@wrangled/cli 0.3.22 → 0.3.24

package/dist/guide.js

@@ -20,6 +20,10 @@ Run \`${cli} guide\` first. It says, in plain words, what this workspace has. It
 
 Before writing new ads or a campaign plan, run \`${cli} brief\`. In one read it gives the offer (from the confirmed offers and funnels map, else the brand notes and landing pages, or it says the offer is unknown), the brand voice, the best ads by spend and by results with their words and a piece of their video transcript, the ads to avoid repeating, the pages those ads send people to, the funnel and top campaigns, who the top Facebook campaigns were aimed at, and what is missing. \`--from\` and \`--to\` pick the dates (default: the last 30 days), \`--channel\` or \`--campaign <id>\` narrow it, and \`--top <n>\` sets how many of each list it shows (default 5, at most 20). It prints plain Markdown to paste into a prompt; \`--format json\` gives the same brief as data. It says what it cut to stay short.
 
+## Which Facebook ads won: winners and flops
+
+\`${cli} winners\` says which Facebook ads won and which flopped, for one Monday-to-Sunday week and for the 30 days ending that Sunday, judged on real sales: each sale goes to the Facebook ad that first brought the buyer in, and payments typed into GoHighLevel by hand never count toward an ad. An ad with no sales that has spent less than twice the account's cost per sale is too new to judge on sales, so it is judged on cost per sign-up and listed apart. Each ad says in a sentence why it is on its list. \`--week <YYYY-MM-DD>\` picks any day of the week (default: the latest complete week), \`--top <n>\` how many ads each list shows (default 5, at most 20), and \`--format json\` gives the report as data.
+
 ## Content: search before SQL
 
 To find content, such as "the 10 emails with the best click rate", "our 5 best videos by return on ad spend" or "every landing page that mentions the guarantee", use \`${cli} content search\` before writing SQL on the \`content_\` tables.
@@ -35,9 +39,10 @@ To read the whole library with your own tools, such as "every email we have" or
 
 ## Know-how library: Creative Cartel
 
-For what has worked for others, such as "how do members test new hooks" or "what do people say about advertorials", search the know-how library: \`${cli} knowledge search "<question>"\`. Today it holds Creative Cartel's library (\`creative-cartel\`) and the team's own Exodus copy and ad frameworks (\`exodus\`), both copied into this server every Monday afternoon. It works only when the key has library access (\`${cli} whoami\` says); without it the answer is \`knowledge_not_allowed\`. It finds snippets by their words and by their meaning, and returns them short, never whole documents: each at most 280 characters, with its title, link, kind, caveat, evidence and the day its library was copied in. Evidence \`reported member statement\` is a member's claim, not a checked result; weigh it that way and say so when you use it. Evidence \`team method\` is the team's own framework: reason with it, never paste it into client work. The library writes no answer for you: read the snippets and decide. Use your own words: saving text that repeats 12 words in a row from anything in the library is refused with \`copies_library_text\`.
+For what has worked for others, such as "how do members test new hooks" or "what do people say about advertorials", search the know-how library: \`${cli} knowledge search "<question>"\`. Today it holds Creative Cartel's library (\`creative-cartel\`) the team's own Exodus copy and ad frameworks (\`exodus\`), and Meta's advertising standards (\`meta-policy\`), all copied into this server every Monday afternoon. It works only when the key has library access (\`${cli} whoami\` says); without it the answer is \`knowledge_not_allowed\`. It finds snippets by their words and by their meaning, and returns them short, never whole documents: each at most 280 characters, with its title, link, kind, caveat, evidence and the day its library was copied in. Evidence \`reported member statement\` is a member's claim, not a checked result; weigh it that way and say so when you use it. Evidence \`team method\` is the team's own framework: reason with it, never paste it into client work. The library writes no answer for you: read the snippets and decide. Use your own words: saving text that repeats 12 words in a row from anything in the library is refused with \`copies_library_text\`.
 
 - The question is 1 to 4000 characters. \`--limit\` is 1 to 12 (default 5). \`--kind\` (message, transcript, document, note) keeps one kind of source. \`--library <slug>\` keeps one library.
+- \`--library meta-policy\` searches Meta's own advertising standards, copied from Meta's site every Monday. Before you write or judge an ad, check its claims there, such as income, health or before-and-after claims: \`${cli} knowledge search --library meta-policy "income claims"\`. Its evidence is \`meta_published_policy\`, Meta's published rule; Meta's reviewers still decide each ad.
 - \`knowledge_not_synced\` means the library has not been copied into this server yet, so nothing was searched; the weekly sync fills it.
 
