@rulemetric/cli 0.7.34 → 0.7.36

package/dist/{chunk-BZUIMHHT.js → chunk-TVHT6H42.js}

@@ -1,4 +1,4 @@
-import{a as S,b as z,c as T}from"./chunk-QFVGUZDO.js";import{a as O}from"./chunk-ZL4MZMUK.js";import{a as j}from"./chunk-5RBLIVBF.js";import{c as A,d as C,f as b}from"./chunk-MODFTCZY.js";import{e as _,f as I,h as u}from"./chunk-H2ORWIEY.js";import{a as $}from"./chunk-ZLLER3HP.js";async function D(f,m){let{mirrorId:o,projectPath:s,projectId:i,config:p}=f,e=$(n=>m.logger.info(n),"log");if(p?.sessionId)return M({mirrorId:o,sessionId:p.sessionId},e);e(`Processing insights job ${o}...`),await u(`/api/insights-jobs/${o}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:n}=await I("/api/sessions/repair-analyzed-at",{});n>0&&e(`  Repaired analyzed_at on ${n} orphan session(s).`)}catch(n){e(`  Warning: analyzed_at repair failed: ${n instanceof Error?n.message:String(n)}`)}let a=await J(s,i);if(!a)throw new j("No session data found \u2014 ensure sessions are captured and enriched.");e(`  Found ${a.stats.sessions} sessions, generating recommendations...`);let w=20,c=200;if(C()!=="none"){let n=0,t=0;try{for(;n<c;){let l=await _(`/api/sessions?hasAnalysis=false&limit=${w}&offset=${t}`);if(l.length===0)break;e(`  Generating facets for ${l.length} un-analyzed sessions (drained ${n} so far)...`);for(let g of l){if(n>=c)break;try{let d=await _(`/api/sessions/${g.id}/events?limit=100`);if(d.length<2){n++,t++;continue}let P=d.filter(h=>h.eventType==="user_message"||h.eventType==="human").map(h=>h.content?.slice(0,200)??"").filter(Boolean).join(`
+import{a as S,b as z,c as T}from"./chunk-QFVGUZDO.js";import{a as j}from"./chunk-5RBLIVBF.js";import{a as O}from"./chunk-ZL4MZMUK.js";import{c as A,d as C,f as b}from"./chunk-B3BHB26P.js";import{e as _,f as I,h as u}from"./chunk-H2ORWIEY.js";import{a as $}from"./chunk-ZLLER3HP.js";async function D(f,m){let{mirrorId:o,projectPath:s,projectId:i,config:p}=f,e=$(n=>m.logger.info(n),"log");if(p?.sessionId)return M({mirrorId:o,sessionId:p.sessionId},e);e(`Processing insights job ${o}...`),await u(`/api/insights-jobs/${o}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:n}=await I("/api/sessions/repair-analyzed-at",{});n>0&&e(`  Repaired analyzed_at on ${n} orphan session(s).`)}catch(n){e(`  Warning: analyzed_at repair failed: ${n instanceof Error?n.message:String(n)}`)}let a=await J(s,i);if(!a)throw new j("No session data found \u2014 ensure sessions are captured and enriched.");e(`  Found ${a.stats.sessions} sessions, generating recommendations...`);let w=20,c=200;if(C()!=="none"){let n=0,t=0;try{for(;n<c;){let l=await _(`/api/sessions?hasAnalysis=false&limit=${w}&offset=${t}`);if(l.length===0)break;e(`  Generating facets for ${l.length} un-analyzed sessions (drained ${n} so far)...`);for(let g of l){if(n>=c)break;try{let d=await _(`/api/sessions/${g.id}/events?limit=100`);if(d.length<2){n++,t++;continue}let P=d.filter(h=>h.eventType==="user_message"||h.eventType==="human").map(h=>h.content?.slice(0,200)??"").filter(Boolean).join(`
 ---
 `);if(!P){n++,t++;continue}let E=await b(T,P,4096);if(E){let h=d.filter(R=>R.eventType==="user_message"||R.eventType==="human").length;await u(`/api/sessions/${g.id}/enrich`,{analysis:{source:"generated_facets",...E},analyzedEventCount:d.length,analyzedTurnCount:h}),e(`    Generated facets for session ${g.id} (${g.tool})`)}else t++;n++}catch(d){e(`    Failed to generate facets for session ${g.id}: ${d instanceof Error?d.message:String(d)}`),n++,t++}}if(l.length<w)break}}catch(l){e(`  Warning: facet generation failed after ${n} sessions: ${l instanceof Error?l.message:String(l)}`)}}let r={},y=S.length+1;for(let n=0;n<S.length;n++){let t=S[n];e(`  [${n+1}/${y}] Generating ${t.name}...`);try{await u(`/api/insights-jobs/${o}/progress`,{status:"running",progress:{step:t.name,current:n+1,total:y,partial:{...r}}})}catch(g){e(`  Warning: progress update failed for ${t.name}: ${g instanceof Error?g.message:String(g)}`)}e(`  Running LLM prompt for ${t.name}...`);let l=await b(t.prompt,a.summary,t.maxTokens);l?(r[t.name]=l,e(`  Got result for ${t.name}`)):e(`  No result for ${t.name} (LLM returned null)`)}e(`  [${y}/${y}] Generating at_a_glance...`),await u(`/api/insights-jobs/${o}/progress`,{status:"running",progress:{step:"at_a_glance",current:y,total:y,partial:{...r}}});let N=[a.summary,r.what_works?`
 ## What Works

package/dist/{chunk-LETTPEG4.js → chunk-VSUL3MNG.js}

@@ -1 +1 @@
-import{a as $}from"./chunk-WKXTY5AM.js";import{e as w,f as k}from"./chunk-H2ORWIEY.js";import{a as u}from"./chunk-ZLLER3HP.js";function h(e){if(e.length===0)return{mean:0,stddev:0,min:0,max:0};let n=e.reduce((s,a)=>s+a,0)/e.length,o=e.reduce((s,a)=>s+(a-n)**2,0)/e.length;return{mean:Math.round(n*1e3)/1e3,stddev:Math.round(Math.sqrt(o)*1e3)/1e3,min:Math.min(...e),max:Math.max(...e)}}u(h,"computeStats");var x="with_target",E="without_target";function _(e){let n=e.grading.summary?.pass_rate;return typeof n=="number"&&n>=1}u(_,"fullPass");function b(e){return e.grading.summary?.pass_rate!=null}u(b,"hasGrading");function R(e){let n=e.timing.total_tokens;return typeof n=="number"&&n>0?n:void 0}u(R,"tokensOf");function v(e){let n=e.timing.duration_ms;return typeof n=="number"&&!isNaN(n)?n/1e3:void 0}u(v,"secondsOf");function y(e,n){let o={};for(let s of e)(o[n(s)]??=[]).push(s);return o}u(y,"groupBy");function C(e,n){let o=y(e.filter(b),i=>i.evalId),s=y(n.filter(b),i=>i.evalId),a=u((i,l)=>i.runNumber-l.runNumber,"byRun"),t=[];for(let i of Object.keys(o)){let l=o[i].sort(a),d=(s[i]??[]).sort(a);for(let r=0;r<Math.min(l.length,d.length);r++){let c={withScore:_(l[r])?1:0,withoutScore:_(d[r])?1:0},f=R(l[r]),g=R(d[r]);f!=null&&g!=null&&(c.withCost=f,c.withoutCost=g);let p=v(l[r]),m=v(d[r]);p!=null&&m!=null&&(c.withLatency=p,c.withoutLatency=m),t.push(c)}}return t}u(C,"pairRunsByEval");function T(e){let n=e.filter(r=>r.status==="completed");if(n.length===0)return{runSummary:{},verdict:null,notes:["No completed runs found"],totalRuns:0};let o=y(n,r=>r.configuration),s={},a=[];for(let[r,c]of Object.entries(o)){let f=c.filter(b).map(m=>m.grading.summary.pass_rate),g=c.map(v).filter(m=>m!=null),p=c.map(R).filter(m=>m!=null);s[r]={},f.length>0&&(s[r].pass_rate=h(f)),g.length>0&&(s[r].time_seconds=h(g)),p.length>0&&(s[r].tokens=h(p))}let t=null,i=o[x]??[],l=o[E]??[],d=C(i,l);if(d.length>0){if(t=$(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 $}from"./chunk-G5BXZT6F.js";import{e as w,f as k}from"./chunk-H2ORWIEY.js";import{a as u}from"./chunk-ZLLER3HP.js";function h(e){if(e.length===0)return{mean:0,stddev:0,min:0,max:0};let n=e.reduce((s,a)=>s+a,0)/e.length,o=e.reduce((s,a)=>s+(a-n)**2,0)/e.length;return{mean:Math.round(n*1e3)/1e3,stddev:Math.round(Math.sqrt(o)*1e3)/1e3,min:Math.min(...e),max:Math.max(...e)}}u(h,"computeStats");var x="with_target",E="without_target";function _(e){let n=e.grading.summary?.pass_rate;return typeof n=="number"&&n>=1}u(_,"fullPass");function b(e){return e.grading.summary?.pass_rate!=null}u(b,"hasGrading");function R(e){let n=e.timing.total_tokens;return typeof n=="number"&&n>0?n:void 0}u(R,"tokensOf");function v(e){let n=e.timing.duration_ms;return typeof n=="number"&&!isNaN(n)?n/1e3:void 0}u(v,"secondsOf");function y(e,n){let o={};for(let s of e)(o[n(s)]??=[]).push(s);return o}u(y,"groupBy");function C(e,n){let o=y(e.filter(b),i=>i.evalId),s=y(n.filter(b),i=>i.evalId),a=u((i,l)=>i.runNumber-l.runNumber,"byRun"),t=[];for(let i of Object.keys(o)){let l=o[i].sort(a),d=(s[i]??[]).sort(a);for(let r=0;r<Math.min(l.length,d.length);r++){let c={withScore:_(l[r])?1:0,withoutScore:_(d[r])?1:0},f=R(l[r]),g=R(d[r]);f!=null&&g!=null&&(c.withCost=f,c.withoutCost=g);let p=v(l[r]),m=v(d[r]);p!=null&&m!=null&&(c.withLatency=p,c.withoutLatency=m),t.push(c)}}return t}u(C,"pairRunsByEval");function T(e){let n=e.filter(r=>r.status==="completed");if(n.length===0)return{runSummary:{},verdict:null,notes:["No completed runs found"],totalRuns:0};let o=y(n,r=>r.configuration),s={},a=[];for(let[r,c]of Object.entries(o)){let f=c.filter(b).map(m=>m.grading.summary.pass_rate),g=c.map(v).filter(m=>m!=null),p=c.map(R).filter(m=>m!=null);s[r]={},f.length>0&&(s[r].pass_rate=h(f)),g.length>0&&(s[r].time_seconds=h(g)),p.length>0&&(s[r].tokens=h(p))}let t=null,i=o[x]??[],l=o[E]??[],d=C(i,l);if(d.length>0){if(t=$(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};

package/dist/chunk-VTA6HDVW.js

@@ -0,0 +1,11 @@
+import{a as R,g as v,h as T,j as I,m as C}from"./chunk-VLXCQ2DC.js";import{e as d}from"./chunk-QWCEZEDO.js";import{d as g}from"./chunk-NTKO4H2K.js";import{e as _}from"./chunk-H2ORWIEY.js";import{g as m,l as k}from"./chunk-O6C2UGBY.js";import{a as i}from"./chunk-ZLLER3HP.js";import{execSync as j}from"node:child_process";import{existsSync as p,readFileSync as y}from"node:fs";import{homedir as f}from"node:os";import{join as a}from"node:path";var P=["install","auth","hooks","worker"],O=["judge"],M=[...P,...O];function $(){return{projectPath:process.cwd(),configDir:k()}}i($,"defaultContext");var A=a(f(),".mitmproxy","mitmproxy-ca-cert.pem");function S(e){if(!e)return"";if(e.length<=8)return"*".repeat(e.length);let t=e.slice(0,4),n=e.slice(-4);return`${t}${"*".repeat(Math.max(4,e.length-8))}${n}`}i(S,"maskToken");function te(e,t){let n=e===null?[]:e.replace(/\n$/,"").split(`
+`),o=t===null?[]:t.replace(/\n$/,"").split(`
+`),s=[],u=Math.max(n.length,o.length);for(let c=0;c<u;c++){let l=n[c],r=o[c];l===r?s.push(` ${l}`):(l!==void 0&&s.push(`-${l}`),r!==void 0&&s.push(`+${r}`))}return s.join(`
+`)}i(te,"renderDiff");function x(e,t){let n=a(e,"env");if(!p(n))return null;try{let s=y(n,"utf-8").match(new RegExp(`^${t}=(.+)$`,"m"));return s?s[1]:null}catch{return null}}i(x,"readEnvValue");async function b(){if(!m())return!1;try{return(await Promise.race([_("/api/setup/status"),new Promise((t,n)=>setTimeout(()=>n(new Error("setup/status timeout")),2500))]))?.workerActive===!0}catch{return!1}}i(b,"workerActive");function E(e,t){let n=a(e,".claude",t),o={},s=null;if(p(n))try{s=y(n,"utf-8").replace(/\n$/,""),o=JSON.parse(s)}catch{o={}}return{settings:o,before:s}}i(E,"readProjectSettings");function D(e){let{settings:t,before:n}=E(e,"settings.local.json");return v(t),{before:n,after:JSON.stringify(t,null,2)}}i(D,"projectHooksSettings");function N(e){let{settings:t,before:n}=E(e,"settings.local.json"),o={...t.env??{}};return o.HTTPS_PROXY=`http://localhost:${d}`,o.NO_PROXY=I,o.NODE_EXTRA_CA_CERTS=A,{before:n,after:JSON.stringify({...t,env:o},null,2)}}i(N,"projectProxyEnvSettings");function U(e){let{settings:t,before:n}=E(e,"settings.json");if(n===null)return null;let{removed:o}=C(t,d);return o?{before:n,after:JSON.stringify(t,null,2)}:null}i(U,"projectSharedSettingsMigration");var Y=a(f(),"Library","LaunchAgents","com.rulemetric.worker.plist"),G=a(f(),".config","systemd","user","rulemetric-worker.service");function h(e){j(e,{stdio:"inherit"})}i(h,"runCommand");var L=[{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",async detect(){return m()||await b()},async plan(){return{id:"install",title:"Install the CLI",command:"npm install -g @rulemetric/cli",artifacts:[{path:"npm install -g @rulemetric/cli",kind:"command",before:null,after:"rulemetric --version   # verifies the binary is on PATH"}]}},async apply(){h("npm install -g @rulemetric/cli")}},{id:"auth",title:"Authenticate",command:"rulemetric auth login",async detect(){return m()},async plan(e){let t=a(e.configDir,"env"),n=null;if(p(t))try{n=X(y(t,"utf-8"))}catch{n="(unreadable)"}let o=x(e.configDir,"RULEMETRIC_ACCESS_TOKEN"),u=[`RULEMETRIC_ACCESS_TOKEN=${S(o||"<token-from-login>")}`],c=x(e.configDir,"RULEMETRIC_API_KEY");c&&u.push(`RULEMETRIC_API_KEY=${S(c)}`);let l=process.env.RULEMETRIC_API_URL;l&&u.push(`RULEMETRIC_API_URL=${l}`);let r=u.join(`
+`);return{id:"auth",title:"Authenticate",command:"rulemetric auth login",artifacts:[{path:t,kind:"file",before:n,after:r,masked:!0},{path:a(e.configDir,"auth.json"),kind:"file",before:p(a(e.configDir,"auth.json"))?"(existing session tokens)":null,after:"{ accessToken, refreshToken, expiresAt, email }  \u2014 OAuth session tokens (secret)",masked:!0}]}},async apply(){h("rulemetric auth login")}},{id:"hooks",title:"Install hooks + capture proxy",command:process.env.RULEMETRIC_SETUP_NO_PROXY==="1"||process.env.RULEMETRIC_SETUP_NO_PROXY==="true"?"rulemetric hooks install --no-proxy":"rulemetric hooks install",async detect(e){for(let t of["settings.local.json","settings.json"]){let n=a(e.projectPath,".claude",t);if(p(n))try{let o=JSON.parse(y(n,"utf-8"));if(T(o))return!0}catch{}}return!1},async plan(e){let t=a(e.projectPath,".claude","settings.local.json"),{before:n,after:o}=D(e.projectPath),s=[{path:t,kind:"file",before:n,after:o},{path:"what these hooks capture",kind:"service",before:null,after:"session-end reimports the FULL transcript to the API; post-tool-use records each tool call and snapshots the working tree to a LOCAL shadow git ref (paper-trail, never pushed); user-prompt posts prompt text. All scoped to sessions in this project."},{path:`${t} (proxy env block)`,kind:"file",...N(e.projectPath)},{path:"system keychain \u2014 mitmproxy root CA",kind:"service",before:p(A)?"(CA generated; trust state unchanged)":"(no CA yet)",after:"Installs + trusts the mitmproxy CA (requires sudo). This lets the proxy DECRYPT Claude Code HTTPS request/response bodies \u2014 the capability that powers rendered-system-prompt / instruction linking. Decline the sudo prompt and setup stays hooks-only (still fully onboarded)."},{path:"launchctl setenv HTTPS_PROXY (machine-wide, all GUI apps)",kind:"command",before:null,after:`http://localhost:${d}  \u2014 routes every GUI app's HTTPS through the local gateway (only when the CA is trusted AND :${d} is live)`},{path:"VS Code + Cursor user settings (http.proxy)",kind:"file",before:null,after:`http.proxy \u2192 http://localhost:${d} (+ http.proxyStrictSSL=false)`},{path:`gateway process on :${d}`,kind:"service",before:null,after:"started (routes Claude Code \u2192 mitmproxy \u2192 Anthropic; falls back to direct if mitmproxy is down)"}],u=U(e.projectPath);u&&s.push({path:`${a(e.projectPath,".claude","settings.json")} (removes RuleMetric machine-local state left by an older install \u2014 see the local-file artifacts above for where it moves)`,kind:"file",...u});for(let c of R(e.projectPath).filter(l=>l==="cursor"||l==="copilot_agent"))s.push({path:`${e.projectPath} (${c} workspace proxy config)`,kind:"file",before:null,after:`proxy \u2192 http://localhost:${d}`});for(let[c,l,r]of[["Cursor",a(f(),".cursor"),a(f(),".cursor","hooks.json")],["Antigravity/Gemini",a(f(),".gemini"),a(f(),".gemini","config","hooks.json")]])p(l)&&s.push({path:`${r} (${c} user hooks)`,kind:"file",before:p(r)?"(existing config)":null,after:"adds RuleMetric session-capture hooks for this tool"});return{id:"hooks",title:"Install hooks + capture proxy",command:"rulemetric hooks install",artifacts:s}},async apply(){let e=process.env.RULEMETRIC_SETUP_NO_PROXY==="1"||process.env.RULEMETRIC_SETUP_NO_PROXY==="true";h(e?"rulemetric hooks install --no-proxy":"rulemetric hooks install")}},{id:"worker",title:"Run the worker",command:"rulemetric service install --worker-only",async detect(){return b()},async plan(){let e=process.platform==="darwin",t=e?Y:G,n=p(t)?`(service already installed at ${t})`:null,o=e?["Label: com.rulemetric.worker","ProgramArguments: <node> <cli>/bin/run.js evals agent","RunAtLoad: true, KeepAlive: true"].join(`
+`):["[Service]","ExecStart=<node> <cli>/bin/run.js evals agent","Restart=on-failure"].join(`
+`);return{id:"worker",title:"Run the worker",command:"rulemetric service install --worker-only",artifacts:[{path:t,kind:"service",before:n,after:o}]}},async apply(){h("rulemetric service install --worker-only")}},{id:"judge",title:"Second LLM engine for cross-vendor eval grading (optional)",command:"npm install -g @openai/codex",optional:!0,async detect(){let e=g();return e.claudeBin!=null&&e.codexBin!=null},async plan(){let e=g();return{id:"judge",title:"Second LLM engine for cross-vendor eval grading (optional)",command:"npm install -g @openai/codex",artifacts:[{path:"installed LLM engines",kind:"command",before:[`claude CLI: ${e.claudeBin??"not found"}`,`codex CLI: ${e.codexBin??"not found"}`].join(`
+`),after:["claude CLI + codex CLI both installed.","Hand labels + one judge already give calibrated verdicts; a second","engine additionally unlocks cross-vendor judge validation","(`rulemetric evals validate-judge`)."].join(`
+`)}]}},async apply(){let e=g();if(e.claudeBin!=null&&e.codexBin!=null){console.log("[OK] Both engines already installed \u2014 nothing to do.");return}let t=e.codexBin==null?"codex":"claude";console.log(`[..] Second GRADING engine not detected (missing: ${t} CLI).`),t==="codex"?(console.log("     This is about GRADING EVALS with a second vendor, not about capturing"),console.log("     your Codex sessions. Codex capture needs no setup: the worker imports"),console.log("     ~/.codex/sessions automatically. Nothing is missing for that."),console.log("     Optional install: npm install -g @openai/codex")):console.log("     Install it yourself when ready: npm install -g @anthropic-ai/claude-code"),console.log("     Then re-run: rulemetric setup judge --show")}}],H=new Map(L.map(e=>[e.id,e]));function ne(e){let t=H.get(e);if(!t)throw new Error(`Unknown setup step: ${e}`);return t}i(ne,"getStep");function oe(e){return M.includes(e)}i(oe,"isStepId");var K=new Set(["RULEMETRIC_API_URL","RULEMETRIC_USER_ID","RULEMETRIC_CLIENT_NAME","RESEND_FROM_EMAIL","SENTRY_DSN","PG_POOL_MAX","PG_IDLE_TIMEOUT","PG_STATEMENT_TIMEOUT_MS","NODE_ENV"]);function X(e){return e.replace(/\n$/,"").split(`
+`).map(t=>{let n=t.match(/^([A-Z0-9_]+)=(.+)$/);if(!n)return t;let[,o,s]=n;return K.has(o)?t:`${o}=${S(s)}`}).join(`
+`)}i(X,"maskEnvFile");async function re(e=$()){let t=[];for(let r of L){let w=await r.detect(e);t.push({id:r.id,done:w})}let n=new Set(P),o=t.find(r=>r.id==="install");o&&!o.done&&(o.done=t.some(r=>r.id!=="install"&&n.has(r.id)&&r.done));let s=new Set(P),u=t.filter(r=>s.has(r.id)),c=u.find(r=>!r.done)?.id??null,l=u.every(r=>r.done);return{steps:t,coreComplete:l,currentStepId:c}}i(re,"computeStatus");export{P as a,O as b,M as c,$ as d,S as e,te as f,L as g,ne as h,oe as i,re as j};

