@rulemetric/cli 0.7.20 → 0.7.22

package/dist/commands/evals/run.js

@@ -1,3 +1,3 @@
-import{a as w}from"../../chunk-IL35G6CK.js";import"../../chunk-S6B2R2LL.js";import{b as _}from"../../chunk-WQ2BJQYB.js";import{a as R}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 h,f as k}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as $}from"../../chunk-ZLLER3HP.js";import{Args as T,Flags as c}from"@oclif/core";var b=class x extends R{static{$(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:T.string({description:"Eval target ID",required:!0})};static flags={parallel:c.integer({description:"Number of parallel eval executions",default:1}),"runs-per-eval":c.integer({description:"Number of runs per eval for variance analysis",default:1}),iteration:c.integer({description:"Iteration number for tracking",default:1}),"skip-baseline":c.boolean({description:"Skip baseline (without_target) runs",default:!1}),"skip-grading":c.boolean({description:"Skip LLM grading step",default:!1}),model:c.string({description:"Model for eval execution (default: user configured)"}),"grader-model":c.string({description:"Model for grading (default: same as execution model)"}),timeout:c.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:f,flags:i}=await this.parse(x);this.requireAuth();let s=await h(`/api/eval-targets/${f.targetId}`),r=await h(`/api/eval-targets/${f.targetId}/evals`);r.length===0&&this.error(`No evals found for target "${s.name}". Create some with \`rulemetric evals create\`.`),this.log(`Running ${r.length} eval(s) for "${s.name}" v${s.version}`),this.log(`Config: ${i["runs-per-eval"]} run(s) per eval, parallel=${i.parallel}, iteration=${i.iteration}`),i["skip-baseline"]&&this.log("Skipping baseline runs"),i["skip-grading"]&&this.log("Skipping grading"),this.log("");let l=[{name:"with_target",content:s.content}];i["skip-baseline"]||l.push({name:"without_target"});let p=[];for(let e of r)for(let a of l)for(let n=1;n<=i["runs-per-eval"];n++)p.push({eval_:e,configuration:a.name,targetContent:a.content,runNumber:n});let o=[],g=i.parallel;for(let e=0;e<p.length;e+=g){let a=p.slice(e,e+g),n=await Promise.all(a.map(m=>this.executeAndGrade(m,s,i)));o.push(...n)}this.log(""),this.log("Posting results to API...");for(let e of o)await k("/api/eval-runs",e);this.log(""),this.log("=== Results Summary ===");for(let e of l){let a=o.filter(t=>t.configuration===e.name),n=a.filter(t=>t.status==="completed"),m=n.filter(t=>Object.keys(t.grading).length>0);if(this.log(`
-${e.name}:`),this.log(`  Runs: ${n.length}/${a.length} completed`),m.length>0){let t=m.map(d=>d.grading.summary?.pass_rate??0),u=t.reduce((d,v)=>d+v,0)/t.length;this.log(`  Avg pass_rate: ${(u*100).toFixed(1)}%`)}for(let t of a){let u=r.find(I=>I.id===t.evalId)?.name??t.evalId,d=t.grading.summary?.pass_rate,v=d!==void 0?` pass_rate=${(d*100).toFixed(0)}%`:"",y=t.timing.duration_ms?` ${(t.timing.duration_ms/1e3).toFixed(1)}s`:"";this.log(`  [${t.status}] ${u} run#${t.runNumber}${v}${y}`)}}if(!i["skip-baseline"]&&o.length>0){let e=o.filter(n=>n.configuration==="with_target"&&Object.keys(n.grading).length>0),a=o.filter(n=>n.configuration==="without_target"&&Object.keys(n.grading).length>0);if(e.length>0&&a.length>0){let n=e.map(u=>u.grading.summary?.pass_rate??0).reduce((u,d)=>u+d,0)/e.length,m=a.map(u=>u.grading.summary?.pass_rate??0).reduce((u,d)=>u+d,0)/a.length,t=n-m;this.log(`
-Delta: ${t>0?"+":""}${(t*100).toFixed(1)}% (with_target vs without_target)`)}}}async executeAndGrade(f,i,s){let{eval_:r,configuration:l,targetContent:p,runNumber:o}=f;this.log(`[${l}] ${r.name} run#${o} \u2014 executing...`);let g=await _({prompt:r.prompt,targetContent:p,timeout:s.timeout,model:s.model});if(g.error)return this.log(`[${l}] ${r.name} run#${o} \u2014 FAILED: ${g.error}`),{evalId:r.id,evalTargetId:i.id,evalTargetVersion:i.version,configuration:l,runNumber:o,iteration:s.iteration,status:"failed",outputs:{error:g.error},timing:{duration_ms:g.durationMs,total_tokens:g.totalTokens},grading:{}};this.log(`[${l}] ${r.name} run#${o} \u2014 completed in ${(g.durationMs/1e3).toFixed(1)}s`);let e={};if(!s["skip-grading"]&&r.expectations.length>0){this.log(`[${l}] ${r.name} run#${o} \u2014 grading...`);let a=await w({prompt:r.prompt,output:g.output,expectations:r.expectations,model:s["grader-model"]??s.model,timeout:s.timeout});e=a;let n=a.summary.pass_rate;this.log(`[${l}] ${r.name} run#${o} \u2014 grading done: pass_rate=${(n*100).toFixed(0)}%`)}return{evalId:r.id,evalTargetId:i.id,evalTargetVersion:i.version,configuration:l,runNumber:o,iteration:s.iteration,status:"completed",outputs:{transcript:g.output},timing:{duration_ms:g.durationMs,total_tokens:g.totalTokens,total_duration_seconds:g.durationMs/1e3},grading:e}}};export{b as default};
+import{a as $}from"../../chunk-UHL5NOWQ.js";import"../../chunk-YDRZDBHG.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-K25S2MGU.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as y}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import{e as v,f as k}from"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import{a as R}from"../../chunk-ZLLER3HP.js";import{Args as I,Flags as g}from"@oclif/core";var h=class _ extends y{static{R(this,"EvalRunCommand")}static description="Run evals against a target, optionally grade results, and post to API";static examples=["<%= config.bin %> evals run <target-id>","<%= config.bin %> evals run <target-id> --parallel 3 --runs-per-eval 2","<%= config.bin %> evals run <target-id> --skip-baseline --skip-grading"];static args={targetId:I.string({description:"Eval target ID",required:!0})};static flags={parallel:g.integer({description:"Number of parallel eval executions",default:1}),"runs-per-eval":g.integer({description:"Number of runs per eval for variance analysis",default:1}),iteration:g.integer({description:"Iteration number for tracking",default:1}),"skip-baseline":g.boolean({description:"Skip baseline (without_target) runs",default:!1}),"skip-grading":g.boolean({description:"Skip LLM grading step",default:!1}),model:g.string({description:"Model for eval execution (default: user configured)"}),"grader-model":g.string({description:"Model for grading (default: same as execution model)"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:c,flags:n}=await this.parse(_);this.requireAuth();let i=await v(`/api/eval-targets/${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: ${n["runs-per-eval"]} run(s) per eval, parallel=${n.parallel}, iteration=${n.iteration}`),n["skip-baseline"]&&this.log("Skipping baseline runs"),n["skip-grading"]&&this.log("Skipping grading"),this.log("");let m=[{name:"with_target",content:i.content}];n["skip-baseline"]||m.push({name:"without_target"});let p=[];for(let r of l)for(let a of m)for(let t=1;t<=n["runs-per-eval"];t++)p.push({eval_:r,configuration:a.name,targetContent:a.content,runNumber:t});let u=[],b=n.parallel;for(let r=0;r<p.length;r+=b){let a=p.slice(r,r+b),t=await Promise.all(a.map(d=>this.executeAndGrade(d,i,n)));u.push(...t)}this.log(""),this.log("Posting results to API...");for(let r of u)await k("/api/eval-runs",r);this.log(""),this.log("=== Results Summary ===");for(let r of m){let a=u.filter(e=>e.configuration===r.name),t=a.filter(e=>e.status==="completed"),d=t.filter(e=>Object.keys(e.grading).length>0);if(this.log(`
+${r.name}:`),this.log(`  Runs: ${t.length}/${a.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: ${(s*100).toFixed(1)}%`)}for(let e of a){let s=l.find(x=>x.id===e.evalId)?.name??e.evalId,o=e.grading.summary?.pass_rate,f=o!==void 0?` pass_rate=${(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(!n["skip-baseline"]&&u.length>0){let r=u.filter(t=>t.configuration==="with_target"&&Object.keys(t.grading).length>0),a=u.filter(t=>t.configuration==="without_target"&&Object.keys(t.grading).length>0);if(r.length>0&&a.length>0){let t=r.map(s=>s.grading.summary?.pass_rate??0).reduce((s,o)=>s+o,0)/r.length,d=a.map(s=>s.grading.summary?.pass_rate??0).reduce((s,o)=>s+o,0)/a.length,e=t-d;this.log(`
+Delta: ${e>0?"+":""}${(e*100).toFixed(1)}% (with_target vs without_target)`)}}}async executeAndGrade(c,n,i){return $(c,{targetId:n.id,targetVersion:n.version,iteration:i.iteration,timeout:i.timeout,model:i.model,graderModel:i["grader-model"],skipGrading:i["skip-grading"]},l=>this.log(l))}};export{h as default};