 ## Query workflow: schema first
@@ -190,7 +195,7 @@ Tag ads, emails, texts and pages only with the words \`${cli} grammar\` lists (\
 
 ## What customers keep saying
 
-Every hour an AI reads the workspace's new Facebook and Instagram comments, Google reviews and Reddit posts and comments, pulls out openers (lines that would open an ad), objections (why people hesitate) and phrases (customers' own words for their problem or result), and groups the ones that say the same thing. Names, emails, phone numbers and handles are taken out first. \`${cli} voice\` lists the groups, most common first (\`--kind opener|objection|phrase\`, \`--limit <n>\` up to 200, \`--json\`), each with how many rows said it, their likes, and up to 3 real quotes. In \`--json\`, each quote's \`{kind, ref, quote}\` goes straight into a creative's \`researchSources\`, and a group itself can be cited as its \`ref\`, \`voice_clusters:<id>\`. The same rows are in the \`voice_clusters\` and \`voice_insights\` tables.
+Every hour an AI reads what is new in the workspace's Facebook and Instagram comments, Google reviews, Reddit posts and comments, recorded calls and meetings (Fathom, Zoom, CallRail) and chats (GoHighLevel texts and chat, Closebot), pulls out openers (lines that would open an ad), objections (why people hesitate), pains (problems they live with), desires (results they want) and phrases (exact wording worth copying into an ad), and groups the ones that say the same thing. From calls and chats only the customer's words count, and names, emails, phone numbers and handles are taken out first. \`${cli} voice\` lists the groups, most common first (\`--kind opener|objection|pain|desire|phrase\`, \`--limit <n>\` up to 200, \`--json\`), each with how many sources said it (a conversation counts once), their likes, and up to 3 real quotes. \`${cli} voice --since <YYYY-MM-DD> --sort useful\` lists single lines said since that day instead, the most usable in an ad first: specific, first-person, recent and often said; \`--json\` adds each line's link. In \`--json\`, each quote's \`kind\` (a line's \`source\`) and \`ref\` go straight into a creative's \`researchSources\`, and a group itself can be cited as its \`ref\`, \`voice_clusters:<id>\`. The same rows are in the \`voice_clusters\` and \`voice_insights\` tables.
 
 ## The weekly story
 
@@ -242,6 +247,7 @@ ${cli} whoami                       show the current workspace and library acces
 ${cli} knowledge search "<question>" [--limit <n>] [--kind <kind>]   short, cited snippets from the Creative Cartel know-how library (keys with library access)
 ${cli} guide                        what this workspace's data holds, how fresh it is, and what is missing: read it first
 ${cli} brief [--from <date>] [--to <date>] [--channel <channel>] [--campaign <id>] [--top <n>]   the offer, voice, best and worst ads, pages, results and audience: start here before writing ads
+${cli} winners [--week <date>] [--top <n>]   which Facebook ads won and flopped this week and over 30 days, judged on first-touch sales
 ${cli} warehouse schema [--summary] [--table <name> ...] [--search <words>]   list tables and columns; --summary for one row per table
 ${cli} warehouse query "<sql>"      run one read-only query (or --file <path>, or stdin); --out <file.csv | file.ndjson | file.json> saves the rows; add --all for every row
 ${cli} funnel [--from <date>] [--to <date>] [--grain day|week|month]   spend, leads, calls and sales per period, with totals