package/dist/{chunk-OZQV5CDT.js → chunk-X64R6JCJ.js}

@@ -1 +1 @@
-import{b as D,c as j}from"./chunk-6UEFCMR2.js";import{a as G}from"./chunk-7AZKPX2J.js";import{c as J}from"./chunk-V5MJ4ICC.js";import{b as B,c as R,d as N,e as z}from"./chunk-OPUI7CM2.js";import{e as f,g as T}from"./chunk-H2ORWIEY.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: ${e instanceof Error?e.message:e}`)}return{targetName:v.name,reviewer:n,judgeModels:Y,reviewerModel:_,report:V,persisted:E,alreadyReviewed:c,deferredPartialRuns:x}}p(oe,"runJudgeReviewPass");export{W as a,Z as b,w as c,oe as d};
+import{b as D,c as j}from"./chunk-6UEFCMR2.js";import{b as B,c as R,d as N,e as z}from"./chunk-NTKO4H2K.js";import{a as G}from"./chunk-4L2ZEIRE.js";import{c as J}from"./chunk-XRTZU7IZ.js";import{e as f,g as T}from"./chunk-H2ORWIEY.js";import{a as p}from"./chunk-ZLLER3HP.js";function W(r,a,s){return s?r.reviewerModel===s||r.reviewerModel===`${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-G6XGDHGC.js → chunk-XA4KSQEC.js}

@@ -1,2 +1,2 @@
-import{c as E,o as m}from"./chunk-7DFNBRLY.js";import{a as d}from"./chunk-H2ORWIEY.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,o as m}from"./chunk-5DIOZDDZ.js";import{a as d}from"./chunk-H2ORWIEY.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-V5MJ4ICC.js → chunk-XRTZU7IZ.js}

@@ -1,3 +1,3 @@
-import{a as w}from"./chunk-MODFTCZY.js";import{a as r}from"./chunk-ZLLER3HP.js";import{spawn as k}from"node:child_process";import{mkdtempSync as v,readFileSync as M,rmSync as S}from"node:fs";import{tmpdir as $}from"node:os";import{join as P}from"node:path";var D=r(o=>`codex:${o}`,"codexModelId");function R(o,i){return["exec","--sandbox","read-only","--skip-git-repo-check","--ephemeral","--model",o,"--output-last-message",i,"-"]}r(R,"buildCodexArgs");async function j(o){let{prompt:i,systemPrompt:l,model:c,timeout:u=12e4,workDir:b=process.cwd()}=o,T=l?`${l}
+import{a as w}from"./chunk-B3BHB26P.js";import{a as r}from"./chunk-ZLLER3HP.js";import{spawn as k}from"node:child_process";import{mkdtempSync as v,readFileSync as M,rmSync as S}from"node:fs";import{tmpdir as $}from"node:os";import{join as P}from"node:path";var D=r(o=>`codex:${o}`,"codexModelId");function R(o,i){return["exec","--sandbox","read-only","--skip-git-repo-check","--ephemeral","--model",o,"--output-last-message",i,"-"]}r(R,"buildCodexArgs");async function j(o){let{prompt:i,systemPrompt:l,model:c,timeout:u=12e4,workDir:b=process.cwd()}=o,T=l?`${l}
 
 ${i}`:i,m=w("codex");if(!m)return{output:"",exitCode:null,durationMs:0,error:"Codex CLI not found on PATH, well-known install dirs, or the node bin dir \u2014 is it installed?"};let p=v(P($(),"rulemetric-codex-")),a=P(p,"last-message.txt"),h=Date.now();return new Promise(E=>{let s={...process.env};delete s.HTTPS_PROXY;let y=["/opt/homebrew/bin","/usr/local/bin"],x=(s.PATH??"").split(":").filter(Boolean),f=y.filter(e=>!x.includes(e));f.length&&(s.PATH=[...x,...f].join(":"));let t=k(m,R(c,a),{cwd:b,env:s,stdio:["pipe","pipe","pipe"]}),g="";t.stdout.on("data",()=>{}),t.stderr.on("data",e=>{g+=e.toString()});let n=r(e=>{S(p,{recursive:!0,force:!0}),E({...e,durationMs:Date.now()-h})},"finish"),C=setTimeout(()=>{t.kill("SIGTERM"),n({output:"",exitCode:null,error:`codex exec timed out after ${u}ms`})},u);t.on("error",e=>{clearTimeout(C),n({output:"",exitCode:null,error:`failed to spawn codex: ${e.message} (is the Codex CLI installed?)`})}),t.on("close",e=>{clearTimeout(C);let d="";try{d=M(a,"utf8").trim()}catch{}if(e!==0||d===""){n({output:d,exitCode:e,error:`codex exec failed (exit ${e}): ${g.trim().slice(-500)||"no last message produced"}`});return}n({output:d,exitCode:e,model:D(c)})}),t.stdin.write(T),t.stdin.end()})}r(j,"executeCodexEval");export{D as a,R as b,j as c};

package/dist/chunk-ZYI66XKV.js

@@ -0,0 +1 @@
+import{f as Z}from"./chunk-RMDLJN4L.js";import{b as B,i as V,j as X,k as J,l as Y,m as z,n as K}from"./chunk-BL4BN6RU.js";import{e as Q}from"./chunk-2ZAKZEXW.js";import{d as G,e as S,f as W}from"./chunk-G5BXZT6F.js";import{e as $,f as q}from"./chunk-H2ORWIEY.js";import{a as c}from"./chunk-ZLLER3HP.js";import{createHash as ie}from"node:crypto";var _=c(e=>ie("sha256").update(e).digest("hex"),"sha256"),ee=1.3,H=4,le=20;function ue(e,r){if(!e.configured||r<=0)return H;let n=Math.ceil(e.minPairs*ee/r);return Math.min(le,Math.max(H,n))}c(ue,"passCapFor");function ce(e,r=10,n=H){if(e<=0)return{passes:0,expectedPairs:0,feasible:!1,reason:"no live cases"};let s=Math.ceil(r*ee),t=Math.min(n,Math.ceil(s/e)),a=t*e;return a<r?{passes:t,expectedPairs:a,feasible:!1,reason:`${e} live case(s) x ${n} passes = ${e*n} pairs at best, below the ${r} a controlled verdict requires \u2014 add cases rather than replaying these`}:{passes:t,expectedPairs:a,feasible:!0}}c(ce,"planPasses");var pe=60,de=c(e=>typeof e=="object"&&e!==null&&typeof e.withScore=="number"&&typeof e.withoutScore=="number","isPair");function ge(e,r,n){let s=[...e].sort((t,a)=>Date.parse(a.createdAt)-Date.parse(t.createdAt));for(let t of s){if(t.applied)continue;let a=t.gates?.accumulation;if(!a||typeof a!="object"||a.priorContentHash!==r||a.targetPairs!==n||typeof a.candidateContentHash!="string"||!Array.isArray(a.pairs))continue;let p=a.pairs.filter(de);if(p.length===a.pairs.length&&!(p.length>=n))return{promotionId:t.id,candidateContentHash:a.candidateContentHash,pairs:p,runs:typeof a.runs=="number"?a.runs:1}}return null}c(ge,"findOpenAccumulation");function me(e,r){let n=r.accumulate??(e.autoEvolveAccumulate===!0||e.autoEvolveAccumulate==="true"),s=r.accumulateTargetPairs??e.autoEvolveAccumulateTarget,t=typeof s=="string"?Number.parseInt(s,10):s,a=typeof t=="number"&&Number.isInteger(t)&&t>0?t:pe;return{enabled:n,targetPairs:a}}c(me,"readAccumulateConfig");function fe(e){let r=e.accumulateEnabled&&e.pairCount<e.targetPairs;return{apply:e.consented&&!r,peeking:r}}c(fe,"resolveApply");async function Ae(e,r){let n=c(o=>r?.logger?.info?.(o)??console.log(o),"log"),{evalTargetId:s}=e,t=`[cron_instruction_evolution] ${s}`,a=await $(`/api/eval-targets/${s}`),p=a.metadata??{};if(X(p))return n(`${t}: PAUSED (metadata.autoEvolvePaused) \u2014 no model budget spent. Opt-in is intact.`),{skipped:"paused"};let y=be(p);if(y==null)return n(`${t}: refusing \u2014 metadata.immutableSections is unset. Set it (use [] to explicitly protect nothing) before letting the loop edit this target unattended.`),{skipped:"immutable-sections-unset"};let f=(await $(`/api/eval-targets/${s}/evals`)).filter(o=>!o.retiredAt);if(f.length===0)return{skipped:"no-live-cases"};let R=V(p);R.note&&n(`${t}: ${R.note}`);let d=ce(f.length,R.minPairs,ue(R,f.length));if(!d.feasible)return n(`${t}: cannot reach a sufficient verdict \u2014 ${d.reason}`),{skipped:"cannot-reach-min-n",reason:d.reason};let te=J({liveCases:f.length,passes:d.passes});n(`${t}: budget for this run \u2014 ${K(te)}`);let E=await $(`/api/eval-targets/${s}/judge-calibration`);if(!E.powerCheck.adequate){let o=(E.powerCheck.reasons??[]).join("; ");return n(`${t}: not adequately powered, skipping before spending budget \u2014 ${o}`),{skipped:"not-adequate",reason:o}}let M=W,ne=Date.now()-M*24*60*60*1e3,T=await $(`/api/eval-targets/${s}/promotions?limit=200`),O=T.promotions.filter(o=>o.applied&&Date.parse(o.createdAt)>=ne).length;if(O>=S)return n(`${t}: change budget looks exhausted (${O}/${S} applied in ${M}d), skipping before spending budget`),{skipped:"budget-exhausted"};let A=G();if(!A.green)return n(`${t}: A/A canary is RED \u2014 halting. ${A.failures.join(" | ")}`),{skipped:"canary-red",reason:A.failures.join(" | ")};let L=a.content,k=_(L),l=me(p,e),b=l.enabled?ge(T.promotions,k,l.targetPairs):null,v,x,w=[],C=0,N=B;if(b){let o=await $(`/api/promotions/${b.promotionId}`);if(_(o.toContent)!==b.candidateContentHash)return n(`${t}: stored candidate does not match its recorded hash \u2014 refusing to continue`),{skipped:"accumulation-hash-mismatch",reason:b.promotionId};v=o.toContent,x=o.proposalRationale??"",w=b.pairs,C=b.runs,N=0,n(`${t}: continuing accumulation ${b.promotionId} \u2014 ${w.length}/${l.targetPairs} pair(s) carried from ${C} prior run(s)`)}else{let o=E.caseDiscrimination?.cases??[],u=await Z({currentContent:a.content,targetType:a.type,targetName:a.name,immutableSections:y,evidence:{nonDiscriminating:o.filter(i=>!i.discriminating).map(i=>`${i.name??i.assertion??"unnamed"} \u2014 ${i.note}`),weakest:o.filter(i=>i.discriminating&&i.withPassRate<i.withoutPassRate).map(i=>({name:i.name??i.assertion??"unnamed",withPassRate:i.withPassRate,withoutPassRate:i.withoutPassRate})),notes:[]}});if(!u.ok)return n(`${t}: no usable proposal \u2014 ${u.reason}`),{skipped:"no-proposal",reason:u.reason};v=u.proposal.candidateContent,x=u.proposal.rationale}n(`${t}: measuring ${f.length} case(s) x ${d.passes} pass(es) = up to ${d.expectedPairs} pairs, candidate vs prior version`);let h=[],j=[],ae=Date.now(),U=0,I=0;for(let o=1;o<=d.passes;o++){let u=await Q({cases:f,candidateContent:v,priorVersionContent:L,concurrency:e.concurrency??2,runNumber:o,opts:{targetId:s,targetVersion:a.version??1,iteration:1,timeout:e.timeout??24e4,model:e.model,graderModel:e.graderModel,skipGrading:!1,targetType:a.type,targetName:a.name},onLog:n});h.push(...u.pairs),j.push(...u.results),U+=u.completed,I+=u.failed,n(`${t}: pass ${o}/${d.passes} \u2014 ${u.pairs.length} pair(s), running total ${h.length}`)}let P=[...w,...h],D=h.length>0,g=Y({runs:j,plannedPairs:d.expectedPairs,usablePairs:h.length,verdictPairs:D?P.length:h.length,proposalCalls:N,passes:d.passes,liveCases:f.length,wallMs:Date.now()-ae});if(n(`${t}: ${U} completed, ${I} failed, ${h.length} usable pair(s) \u2014 spent ~${g.modelCalls} model calls (${g.proposalCalls} proposal + ${g.executorCalls} executor + >=${g.graderCalls} grader) in ${(g.wallMs/6e4).toFixed(1)} min`+(D?`; the ${g.verdictPairs} pair(s) behind this verdict resolve +/-${z(g.mde)} at 80% power`:"")),!D)return{skipped:"no-pairs",reason:`${I} run(s) failed or were ungradeable`,cost:g};l.enabled&&n(`${t}: accumulated ${P.length}/${l.targetPairs} pair(s) over ${C+1} run(s) on one candidate`);let F=e.apply??(p.autoEvolveApply===!0||p.autoEvolveApply==="true"),{apply:oe,peeking:re}=fe({consented:F,accumulateEnabled:l.enabled,pairCount:P.length,targetPairs:l.targetPairs});F&&re&&n(`${t}: interim readout at ${P.length}/${l.targetPairs} pairs \u2014 recording the decision but not acting until the target is reached`);let m=await q(`/api/eval-targets/${s}/promotions`,{candidateContent:v,proposalSource:"llm",proposalRationale:x,pairs:P,canaryGreen:A.green,measuredAgainstContentHash:k,immutableSections:y,apply:oe,measurementCost:g,...l.enabled?{accumulation:{candidateContentHash:_(v),priorContentHash:k,targetPairs:l.targetPairs,carriedPairs:w.length,runs:C+1}}:{}});return n(`${t}: ${m.decision}${m.applied?" (APPLIED)":""} \u2014 ${m.rationale}`+(m.refusals.length>0?` | refusals: ${m.refusals.join("; ")}`:"")),{decision:m.decision,applied:m.applied,nPairs:P.length,promotionId:m.id,cost:g,...l.enabled?{carriedPairs:w.length,accumulationRuns:C+1}:{}}}c(Ae,"cronInstructionEvolution");function be(e){let r=e.immutableSections;return r==null||!Array.isArray(r)?null:r.filter(n=>typeof n=="string")}c(be,"readImmutableSections");export{ee as a,H as b,le as c,ue as d,ce as e,pe as f,ge as g,me as h,fe as i,Ae as j,be as k};

package/dist/{chunk-JEBHSV2M.js → chunk-ZYVWALMS.js}

@@ -1,4 +1,4 @@
-import{f}from"./chunk-MODFTCZY.js";import{b as S,e as a,h as y}from"./chunk-H2ORWIEY.js";import{a as p}from"./chunk-ZLLER3HP.js";var _=1800*1e3,G=300*1e3,T=10,h=`Given these messages from an AI coding session, write a 1-2 sentence summary of what the user is trying to accomplish. Focus on the goal, not the steps. No preamble.
+import{f}from"./chunk-B3BHB26P.js";import{b as S,e as a,h as y}from"./chunk-H2ORWIEY.js";import{a as p}from"./chunk-ZLLER3HP.js";var _=1800*1e3,G=300*1e3,T=10,h=`Given these messages from an AI coding session, write a 1-2 sentence summary of what the user is trying to accomplish. Focus on the goal, not the steps. No preamble.
 