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

@@ -0,0 +1 @@
+import{b as N,c as O}from"../../chunk-6UEFCMR2.js";import{a as T}from"../../chunk-YDRZDBHG.js";import"../../chunk-S6B2R2LL.js";import{c as L}from"../../chunk-K25S2MGU.js";import"../../chunk-WQ2BJQYB.js";import{b as F,c as m,d as _,e as H}from"../../chunk-3BVHPQZW.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as B}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 j}from"../../chunk-ZLLER3HP.js";import{Args as J,Flags as d}from"@oclif/core";var I=100,p=j(h=>h==null?"\u2014":`${(h*100).toFixed(0)}%`,"pct"),A=class h extends B{static{j(this,"EvalValidateJudge")}static description="Score a second model against your human judge-review labels (\xA76g validation pass)";static examples=["<%= config.bin %> evals validate-judge <target-id>                       # auto-detect installed CLIs","<%= config.bin %> evals validate-judge <target-id> --engine codex --model gpt-5.5","<%= config.bin %> evals validate-judge <target-id> --model claude-sonnet-5 --json","<%= config.bin %> evals validate-judge <target-id> --limit 5             # quick sample"];static args={targetId:J.string({description:"Eval target ID",required:!0})};static flags={model:d.string({description:"Second reviewer model. Default: resolved from your installed tooling (codex: your ~/.codex/config.toml model; claude: a tier that differs from the original judge)."}),engine:d.string({description:"CLI that runs the second reviewer. `auto` prefers your installed codex for a cross-vendor reviewer \u2014 two Claude models share a training lineage, so their agreement overstates independence.",options:["auto","claude","codex"],default:"auto"}),limit:d.integer({description:"Sample only the N most-recently-annotated runs (cost control / smoke)"}),parallel:d.integer({description:"Concurrent re-grades",default:3}),timeout:d.integer({description:"Timeout per re-grade in ms",default:12e4}),json:d.boolean({description:"Emit the full report as JSON"})};async run(){let{args:z,flags:a}=await this.parse(h);this.requireAuth();let f=z.targetId,b=await c(`/api/eval-targets/${f}`),v=(await c(`/api/eval-targets/${f}/annotations?limit=500`)).filter(e=>e.judgeMethod==="llm");v.length===0&&this.error(`No human labels on "${b.name}" yet \u2014 label some runs in the Review tab first; this command scores a second model against those labels.`);let C=new Map;for(let e=0;;e+=I){let o=await c(`/api/eval-targets/${f}/runs?limit=${I}&offset=${e}`);for(let g of o)C.set(g.id,g);if(o.length<I)break}let D=await c(`/api/eval-targets/${f}/evals`),V=new Map(D.map(e=>[e.id,e.prompt])),s=new Map;for(let e of v){let o=s.get(e.evalRunId)??[];o.push(e),s.set(e.evalRunId,o)}let r=[...new Set(v.map(e=>e.judgeModel).filter(e=>e!=null))],n;try{n=H({engineFlag:a.engine,modelFlag:a.model,availability:_(),judgeModels:r})}catch(e){this.error(e instanceof Error?e.message:String(e))}if(this.log(`Reviewer: ${n.rationale}`),n.engine==="codex"&&n.model){let e=await L({prompt:"Reply with exactly: OK",model:n.model,timeout:6e4});e.error&&F(e.error)?(n.model===m&&this.error(`Codex fallback model ${m} is also unrunnable: ${e.error}`),this.warn(`Your codex model "${n.model}" can't run on the installed CLI/account \u2014 falling back to ${m}. (${e.error.slice(0,160)})`),n={...n,model:m}):e.error&&this.error(`Codex preflight failed: ${e.error}`)}this.log(`Re-grading ${s.size} annotated run(s) / ${v.length} assertion(s) on "${b.name}"`+(r.length?` (original judge: ${r.join(", ")})`:"")+".");let M=[],x=0,w=new Set,l=a.limit!=null?[...s.entries()].slice(0,Math.max(1,a.limit)):[...s.entries()];l.length<s.size&&this.log(`Sampling ${l.length} of ${s.size} annotated runs (--limit).`);let q=0,G=j(async()=>{for(;;){let e=q++;if(e>=l.length)return;let[o,g]=l[e],u=C.get(o),y=u?.grading?.expectations,k=u?.outputs?.transcript,E=u?V.get(u.evalId):void 0;if(!u||!y?.length||!k||!E){x+=g.length,this.warn(`run ${o.slice(0,8)}: missing output/expectations/prompt \u2014 skipped`);continue}let $=await T({prompt:E,output:k,expectations:y.map(i=>i.text),model:n.model??void 0,timeout:a.timeout,engine:n.engine});$.judge_model&&w.add($.judge_model);for(let i of g){let S=y[i.expectationIndex],R=$.expectations[i.expectationIndex];if(!S||!R||S.text!==i.expectationText){x+=1;continue}if($.judge_model==null&&R.method==="llm"){x+=1;continue}M.push({runId:o,expectationIndex:i.expectationIndex,text:i.expectationText,humanPassed:N(i.verdict,i.judgePassed),judgeAPassed:i.judgePassed,judgeBPassed:R.passed})}this.log(`  ${e+1}/${l.length} re-graded`)}},"worker");await Promise.all(Array.from({length:Math.max(1,Math.min(a.parallel,l.length))},G));let t=O(M,x),P=[...w].join(", ")||n.model||`${n.engine} (CLI default)`;if(r.length===1&&w.size===1&&r[0]===[...w][0]&&this.warn(`Second reviewer resolved to the SAME model as the original judge (${r[0]}) \u2014 this measures self-consistency, not independent review.`),a.json){this.log(JSON.stringify({target:b.name,judgeA:r,judgeB:P,...t},null,2));return}if(this.log(""),this.log(`Second reviewer: ${P} \u2014 scored against ${t.pairs} human-labeled assertion(s)`+(t.skipped?` (${t.skipped} skipped)`:"")),this.log(""),this.log(`  reviewer vs HUMAN:   ${p(t.bVsHuman.overall.rate)} overall \xB7 ${p(t.bVsHuman.onJudgePass.rate)} where judge passed \xB7 ${p(t.bVsHuman.onJudgeFail.rate)} where judge failed`),this.log(`  original judge vs human: ${p(t.aVsHuman.overall.rate)} (your approve rate)`),this.log(`  reviewer vs judge:   ${p(t.bVsA.overall.rate)} (inter-model, not truth)`),this.log(`  Cohen's \u03BA (reviewer vs human): ${t.kappaBvsHuman==null?"undefined (degenerate marginals)":t.kappaBvsHuman.toFixed(2)}`),this.log(""),t.suggestedAnchorWeight!=null&&this.log(`  suggested \xA76g anchor weight if used as a passive label source: ${t.suggestedAnchorWeight.toFixed(2)} (misclassification attenuation (2\xB7acc\u22121)\xB2; validated:false either way)`),t.disagreements.length>0){this.log(""),this.log(`  ${t.disagreements.length} disagreement(s) with your labels \u2014 worth reading:`);for(let e of t.disagreements.slice(0,10))this.log(`    run ${e.runId.slice(0,8)} #${e.expectationIndex}: human=${e.humanPassed?"pass":"fail"} reviewer=${e.judgeBPassed?"pass":"fail"} \u2014 ${e.text.slice(0,80)}`);t.disagreements.length>10&&this.log(`    \u2026 and ${t.disagreements.length-10} more (use --json for all)`)}}};export{A as default};