package/dist/main.js

@@ -10,11 +10,13 @@ import { firstSymlink, inRunFolder, writeFileAtomic } from "./files.js";
 import { guide, skill } from "./guide.js";
 import { PULL_STATE_FILE, pullContent } from "./pull.js";
 import { AGENTS_MD_MARKER, metaAdsAgentsMd, PLAYBOOK_CHOICES, playbooks } from "./playbooks.js";
+import { OPENROUTER_KEY_ENV } from "./runner/codex.js";
 import { AGENT_ENV, HARNESS_COMMAND_ENV, RUNNER_TOKEN_ENV, runLoop, runOnce } from "./runner/run.js";
 import { SKILL_NAME_PATTERN } from "./skill-zip.js";
 import { installSkillFolder, operatorConnection, packSkill, parseSkillRef, readSkillFolder, unpackSkill, } from "./skills.js";
 const DEFAULT_WAIT_SECONDS = 300;
 const SERVER_WAIT_CAP = 50;
+const VOICE_USAGE = "voice [--kind opener|objection|pain|desire|phrase] [--since <YYYY-MM-DD>] [--sort common|useful] [--limit <n>] [--json]";
 const commands = {
     login: {
         usage: `login   (paste the key when asked, pipe it on stdin, or set ${envPrefix}_API_KEY)`,
@@ -608,6 +610,21 @@ const commands = {
             process.stdout.write(brief.text);
         },
     },
+    winners: {
+        usage: "winners [--week <YYYY-MM-DD>] [--top <n>]   (which Facebook ads won and flopped, judged on real first-touch sales)",
+        async run(args, options) {
+            if (args.length > 0)
+                throw new CliError("Usage: winners [--week <date>] [--top <n>]", exitCodes.usage);
+            const query = new URLSearchParams();
+            for (const [name, value] of Object.entries({ week: options.week, top: options.top }))
+                if (value !== undefined)
+                    query.set(name, value);
+            const report = await api(await resolveConnection(options.apiUrl), `/api/v1/winners${query.size ? `?${query}` : ""}`);
+            if (options.format === "json")
+                return printJson(report);
+            process.stdout.write(winnersText(report));
+        },
+    },
     examples: {
         usage: "examples   (plain questions with SQL that answers them, for the tables this workspace has)",
         async run(args, options) {
@@ -1125,21 +1142,30 @@ const commands = {
         },
     },
     voice: {
-        usage: "voice [--kind opener|objection|phrase] [--limit <n>] [--json]   what customers keep saying, most common first",
+        usage: `${VOICE_USAGE}   what customers keep saying, most common first, or with --sort useful the most usable single lines`,
         async run(args, options) {
             if (args.length > 0)
-                throw new CliError("Usage: voice [--kind opener|objection|phrase] [--limit <n>] [--json]", exitCodes.usage);
+                throw new CliError(`Usage: ${VOICE_USAGE}`, exitCodes.usage);
             const query = new URLSearchParams();
             if (options.kind)
                 query.set("kind", options.kind);
+            if (options.since)
+                query.set("since", options.since);
+            if (options.sort)
+                query.set("sort", options.sort);
             if (options.limit)
                 query.set("limit", options.limit);
-            const { clusters } = await api(await resolveConnection(options.apiUrl), `${VOICE_PATH}${query.size ? `?${query}` : ""}`);
+            const read = await api(await resolveConnection(options.apiUrl), `${VOICE_PATH}${query.size ? `?${query}` : ""}`);
             if (options.format === "json")
-                return printJson({ clusters });
-            if (clusters.length === 0)
+                return printJson(read);
+            if ("lines" in read) {
+                if (read.lines.length === 0)
+                    return console.log("Nothing said in that time yet.");
+                return console.log(renderTable(VOICE_LINE_COLUMNS, read.lines.map(voiceLineRow)));
+            }
+            if (read.clusters.length === 0)
                 return console.log("Nothing grouped yet.");
-            console.log(renderTable(VOICE_COLUMNS, clusters.map(voiceRow)));
+            console.log(renderTable(VOICE_COLUMNS, read.clusters.map(voiceRow)));
         },
     },
     "ads tags": {
@@ -1199,6 +1225,8 @@ const commands = {
                 console.log(`Failed for ${failure.adIds.join(", ")}: ${failure.error}`);
             if (result.stopped === "tool_cap")