 Return strict JSON in the form: {"goalSummary": "..."}`;async function k(n,e={}){let d=e.llm??f,o=e.now??Date.now,w=e.getSession??(async s=>{try{return await a(`/api/sessions/${s}`)}catch(t){if(t instanceof S&&t.statusCode===404)return null;throw t}}),m=e.getEvents??((s,t)=>a(`/api/sessions/${s}/events?limit=${t.limit}&sort=${t.sort}`)),v=e.saveGoal??(async(s,t)=>{await y(`/api/sessions/${s}/enrich`,{analysis:t})}),i=await w(n.sessionId);if(!i)return{skipped:"not_found"};let[P]=await m(n.sessionId,{limit:1,sort:"desc"}),A=new Date(P?.timestamp??i.startedAt).getTime();if(o()-A>_)return{skipped:"stale"};let l=i.analysis??{},c=l.goalSummaryAt?new Date(l.goalSummaryAt).getTime():0;if(c&&o()-c<G)return{skipped:"recent"};let g=(await m(n.sessionId,{limit:T*2,sort:"asc"})).filter(s=>s.content&&(s.eventType==="user_message"||s.eventType==="assistant_response")).map(s=>`${s.eventType==="user_message"?"USER":"ASSISTANT"}: ${s.content}`).join(`
 

package/dist/commands/capture/scope.js

@@ -0,0 +1,2 @@
+import{a as l,b as w,c as C,d as b,e as v,f as a,g as x}from"../../chunk-RUZ36O55.js";import{b as m,d as u}from"../../chunk-VX2GJCQ4.js";import{a as g}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import{l as h}from"../../chunk-O6C2UGBY.js";import{a as p}from"../../chunk-ZLLER3HP.js";import{existsSync as S}from"node:fs";import{Flags as s}from"@oclif/core";var c=class y extends g{static{p(this,"CaptureScope")}static description="Show or set which projects the worker sweeps for Codex transcripts";static examples=["<%= config.bin %> <%= command.id %>","<%= config.bin %> <%= command.id %> --add ~/work/client-a --add ~/work/client-b","<%= config.bin %> <%= command.id %> --remove ~/work/client-a","<%= config.bin %> <%= command.id %> --off","<%= config.bin %> <%= command.id %> --all"];static flags={add:s.string({description:"Capture Codex sessions under this path (repeatable). Implies an allowlist.",multiple:!0}),remove:s.string({description:"Stop capturing under this path (repeatable).",multiple:!0}),off:s.boolean({description:"Skip the Codex sweep entirely.",exclusive:["all"]}),all:s.boolean({description:"Capture every project (the default).",exclusive:["off"]}),clear:s.boolean({description:"Remove the codex scope, restoring the default."})};async run(){let{flags:e}=await this.parse(y),i=h(),r=l(i);if(!((e.add?.length??0)>0||(e.remove?.length??0)>0||e.off||e.all||e.clear)){await this.show(i,r);return}for(let t of[...e.add??[],...e.remove??[]])v(t)||this.error(`Not an absolute path: ${t}
+Use an absolute path or ~/\u2026 \u2014 relative paths cannot be matched against a session's cwd.`);for(let t of e.add??[]){let n=b(t);S(n)||this.warn(`${n} does not exist yet \u2014 added anyway, but nothing will match until it does.`)}let o;try{o=w(i)}catch(t){e.clear||this.error(t instanceof Error?t.message:String(t));let n=x(i);this.warn(`${l(i)} was not valid JSON \u2014 saved it to ${n} and starting fresh.`),o={}}let f=!1;if(e.clear&&({config:o}=a(o,{op:"clear"})),e.off&&({config:o}=a(o,{op:"mode",mode:"off"})),e.all&&({config:o}=a(o,{op:"mode",mode:"all"})),e.remove?.length){let t=a(o,{op:"remove",roots:e.remove});o=t.config,f=t.emptiedToOff}e.add?.length&&({config:o}=a(o,{op:"add",roots:e.add})),C(i,o),f&&(this.log("Removed the last root \u2014 Codex capture is now OFF rather than silently capturing everything."),this.log("Use --all to go back to capturing every project.")),await this.show(i,r),this.log(""),this.log("Takes effect on the next sweep (within ~30s). No restart needed.")}async show(e,i){let r=await m({configDir:e});if(this.log(`Codex capture: ${u(r)}`),r.roots.length)for(let d of r.roots)this.log(`  \u2022 ${d}`);this.log(`Config: ${i}`),r.error&&(this.log(""),this.log("This scope is FAILING CLOSED \u2014 nothing is being captured from Codex."),this.log("Fix the file, or run `rulemetric capture scope --clear` to start over.")),r.source==="env"&&(this.log(""),this.log("Set by environment variable, which outranks the config file."),this.log("Unset RULEMETRIC_CODEX_CAPTURE / RULEMETRIC_CODEX_ROOTS for the file to apply."))}};export{c as default};

package/dist/commands/diagnostics.js

@@ -1 +1 @@
-import{b as A,c as I}from"../chunk-V5U3E4BZ.js";import{a as L}from"../chunk-VLXCQ2DC.js";import"../chunk-I2DZGJRR.js";import{c as b,d as x,e as $}from"../chunk-MVM6OGAF.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import{a as E,b as H,d as X}from"../chunk-VX2GJCQ4.js";import{d as U}from"../chunk-OPUI7CM2.js";import"../chunk-MODFTCZY.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import{a as C}from"../chunk-PPKPCXJD.js";import"../chunk-IZTIYR5Q.js";import"../chunk-V44SRRYR.js";import"../chunk-U6XEU573.js";import"../chunk-VLOX6KOM.js";import"../chunk-3AV3GTMP.js";import{a as O,e as S}from"../chunk-H2ORWIEY.js";import{a as _,g,l as w}from"../chunk-O6C2UGBY.js";import{a as v}from"../chunk-ZLLER3HP.js";import{spawnSync as B}from"node:child_process";import{existsSync as D,readFileSync as G}from"node:fs";import{join as Y}from"node:path";var R=class u extends C{static{v(this,"Diagnostics")}static description='Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';static examples=["<%= config.bin %> diagnostics"];baseUrl(){return process.env.RULEMETRIC_API_URL??_().apiUrl??O}async checkDirect(){let o=process.env.HTTPS_PROXY,t=process.env.HTTP_PROXY;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let i=new AbortController,s=setTimeout(()=>i.abort(),8e3);try{return(await globalThis.fetch(`${this.baseUrl()}/health`,{signal:i.signal})).ok}catch{return!1}finally{clearTimeout(s),o!==void 0&&(process.env.HTTPS_PROXY=o),t!==void 0&&(process.env.HTTP_PROXY=t)}}configuredProxy(){if(process.env.HTTPS_PROXY)return process.env.HTTPS_PROXY;for(let o of["settings.local.json","settings.json"]){let t=Y(process.cwd(),".claude",o);if(D(t))try{let i=JSON.parse(G(t,"utf8"));if(i.env?.HTTPS_PROXY)return i.env.HTTPS_PROXY}catch{}}return null}checkThroughProxy(o){let t=B("curl",["-sS","-x",o,"--max-time","8","-o","/dev/null","-w","%{http_code}",`${this.baseUrl()}/health`],{encoding:"utf8",stdio:["ignore","pipe","ignore"]});return t.error||typeof t.stdout!="string"?null:t.stdout.trim()==="200"}async run(){await this.parse(u);let o=process.version,t=this.configuredProxy(),i=t!=null,[s,p,P]=await Promise.all([this.checkDirect(),$(b),$(x)]),f=i&&t?this.checkThroughProxy(t):null,n=v((e,a,d)=>{let h=a===null?"\u2013":a?"[OK]":"[!!]";this.log(`  ${h} ${e.padEnd(22)} ${d}`)},"line");this.log("RuleMetric diagnostics"),this.log(""),n("Node",!A(o),`${o}${A(o)?" (< 20.11 \u2014 breaks the bundled proxy)":""}`),n("API (direct)",s,`${this.baseUrl()}/health ${s?"reachable":"UNREACHABLE"}`),n("Authenticated",g(),g()?"yes":"no \u2014 run `rulemetric auth login`"),n("Proxy configured",i?!0:null,i?t:"no HTTPS_PROXY set"),i&&(n("Gateway :"+b,p,p?"listening":"NOT listening"),n("mitmproxy :"+x,P,P?"listening":"not listening (gateway should route direct)"),n("API (via proxy)",f,f===null?"inconclusive (curl unavailable)":f?"reachable":"FAILED \u2014 502/blocked through your proxy"));let l=U();n("claude CLI",l.claudeBin!=null,l.claudeBin??"not found \u2014 LLM features (grading, distill) will no-op"),n("codex CLI",l.codexBin!=null?!0:null,l.codexBin?`${l.codexBin}${l.codexConfiguredModel?` (model ${l.codexConfiguredModel})`:""}`:"not found \u2014 optional; enables the cross-vendor reviewer (evals validate-judge)");let r=null;if(g()&&s)try{r=await S("/api/workers/health")}catch{r=null}if(r){let e=r.worker;n("Worker",!!e.active,e.present?e.active?`online (last heartbeat ${e.ageSeconds??"?"}s ago${e.clientVersion?`, v${e.clientVersion}`:""})`:`NOT RUNNING \u2014 last heartbeat ${j(e.ageSeconds)} ago`:"never connected \u2014 no worker has ever reported in");let a=r.pending.reduce((d,h)=>d+h.count,0);if(a>0){let d=r.verdict.stuck.reduce((k,T)=>k+T.count,0),h=r.verdict.unclaimable.reduce((k,T)=>k+T.count,0);n("Job queue",d===0,`${a} pending`+(d>0?` \u2014 ${d} waiting on this worker`:"")+(h>0?` \u2014 ${h} need a worker with DATABASE_URL`:""))}else n("Job queue",!0,"empty")}else g()&&s&&n("Worker",null,"could not read /workers/health (older API?)");let y=L(process.cwd()),c=await H({configDir:w()}),M={claude_code:"session hooks (+ proxy for instruction linking)",codex:`worker transcript sync of ~/.codex/sessions \u2014 ${X(c)}`,cursor:"user hooks (~/.cursor/hooks.json)",copilot_agent:"workspace hooks",antigravity:"user hooks (~/.gemini/config/hooks.json)"};this.log(""),this.log("  Capture coverage (agents found on this machine):");for(let e of y)this.log(`    \u2022 ${e.padEnd(15)} ${M[e]??"captured"}`);if(y.includes("codex"))if(c.error)this.log(`    ! ${c.error}`);else if(c.mode==="allowlist")for(let e of c.roots)this.log(`        \u21B3 ${e}`);else c.mode==="all"&&this.log(`        \u21B3 scope it with ${Y(w(),E)} ({"codex":{"roots":["/path/to/repo"]}}) or RULEMETRIC_CODEX_CAPTURE=off`);r&&!r.worker.active&&y.includes("codex")&&c.mode!=="off"&&this.log("    ! Codex capture runs INSIDE the worker \u2014 it is stopped while the worker is down.");let m=I({nodeVersion:o,directApiOk:s,httpsProxySet:i,gatewayAlive:p,mitmproxyAlive:P,proxyPathOk:f});this.log("");let N=m.severity==="ok"?"[OK]":m.severity==="warn"?"[!!]":"[XX]";this.log(`${N} ${m.headline}`);for(let e of m.hints)this.log(`   \u2022 ${e}`);if(r&&r.verdict.severity!=="ok"){let e=r.verdict.severity==="warn"?"[!!]":"[XX]";this.log(""),this.log(`${e} ${r.verdict.headline}`);for(let a of r.verdict.hints)this.log(`   \u2022 ${a}`)}}};function j(u){if(u===null)return"never";if(u<90)return`${Math.round(u)}s`;let o=Math.round(u/60);return o<90?`${o}m`:`${Math.round(o/60)}h`}v(j,"formatAge");export{R as default};
+import{b as A,c as U}from"../chunk-V5U3E4BZ.js";import{a as L}from"../chunk-VLXCQ2DC.js";import"../chunk-I2DZGJRR.js";import{e as T,f as E,g as x}from"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import{d as F,e as G}from"../chunk-ZYI66XKV.js";import"../chunk-RMDLJN4L.js";import{i as M,j as N,k as D,m as Y,n as B}from"../chunk-BL4BN6RU.js";import"../chunk-2ZAKZEXW.js";import{d as X}from"../chunk-NTKO4H2K.js";import"../chunk-KQCAELPN.js";import"../chunk-L2CFSMBC.js";import"../chunk-Z72VHGV7.js";import"../chunk-4L2ZEIRE.js";import"../chunk-CDY7TQFZ.js";import"../chunk-XRTZU7IZ.js";import"../chunk-B3BHB26P.js";import"../chunk-N7ODX6PO.js";import"../chunk-PJ3WHQ67.js";import"../chunk-G5BXZT6F.js";import"../chunk-HMMO27U4.js";import{a as O,b as I,d as H}from"../chunk-VX2GJCQ4.js";import{a as _}from"../chunk-PPKPCXJD.js";import"../chunk-IZTIYR5Q.js";import"../chunk-V44SRRYR.js";import"../chunk-U6XEU573.js";import"../chunk-VLOX6KOM.js";import"../chunk-3AV3GTMP.js";import{a as C,e as P}from"../chunk-H2ORWIEY.js";import{a as R,g as v,l as k}from"../chunk-O6C2UGBY.js";import{a as $}from"../chunk-ZLLER3HP.js";import{spawnSync as J}from"node:child_process";import{existsSync as q,readFileSync as K}from"node:fs";import{join as j}from"node:path";var S=class h extends _{static{$(this,"Diagnostics")}static description='Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';static examples=["<%= config.bin %> diagnostics"];baseUrl(){return process.env.RULEMETRIC_API_URL??R().apiUrl??C}async checkDirect(){let o=process.env.HTTPS_PROXY,i=process.env.HTTP_PROXY;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let t=new AbortController,a=setTimeout(()=>t.abort(),8e3);try{return(await globalThis.fetch(`${this.baseUrl()}/health`,{signal:t.signal})).ok}catch{return!1}finally{clearTimeout(a),o!==void 0&&(process.env.HTTPS_PROXY=o),i!==void 0&&(process.env.HTTP_PROXY=i)}}configuredProxy(){if(process.env.HTTPS_PROXY)return process.env.HTTPS_PROXY;for(let o of["settings.local.json","settings.json"]){let i=j(process.cwd(),".claude",o);if(q(i))try{let t=JSON.parse(K(i,"utf8"));if(t.env?.HTTPS_PROXY)return t.env.HTTPS_PROXY}catch{}}return null}checkThroughProxy(o){let i=J("curl",["-sS","-x",o,"--max-time","8","-o","/dev/null","-w","%{http_code}",`${this.baseUrl()}/health`],{encoding:"utf8",stdio:["ignore","pipe","ignore"]});return i.error||typeof i.stdout!="string"?null:i.stdout.trim()==="200"}async run(){await this.parse(h);let o=process.version,i=this.configuredProxy(),t=i!=null,[a,m,c]=await Promise.all([this.checkDirect(),x(T),x(E)]),u=t&&i?this.checkThroughProxy(i):null,e=$((n,g,p)=>{let f=g===null?"\u2013":g?"[OK]":"[!!]";this.log(`  ${f} ${n.padEnd(22)} ${p}`)},"line");this.log("RuleMetric diagnostics"),this.log(""),e("Node",!A(o),`${o}${A(o)?" (< 20.11 \u2014 breaks the bundled proxy)":""}`),e("API (direct)",a,`${this.baseUrl()}/health ${a?"reachable":"UNREACHABLE"}`),e("Authenticated",v(),v()?"yes":"no \u2014 run `rulemetric auth login`"),e("Proxy configured",t?!0:null,t?i:"no HTTPS_PROXY set"),t&&(e("Gateway :"+T,m,m?"listening":"NOT listening"),e("mitmproxy :"+E,c,c?"listening":"not listening (gateway should route direct)"),e("API (via proxy)",u,u===null?"inconclusive (curl unavailable)":u?"reachable":"FAILED \u2014 502/blocked through your proxy"));let s=X();e("claude CLI",s.claudeBin!=null,s.claudeBin??"not found \u2014 LLM features (grading, distill) will no-op"),e("codex CLI",s.codexBin!=null?!0:null,s.codexBin?`${s.codexBin}${s.codexConfiguredModel?` (model ${s.codexConfiguredModel})`:""}`:"not found \u2014 optional; enables the cross-vendor reviewer (evals validate-judge)");let r=null;if(v()&&a)try{r=await P("/api/workers/health")}catch{r=null}if(r){let n=r.worker;e("Worker",!!n.active,n.present?n.active?`online (last heartbeat ${n.ageSeconds??"?"}s ago${n.clientVersion?`, v${n.clientVersion}`:""})`:`NOT RUNNING \u2014 last heartbeat ${z(n.ageSeconds)} ago`:"never connected \u2014 no worker has ever reported in");let g=r.pending.reduce((p,f)=>p+f.count,0);if(g>0){let p=r.verdict.stuck.reduce((b,w)=>b+w.count,0),f=r.verdict.unclaimable.reduce((b,w)=>b+w.count,0);e("Job queue",p===0,`${g} pending`+(p>0?` \u2014 ${p} waiting on this worker`:"")+(f>0?` \u2014 ${f} need a worker with DATABASE_URL`:""))}else e("Job queue",!0,"empty")}else v()&&a&&e("Worker",null,"could not read /workers/health (older API?)");let l=L(process.cwd()),d=await I({configDir:k()}),W={claude_code:"session hooks (+ proxy for instruction linking)",codex:`worker transcript sync of ~/.codex/sessions \u2014 ${H(d)}`,cursor:"user hooks (~/.cursor/hooks.json)",copilot_agent:"workspace hooks",antigravity:"user hooks (~/.gemini/config/hooks.json)"};this.log(""),this.log("  Capture coverage (agents found on this machine):");for(let n of l)this.log(`    \u2022 ${n.padEnd(15)} ${W[n]??"captured"}`);if(l.includes("codex"))if(d.error)this.log(`    ! ${d.error}`);else if(d.mode==="allowlist")for(let n of d.roots)this.log(`        \u21B3 ${n}`);else d.mode==="all"&&this.log(`        \u21B3 scope it with ${j(k(),O)} ({"codex":{"roots":["/path/to/repo"]}}) or RULEMETRIC_CODEX_CAPTURE=off`);r&&!r.worker.active&&l.includes("codex")&&d.mode!=="off"&&this.log("    ! Codex capture runs INSIDE the worker \u2014 it is stopped while the worker is down."),v()&&a&&await this.reportEvolutionSpend();let y=U({nodeVersion:o,directApiOk:a,httpsProxySet:t,gatewayAlive:m,mitmproxyAlive:c,proxyPathOk:u});this.log("");let V=y.severity==="ok"?"[OK]":y.severity==="warn"?"[!!]":"[XX]";this.log(`${V} ${y.headline}`);for(let n of y.hints)this.log(`   \u2022 ${n}`);if(r&&r.verdict.severity!=="ok"){let n=r.verdict.severity==="warn"?"[!!]":"[XX]";this.log(""),this.log(`${n} ${r.verdict.headline}`);for(let g of r.verdict.hints)this.log(`   \u2022 ${g}`)}}async reportEvolutionSpend(){let o;try{o=await P("/api/eval-targets?type=instruction&limit=200")}catch{return}let i=o.filter(t=>{let a=t.metadata??{};return a.autoEvolveEnabled===!0||a.autoEvolveEnabled==="true"});if(i.length!==0){this.log(""),this.log("  Nightly instruction evolution (cron_instruction_evolution, 07:30 UTC):");for(let t of i){let a=t.metadata??{},m=N(a),c=M(a);if(m){this.log(`    \u2022 ${t.name} \u2014 PAUSED (metadata.autoEvolvePaused). Opt-in preserved.`),this.log("        \u21B3 resume: set autoEvolvePaused=false");continue}let u=null;try{u=(await P(`/api/eval-targets/${t.id}/evals`)).filter(s=>!s.retiredAt).length}catch{}if(this.log(`    \u2022 ${t.name}`),u!=null){let e=G(u,c.minPairs,F(c,u));if(!e.feasible)this.log(`        \u21B3 cannot reach a verdict \u2014 ${e.reason}`);else{let s=D({liveCases:u,passes:e.passes});this.log(`        \u21B3 each night: ${B(s)}`)}}this.log(`        \u21B3 pairs budget ${c.minPairs}`+(c.configured?" (metadata.autoEvolvePairs)":" (default)")+` \xB7 writes ${a.autoEvolveApply===!0||a.autoEvolveApply==="true"?"ENABLED":"off (dry run)"}`),c.note&&this.log(`        ! ${c.note}`);try{let e=await P(`/api/eval-targets/${t.id}/promotions?limit=1`),s=e.promotions[0];if(s){let l=s.gates?.cost,d=l?.modelCalls?`${l.modelCalls} model calls, ${l.usablePairs??"?"} usable pair(s)`+(l.mde!=null?`, resolved +/-${Y(l.mde)}`:"")+(l.wallMs?`, ${(l.wallMs/6e4).toFixed(0)} min`:""):"cost not recorded (decision predates cost tracking)";this.log(`        \u21B3 last run ${new Date(s.createdAt).toISOString().slice(0,10)}: ${s.decision}${s.applied?" (APPLIED)":""} \u2014 ${d}`)}let r=e.counts??{};this.log(`        \u21B3 lifetime: ${r.promote??0} promote / ${r.revert??0} revert / ${r.refuse??0} refuse`)}catch{}}this.log("        \u21B3 pause the spend: metadata.autoEvolvePaused=true (keeps the opt-in)")}}};function z(h){if(h===null)return"never";if(h<90)return`${Math.round(h)}s`;let o=Math.round(h/60);return o<90?`${o}m`:`${Math.round(o/60)}h`}$(z,"formatAge");export{S as default};