package/dist/commands/hooks/install.js

@@ -1 +1 @@
-import{a as pt}from"../../chunk-6XEC6474.js";import{B as ut,E as dt,H as gt,I as ht,e as z,f as Q,g as Z,i as tt,j as et,k as st,n as ot,p as it,q as rt,r as at,u as nt,v as lt,y as ct}from"../../chunk-2JRDG4LJ.js";import{a as q,b as W,c as m,d as J,g as A}from"../../chunk-I2DZGJRR.js";import"../../chunk-N7ODX6PO.js";import{a as f}from"../../chunk-67QCYUJU.js";import{c as s,g as X,h as D,i as F,k as O,l as V}from"../../chunk-MVM6OGAF.js";import{e as B}from"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a as Y}from"../../chunk-XICLXXRH.js";import{a as M}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import{c as N}from"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import{b as L}from"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import{a as G}from"../../chunk-O6C2UGBY.js";import{a as v}from"../../chunk-ZLLER3HP.js";import{Flags as w}from"@oclif/core";import{execFileSync as ft,spawnSync as l}from"node:child_process";import{existsSync as c}from"node:fs";import{homedir as mt}from"node:os";import{join as _,dirname as Ct}from"node:path";function vt(){try{let d=ft("/bin/sh",["-c","command -v rulemetric"],{encoding:"utf-8"}).trim();if(d&&c(d))return Ct(d)}catch{}}v(vt,"resolveHookBinDir");function wt(){try{return ft("launchctl",["list","com.rulemetric.gateway"],{stdio:["ignore","pipe","pipe"]}),!0}catch{return!1}}v(wt,"isGatewayLaunchdManaged");var o=_(mt(),".mitmproxy","mitmproxy-ca-cert.pem"),R=class d extends M{static{v(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":w.boolean({default:!1,description:"Skip proxy/gateway configuration (hooks only)"}),"project-dir":w.string({description:"Project directory to install hooks in (defaults to cwd)"}),global:w.boolean({char:"g",default:!1,description:"Install hooks globally (~/.claude/settings.json) so all projects are tracked"}),"statusline-usage":w.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(d),r=i.global,n=r?mt():i["project-dir"]??process.cwd(),x=_(n,".claude"),y=st(x,r),K=r?"settings.json":"settings.local.json",k=i["no-proxy"],P=!1,g=k?"no_proxy_flag":void 0,h=!1,S=[];r&&this.log(`Installing hooks globally \u2192 ${y}`);let{settings:a,recoveredBackup:$}=z(y);if($&&this.log(`[!!] ${y} was not valid JSON \u2014 backed it up to ${$} and continued with a fresh file.`),!r){let e=ot(_(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 ${e.backupPath}`),e.removedStatusline&&c(q)&&!A(a)&&(a.statusLine=m))}let I=G();(I.accessToken||process.env.RULEMETRIC_ACCESS_TOKEN)&&!I.apiKey&&!process.env.RULEMETRIC_API_KEY&&await Y(this.log.bind(this),this.warn.bind(this),this.config.version);let yt=tt(a),C=vt();if(Z(a,C),this.log(yt?`[OK] Refreshed RuleMetric Claude Code session hooks \u2192 .claude/${K}`:`[OK] Installed RuleMetric Claude Code session hooks \u2192 .claude/${K}`),!k){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"))),c(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"}),c(o)?this.log(`[OK] CA certificate generated at ${o}`):this.log("[!!] Could not generate CA cert"))}let u=c(o);u&&!this.isCACertTrusted()?(this.log("[..] Installing CA certificate into system trust store (requires admin)..."),h=this.trustCACert(),h?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`))):u&&(this.log("[OK] CA certificate already trusted"),h=!0);let p=wt(),b=await X(s);if(p)this.log("[OK] Gateway managed by launchd (com.rulemetric.gateway)");else if(b.status==="ours-live")this.log("[OK] Gateway already running");else if(b.status==="foreign"&&F()){this.log("[..] Restarting gateway (running build predates the identity handshake)..."),V(),await new Promise(t=>setTimeout(t,500));try{let t=O(this.config.version);this.log(`[OK] Gateway restarted (PID ${t})`)}catch(t){this.log(`[!!] Could not restart gateway: ${t.message}`)}}else if(b.status!=="foreign")try{let t=O(this.config.version);this.log(`[OK] Gateway started (PID ${t})`)}catch(t){this.log(`[!!] Could not start gateway: ${t.message}`)}let j=await D(s,{timeoutMs:4e3});if(!u)g=e?"ca_not_generated":S.length===0?"no_package_manager":"mitmproxy_install_failed",this.log("[!!] No CA certificate \u2014 NOT setting HTTPS_PROXY (capture stays hooks-only)."),this.log("     Install mitmproxy, then re-run: rulemetric hooks install");else if(j.status==="foreign"&&!p){g="gateway_port_foreign";let t=B(s);this.log(`[!!] Port :${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(j.status==="dead")g="gateway_not_live",this.log(`[!!] Gateway is not answering on :${s} \u2014 NOT setting HTTPS_PROXY.`),this.log("     Capture still works via hooks. To enable deep (proxy) capture later:"),this.log("     rulemetric proxy start  &&  rulemetric hooks install");else{let t=a.env??{};t.HTTPS_PROXY=`http://localhost:${s}`,t.NO_PROXY=et,c(o)&&(t.NODE_EXTRA_CA_CERTS=o),a.env=t,P=!0,this.log(`[OK] HTTPS_PROXY set to gateway on :${s} (verified live)`),nt(n,s,o)&&this.log("[OK] Cursor proxy configured \u2192 .cursor/settings.json"),lt(n,s,o)&&this.log("[OK] VS Code workspace proxy configured \u2192 .vscode/settings.json"),ut(s)&&this.log("[OK] VS Code user proxy configured \u2192 ~/Library/Application Support/Code/User/settings.json"),dt(s)&&this.log("[OK] Cursor user proxy configured \u2192 ~/Library/Application Support/Cursor/User/settings.json"),h?ct(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."))}}if(!r){let e=it(n,C);e&&this.log(`[OK] Cursor hooks configured \u2192 ${e}`);let u=at(n,C);u&&this.log(`[OK] Copilot Agent hooks configured \u2192 ${u}`);let p=gt(n);p&&this.log(`[OK] Antigravity workspace hooks configured \u2192 ${p}`)}let E=ht();E&&this.log(`[OK] Antigravity user hooks configured \u2192 ${E}`);let H=rt(C);if(H&&this.log(`[OK] Cursor user hooks configured \u2192 ${H}`),i["statusline-usage"])if(A(a))this.log("[OK] Statusline shim already installed");else{let e=J(a);W(e),a.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:U}=Q(y,a);U&&this.log(`[OK] Backed up previous settings \u2192 ${U}`),this.log(""),this.log("Setup complete! Next steps:"),r?(this.log("  1. Install always-on API:    rulemetric service install"),this.log("  2. Start Claude Code in any project \u2014 all sessions tracked automatically"),this.log(""),this.log("The API service starts on login and restarts on crash.")):k?(this.log("  1. Start Claude Code \u2014 sessions are captured by the hooks just installed"),this.log(""),this.log("No proxy means no mitmproxy / CA cert needed. The hooks capture full"),this.log("sessions on their own; re-run without --no-proxy to also capture system"),this.log("prompts for instruction-effectiveness linking.")):(this.log("  1. Start Claude Code \u2014 sessions are captured by hooks immediately"),P?this.log("  2. Full-depth capture is ON \u2014 the proxy is live on :"+s+"."):this.log("  2. Enable full depth (proxy not yet live):  rulemetric proxy start"),this.log(""),this.log("Hooks capture every session (lifecycle + transcript). The proxy adds the"),this.log("rendered system prompt (instruction-effectiveness linking) + cross-tool capture."),this.log(""),this.log("To capture traffic from other clients (Python, curl, etc.):"),this.log('  eval "$(rulemetric proxy env --all)"')),await L(),N("hooks_installed",{proxy:P,caTrusted:h,reason:g,installers:S,global:r},this.config.version);let T=await pt();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 n=f(i);if(n&&(this.log(`     Installing via ${i}...`),l(n,[...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"?c("/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{R as default};
+import{a as pt}from"../../chunk-6XEC6474.js";import{B as ut,E as dt,H as gt,I as ht,e as z,f as Q,g as Z,i as tt,j as et,k as st,n as ot,p as it,q as rt,r as at,u as nt,v as lt,y as ct}from"../../chunk-2JRDG4LJ.js";import{a as q,b as W,c as m,d as J,g as A}from"../../chunk-I2DZGJRR.js";import{c as s,g as X,h as D,i as F,k as O,l as V}from"../../chunk-MVM6OGAF.js";import{e as B}from"../../chunk-KKDHD4XA.js";import"../../chunk-SZ62CP55.js";import{a as f}from"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import{a as Y}from"../../chunk-XICLXXRH.js";import{a as M}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import{c as N}from"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import{b as L}from"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import{a as G}from"../../chunk-O6C2UGBY.js";import{a as v}from"../../chunk-ZLLER3HP.js";import{Flags as w}from"@oclif/core";import{execFileSync as ft,spawnSync as l}from"node:child_process";import{existsSync as c}from"node:fs";import{homedir as mt}from"node:os";import{join as _,dirname as Ct}from"node:path";function vt(){try{let d=ft("/bin/sh",["-c","command -v rulemetric"],{encoding:"utf-8"}).trim();if(d&&c(d))return Ct(d)}catch{}}v(vt,"resolveHookBinDir");function wt(){try{return ft("launchctl",["list","com.rulemetric.gateway"],{stdio:["ignore","pipe","pipe"]}),!0}catch{return!1}}v(wt,"isGatewayLaunchdManaged");var o=_(mt(),".mitmproxy","mitmproxy-ca-cert.pem"),R=class d extends M{static{v(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":w.boolean({default:!1,description:"Skip proxy/gateway configuration (hooks only)"}),"project-dir":w.string({description:"Project directory to install hooks in (defaults to cwd)"}),global:w.boolean({char:"g",default:!1,description:"Install hooks globally (~/.claude/settings.json) so all projects are tracked"}),"statusline-usage":w.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(d),r=i.global,n=r?mt():i["project-dir"]??process.cwd(),x=_(n,".claude"),y=st(x,r),K=r?"settings.json":"settings.local.json",k=i["no-proxy"],P=!1,g=k?"no_proxy_flag":void 0,h=!1,S=[];r&&this.log(`Installing hooks globally \u2192 ${y}`);let{settings:a,recoveredBackup:$}=z(y);if($&&this.log(`[!!] ${y} was not valid JSON \u2014 backed it up to ${$} and continued with a fresh file.`),!r){let e=ot(_(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 ${e.backupPath}`),e.removedStatusline&&c(q)&&!A(a)&&(a.statusLine=m))}let I=G();(I.accessToken||process.env.RULEMETRIC_ACCESS_TOKEN)&&!I.apiKey&&!process.env.RULEMETRIC_API_KEY&&await Y(this.log.bind(this),this.warn.bind(this),this.config.version);let yt=tt(a),C=vt();if(Z(a,C),this.log(yt?`[OK] Refreshed RuleMetric Claude Code session hooks \u2192 .claude/${K}`:`[OK] Installed RuleMetric Claude Code session hooks \u2192 .claude/${K}`),!k){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"))),c(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"}),c(o)?this.log(`[OK] CA certificate generated at ${o}`):this.log("[!!] Could not generate CA cert"))}let u=c(o);u&&!this.isCACertTrusted()?(this.log("[..] Installing CA certificate into system trust store (requires admin)..."),h=this.trustCACert(),h?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`))):u&&(this.log("[OK] CA certificate already trusted"),h=!0);let p=wt(),b=await X(s);if(p)this.log("[OK] Gateway managed by launchd (com.rulemetric.gateway)");else if(b.status==="ours-live")this.log("[OK] Gateway already running");else if(b.status==="foreign"&&F()){this.log("[..] Restarting gateway (running build predates the identity handshake)..."),V(),await new Promise(t=>setTimeout(t,500));try{let t=O(this.config.version);this.log(`[OK] Gateway restarted (PID ${t})`)}catch(t){this.log(`[!!] Could not restart gateway: ${t.message}`)}}else if(b.status!=="foreign")try{let t=O(this.config.version);this.log(`[OK] Gateway started (PID ${t})`)}catch(t){this.log(`[!!] Could not start gateway: ${t.message}`)}let j=await D(s,{timeoutMs:4e3});if(!u)g=e?"ca_not_generated":S.length===0?"no_package_manager":"mitmproxy_install_failed",this.log("[!!] No CA certificate \u2014 NOT setting HTTPS_PROXY (capture stays hooks-only)."),this.log("     Install mitmproxy, then re-run: rulemetric hooks install");else if(j.status==="foreign"&&!p){g="gateway_port_foreign";let t=B(s);this.log(`[!!] Port :${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(j.status==="dead")g="gateway_not_live",this.log(`[!!] Gateway is not answering on :${s} \u2014 NOT setting HTTPS_PROXY.`),this.log("     Capture still works via hooks. To enable deep (proxy) capture later:"),this.log("     rulemetric proxy start  &&  rulemetric hooks install");else{let t=a.env??{};t.HTTPS_PROXY=`http://localhost:${s}`,t.NO_PROXY=et,c(o)&&(t.NODE_EXTRA_CA_CERTS=o),a.env=t,P=!0,this.log(`[OK] HTTPS_PROXY set to gateway on :${s} (verified live)`),nt(n,s,o)&&this.log("[OK] Cursor proxy configured \u2192 .cursor/settings.json"),lt(n,s,o)&&this.log("[OK] VS Code workspace proxy configured \u2192 .vscode/settings.json"),ut(s)&&this.log("[OK] VS Code user proxy configured \u2192 ~/Library/Application Support/Code/User/settings.json"),dt(s)&&this.log("[OK] Cursor user proxy configured \u2192 ~/Library/Application Support/Cursor/User/settings.json"),h?ct(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."))}}if(!r){let e=it(n,C);e&&this.log(`[OK] Cursor hooks configured \u2192 ${e}`);let u=at(n,C);u&&this.log(`[OK] Copilot Agent hooks configured \u2192 ${u}`);let p=gt(n);p&&this.log(`[OK] Antigravity workspace hooks configured \u2192 ${p}`)}let E=ht();E&&this.log(`[OK] Antigravity user hooks configured \u2192 ${E}`);let H=rt(C);if(H&&this.log(`[OK] Cursor user hooks configured \u2192 ${H}`),i["statusline-usage"])if(A(a))this.log("[OK] Statusline shim already installed");else{let e=J(a);W(e),a.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:U}=Q(y,a);U&&this.log(`[OK] Backed up previous settings \u2192 ${U}`),this.log(""),this.log("Setup complete! Next steps:"),r?(this.log("  1. Install always-on API:    rulemetric service install"),this.log("  2. Start Claude Code in any project \u2014 all sessions tracked automatically"),this.log(""),this.log("The API service starts on login and restarts on crash.")):k?(this.log("  1. Start Claude Code \u2014 sessions are captured by the hooks just installed"),this.log(""),this.log("No proxy means no mitmproxy / CA cert needed. The hooks capture full"),this.log("sessions on their own; re-run without --no-proxy to also capture system"),this.log("prompts for instruction-effectiveness linking.")):(this.log("  1. Start Claude Code \u2014 sessions are captured by hooks immediately"),P?this.log("  2. Full-depth capture is ON \u2014 the proxy is live on :"+s+"."):this.log("  2. Enable full depth (proxy not yet live):  rulemetric proxy start"),this.log(""),this.log("Hooks capture every session (lifecycle + transcript). The proxy adds the"),this.log("rendered system prompt (instruction-effectiveness linking) + cross-tool capture."),this.log(""),this.log("To capture traffic from other clients (Python, curl, etc.):"),this.log('  eval "$(rulemetric proxy env --all)"')),await L(),N("hooks_installed",{proxy:P,caTrusted:h,reason:g,installers:S,global:r},this.config.version);let T=await pt();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 n=f(i);if(n&&(this.log(`     Installing via ${i}...`),l(n,[...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"?c("/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{R as default};

