFNM.js";import{a as p}from"./chunk-ZLLER3HP.js";function W(r,a,s){return s?r.reviewerModel===s||r.reviewerModel===`${a}:${s}`:r.reviewerEngine===a}p(W,"reviewMatchesReviewer");var C=100;function Z(r){let a=new Set;return r.forEach((s,d)=>{s.method!=="deterministic"&&a.add(d)}),a}p(Z,"llmExpectationIndexes");var w=class extends Error{constructor(s,d){super(d);this.reason=s}static{p(this,"JudgeReviewSkip")}};async function oe(r){let a=r.log??(()=>{}),s=r.warn??a,d=r.targetId,K=r.parallel??3,U=r.timeoutMs??12e4,v=await f(`/api/eval-targets/${d}`),u=(await f(`/api/eval-targets/${d}/annotations?limit=500`)).filter(e=>e.judgeMethod==="llm");if(u.length===0)throw new w("no-labels",`No human labels on "${v.name}" yet \u2014 label some runs in the Review tab first.`);let n;try{n=z({engineFlag:r.engineFlag??"auto",modelFlag:r.modelFlag,availability:N(),judgeModels:[...new Set(u.map(e=>e.judgeModel).filter(e=>e!=null))]})}catch(e){throw new w("no-engine",e instanceof Error?e.message:String(e))}a(`Reviewer: ${n.rationale}`);let c=0,b=u;if(r.skipReviewed){let e=await f(`/api/eval-targets/${d}/judge-reviews`),t=new Set(e.filter(i=>W(i,n.engine,r.modelFlag)).map(i=>`${i.evalRunId}:${i.expectationIndex}`));if(b=u.filter(i=>!t.has(`${i.evalRunId}:${i.expectationIndex}`)),c=u.length-b.length,b.length===0)return{targetName:v.name,reviewer:n,judgeModels:[],reviewerModel:n.model??`${n.engine} (CLI default)`,report:j([],0),persisted:0,alreadyReviewed:c,deferredPartialRuns:0}}let y=new Map;for(let e=0;;e+=C){let t=await f(`/api/eval-targets/${d}/runs?limit=${C}&offset=${e}`);for(let i of t)y.set(i.id,i);if(t.length<C)break}let q=await f(`/api/eval-targets/${d}/evals`),X=new Map(q.map(e=>[e.id,e.prompt])),o=new Map;for(let e of b){let t=o.get(e.evalRunId)??[];t.push(e),o.set(e.evalRunId,t)}let P=new Map;for(let e of u){let t=P.get(e.evalRunId)??new Set;t.add(e.expectationIndex),P.set(e.evalRunId,t)}let x=0;for(let[e]of[...o]){let t=y.get(e)?.grading?.expectations;if(!t)continue;let i=Z(t),g=P.get(e)??new Set;[...i].every(h=>g.has(h))||(o.delete(e),x+=1)}if(x>0&&a(`${x} partially-labeled run(s) deferred \u2014 a run is re-graded once, after all its assertions are labeled.`),o.size===0)return{targetName:v.name,reviewer:n,judgeModels:[],reviewerModel:n.model??`${n.engine} (CLI default)`,report:j([],0),persisted:0,alreadyReviewed:c,deferredPartialRuns:x};if(n.engine==="codex"&&n.model){let e=await J({prompt:"Reply with exactly: OK",model:n.model,timeout:6e4});if(e.error&&B(e.error)){if(n.model===R)throw new w("preflight-failed",`Codex fallback model ${R} is also unrunnable: ${e.error}`);s(`Your codex model "${n.model}" can't run on the installed CLI/account \u2014 falling back to ${R}.`),n={...n,model:R}}else if(e.error)throw new w("preflight-failed",`Codex preflight failed: ${e.error}`)}let S=[...o.values()].flat(),Y=[...new Set(S.map(e=>e.judgeModel).filter(e=>e!=null))];a(`Re-grading ${o.size} annotated run(s) / ${S.length} assertion(s) on "${v.name}"`+(c?` (${c} already reviewed \u2014 skipped)`:"")+".");let $=[],I=0,k=new Set,m=r.limit!=null?[...o.entries()].slice(0,Math.max(1,r.limit)):[...o.entries()];m.length<o.size&&a(`Sampling ${m.length} of ${o.size} annotated runs (limit).