package/dist/commands/evals/agent.js

@@ -1,3 +1,3 @@
-import{a as C,b as L,c as D}from"../../chunk-3HSEVXMQ.js";import{e as I}from"../../chunk-VXIRJ7YY.js";import"../../chunk-TIOKY2IA.js";import"../../chunk-LEXVC7VH.js";import"../../chunk-W2MOTOQE.js";import"../../chunk-3FJT6LL4.js";import"../../chunk-DAGDGAPU.js";import"../../chunk-6QKI756T.js";import"../../chunk-RSKQRVPJ.js";import"../../chunk-JEBHSV2M.js";import"../../chunk-UF6US7VU.js";import"../../chunk-L5ORQB25.js";import"../../chunk-E3MQVVH6.js";import"../../chunk-WEF5LPGS.js";import"../../chunk-N7TGRFVZ.js";import{j as d}from"../../chunk-3DLVY7LD.js";import"../../chunk-ZAWAPLIE.js";import"../../chunk-AIEM6VXO.js";import"../../chunk-BZUIMHHT.js";import"../../chunk-QFVGUZDO.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-5W55CRBJ.js";import"../../chunk-JR4CWY3J.js";import"../../chunk-ZERPMWRT.js";import"../../chunk-UR6GZGFK.js";import"../../chunk-ZWKKFAIX.js";import"../../chunk-YQ56FQEO.js";import"../../chunk-JGYCLDYF.js";import"../../chunk-TDD7IOV2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-7DFNBRLY.js";import"../../chunk-4VZLIGMS.js";import"../../chunk-3VYV6ZYB.js";import"../../chunk-OZQV5CDT.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-CJRYYV76.js";import"../../chunk-BSJBNOJT.js";import"../../chunk-L2CFSMBC.js";import"../../chunk-Z72VHGV7.js";import"../../chunk-7AZKPX2J.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-V5MJ4ICC.js";import"../../chunk-WKXTY5AM.js";import"../../chunk-HMMO27U4.js";import{b as k,d as h,e as A}from"../../chunk-VX2GJCQ4.js";import"../../chunk-OPUI7CM2.js";import"../../chunk-MODFTCZY.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as O}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{f as b}from"../../chunk-H2ORWIEY.js";import{e as P,l as f}from"../../chunk-O6C2UGBY.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as se,context as ie}from"@opentelemetry/api";import{NodeSDK as j}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as q}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as F,ConsoleSpanExporter as U,SimpleSpanProcessor as X}from"@opentelemetry/sdk-trace-node";import{Resource as B}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as H,ATTR_SERVICE_VERSION as K}from"@opentelemetry/semantic-conventions";function $(t){if(!t)return{};let r={};for(let n of t.split(",")){let s=n.indexOf("=");if(s===-1)continue;let i=n.slice(0,s).trim(),a=n.slice(s+1).trim();i&&(r[i]=a)}return r}e($,"parseOtlpHeaders");var p=null;function y(){return process.env.TELEMETRY_ENABLED!=="false"}e(y,"isTelemetryEnabled");function S(t){if(!y()||p!==null)return"off";let r=new B({[H]:t.serviceName,[K]:t.serviceVersion??"0.1.0"}),n=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(n){let a=$(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new q({url:`${n}/v1/traces`,headers:a});i=new F(l),s="otlp"}else if(process.env.OTEL_SPAN_EXPORT==="none")i=void 0,s="silent";else{let a=new U;i=new X(a),s="console"}return p=new j({resource:r,spanProcessors:i?[i]:[]}),p.start(),s}e(S,"initTelemetry");async function m(){p&&(await p.shutdown(),p=null)}e(m,"shutdownTelemetry");import{SpanKind as he,SpanStatusCode as ye}from"@opentelemetry/api";import{trace as ve,context as we}from"@opentelemetry/api";import*as u from"@sentry/node";var E=!1;function T(t){if(E)return!1;let r=process.env.SENTRY_DSN;return r?(u.init({dsn:r,serverName:t.serviceName,environment:process.env.NODE_ENV??"development",tracesSampleRate:0}),E=!0,!0):!1}e(T,"initSentry");async function g(t){E&&await u.flush(t)}e(g,"flushSentry");import{existsSync as J,mkdirSync as W,readFileSync as R,unlinkSync as V,writeFileSync as M}from"node:fs";import{dirname as z,join as v}from"node:path";import{execFileSync as Q}from"node:child_process";var c=v(f(),"evals-agent.lock");function Z(t){try{return process.kill(t,0),!0}catch(r){return r.code==="EPERM"}}e(Z,"pidAlive");function ee(t){if(!Z(t))return!1;try{let r=Q("ps",["-p",String(t),"-o","command="],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return/run\.js\s+evals\s+agent|evals\s+agent/.test(r)}catch{return!0}}e(ee,"isLiveAgent");function te(){try{W(z(c),{recursive:!0})}catch{}for(let t=0;t<2;t++)try{return M(c,String(process.pid),{flag:"wx"}),{ok:!0}}catch(r){if(r.code!=="EEXIST")throw r;let n=0;try{n=Number(R(c,"utf-8").trim())}catch{}if(n&&n!==process.pid&&ee(n))return{ok:!1,holder:n};try{V(c)}catch{}}try{M(c,String(process.pid))}catch{}return{ok:!0}}e(te,"acquireAgentLock");function x(){try{J(c)&&Number(R(c,"utf-8").trim())===process.pid&&V(c)}catch{}}e(x,"releaseAgentLock");var w=class t extends O{static{e(this,"EvalAgentCommand")}static description="The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.";static examples=["<%= config.bin %> evals agent"];async run(){await this.parse(t),this.requireAuth();let r=te();if(!r.ok){this.log(`Another rulemetric evals agent is already running (pid ${r.holder}). Refusing to start a second instance.`);return}process.env.PG_APPLICATION_NAME||="rulemetric-agent",T({serviceName:"rulemetric-agent"})&&this.log("[sentry] error capture enabled");let s=S({serviceName:"rulemetric-agent"});s==="otlp"?this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`):s==="console"?this.log("[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)"):this.log("[otel] disabled (TELEMETRY_ENABLED=false)"),!P()&&!process.env.RULEMETRIC_API_KEY&&this.error("No auth token found. Run `rulemetric auth login` first."),this.log("Starting thin agent (Phase 4b parallel-run)..."),this.log(`Long-polling /api/work/next. Press Ctrl+C to stop.
+import{a as C,b as L,c as D}from"../../chunk-3HSEVXMQ.js";import{e as I}from"../../chunk-6NML7IVJ.js";import"../../chunk-TIOKY2IA.js";import"../../chunk-ZYVWALMS.js";import"../../chunk-NKPLTT7J.js";import"../../chunk-HUADECKZ.js";import"../../chunk-TVHT6H42.js";import"../../chunk-QFVGUZDO.js";import"../../chunk-H7J27U4P.js";import"../../chunk-BUWWFHGO.js";import"../../chunk-IQTSUAGM.js";import"../../chunk-CRIFP3QP.js";import"../../chunk-MDPN46EY.js";import"../../chunk-RSKQRVPJ.js";import"../../chunk-SWSGNEVO.js";import"../../chunk-BPQ66YTY.js";import"../../chunk-UF6US7VU.js";import"../../chunk-DKGKOSW3.js";import"../../chunk-IXLDCX7M.js";import"../../chunk-JIJKIPE6.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-N7TGRFVZ.js";import{j as d}from"../../chunk-JGLVVIEM.js";import"../../chunk-Q4APAZUC.js";import"../../chunk-YQ56FQEO.js";import"../../chunk-TDD7IOV2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-5DIOZDDZ.js";import"../../chunk-4VZLIGMS.js";import"../../chunk-CNOD6HWU.js";import"../../chunk-X64R6JCJ.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-ZYI66XKV.js";import"../../chunk-RMDLJN4L.js";import"../../chunk-BL4BN6RU.js";import"../../chunk-2ZAKZEXW.js";import"../../chunk-NTKO4H2K.js";import"../../chunk-CJRYYV76.js";import"../../chunk-KQCAELPN.js";import"../../chunk-L2CFSMBC.js";import"../../chunk-Z72VHGV7.js";import"../../chunk-4L2ZEIRE.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-XRTZU7IZ.js";import"../../chunk-B3BHB26P.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-G5BXZT6F.js";import"../../chunk-HMMO27U4.js";import{b as k,d as h,e as A}from"../../chunk-VX2GJCQ4.js";import{a as O}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{f as b}from"../../chunk-H2ORWIEY.js";import{e as P,l as f}from"../../chunk-O6C2UGBY.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as se,context as ie}from"@opentelemetry/api";import{NodeSDK as j}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as q}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as F,ConsoleSpanExporter as U,SimpleSpanProcessor as X}from"@opentelemetry/sdk-trace-node";import{Resource as B}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as H,ATTR_SERVICE_VERSION as K}from"@opentelemetry/semantic-conventions";function $(t){if(!t)return{};let r={};for(let n of t.split(",")){let s=n.indexOf("=");if(s===-1)continue;let i=n.slice(0,s).trim(),a=n.slice(s+1).trim();i&&(r[i]=a)}return r}e($,"parseOtlpHeaders");var p=null;function y(){return process.env.TELEMETRY_ENABLED!=="false"}e(y,"isTelemetryEnabled");function S(t){if(!y()||p!==null)return"off";let r=new B({[H]:t.serviceName,[K]:t.serviceVersion??"0.1.0"}),n=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(n){let a=$(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new q({url:`${n}/v1/traces`,headers:a});i=new F(l),s="otlp"}else if(process.env.OTEL_SPAN_EXPORT==="none")i=void 0,s="silent";else{let a=new U;i=new X(a),s="console"}return p=new j({resource:r,spanProcessors:i?[i]:[]}),p.start(),s}e(S,"initTelemetry");async function m(){p&&(await p.shutdown(),p=null)}e(m,"shutdownTelemetry");import{SpanKind as he,SpanStatusCode as ye}from"@opentelemetry/api";import{trace as ve,context as we}from"@opentelemetry/api";import*as u from"@sentry/node";var E=!1;function T(t){if(E)return!1;let r=process.env.SENTRY_DSN;return r?(u.init({dsn:r,serverName:t.serviceName,environment:process.env.NODE_ENV??"development",tracesSampleRate:0}),E=!0,!0):!1}e(T,"initSentry");async function g(t){E&&await u.flush(t)}e(g,"flushSentry");import{existsSync as J,mkdirSync as W,readFileSync as R,unlinkSync as V,writeFileSync as M}from"node:fs";import{dirname as z,join as v}from"node:path";import{execFileSync as Q}from"node:child_process";var c=v(f(),"evals-agent.lock");function Z(t){try{return process.kill(t,0),!0}catch(r){return r.code==="EPERM"}}e(Z,"pidAlive");function ee(t){if(!Z(t))return!1;try{let r=Q("ps",["-p",String(t),"-o","command="],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return/run\.js\s+evals\s+agent|evals\s+agent/.test(r)}catch{return!0}}e(ee,"isLiveAgent");function te(){try{W(z(c),{recursive:!0})}catch{}for(let t=0;t<2;t++)try{return M(c,String(process.pid),{flag:"wx"}),{ok:!0}}catch(r){if(r.code!=="EEXIST")throw r;let n=0;try{n=Number(R(c,"utf-8").trim())}catch{}if(n&&n!==process.pid&&ee(n))return{ok:!1,holder:n};try{V(c)}catch{}}try{M(c,String(process.pid))}catch{}return{ok:!0}}e(te,"acquireAgentLock");function x(){try{J(c)&&Number(R(c,"utf-8").trim())===process.pid&&V(c)}catch{}}e(x,"releaseAgentLock");var w=class t extends O{static{e(this,"EvalAgentCommand")}static description="The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.";static examples=["<%= config.bin %> evals agent"];async run(){await this.parse(t),this.requireAuth();let r=te();if(!r.ok){this.log(`Another rulemetric evals agent is already running (pid ${r.holder}). Refusing to start a second instance.`);return}process.env.PG_APPLICATION_NAME||="rulemetric-agent",T({serviceName:"rulemetric-agent"})&&this.log("[sentry] error capture enabled");let s=S({serviceName:"rulemetric-agent"});s==="otlp"?this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`):s==="console"?this.log("[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)"):this.log("[otel] disabled (TELEMETRY_ENABLED=false)"),!P()&&!process.env.RULEMETRIC_API_KEY&&this.error("No auth token found. Run `rulemetric auth login` first."),this.log("Starting thin agent (Phase 4b parallel-run)..."),this.log(`Long-polling /api/work/next. Press Ctrl+C to stop.
 `);let a=new AbortController,l=await k({configDir:f()});if(this.log(`[transcript-sync] Codex capture: ${h(l)}`),l.error)this.log(`[transcript-sync] WARNING: ${l.error}`);else if(l.mode==="allowlist")for(let o of l.roots)this.log(`[transcript-sync]   \u2022 ${o}`);let G=new C({providers:[L(void 0,void 0,A({configDir:f(),onChange:e(o=>this.log(`[transcript-sync] Codex capture changed: ${h(o)}`),"onChange")}))],statePath:v(f(),"transcript-sync-state.json"),importTranscript:e(o=>b("/api/sessions/import",o),"importTranscript"),log:e(o=>this.log(`[transcript-sync] ${o}`),"log")});D(G,{signal:a.signal,log:e(o=>this.log(`[transcript-sync] ${o}`),"log")});let N=e(async o=>{this.log(`
 [${o}] shutting down agent...`),a.abort(),await Promise.allSettled([g(2e3),m(),d()]),x(),process.exit(0)},"shutdown");process.on("SIGINT",()=>{N("SIGINT")}),process.on("SIGTERM",()=>{N("SIGTERM")});let _=!1;try{await I({signal:a.signal,log:e(o=>this.log(o),"log"),version:this.config.version,readInstalledVersion:e(()=>{try{return JSON.parse(R(v(this.config.root,"package.json"),"utf-8")).version??null}catch{return null}},"readInstalledVersion"),onVersionDrift:e(()=>{_=!0},"onVersionDrift")}),_&&(this.log("[upgrade] Installed CLI version changed \u2014 restarting worker onto the new code."),await Promise.allSettled([g(2e3),m(),d()]),x(),process.exit(1))}finally{x(),await d()}}};export{w as default};

package/dist/commands/evals/auto-run.js

@@ -1,4 +1,4 @@
-import{a as k}from"../../chunk-CJRYYV76.js";import{b as $}from"../../chunk-BSJBNOJT.js";import"../../chunk-L2CFSMBC.js";import"../../chunk-Z72VHGV7.js";import"../../chunk-7AZKPX2J.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-V5MJ4ICC.js";import"../../chunk-HMMO27U4.js";import"../../chunk-MODFTCZY.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as w}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as A,Flags as r}from"@oclif/core";var m=class y extends w{static{h(this,"EvalAutoRun")}static description="Run evals in a loop until the effect verdict is statistically powered (\xA76f auto-run-to-power)";static examples=["<%= config.bin %> evals auto-run <target-id>            # preview (cost + plan)","<%= config.bin %> evals auto-run <target-id> --yes      # run until powered / budget","<%= config.bin %> evals auto-run <target-id> --yes --max-runs 120 --batch-parallel 4"];static args={targetId:A.string({description:"Eval target ID",required:!0})};static flags={yes:r.boolean({char:"y",description:"Actually run. Without this, only previews the plan + cost."}),"max-runs":r.integer({description:"Hard ceiling on completed runs",default:200}),"batch-parallel":r.integer({description:"Concurrent executions within a batch",default:3}),timeout:r.integer({description:"Timeout per eval in ms",default:12e4}),model:r.string({description:"Model for eval execution"}),"grader-model":r.string({description:"Model for grading"}),"skip-baseline":r.boolean({description:"Skip without_target runs (no delta \u2014 not recommended)",default:!1})};async run(){let{args:C,flags:e}=await this.parse(y);this.requireAuth();let i=C.targetId,c=20,a=await n(`/api/eval-targets/${i}`),l=await n(`/api/eval-targets/${i}/evals`);l.length===0&&this.error(`No active cases on "${a.name}". Generate some (\`evals create\` / the web Generate button) or run \`evals refine-cases\` first.`);let b=l.length*(e["skip-baseline"]?1:2),f=Math.ceil(e["max-runs"]/b);if(this.log(`Auto-run "${a.name}" v${a.version} toward a powered \xB1${c}pp verdict.`),this.log(`Plan: up to ${e["max-runs"]} runs (${l.length} active case(s) \xD7 ${e["skip-baseline"]?1:2}/batch = ${b}/batch, \u2264${f} batches). Each run is a synthetic \`claude -p\` execution + grade.`),this.log('Note: this is a controlled synthetic A/B. For a DEPLOYED instruction with real usage, the Insights observational verdict is cheaper and uses real sessions \u2014 auto-run is for pre-deploy, catalog skills, or targets with no real "without" cohort.'),!e.yes){this.log(`
+import{a as k}from"../../chunk-CJRYYV76.js";import{b as $}from"../../chunk-KQCAELPN.js";import"../../chunk-L2CFSMBC.js";import"../../chunk-Z72VHGV7.js";import"../../chunk-4L2ZEIRE.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-XRTZU7IZ.js";import"../../chunk-B3BHB26P.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-HMMO27U4.js";import{a as w}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as A,Flags as r}from"@oclif/core";var m=class y extends w{static{h(this,"EvalAutoRun")}static description="Run evals in a loop until the effect verdict is statistically powered (\xA76f auto-run-to-power)";static examples=["<%= config.bin %> evals auto-run <target-id>            # preview (cost + plan)","<%= config.bin %> evals auto-run <target-id> --yes      # run until powered / budget","<%= config.bin %> evals auto-run <target-id> --yes --max-runs 120 --batch-parallel 4"];static args={targetId:A.string({description:"Eval target ID",required:!0})};static flags={yes:r.boolean({char:"y",description:"Actually run. Without this, only previews the plan + cost."}),"max-runs":r.integer({description:"Hard ceiling on completed runs",default:200}),"batch-parallel":r.integer({description:"Concurrent executions within a batch",default:3}),timeout:r.integer({description:"Timeout per eval in ms",default:12e4}),model:r.string({description:"Model for eval execution"}),"grader-model":r.string({description:"Model for grading"}),"skip-baseline":r.boolean({description:"Skip without_target runs (no delta \u2014 not recommended)",default:!1})};async run(){let{args:C,flags:e}=await this.parse(y);this.requireAuth();let i=C.targetId,c=20,a=await n(`/api/eval-targets/${i}`),l=await n(`/api/eval-targets/${i}/evals`);l.length===0&&this.error(`No active cases on "${a.name}". Generate some (\`evals create\` / the web Generate button) or run \`evals refine-cases\` first.`);let b=l.length*(e["skip-baseline"]?1:2),f=Math.ceil(e["max-runs"]/b);if(this.log(`Auto-run "${a.name}" v${a.version} toward a powered \xB1${c}pp verdict.`),this.log(`Plan: up to ${e["max-runs"]} runs (${l.length} active case(s) \xD7 ${e["skip-baseline"]?1:2}/batch = ${b}/batch, \u2264${f} batches). Each run is a synthetic \`claude -p\` execution + grade.`),this.log('Note: this is a controlled synthetic A/B. For a DEPLOYED instruction with real usage, the Insights observational verdict is cheaper and uses real sessions \u2014 auto-run is for pre-deploy, catalog skills, or targets with no real "without" cohort.'),!e.yes){this.log(`
 Preview only. Re-run with --yes to start.`);return}let t={runsCompleted:0,maxRuns:e["max-runs"],batchesRun:0,maxBatches:f},x={targetId:i,targetVersion:a.version,iteration:1,timeout:e.timeout,model:e.model,graderModel:e["grader-model"],skipGrading:!1,targetType:a.type,targetName:a.name};for(;;){let s;try{s=await n(`/api/eval-targets/${i}/judge-calibration`)}catch(o){this.log(`
 \u25A0 stopping \u2014 could not read the power check: ${o instanceof Error?o.message:String(o)}`);break}let E={powered:s.powerCheck?.powered??!1,mde:s.powerCheck?.mde==null?1/0:s.powerCheck.mde},B=s.caseDiscrimination??{total:0,discriminating:0,nonDiscriminating:0,unpaired:0},p=k(E,B,t,{targetEffectPp:c});if(this.log(`
 ${p.action==="stop"?"\u25A0":"\u25B8"} ${p.reason}`),p.action==="stop")break;let v=await n(`/api/eval-targets/${i}/evals`);if(v.length===0){this.log("\u25A0 no active cases left \u2014 stopping.");break}let{completed:g,failed:D}=await $({cases:v,targetContent:a.content,concurrency:e["batch-parallel"],runNumber:t.batchesRun+1,opts:x,skipBaseline:e["skip-baseline"],onLog:h(o=>this.log(o),"onLog")});if(t.runsCompleted+=g,t.batchesRun+=1,this.log(`Batch ${t.batchesRun}: ${g} completed, ${D} failed \u2014 ${t.runsCompleted}/${t.maxRuns} runs, ${t.batchesRun}/${t.maxBatches} batches.`),g===0){this.log("\u25A0 batch produced 0 completed runs \u2014 the executor appears broken (check the model / `claude -p`). Aborting.");break}}let d=await n(`/api/eval-targets/${i}/judge-calibration`),u=d.powerCheck?.mde,R=u==null||!Number.isFinite(u)?"unknown":`\xB1${Math.round(u*100)}pp`;this.log(`