                 console.log("Stopped early: the run's tool spending cap was reached.");
+            if (result.nothingToRead > 0)
+                console.log(`Left out ${plural(result.nothingToRead)} with nothing to read: no copy, headline, transcript or picture description.`);
         },
     },
     "content tags": {
@@ -1683,6 +1711,7 @@ async function withRunner(options, work) {
             apiUrl: resolveApiUrl(options.apiUrl, null),
             token,
             harnessCommand: process.env[HARNESS_COMMAND_ENV] || undefined,
+            openRouterKey: process.env[OPENROUTER_KEY_ENV] || undefined,
             maxSeconds: options.maxSeconds,
             signal: stop.signal,
             say: (line) => console.log(line),
@@ -1759,6 +1788,59 @@ const CAMPAIGN_NUMBERS = [
     "roas_first_touch",
     "roas_last_touch",
 ];
+function money(value, currency) {
+    try {
+        return new Intl.NumberFormat("en-US", { style: "currency", currency }).format(value);
+    }
+    catch {
+        return `${value.toFixed(2)} ${currency}`;
+    }
+}
+const counted = (n, one) => `${n} ${n === 1 ? one : `${one}s`}`;
+/** An ad's line: name, SKU and campaign, its numbers, its tags, then why it is on the list. */
+function winnersAdLines(ad, currency) {
+    const n = ad.numbers;
+    const onSales = ad.judgedOn === "sales";
+    const cost = onSales
+        ? n.costPerSale === null ? "no sales" : `${money(n.costPerSale, currency)} per sale`
+        : n.costPerSignup === null ? "no sign-ups" : `${money(n.costPerSignup, currency)} per sign-up`;
+    const name = [ad.name ?? `Ad ${ad.adId}`, ad.sku && `[${ad.sku}]`, ad.campaignName && `(${ad.campaignName})`].filter(Boolean).join(" ");
+    const tags = Object.entries(ad.tags).map(([field, value]) => `${field}: ${value}`).join(", ");
+    const count = onSales ? n.sales !== null && counted(n.sales, "sale") : n.signups !== null && counted(n.signups, "sign-up");
+    return `- ${name}: spent ${money(n.spend, currency)}, ${count ? `${count}, ` : ""}${cost}${tags ? `. Tags: ${tags}` : ""}\n  ${ad.pickedBecause}\n`;
+}
+function winnersWindowText(title, window) {
+    const head = `\n## ${title}, ${window.from} to ${window.to}\n`;
+    if (window.sections.length === 0)
+        return `${head}\nNo Facebook ad spent in these dates.\n`;
+    return (head +
+        window.sections
+            .map((s) => {
+            const list = (heading, ads) => `\n${heading}\n${ads.length === 0 ? "- none\n" : ads.map((ad) => winnersAdLines(ad, s.currency)).join("")}`;
+            const account = s.salesSource === "not_connected"
+                ? "sales are not connected"
+                : s.accountCostPerSale === null
+                    ? "no sale credited to a Facebook ad in these dates"
+                    : `the account paid ${money(s.accountCostPerSale, s.currency)} per sale`;
+            const tests = s.tests.map((t) => `- Test ${t.testId}${t.hypothesisId ? ` (idea ${t.hypothesisId})` : ""}: ${counted(t.ads.length, "ad")}, spent ${money(t.spend, s.currency)}, ${counted(t.sales, "sale")}` +
+                `${t.costPerSale === null ? "" : `, ${money(t.costPerSale, s.currency)} per sale`}\n`);
+            return (`\nIn ${s.currency}: ${account}. ${s.adsJudgedOnSales} judged on sales, ${s.adsJudgedOnSignups} judged on sign-ups.\n` +
+                s.notes.map((note) => `${note}\n`).join("") +
+                list("Winners", s.winners) +
+                list("Flops", s.flops) +
+                list(`${s.onSignups.title}: best`, s.onSignups.best) +
+                list(`${s.onSignups.title}: worst`, s.onSignups.worst) +
+                (tests.length > 0 ? `\nAdzara tests\n${tests.join("")}` : ""));
+        })
+            .join(""));
+}
+/** The winners report as plain Markdown: what a sale means, then this week and the last 30 days, each currency apart. */
+function winnersText(report) {
+    return (`# Winners and flops for ${report.workspace}\n\n` +
+        `A sale means ${report.saleMeans}. Each sale goes to the Facebook ad that first brought the buyer in; payments typed into GoHighLevel by hand never count toward an ad.\n` +
+        winnersWindowText("This week", report.week) +
+        winnersWindowText("Last 30 days", report.last30Days));
+}
 function reportQuery(options, extra) {
     const query = new URLSearchParams();