package/dist/commands/proxy/setup.js

@@ -1,4 +1,4 @@
-import{a as g}from"../../chunk-67QCYUJU.js";import{a as S}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{g as x}from"../../chunk-O6C2UGBY.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Flags as _}from"@oclif/core";import{execSync as P,spawnSync as m}from"node:child_process";import{existsSync as f,readFileSync as y,appendFileSync as T,writeFileSync as k,mkdirSync as v}from"node:fs";import{homedir as d}from"node:os";import{dirname as E,join as u}from"node:path";var c=u(d(),".mitmproxy","mitmproxy-ca-cert.pem"),$=u(d(),".config","rulemetric"),A=8787,R=class C extends S{static{h(this,"ProxySetup")}static description="Set up context capture proxy (installs mitmproxy, configures shell)";static examples=["<%= config.bin %> proxy setup","<%= config.bin %> proxy setup --shell zsh","<%= config.bin %> proxy setup --check"];static flags={shell:_.string({description:"Shell to configure (auto-detected if omitted)",options:["zsh","bash","fish","powershell"]}),port:_.integer({char:"p",default:A,description:"Proxy listen port"}),check:_.boolean({description:"Only check prerequisites, do not install anything",default:!1})};async run(){let{flags:e}=await this.parse(C),i=e.check,t=e.port;this.log("RuleMetric Proxy Setup"),this.log(`=====================
+import{a as g}from"../../chunk-PJ3WHQ67.js";import{a as S}from"../../chunk-PPKPCXJD.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-V44SRRYR.js";import"../../chunk-U6XEU573.js";import"../../chunk-VLOX6KOM.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-H2ORWIEY.js";import{g as x}from"../../chunk-O6C2UGBY.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Flags as _}from"@oclif/core";import{execSync as P,spawnSync as m}from"node:child_process";import{existsSync as f,readFileSync as y,appendFileSync as T,writeFileSync as k,mkdirSync as v}from"node:fs";import{homedir as d}from"node:os";import{dirname as E,join as u}from"node:path";var c=u(d(),".mitmproxy","mitmproxy-ca-cert.pem"),$=u(d(),".config","rulemetric"),A=8787,R=class C extends S{static{h(this,"ProxySetup")}static description="Set up context capture proxy (installs mitmproxy, configures shell)";static examples=["<%= config.bin %> proxy setup","<%= config.bin %> proxy setup --shell zsh","<%= config.bin %> proxy setup --check"];static flags={shell:_.string({description:"Shell to configure (auto-detected if omitted)",options:["zsh","bash","fish","powershell"]}),port:_.integer({char:"p",default:A,description:"Proxy listen port"}),check:_.boolean({description:"Only check prerequisites, do not install anything",default:!1})};async run(){let{flags:e}=await this.parse(C),i=e.check,t=e.port;this.log("RuleMetric Proxy Setup"),this.log(`=====================
 `);let l=[await this.checkAuth(),await this.checkMitmproxy(),await this.checkCACert(),await this.checkHooks(),this.checkShellEnv(t)],r=!0;for(let n of l){let p=n.ok?"[OK]":"[!!]";this.log(`${p} ${n.name}: ${n.message}`),n.ok||(r=!1)}if(r){this.log("\nAll checks passed. Run `rulemetric proxy start` to begin capturing.");return}if(i){this.log("\nSome checks failed. Run `rulemetric proxy setup` (without --check) to fix.");return}this.log(`
 Fixing issues...
 `);for(let n of l){if(n.ok||!n.fix)continue;this.log(`Fixing: ${n.name}...`),await n.fix()?this.log(`  Fixed.

package/dist/commands/proxy/start.js

@@ -1,4 +1,4 @@
-import{a as K}from"../../chunk-OMXUK4M6.js";import{a as U}from"../../chunk-67QCYUJU.js";import{b as F,d as N}from"../../chunk-SZ62CP55.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"../../chunk-H2ORWIEY.js";import{g as $}from"../../chunk-O6C2UGBY.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Flags as _}from"@oclif/core";import{execSync as H,spawn as q,spawnSync as m}from"node:child_process";import{appendFileSync as G,existsSync as l,mkdirSync as P,openSync as V,readFileSync as b,writeFileSync as Y}from"node:fs";import{createRequire as W}from"node:module";import{homedir as x}from"node:os";import{dirname as A,join as c,resolve as f}from"node:path";import{fileURLToPath as X}from"node:url";var g=A(X(import.meta.url)),J=8788,S=c(x(),".config","rulemetric"),T=c(S,"proxy.pid"),D=c(S,"proxy.log"),w="rulemetric/proxy",v="rulemetric-proxy";function j(s,o){try{let d=W(import.meta.url).resolve(`@rulemetric/proxy/${s}`);if(l(d))return d}catch{}for(let t of o)if(l(t))return t;return null}u(j,"resolveProxyFile");function B(){let s=[f(g,"../../../../../packages/proxy/addon/rulemetric_addon.py"),f(g,"../../../node_modules/@rulemetric/proxy/addon/rulemetric_addon.py"),f(g,"../../../../../node_modules/@rulemetric/proxy/addon/rulemetric_addon.py")],o=j("addon/rulemetric_addon.py",s);if(o)return o;throw new Error(`Could not find rulemetric_addon.py. Is @rulemetric/proxy installed?
+import{a as K}from"../../chunk-OMXUK4M6.js";import{b as F,d as N}from"../../chunk-SZ62CP55.js";import{a as U}from"../../chunk-PJ3WHQ67.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"../../chunk-H2ORWIEY.js";import{g as $}from"../../chunk-O6C2UGBY.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Flags as _}from"@oclif/core";import{execSync as H,spawn as q,spawnSync as m}from"node:child_process";import{appendFileSync as G,existsSync as l,mkdirSync as P,openSync as V,readFileSync as b,writeFileSync as Y}from"node:fs";import{createRequire as W}from"node:module";import{homedir as x}from"node:os";import{dirname as A,join as c,resolve as f}from"node:path";import{fileURLToPath as X}from"node:url";var g=A(X(import.meta.url)),J=8788,S=c(x(),".config","rulemetric"),T=c(S,"proxy.pid"),D=c(S,"proxy.log"),w="rulemetric/proxy",v="rulemetric-proxy";function j(s,o){try{let d=W(import.meta.url).resolve(`@rulemetric/proxy/${s}`);if(l(d))return d}catch{}for(let t of o)if(l(t))return t;return null}u(j,"resolveProxyFile");function B(){let s=[f(g,"../../../../../packages/proxy/addon/rulemetric_addon.py"),f(g,"../../../node_modules/@rulemetric/proxy/addon/rulemetric_addon.py"),f(g,"../../../../../node_modules/@rulemetric/proxy/addon/rulemetric_addon.py")],o=j("addon/rulemetric_addon.py",s);if(o)return o;throw new Error(`Could not find rulemetric_addon.py. Is @rulemetric/proxy installed?
 Searched (after module resolution):
 `+s.map(t=>`  ${t}`).join(`
 `))}u(B,"findAddonPath");function z(){let s=[f(g,"../../../../../packages/proxy/Dockerfile"),f(g,"../../../node_modules/@rulemetric/proxy/Dockerfile"),f(g,"../../../../../node_modules/@rulemetric/proxy/Dockerfile")],o=j("Dockerfile",s);if(o)return A(o);throw new Error("Could not find proxy Dockerfile. Is @rulemetric/proxy installed?")}u(z,"findDockerfilePath");function Q(){return m("docker",["image","inspect",w],{stdio:"ignore"}).status===0}u(Q,"dockerImageExists");function Z(){let s=m("docker",["container","inspect","-f","{{.State.Running}}",v],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return s.status===0&&s.stdout?.trim()==="true"}u(Z,"dockerContainerRunning");var L=class s extends M{static{u(this,"ProxyStart")}static description="Start context capture proxy";static examples=["<%= config.bin %> proxy start","<%= config.bin %> proxy start --docker","<%= config.bin %> proxy start --port 9090","<%= config.bin %> proxy start --web"];static flags={port:_.integer({char:"p",default:J,description:"Proxy listen port"}),docker:_.boolean({char:"d",default:!1,description:"Run proxy in Docker (no local mitmproxy needed)"}),web:_.boolean({default:!1,description:"Launch mitmweb debug UI instead of mitmdump (local mode only)"}),"skip-checks":_.boolean({default:!1,description:"Skip prerequisite checks"}),build:_.boolean({default:!1,description:"Force rebuild Docker image"})};async run(){let{flags:o}=await this.parse(s),t=o.port;o.docker?await this.startDocker(t,o.build):await this.startLocal(t,o.web,o["skip-checks"])}async startDocker(o,t){try{H("docker info",{stdio:"ignore"})}catch{this.error("Docker is not running. Start Docker Desktop and try again.")}Z()&&this.error("Proxy container already running. Stop it first: rulemetric proxy stop");let d=this.findComposeFile(),k=c(x(),".mitmproxy");P(k,{recursive:!0});let E=process.env.TMPDIR||"/tmp";P(c(E,"rulemetric"),{recursive:!0});let y={PROXY_PORT:String(o),RULEMETRIC_PROJECT_PATH:process.cwd()};if(!process.env.RULEMETRIC_API_KEY&&!process.env.RULEMETRIC_ACCESS_TOKEN){let n=c(x(),".config","rulemetric","env");if(l(n))for(let h of b(n,"utf-8").split(`