package/dist/commands/evals/benchmark.js

@@ -1,4 +1,4 @@
-import{b as h,c as f}from"../../chunk-LETTPEG4.js";import"../../chunk-WKXTY5AM.js";import{a as m}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as c}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as v}from"@oclif/core";var d=class b extends m{static{u(this,"EvalBenchmarkCommand")}static description="Aggregate eval runs into a benchmark for a target";static examples=["<%= config.bin %> evals benchmark <target-id>","<%= config.bin %> evals benchmark <target-id> --iteration 2"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={iteration:v.integer({description:"Filter runs by iteration",default:1})};async run(){let{args:o,flags:l}=await this.parse(b);this.requireAuth();let g=await c(`/api/eval-targets/${o.targetId}`);this.log(`Aggregating benchmark for "${g.name}" v${g.version}, iteration ${l.iteration}`);let a=await h(o.targetId,l.iteration);a.totalRuns===0&&this.error("No completed runs found. Run evals first with `rulemetric evals run`."),this.log(`
+import{b as h,c as f}from"../../chunk-VSUL3MNG.js";import"../../chunk-G5BXZT6F.js";import{a as m}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as c}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as v}from"@oclif/core";var d=class b extends m{static{u(this,"EvalBenchmarkCommand")}static description="Aggregate eval runs into a benchmark for a target";static examples=["<%= config.bin %> evals benchmark <target-id>","<%= config.bin %> evals benchmark <target-id> --iteration 2"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={iteration:v.integer({description:"Filter runs by iteration",default:1})};async run(){let{args:o,flags:l}=await this.parse(b);this.requireAuth();let g=await c(`/api/eval-targets/${o.targetId}`);this.log(`Aggregating benchmark for "${g.name}" v${g.version}, iteration ${l.iteration}`);let a=await h(o.targetId,l.iteration);a.totalRuns===0&&this.error("No completed runs found. Run evals first with `rulemetric evals run`."),this.log(`
 Aggregated ${a.totalRuns} runs
 `);for(let[e,n]of Object.entries(a.runSummary)){this.log(`${e}:`);for(let[r,i]of Object.entries(n)){let s=i;this.log(`  ${r}: mean=${s.mean} stddev=${s.stddev} [${s.min}, ${s.max}]`)}}let t=a.verdict;if(t)if(this.log(`
 Verdict (${t.mode}, baseline=${t.baseline}):`),!t.sufficient)this.log(`  ${t.insufficientReason??"insufficient data"} \u2014 no effect claimed`);else{let e=t.delta>=0?"+":"";this.log(`  pass-rate delta: ${e}${(t.delta*100).toFixed(0)}% (n=${t.nWith} paired)`),this.log(`  95% CI: ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%`),t.pValue!=null&&this.log(`  McNemar p: ${t.pValue.toFixed(4)}`),this.log(`  significant: ${t.significant?"YES":"no"}`)}if(a.notes.length>0){this.log(`

package/dist/commands/evals/optimize.js

@@ -1,4 +1,4 @@
-import{b as P,c as C}from"../../chunk-QH7PCQHM.js";import{a as V}from"../../chunk-4VZLIGMS.js";import{a as N}from"../../chunk-7AZKPX2J.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-V5MJ4ICC.js";import"../../chunk-WKXTY5AM.js";import{c as F}from"../../chunk-HMMO27U4.js";import"../../chunk-MODFTCZY.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as A}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as w,f as m,g as S}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as k}from"../../chunk-ZLLER3HP.js";import{createInterface as _}from"node:readline/promises";import{Args as q,Flags as g}from"@oclif/core";var $=class O extends A{static{k(this,"EvalOptimizeCommand")}static description="Iteratively optimize eval target content using eval results as feedback";static examples=["<%= config.bin %> evals optimize <target-id>","<%= config.bin %> evals optimize <target-id> --auto --max-iterations 3"];static args={targetId:q.string({description:"Eval target ID",required:!0})};static flags={auto:g.boolean({description:"Skip interactive review, apply suggestions automatically",default:!1}),"max-iterations":g.integer({description:"Maximum optimization iterations",default:5}),threshold:g.string({description:"Stop if pass rate reaches this threshold (0-1)",default:"0.95"}),model:g.string({description:"Model for eval execution"}),"analyzer-model":g.string({description:"Model for analysis/suggestions"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:n,flags:e}=await this.parse(O);this.requireAuth();let t=await w(`/api/eval-targets/${n.targetId}`),i=await w(`/api/eval-targets/${n.targetId}/evals`);i.length===0&&this.error("No evals found. Create some first with `rulemetric evals create`.");let s=[...i].sort(()=>Math.random()-.5),c=Math.ceil(s.length*.6),u=s.slice(0,c),d=s.slice(c);this.log(`Optimizing "${t.name}" v${t.version}`),this.log(`Evals: ${u.length} train, ${d.length} test`);let h=parseFloat(e.threshold);this.log(`Config: max ${e["max-iterations"]} iterations, threshold ${(h*100).toFixed(0)}%`),e.auto&&this.log("Mode: auto (no interactive review)"),this.log("");let T=t.version,f=0,x=0;for(let l=1;l<=e["max-iterations"];l++){this.log(`
+import{b as P,c as C}from"../../chunk-MQ47NLPC.js";import{a as V}from"../../chunk-4VZLIGMS.js";import{a as N}from"../../chunk-4L2ZEIRE.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-XRTZU7IZ.js";import"../../chunk-B3BHB26P.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-G5BXZT6F.js";import{c as F}from"../../chunk-HMMO27U4.js";import{a as A}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as w,f as m,g as S}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as k}from"../../chunk-ZLLER3HP.js";import{createInterface as _}from"node:readline/promises";import{Args as q,Flags as g}from"@oclif/core";var $=class O extends A{static{k(this,"EvalOptimizeCommand")}static description="Iteratively optimize eval target content using eval results as feedback";static examples=["<%= config.bin %> evals optimize <target-id>","<%= config.bin %> evals optimize <target-id> --auto --max-iterations 3"];static args={targetId:q.string({description:"Eval target ID",required:!0})};static flags={auto:g.boolean({description:"Skip interactive review, apply suggestions automatically",default:!1}),"max-iterations":g.integer({description:"Maximum optimization iterations",default:5}),threshold:g.string({description:"Stop if pass rate reaches this threshold (0-1)",default:"0.95"}),model:g.string({description:"Model for eval execution"}),"analyzer-model":g.string({description:"Model for analysis/suggestions"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:n,flags:e}=await this.parse(O);this.requireAuth();let t=await w(`/api/eval-targets/${n.targetId}`),i=await w(`/api/eval-targets/${n.targetId}/evals`);i.length===0&&this.error("No evals found. Create some first with `rulemetric evals create`.");let s=[...i].sort(()=>Math.random()-.5),c=Math.ceil(s.length*.6),u=s.slice(0,c),d=s.slice(c);this.log(`Optimizing "${t.name}" v${t.version}`),this.log(`Evals: ${u.length} train, ${d.length} test`);let h=parseFloat(e.threshold);this.log(`Config: max ${e["max-iterations"]} iterations, threshold ${(h*100).toFixed(0)}%`),e.auto&&this.log("Mode: auto (no interactive review)"),this.log("");let T=t.version,f=0,x=0;for(let l=1;l<=e["max-iterations"];l++){this.log(`
 === Iteration ${l} ===`),this.log("Running train evals...");let b=await this.runEvalsOnTarget(t,u,e),p=this.computePassRate(b);if(this.log(`Train pass rate: ${(p*100).toFixed(1)}%`),p>=h){this.log(`Train score ${(p*100).toFixed(1)}% >= threshold ${(h*100).toFixed(0)}%. Stopping.`);let o=await this.runEvalsOnTarget(t,d,e),z=this.computePassRate(o);this.log(`Final test score: ${(z*100).toFixed(1)}%`),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,trainScore:p,testScore:z,analysis:{reason:"threshold_reached"},status:"completed"});break}this.log("Analyzing results...");let M=b.map(o=>({evalName:o.evalName,prompt:o.prompt,expectations:o.expectations,grading:o.grading})),a=await V({content:t.content,evalResults:M,model:e["analyzer-model"]??e.model,timeout:e.timeout});this.log(`
 Assessment: ${a.overallAssessment}`),this.log(`Suggestions: ${a.suggestions.length}`);for(let o of a.suggestions)this.log(`  - [${o.section}] ${o.reasoning}`);if(a.proposedContent===t.content){this.log("No changes proposed. Stopping."),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,trainScore:p,analysis:{reason:"no_changes"},status:"completed"});break}if(!e.auto&&!await this.confirm(`
 Apply the proposed changes as a new version? [y/N] `)){this.log("Changes not applied. Stopping."),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,trainScore:p,analysis:{reason:"user_declined",suggestions:a.suggestions},status:"completed"});break}let r=await S(`/api/eval-targets/${t.id}`,{content:a.proposedContent});this.log(`Created v${r.version}`),this.log(`Running test evals on v${t.version} (baseline) and v${r.version}...`);let E=await this.runEvalsOnTarget(t,d,e),R=await this.runEvalsOnTarget(r,d,e),v=this.computePassRate(R),I=this.computePassRate(E);this.log(`Test pass rate: v${t.version} ${(I*100).toFixed(1)}% \u2192 v${r.version} ${(v*100).toFixed(1)}%`);let y=P(E,R);if(y&&this.log(C(y,t.version,r.version)),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,outputVersion:r.version,trainScore:p,testScore:v,analysis:{suggestions:a.suggestions,overallAssessment:a.overallAssessment,priorTestScore:I,verdict:y},status:"completed"}),v>f)f=v,T=r.version,x=0;else if(x++,x>=2){this.log("No improvement for 2 consecutive iterations. Stopping.");break}t=r}this.log(`

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

@@ -1,4 +1,4 @@
-import{b as w}from"../../chunk-3VYV6ZYB.js";import"../../chunk-MODFTCZY.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as f}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as g,f as p,h as v}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as d}from"../../chunk-ZLLER3HP.js";import{Args as P,Flags as b}from"@oclif/core";var r=d(i=>i==null?"n/a":`${Math.round(i*100)}%`,"pct"),c=class i extends f{static{d(this,"EvalRefineCases")}static description="Retire non-discriminating eval cases and regenerate discriminating replacements (\xA76e)";static examples=["<%= config.bin %> evals refine-cases <target-id>            # preview","<%= config.bin %> evals refine-cases <target-id> --yes      # apply: retire + regenerate","<%= config.bin %> evals refine-cases <target-id> --yes --no-generate"];static args={targetId:P.string({description:"Eval target ID",required:!0})};static flags={yes:b.boolean({char:"y",description:"Apply (retire + regenerate). Without this, only previews."}),"no-generate":b.boolean({description:"Retire dead cases but do not regenerate replacements."})};async run(){let{args:$,flags:o}=await this.parse(i);this.requireAuth();let n=$.targetId,s=await g(`/api/eval-targets/${n}`),m=await g(`/api/eval-targets/${n}/judge-calibration`),t=m.caseDiscrimination;if(!t||t.cases.length===0&&t.unpairedCases.length===0){this.log("No case-discrimination data yet \u2014 run some evals with AND without the target first (`evals run` / the web Run Evals button), then refine.");return}let u=t.cases.filter(e=>!e.discriminating),a=[...u.map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate,withoutPassRate:e.withoutPassRate,reason:e.note}})),...t.unpairedCases.map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate??0,withoutPassRate:e.withoutPassRate,reason:"runs in only one cohort (no with-vs-without comparison possible)"}}))];if(this.log(`Target "${s.name}" v${s.version} \u2014 judge ${m.judge.currentModel??"(unstamped)"}`),this.log(`${t.total} diagnosable case(s): ${t.discriminating} discriminate, ${t.nonDiscriminating} flat/wrong-sign, ${t.unpaired} unpaired.`),a.length===0){this.log("Every case discriminates \u2014 nothing to retire. \u{1F389}");return}this.log(`
+import{b as w}from"../../chunk-CNOD6HWU.js";import"../../chunk-B3BHB26P.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as g,f as p,h as v}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as d}from"../../chunk-ZLLER3HP.js";import{Args as P,Flags as b}from"@oclif/core";var r=d(i=>i==null?"n/a":`${Math.round(i*100)}%`,"pct"),c=class i extends f{static{d(this,"EvalRefineCases")}static description="Retire non-discriminating eval cases and regenerate discriminating replacements (\xA76e)";static examples=["<%= config.bin %> evals refine-cases <target-id>            # preview","<%= config.bin %> evals refine-cases <target-id> --yes      # apply: retire + regenerate","<%= config.bin %> evals refine-cases <target-id> --yes --no-generate"];static args={targetId:P.string({description:"Eval target ID",required:!0})};static flags={yes:b.boolean({char:"y",description:"Apply (retire + regenerate). Without this, only previews."}),"no-generate":b.boolean({description:"Retire dead cases but do not regenerate replacements."})};async run(){let{args:$,flags:o}=await this.parse(i);this.requireAuth();let n=$.targetId,s=await g(`/api/eval-targets/${n}`),m=await g(`/api/eval-targets/${n}/judge-calibration`),t=m.caseDiscrimination;if(!t||t.cases.length===0&&t.unpairedCases.length===0){this.log("No case-discrimination data yet \u2014 run some evals with AND without the target first (`evals run` / the web Run Evals button), then refine.");return}let u=t.cases.filter(e=>!e.discriminating),a=[...u.map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate,withoutPassRate:e.withoutPassRate,reason:e.note}})),...t.unpairedCases.map(e=>({evalId:e.evalId,name:e.name,feedback:{name:e.name,withPassRate:e.withPassRate??0,withoutPassRate:e.withoutPassRate,reason:"runs in only one cohort (no with-vs-without comparison possible)"}}))];if(this.log(`Target "${s.name}" v${s.version} \u2014 judge ${m.judge.currentModel??"(unstamped)"}`),this.log(`${t.total} diagnosable case(s): ${t.discriminating} discriminate, ${t.nonDiscriminating} flat/wrong-sign, ${t.unpaired} unpaired.`),a.length===0){this.log("Every case discriminates \u2014 nothing to retire. \u{1F389}");return}this.log(`
 Would retire ${a.length} non-discriminating case(s):`);for(let e of u)this.log(`  \u2717 ${e.name} \u2014 ${r(e.withPassRate)} with vs ${r(e.withoutPassRate)} without (\u0394${(e.separation*100).toFixed(0)}pp) \u2014 ${e.note}`);for(let e of t.unpairedCases)this.log(`  \u2717 ${e.name} \u2014 ${r(e.withPassRate)} with vs ${r(e.withoutPassRate)} without \u2014 unpaired`);if(!o.yes){this.log(`
 Preview only. Re-run with --yes to retire ${a.length} case(s)${o["no-generate"]?"":" and generate replacements"}.`);return}for(let e of a)await v(`/api/eval-targets/${n}/evals/${e.evalId}`,{retire:!0,retiredReason:e.feedback.reason}),this.log(`  retired ${e.name}`);if(o["no-generate"]){this.log(`
 Retired ${a.length} case(s). Skipped regeneration (--no-generate).`);return}this.log(`