     for (const [name, value] of Object.entries({ from: options.from, to: options.to, ...extra }))
@@ -1798,6 +1880,7 @@ const VOICE_SUGGESTIONS_PATH = "/api/v1/voice-lists/suggestions";
 const FACT_COLUMNS = ["id", "kind", "subject", "key", "status", "summary"];
 const VOICE_PATH = "/api/v1/voice";
 const VOICE_COLUMNS = ["ref", "kind", "said by", "likes", "label", "top quote"];
+const VOICE_LINE_COLUMNS = ["date", "kind", "source", "quote", "ref"];
 function knowledgeBlock(r, i) {
     const lines = [`${i + 1}. ${r.title}`, `   ${r.kind}, ${r.evidence.replaceAll("_", " ")} (${r.source}${r.syncedAt ? `, copied in ${r.syncedAt.slice(0, 10)}` : ""})`];
     if (r.url)
@@ -1818,6 +1901,9 @@ function contentTagRow(t) {
 function voiceRow(c) {
     return { ref: c.ref, kind: c.kind, "said by": c.supportCount, likes: c.likeSum, label: excerpt(c.label, 60), "top quote": c.evidence[0] ? excerpt(c.evidence[0].quote, 60) : null };
 }
+function voiceLineRow(l) {
+    return { date: l.date.slice(0, 10), kind: l.kind, source: l.source, quote: excerpt(l.quote, 80), ref: l.ref };
+}
 function factRow(f) {
     return { id: f.id, kind: f.kind, subject: `${f.subjectType}:${f.subjectRef}`, key: f.key, status: f.status, summary: excerpt(f.summary, 80) };
 }
@@ -1901,7 +1987,8 @@ Options:
   --channel <channel>   only this channel's campaigns (and, with brief, its ads): meta, google, affiliate, a utm_source, unknown or recorded_in_crm;
                         for content track-link, where the link goes: facebook, email, sms or web (default)
   --campaign <id>       the campaign content track-link puts in utm_campaign; with brief, the one campaign it covers
-  --top <n>             how many best ads, worst ads, pages and campaigns brief shows (default: 5, at most 20)
+  --top <n>             how many best ads, worst ads, pages and campaigns brief shows, or ads in each winners list (default: 5, at most 20)
+  --week <YYYY-MM-DD>   any day of the week winners covers, Monday to Sunday (default: the latest complete week)
   --body, --html <text | @file>   the words, or the full HTML, content create, capture and edit save
   --file <path>         an image, video or audio file, up to 50 MB, content create, capture and edit save; repeat for several
   --transcript <text | @file>   what is said in the first video or audio file content capture saves, so none is paid for
@@ -1934,7 +2021,7 @@ Options:
   --order desc|asc      content search's sort direction (default: desc, best or newest first)
   --limit <n>           how many results content search, creatives list, campaigns, voice, ads tags, content tags and content to-check
                         return, or how many ads ads autotag tags and how many items content autotag tags
-  --library <slug>      with knowledge search, only this library: creative-cartel or exodus
+  --library <slug>      with knowledge search, only this library: creative-cartel, exodus or meta-policy
   --ad <id>             with ads tags, only this Meta ad's tags; with ads autotag, only this ad
   --kind email|text|page   with content tags and content autotag, only that kind (content autotag takes a comma list)
   --max-seconds <n>     stop a runner's run after this many seconds, if that comes before its deadline
@@ -1987,11 +2074,13 @@ function parseOptions(argv) {
             force: { type: "boolean", default: false },
             yes: { type: "boolean", default: false },
             from: { type: "string" },
+            since: { type: "string" },
             to: { type: "string" },
             grain: { type: "string" },
             channel: { type: "string" },
             campaign: { type: "string" },
             top: { type: "string" },
+            week: { type: "string" },
             name: { type: "string" },
             subject: { type: "string" },
             preview: { type: "string" },
@@ -2082,11 +2171,13 @@ function parseOptions(argv) {
             force: values.force,
             yes: values.yes,
             from: values.from,
+            since: values.since,
             to: values.to,
             grain: values.grain,
             channel: values.channel,
             campaign: values.campaign,
             top: values.top,
+            week: values.week,
             name: values.name,
             subject: values.subject,
             preview: values.preview,

