@rulemetric/cli 0.13.0 → 0.14.1

package/dist/commands/suggestions/auto.js

@@ -1 +1 @@
-import{a as p}from"../../chunk-ZGW3N3A6.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5XK6LB2T.js";import"../../chunk-P2WSDHIA.js";import"../../chunk-25RDT32M.js";import"../../chunk-3AV3GTMP.js";import{e as u,h as l}from"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import{a as c}from"../../chunk-ZLLER3HP.js";import{Flags as t}from"@oclif/core";var d=class r extends p{static{c(this,"SuggestionsAuto")}static description="Show or change automatic adoption of high-confidence suggestions into CLAUDE.md";static examples=["<%= config.bin %> suggestions auto","<%= config.bin %> suggestions auto --off","<%= config.bin %> suggestions auto --on","<%= config.bin %> suggestions auto --off --project <project-id>"];static flags={on:t.boolean({description:"Enable automatic adoption",exclusive:["off"]}),off:t.boolean({description:"Disable automatic adoption",exclusive:["on"]}),project:t.string({description:"Apply to one project only (overrides the account default)"}),"max-per-window":t.integer({description:"How many rules the loop may adopt per project per window (0 stops it adopting)"}),"max-per-run":t.integer({description:"How many rules one nightly run may adopt"}),"window-days":t.integer({description:"Length of the budget window, in days"}),"min-score":t.string({description:"Minimum suggestion score the loop may adopt (0\u20131)"}),reset:t.boolean({description:"Clear the budget overrides at this level and fall back to the shipped defaults"})};async run(){let{flags:e}=await this.parse(r);this.requireAuth();let i=e.reset?{minScore:null,maxPerRun:null,maxPerWindow:null,windowDays:null}:{...e["min-score"]!==void 0?{minScore:Number(e["min-score"])}:{},...e["max-per-run"]!==void 0?{maxPerRun:e["max-per-run"]}:{},...e["max-per-window"]!==void 0?{maxPerWindow:e["max-per-window"]}:{},...e["window-days"]!==void 0?{windowDays:e["window-days"]}:{}},n=Object.keys(i).length>0;if(e.on||e.off||n){let s=e.on||e.off?!!e.on:void 0,g=await l("/api/settings/auto-accept",{...s!==void 0?{enabled:s}:{},...n?{limits:i}:{},...e.project?{projectId:e.project}:{}});if(n){let a=g.limits;this.log(`Budget for ${e.project?`project ${e.project}`:"your account"}: ${a.maxPerRun} per night, ${a.maxPerWindow} per ${a.windowDays} days, score >= ${a.minScore}.`)}if(s===void 0)return;let m=e.project?`project ${e.project}`:"your account";this.log(`Automatic adoption is now ${s?"ON":"OFF"} for ${m}.`),s||(this.log("Already-adopted rules stay in CLAUDE.md \u2014 remove one with:"),this.log("  rulemetric suggestions undo <instruction-id>"));return}let o=await u("/api/settings/auto-accept"+(e.project?`?projectId=${encodeURIComponent(e.project)}`:""));this.log(`Automatic adoption: ${o.effective?"ON":"OFF"} (from ${o.source})`),this.log(`  account default:  ${o.accountDefault?"on":"off"}`),e.project&&this.log(`  project override: ${o.projectOverride===null?"none":o.projectOverride?"on":"off"}`),this.log(""),this.log("When on, the nightly loop may add \u2014 and only add \u2014 suggestions that:"),this.log(`  \u2022 score >= ${o.limits.minScore} (already injected into every session)`),this.log(`  \u2022 are '${o.limits.kind}' proposals; removals and trims always wait for you`),this.log("  \u2022 contain no unfilled placeholders"),this.log(`  \u2022 at most ${o.limits.maxPerRun} per night, ${o.limits.maxPerWindow} per ${o.limits.windowDays} days \u2014 per PROJECT, counting every checkout of it (${f(o.limits.source)})`),this.log(""),this.log("Change the budget:  rulemetric suggestions auto --max-per-window 20"),this.log("Back to defaults:   rulemetric suggestions auto --reset"),this.log(""),this.log("Every adopted rule is a tagged block in CLAUDE.md; undo one with:"),this.log("  rulemetric suggestions undo <instruction-id>"),this.log("Turn it off entirely with: rulemetric suggestions auto --off")}};function f(r){let e=new Set(Object.values(r));if(e.size===1){let i=[...e][0];return i==="default"?"shipped defaults":`set on your ${i}`}return Object.entries(r).map(([i,n])=>`${i}: ${n}`).join(", ")}c(f,"describeSource");export{d as default};
+import{a as p}from"../../chunk-MU7KKU3O.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-FKTT7A4H.js";import"../../chunk-VIFOLTZT.js";import"../../chunk-4FJYVYCU.js";import"../../chunk-3AV3GTMP.js";import{e as u,h as l}from"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as c}from"../../chunk-ZLLER3HP.js";import{Flags as t}from"@oclif/core";var d=class r extends p{static{c(this,"SuggestionsAuto")}static description="Show or change automatic adoption of high-confidence suggestions into CLAUDE.md";static examples=["<%= config.bin %> suggestions auto","<%= config.bin %> suggestions auto --off","<%= config.bin %> suggestions auto --on","<%= config.bin %> suggestions auto --off --project <project-id>"];static flags={on:t.boolean({description:"Enable automatic adoption",exclusive:["off"]}),off:t.boolean({description:"Disable automatic adoption",exclusive:["on"]}),project:t.string({description:"Apply to one project only (overrides the account default)"}),"max-per-window":t.integer({description:"How many rules the loop may adopt per project per window (0 stops it adopting)"}),"max-per-run":t.integer({description:"How many rules one nightly run may adopt"}),"window-days":t.integer({description:"Length of the budget window, in days"}),"min-score":t.string({description:"Minimum suggestion score the loop may adopt (0\u20131)"}),reset:t.boolean({description:"Clear the budget overrides at this level and fall back to the shipped defaults"})};async run(){let{flags:e}=await this.parse(r);this.requireAuth();let i=e.reset?{minScore:null,maxPerRun:null,maxPerWindow:null,windowDays:null}:{...e["min-score"]!==void 0?{minScore:Number(e["min-score"])}:{},...e["max-per-run"]!==void 0?{maxPerRun:e["max-per-run"]}:{},...e["max-per-window"]!==void 0?{maxPerWindow:e["max-per-window"]}:{},...e["window-days"]!==void 0?{windowDays:e["window-days"]}:{}},n=Object.keys(i).length>0;if(e.on||e.off||n){let s=e.on||e.off?!!e.on:void 0,g=await l("/api/settings/auto-accept",{...s!==void 0?{enabled:s}:{},...n?{limits:i}:{},...e.project?{projectId:e.project}:{}});if(n){let a=g.limits;this.log(`Budget for ${e.project?`project ${e.project}`:"your account"}: ${a.maxPerRun} per night, ${a.maxPerWindow} per ${a.windowDays} days, score >= ${a.minScore}.`)}if(s===void 0)return;let m=e.project?`project ${e.project}`:"your account";this.log(`Automatic adoption is now ${s?"ON":"OFF"} for ${m}.`),s||(this.log("Already-adopted rules stay in CLAUDE.md \u2014 remove one with:"),this.log("  rulemetric suggestions undo <instruction-id>"));return}let o=await u("/api/settings/auto-accept"+(e.project?`?projectId=${encodeURIComponent(e.project)}`:""));this.log(`Automatic adoption: ${o.effective?"ON":"OFF"} (from ${o.source})`),this.log(`  account default:  ${o.accountDefault?"on":"off"}`),e.project&&this.log(`  project override: ${o.projectOverride===null?"none":o.projectOverride?"on":"off"}`),this.log(""),this.log("When on, the nightly loop may add \u2014 and only add \u2014 suggestions that:"),this.log(`  \u2022 score >= ${o.limits.minScore} (already injected into every session)`),this.log(`  \u2022 are '${o.limits.kind}' proposals; removals and trims always wait for you`),this.log("  \u2022 contain no unfilled placeholders"),this.log(`  \u2022 at most ${o.limits.maxPerRun} per night, ${o.limits.maxPerWindow} per ${o.limits.windowDays} days \u2014 per PROJECT, counting every checkout of it (${f(o.limits.source)})`),this.log(""),this.log("Change the budget:  rulemetric suggestions auto --max-per-window 20"),this.log("Back to defaults:   rulemetric suggestions auto --reset"),this.log(""),this.log("Every adopted rule is a tagged block in CLAUDE.md; undo one with:"),this.log("  rulemetric suggestions undo <instruction-id>"),this.log("Turn it off entirely with: rulemetric suggestions auto --off")}};function f(r){let e=new Set(Object.values(r));if(e.size===1){let i=[...e][0];return i==="default"?"shipped defaults":`set on your ${i}`}return Object.entries(r).map(([i,n])=>`${i}: ${n}`).join(", ")}c(f,"describeSource");export{d as default};