package/dist/commands/evals/run.js

@@ -1,3 +1,3 @@
-import{a as k}from"../../chunk-BSJBNOJT.js";import"../../chunk-L2CFSMBC.js";import"../../chunk-Z72VHGV7.js";import"../../chunk-7AZKPX2J.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-V5MJ4ICC.js";import"../../chunk-HMMO27U4.js";import"../../chunk-MODFTCZY.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as $}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as v,f as h}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as R}from"../../chunk-ZLLER3HP.js";import{Args as I,Flags as g}from"@oclif/core";var b=class _ extends ${static{R(this,"EvalRunCommand")}static description="Run evals against a target, optionally grade results, and post to API";static examples=["<%= config.bin %> evals run <target-id>","<%= config.bin %> evals run <target-id> --parallel 3 --runs-per-eval 2","<%= config.bin %> evals run <target-id> --skip-baseline --skip-grading"];static args={targetId:I.string({description:"Eval target ID",required:!0})};static flags={parallel:g.integer({description:"Number of parallel eval executions",default:1}),"runs-per-eval":g.integer({description:"Number of runs per eval for variance analysis",default:1}),iteration:g.integer({description:"Iteration number for tracking",default:1}),"skip-baseline":g.boolean({description:"Skip baseline (without_target) runs",default:!1}),"skip-grading":g.boolean({description:"Skip LLM grading step",default:!1}),model:g.string({description:"Model for eval execution (default: user configured)"}),"grader-model":g.string({description:"Model for grading (default: same as execution model)"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:c,flags:r}=await this.parse(_);this.requireAuth();let i=await v(`/api/eval-targets/${c.targetId}`),l=await v(`/api/eval-targets/${c.targetId}/evals`);l.length===0&&this.error(`No evals found for target "${i.name}". Create some with \`rulemetric evals create\`.`),this.log(`Running ${l.length} eval(s) for "${i.name}" v${i.version}`),this.log(`Config: ${r["runs-per-eval"]} run(s) per eval, parallel=${r.parallel}, iteration=${r.iteration}`),r["skip-baseline"]&&this.log("Skipping baseline runs"),r["skip-grading"]&&this.log("Skipping grading"),this.log("");let m=[{name:"with_target",content:i.content}];r["skip-baseline"]||m.push({name:"without_target"});let p=[];for(let a of l)for(let n of m)for(let t=1;t<=r["runs-per-eval"];t++)p.push({eval_:a,configuration:n.name,targetContent:n.content,runNumber:t});let u=[],y=r.parallel;for(let a=0;a<p.length;a+=y){let n=p.slice(a,a+y),t=await Promise.all(n.map(d=>this.executeAndGrade(d,i,r)));u.push(...t)}this.log(""),this.log("Posting results to API...");for(let a of u)await h("/api/eval-runs",a);try{await h(`/api/eval-targets/${c.targetId}/label-state`,{})}catch{}this.log(""),this.log("=== Results Summary ===");for(let a of m){let n=u.filter(e=>e.configuration===a.name),t=n.filter(e=>e.status==="completed"),d=t.filter(e=>Object.keys(e.grading).length>0);if(this.log(`
+import{a as k}from"../../chunk-KQCAELPN.js";import"../../chunk-L2CFSMBC.js";import"../../chunk-Z72VHGV7.js";import"../../chunk-4L2ZEIRE.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-XRTZU7IZ.js";import"../../chunk-B3BHB26P.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-HMMO27U4.js";import{a as $}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as v,f as h}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as R}from"../../chunk-ZLLER3HP.js";import{Args as I,Flags as g}from"@oclif/core";var b=class _ extends ${static{R(this,"EvalRunCommand")}static description="Run evals against a target, optionally grade results, and post to API";static examples=["<%= config.bin %> evals run <target-id>","<%= config.bin %> evals run <target-id> --parallel 3 --runs-per-eval 2","<%= config.bin %> evals run <target-id> --skip-baseline --skip-grading"];static args={targetId:I.string({description:"Eval target ID",required:!0})};static flags={parallel:g.integer({description:"Number of parallel eval executions",default:1}),"runs-per-eval":g.integer({description:"Number of runs per eval for variance analysis",default:1}),iteration:g.integer({description:"Iteration number for tracking",default:1}),"skip-baseline":g.boolean({description:"Skip baseline (without_target) runs",default:!1}),"skip-grading":g.boolean({description:"Skip LLM grading step",default:!1}),model:g.string({description:"Model for eval execution (default: user configured)"}),"grader-model":g.string({description:"Model for grading (default: same as execution model)"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:c,flags:r}=await this.parse(_);this.requireAuth();let i=await v(`/api/eval-targets/${c.targetId}`),l=await v(`/api/eval-targets/${c.targetId}/evals`);l.length===0&&this.error(`No evals found for target "${i.name}". Create some with \`rulemetric evals create\`.`),this.log(`Running ${l.length} eval(s) for "${i.name}" v${i.version}`),this.log(`Config: ${r["runs-per-eval"]} run(s) per eval, parallel=${r.parallel}, iteration=${r.iteration}`),r["skip-baseline"]&&this.log("Skipping baseline runs"),r["skip-grading"]&&this.log("Skipping grading"),this.log("");let m=[{name:"with_target",content:i.content}];r["skip-baseline"]||m.push({name:"without_target"});let p=[];for(let a of l)for(let n of m)for(let t=1;t<=r["runs-per-eval"];t++)p.push({eval_:a,configuration:n.name,targetContent:n.content,runNumber:t});let u=[],y=r.parallel;for(let a=0;a<p.length;a+=y){let n=p.slice(a,a+y),t=await Promise.all(n.map(d=>this.executeAndGrade(d,i,r)));u.push(...t)}this.log(""),this.log("Posting results to API...");for(let a of u)await h("/api/eval-runs",a);try{await h(`/api/eval-targets/${c.targetId}/label-state`,{})}catch{}this.log(""),this.log("=== Results Summary ===");for(let a of m){let n=u.filter(e=>e.configuration===a.name),t=n.filter(e=>e.status==="completed"),d=t.filter(e=>Object.keys(e.grading).length>0);if(this.log(`
 ${a.name}:`),this.log(`  Runs: ${t.length}/${n.length} completed`),d.length>0){let e=d.map(o=>o.grading.summary?.pass_rate??0),s=e.reduce((o,f)=>o+f,0)/e.length;this.log(`  Avg pass_rate: ${(s*100).toFixed(1)}%`)}for(let e of n){let s=l.find(x=>x.id===e.evalId)?.name??e.evalId,o=e.grading.summary?.pass_rate,f=o!==void 0?` pass_rate=${(o*100).toFixed(0)}%`:"",w=e.timing.duration_ms?` ${(e.timing.duration_ms/1e3).toFixed(1)}s`:"";this.log(`  [${e.status}] ${s} run#${e.runNumber}${f}${w}`)}}if(!r["skip-baseline"]&&u.length>0){let a=u.filter(t=>t.configuration==="with_target"&&Object.keys(t.grading).length>0),n=u.filter(t=>t.configuration==="without_target"&&Object.keys(t.grading).length>0);if(a.length>0&&n.length>0){let t=a.map(s=>s.grading.summary?.pass_rate??0).reduce((s,o)=>s+o,0)/a.length,d=n.map(s=>s.grading.summary?.pass_rate??0).reduce((s,o)=>s+o,0)/n.length,e=t-d;this.log(`
 Delta: ${e>0?"+":""}${(e*100).toFixed(1)}% (with_target vs without_target)`)}}}async executeAndGrade(c,r,i){return k(c,{targetId:r.id,targetVersion:r.version,iteration:i.iteration,timeout:i.timeout,model:i.model,graderModel:i["grader-model"],skipGrading:i["skip-grading"],targetType:r.type,targetName:r.name},l=>this.log(l))}};export{b as default};

package/dist/commands/evals/validate-judge.js

@@ -1 +1 @@
-import{c,d as m}from"../../chunk-OZQV5CDT.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-7AZKPX2J.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-V5MJ4ICC.js";import"../../chunk-HMMO27U4.js";import"../../chunk-OPUI7CM2.js";import"../../chunk-MODFTCZY.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as u}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as r}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as s}from"@oclif/core";var o=r(n=>n==null?"\u2014":`${(n*100).toFixed(0)}%`,"pct"),g=class n extends u{static{r(this,"EvalValidateJudge")}static description="Score a second model against your human judge-review labels (\xA76g validation pass)";static examples=["<%= config.bin %> evals validate-judge <target-id>                       # auto-detect installed CLIs","<%= config.bin %> evals validate-judge <target-id> --engine codex --model gpt-5.5","<%= config.bin %> evals validate-judge <target-id> --model claude-sonnet-5 --json","<%= config.bin %> evals validate-judge <target-id> --limit 5             # quick sample"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={model:s.string({description:"Second reviewer model. Default: resolved from your installed tooling (codex: your ~/.codex/config.toml model; claude: a tier that differs from the original judge)."}),engine:s.string({description:"CLI that runs the second reviewer. `auto` prefers your installed codex for a cross-vendor reviewer \u2014 two Claude models share a training lineage, so their agreement overstates independence.",options:["auto","claude","codex"],default:"auto"}),limit:s.integer({description:"Sample only the N most-recently-annotated runs (cost control / smoke)"}),"skip-reviewed":s.boolean({description:"Only re-grade assertions this reviewer model has not ruled on yet",default:!1}),parallel:s.integer({description:"Concurrent re-grades",default:3}),timeout:s.integer({description:"Timeout per re-grade in ms",default:12e4}),json:s.boolean({description:"Emit the full report as JSON"})};async run(){let{args:h,flags:i}=await this.parse(n);this.requireAuth();let a;try{a=await m({targetId:h.targetId,engineFlag:i.engine,modelFlag:i.model,limit:i.limit,parallel:i.parallel,timeoutMs:i.timeout,skipReviewed:i["skip-reviewed"],log:r(t=>this.log(t),"log"),warn:r(t=>this.warn(t),"warn")})}catch(t){throw t instanceof c&&this.error(t.message),t}let{report:e,judgeModels:l,reviewerModel:d}=a;if(l.length===1&&l[0]===d&&this.warn(`Second reviewer resolved to the SAME model as the original judge (${l[0]}) \u2014 this measures self-consistency, not independent review.`),i.json){this.log(JSON.stringify({target:a.targetName,judgeA:l,judgeB:d,persisted:a.persisted,alreadyReviewed:a.alreadyReviewed,...e},null,2));return}if(e.pairs===0&&a.alreadyReviewed>0){this.log(`All ${a.alreadyReviewed} labeled assertion(s) already reviewed by ${d} \u2014 nothing new to grade.`);return}if(this.log(""),this.log(`Second reviewer: ${d} \u2014 scored against ${e.pairs} human-labeled assertion(s)`+(e.skipped?` (${e.skipped} skipped)`:"")),this.log(""),this.log(`  reviewer vs HUMAN:   ${o(e.bVsHuman.overall.rate)} overall \xB7 ${o(e.bVsHuman.onJudgePass.rate)} where judge passed \xB7 ${o(e.bVsHuman.onJudgeFail.rate)} where judge failed`),this.log(`  original judge vs human: ${o(e.aVsHuman.overall.rate)} (your approve rate)`),this.log(`  reviewer vs judge:   ${o(e.bVsA.overall.rate)} (inter-model, not truth)`),this.log(`  Cohen's \u03BA (reviewer vs human): ${e.kappaBvsHuman==null?"undefined (degenerate marginals)":e.kappaBvsHuman.toFixed(2)}`),this.log(""),e.suggestedAnchorWeight!=null&&this.log(`  suggested \xA76g anchor weight if used as a passive label source: ${e.suggestedAnchorWeight.toFixed(2)} (misclassification attenuation (2\xB7acc\u22121)\xB2; validated:false either way)`),e.disagreements.length>0){this.log(""),this.log(`  ${e.disagreements.length} disagreement(s) with your labels \u2014 flagged on the Review tab:`);for(let t of e.disagreements.slice(0,10))this.log(`    run ${t.runId.slice(0,8)} #${t.expectationIndex}: human=${t.humanPassed?"pass":"fail"} reviewer=${t.judgeBPassed?"pass":"fail"} \u2014 ${t.text.slice(0,80)}`);e.disagreements.length>10&&this.log(`    \u2026 and ${e.disagreements.length-10} more (use --json for all)`)}}};export{g as default};
+import{c,d as m}from"../../chunk-X64R6JCJ.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-NTKO4H2K.js";import"../../chunk-4L2ZEIRE.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-XRTZU7IZ.js";import"../../chunk-B3BHB26P.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-HMMO27U4.js";import{a as u}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as r}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as s}from"@oclif/core";var o=r(n=>n==null?"\u2014":`${(n*100).toFixed(0)}%`,"pct"),g=class n extends u{static{r(this,"EvalValidateJudge")}static description="Score a second model against your human judge-review labels (\xA76g validation pass)";static examples=["<%= config.bin %> evals validate-judge <target-id>                       # auto-detect installed CLIs","<%= config.bin %> evals validate-judge <target-id> --engine codex --model gpt-5.5","<%= config.bin %> evals validate-judge <target-id> --model claude-sonnet-5 --json","<%= config.bin %> evals validate-judge <target-id> --limit 5             # quick sample"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={model:s.string({description:"Second reviewer model. Default: resolved from your installed tooling (codex: your ~/.codex/config.toml model; claude: a tier that differs from the original judge)."}),engine:s.string({description:"CLI that runs the second reviewer. `auto` prefers your installed codex for a cross-vendor reviewer \u2014 two Claude models share a training lineage, so their agreement overstates independence.",options:["auto","claude","codex"],default:"auto"}),limit:s.integer({description:"Sample only the N most-recently-annotated runs (cost control / smoke)"}),"skip-reviewed":s.boolean({description:"Only re-grade assertions this reviewer model has not ruled on yet",default:!1}),parallel:s.integer({description:"Concurrent re-grades",default:3}),timeout:s.integer({description:"Timeout per re-grade in ms",default:12e4}),json:s.boolean({description:"Emit the full report as JSON"})};async run(){let{args:h,flags:i}=await this.parse(n);this.requireAuth();let a;try{a=await m({targetId:h.targetId,engineFlag:i.engine,modelFlag:i.model,limit:i.limit,parallel:i.parallel,timeoutMs:i.timeout,skipReviewed:i["skip-reviewed"],log:r(t=>this.log(t),"log"),warn:r(t=>this.warn(t),"warn")})}catch(t){throw t instanceof c&&this.error(t.message),t}let{report:e,judgeModels:l,reviewerModel:d}=a;if(l.length===1&&l[0]===d&&this.warn(`Second reviewer resolved to the SAME model as the original judge (${l[0]}) \u2014 this measures self-consistency, not independent review.`),i.json){this.log(JSON.stringify({target:a.targetName,judgeA:l,judgeB:d,persisted:a.persisted,alreadyReviewed:a.alreadyReviewed,...e},null,2));return}if(e.pairs===0&&a.alreadyReviewed>0){this.log(`All ${a.alreadyReviewed} labeled assertion(s) already reviewed by ${d} \u2014 nothing new to grade.`);return}if(this.log(""),this.log(`Second reviewer: ${d} \u2014 scored against ${e.pairs} human-labeled assertion(s)`+(e.skipped?` (${e.skipped} skipped)`:"")),this.log(""),this.log(`  reviewer vs HUMAN:   ${o(e.bVsHuman.overall.rate)} overall \xB7 ${o(e.bVsHuman.onJudgePass.rate)} where judge passed \xB7 ${o(e.bVsHuman.onJudgeFail.rate)} where judge failed`),this.log(`  original judge vs human: ${o(e.aVsHuman.overall.rate)} (your approve rate)`),this.log(`  reviewer vs judge:   ${o(e.bVsA.overall.rate)} (inter-model, not truth)`),this.log(`  Cohen's \u03BA (reviewer vs human): ${e.kappaBvsHuman==null?"undefined (degenerate marginals)":e.kappaBvsHuman.toFixed(2)}`),this.log(""),e.suggestedAnchorWeight!=null&&this.log(`  suggested \xA76g anchor weight if used as a passive label source: ${e.suggestedAnchorWeight.toFixed(2)} (misclassification attenuation (2\xB7acc\u22121)\xB2; validated:false either way)`),e.disagreements.length>0){this.log(""),this.log(`  ${e.disagreements.length} disagreement(s) with your labels \u2014 flagged on the Review tab:`);for(let t of e.disagreements.slice(0,10))this.log(`    run ${t.runId.slice(0,8)} #${t.expectationIndex}: human=${t.humanPassed?"pass":"fail"} reviewer=${t.judgeBPassed?"pass":"fail"} \u2014 ${t.text.slice(0,80)}`);e.disagreements.length>10&&this.log(`    \u2026 and ${e.disagreements.length-10} more (use --json for all)`)}}};export{g as default};

package/dist/commands/gateway/ensure.js

@@ -1 +1 @@
-import{b as y,c as h}from"../../chunk-CYKELSHS.js";import{B as l,E as g}from"../../chunk-VLXCQ2DC.js";import"../../chunk-I2DZGJRR.js";import{c as e,g as p,i as m,k as f}from"../../chunk-MVM6OGAF.js";import"../../chunk-KKDHD4XA.js";import{b as d,d as u}from"../../chunk-SZ62CP55.js";import{a as c}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a}from"../../chunk-ZLLER3HP.js";import{spawn as S}from"node:child_process";import{existsSync as v,readFileSync as R,statSync as C,writeFileSync as M}from"node:fs";import{homedir as T}from"node:os";import{join as o}from"node:path";var s=o(T(),".config","rulemetric"),x=o(s,"proxy-addon.version"),_=o(s,"proxy.pid"),w=o(s,"proxy-drift-restart.at"),E=10*6e4,n=class P extends c{static{a(this,"GatewayEnsure")}static description="Ensure the gateway proxy is running (called by hooks, not for direct use)";static hidden=!0;async run(){await this.parse(P),this.maybeRestartStaleProxy();let t=await p(e);if(t.status==="ours-live"){let r=h(i=>{i==="vscode"?l(e):g(e)});r.length>0&&console.error(`[RuleMetric] Gateway healthy again \u2014 restored ${r.join(", ")} proxy settings.`);return}if(t.status==="foreign"){if(m())return;let r=y(e);console.error(`[RuleMetric] A non-RuleMetric process holds :${e} \u2014 cannot start gateway. `+(r.length>0?`Unpinned ${r.join(", ")} proxy settings so editors are not routed through it. `:"")+"Run `rulemetric diagnostics` if Claude Code HTTPS requests fail.");return}try{f(this.config.version)}catch(r){console.error(`[RuleMetric] Could not start gateway: ${r.message}`)}}maybeRestartStaleProxy(){try{if(!v(x))return;let t=R(x,"utf-8").trim();if(!t||t===this.config.version)return;let r=R(_,"utf-8").trim(),i=Number(r);if(!Number.isFinite(i)||!u(i,d))return;try{if(Date.now()-C(w).mtimeMs<E)return}catch{}M(w,new Date().toISOString()),S(process.execPath,[process.argv[1],"proxy","restart"],{detached:!0,stdio:"ignore"}).unref(),console.error(`[RuleMetric] Capture proxy is running CLI ${t} but ${this.config.version} is installed \u2014 restarting it in the background.`)}catch{}}};export{n as default};
+import{b as y,c as h}from"../../chunk-CYKELSHS.js";import{B as l,E as g}from"../../chunk-VLXCQ2DC.js";import"../../chunk-I2DZGJRR.js";import{e,i as p,k as m,m as f}from"../../chunk-QWCEZEDO.js";import"../../chunk-KKDHD4XA.js";import{b as d,d as u}from"../../chunk-SZ62CP55.js";import{a as c}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a}from"../../chunk-ZLLER3HP.js";import{spawn as S}from"node:child_process";import{existsSync as v,readFileSync as R,statSync as C,writeFileSync as M}from"node:fs";import{homedir as T}from"node:os";import{join as o}from"node:path";var s=o(T(),".config","rulemetric"),x=o(s,"proxy-addon.version"),_=o(s,"proxy.pid"),w=o(s,"proxy-drift-restart.at"),E=10*6e4,n=class P extends c{static{a(this,"GatewayEnsure")}static description="Ensure the gateway proxy is running (called by hooks, not for direct use)";static hidden=!0;async run(){await this.parse(P),this.maybeRestartStaleProxy();let t=await p(e);if(t.status==="ours-live"){let r=h(i=>{i==="vscode"?l(e):g(e)});r.length>0&&console.error(`[RuleMetric] Gateway healthy again \u2014 restored ${r.join(", ")} proxy settings.`);return}if(t.status==="foreign"){if(m())return;let r=y(e);console.error(`[RuleMetric] A non-RuleMetric process holds :${e} \u2014 cannot start gateway. `+(r.length>0?`Unpinned ${r.join(", ")} proxy settings so editors are not routed through it. `:"")+"Run `rulemetric diagnostics` if Claude Code HTTPS requests fail.");return}try{f(this.config.version)}catch(r){console.error(`[RuleMetric] Could not start gateway: ${r.message}`)}}maybeRestartStaleProxy(){try{if(!v(x))return;let t=R(x,"utf-8").trim();if(!t||t===this.config.version)return;let r=R(_,"utf-8").trim(),i=Number(r);if(!Number.isFinite(i)||!u(i,d))return;try{if(Date.now()-C(w).mtimeMs<E)return}catch{}M(w,new Date().toISOString()),S(process.execPath,[process.argv[1],"proxy","restart"],{detached:!0,stdio:"ignore"}).unref(),console.error(`[RuleMetric] Capture proxy is running CLI ${t} but ${this.config.version} is installed \u2014 restarting it in the background.`)}catch{}}};export{n as default};