package/dist/dashboard/Dashboard.js

@@ -1 +1 @@
-import{a as F}from"../chunk-DVN7GFJG.js";import{a as M}from"../chunk-O2SDD2WW.js";import{a as E}from"../chunk-GBMF3TP6.js";import{a as O}from"../chunk-WSCK4WNK.js";import{a as D}from"../chunk-4Z5ECCJE.js";import"../chunk-ARNZQBMX.js";import{a as B,b as S,c as R}from"../chunk-G6XGDHGC.js";import{a as $}from"../chunk-SJGLRC2R.js";import"../chunk-RDET7CKR.js";import"../chunk-NBTC5FNV.js";import"../chunk-TU2AR6XO.js";import{a as u,b as T,d as a,e as v}from"../chunk-NJ5UED3H.js";import"../chunk-ZR3MUNLS.js";import"../chunk-ZXF5NE54.js";import"../chunk-7DFNBRLY.js";import{a as X}from"../chunk-3AV3GTMP.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import{a as m}from"../chunk-ZLLER3HP.js";import{useState as w,useEffect as nt,useCallback as G,useRef as st}from"react";import{render as U,Box as c,Text as o,useInput as at,useApp as lt}from"ink";import{Fragment as ht,jsx as t,jsxs as h}from"react/jsx-runtime";var dt=90;function J(){let s=process.stdout.columns||80,n=Math.min(160,s),i=n>=dt,d=i?Math.floor(n/2):n,b=Math.max(6,Math.min(10,d-34));return{dashWidth:n,wide:i,halfWidth:d,barWidth:b}}m(J,"getLayout");function z({active:s,providerName:n}){return h(c,{justifyContent:"space-between",paddingX:1,children:[t(c,{gap:1,children:u.map(i=>t(o,{bold:s===i,color:s===i?a:v,children:s===i?`[ ${T[i]} ]`:`  ${T[i]}  `},i))}),n&&t(o,{color:a,bold:!0,children:n})]})}m(z,"PeriodTabs");function H({width:s,insightsFocused:n=!1}){return t(c,{borderStyle:"round",borderColor:v,width:s,justifyContent:"center",paddingX:1,children:h(o,{children:[t(o,{color:a,bold:!0,children:"<"}),t(o,{color:a,children:">"}),t(o,{dimColor:!0,children:" switch   "}),t(o,{color:a,bold:!0,children:"q"}),t(o,{dimColor:!0,children:" quit   "}),t(o,{color:a,bold:!0,children:"s"}),h(o,{dimColor:!0,children:[" ",n?"back":"suggestions","   "]}),t(o,{color:a,bold:!0,children:"1"}),t(o,{dimColor:!0,children:"-"}),t(o,{color:a,bold:!0,children:"5"}),t(o,{dimColor:!0,children:" period   "}),t(o,{color:a,bold:!0,children:"r"}),t(o,{dimColor:!0,children:" refresh   "}),t(o,{color:a,bold:!0,children:"n"}),t(o,{dimColor:!0,children:" new session   "}),t(o,{color:a,bold:!0,children:"l"}),t(o,{dimColor:!0,children:" launcher"})]})})}m(H,"StatusBar");function ct({wide:s,width:n,children:i}){return s?t(c,{width:n,children:i}):t(ht,{children:i})}m(ct,"Row");function ut({initialLocal:s,initialApi:n,initialInstructions:i,initialPeriod:d,initialLauncherFocused:b=!1}){let{exit:j}=lt(),[e,L]=w(d),[K,Q]=w(s),[W,V]=w(n),[Z,tt]=w(i),[ot,k]=w(!1),[p,I]=w(!1),[Y,_]=w(b),et=X()?.orgSlug??null,y=st(null),it=J(),g=G(async()=>{k(!0);let[r,x,P]=await Promise.all([B(),S(),R()]);Q(r),V(x),tt(P),k(!1)},[]);nt(()=>{let r=setInterval(g,3e4);return()=>clearInterval(r)},[g]);let f=G(r=>{r!==e&&(L(r),y.current&&clearTimeout(y.current),y.current=setTimeout(g,300))},[e,g]);at((r,x)=>{if(Y)return;if(r==="q"){if(p){I(!1);return}j();return}if(r==="s"){I(A=>!A);return}if(x.escape&&p){I(!1);return}if(r==="l"&&!p){_(!0);return}if(p)return;if(r==="r"){g();return}if(r==="n"){import("../dist-FAVALJ5C.js").then(({getLauncher:A})=>{A("claude_code").launch({projectPath:process.cwd(),mode:"interactive"})});return}let P=u.indexOf(e);x.leftArrow?f(u[(P-1+u.length)%u.length]):x.rightArrow||x.tab?f(u[(P+1)%u.length]):r==="1"?f("today"):r==="2"?f("week"):r==="3"?f("30days"):r==="4"?f("month"):r==="5"&&f("all")});let{dashWidth:l,wide:q,halfWidth:rt,barWidth:C}=it,N=q?rt:l;return ot?h(c,{flexDirection:"column",width:l,children:[t(z,{active:e}),t(c,{borderStyle:"round",borderColor:a,paddingX:1,width:l,children:h(o,{dimColor:!0,children:["Loading ",T[e],"..."]})}),t(H,{width:l})]}):Y?t(c,{flexDirection:"column",width:l,children:t($,{onClose:()=>_(!1),activeOrg:et,defaultProjectPath:process.cwd(),width:l})}):h(c,{flexDirection:"column",width:l,children:[t(z,{active:e}),t(D,{local:K,api:W,label:T[e],width:l}),h(ct,{wide:q,width:l,children:[t(O,{api:W,width:N,barWidth:C}),t(F,{api:W,width:N,barWidth:C})]}),t(E,{instructions:Z,width:l,barWidth:C}),t(M,{api:W,width:l,focused:p}),t(H,{width:l,insightsFocused:p})]})}m(ut,"InteractiveDashboard");async function yt(s="week",n={}){let[i,d,b]=await Promise.all([B(),S(),R()]);if(process.stdin.isTTY&&process.stdout.isTTY){let{waitUntilExit:e}=U(t(ut,{initialLocal:i,initialApi:d,initialInstructions:b,initialPeriod:s,initialLauncherFocused:n.openLauncher??!1}));await e()}else{let e=J(),{unmount:L}=U(h(c,{flexDirection:"column",width:e.dashWidth,children:[t(D,{local:i,api:d,label:T[s],width:e.dashWidth}),t(O,{api:d,width:e.dashWidth,barWidth:e.barWidth}),t(F,{api:d,width:e.dashWidth,barWidth:e.barWidth}),t(E,{instructions:b,width:e.dashWidth,barWidth:e.barWidth}),t(M,{api:d,width:e.dashWidth})]}),{patchConsole:!1});L()}}m(yt,"renderDashboard");export{yt as renderDashboard};
+import{a as F}from"../chunk-DVN7GFJG.js";import{a as M}from"../chunk-O2SDD2WW.js";import{a as E}from"../chunk-GBMF3TP6.js";import{a as O}from"../chunk-WSCK4WNK.js";import{a as D}from"../chunk-4Z5ECCJE.js";import"../chunk-ARNZQBMX.js";import{a as B,b as S,c as R}from"../chunk-G6XGDHGC.js";import{a as $}from"../chunk-BMGQFKG6.js";import"../chunk-RDET7CKR.js";import"../chunk-X23L2MF6.js";import"../chunk-B22HYCBV.js";import"../chunk-ZXF5NE54.js";import{a as u,b as T,d as a,e as v}from"../chunk-NJ5UED3H.js";import"../chunk-ZR3MUNLS.js";import"../chunk-7DFNBRLY.js";import{a as X}from"../chunk-3AV3GTMP.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import{a as m}from"../chunk-ZLLER3HP.js";import{useState as w,useEffect as nt,useCallback as G,useRef as st}from"react";import{render as U,Box as c,Text as o,useInput as at,useApp as lt}from"ink";import{Fragment as ht,jsx as t,jsxs as h}from"react/jsx-runtime";var dt=90;function J(){let s=process.stdout.columns||80,n=Math.min(160,s),i=n>=dt,d=i?Math.floor(n/2):n,b=Math.max(6,Math.min(10,d-34));return{dashWidth:n,wide:i,halfWidth:d,barWidth:b}}m(J,"getLayout");function z({active:s,providerName:n}){return h(c,{justifyContent:"space-between",paddingX:1,children:[t(c,{gap:1,children:u.map(i=>t(o,{bold:s===i,color:s===i?a:v,children:s===i?`[ ${T[i]} ]`:`  ${T[i]}  `},i))}),n&&t(o,{color:a,bold:!0,children:n})]})}m(z,"PeriodTabs");function H({width:s,insightsFocused:n=!1}){return t(c,{borderStyle:"round",borderColor:v,width:s,justifyContent:"center",paddingX:1,children:h(o,{children:[t(o,{color:a,bold:!0,children:"<"}),t(o,{color:a,children:">"}),t(o,{dimColor:!0,children:" switch   "}),t(o,{color:a,bold:!0,children:"q"}),t(o,{dimColor:!0,children:" quit   "}),t(o,{color:a,bold:!0,children:"s"}),h(o,{dimColor:!0,children:[" ",n?"back":"suggestions","   "]}),t(o,{color:a,bold:!0,children:"1"}),t(o,{dimColor:!0,children:"-"}),t(o,{color:a,bold:!0,children:"5"}),t(o,{dimColor:!0,children:" period   "}),t(o,{color:a,bold:!0,children:"r"}),t(o,{dimColor:!0,children:" refresh   "}),t(o,{color:a,bold:!0,children:"n"}),t(o,{dimColor:!0,children:" new session   "}),t(o,{color:a,bold:!0,children:"l"}),t(o,{dimColor:!0,children:" launcher"})]})})}m(H,"StatusBar");function ct({wide:s,width:n,children:i}){return s?t(c,{width:n,children:i}):t(ht,{children:i})}m(ct,"Row");function ut({initialLocal:s,initialApi:n,initialInstructions:i,initialPeriod:d,initialLauncherFocused:b=!1}){let{exit:j}=lt(),[e,L]=w(d),[K,Q]=w(s),[W,V]=w(n),[Z,tt]=w(i),[ot,k]=w(!1),[p,I]=w(!1),[Y,_]=w(b),et=X()?.orgSlug??null,y=st(null),it=J(),g=G(async()=>{k(!0);let[r,x,P]=await Promise.all([B(),S(),R()]);Q(r),V(x),tt(P),k(!1)},[]);nt(()=>{let r=setInterval(g,3e4);return()=>clearInterval(r)},[g]);let f=G(r=>{r!==e&&(L(r),y.current&&clearTimeout(y.current),y.current=setTimeout(g,300))},[e,g]);at((r,x)=>{if(Y)return;if(r==="q"){if(p){I(!1);return}j();return}if(r==="s"){I(A=>!A);return}if(x.escape&&p){I(!1);return}if(r==="l"&&!p){_(!0);return}if(p)return;if(r==="r"){g();return}if(r==="n"){import("../dist-FAVALJ5C.js").then(({getLauncher:A})=>{A("claude_code").launch({projectPath:process.cwd(),mode:"interactive"})});return}let P=u.indexOf(e);x.leftArrow?f(u[(P-1+u.length)%u.length]):x.rightArrow||x.tab?f(u[(P+1)%u.length]):r==="1"?f("today"):r==="2"?f("week"):r==="3"?f("30days"):r==="4"?f("month"):r==="5"&&f("all")});let{dashWidth:l,wide:q,halfWidth:rt,barWidth:C}=it,N=q?rt:l;return ot?h(c,{flexDirection:"column",width:l,children:[t(z,{active:e}),t(c,{borderStyle:"round",borderColor:a,paddingX:1,width:l,children:h(o,{dimColor:!0,children:["Loading ",T[e],"..."]})}),t(H,{width:l})]}):Y?t(c,{flexDirection:"column",width:l,children:t($,{onClose:()=>_(!1),activeOrg:et,defaultProjectPath:process.cwd(),width:l})}):h(c,{flexDirection:"column",width:l,children:[t(z,{active:e}),t(D,{local:K,api:W,label:T[e],width:l}),h(ct,{wide:q,width:l,children:[t(O,{api:W,width:N,barWidth:C}),t(F,{api:W,width:N,barWidth:C})]}),t(E,{instructions:Z,width:l,barWidth:C}),t(M,{api:W,width:l,focused:p}),t(H,{width:l,insightsFocused:p})]})}m(ut,"InteractiveDashboard");async function yt(s="week",n={}){let[i,d,b]=await Promise.all([B(),S(),R()]);if(process.stdin.isTTY&&process.stdout.isTTY){let{waitUntilExit:e}=U(t(ut,{initialLocal:i,initialApi:d,initialInstructions:b,initialPeriod:s,initialLauncherFocused:n.openLauncher??!1}));await e()}else{let e=J(),{unmount:L}=U(h(c,{flexDirection:"column",width:e.dashWidth,children:[t(D,{local:i,api:d,label:T[s],width:e.dashWidth}),t(O,{api:d,width:e.dashWidth,barWidth:e.barWidth}),t(F,{api:d,width:e.dashWidth,barWidth:e.barWidth}),t(E,{instructions:b,width:e.dashWidth,barWidth:e.barWidth}),t(M,{api:d,width:e.dashWidth})]}),{patchConsole:!1});L()}}m(yt,"renderDashboard");export{yt as renderDashboard};