package/dist/commands/suggestions/dismiss.js

@@ -1 +1 @@
-import{a}from"../../chunk-WLVQJH4E.js";import{a as r}from"../../chunk-MODXEGCU.js";import{a as o}from"../../chunk-ZGW3N3A6.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5XK6LB2T.js";import"../../chunk-P2WSDHIA.js";import"../../chunk-25RDT32M.js";import"../../chunk-3AV3GTMP.js";import{f as t}from"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{Args as m}from"@oclif/core";var s=class n extends o{static{e(this,"SuggestionsDismiss")}static description="Dismiss a suggestion so it will not re-surface for 30 days";static examples=["<%= config.bin %> suggestions dismiss <id>"];static args={id:m.string({description:"Suggestion UUID from `suggestions list`",required:!0})};async run(){let{args:i}=await this.parse(n);this.requireAuth(),a(i.id,"suggestion ID");let d=r(process.cwd());await t(`/api/instruction-suggestions/${encodeURIComponent(i.id)}/events`,{eventType:"dismissed",projectPath:d}),this.log("Dismissed. Won't re-surface for 30 days.")}};export{s as default};
+import{a}from"../../chunk-WLVQJH4E.js";import{a as r}from"../../chunk-MODXEGCU.js";import{a as o}from"../../chunk-MU7KKU3O.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-FKTT7A4H.js";import"../../chunk-VIFOLTZT.js";import"../../chunk-4FJYVYCU.js";import"../../chunk-3AV3GTMP.js";import{f as t}from"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{Args as m}from"@oclif/core";var s=class n extends o{static{e(this,"SuggestionsDismiss")}static description="Dismiss a suggestion so it will not re-surface for 30 days";static examples=["<%= config.bin %> suggestions dismiss <id>"];static args={id:m.string({description:"Suggestion UUID from `suggestions list`",required:!0})};async run(){let{args:i}=await this.parse(n);this.requireAuth(),a(i.id,"suggestion ID");let d=r(process.cwd());await t(`/api/instruction-suggestions/${encodeURIComponent(i.id)}/events`,{eventType:"dismissed",projectPath:d}),this.log("Dismissed. Won't re-surface for 30 days.")}};export{s as default};