package/dist/playbooks.js

@@ -864,11 +864,11 @@ ${cli} ads autotag --limit ${BACK_TAG_BATCH} --json > batch.json
 
 Each call takes the ${BACK_TAG_BATCH} biggest-spending ads with no tags yet, so the next call moves on to the next ones. Open \`batch.json\` after each one. An ad Jev says is not marketing (\`not_marketing\`), or can't tell about (\`gate_unclear\`), is listed under \`skipped\` but was saved with only that answer, so it counts as progress. Run it again until one of these happens:
 
-- A batch saved nothing new: \`proposed\` is empty, and every ad under \`skipped\` is \`no_content\` (nothing to read: no copy, transcript or picture description) or \`not_found\`, with the rest under \`failed\`. The same ads would come back next time, so asking again won't help.
+- A batch saved nothing new: \`proposed\` is empty, and every ad under \`skipped\` is \`no_content\` (nothing to read: no copy, transcript or picture description) or \`not_found\`, with the rest under \`failed\`. Asking again won't help.
 - You have run ${BACK_TAG_MAX_BATCHES} batches, or the person's spending limit would be passed.
 - \`stopped\` is \`tool_cap\`: the run's tool spending cap was reached.
 
-Keep a running count of ads proposed, ads skipped (and why), failures and \`costUsd\` from each batch.
+Keep a running count of ads proposed, ads skipped (and why), failures and \`costUsd\` from each batch. \`nothingToRead\` counts the untagged ads the tagger left out because they have nothing to read; keep the last batch's number.
 
 ## 3. Report
 
@@ -876,7 +876,7 @@ Run the coverage query from step 1 again, then finish with a short report in pla
 
 - How many ads you tagged this run and what it cost.
 - How many ads that spent anything now have tags, and what share of the spend that is (\`pct_spend_tagged\`); what share of spend a person has confirmed (\`pct_spend_confirmed\`). Give each currency's shares on their own; never add them together.
-- How many ads are still untagged and why, if the batches said (nothing to read, failures).
+- How many ads are still untagged and why, if the batches said: \`nothingToRead\` ads have nothing to read (no copy, headline, transcript or picture description), plus any failures.
 - What to do next: open the Ad tags page in the ${brand.name} web app, press Make the list (or Make a new list) under Check these first, and confirm, fix or skip the ads that carry the top 20% of spend. Everything else stays as the tagger's best guess, marked not yet confirmed.
 
 ## Commands

package/dist/runner/claim.js

@@ -1,6 +1,7 @@
 import { CliError, exitCodes } from "../api.js";
 import { SKILL_NAME_PATTERN } from "../skill-zip.js";
 export const harnessNames = ["claude", "codex"];
+const aiBillings = ["run_key", "shared_key", "subscription"];
 const isObject = (value) => typeof value === "object" && value !== null && !Array.isArray(value);
 /** The claim response as the runner uses it. Never echoes the body: it holds the run key. */
 export function parseClaim(body) {
@@ -28,6 +29,12 @@ export function parseClaim(body) {
         throw bad("deadline");
     if (body.openRouterKey !== undefined && (typeof body.openRouterKey !== "string" || body.openRouterKey.length === 0))
         throw bad("openRouterKey");
+    // A server from before `aiBilling` says run_key only by sending a key.
+    const aiBilling = body.aiBilling ?? (body.openRouterKey === undefined ? "subscription" : "run_key");
+    if (!aiBillings.includes(aiBilling))
+        throw bad("aiBilling");
+    if ((aiBilling === "run_key") !== (typeof body.openRouterKey === "string"))
+        throw bad("openRouterKey");
     return {
         runId: run.id,
         runKey: body.runKey,
@@ -37,6 +44,6 @@ export function parseClaim(body) {
         prompt: body.prompt,
         aiCapUsd: caps.aiCents / 100,
         deadline,
-        billing: typeof body.openRouterKey === "string" ? { kind: "openrouter", key: body.openRouterKey } : { kind: "subscription" },
+        billing: typeof body.openRouterKey === "string" ? { kind: "openrouter", key: body.openRouterKey } : { kind: aiBilling },
     };
 }