`);let H=0,Q=p(async()=>{for(;;){let e=H++;if(e>=m.length)return;let[t,i]=m[e],g=y.get(t),h=g?.grading?.expectations,O=g?.outputs?.transcript,F=g?X.get(g.evalId):void 0;if(!g||!h?.length||!O||!F){I+=i.length,s(`run ${t.slice(0,8)}: missing output/expectations/prompt \u2014 skipped`);continue}let M=await G({prompt:F,output:O,expectations:h.map(l=>l.text),model:n.model??void 0,timeout:U,engine:n.engine});M.judge_model&&k.add(M.judge_model);for(let l of i){let L=h[l.expectationIndex],A=M.expectations[l.expectationIndex];if(!L||!A||L.text!==l.expectationText){I+=1;continue}if(M.judge_model==null&&A.method==="llm"){I+=1;continue}$.push({runId:t,expectationIndex:l.expectationIndex,text:l.expectationText,humanPassed:D(l.verdict,l.judgePassed),judgeAPassed:l.judgePassed,judgeBPassed:A.passed})}a(`  ${e+1}/${m.length} re-graded`)}},"workerFn");await Promise.all(Array.from({length:Math.max(1,Math.min(K,m.length))},Q));let V=j($,I),_=[...k][0]??n.model??`${n.engine} (CLI default)`,E=0;if($.length>0)try{E=(await T(`/api/eval-targets/${d}/judge-reviews`,{reviewerEngine:n.engine,reviewerModel:_,verdicts:$.map(t=>({evalRunId:t.runId,expectationIndex:t.expectationIndex,expectationText:t.text,judgePassed:t.judgeAPassed,humanPassed:t.humanPassed,reviewerPassed:t.judgeBPassed}))})).upserted,a(`Persisted ${E} reviewer verdict(s) \u2014 disagreements appear on the Review tab.`)}catch(e){s(`Could not persist reviewer verdicts: ${e instanceof Error?e.message:e}`)}return{targetName:v.name,reviewer:n,judgeModels:Y,reviewerModel:_,report:V,persisted:E,alreadyReviewed:c,deferredPartialRuns:x}}p(oe,"runJudgeReviewPass");export{W as a,Z as b,w as c,oe as d};

package/dist/{chunk-FJX7WDPG.js → chunk-I6ZQOMZS.js}

@@ -1 +1 @@
-import{b as i}from"./chunk-ZZHIBOMT.js";import{a as n}from"./chunk-ZLLER3HP.js";var d=new Set([408,429,502,503,504]),R=new Set([500]),l=6e4,_=1e3,h=8e3;function T(e){return e instanceof i&&d.has(e.statusCode)}n(T,"isShedError");function w(e){return e instanceof i&&R.has(e.statusCode)}n(w,"isRetryableServerError");async function g(e,r={}){let m=r.budgetMs??l,u=r.sleep??(o=>new Promise(t=>setTimeout(t,o))),S=Date.now()+m,s=_;for(let o=1;;o++)try{return await e()}catch(t){let a=S-Date.now(),E=r.idempotent===!0&&w(t);if(!T(t)&&!E||a<=0)throw t;let p=t.statusCode,c=Math.min(s,a);r.log?.(`[shed-retry] ${r.label??"api call"} ${E?"hit a server error":"shed"} with ${p} (${t.message}) \u2014 attempt ${o}, retrying in ${c}ms`),await u(c),s=Math.min(s*2,h)}}n(g,"withShedRetry");export{d as a,R as b,l as c,_ as d,h as e,T as f,w as g,g as h};
+import{b as i}from"./chunk-76UKAFNM.js";import{a as n}from"./chunk-ZLLER3HP.js";var d=new Set([408,429,502,503,504]),R=new Set([500]),l=6e4,_=1e3,h=8e3;function T(e){return e instanceof i&&d.has(e.statusCode)}n(T,"isShedError");function w(e){return e instanceof i&&R.has(e.statusCode)}n(w,"isRetryableServerError");async function g(e,r={}){let m=r.budgetMs??l,u=r.sleep??(o=>new Promise(t=>setTimeout(t,o))),S=Date.now()+m,s=_;for(let o=1;;o++)try{return await e()}catch(t){let a=S-Date.now(),E=r.idempotent===!0&&w(t);if(!T(t)&&!E||a<=0)throw t;let p=t.statusCode,c=Math.min(s,a);r.log?.(`[shed-retry] ${r.label??"api call"} ${E?"hit a server error":"shed"} with ${p} (${t.message}) \u2014 attempt ${o}, retrying in ${c}ms`),await u(c),s=Math.min(s*2,h)}}n(g,"withShedRetry");export{d as a,R as b,l as c,_ as d,h as e,T as f,w as g,g as h};

package/dist/{chunk-AT2MMCOJ.js → chunk-IO46ICZY.js}

@@ -1 +1 @@