package/dist/commands/suggestions/list.js

@@ -1,3 +1,3 @@
-import{a as u}from"../../chunk-MODXEGCU.js";import{a as g}from"../../chunk-ZGW3N3A6.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5XK6LB2T.js";import"../../chunk-P2WSDHIA.js";import"../../chunk-25RDT32M.js";import"../../chunk-3AV3GTMP.js";import{e as d}from"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Flags as j}from"@oclif/core";import{printTable as v}from"@oclif/table";var r=class p extends g{static{a(this,"SuggestionsList")}static description="List instruction suggestions for a project";static examples=["<%= config.bin %> suggestions list","<%= config.bin %> suggestions list --project /path/to/project"];static flags={project:j.string({char:"p",description:"Project path (defaults to current directory)"})};async run(){let{flags:m}=await this.parse(p);this.requireAuth();let l=u(m.project??process.cwd()),f=encodeURIComponent(l),s=await d(`/api/instruction-suggestions?projectPath=${f}&limit=10&groupByKind=true`);if(s.suggestions.length===0){this.log("No suggestions for this project. They are recomputed daily at 04:30 UTC by a server-side cron and run on your background worker \u2014 make sure one is running (`rulemetric evals agent`, or `rulemetric service install --worker-only`).");return}v({data:s.suggestions.map(e=>({ID:e.id,Action:(e.kind??"add")==="add"?"add":(e.kind??"add").toUpperCase(),Scope:e.instructionScope,Name:e.name,Tokens:e.estimatedTokens?String(e.estimatedTokens):"-",Score:e.score.toFixed(2),Reason:e.reason.length>60?e.reason.slice(0,59)+"\u2026":e.reason})),columns:["ID","Action","Scope","Name","Tokens","Score","Reason"]});let o=s.suggestions.filter(e=>e.kind==="remove");if(o.length>0){let e=o.reduce((h,k)=>h+(k.estimatedTokens??0),0);this.log(`
+import{a as u}from"../../chunk-MODXEGCU.js";import{a as g}from"../../chunk-MU7KKU3O.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-FKTT7A4H.js";import"../../chunk-VIFOLTZT.js";import"../../chunk-4FJYVYCU.js";import"../../chunk-3AV3GTMP.js";import{e as d}from"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Flags as j}from"@oclif/core";import{printTable as v}from"@oclif/table";var r=class p extends g{static{a(this,"SuggestionsList")}static description="List instruction suggestions for a project";static examples=["<%= config.bin %> suggestions list","<%= config.bin %> suggestions list --project /path/to/project"];static flags={project:j.string({char:"p",description:"Project path (defaults to current directory)"})};async run(){let{flags:m}=await this.parse(p);this.requireAuth();let l=u(m.project??process.cwd()),f=encodeURIComponent(l),s=await d(`/api/instruction-suggestions?projectPath=${f}&limit=10&groupByKind=true`);if(s.suggestions.length===0){this.log("No suggestions for this project. They are recomputed daily at 04:30 UTC by a server-side cron and run on your background worker \u2014 make sure one is running (`rulemetric evals agent`, or `rulemetric service install --worker-only`).");return}v({data:s.suggestions.map(e=>({ID:e.id,Action:(e.kind??"add")==="add"?"add":(e.kind??"add").toUpperCase(),Scope:e.instructionScope,Name:e.name,Tokens:e.estimatedTokens?String(e.estimatedTokens):"-",Score:e.score.toFixed(2),Reason:e.reason.length>60?e.reason.slice(0,59)+"\u2026":e.reason})),columns:["ID","Action","Scope","Name","Tokens","Score","Reason"]});let o=s.suggestions.filter(e=>e.kind==="remove");if(o.length>0){let e=o.reduce((h,k)=>h+(k.estimatedTokens??0),0);this.log(`
 ${o.length} retirement proposal(s) \u2014 accepting one archives the instruction (reversible: rulemetric instructions restore <instruction-id>).`+(e>0?` Together they cost roughly ${e} tokens of context per session.`:""))}let t=s.stats??{},i=(t.surfaced??0)+(t.surfaced_with_content??0)+(t.proxy_injected??0),n=(t.accepted??0)+(t.promoted??0),c=t.dismissed??0;i+n+c>0&&this.log(`
 Lifetime for this project: ${i} surfaced \xB7 ${n} accepted \xB7 ${c} dismissed`)}};export{r as default};

package/dist/commands/suggestions/refresh.js

@@ -1 +1 @@
-import{a as i}from"../../chunk-ZGW3N3A6.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5XK6LB2T.js";import"../../chunk-P2WSDHIA.js";import"../../chunk-25RDT32M.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import{a as r}from"../../chunk-ZLLER3HP.js";var e=class extends i{static{r(this,"SuggestionsRefresh")}static description="Information about refreshing instruction suggestions";static examples=["<%= config.bin %> suggestions refresh"];async run(){this.log("Suggestions are recomputed automatically: a server-side cron enqueues a job daily at 04:30 UTC, and your background worker picks it up \u2014 there is no manual trigger."),this.log("Just make sure a worker is running: `rulemetric evals agent` (foreground), or install it persistently with `rulemetric service install --worker-only`.")}};export{e as default};
+import{a as i}from"../../chunk-MU7KKU3O.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-FKTT7A4H.js";import"../../chunk-VIFOLTZT.js";import"../../chunk-4FJYVYCU.js";import"../../chunk-3AV3GTMP.js";import"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as r}from"../../chunk-ZLLER3HP.js";var e=class extends i{static{r(this,"SuggestionsRefresh")}static description="Information about refreshing instruction suggestions";static examples=["<%= config.bin %> suggestions refresh"];async run(){this.log("Suggestions are recomputed automatically: a server-side cron enqueues a job daily at 04:30 UTC, and your background worker picks it up \u2014 there is no manual trigger."),this.log("Just make sure a worker is running: `rulemetric evals agent` (foreground), or install it persistently with `rulemetric service install --worker-only`.")}};export{e as default};