package/dist/dashboard/launcher/ActiveJobsList.js

@@ -1 +1 @@
-import{a}from"../../chunk-NBTC5FNV.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-ZLLER3HP.js";export{a as ActiveJobsList};
+import{a}from"../../chunk-X23L2MF6.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-ZLLER3HP.js";export{a as ActiveJobsList};

package/dist/dashboard/launcher/LauncherOverlay.js

@@ -1 +1 @@
-import{a}from"../../chunk-SJGLRC2R.js";import"../../chunk-RDET7CKR.js";import"../../chunk-NBTC5FNV.js";import"../../chunk-TU2AR6XO.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-ZR3MUNLS.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as LauncherOverlay};
+import{a}from"../../chunk-BMGQFKG6.js";import"../../chunk-RDET7CKR.js";import"../../chunk-X23L2MF6.js";import"../../chunk-B22HYCBV.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-ZR3MUNLS.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as LauncherOverlay};

package/dist/dashboard/launcher/NewLaunchForm.js

@@ -1 +1 @@
-import{a}from"../../chunk-TU2AR6XO.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-ZLLER3HP.js";export{a as NewLaunchForm};
+import{a}from"../../chunk-B22HYCBV.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-ZLLER3HP.js";export{a as NewLaunchForm};

package/dist/lib/agent-loop.js

@@ -1 +1 @@
-import{a,b,c,d,e}from"../chunk-25C5P7BF.js";import"../chunk-TIOKY2IA.js";import"../chunk-EXUPWAVK.js";import"../chunk-RSKQRVPJ.js";import"../chunk-XY4AK3TR.js";import"../chunk-J66PBCNP.js";import"../chunk-O2UOCI3Q.js";import"../chunk-N7TGRFVZ.js";import"../chunk-65TC72MN.js";import"../chunk-YL3T4367.js";import"../chunk-HDU5QPEO.js";import"../chunk-QFVGUZDO.js";import"../chunk-W2MOTOQE.js";import"../chunk-6QKI756T.js";import"../chunk-5RBLIVBF.js";import"../chunk-4SFXTUFE.js";import"../chunk-KXW3IWKQ.js";import"../chunk-6V4WM73L.js";import"../chunk-KAF3CYRW.js";import"../chunk-UF6US7VU.js";import"../chunk-KYDATXAB.js";import"../chunk-ZL4MZMUK.js";import"../chunk-CC4HVDUK.js";import"../chunk-GTOEWPIM.js";import"../chunk-TDD7IOV2.js";import"../chunk-7DFNBRLY.js";import"../chunk-MODXEGCU.js";import"../chunk-VNYMMINL.js";import"../chunk-RTWRHIKY.js";import"../chunk-N7ODX6PO.js";import"../chunk-67QCYUJU.js";import"../chunk-GQRCSR2A.js";import"../chunk-IL35G6CK.js";import"../chunk-S6B2R2LL.js";import"../chunk-EL5M4MAQ.js";import"../chunk-WQ2BJQYB.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{b as CAPABILITIES,a as HANDLERS,c as availableCapabilities,d as pollForWork,e as runAgentLoop};
+import{a,b,c,d,e}from"../chunk-FS2Q2J7V.js";import"../chunk-TIOKY2IA.js";import"../chunk-MYPDODRJ.js";import"../chunk-O23Z6LTP.js";import"../chunk-PRDYXO6Z.js";import"../chunk-QFVGUZDO.js";import"../chunk-W2MOTOQE.js";import"../chunk-F53QW6VG.js";import"../chunk-EXUPWAVK.js";import"../chunk-6QKI756T.js";import"../chunk-RSKQRVPJ.js";import"../chunk-AW6LN6ZU.js";import"../chunk-DO3KHK2X.js";import"../chunk-UF6US7VU.js";import"../chunk-MSTF7A7D.js";import"../chunk-TSS6ESRK.js";import"../chunk-SD4SGJZY.js";import"../chunk-5RBLIVBF.js";import"../chunk-ZL4MZMUK.js";import"../chunk-N7TGRFVZ.js";import"../chunk-KXW3IWKQ.js";import"../chunk-GTOEWPIM.js";import"../chunk-6V4WM73L.js";import"../chunk-JUF4JLZQ.js";import"../chunk-TDD7IOV2.js";import"../chunk-7DFNBRLY.js";import"../chunk-MODXEGCU.js";import"../chunk-GQRCSR2A.js";import"../chunk-UUMW6X2W.js";import"../chunk-YDRZDBHG.js";import"../chunk-S6B2R2LL.js";import"../chunk-K25S2MGU.js";import"../chunk-CHA5SX4Y.js";import"../chunk-WQ2BJQYB.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{b as CAPABILITIES,a as HANDLERS,c as availableCapabilities,d as pollForWork,e as runAgentLoop};

package/dist/lib/auto-run.js

@@ -0,0 +1 @@
+import{a}from"../chunk-CJRYYV76.js";import"../chunk-ZLLER3HP.js";export{a as decideAutoRunStep};

package/dist/lib/codex-exec.js

@@ -0,0 +1 @@
+import{a,b,c}from"../chunk-K25S2MGU.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-ZLLER3HP.js";export{b as buildCodexArgs,a as codexModelId,c as executeCodexEval};

package/dist/lib/eval-benchmark.js

@@ -1 +1 @@
-import{a,b,c}from"../chunk-P6HUBBOD.js";import"../chunk-EL5M4MAQ.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{b as aggregateBenchmark,c as postBenchmark,a as summarizeRuns};
+import{a,b,c}from"../chunk-WXFFD4IV.js";import"../chunk-CHA5SX4Y.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{b as aggregateBenchmark,c as postBenchmark,a as summarizeRuns};

package/dist/lib/eval-case-generator.js

@@ -1 +1 @@
-import{a,b}from"../chunk-VNYMMINL.js";import"../chunk-RTWRHIKY.js";import"../chunk-N7ODX6PO.js";import"../chunk-67QCYUJU.js";import"../chunk-ZLLER3HP.js";export{a as buildFeedbackBlock,b as generateEvalCases};
+import{a,b}from"../chunk-UUMW6X2W.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-ZLLER3HP.js";export{a as buildFeedbackBlock,b as generateEvalCases};

package/dist/lib/eval-conversation-handler.js

@@ -1 +1 @@
-import{a}from"../chunk-CC4HVDUK.js";import"../chunk-VNYMMINL.js";import"../chunk-RTWRHIKY.js";import"../chunk-N7ODX6PO.js";import"../chunk-67QCYUJU.js";import"../chunk-GQRCSR2A.js";import"../chunk-IL35G6CK.js";import"../chunk-S6B2R2LL.js";import"../chunk-WQ2BJQYB.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{a as processConversationJob};
+import{a}from"../chunk-JUF4JLZQ.js";import"../chunk-GQRCSR2A.js";import"../chunk-UUMW6X2W.js";import"../chunk-YDRZDBHG.js";import"../chunk-S6B2R2LL.js";import"../chunk-K25S2MGU.js";import"../chunk-WQ2BJQYB.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{a as processConversationJob};

package/dist/lib/eval-grader.js

@@ -1 +1 @@
-import{a}from"../chunk-IL35G6CK.js";import"../chunk-S6B2R2LL.js";import"../chunk-WQ2BJQYB.js";import"../chunk-ZLLER3HP.js";export{a as gradeEvalOutput};
+import{a}from"../chunk-YDRZDBHG.js";import"../chunk-S6B2R2LL.js";import"../chunk-K25S2MGU.js";import"../chunk-WQ2BJQYB.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-ZLLER3HP.js";export{a as gradeEvalOutput};

package/dist/lib/eval-run-batch.js

@@ -0,0 +1 @@
+import{a,b}from"../chunk-UHL5NOWQ.js";import"../chunk-YDRZDBHG.js";import"../chunk-S6B2R2LL.js";import"../chunk-K25S2MGU.js";import"../chunk-WQ2BJQYB.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import"../chunk-ZLLER3HP.js";export{a as executeAndGradeCase,b as runBatchAndPost};

package/dist/lib/handlers/cron-refresh-pricing.js

@@ -1 +1 @@
-import{a}from"../../chunk-KAF3CYRW.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as cronRefreshPricing};
+import{a}from"../../chunk-DO3KHK2X.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as cronRefreshPricing};