package/dist/commands/hooks/install.js

@@ -1 +1 @@
-import{b as mt}from"../../chunk-UGIUIZ2V.js";import{a as yt}from"../../chunk-6XEC6474.js";import{B as gt,E as ht,H as pt,I as ft,e as Z,f as tt,g as et,i as st,j as ot,k as it,n as rt,p as at,q as nt,r as lt,u as ct,v as ut,y as dt}from"../../chunk-VLXCQ2DC.js";import{a as J,b as z,c as m,d as Q,g as _}from"../../chunk-I2DZGJRR.js";import{c as s,g as F,h as V,i as q,k as A,l as W}from"../../chunk-MVM6OGAF.js";import{e as D}from"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a as f}from"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as X}from"../../chunk-XICLXXRH.js";import{a as B}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import{c as Y}from"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import{b as M}from"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import{a as L}from"../../chunk-O6C2UGBY.js";import{a as w}from"../../chunk-ZLLER3HP.js";import{Flags as x}from"@oclif/core";import{execFileSync as Ct,spawnSync as l}from"node:child_process";import{existsSync as u}from"node:fs";import{homedir as vt}from"node:os";import{join as R,dirname as xt}from"node:path";function kt(){try{let g=Ct("/bin/sh",["-c","command -v rulemetric"],{encoding:"utf-8"}).trim();if(g&&u(g))return xt(g)}catch{}}w(kt,"resolveHookBinDir");function Pt(){try{return Ct("launchctl",["list","com.rulemetric.gateway"],{stdio:["ignore","pipe","pipe"]}),!0}catch{return!1}}w(Pt,"isGatewayLaunchdManaged");var o=R(vt(),".mitmproxy","mitmproxy-ca-cert.pem"),K=class g extends B{static{w(this,"HooksInstall")}static description="Install session tracking hooks and context capture proxy for Claude Code";static examples=["<%= config.bin %> hooks install","<%= config.bin %> hooks install --global","<%= config.bin %> hooks install --no-proxy"];static flags={"no-proxy":x.boolean({default:!1,description:"Skip proxy/gateway configuration (hooks only)"}),"project-dir":x.string({description:"Project directory to install hooks in (defaults to cwd)"}),global:x.boolean({char:"g",default:!1,description:"Install hooks globally (~/.claude/settings.json) so all projects are tracked"}),"statusline-usage":x.boolean({default:!1,description:"Install the statusline shim so Claude Code pushes live rate_limits on every refresh (makes /usage rings update in real time without needing the proxy)"})};async run(){let{flags:i}=await this.parse(g),r=i.global,a=r?vt():i["project-dir"]??process.cwd(),k=R(a,".claude"),y=it(k,r),$=r?"settings.json":"settings.local.json",P=i["no-proxy"],S=!1,h=P?"no_proxy_flag":void 0,p=!1,T=[];r&&this.log(`Installing hooks globally \u2192 ${y}`);let{settings:n,recoveredBackup:I}=Z(y);if(I&&this.log(`[!!] ${y} was not valid JSON \u2014 backed it up to ${I} and continued with a fresh file.`),!r){let e=rt(R(k,"settings.json"),s);e.migrated&&(this.log("[OK] Moved RuleMetric machine-local state out of the shared .claude/settings.json (it now lives in .claude/settings.local.json)"),e.backupPath&&this.log(`     Previous shared settings backed up \u2192 ${e.backupPath}`),e.removedStatusline&&u(J)&&!_(n)&&(n.statusLine=m))}let E=L();(E.accessToken||process.env.RULEMETRIC_ACCESS_TOKEN)&&!E.apiKey&&!process.env.RULEMETRIC_API_KEY&&await X(this.log.bind(this),this.warn.bind(this),this.config.version);let wt=st(n),C=kt();if(et(n,C),this.log(wt?`[OK] Refreshed RuleMetric Claude Code session hooks \u2192 .claude/${$}`:`[OK] Installed RuleMetric Claude Code session hooks \u2192 .claude/${$}`),!P){let e=!!f("mitmdump");if(e||(this.log(""),this.log("[!!] mitmproxy not found \u2014 installing..."),T=["brew","pipx","uv"].filter(t=>!!f(t)),e=this.installMitmproxy(),e||(this.log("[!!] Could not install mitmproxy. Skipping proxy config."),this.log("     Install manually: pipx install mitmproxy"),this.log("     Then re-run: rulemetric hooks install"))),u(o))this.log("[OK] CA certificate found");else{this.log("[..] Generating CA certificate...");let t=f("mitmdump");t&&(l(t,["--set","listen_port=0","-q"],{timeout:5e3,stdio:"ignore"}),u(o)?this.log(`[OK] CA certificate generated at ${o}`):this.log("[!!] Could not generate CA cert"))}let c=u(o);c&&!this.isCACertTrusted()?(this.log("[..] Installing CA certificate into system trust store (requires admin)..."),p=this.trustCACert(),p?this.log("[OK] CA certificate trusted by system"):(this.log("[!!] Could not install CA cert automatically."),process.platform==="darwin"?this.log(`     Run manually: sudo security add-trusted-cert -d -r trustRoot -k /Library/Keychains/System.keychain ${o}`):process.platform==="linux"&&this.log(`     Run manually: sudo cp ${o} /usr/local/share/ca-certificates/mitmproxy-ca.crt && sudo update-ca-certificates`))):c&&(this.log("[OK] CA certificate already trusted"),p=!0);let d=Pt(),O=await F(s);if(d)this.log("[OK] Gateway managed by launchd (com.rulemetric.gateway)");else if(O.status==="ours-live")this.log("[OK] Gateway already running");else if(O.status==="foreign"&&q()){this.log("[..] Restarting gateway (running build predates the identity handshake)..."),W(),await new Promise(t=>setTimeout(t,500));try{let t=A(this.config.version);this.log(`[OK] Gateway restarted (PID ${t})`)}catch(t){this.log(`[!!] Could not restart gateway: ${t.message}`)}}else if(O.status!=="foreign")try{let t=A(this.config.version);this.log(`[OK] Gateway started (PID ${t})`)}catch(t){this.log(`[!!] Could not start gateway: ${t.message}`)}let G=await V(s,{timeoutMs:4e3});if(!c)h=e?"ca_not_generated":T.length===0?"no_package_manager":"mitmproxy_install_failed",this.log("[!!] No CA certificate \u2014 NOT setting HTTPS_PROXY (capture stays hooks-only)."),this.log("     Install mitmproxy, then re-run: rulemetric hooks install");else if(G.status==="foreign"&&!d){h="gateway_port_foreign";let t=D(s);this.log(`[!!] Port :${s} is held by a process that is NOT the RuleMetric gateway \u2014 NOT setting HTTPS_PROXY.`),this.log(t?`     Holding the port: ${t}`:`     Could not identify the holder (try: lsof -nP -iTCP:${s} -sTCP:LISTEN)`),this.log("     Capture still works via hooks. Free the port, then re-run: rulemetric hooks install")}else if(G.status==="dead")h="gateway_not_live",this.log(`[!!] Gateway is not answering on :${s} \u2014 NOT setting HTTPS_PROXY.`),this.log("     Capture still works via hooks. To enable deep (proxy) capture later:"),this.log("     rulemetric proxy start  &&  rulemetric hooks install");else{let t=n.env??{};t.HTTPS_PROXY=`http://localhost:${s}`,t.NO_PROXY=ot,u(o)&&(t.NODE_EXTRA_CA_CERTS=o),n.env=t,S=!0,this.log(`[OK] HTTPS_PROXY set to gateway on :${s} (verified live)`),ct(a,s,o)&&this.log("[OK] Cursor proxy configured \u2192 .cursor/settings.json"),ut(a,s,o)&&this.log("[OK] VS Code workspace proxy configured \u2192 .vscode/settings.json"),gt(s)&&this.log("[OK] VS Code user proxy configured \u2192 ~/Library/Application Support/Code/User/settings.json"),ht(s)&&this.log("[OK] Cursor user proxy configured \u2192 ~/Library/Application Support/Cursor/User/settings.json"),p?dt(s,o)&&this.log("[OK] HTTPS_PROXY set for GUI apps (launchctl + LaunchAgent)"):(h="ca_untrusted",this.log("[!!] CA not trusted \u2014 skipping machine-wide GUI proxy (would break native-app TLS)."),this.log("     Claude Code capture is unaffected. Trust the CA + re-run to enable cross-app capture."))}}let v=[];if(!r){let e=at(a,C);e&&(this.log(`[OK] Cursor hooks configured \u2192 ${e}`),v.push(e));let c=lt(a,C);c&&(this.log(`[OK] Copilot Agent hooks configured \u2192 ${c}`),v.push(c));let d=pt(a);d&&(this.log(`[OK] Antigravity workspace hooks configured \u2192 ${d}`),v.push(d))}let H=ft();H&&this.log(`[OK] Antigravity user hooks configured \u2192 ${H}`);let U=nt(C);if(U&&this.log(`[OK] Cursor user hooks configured \u2192 ${U}`),i["statusline-usage"])if(_(n))this.log("[OK] Statusline shim already installed");else{let e=Q(n);z(e),n.statusLine=m,e?this.log(`[OK] Statusline shim installed \u2192 ${m.command} (chains to: ${e})`):this.log(`[OK] Statusline shim installed \u2192 ${m.command}`),this.log("     Rate-limit rings will now update on Claude Code's statusline refresh cadence")}let{backupPath:N}=tt(y,n);N&&this.log(`[OK] Backed up previous settings \u2192 ${N}`),this.log(""),this.log("Setup complete! Next steps:"),r?(this.log("  1. Install always-on API:    rulemetric service install"),this.log("  2. Start Claude Code in any project \u2014 all sessions tracked automatically"),this.log(""),this.log("The API service starts on login and restarts on crash.")):P?(this.log("  1. Start Claude Code \u2014 sessions are captured by the hooks just installed"),this.log(""),this.log("No proxy means no mitmproxy / CA cert needed. The hooks capture full"),this.log("sessions on their own; re-run without --no-proxy to also capture system"),this.log("prompts for instruction-effectiveness linking.")):(this.log("  1. Start Claude Code \u2014 sessions are captured by hooks immediately"),S?this.log("  2. Full-depth capture is ON \u2014 the proxy is live on :"+s+"."):this.log("  2. Enable full depth (proxy not yet live):  rulemetric proxy start"),this.log(""),this.log("Hooks capture every session (lifecycle + transcript). The proxy adds the"),this.log("rendered system prompt (instruction-effectiveness linking) + cross-tool capture."),this.log(""),this.log("To capture traffic from other clients (Python, curl, etc.):"),this.log('  eval "$(rulemetric proxy env --all)"'));let j=mt(a,v);j&&this.log(j),await M(),Y("hooks_installed",{proxy:S,caTrusted:p,reason:h,installers:T,global:r},this.config.version);let b=await yt();this.log(""),b.available?this.log(`[OK] tmux found${b.version?` (v${b.version})`:""} \u2014 live send enabled`):this.log("[!!] tmux not found \u2014 live send will fall back to opening Terminal. Install with: brew install tmux")}installMitmproxy(){for(let[i,r]of[["pipx",["install","mitmproxy"]],["uv",["tool","install","mitmproxy"]],["brew",["install","mitmproxy"]]]){let a=f(i);if(a&&(this.log(`     Installing via ${i}...`),l(a,[...r],{stdio:"inherit"}).status===0))return!0}return!1}isCACertTrusted(){return process.platform==="darwin"?l("security",["verify-cert","-c",o],{stdio:"pipe"}).status===0:process.platform==="linux"?u("/usr/local/share/ca-certificates/mitmproxy-ca.crt"):process.platform==="win32"?l("certutil",["-verify",o],{stdio:"pipe"}).status===0:!1}trustCACert(){return process.platform==="darwin"?l("sudo",["security","add-trusted-cert","-d","-r","trustRoot","-k","/Library/Keychains/System.keychain",o],{stdio:"inherit"}).status===0:process.platform==="linux"?l("sudo",["cp",o,"/usr/local/share/ca-certificates/mitmproxy-ca.crt"],{stdio:"inherit"}).status!==0?!1:l("sudo",["update-ca-certificates"],{stdio:"inherit"}).status===0:process.platform==="win32"?l("certutil",["-addstore","-f","Root",o],{stdio:"inherit"}).status===0:!1}};export{K as default};
+import{b as mt}from"../../chunk-UGIUIZ2V.js";import{a as yt}from"../../chunk-6XEC6474.js";import{B as ht,E as gt,H as pt,I as ft,e as Z,f as tt,g as et,i as st,j as ot,k as it,n as rt,p as at,q as nt,r as lt,u as ct,v as ut,y as dt}from"../../chunk-VLXCQ2DC.js";import{a as J,b as z,c as m,d as Q,g as R}from"../../chunk-I2DZGJRR.js";import{e as s,i as F,j as V,k as q,m as A,n as W}from"../../chunk-QWCEZEDO.js";import{e as D}from"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import"../../chunk-N7ODX6PO.js";import{a as f}from"../../chunk-PJ3WHQ67.js";import{a as X}from"../../chunk-XICLXXRH.js";import{a as B}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import{c as Y}from"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import{b as M}from"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import{a as L}from"../../chunk-O6C2UGBY.js";import{a as w}from"../../chunk-ZLLER3HP.js";import{Flags as y}from"@oclif/core";import{execFileSync as Ct,spawnSync as l}from"node:child_process";import{existsSync as u}from"node:fs";import{homedir as kt}from"node:os";import{join as _,dirname as wt}from"node:path";function xt(){try{let h=Ct("/bin/sh",["-c","command -v rulemetric"],{encoding:"utf-8"}).trim();if(h&&u(h))return wt(h)}catch{}}w(xt,"resolveHookBinDir");function Pt(){try{return Ct("launchctl",["list","com.rulemetric.gateway"],{stdio:["ignore","pipe","pipe"]}),!0}catch{return!1}}w(Pt,"isGatewayLaunchdManaged");var o=_(kt(),".mitmproxy","mitmproxy-ca-cert.pem"),K=class h extends B{static{w(this,"HooksInstall")}static description="Install session tracking hooks and context capture proxy for Claude Code";static examples=["<%= config.bin %> hooks install","<%= config.bin %> hooks install --global","<%= config.bin %> hooks install --no-proxy"];static flags={"no-proxy":y.boolean({default:!1,description:"Skip proxy/gateway configuration (hooks only)"}),"project-dir":y.string({description:"Project directory to install hooks in (defaults to cwd)"}),"user-hooks":y.boolean({description:"Write user-level Cursor/Antigravity hooks (~/.cursor, ~/.gemini). Required for Cursor capture \u2014 its traffic bypasses the proxy.",default:!0,allowNo:!0}),global:y.boolean({char:"g",default:!1,description:"Install hooks globally (~/.claude/settings.json) so all projects are tracked"}),"statusline-usage":y.boolean({default:!1,description:"Install the statusline shim so Claude Code pushes live rate_limits on every refresh (makes /usage rings update in real time without needing the proxy)"})};async run(){let{flags:i}=await this.parse(h),r=i.global,a=r?kt():i["project-dir"]??process.cwd(),x=_(a,".claude"),C=it(x,r),$=r?"settings.json":"settings.local.json",P=i["no-proxy"],b=!1,g=P?"no_proxy_flag":void 0,p=!1,S=[];r&&this.log(`Installing hooks globally \u2192 ${C}`);let{settings:n,recoveredBackup:I}=Z(C);if(I&&this.log(`[!!] ${C} was not valid JSON \u2014 backed it up to ${I} and continued with a fresh file.`),!r){let e=rt(_(x,"settings.json"),s);e.migrated&&(this.log("[OK] Moved RuleMetric machine-local state out of the shared .claude/settings.json (it now lives in .claude/settings.local.json)"),e.backupPath&&this.log(`     Previous shared settings backed up \u2192 ${e.backupPath}`),e.removedStatusline&&u(J)&&!R(n)&&(n.statusLine=m))}let E=L();(E.accessToken||process.env.RULEMETRIC_ACCESS_TOKEN)&&!E.apiKey&&!process.env.RULEMETRIC_API_KEY&&await X(this.log.bind(this),this.warn.bind(this),this.config.version);let vt=st(n),k=xt();if(et(n,k),this.log(vt?`[OK] Refreshed RuleMetric Claude Code session hooks \u2192 .claude/${$}`:`[OK] Installed RuleMetric Claude Code session hooks \u2192 .claude/${$}`),!P){let e=!!f("mitmdump");if(e||(this.log(""),this.log("[!!] mitmproxy not found \u2014 installing..."),S=["brew","pipx","uv"].filter(t=>!!f(t)),e=this.installMitmproxy(),e||(this.log("[!!] Could not install mitmproxy. Skipping proxy config."),this.log("     Install manually: pipx install mitmproxy"),this.log("     Then re-run: rulemetric hooks install"))),u(o))this.log("[OK] CA certificate found");else{this.log("[..] Generating CA certificate...");let t=f("mitmdump");t&&(l(t,["--set","listen_port=0","-q"],{timeout:5e3,stdio:"ignore"}),u(o)?this.log(`[OK] CA certificate generated at ${o}`):this.log("[!!] Could not generate CA cert"))}let c=u(o);c&&!this.isCACertTrusted()?(this.log("[..] Installing CA certificate into system trust store (requires admin)..."),p=this.trustCACert(),p?this.log("[OK] CA certificate trusted by system"):(this.log("[!!] Could not install CA cert automatically."),process.platform==="darwin"?this.log(`     Run manually: sudo security add-trusted-cert -d -r trustRoot -k /Library/Keychains/System.keychain ${o}`):process.platform==="linux"&&this.log(`     Run manually: sudo cp ${o} /usr/local/share/ca-certificates/mitmproxy-ca.crt && sudo update-ca-certificates`))):c&&(this.log("[OK] CA certificate already trusted"),p=!0);let d=Pt(),O=await F(s);if(d)this.log("[OK] Gateway managed by launchd (com.rulemetric.gateway)");else if(O.status==="ours-live")this.log("[OK] Gateway already running");else if(O.status==="foreign"&&q()){this.log("[..] Restarting gateway (running build predates the identity handshake)..."),W(),await new Promise(t=>setTimeout(t,500));try{let t=A(this.config.version);this.log(`[OK] Gateway restarted (PID ${t})`)}catch(t){this.log(`[!!] Could not restart gateway: ${t.message}`)}}else if(O.status!=="foreign")try{let t=A(this.config.version);this.log(`[OK] Gateway started (PID ${t})`)}catch(t){this.log(`[!!] Could not start gateway: ${t.message}`)}let G=await V(s,{timeoutMs:4e3});if(!c)g=e?"ca_not_generated":S.length===0?"no_package_manager":"mitmproxy_install_failed",this.log("[!!] No CA certificate \u2014 NOT setting HTTPS_PROXY (capture stays hooks-only)."),this.log("     Install mitmproxy, then re-run: rulemetric hooks install");else if(G.status==="foreign"&&!d){g="gateway_port_foreign";let t=D(s);this.log(`[!!] Port :${s} is held by a process that is NOT the RuleMetric gateway \u2014 NOT setting HTTPS_PROXY.`),this.log(t?`     Holding the port: ${t}`:`     Could not identify the holder (try: lsof -nP -iTCP:${s} -sTCP:LISTEN)`),this.log("     Capture still works via hooks. Free the port, then re-run: rulemetric hooks install")}else if(G.status==="dead")g="gateway_not_live",this.log(`[!!] Gateway is not answering on :${s} \u2014 NOT setting HTTPS_PROXY.`),this.log("     Capture still works via hooks. To enable deep (proxy) capture later:"),this.log("     rulemetric proxy start  &&  rulemetric hooks install");else{let t=n.env??{};t.HTTPS_PROXY=`http://localhost:${s}`,t.NO_PROXY=ot,u(o)&&(t.NODE_EXTRA_CA_CERTS=o),n.env=t,b=!0,this.log(`[OK] HTTPS_PROXY set to gateway on :${s} (verified live)`),ct(a,s,o)&&this.log("[OK] Cursor proxy configured \u2192 .cursor/settings.json"),ut(a,s,o)&&this.log("[OK] VS Code workspace proxy configured \u2192 .vscode/settings.json"),ht(s)&&this.log("[OK] VS Code user proxy configured \u2192 ~/Library/Application Support/Code/User/settings.json"),gt(s)&&this.log("[OK] Cursor user proxy configured \u2192 ~/Library/Application Support/Cursor/User/settings.json"),p?dt(s,o)&&this.log("[OK] HTTPS_PROXY set for GUI apps (launchctl + LaunchAgent)"):(g="ca_untrusted",this.log("[!!] CA not trusted \u2014 skipping machine-wide GUI proxy (would break native-app TLS)."),this.log("     Claude Code capture is unaffected. Trust the CA + re-run to enable cross-app capture."))}}let v=[];if(!r){let e=at(a,k);e&&(this.log(`[OK] Cursor hooks configured \u2192 ${e}`),v.push(e));let c=lt(a,k);c&&(this.log(`[OK] Copilot Agent hooks configured \u2192 ${c}`),v.push(c));let d=pt(a);d&&(this.log(`[OK] Antigravity workspace hooks configured \u2192 ${d}`),v.push(d))}let H=i["user-hooks"]?ft():null;H&&this.log(`[OK] Antigravity user hooks configured \u2192 ${H}`);let U=i["user-hooks"]?nt(k):null;if(U&&this.log(`[OK] Cursor user hooks configured \u2192 ${U}`),i["user-hooks"]||(this.log("[--] Skipped user-level Cursor/Antigravity hooks (--no-user-hooks)."),this.log("     Cursor capture will not work: its chat traffic bypasses the proxy,"),this.log("     so hooks are the only path. Claude Code capture is unaffected.")),i["statusline-usage"])if(R(n))this.log("[OK] Statusline shim already installed");else{let e=Q(n);z(e),n.statusLine=m,e?this.log(`[OK] Statusline shim installed \u2192 ${m.command} (chains to: ${e})`):this.log(`[OK] Statusline shim installed \u2192 ${m.command}`),this.log("     Rate-limit rings will now update on Claude Code's statusline refresh cadence")}let{backupPath:N}=tt(C,n);N&&this.log(`[OK] Backed up previous settings \u2192 ${N}`),this.log(""),this.log("Setup complete! Next steps:"),r?(this.log("  1. Install always-on API:    rulemetric service install"),this.log("  2. Start Claude Code in any project \u2014 all sessions tracked automatically"),this.log(""),this.log("The API service starts on login and restarts on crash.")):P?(this.log("  1. Start Claude Code \u2014 sessions are captured by the hooks just installed"),this.log(""),this.log("No proxy means no mitmproxy / CA cert needed. The hooks capture full"),this.log("sessions on their own; re-run without --no-proxy to also capture system"),this.log("prompts for instruction-effectiveness linking.")):(this.log("  1. Start Claude Code \u2014 sessions are captured by hooks immediately"),b?this.log("  2. Full-depth capture is ON \u2014 the proxy is live on :"+s+"."):this.log("  2. Enable full depth (proxy not yet live):  rulemetric proxy start"),this.log(""),this.log("Hooks capture every session (lifecycle + transcript). The proxy adds the"),this.log("rendered system prompt (instruction-effectiveness linking) + cross-tool capture."),this.log(""),this.log("To capture traffic from other clients (Python, curl, etc.):"),this.log('  eval "$(rulemetric proxy env --all)"'));let j=mt(a,v);j&&this.log(j),await M(),Y("hooks_installed",{proxy:b,caTrusted:p,reason:g,installers:S,global:r},this.config.version);let T=await yt();this.log(""),T.available?this.log(`[OK] tmux found${T.version?` (v${T.version})`:""} \u2014 live send enabled`):this.log("[!!] tmux not found \u2014 live send will fall back to opening Terminal. Install with: brew install tmux")}installMitmproxy(){for(let[i,r]of[["pipx",["install","mitmproxy"]],["uv",["tool","install","mitmproxy"]],["brew",["install","mitmproxy"]]]){let a=f(i);if(a&&(this.log(`     Installing via ${i}...`),l(a,[...r],{stdio:"inherit"}).status===0))return!0}return!1}isCACertTrusted(){return process.platform==="darwin"?l("security",["verify-cert","-c",o],{stdio:"pipe"}).status===0:process.platform==="linux"?u("/usr/local/share/ca-certificates/mitmproxy-ca.crt"):process.platform==="win32"?l("certutil",["-verify",o],{stdio:"pipe"}).status===0:!1}trustCACert(){return process.platform==="darwin"?l("sudo",["security","add-trusted-cert","-d","-r","trustRoot","-k","/Library/Keychains/System.keychain",o],{stdio:"inherit"}).status===0:process.platform==="linux"?l("sudo",["cp",o,"/usr/local/share/ca-certificates/mitmproxy-ca.crt"],{stdio:"inherit"}).status!==0?!1:l("sudo",["update-ca-certificates"],{stdio:"inherit"}).status===0:process.platform==="win32"?l("certutil",["-addstore","-f","Root",o],{stdio:"inherit"}).status===0:!1}};export{K as default};