package/dist/commands/suggestions/undo.js

@@ -1 +1 @@
-import{a as f,c as g}from"../../chunk-6QLZT7EP.js";import{a as l}from"../../chunk-WLVQJH4E.js";import{d as p}from"../../chunk-4FSQSETR.js";import"../../chunk-VH2EAP4J.js";import"../../chunk-J6QMBMC2.js";import"../../chunk-UYPBIXF2.js";import{d as u}from"../../chunk-OZNKMJ5B.js";import"../../chunk-LJRE3LZ2.js";import{a as m}from"../../chunk-MODXEGCU.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-UPFH72PF.js";import{a as d}from"../../chunk-ZGW3N3A6.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-5XK6LB2T.js";import"../../chunk-P2WSDHIA.js";import"../../chunk-25RDT32M.js";import"../../chunk-3AV3GTMP.js";import{f as a}from"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import{a as c}from"../../chunk-ZLLER3HP.js";import{relative as h}from"node:path";import{Args as P}from"@oclif/core";var o=class I extends d{static{c(this,"SuggestionsUndo")}static description="Remove an adopted instruction from this project's agent instruction files and stop it being re-proposed";static examples=["<%= config.bin %> suggestions undo <instruction-id>"];static args={instructionId:P.string({description:"Instruction ID from the rulemetric:start marker in the instruction file",required:!0})};async run(){let{args:r}=await this.parse(I);this.requireAuth(),l(r.instructionId,"instruction ID");let e=m(process.cwd()),i=process.cwd(),v=u(i).map(t=>h(i,t.path)),w=p(i,r.instructionId).map(t=>h(i,t.path)),A=await f({dir:process.cwd(),instructionId:r.instructionId,projectPath:e}),n=null,s=null;try{n=(await a("/api/instruction-suggestions/undo",{projectPath:e,instructionId:r.instructionId})).name}catch(t){s=t instanceof Error?t.message:String(t)}for(let t of g({stripped:w,searched:v,reversals:A,name:n,instructionId:r.instructionId,dismissalError:s}))this.log(t);s&&this.exit(1)}};export{o as default};
+import{a as f,c as g}from"../../chunk-VU24IH5S.js";import{a as l}from"../../chunk-WLVQJH4E.js";import{d as p}from"../../chunk-2UK4DZFW.js";import"../../chunk-3FNTV2TW.js";import"../../chunk-JTYVBR5H.js";import"../../chunk-B5X6HCBK.js";import"../../chunk-LJRE3LZ2.js";import{a as m}from"../../chunk-MODXEGCU.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import{d as u}from"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import{a as d}from"../../chunk-MU7KKU3O.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-FKTT7A4H.js";import"../../chunk-VIFOLTZT.js";import"../../chunk-4FJYVYCU.js";import"../../chunk-3AV3GTMP.js";import{f as a}from"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as c}from"../../chunk-ZLLER3HP.js";import{relative as h}from"node:path";import{Args as P}from"@oclif/core";var o=class I extends d{static{c(this,"SuggestionsUndo")}static description="Remove an adopted instruction from this project's agent instruction files and stop it being re-proposed";static examples=["<%= config.bin %> suggestions undo <instruction-id>"];static args={instructionId:P.string({description:"Instruction ID from the rulemetric:start marker in the instruction file",required:!0})};async run(){let{args:r}=await this.parse(I);this.requireAuth(),l(r.instructionId,"instruction ID");let e=m(process.cwd()),i=process.cwd(),v=u(i).map(t=>h(i,t.path)),w=p(i,r.instructionId).map(t=>h(i,t.path)),A=await f({dir:process.cwd(),instructionId:r.instructionId,projectPath:e}),n=null,s=null;try{n=(await a("/api/instruction-suggestions/undo",{projectPath:e,instructionId:r.instructionId})).name}catch(t){s=t instanceof Error?t.message:String(t)}for(let t of g({stripped:w,searched:v,reversals:A,name:n,instructionId:r.instructionId,dismissalError:s}))this.log(t);s&&this.exit(1)}};export{o as default};