package/dist/lib/handlers/cron-suggest-instructions.js

@@ -1 +1 @@
-import{a}from"../../chunk-KYDATXAB.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-TDD7IOV2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as cronSuggestInstructions};
+import{a}from"../../chunk-MSTF7A7D.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-TDD7IOV2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as cronSuggestInstructions};

package/dist/lib/handlers/distill-memories.js

@@ -1 +1 @@
-import{a,b}from"../../chunk-J66PBCNP.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-ZLLER3HP.js";export{b as distillMemories,a as selectNewMemories};
+import{a,b}from"../../chunk-TSS6ESRK.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-ZLLER3HP.js";export{b as distillMemories,a as selectNewMemories};

package/dist/lib/handlers/process-announcement.js

@@ -1 +1 @@
-import{a}from"../../chunk-O2UOCI3Q.js";import"../../chunk-N7TGRFVZ.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-ZLLER3HP.js";export{a as processAnnouncement};
+import{a}from"../../chunk-SD4SGJZY.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-N7TGRFVZ.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-ZLLER3HP.js";export{a as processAnnouncement};

package/dist/lib/handlers/process-changelog.js

@@ -1 +1 @@
-import{a}from"../../chunk-YLVPZ3BS.js";import"../../chunk-N7TGRFVZ.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-ZLLER3HP.js";export{a as processChangelog};
+import{a}from"../../chunk-NZ7NZEQV.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-N7TGRFVZ.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-ZLLER3HP.js";export{a as processChangelog};

package/dist/lib/handlers/process-conversation.js

@@ -1 +1 @@
-import{a}from"../../chunk-65TC72MN.js";import"../../chunk-CC4HVDUK.js";import"../../chunk-VNYMMINL.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-GQRCSR2A.js";import"../../chunk-IL35G6CK.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as processConversation};
+import{a}from"../../chunk-MYPDODRJ.js";import"../../chunk-JUF4JLZQ.js";import"../../chunk-GQRCSR2A.js";import"../../chunk-UUMW6X2W.js";import"../../chunk-YDRZDBHG.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-K25S2MGU.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as processConversation};

package/dist/lib/handlers/process-eval.js

@@ -1 +1 @@
-import{a}from"../../chunk-YL3T4367.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-6V4WM73L.js";import"../../chunk-GTOEWPIM.js";import"../../chunk-VNYMMINL.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-IL35G6CK.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as processEval};
+import{a}from"../../chunk-O23Z6LTP.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-GTOEWPIM.js";import"../../chunk-6V4WM73L.js";import"../../chunk-UUMW6X2W.js";import"../../chunk-YDRZDBHG.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-K25S2MGU.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as processEval};

package/dist/lib/handlers/process-insights.js

@@ -1 +1 @@
-import{a}from"../../chunk-HDU5QPEO.js";import"../../chunk-QFVGUZDO.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as processInsights};
+import{a}from"../../chunk-PRDYXO6Z.js";import"../../chunk-QFVGUZDO.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-ZL4MZMUK.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-H2ORWIEY.js";import"../../chunk-O6C2UGBY.js";import"../../chunk-ZLLER3HP.js";export{a as processInsights};

package/dist/lib/handlers/process-memory-ab.js

@@ -1 +1 @@
-import{a}from"../../chunk-4SFXTUFE.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-IL35G6CK.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-EL5M4MAQ.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-ZLLER3HP.js";export{a as processMemoryAb};
+import{a}from"../../chunk-F53QW6VG.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-YDRZDBHG.js";import"../../chunk-S6B2R2LL.js";import"../../chunk-K25S2MGU.js";import"../../chunk-CHA5SX4Y.js";import"../../chunk-WQ2BJQYB.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-ZLLER3HP.js";export{a as processMemoryAb};

package/dist/lib/handlers/process-session-goal.js

@@ -1 +1 @@
-import{a}from"../../chunk-XY4AK3TR.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-RTWRHIKY.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-67QCYUJU.js";import"../../chunk-ZLLER3HP.js";export{a as processSessionGoal};
+import{a}from"../../chunk-AW6LN6ZU.js";import"../../chunk-KXW3IWKQ.js";import"../../chunk-SMGSLY5N.js";import"../../chunk-PJ3WHQ67.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-ZLLER3HP.js";export{a as processSessionGoal};

package/dist/lib/judge-agreement.js

@@ -0,0 +1 @@
+import{a,b,c}from"../chunk-6UEFCMR2.js";import"../chunk-ZLLER3HP.js";export{c as buildAgreementReport,a as cohensKappa,b as humanTruth};

package/dist/lib/llm-client.js

@@ -1 +1 @@
-import{a,b,c,d,e,f}from"../chunk-RTWRHIKY.js";import"../chunk-N7ODX6PO.js";import"../chunk-67QCYUJU.js";import"../chunk-ZLLER3HP.js";export{d as detectLlmBackend,b as getToolCli,e as resetLlmBackendCache,a as resolveToolBin,f as runLlmPrompt,c as runToolPrompt};
+import{a,b,c,d,e,f}from"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-ZLLER3HP.js";export{d as detectLlmBackend,b as getToolCli,e as resetLlmBackendCache,a as resolveToolBin,f as runLlmPrompt,c as runToolPrompt};

package/dist/lib/manual-tasks.js

@@ -1 +1 @@
-import{a as l}from"../chunk-YLVPZ3BS.js";import"../chunk-N7TGRFVZ.js";import"../chunk-KXW3IWKQ.js";import{a as u}from"../chunk-UF6US7VU.js";import{a as g}from"../chunk-KYDATXAB.js";import"../chunk-ZL4MZMUK.js";import"../chunk-TDD7IOV2.js";import"../chunk-MODXEGCU.js";import"../chunk-RTWRHIKY.js";import"../chunk-N7ODX6PO.js";import"../chunk-67QCYUJU.js";import{a as i}from"../chunk-GY5IJ4RS.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import{a as o}from"../chunk-ZLLER3HP.js";var c={cron_check_changelog:o(async(r,e)=>{let s=await l({logger:{info:o(t=>e.logger.info(t),"info"),warn:o(t=>e.logger.warn(t),"warn")}});e.logger.info(`[cron_check_changelog] result: created=${s.createdCount} skipped=${s.skippedCount}`)},"cron_check_changelog"),cron_suggest_instructions:o((r,e)=>g(r,e),"cron_suggest_instructions"),cron_refresh_skills:o((r,e)=>u(r,e),"cron_refresh_skills")},p=i.map(r=>{let e=c[r.name];if(!e)throw new Error(`manual-tasks: no handler registered for "${r.name}"`);return{...r,handler:e}});async function m(r){let e=p.find(n=>n.name===r);if(!e)return{name:r,durationMs:0,ok:!1,error:`Unknown task: ${r}`,logs:[]};let s=[],t={logger:{info:o(n=>s.push({level:"info",message:n}),"info"),warn:o(n=>s.push({level:"warn",message:n}),"warn"),error:o(n=>s.push({level:"error",message:n}),"error")},job:{attempts:0}},a=Date.now();try{return await e.handler({},t),{name:r,durationMs:Date.now()-a,ok:!0,logs:s}}catch(n){return{name:r,durationMs:Date.now()-a,ok:!1,error:n instanceof Error?n.message:String(n),logs:s}}}o(m,"runManualTask");export{p as MANUAL_TASKS,m as runManualTask};
+import{a as l}from"../chunk-NZ7NZEQV.js";import{a as u}from"../chunk-UF6US7VU.js";import{a as g}from"../chunk-MSTF7A7D.js";import"../chunk-ZL4MZMUK.js";import"../chunk-N7TGRFVZ.js";import"../chunk-KXW3IWKQ.js";import"../chunk-TDD7IOV2.js";import"../chunk-MODXEGCU.js";import{a as i}from"../chunk-GY5IJ4RS.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-H2ORWIEY.js";import"../chunk-O6C2UGBY.js";import{a as o}from"../chunk-ZLLER3HP.js";var c={cron_check_changelog:o(async(r,e)=>{let s=await l({logger:{info:o(t=>e.logger.info(t),"info"),warn:o(t=>e.logger.warn(t),"warn")}});e.logger.info(`[cron_check_changelog] result: created=${s.createdCount} skipped=${s.skippedCount}`)},"cron_check_changelog"),cron_suggest_instructions:o((r,e)=>g(r,e),"cron_suggest_instructions"),cron_refresh_skills:o((r,e)=>u(r,e),"cron_refresh_skills")},p=i.map(r=>{let e=c[r.name];if(!e)throw new Error(`manual-tasks: no handler registered for "${r.name}"`);return{...r,handler:e}});async function m(r){let e=p.find(n=>n.name===r);if(!e)return{name:r,durationMs:0,ok:!1,error:`Unknown task: ${r}`,logs:[]};let s=[],t={logger:{info:o(n=>s.push({level:"info",message:n}),"info"),warn:o(n=>s.push({level:"warn",message:n}),"warn"),error:o(n=>s.push({level:"error",message:n}),"error")},job:{attempts:0}},a=Date.now();try{return await e.handler({},t),{name:r,durationMs:Date.now()-a,ok:!0,logs:s}}catch(n){return{name:r,durationMs:Date.now()-a,ok:!1,error:n instanceof Error?n.message:String(n),logs:s}}}o(m,"runManualTask");export{p as MANUAL_TASKS,m as runManualTask};

package/dist/lib/optimize-compare.js

@@ -1 +1 @@
-import{a,b,c}from"../chunk-XFNAFWWW.js";import"../chunk-EL5M4MAQ.js";import"../chunk-ZLLER3HP.js";export{b as buildVersionVerdict,c as formatVersionVerdict,a as pairVersionResults};
+import{a,b,c}from"../chunk-DMLAMJLV.js";import"../chunk-CHA5SX4Y.js";import"../chunk-ZLLER3HP.js";export{b as buildVersionVerdict,c as formatVersionVerdict,a as pairVersionResults};

package/dist/lib/reviewer-detect.js

@@ -0,0 +1 @@
+import{a,b,c,d,e}from"../chunk-3BVHPQZW.js";import"../chunk-SMGSLY5N.js";import"../chunk-PJ3WHQ67.js";import"../chunk-N7ODX6PO.js";import"../chunk-ZLLER3HP.js";export{c as CODEX_FALLBACK_MODEL,e as chooseReviewer,d as detectEngines,b as isCodexModelUnsupported,a as parseCodexConfigModel};

package/dist/lib/which.js

@@ -1 +1 @@
-import{a}from"../chunk-67QCYUJU.js";import"../chunk-ZLLER3HP.js";export{a as findBinary};
+import{a}from"../chunk-PJ3WHQ67.js";import"../chunk-ZLLER3HP.js";export{a as findBinary};