package/dist/commands/hooks/uninstall.js

@@ -1,2 +1,2 @@
-import{C as x,F as k,J as C,K as S,o as h,s as p,t as g,w as v,x as y,z as R}from"../../chunk-VLXCQ2DC.js";import"../../chunk-I2DZGJRR.js";import{c as f,l as d}from"../../chunk-MVM6OGAF.js";import"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as i}from"../../chunk-ZLLER3HP.js";import{existsSync as n,readFileSync as A,writeFileSync as j}from"node:fs";import{spawnSync as t}from"node:child_process";import{homedir as O}from"node:os";import{join as c}from"node:path";var r=c(O(),".mitmproxy","mitmproxy-ca-cert.pem");function N(){return n(r)?process.platform==="darwin"?t("sudo",["security","remove-trusted-cert","-d",r],{stdio:"inherit"}).status===0:process.platform==="linux"?t("sudo",["rm","-f","/usr/local/share/ca-certificates/mitmproxy-ca.crt"],{stdio:"inherit"}).status!==0?!1:t("sudo",["update-ca-certificates","--fresh"],{stdio:"inherit"}).status===0:process.platform==="win32"?t("certutil",["-delstore","Root","mitmproxy"],{stdio:"inherit"}).status===0:!1:!1}i(N,"untrustCACert");var u=class extends a{static{i(this,"HooksUninstall")}static description="Remove session tracking hooks from the current project";static examples=["<%= config.bin %> hooks uninstall"];async run(){let w=c(process.cwd(),".claude"),l=h(w,f);if(l.length===0){this.log("No .claude/settings.json or .claude/settings.local.json found \u2014 nothing to uninstall.");return}let e=!1,m=!1;for(let o of l)o.recoveredBackup&&this.log(`[!!] ${o.configPath} was not valid JSON \u2014 backed up to ${o.recoveredBackup}; continuing.`),o.removedHooks&&this.log(`Removed RuleMetric hooks from .claude/${o.fileName}`),o.removedProxyEnv&&(this.log(`Removed HTTPS_PROXY, NO_PROXY and NODE_EXTRA_CA_CERTS from .claude/${o.fileName}`),m=!0),o.removedStatusline&&this.log(`Removed RuleMetric statusline shim from .claude/${o.fileName}`),e=e||o.changed;e||this.log("No hooks or proxy config found \u2014 nothing to uninstall."),p(process.cwd())&&(this.log("Removed RuleMetric hooks from .cursor/hooks.json"),e=!0),g()&&(this.log("Removed RuleMetric hooks from ~/.cursor/hooks.json"),e=!0);let s=c(process.cwd(),".cursor","settings.json");if(n(s))try{let o=JSON.parse(A(s,"utf-8"));o["http.proxy"]?.includes("localhost")&&(delete o["http.proxy"],delete o["http.proxyStrictSSL"],j(s,JSON.stringify(o,null,2)+`
-`),this.log("Removed proxy config from .cursor/settings.json"),e=!0)}catch{}y(process.cwd())&&(this.log("Removed RuleMetric hooks from .github/hooks/"),e=!0),C(process.cwd())&&(this.log("Removed RuleMetric block from .agents/hooks.json"),e=!0),S()&&(this.log("Removed RuleMetric block from ~/.gemini/config/hooks.json"),e=!0),v(process.cwd())&&(this.log("Removed proxy config from .vscode/settings.json"),e=!0),x()&&(this.log("Removed proxy config from VS Code user settings"),e=!0),k()&&(this.log("Removed proxy config from Cursor user settings"),e=!0),R()&&(this.log("Removed HTTPS_PROXY LaunchAgent"),e=!0),m&&n(r)&&(this.log("Removing the mitmproxy CA from the system trust store (requires sudo)..."),N()?this.log("[OK] mitmproxy CA untrusted."):(this.log("[!!] Could not untrust the CA automatically. Remove it manually:"),process.platform==="darwin"?this.log(`     sudo security remove-trusted-cert -d ${r}`):process.platform==="linux"&&this.log("     sudo rm -f /usr/local/share/ca-certificates/mitmproxy-ca.crt && sudo update-ca-certificates --fresh"))),d()&&this.log("Gateway stopped.")}};export{u as default};
+import{C as j,F as b,J as A,K as w,o as R,s as k,t as x,w as C,x as S,z as P}from"../../chunk-VLXCQ2DC.js";import"../../chunk-I2DZGJRR.js";import{e as v,n as y}from"../../chunk-QWCEZEDO.js";import"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a as p}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as l}from"../../chunk-ZLLER3HP.js";import{existsSync as c,readFileSync as N,writeFileSync as T}from"node:fs";import{spawnSync as s}from"node:child_process";import{homedir as O}from"node:os";import{join as u}from"node:path";import{Flags as a}from"@oclif/core";var i=u(O(),".mitmproxy","mitmproxy-ca-cert.pem");function E(){return c(i)?process.platform==="darwin"?s("sudo",["security","remove-trusted-cert","-d",i],{stdio:"inherit"}).status===0:process.platform==="linux"?s("sudo",["rm","-f","/usr/local/share/ca-certificates/mitmproxy-ca.crt"],{stdio:"inherit"}).status!==0?!1:s("sudo",["update-ca-certificates","--fresh"],{stdio:"inherit"}).status===0:process.platform==="win32"?s("certutil",["-delstore","Root","mitmproxy"],{stdio:"inherit"}).status===0:!1:!1}l(E,"untrustCACert");var m=class f extends p{static{l(this,"HooksUninstall")}static flags={global:a.boolean({description:"Remove the GLOBAL install (~/.claude) rather than this project",default:!1}),"project-dir":a.string({description:"Project directory to clean (defaults to cwd)"}),"user-hooks":a.boolean({description:"Also strip user-level Cursor/Antigravity hooks (~/.cursor, ~/.gemini)",default:!0,allowNo:!0})};static description="Remove session tracking hooks from a project (or the global install)";static examples=["<%= config.bin %> hooks uninstall","<%= config.bin %> hooks uninstall --global","<%= config.bin %> hooks uninstall --project-dir ~/work/repo-a","<%= config.bin %> hooks uninstall --no-user-hooks"];async run(){let{flags:r}=await this.parse(f),t=r.global?O():r["project-dir"]??process.cwd(),d=u(t,".claude");r.global&&this.log(`Removing the GLOBAL install \u2192 ${d}`);let h=R(d,v);if(h.length===0){this.log("No .claude/settings.json or .claude/settings.local.json found \u2014 nothing to uninstall.");return}let e=!1,g=!1;for(let o of h)o.recoveredBackup&&this.log(`[!!] ${o.configPath} was not valid JSON \u2014 backed up to ${o.recoveredBackup}; continuing.`),o.removedHooks&&this.log(`Removed RuleMetric hooks from .claude/${o.fileName}`),o.removedProxyEnv&&(this.log(`Removed HTTPS_PROXY, NO_PROXY and NODE_EXTRA_CA_CERTS from .claude/${o.fileName}`),g=!0),o.removedStatusline&&this.log(`Removed RuleMetric statusline shim from .claude/${o.fileName}`),e=e||o.changed;e||this.log("No hooks or proxy config found \u2014 nothing to uninstall."),k(t)&&(this.log("Removed RuleMetric hooks from .cursor/hooks.json"),e=!0),r["user-hooks"]&&x()&&(this.log("Removed RuleMetric hooks from ~/.cursor/hooks.json"),e=!0);let n=u(t,".cursor","settings.json");if(c(n))try{let o=JSON.parse(N(n,"utf-8"));o["http.proxy"]?.includes("localhost")&&(delete o["http.proxy"],delete o["http.proxyStrictSSL"],T(n,JSON.stringify(o,null,2)+`
+`),this.log("Removed proxy config from .cursor/settings.json"),e=!0)}catch{}S(t)&&(this.log("Removed RuleMetric hooks from .github/hooks/"),e=!0),A(t)&&(this.log("Removed RuleMetric block from .agents/hooks.json"),e=!0),r["user-hooks"]&&w()&&(this.log("Removed RuleMetric block from ~/.gemini/config/hooks.json"),e=!0),C(t)&&(this.log("Removed proxy config from .vscode/settings.json"),e=!0),j()&&(this.log("Removed proxy config from VS Code user settings"),e=!0),b()&&(this.log("Removed proxy config from Cursor user settings"),e=!0),P()&&(this.log("Removed HTTPS_PROXY LaunchAgent"),e=!0),g&&c(i)&&(this.log("Removing the mitmproxy CA from the system trust store (requires sudo)..."),E()?this.log("[OK] mitmproxy CA untrusted."):(this.log("[!!] Could not untrust the CA automatically. Remove it manually:"),process.platform==="darwin"?this.log(`     sudo security remove-trusted-cert -d ${i}`):process.platform==="linux"&&this.log("     sudo rm -f /usr/local/share/ca-certificates/mitmproxy-ca.crt && sudo update-ca-certificates --fresh"))),y()&&this.log("Gateway stopped.")}};export{m as default};

package/dist/commands/proxy/env.js

@@ -1,3 +1,3 @@
-import{c as f}from"../../chunk-MVM6OGAF.js";import"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a as h}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as m}from"../../chunk-ZLLER3HP.js";import{Flags as i}from"@oclif/core";import{existsSync as v}from"node:fs";import{homedir as x}from"node:os";import{join as g}from"node:path";var o=g(x(),".mitmproxy","mitmproxy-ca-cert.pem"),l=class d extends h{static{m(this,"ProxyEnv")}static description="Print shell env vars to route any client through the RuleMetric proxy";static examples=["<%= config.bin %> proxy env","<%= config.bin %> proxy env --shell fish",'eval "$(rulemetric proxy env)"',"rulemetric proxy env --python"];static flags={shell:i.string({char:"s",default:"posix",description:"Shell format (posix, fish, powershell)",options:["posix","fish","powershell"]}),python:i.boolean({default:!1,description:"Include Python-specific env vars (REQUESTS_CA_BUNDLE, SSL_CERT_FILE)"}),node:i.boolean({default:!1,description:"Include Node.js-specific env vars (NODE_EXTRA_CA_CERTS)"}),all:i.boolean({default:!1,description:"Include env vars for all runtimes"}),vscode:i.boolean({default:!1,description:"Output VS Code settings.json snippet for Copilot proxy"})};async run(){let{flags:s}=await this.parse(d),y=s.shell,c=s.all,a=s.python||c,p=s.node||c,u=v(o),n=`http://localhost:${f}`;if(u||(this.warn(`CA certificate not found at ${o}
+import{e as f}from"../../chunk-QWCEZEDO.js";import"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a as h}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as m}from"../../chunk-ZLLER3HP.js";import{Flags as i}from"@oclif/core";import{existsSync as v}from"node:fs";import{homedir as x}from"node:os";import{join as g}from"node:path";var o=g(x(),".mitmproxy","mitmproxy-ca-cert.pem"),l=class d extends h{static{m(this,"ProxyEnv")}static description="Print shell env vars to route any client through the RuleMetric proxy";static examples=["<%= config.bin %> proxy env","<%= config.bin %> proxy env --shell fish",'eval "$(rulemetric proxy env)"',"rulemetric proxy env --python"];static flags={shell:i.string({char:"s",default:"posix",description:"Shell format (posix, fish, powershell)",options:["posix","fish","powershell"]}),python:i.boolean({default:!1,description:"Include Python-specific env vars (REQUESTS_CA_BUNDLE, SSL_CERT_FILE)"}),node:i.boolean({default:!1,description:"Include Node.js-specific env vars (NODE_EXTRA_CA_CERTS)"}),all:i.boolean({default:!1,description:"Include env vars for all runtimes"}),vscode:i.boolean({default:!1,description:"Output VS Code settings.json snippet for Copilot proxy"})};async run(){let{flags:s}=await this.parse(d),y=s.shell,c=s.all,a=s.python||c,p=s.node||c,u=v(o),n=`http://localhost:${f}`;if(u||(this.warn(`CA certificate not found at ${o}
 Run 'rulemetric hooks install' or 'rulemetric proxy setup' to generate it.`),this.exit(1)),s.vscode){let e={"http.proxy":n,"http.proxyStrictSSL":u,"github.copilot.advanced":{"debug.overrideProxyUrl":n}};this.log(JSON.stringify(e,null,2));return}let r=[{key:"HTTPS_PROXY",value:n,comment:"Route HTTPS through RuleMetric gateway"},{key:"HTTP_PROXY",value:n,comment:"Route HTTP through RuleMetric gateway"}];r.push({key:"SSL_CERT_FILE",value:o,comment:"Trust mitmproxy CA (generic \u2014 works with most libraries)"}),(p||!a&&!p)&&r.push({key:"NODE_EXTRA_CA_CERTS",value:o,comment:"Trust mitmproxy CA (Node.js)"}),a&&(r.push({key:"REQUESTS_CA_BUNDLE",value:o,comment:"Trust mitmproxy CA (Python requests / httpx)"}),r.push({key:"CURL_CA_BUNDLE",value:o,comment:"Trust mitmproxy CA (curl / libcurl)"}));let t=[];for(let e of r)switch(y){case"fish":e.comment&&t.push(`# ${e.comment}`),t.push(`set -gx ${e.key} "${e.value}"`);break;case"powershell":e.comment&&t.push(`# ${e.comment}`),t.push(`$env:${e.key} = "${e.value}"`);break;default:e.comment&&t.push(`# ${e.comment}`),t.push(`export ${e.key}="${e.value}"`)}this.log(t.join(`
 `))}};export{l as default};

package/dist/commands/proxy/status.js

@@ -1 +1 @@
-import{c as m,d as o,i as l,j as d}from"../../chunk-MVM6OGAF.js";import"../../chunk-KKDHD4XA.js";import{b as p,c as u,d as g}from"../../chunk-SZ62CP55.js";import{a as c}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a}from"../../chunk-ZLLER3HP.js";import{spawnSync as x}from"node:child_process";import{existsSync as M,readFileSync as P}from"node:fs";import{homedir as I}from"node:os";import{join as y}from"node:path";var D=y(I(),".config","rulemetric"),f=y(D,"proxy.pid"),e=class h extends c{static{a(this,"ProxyStatus")}static description="Check gateway and context capture proxy status";static examples=["<%= config.bin %> proxy status"];async run(){if(await this.parse(h),l()){let n=d();this.log(`Gateway:   running (PID ${n}) on :${m}`)}else this.log("Gateway:   not running");if(!M(f)){this.log("Mitmproxy: not running \u2014 routing direct (no capture)");return}let i=P(f,"utf-8").trim();if(i.startsWith("docker:")){let n=i.slice(7),r=x("docker",["container","inspect","-f","{{.State.Status}}",n],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]}),s=r.stdout?.trim();r.status===0&&s==="running"?this.log(`Mitmproxy: running in Docker (container: ${n}) on :${o} \u2014 capturing context`):this.log(`Mitmproxy: container ${s||"not running"} (stale)`);return}let t=Number.parseInt(i,10);if(Number.isNaN(t)){this.log("Mitmproxy: not running (invalid PID file)");return}g(t,p)?this.log(`Mitmproxy: running (PID ${t}) on :${o} \u2014 capturing context`):u(t)?this.log(`Mitmproxy: not running (PID ${t} was reused by another process \u2014 stale PID file)`):this.log("Mitmproxy: not running (stale PID file)")}};export{e as default};
+import{e as m,f as o,k as l,l as d}from"../../chunk-QWCEZEDO.js";import"../../chunk-KKDHD4XA.js";import{b as p,c as u,d as g}from"../../chunk-SZ62CP55.js";import{a as c}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a}from"../../chunk-ZLLER3HP.js";import{spawnSync as x}from"node:child_process";import{existsSync as M,readFileSync as P}from"node:fs";import{homedir as I}from"node:os";import{join as y}from"node:path";var D=y(I(),".config","rulemetric"),f=y(D,"proxy.pid"),e=class h extends c{static{a(this,"ProxyStatus")}static description="Check gateway and context capture proxy status";static examples=["<%= config.bin %> proxy status"];async run(){if(await this.parse(h),l()){let n=d();this.log(`Gateway:   running (PID ${n}) on :${m}`)}else this.log("Gateway:   not running");if(!M(f)){this.log("Mitmproxy: not running \u2014 routing direct (no capture)");return}let i=P(f,"utf-8").trim();if(i.startsWith("docker:")){let n=i.slice(7),r=x("docker",["container","inspect","-f","{{.State.Status}}",n],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]}),s=r.stdout?.trim();r.status===0&&s==="running"?this.log(`Mitmproxy: running in Docker (container: ${n}) on :${o} \u2014 capturing context`):this.log(`Mitmproxy: container ${s||"not running"} (stale)`);return}let t=Number.parseInt(i,10);if(Number.isNaN(t)){this.log("Mitmproxy: not running (invalid PID file)");return}g(t,p)?this.log(`Mitmproxy: running (PID ${t}) on :${o} \u2014 capturing context`):u(t)?this.log(`Mitmproxy: not running (PID ${t} was reused by another process \u2014 stale PID file)`):this.log("Mitmproxy: not running (stale PID file)")}};export{e as default};