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-3UPHCQWI.js";import{a as $}from"../chunk-WJZY5Y22.js";import"../chunk-7AWFSFR7.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-CYYDQFRY.js";import"../chunk-ZXF5NE54.js";import"../chunk-F6DAHCJC.js";import{a as X}from"../chunk-3AV3GTMP.js";import"../chunk-APDRR6FS.js";import"../chunk-J5HW45PZ.js";import"../chunk-PZNWH4VQ.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-45KJH5M7.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 O}from"../chunk-DVN7GFJG.js";import{a as E}from"../chunk-O2SDD2WW.js";import{a as M}from"../chunk-GBMF3TP6.js";import{a as U}from"../chunk-IQEZ3T5A.js";import{a as D}from"../chunk-WSCK4WNK.js";import{a as R}from"../chunk-4Z5ECCJE.js";import"../chunk-ARNZQBMX.js";import{a as B,b as S,c as F}from"../chunk-35YVHBLL.js";import{a as z}from"../chunk-XMTKP74H.js";import"../chunk-5Z3H5JHK.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-44TYRMIS.js";import"../chunk-ZXF5NE54.js";import"../chunk-T75KAEJO.js";import{a as G}from"../chunk-3AV3GTMP.js";import"../chunk-76UKAFNM.js";import"../chunk-42A4PUW7.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import{a as b}from"../chunk-ZLLER3HP.js";import{useState as h,useEffect as lt,useCallback as H,useRef as dt}from"react";import{render as J,Box as c,Text as e,useInput as ct,useApp as ut}from"ink";import{Fragment as wt,jsx as t,jsxs as f}from"react/jsx-runtime";var ht=90;function V(){let s=process.stdout.columns||80,n=Math.min(160,s),r=n>=ht,d=r?Math.floor(n/2):n,x=Math.max(6,Math.min(10,d-34));return{dashWidth:n,wide:r,halfWidth:d,barWidth:x}}b(V,"getLayout");function K({active:s,providerName:n}){return f(c,{justifyContent:"space-between",paddingX:1,children:[t(c,{gap:1,children:u.map(r=>t(e,{bold:s===r,color:s===r?a:v,children:s===r?`[ ${T[r]} ]`:`  ${T[r]}  `},r))}),n&&t(e,{color:a,bold:!0,children:n})]})}b(K,"PeriodTabs");function Q({width:s,insightsFocused:n=!1}){return t(c,{borderStyle:"round",borderColor:v,width:s,justifyContent:"center",paddingX:1,children:f(e,{children:[t(e,{color:a,bold:!0,children:"<"}),t(e,{color:a,children:">"}),t(e,{dimColor:!0,children:" switch   "}),t(e,{color:a,bold:!0,children:"q"}),t(e,{dimColor:!0,children:" quit   "}),t(e,{color:a,bold:!0,children:"s"}),f(e,{dimColor:!0,children:[" ",n?"back":"suggestions","   "]}),t(e,{color:a,bold:!0,children:"1"}),t(e,{dimColor:!0,children:"-"}),t(e,{color:a,bold:!0,children:"5"}),t(e,{dimColor:!0,children:" period   "}),t(e,{color:a,bold:!0,children:"r"}),t(e,{dimColor:!0,children:" refresh   "}),t(e,{color:a,bold:!0,children:"n"}),t(e,{dimColor:!0,children:" new session   "}),t(e,{color:a,bold:!0,children:"m"}),t(e,{dimColor:!0,children:" measurement   "}),t(e,{color:a,bold:!0,children:"l"}),t(e,{dimColor:!0,children:" launcher"})]})})}b(Q,"StatusBar");function ft({wide:s,width:n,children:r}){return s?t(c,{width:n,children:r}):t(wt,{children:r})}b(ft,"Row");function mt({initialLocal:s,initialApi:n,initialInstructions:r,initialPeriod:d,initialLauncherFocused:x=!1}){let{exit:j}=ut(),[o,L]=h(d),[Z,tt]=h(s),[W,et]=h(n),[ot,it]=h(r),[rt,k]=h(!1),[Y,_]=h(!1),[m,I]=h(!1),[q,N]=h(x),nt=G()?.orgSlug??null,y=dt(null),st=V(),g=H(async()=>{k(!0);let[i,p,P]=await Promise.all([B(),S(),F()]);tt(i),et(p),it(P),k(!1)},[]);lt(()=>{let i=setInterval(g,3e4);return()=>clearInterval(i)},[g]);let w=H(i=>{i!==o&&(L(i),y.current&&clearTimeout(y.current),y.current=setTimeout(g,300))},[o,g]);ct((i,p)=>{if(q||Y)return;if(i==="q"){if(m){I(!1);return}j();return}if(i==="s"){I(A=>!A);return}if(p.escape&&m){I(!1);return}if(i==="m"&&!m){_(!0);return}if(i==="l"&&!m){N(!0);return}if(m)return;if(i==="r"){g();return}if(i==="n"){import("../dist-GSHUC7UU.js").then(({getLauncher:A})=>{A("claude_code").launch({projectPath:process.cwd(),mode:"interactive"})});return}let P=u.indexOf(o);p.leftArrow?w(u[(P-1+u.length)%u.length]):p.rightArrow||p.tab?w(u[(P+1)%u.length]):i==="1"?w("today"):i==="2"?w("week"):i==="3"?w("30days"):i==="4"?w("month"):i==="5"&&w("all")});let{dashWidth:l,wide:X,halfWidth:at,barWidth:C}=st,$=X?at:l;return Y?t(U,{width:l,onClose:()=>_(!1)}):rt?f(c,{flexDirection:"column",width:l,children:[t(K,{active:o}),t(c,{borderStyle:"round",borderColor:a,paddingX:1,width:l,children:f(e,{dimColor:!0,children:["Loading ",T[o],"..."]})}),t(Q,{width:l})]}):q?t(c,{flexDirection:"column",width:l,children:t(z,{onClose:()=>N(!1),activeOrg:nt,defaultProjectPath:process.cwd(),width:l})}):f(c,{flexDirection:"column",width:l,children:[t(K,{active:o}),t(R,{local:Z,api:W,label:T[o],width:l}),f(ft,{wide:X,width:l,children:[t(D,{api:W,width:$,barWidth:C}),t(O,{api:W,width:$,barWidth:C})]}),t(M,{instructions:ot,width:l,barWidth:C}),t(E,{api:W,width:l,focused:m}),t(Q,{width:l,insightsFocused:m})]})}b(mt,"InteractiveDashboard");async function Bt(s="week",n={}){let[r,d,x]=await Promise.all([B(),S(),F()]);if(process.stdin.isTTY&&process.stdout.isTTY){let{waitUntilExit:o}=J(t(mt,{initialLocal:r,initialApi:d,initialInstructions:x,initialPeriod:s,initialLauncherFocused:n.openLauncher??!1}));await o()}else{let o=V(),{unmount:L}=J(f(c,{flexDirection:"column",width:o.dashWidth,children:[t(R,{local:r,api:d,label:T[s],width:o.dashWidth}),t(D,{api:d,width:o.dashWidth,barWidth:o.barWidth}),t(O,{api:d,width:o.dashWidth,barWidth:o.barWidth}),t(M,{instructions:x,width:o.dashWidth,barWidth:o.barWidth}),t(E,{api:d,width:o.dashWidth})]}),{patchConsole:!1});L()}}b(Bt,"renderDashboard");export{Bt as renderDashboard};

package/dist/dashboard/data.js

@@ -1 +1 @@
-import{a,b,c}from"../chunk-3UPHCQWI.js";import"../chunk-F6DAHCJC.js";import"../chunk-APDRR6FS.js";import"../chunk-J5HW45PZ.js";import"../chunk-PZNWH4VQ.js";import"../chunk-ZLLER3HP.js";export{b as loadApiInsights,c as loadLinkedInstructions,a as loadLocalStats};
+import{a,b,c}from"../chunk-35YVHBLL.js";import"../chunk-T75KAEJO.js";import"../chunk-76UKAFNM.js";import"../chunk-42A4PUW7.js";import"../chunk-XQYAVSUQ.js";import"../chunk-PZNWH4VQ.js";import"../chunk-ZLLER3HP.js";export{b as loadApiInsights,c as loadLinkedInstructions,a as loadLocalStats};

package/dist/dashboard/launcher/LauncherOverlay.js

@@ -1 +1 @@
-import{a}from"../../chunk-WJZY5Y22.js";import"../../chunk-7AWFSFR7.js";import"../../chunk-NBTC5FNV.js";import"../../chunk-TU2AR6XO.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-CYYDQFRY.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{a as LauncherOverlay};
+import{a}from"../../chunk-XMTKP74H.js";import"../../chunk-5Z3H5JHK.js";import"../../chunk-NBTC5FNV.js";import"../../chunk-TU2AR6XO.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-44TYRMIS.js";import"../../chunk-ZXF5NE54.js";import"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{a as LauncherOverlay};

package/dist/dashboard/launcher/api.js

@@ -1 +1 @@
-import{a,b,c,d}from"../../chunk-CYYDQFRY.js";import"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{c as cancelLaunchJob,b as createLaunchJob,a as fetchLaunchJobs,d as listLaunchers};
+import{a,b,c,d}from"../../chunk-44TYRMIS.js";import"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{c as cancelLaunchJob,b as createLaunchJob,a as fetchLaunchJobs,d as listLaunchers};

package/dist/dashboard/launcher/useLaunchJobs.js

@@ -1 +1 @@
-import{a}from"../../chunk-7AWFSFR7.js";import"../../chunk-CYYDQFRY.js";import"../../chunk-APDRR6FS.js";import"../../chunk-J5HW45PZ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{a as useLaunchJobs};
+import{a}from"../../chunk-5Z3H5JHK.js";import"../../chunk-44TYRMIS.js";import"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{a as useLaunchJobs};