package/dist/runner/run.js

@@ -8,6 +8,7 @@ import { brand, envPrefix } from "../brand.js";
 import { RUN_ROOT_ENV } from "../files.js";
 import { installSkillFolder, unpackSkill } from "../skills.js";
 import { parseClaim } from "./claim.js";
+import { OPENROUTER_KEY_ENV } from "./codex.js";
 import { MAX_TURNS } from "./harness.js";
 import { commandHarness, harnesses } from "./harnesses.js";
 import { LogTail, redact, RUN_ERROR_MAX_BYTES, tailBytes } from "./redact.js";
@@ -68,7 +69,7 @@ export async function runOnce(options) {
     if (answer === "")
         return { kind: "idle" };
     const claim = parseClaim(JSON.parse(answer));
-    const secrets = [claim.runKey, options.token, ...(claim.billing.kind === "openrouter" ? [claim.billing.key] : [])];
+    const secrets = [claim.runKey, options.token, ...(claim.billing.kind === "openrouter" ? [claim.billing.key] : []), ...(options.openRouterKey ? [options.openRouterKey] : [])];
     const say = (line) => options.say(redact(line, secrets));
     say(`Claimed run ${claim.runId}: ${claim.skill.name} version ${claim.skill.version} with ${claim.harness}.`);
     const stop = new AbortController();
@@ -90,6 +91,9 @@ export async function runOnce(options) {
         dir = await mkdtemp(join(tmpdir(), `${brand.slug}-run-`));
         const result = await execute(claim, dir, options, stop.signal, (text) => log.append(text));
         report = finishReport(result, stop.signal.reason);
+        const aiCents = reportedAiCents(claim.billing, result.tokens);
+        if (aiCents !== undefined)
+            report.aiCents = aiCents;
     }
     catch (error) {
         report = { status: "failed", error: stop.signal.aborted ? stopMessages[stop.signal.reason] : `The runner failed: ${error.message}` };
@@ -143,6 +147,7 @@ function finishReport(result, stopped) {
 }
 /** Puts the skill, the run key's login and a `wr` in `dir`, then runs the harness there. */
 async function execute(claim, dir, options, signal, log) {
+    const billing = harnessBilling(claim, options.openRouterKey);
     const skillsDir = join(dir, claim.harness === "claude" ? ".claude" : ".codex", "skills");
     await installSkillFolder(skillsDir, claim.skill.name, unpackSkill(claim.skill, await getSigned(claim.skill.url)), []);
     const configHome = join(dir, ".config");
@@ -158,7 +163,7 @@ async function execute(claim, dir, options, signal, log) {
         skill: claim.skill.name,
         prompt: runPrompt(claim.prompt),
         model: claim.model,
-        billing: claim.billing,
+        billing,
         aiCapUsd: claim.aiCapUsd,
         maxTurns: MAX_TURNS,
         env: harnessEnv(process.env, { dir, binDir, configHome, apiUrl: options.apiUrl }),
@@ -166,6 +171,30 @@ async function execute(claim, dir, options, signal, log) {
         log,
     });
 }
+/**
+ * How the harness pays. A shared_key run pays OpenRouter with the runner's own key, and only Claude Code, whose budget
+ * stops it at the AI cap, may run that way: nothing would hold a Codex run to its cap on a key shared by every run.
+ */
+export function harnessBilling(claim, openRouterKey) {
+    if (claim.billing.kind !== "shared_key")
+        return claim.billing;
+    if (!openRouterKey) {
+        throw new Error(`This run pays for AI with the runner's own OpenRouter key, but ${OPENROUTER_KEY_ENV} is not set on this runner. Set it, or set OPENROUTER_PROVISIONING_KEY on the server to an OpenRouter management key so each run gets its own key.`);
+    }
+    if (claim.harness === "codex") {
+        throw new Error(`A Codex run cannot keep to its AI cap on the runner's shared OpenRouter key. Set OPENROUTER_PROVISIONING_KEY on the server to an OpenRouter management key so each run gets its own capped key.`);
+    }
+    return { kind: "openrouter", key: openRouterKey };
+}
+/**
+ * The AI spend a finish report carries: only a shared_key run's, whose spend the server cannot read from OpenRouter
+ * itself, from the harness's own estimate. A zero estimate may mean the harness could not price the model, so it is not
+ * sent; a run without a report counts at its full cap.
+ */
+export function reportedAiCents(billing, tokens) {
+    const costUsd = tokens?.costUsd;
+    return billing.kind === "shared_key" && costUsd !== undefined && costUsd > 0 ? Math.ceil(costUsd * 100) : undefined;
+}
 /** Beats now (moving the run to running) and every minute; a cancel, or a server that no longer knows the run, stops it. */
 function startHeartbeat(connection, runId, cancel, say) {
     const beat = async () => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wrangled/cli",
-  "version": "0.3.22",
+  "version": "0.3.24",
   "description": "Command-line access to your marketing data warehouse, for people and coding agents",
   "type": "module",
   "bin": {