-import{b as $}from"./chunk-VBXL7XO7.js";import{e as w,f as k}from"./chunk-ZZHIBOMT.js";import{a as u}from"./chunk-ZLLER3HP.js";function h(e){if(e.length===0)return{mean:0,stddev:0,min:0,max:0};let n=e.reduce((s,a)=>s+a,0)/e.length,o=e.reduce((s,a)=>s+(a-n)**2,0)/e.length;return{mean:Math.round(n*1e3)/1e3,stddev:Math.round(Math.sqrt(o)*1e3)/1e3,min:Math.min(...e),max:Math.max(...e)}}u(h,"computeStats");var x="with_target",E="without_target";function _(e){let n=e.grading.summary?.pass_rate;return typeof n=="number"&&n>=1}u(_,"fullPass");function b(e){return e.grading.summary?.pass_rate!=null}u(b,"hasGrading");function R(e){let n=e.timing.total_tokens;return typeof n=="number"&&n>0?n:void 0}u(R,"tokensOf");function v(e){let n=e.timing.duration_ms;return typeof n=="number"&&!isNaN(n)?n/1e3:void 0}u(v,"secondsOf");function y(e,n){let o={};for(let s of e)(o[n(s)]??=[]).push(s);return o}u(y,"groupBy");function C(e,n){let o=y(e.filter(b),i=>i.evalId),s=y(n.filter(b),i=>i.evalId),a=u((i,l)=>i.runNumber-l.runNumber,"byRun"),t=[];for(let i of Object.keys(o)){let l=o[i].sort(a),d=(s[i]??[]).sort(a);for(let r=0;r<Math.min(l.length,d.length);r++){let c={withScore:_(l[r])?1:0,withoutScore:_(d[r])?1:0},f=R(l[r]),g=R(d[r]);f!=null&&g!=null&&(c.withCost=f,c.withoutCost=g);let p=v(l[r]),m=v(d[r]);p!=null&&m!=null&&(c.withLatency=p,c.withoutLatency=m),t.push(c)}}return t}u(C,"pairRunsByEval");function T(e){let n=e.filter(r=>r.status==="completed");if(n.length===0)return{runSummary:{},verdict:null,notes:["No completed runs found"],totalRuns:0};let o=y(n,r=>r.configuration),s={},a=[];for(let[r,c]of Object.entries(o)){let f=c.filter(b).map(m=>m.grading.summary.pass_rate),g=c.map(v).filter(m=>m!=null),p=c.map(R).filter(m=>m!=null);s[r]={},f.length>0&&(s[r].pass_rate=h(f)),g.length>0&&(s[r].time_seconds=h(g)),p.length>0&&(s[r].tokens=h(p))}let t=null,i=o[x]??[],l=o[E]??[],d=C(i,l);if(d.length>0){if(t=$(d),!t.sufficient)a.push(t.insufficientReason??"Insufficient data for a significance verdict.");else if(t.significant){let r=t.delta>=0?"improves":"reduces",c=t.pValue!=null?`, McNemar p=${t.pValue.toFixed(4)}`:"";a.push(`Target ${r} the pass rate by ${Math.abs(t.delta*100).toFixed(0)}% (95% CI ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%, n=${t.nWith}${c}).`)}else a.push(`No significant effect on pass rate (95% CI ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%, n=${t.nWith}).`);t.costDelta!=null&&Math.round(t.costDelta)!==0&&a.push(`Token cost delta: ${t.costDelta>=0?"+":""}${Math.round(t.costDelta)} tokens/task.`),t.latencyDelta!=null&&Math.abs(t.latencyDelta)>1&&a.push(`Target ${t.latencyDelta>0?"adds":"saves"} ${Math.abs(t.latencyDelta).toFixed(1)}s/task.`)}for(let r of Object.values(s))if(r.pass_rate&&r.pass_rate.stddev>.2){a.push(`High pass-rate variance (stddev=${r.pass_rate.stddev.toFixed(2)}) \u2014 results may be unreliable.`);break}return{runSummary:s,verdict:t,notes:a,totalRuns:n.length}}u(T,"summarizeRuns");async function S(e,n){let o=n?`&iteration=${n}`:"",s=100,a=[];for(let t=0;t<1e5;t+=s){let i=await w(`/api/eval-targets/${e}/runs?limit=${s}&offset=${t}&activeCasesOnly=1${o}`);if(a.push(...i),i.length<s)break}return T(a)}u(S,"aggregateBenchmark");async function B(e,n,o){return k(`/api/eval-targets/${e}/benchmarks`,{iteration:n,runSummary:o.runSummary,notes:o.notes})}u(B,"postBenchmark");export{T as a,S as b,B as c};