package/dist/dashboard/panels/measurement.js

@@ -0,0 +1 @@
+import{a}from"../../chunk-IQEZ3T5A.js";import"../../chunk-NJ5UED3H.js";import"../../chunk-76UKAFNM.js";import"../../chunk-42A4PUW7.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import"../../chunk-ZLLER3HP.js";export{a as MeasurementPanel};

package/dist/dist-GSHUC7UU.js

@@ -0,0 +1 @@
+import{a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v,w,x}from"./chunk-T75KAEJO.js";import"./chunk-ZLLER3HP.js";export{c as ClaudeCodeAdapter,i as CodexAdapter,f as CopilotAgentAdapter,d as CursorAdapter,j as PiAdapter,g as RuleMetricMcpAdapter,b as SESSION_EVENT_TYPES,a as SESSION_TOOLS,e as VSCodeCopilotAdapter,t as buildTmuxAttachCommand,s as buildTmuxHasSessionArgs,q as buildTmuxLaunchArgs,r as buildTmuxSendKeysArgs,l as detectTool,n as discoverAllSessions,h as freshCodexState,k as getAdapter,m as getAllProviders,v as getAvailableLaunchers,u as getLauncher,o as getProvider,w as sanitizeBranch,p as sanitizeTmuxName,x as worktreePathFor};