package/oclif.manifest.json

@@ -1095,6 +1095,99 @@
         "agent.js"
       ]
     },
+    "evals:auto-run": {
+      "aliases": [],
+      "args": {
+        "targetId": {
+          "description": "Eval target ID",
+          "name": "targetId",
+          "required": true
+        }
+      },
+      "description": "Run evals in a loop until the effect verdict is statistically powered (§6f auto-run-to-power)",
+      "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"
+      ],
+      "flags": {
+        "api-url": {
+          "description": "Override API base URL",
+          "env": "RULEMETRIC_API_URL",
+          "hidden": true,
+          "name": "api-url",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "yes": {
+          "char": "y",
+          "description": "Actually run. Without this, only previews the plan + cost.",
+          "name": "yes",
+          "allowNo": false,
+          "type": "boolean"
+        },
+        "max-runs": {
+          "description": "Hard ceiling on completed runs",
+          "name": "max-runs",
+          "default": 200,
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "batch-parallel": {
+          "description": "Concurrent executions within a batch",
+          "name": "batch-parallel",
+          "default": 3,
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "timeout": {
+          "description": "Timeout per eval in ms",
+          "name": "timeout",
+          "default": 120000,
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "model": {
+          "description": "Model for eval execution",
+          "name": "model",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "grader-model": {
+          "description": "Model for grading",
+          "name": "grader-model",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "skip-baseline": {
+          "description": "Skip without_target runs (no delta — not recommended)",
+          "name": "skip-baseline",
+          "allowNo": false,
+          "type": "boolean"
+        }
+      },
+      "hasDynamicHelp": false,
+      "hiddenAliases": [],
+      "id": "evals:auto-run",
+      "pluginAlias": "@rulemetric/cli",
+      "pluginName": "@rulemetric/cli",
+      "pluginType": "core",
+      "strict": true,
+      "enableJsonFlag": false,
+      "isESM": true,
+      "relativePath": [
+        "dist",
+        "commands",
+        "evals",
+        "auto-run.js"
+      ]
+    },
     "evals:benchmark": {
       "aliases": [],
       "args": {
@@ -1604,6 +1697,98 @@
         "run.js"
       ]
     },
+    "evals:validate-judge": {
+      "aliases": [],
+      "args": {
+        "targetId": {
+          "description": "Eval target ID",
+          "name": "targetId",
+          "required": true
+        }
+      },
+      "description": "Score a second model against your human judge-review labels (§6g validation pass)",
+      "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"
+      ],
+      "flags": {
+        "api-url": {
+          "description": "Override API base URL",
+          "env": "RULEMETRIC_API_URL",
+          "hidden": true,
+          "name": "api-url",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "model": {
+          "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).",
+          "name": "model",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "engine": {
+          "description": "CLI that runs the second reviewer. `auto` prefers your installed codex for a cross-vendor reviewer — two Claude models share a training lineage, so their agreement overstates independence.",
+          "name": "engine",
+          "default": "auto",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "options": [
+            "auto",
+            "claude",
+            "codex"
+          ],
+          "type": "option"
+        },
+        "limit": {
+          "description": "Sample only the N most-recently-annotated runs (cost control / smoke)",
+          "name": "limit",
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "parallel": {
+          "description": "Concurrent re-grades",
+          "name": "parallel",
+          "default": 3,
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "timeout": {
+          "description": "Timeout per re-grade in ms",
+          "name": "timeout",
+          "default": 120000,
+          "hasDynamicHelp": false,
+          "multiple": false,
+          "type": "option"
+        },
+        "json": {
+          "description": "Emit the full report as JSON",
+          "name": "json",
+          "allowNo": false,
+          "type": "boolean"
+        }
+      },
+      "hasDynamicHelp": false,
+      "hiddenAliases": [],
+      "id": "evals:validate-judge",
+      "pluginAlias": "@rulemetric/cli",
+      "pluginName": "@rulemetric/cli",
+      "pluginType": "core",
+      "strict": true,
+      "enableJsonFlag": false,
+      "isESM": true,
+      "relativePath": [
+        "dist",
+        "commands",
+        "evals",
+        "validate-judge.js"
+      ]
+    },
     "gateway:ensure": {
       "aliases": [],
       "args": {},
@@ -4132,5 +4317,5 @@
       ]
     }
   },
-  "version": "0.7.20"
+  "version": "0.7.22"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rulemetric/cli",
-  "version": "0.7.20",
+  "version": "0.7.22",
   "publishConfig": {
     "access": "public"
   },
@@ -74,9 +74,9 @@
     "postgres": "^3.4.0",
     "react": "^19.0.0",
     "zod": "^3.24.0",
-    "@rulemetric/hooks": "0.7.20",
-    "@rulemetric/proxy": "0.7.20",
-    "@rulemetric/skills-registry": "0.7.20"
+    "@rulemetric/hooks": "0.7.22",
+    "@rulemetric/skills-registry": "0.7.22",
+    "@rulemetric/proxy": "0.7.22"
   },
   "//deps": "Worker (`evals agent`) runtime deps: the @opentelemetry/* set, @sentry/node, and postgres are STATIC top-level imports loaded at worker startup (via bundled @rulemetric/{telemetry,db}) — they MUST be runtime deps or a clean `npm i -g` crashes with ModuleLoadError. `better-sqlite3` is optionalDependencies (below): it's a native module, lazy `require()`'d inside a try/catch (packages/session providers cursor/opencode) so a missing binary degrades gracefully — making it a hard dep would break `npm i -g` on platforms without prebuilt binaries. `resend` stays a devDep: lazy `await import('resend')` guarded by RESEND_API_KEY, which end-user workers never set (provider falls back to noop). Classification verified via import-graph scan of dist/commands/evals/agent.js + per-importer static/lazy check.",
   "optionalDependencies": {
@@ -94,10 +94,10 @@
     "tsx": "^4.19.0",
     "typescript": "^5.9.0",
     "vitest": "^3.2.0",
-    "@rulemetric/api": "0.0.1",
     "@rulemetric/core": "0.0.1",
     "@rulemetric/db": "0.0.1",
     "@rulemetric/effectiveness": "0.0.1",
+    "@rulemetric/api": "0.0.1",
     "@rulemetric/email": "0.0.1",
     "@rulemetric/session": "0.2.1",
     "@rulemetric/telemetry": "0.0.1",

package/dist/chunk-67QCYUJU.js

@@ -1 +0,0 @@
-import{a as i}from"./chunk-ZLLER3HP.js";import{existsSync as p}from"node:fs";import{homedir as u}from"node:os";import{join as o,delimiter as d}from"node:path";function b(r,n){return r==="darwin"?["/opt/homebrew/bin","/opt/homebrew/sbin","/usr/local/bin",o(n,".local/bin")]:r==="linux"?["/usr/local/bin","/usr/bin",o(n,".local/bin")]:[]}i(b,"wellKnownBinDirs");function g(r,n={}){let t=n.platform??process.platform,s=n.home??u(),l=(n.pathEnv??process.env.PATH??"").split(d).filter(Boolean),a=[...new Set([...l,...b(t,s)])],f=t==="win32"?["",...(process.env.PATHEXT||".EXE;.CMD;.BAT;.COM").split(";").filter(Boolean)]:[""];for(let c of a)for(let m of f){let e=o(c,r+m);if(p(e))return e}return null}i(g,"findBinary");export{g as a};

package/dist/chunk-EL5M4MAQ.js

@@ -1 +0,0 @@
-import{a as r}from"./chunk-ZLLER3HP.js";function g(e,o){let t=e,i=o,n=t+i;if(n===0)return{pValue:1,nDiscordant:0};let s=Math.min(t,i),d=Math.pow(.5,n),a=d;for(let c=1;c<=s;c++)d*=(n-c+1)/c,a+=d;return{pValue:Math.min(1,2*a),nDiscordant:n}}r(g,"mcnemarExact");function w(e,o,t=10){return{sufficient:e>=t&&o>=t,minN:t,nWith:e,nWithout:o,shortfall:{with:Math.max(0,t-e),without:Math.max(0,t-o)}}}r(w,"assessSufficiency");function R(e){let o=e>>>0;return()=>{o=o+1831565813|0;let t=Math.imul(o^o>>>15,1|o);return t=t+Math.imul(t^t>>>7,61|t)^t,((t^t>>>14)>>>0)/4294967296}}r(R,"mulberry32");function M(e,o){let t=e.length;if(t===1)return e[0];let i=(t-1)*o,n=Math.floor(i),s=Math.ceil(i);return n===s?e[n]:e[n]+(i-n)*(e[s]-e[n])}r(M,"sortedQuantile");function x(e,o={}){let t=e.length;if(t===0)return{estimate:0,lower:0,upper:0,iterations:0};let i=o.iterations??1e4,n=o.ciLevel??.95,s=R(o.seed??1),d=e.reduce((u,f)=>u+f,0)/t,a=new Array(i);for(let u=0;u<i;u++){let f=0;for(let h=0;h<t;h++)f+=e[s()*t|0];a[u]=f/t}a.sort((u,f)=>u-f);let c=(1-n)/2;return{estimate:d,lower:M(a,c),upper:M(a,1-c),iterations:i}}r(x,"pairedBootstrapCI");var v=r(e=>e===0||e===1,"isBinary");function _(e,o,t){return e.every(n=>n[o]!=null&&n[t]!=null)?e.reduce((n,s)=>n+(s[o]-s[t]),0)/e.length:null}r(_,"meanPairDelta");function $(e,o={}){let t=o.baseline??"no-rule",i=o.minN??10,n=e.length,s=e.map(l=>l.withScore-l.withoutScore),d=n===0?0:s.reduce((l,m)=>l+m,0)/n,a=x(s,{seed:o.seed,iterations:o.iterations}),c={lower:a.lower,upper:a.upper},u=e.every(l=>v(l.withScore)&&v(l.withoutScore)),f=null;if(u&&n>0){let l=0,m=0;for(let p of e)p.withScore===1&&p.withoutScore===0?l++:p.withScore===0&&p.withoutScore===1&&m++;f=g(l,m).pValue}let{sufficient:h}=w(n,n,i),E=c.lower>0||c.upper<0,b={mode:"controlled",baseline:t,delta:d,ci:c,pValue:f,significant:h&&E,sufficient:h,nWith:n,nWithout:n,costDelta:_(e,"withCost","withoutCost"),latencyDelta:_(e,"withLatency","withoutLatency"),confounded:!1};return h||(b.insufficientReason=`insufficient data: need >= ${i} paired tasks per cohort, have ${n}`),b}r($,"buildControlledVerdict");export{$ as a};