+import{b as $}from"./chunk-VBXL7XO7.js";import{e as w,f as k}from"./chunk-76UKAFNM.js";import{a as u}from"./chunk-ZLLER3HP.js";function h(e){if(e.length===0)return{mean:0,stddev:0,min:0,max:0};let n=e.reduce((s,a)=>s+a,0)/e.length,o=e.reduce((s,a)=>s+(a-n)**2,0)/e.length;return{mean:Math.round(n*1e3)/1e3,stddev:Math.round(Math.sqrt(o)*1e3)/1e3,min:Math.min(...e),max:Math.max(...e)}}u(h,"computeStats");var x="with_target",E="without_target";function _(e){let n=e.grading.summary?.pass_rate;return typeof n=="number"&&n>=1}u(_,"fullPass");function b(e){return e.grading.summary?.pass_rate!=null}u(b,"hasGrading");function R(e){let n=e.timing.total_tokens;return typeof n=="number"&&n>0?n:void 0}u(R,"tokensOf");function v(e){let n=e.timing.duration_ms;return typeof n=="number"&&!isNaN(n)?n/1e3:void 0}u(v,"secondsOf");function y(e,n){let o={};for(let s of e)(o[n(s)]??=[]).push(s);return o}u(y,"groupBy");function C(e,n){let o=y(e.filter(b),i=>i.evalId),s=y(n.filter(b),i=>i.evalId),a=u((i,l)=>i.runNumber-l.runNumber,"byRun"),t=[];for(let i of Object.keys(o)){let l=o[i].sort(a),d=(s[i]??[]).sort(a);for(let r=0;r<Math.min(l.length,d.length);r++){let c={withScore:_(l[r])?1:0,withoutScore:_(d[r])?1:0},f=R(l[r]),g=R(d[r]);f!=null&&g!=null&&(c.withCost=f,c.withoutCost=g);let p=v(l[r]),m=v(d[r]);p!=null&&m!=null&&(c.withLatency=p,c.withoutLatency=m),t.push(c)}}return t}u(C,"pairRunsByEval");function T(e){let n=e.filter(r=>r.status==="completed");if(n.length===0)return{runSummary:{},verdict:null,notes:["No completed runs found"],totalRuns:0};let o=y(n,r=>r.configuration),s={},a=[];for(let[r,c]of Object.entries(o)){let f=c.filter(b).map(m=>m.grading.summary.pass_rate),g=c.map(v).filter(m=>m!=null),p=c.map(R).filter(m=>m!=null);s[r]={},f.length>0&&(s[r].pass_rate=h(f)),g.length>0&&(s[r].time_seconds=h(g)),p.length>0&&(s[r].tokens=h(p))}let t=null,i=o[x]??[],l=o[E]??[],d=C(i,l);if(d.length>0){if(t=$(d),!t.sufficient)a.push(t.insufficientReason??"Insufficient data for a significance verdict.");else if(t.significant){let r=t.delta>=0?"improves":"reduces",c=t.pValue!=null?`, McNemar p=${t.pValue.toFixed(4)}`:"";a.push(`Target ${r} the pass rate by ${Math.abs(t.delta*100).toFixed(0)}% (95% CI ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%, n=${t.nWith}${c}).`)}else a.push(`No significant effect on pass rate (95% CI ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%, n=${t.nWith}).`);t.costDelta!=null&&Math.round(t.costDelta)!==0&&a.push(`Token cost delta: ${t.costDelta>=0?"+":""}${Math.round(t.costDelta)} tokens/task.`),t.latencyDelta!=null&&Math.abs(t.latencyDelta)>1&&a.push(`Target ${t.latencyDelta>0?"adds":"saves"} ${Math.abs(t.latencyDelta).toFixed(1)}s/task.`)}for(let r of Object.values(s))if(r.pass_rate&&r.pass_rate.stddev>.2){a.push(`High pass-rate variance (stddev=${r.pass_rate.stddev.toFixed(2)}) \u2014 results may be unreliable.`);break}return{runSummary:s,verdict:t,notes:a,totalRuns:n.length}}u(T,"summarizeRuns");async function S(e,n){let o=n?`&iteration=${n}`:"",s=100,a=[];for(let t=0;t<1e5;t+=s){let i=await w(`/api/eval-targets/${e}/runs?limit=${s}&offset=${t}&activeCasesOnly=1${o}`);if(a.push(...i),i.length<s)break}return T(a)}u(S,"aggregateBenchmark");async function B(e,n,o){return k(`/api/eval-targets/${e}/benchmarks`,{iteration:n,runSummary:o.runSummary,notes:o.notes})}u(B,"postBenchmark");export{T as a,S as b,B as c};