package/dist/harbor/cost_delta.py

@@ -0,0 +1,377 @@
+#!/usr/bin/env python3
+"""The COST half of a Harbor A/B: a per-trial cost delta with an interval.
+
+    python3 cost_delta.py <baseline-job-dir> <with-job-dir> [--json]
+
+`verdict_bridge.py` gives the REWARD half of an A/B a Newcombe CI, a Fisher p
+and a graduation gate that refuses rather than publish an undecided result.
+Until 2026-09-02 the cost half had none of that: `summarize.py` summed cost over
+trials and `run.sh` printed the difference of two sums. That is wrong twice.
+
+  1. A sum has no variance, so "the with-arm was $0.09 cheaper" could never be
+     told apart from noise. Cost is right-skewed -- one flailing trial costs
+     several times a clean one -- so a difference of totals is exactly the shape
+     that reads as a finding and is a coin flip.
+  2. A sum is confounded by TRIAL COUNT. Unequal `n_trials` between arms (a
+     crashed trial, a different `-l` cap) puts "this arm ran more trials" into
+     the number and calls it cost. The per-trial MEAN is the comparable unit.
+
+Estimator: percentile bootstrap on the difference of per-trial means. Not a
+t-interval -- that assumes a symmetry cost does not have. Seeded, so a published
+verdict is reproducible from the same job dirs.
+
+The gate mirrors verdict_bridge.py on purpose: below COST_MIN_N per arm, or a CI
+spanning zero, the decision is `undecided`. "We could not tell" must never be
+published as "it is cheaper".
+
+Two honesty rules are load-bearing:
+  * An arm with no per-trial cost output (the remote agent bills the key
+    directly) is UNKNOWN, and this tool exits non-zero. Reading it as $0 would
+    print the largest saving the tool is capable of printing.
+  * The REWARD delta travels with the cost number. Cheaper because the with-arm
+    gave up early is not a saving, and the person writing the case study must
+    not have to look that up separately.
+
+Dependency-free and importable, so the contract test can drive it on synthetic
+job dirs with no harbor install.
+"""
+
+from __future__ import annotations
+
+import json
+import random
+import sys
+from pathlib import Path
+from typing import Any
+
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+from summarize import summarize  # noqa: E402
+
+# Same floor as verdict_bridge.py's GRADUATION_MIN_N. A cost saving and a reward
+# gain are published side by side; letting cost graduate on a smaller n would
+# make the cheaper number the easy one to quote.
+import provenance
+
+COST_MIN_N = 25
+# Harbor rewards on these tasks are pass/fail. A partial reward is not a success:
+# you cannot redo 60% of a task, so the redo it would price is a whole trial.
+SUCCESS_THRESHOLD = 1.0
+# Above this share of bootstrap resamples with no successes in an arm, the ratio
+# has no sampling distribution worth reporting and the verdict is refused by name.
+MAX_UNDEFINED_FRACTION = 0.05
+RESAMPLES = 10000
+SEED = 20260902
+
+
+def arm(job_dir: Path) -> dict[str, Any]:
+    """Per-trial cost series for one arm, plus its reward."""
+    s = summarize(job_dir)
+    trials = s.get("trials") or []
+    costs = [float(t["cost_usd"]) for t in trials]
+    # Paired (cost, reward) for the per-success ratio. A trial whose reward is
+    # UNKNOWN (harbor wrote no verdict) is dropped from this pairing rather than
+    # scored as a failure -- a crashed trial must not read as an instruction that
+    # did not work. It still counts in the per-trial cost above: it was paid for.
+    paired = [(float(t["cost_usd"]), float(t["reward"])) for t in trials if t.get("reward") is not None]
+    successes = sum(1 for _, r in paired if r >= SUCCESS_THRESHOLD)
+    return {
+        "job": s["job"],
+        "n": len(costs),
+        "n_trials_reported": s.get("n_trials"),
+        "costs": costs,
+        "cost_sum_usd": round(sum(costs), 6),
+        "cost_mean_usd": round(sum(costs) / len(costs), 6) if costs else None,
+        "duration_mean_s": round(sum(t["duration_ms"] for t in trials) / len(trials) / 1000, 3)
+        if trials
+        else None,
+        "turns_mean": round(sum(t["turns"] for t in trials) / len(trials), 3) if trials else None,
+        "agent_errors": sum(1 for t in trials if t["is_error"]) if trials else 0,
+        "reward_mean": s.get("reward_mean"),
+        "paired": paired,
+        "n_scored": len(paired),
+        "successes": successes,
+        # None, not infinity: an arm that never passed has no cost PER SUCCESS.
+        # Every trial was paid for and none of them produced one.
+        "cost_per_success_usd": round(sum(c for c, _ in paired) / successes, 6) if successes else None,
+    }
+
+
+def bootstrap_ci(a: list[float], b: list[float], resamples: int = RESAMPLES, seed: int = SEED):
+    """Percentile 95% CI for mean(b) - mean(a). Seeded: same input, same interval."""
+    rng = random.Random(seed)
+    na, nb = len(a), len(b)
+    deltas = []
+    for _ in range(resamples):
+        ma = sum(a[rng.randrange(na)] for _ in range(na)) / na
+        mb = sum(b[rng.randrange(nb)] for _ in range(nb)) / nb
+        deltas.append(mb - ma)
+    deltas.sort()
+    lo = deltas[int(0.025 * (resamples - 1))]
+    hi = deltas[int(0.975 * (resamples - 1))]
+    return round(lo, 6), round(hi, 6)
+
+
+
+def per_success_ci(
+    a: list[tuple[float, float]],
+    b: list[tuple[float, float]],
+    resamples: int = RESAMPLES,
+    seed: int = SEED,
+):
+    """95% CI for cost-per-success(b) - cost-per-success(a), and the share of
+    resamples where the ratio was undefined.
+
+    Resampling is PAIRED within an arm -- the trial that cost this much is the
+    trial that did or did not pass -- so the ratio's numerator and denominator
+    move together the way they do in a real run. Resamples where either arm drew
+    no successes have no ratio; they are counted, not quietly dropped, because
+    "the interval came from the 40% of resamples that happened to work" is not an
+    interval anyone should publish.
+    """
+    rng = random.Random(seed + 1)
+    na, nb = len(a), len(b)
+    deltas: list[float] = []
+    undefined = 0
+    for _ in range(resamples):
+        sa = [a[rng.randrange(na)] for _ in range(na)]
+        sb = [b[rng.randrange(nb)] for _ in range(nb)]
+        ka = sum(1 for _, r in sa if r >= SUCCESS_THRESHOLD)
+        kb = sum(1 for _, r in sb if r >= SUCCESS_THRESHOLD)
+        if ka == 0 or kb == 0:
+            undefined += 1
+            continue
+        deltas.append(sum(c for c, _ in sb) / kb - sum(c for c, _ in sa) / ka)
+    fraction = undefined / resamples
+    if not deltas:
+        return None, None, fraction
+    deltas.sort()
+    n = len(deltas)
+    return round(deltas[int(0.025 * (n - 1))], 6), round(deltas[int(0.975 * (n - 1))], 6), fraction
+
+
+def per_success_block(b: dict[str, Any], w: dict[str, Any], min_n: int, resamples: int):
+    """The cost-per-success half of the report, with its own gate.
+
+    Separate from the per-trial gate on purpose: an instruction routinely makes a
+    trial cost MORE and a success cost LESS, and one decision cannot carry both
+    without one of them being buried.
+    """
+    reasons: list[str] = []
+    bs, ws = b["cost_per_success_usd"], w["cost_per_success_usd"]
+
+    if b["n_scored"] < min_n or w["n_scored"] < min_n:
+        reasons.append(
+            f"below COST_MIN_N={min_n} scored trial(s) per arm "
+            f"(baseline {b['n_scored']}, with {w['n_scored']})"
+        )
+
+    delta = ci = pct = None
+    ci_pct = [None, None]
+    fraction = None
+    if b["successes"] == 0 or w["successes"] == 0:
+        which = "baseline" if b["successes"] == 0 else "with-arm"
+        reasons.append(
+            f"{which} never succeeded — its cost per success is unbounded, not a ratio "
+            f"(every trial was paid for and none passed); compare the pass rates instead"
+        )
+    else:
+        delta = round(ws - bs, 6)
+        pct = round(delta / bs * 100, 4) if bs else None
+        lo, hi, fraction = per_success_ci(b["paired"], w["paired"], resamples=resamples)
+        if lo is None:
+            reasons.append("cost per success was undefined in every resample")
+        else:
+            ci = [lo, hi]
+            ci_pct = [round(lo / bs * 100, 4), round(hi / bs * 100, 4)] if bs else [None, None]
+            if fraction > MAX_UNDEFINED_FRACTION:
+                reasons.append(
+                    f"cost per success undefined in {fraction:.1%} of resamples "
+                    f"(> {MAX_UNDEFINED_FRACTION:.0%}) — too few successes to price one"
+                )
+            elif lo <= 0 <= hi:
+                reasons.append("CI spans zero — undecided, refusing rather than claiming a saving")
+
+    if reasons:
+        decision = "undecided"
+    else:
+        decision = "cheaper" if delta < 0 else "costlier"
+
+    return {
+        "baseline_usd": bs,
+        "with_usd": ws,
+        "baseline_successes": f"{b['successes']}/{b['n_scored']}",
+        "with_successes": f"{w['successes']}/{w['n_scored']}",
+        "delta_usd": delta,
+        "delta_pct": pct,
+        "ci_usd": ci,
+        "ci_pct": ci_pct,
+        "undefined_fraction": round(fraction, 4) if fraction is not None else None,
+        "decision": decision,
+        "refused_reasons": reasons,
+    }
+
+
+def compare(baseline_dir: Path, with_dir: Path, min_n: int = COST_MIN_N, resamples: int = RESAMPLES,
+            allow_cross=(), allow_unstamped: bool = False):
+    # Provenance FIRST: if the two arms were not the same harness, the
+    # numbers below are a comparison of harnesses, and no interval around
+    # them means what it appears to mean (#415, 2026-08-31 -> 2026-09-02).
+    prov = provenance.compare(
+        baseline_dir, with_dir, allow_cross=allow_cross, allow_unstamped=allow_unstamped
+    )
+    b = arm(baseline_dir)
+    w = arm(with_dir)
+
+    for a in (b, w):
+        if a["n"] == 0:
+            raise ValueError(
+                f"no per-trial cost measured in {a['job']} — an unmeasured arm is UNKNOWN, "
+                "not $0.00 (the remote agent bills the key directly and writes no "
+                "claude-output.json; a host claude arm would have written one per trial)"
+            )
+
+    delta = w["cost_mean_usd"] - b["cost_mean_usd"]
+    lo, hi = bootstrap_ci(b["costs"], w["costs"], resamples=resamples)
+
+    def pct(x):
+        return round(x / b["cost_mean_usd"] * 100, 4) if b["cost_mean_usd"] else None
+
+    reasons = []
+    if b["n"] < min_n or w["n"] < min_n:
+        reasons.append(f"below COST_MIN_N={min_n} per arm (baseline {b['n']}, with {w['n']})")
+    if lo <= 0 <= hi:
+        reasons.append("CI spans zero — undecided, refusing rather than claiming a saving")
+    if not prov["comparable"]:
+        reasons.extend(prov["reasons"])
+
+    if reasons:
+        decision = "undecided"
+    else:
+        decision = "cheaper" if delta < 0 else "costlier"
+
+    # per_success_block builds its OWN decision, so the provenance refusal has to
+    # reach it too. Otherwise an incomparable pair could refuse per-trial and
+    # still carry a confident "cheaper per success" onto the job row.
+    per_success = per_success_block(b, w, min_n=min_n, resamples=resamples)
+    if per_success is not None and not prov["comparable"]:
+        per_success = dict(
+            per_success,
+            decision="undecided",
+            refused_reasons=list(per_success.get("refused_reasons", [])) + prov["reasons"],
+        )
+
+    rb, rw = b["reward_mean"], w["reward_mean"]
+    regressed = rb is not None and rw is not None and rw < rb - 1e-9
+
+    drop = ("costs", "paired")
+    return {
+        "baseline": {k: v for k, v in b.items() if k not in drop},
+        "with": {k: v for k, v in w.items() if k not in drop},
+        # The metric that can show an instruction making work CHEAPER. Per-trial
+        # cost prices a failed trial and a passing one alike; a baseline that
+        # fails is not cheap, it is a trial you pay for and then pay again.
+        "per_success": per_success,
+        "delta_usd": round(delta, 6),
+        "delta_pct": pct(delta),
+        "ci_usd": [lo, hi],
+        "ci_pct": [pct(lo), pct(hi)],
+        "decision": decision,
+        "refused_reasons": reasons,
+        # Carried whether or not it refused: a verdict that pooled across a key
+        # by --allow-cross, or could not verify at all, must say so wherever it
+        # is read -- including the scheduled handler, which reads this JSON.
+        "provenance": prov,
+        "reward": {
+            "baseline": rb,
+            "with": rw,
+            "delta": round(rw - rb, 6) if rb is not None and rw is not None else None,
+            # A cost win under a reward regression is not a win. Stated here AND
+            # in the headline so it cannot be dropped on the way to a slide.
+            "regressed": regressed,
+        },
+        "min_n": min_n,
+        "resamples": resamples,
+        "seed": SEED,
+    }
+
+
+def pretty(r: dict[str, Any], brief: bool = False) -> str:
+    """Human form. `brief` drops the per-arm lines for run.sh, which already
+    prints its own arm summaries — the cost verdict is what it lacks."""
+    b, w = r["baseline"], r["with"]
+    lines = [] if brief else [
+        f"baseline : {b['n']} trial(s) · ${b['cost_mean_usd']:.4f}/trial "
+        f"(${b['cost_sum_usd']:.4f} total) · {b['duration_mean_s']:.0f}s · {b['turns_mean']:.1f} turns "
+        f"· reward {b['reward_mean']}",
+        f"with-arm : {w['n']} trial(s) · ${w['cost_mean_usd']:.4f}/trial "
+        f"(${w['cost_sum_usd']:.4f} total) · {w['duration_mean_s']:.0f}s · {w['turns_mean']:.1f} turns "
+        f"· reward {w['reward_mean']}",
+    ]
+    lines += [
+        f"cost     : {r['delta_usd']:+.4f} USD/trial ({r['delta_pct']:+.1f}%)  "
+        f"95% CI [{r['ci_usd'][0]:+.4f}, {r['ci_usd'][1]:+.4f}]",
+        f"DECISION : {r['decision']} (per trial)",
+    ]
+    for reason in r["refused_reasons"]:
+        lines.append(f"  refused: {reason}")
+
+    ps = r["per_success"]
+    bs = "n/a" if ps["baseline_usd"] is None else f"${ps['baseline_usd']:.4f}"
+    ws = "n/a" if ps["with_usd"] is None else f"${ps['with_usd']:.4f}"
+    lines.append(
+        f"per-succ : {bs} → {ws} per success "
+        f"(passed {ps['baseline_successes']} → {ps['with_successes']})"
+    )
+    if ps["delta_usd"] is not None and ps["ci_usd"]:
+        lines.append(
+            f"           {ps['delta_usd']:+.4f} USD/success ({ps['delta_pct']:+.1f}%)  "
+            f"95% CI [{ps['ci_usd'][0]:+.4f}, {ps['ci_usd'][1]:+.4f}]"
+        )
+    lines.append(f"DECISION : {ps['decision']} (per success)")
+    for reason in ps["refused_reasons"]:
+        lines.append(f"  refused: {reason}")
+    if r["reward"]["regressed"]:
+        lines.append(
+            f"  WARNING: reward regressed ({r['reward']['baseline']} → {r['reward']['with']}) — "
+            "a cheaper arm that solves less is not a saving; do not quote the cost delta alone"
+        )
+    return "\n".join(lines)
+
+
+def main(argv: list[str]) -> int:
+    args = [a for a in argv[1:] if not a.startswith("--")]
+    if len(args) < 2:
+        print("usage: cost_delta.py <baseline-job-dir> <with-job-dir> [--json]", file=sys.stderr)
+        return 2
+
+    def opt(name, default):
+        for i, a in enumerate(argv):
+            if a == name and i + 1 < len(argv):
+                return int(argv[i + 1])
+        return default
+
+    try:
+        def strval(name, default=""):
+            for i, a in enumerate(argv):
+                if a == name and i + 1 < len(argv):
+                    return argv[i + 1]
+            return default
+
+        r = compare(
+            Path(args[0]),
+            Path(args[1]),
+            min_n=opt("--min-n", COST_MIN_N),
+            resamples=opt("--resamples", RESAMPLES),
+            allow_cross=[k.strip() for k in strval("--allow-cross").split(",") if k.strip()],
+            allow_unstamped="--allow-unstamped" in argv,
+        )
+    except ValueError as e:
+        print(str(e), file=sys.stderr)
+        return 3
+
+    print(json.dumps(r) if "--json" in argv[1:] else pretty(r, brief="--brief" in argv[1:]))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv))