@koda-sl/baker-cli 0.536.0 → 0.538.0-dev.fc2958a1d

package/dist/chunk-2EVHGBQY.js

@@ -0,0 +1,319 @@
+import {
+  RESEARCH_GOOGLE_ADS_MAX_COPY
+} from "./chunk-XXS6SUVM.js";
+
+// src/commands/research/hints.ts
+var AD_COPY_ROUTE = 'This returns competing domains and their SERP economics \u2014 no ad copy, no headlines, no creative. For a competitor\'s Google ads and their copy: `baker research google-ads <competitor.com> --copy` (Google\'s Ads Transparency Center: every ad, how long it has run, and the text ads\' headline and description). For Meta and LinkedIn: `baker winning-ads advertisers "<brand>"` \u2192 `baker winning-ads search "<brief>" --advertiser-id <id>` \u2192 `baker winning-ads content <adId>` (`primary_text` / `headline` / `cta`, plus the spoken transcript and on-screen text for video). Do not keep querying this command for copy.';
+function noSerpRowsHints(rowCount, queryContext) {
+  if (rowCount > 0) {
+    return [];
+  }
+  const defaulted = queryContext.location_is_default || queryContext.language_is_default;
+  if (defaulted) {
+    return [
+      `No rows for ${queryContext.location} / ${queryContext.language} \u2014 the market you did not set. Set --location and --language for the client's actual market and re-run once (\`baker research countries\` / \`baker research languages\` for the accepted codes) before you touch the keyword.`
+    ];
+  }
+  return [
+    `No rows for ${queryContext.location} / ${queryContext.language}, which you set explicitly. This term has no third-party SERP coverage in that market \u2014 move on rather than re-querying it. For demand signal try \`baker research autocomplete\`; for a competitor's actual terms, \`baker research keywords-for-site <domain>\`.`
+  ];
+}
+function advertisersHints(rowCount, queryContext) {
+  return [...noSerpRowsHints(rowCount, queryContext), AD_COPY_ROUTE];
+}
+function keywordsForSiteHints(rowCount, queryContext, typeFilter) {
+  if (rowCount > 0) {
+    return [];
+  }
+  if (typeFilter === "paid" || typeFilter === "organic") {
+    return [
+      `No rows with --type ${typeFilter} in ${queryContext.location} / ${queryContext.language}. A domain with zero ${typeFilter} keywords in a market is itself the finding \u2014 record it and move on. Confirm once with --type all (nothing at all there means the domain has no SERP coverage, not that it has no ads) instead of re-running with a different --sort or --limit.`
+    ];
+  }
+  return noSerpRowsHints(rowCount, queryContext);
+}
+var THIN_ANSWER_CHARS = 600;
+var THIN_SOURCE_COUNT = 2;
+var ESCALATION = {
+  medium: "`medium` is a quick lookup, not a search. Re-ask the SAME question once with `--depth high` before concluding the web has nothing on it.",
+  high: "`high` already thought hard about this. `--depth xhigh` is the only step left and it can run 15 minutes \u2014 take it only if this answer blocks a deliverable; otherwise record the gap and move on."
+};
+function fetchHints({
+  truncated,
+  contentChars,
+  shownChars
+}) {
+  const hints = [];
+  if (shownChars < contentChars) {
+    hints.push(
+      `Showing the first ${shownChars.toLocaleString("en-US")} of ${contentChars.toLocaleString("en-US")} characters. Re-run the same command with \`--full\` for the whole page \u2014 the read is cached for 24h, so it costs no extra time and no extra credit.`
+    );
+  }
+  if (truncated) {
+    hints.push(
+      "This page is longer than one read returns, so the end of it is missing. Don't present this as the complete page; if the part you need is further down, fetch the deeper URL directly."
+    );
+  }
+  return hints;
+}
+function webDepthHints({
+  answer,
+  sourceCount,
+  depth
+}) {
+  const thin = answer.length < THIN_ANSWER_CHARS || sourceCount <= THIN_SOURCE_COUNT;
+  const escalation = ESCALATION[depth];
+  if (!thin || !escalation) {
+    return [];
+  }
+  return [`Thin answer (${answer.length} chars, ${sourceCount} source(s)). ${escalation}`];
+}
+function keywordMetricsHints({
+  rows,
+  asked,
+  clickstream,
+  queryContext
+}) {
+  if (rows.length === 0) {
+    return noSerpRowsHints(0, queryContext);
+  }
+  const provenance = "DataForSEO resells Google's Keyword Planner volume \u2014 planner_search_volume here is the SAME number `baker ads google keywords metrics` returns, not an independent reading of it. Quoting the two as agreeing sources is not corroboration." + (clickstream ? " clickstream_search_volume IS independent (DataForSEO's own panels); when the two disagree, say which one you are quoting." : " For a genuinely independent estimate of demand, re-run with --clickstream (costs 2x), or use first-party data: `baker gsc query` for what already reaches the site, or the account's own search terms.");
+  const hints = [provenance];
+  const empty = rows.filter((row) => row.search_volume === null).map((row) => row.keyword);
+  if (empty.length > 0) {
+    hints.push(
+      `${empty.length} of ${asked.length} keywords came back with no data in ${queryContext.location} / ${queryContext.language}: ${empty.join(", ")}. That is missing coverage, NOT zero demand \u2014 do not report them as unsearched, and do not re-query them reworded. \`baker research autocomplete\` on the head term is the way to find the phrasing that market uses.`
+    );
+  }
+  return hints;
+}
+var ARCHIVE_NOTE = "Read from the Reddit archive: the words are as posted, but points and comment counts on anything under ~36h old are as first captured \u2014 don't quote those numbers as current.";
+function redditSearchHints({
+  resultCount,
+  kind,
+  source
+}) {
+  const hints = [];
+  if (resultCount === 0) {
+    hints.push(
+      'Nothing matched. Reddit search matches words, not meaning: try the audience\'s own phrasing (one distinctive word, or a seed from `baker research autocomplete`), a longer --since window, or other subreddits \u2014 `baker research reddit subreddits "<topic>"` finds where the audience actually talks' + (kind === "posts" ? ", and try --kind comments \u2014 most of the voice is in the replies." : ".")
+    );
+  } else if (kind === "posts") {
+    hints.push(
+      "Titles rarely carry the voice. Read the promising threads with `baker research reddit thread <permalink>` for the replies, or search the replies directly with --kind comments."
+    );
+  }
+  if (source === "archive" && resultCount > 0) hints.push(ARCHIVE_NOTE);
+  return hints;
+}
+function redditThreadHints({
+  commentsShown,
+  commentsAsked,
+  commentCount,
+  source
+}) {
+  const hints = [];
+  if (commentsAsked > 0 && commentsShown === 0 && (commentCount ?? 0) > 0) {
+    hints.push(
+      `This thread has ${commentCount} comments but none could be read just now. Quote the post only, and don't describe the replies.`
+    );
+  }
+  if (source === "archive") hints.push(ARCHIVE_NOTE);
+  return hints;
+}
+var REDDIT_HOST = /(^|\.)reddit\.com$|^redd\.it$/i;
+function redditSubredditsHints({ resultCount }) {
+  if (resultCount === 0) {
+    return [
+      'No subreddits matched. Try the audience\'s role or the platform they work on ("amazon flex", "deliveroo", "nurses"), one word at a time, or a web search for `site:reddit.com <topic>`.'
+    ];
+  }
+  return [
+    'Members say how big a subreddit is, not how active it is. Before mining one, run `baker research reddit search "<word>" --subreddit <name> --since <a recent date>`: nothing back means it is quiet, so pick another. `restricted` can still be read; `private` cannot.'
+  ];
+}
+var REDDIT_LINK_REJECTED = {
+  action: "reject",
+  explanation: "Do not quote anything attributed to this link \u2014 it does not lead to those words. Find the real thread with `baker research reddit search`, and if you already wrote a quote from it, remove it."
+};
+function redditPageFix(url, code) {
+  let host;
+  try {
+    host = new URL(url).hostname;
+  } catch {
+    return null;
+  }
+  if (!REDDIT_HOST.test(host)) return null;
+  if (code === "NOT_FOUND") return REDDIT_LINK_REJECTED;
+  if (code !== "FORBIDDEN") return null;
+  return {
+    action: "retry_with_changes",
+    explanation: 'Read Reddit with its own commands: `baker research reddit search "<words>" --subreddit <name>` to find what people said, `baker research reddit thread <url>` for one thread.'
+  };
+}
+function googleAdsHints({
+  target,
+  adCount,
+  located,
+  textWithoutCopy,
+  copyAsked,
+  copyUnread,
+  visual
+}) {
+  if (adCount === 0) {
+    return [
+      `No Google ads point to ${target}${located ? " in that region" : ""} \u2014 that is the answer, not a failed call: this competitor is not advertising on Google there, or its ads point to another domain. ` + (located ? "Re-run once without --location to see every region; then move on. " : "Move on. ") + `Its paid search terms, if any: \`baker research keywords-for-site ${target} --type paid\`.`
+    ];
+  }
+  const hints = [];
+  if (!copyAsked && textWithoutCopy > 0) {
+    hints.push(
+      `${textWithoutCopy} text ads came back without their words. Re-run with --copy to read the headline and description of up to ${RESEARCH_GOOGLE_ADS_MAX_COPY} (cached, so only the reading is new).`
+    );
+  }
+  const pastCap = copyAsked ? textWithoutCopy - copyUnread : 0;
+  if (pastCap > 0) {
+    hints.push(
+      `${pastCap} text ads past the first ${RESEARCH_GOOGLE_ADS_MAX_COPY} were not read \u2014 one request reads at most ${RESEARCH_GOOGLE_ADS_MAX_COPY}. Narrow with --since, --location or --platform google_search and re-run for their words.`
+    );
+  }
+  if (visual > 0 || copyUnread > 0) {
+    const without = [
+      visual > 0 ? `${visual} image or video ads (Google gives no text for those)` : "",
+      copyUnread > 0 ? `${copyUnread} text ads whose image could not be read` : ""
+    ].filter(Boolean);
+    hints.push(
+      `No copy for ${without.join(" and ")}: show those by their \`transparency_url\`. Do not open the Transparency Center in agent-browser to read them \u2014 it is the browser load this command exists to avoid.`
+    );
+  }
+  return hints;
+}
+var LINKEDIN_PERSON_USE = "For ad and page messaging only; nothing is stored \u2014 quote their stated priorities into the creative brief for this account, and don't save the person to Knowledge or read them again to keep track of them.";
+function linkedinPersonHints({ isPrivate, postCount }) {
+  if (isPrivate) {
+    return [
+      "LinkedIn doesn't show this profile publicly, and reading it again won't change that. Find their own words elsewhere \u2014 a keynote, podcast or earnings call through `baker research transcript <url>`, or an interview through `baker research fetch <url>` \u2014 or write the brief from the account's company-level priorities."
+    ];
+  }
+  if (postCount === 0) {
+    return [
+      LINKEDIN_PERSON_USE,
+      "They have no recent posts, so their `about` is the only first-person line here. Their priorities are more often said aloud: a keynote, podcast or earnings call through `baker research transcript <url>`."
+    ];
+  }
+  return [LINKEDIN_PERSON_USE];
+}
+function transcriptHints({ excerpted, truncated }) {
+  const hints = [];
+  if (excerpted) {
+    hints.push(
+      "This is the opening excerpt. Re-run with `--out <file>` to write the whole text to a file and read the parts you need, or `--full` to have it all here \u2014 the transcript is cached for 7 days, so either costs nothing more."
+    );
+  }
+  if (truncated) {
+    hints.push(
+      "The recording is longer than one transcript carries, so the end of the recording is missing. Don't quote this as the whole talk."
+    );
+  }
+  return hints;
+}
+function thinProfileHint(domain, profile) {
+  const missing = [
+    profile.linkedinUrl ? null : "no LinkedIn page",
+    profile.hqCountry ? null : "no HQ country",
+    profile.employees ? null : "no employee band"
+  ].filter((gap) => gap !== null);
+  const tiny = profile.employees === "1-10";
+  if (missing.length < 2 && !tiny) return null;
+  const why = [...missing, ...tiny ? ["1-10 employees"] : []].join(", ");
+  return `This profile is thin (${why}). Check ${domain} is the company's main website \u2014 companies often run on another domain (e.g. factorial.es vs factorialhr.com) \u2014 before using these facts for tiering.`;
+}
+function companyHints({
+  domain,
+  profile,
+  hiring
+}) {
+  const hints = [];
+  if (!profile) {
+    hints.push(
+      "There's no company profile for this domain, only its hiring. Check it is the company's main website; otherwise take size, industry and location from its LinkedIn page or its own site."
+    );
+  } else {
+    const thin = thinProfileHint(domain, profile);
+    if (thin) hints.push(thin);
+  }
+  if (hiring.status === "unavailable" && hiring.reason === "not_configured") {
+    hints.push(
+      "Hiring signals aren't switched on for this workspace yet, so these are the firmographics alone. Score the account on them and leave hiring out of the reason."
+    );
+  }
+  if (hiring.status === "unavailable" && hiring.reason === "failed") {
+    hints.push(
+      "Hiring signals couldn't be read this time; don't retry in a loop. Use the firmographics and say hiring is unknown for this account."
+    );
+  }
+  if (hiring.status === "not_tracked") {
+    hints.push(
+      "No hiring data is kept for this company \u2014 that means unknown, not that it isn't hiring. Don't score it down for it."
+    );
+  }
+  return hints;
+}
+function industryList(industries) {
+  return industries.map(({ slug, companies }) => `${slug} (${companies.toLocaleString("en-US")})`).join(", ");
+}
+function relatedIndustriesHint(related) {
+  if (related.length === 0) return null;
+  return `Related industries not included (companies each adds with your other filters): ${industryList(related)} \u2014 add them with --industry a,b to widen.`;
+}
+function companySearchHints({
+  total,
+  returned,
+  page,
+  lastPage,
+  related
+}) {
+  const hints = [];
+  if (total === 0) {
+    hints.push(
+      "No company matches every filter. Widen one at a time \u2014 a neighbouring size band, a related industry, one more country."
+    );
+  } else if (page < lastPage && returned > 0) {
+    hints.push(
+      `${total} companies match and ${returned} came back (each one returned is billed). Use \`total\` to size the market; ask for \`--page ${page + 1}\` only if you need the next rows for the list.`
+    );
+  }
+  const widen = relatedIndustriesHint(related);
+  if (widen) hints.push(widen);
+  return hints;
+}
+function unknownIndustriesFix(industries) {
+  const misses = industries.map(
+    ({ input, closest }) => closest.length > 0 ? `"${input}" is not an industry. Closest (companies each matches with your other filters): ${industryList(closest)}.` : `"${input}" is not an industry, and none is named like it \u2014 try a broader word (software, logistics, retail).`
+  );
+  return {
+    action: "retry_with_changes",
+    explanation: `${misses.join(" ")} Pass one or several of these slugs with --industry a,b \u2014 several widen the search, any of them matching. Nothing was charged.`
+  };
+}
+
+export {
+  noSerpRowsHints,
+  advertisersHints,
+  keywordsForSiteHints,
+  fetchHints,
+  webDepthHints,
+  keywordMetricsHints,
+  redditSearchHints,
+  redditThreadHints,
+  redditSubredditsHints,
+  REDDIT_LINK_REJECTED,
+  redditPageFix,
+  googleAdsHints,
+  linkedinPersonHints,
+  transcriptHints,
+  companyHints,
+  companySearchHints,
+  unknownIndustriesFix
+};
+//# sourceMappingURL=chunk-2EVHGBQY.js.map

package/dist/chunk-2JFCDUZO.js

@@ -0,0 +1,156 @@
+import {
+  Phrases,
+  RULE_APPLIES,
+  RULE_KINDS,
+  RULE_STRENGTHS
+} from "./chunk-YAX5CGUY.js";
+import {
+  holdWorkspaceSpend
+} from "./chunk-W3PJ7V4W.js";
+import {
+  askAboutText
+} from "./chunk-KAUW3OFC.js";
+import {
+  CREATIVE_FEEDBACK_REASON_LABELS
+} from "./chunk-XXS6SUVM.js";
+
+// src/commands/creatives/rule-candidate.ts
+import { createHash } from "crypto";
+import { mkdir, readFile, writeFile } from "fs/promises";
+import path from "path";
+import { z } from "zod";
+var optionalPhrases = Phrases.optional().catch(void 0);
+var Draft = z.object({
+  because: z.string().trim().min(3).max(400),
+  text: z.string().trim().min(1).max(300),
+  strength: z.enum(RULE_STRENGTHS).catch("must"),
+  applies: z.enum(RULE_APPLIES).catch("any"),
+  language: z.string().trim().regex(/^[a-z]{2,3}$/).optional().catch(void 0),
+  kinds: z.array(z.enum(RULE_KINDS)).optional().catch(void 0),
+  platforms: optionalPhrases,
+  forbid: optionalPhrases,
+  require: optionalPhrases,
+  allow: optionalPhrases,
+  people: z.literal("real").optional().catch(void 0)
+});
+function parseDrafts(answer) {
+  const rules = answer?.rules;
+  if (!Array.isArray(rules)) return [];
+  return rules.slice(0, MAX_DRAFTS).flatMap((raw) => {
+    const draft = Draft.safeParse(raw);
+    if (!draft.success) return [];
+    const { because: _because, ...rule } = draft.data;
+    return [
+      Object.fromEntries(
+        Object.entries(rule).filter(([, v]) => v !== void 0 && !(Array.isArray(v) && v.length === 0))
+      )
+    ];
+  });
+}
+var MAX_DRAFTS = 3;
+function draftRulesPrompt(note, ad) {
+  const where = [ad.kind ? `a ${ad.kind} ad` : "an ad", ad.platform ? `for ${ad.platform}` : ""].filter(Boolean).join(" ");
+  return [
+    `A person rejected ${where} and said, in their own words:`,
+    `"""${note}"""`,
+    "",
+    "List only the standing rules this note states for EVERY future ad of the company (at most 3). Most notes state none: they ask for changes to this ad (fix a word, move the logo, speed up a scene, redo the pieces), and a change to this ad is never a rule, however it is phrased. A rule is standing only when the note itself says it holds beyond this ad (always, never, in every video, in these creatives, these mistakes keep happening, it is the format of the programme, the rule of the other pieces), or it corrects how the company or what it sells must always be named.",
+    "For each rule, quote in `because` the note's own words that make it hold beyond this ad; a rule you cannot quote that for is not a rule. Return JSON only:",
+    '{"rules": [{"because": string, "text": string, "strength": "never" | "must" | "prefer", "applies": "copy" | "picture" | "cast" | "logo" | "voice" | "sound" | "subtitles" | "any", "language": string, "kinds": ["static" | "video"], "platforms": string[], "forbid": string[], "require": string[], "allow": string[], "people": "real" | null}]}',
+    "- text: the rule in one short sentence, in the note's own language and close to its words.",
+    '- strength: "never" for a ban, "must" for something every ad has to do, "prefer" for a taste.',
+    '- applies: what of the ad it is about ("copy" for its words, "picture", "cast" for who appears, "logo", "voice", "sound", "subtitles"), else "any".',
+    `- language: the note's language as a two-letter code ("es").`,
+    '- kinds and platforms: only when the note limits the rule to them ("en estos creativos de linkedin" gives platforms ["linkedin"]); else [].',
+    `- forbid: the exact words the rule bans in the ad's words, only when a word or phrase decides it ("100 % online", not "online" gives forbid ["online"], allow ["100 % online"]); else []. require: words every ad must say; usually []. allow: the right phrase that contains a banned word.`,
+    '- people: "real" only when the rule is that no AI-made or invented people appear; else null.',
+    'Return {"rules": []} when the note states no standing rule.'
+  ].join("\n");
+}
+var RULE_DRAFT_MODEL = "google/gemini-3.5-flash";
+var DRAFT_REQUEST = {
+  provider: { order: ["google-vertex"], allow_fallbacks: false },
+  seed: 0,
+  reasoning: { effort: "low" },
+  response_format: { type: "json_object" }
+};
+var DRAFT_TIMEOUT_MS = 3e4;
+function modelRuleDrafter(cacheDir) {
+  return async (note, ad) => {
+    const prompt = draftRulesPrompt(note, ad);
+    const key = createHash("sha256").update(JSON.stringify([RULE_DRAFT_MODEL, prompt])).digest("hex").slice(0, 16);
+    const file = path.join(cacheDir, "rule-drafts", `${key}.json`);
+    const cached = await readFile(file, "utf8").then(
+      (raw) => parseDrafts(JSON.parse(raw)),
+      () => null
+    );
+    if (cached) return cached;
+    const spent = [];
+    const answer = await askAboutText(RULE_DRAFT_MODEL, prompt, (u) => spent.push(u), {
+      timeoutMs: DRAFT_TIMEOUT_MS,
+      request: DRAFT_REQUEST
+    }).catch(() => null);
+    await holdWorkspaceSpend(spent);
+    if (answer === null) return [];
+    await mkdir(path.dirname(file), { recursive: true }).catch(() => void 0);
+    await writeFile(file, JSON.stringify(answer), "utf8").catch(() => void 0);
+    return parseDrafts(answer);
+  };
+}
+async function ruleCandidates(note, ad, draft) {
+  const text = note?.trim();
+  if (!text) return [];
+  return await draft(text, ad).catch(() => []);
+}
+var candidatesDir = (cacheDir) => path.join(cacheDir, "rule-candidates");
+async function saveCandidates(cacheDir, candidates) {
+  await mkdir(candidatesDir(cacheDir), { recursive: true });
+  return await Promise.all(
+    candidates.map(async (c) => {
+      const id = `rc-${createHash("sha256").update(JSON.stringify(c)).digest("hex").slice(0, 8)}`;
+      await writeFile(path.join(candidatesDir(cacheDir), `${id}.json`), JSON.stringify(c), "utf8");
+      return id;
+    })
+  );
+}
+async function loadCandidate(cacheDir, id) {
+  if (!/^rc-[0-9a-f]{8}$/.test(id)) return null;
+  return await readFile(path.join(candidatesDir(cacheDir), `${id}.json`), "utf8").then(
+    (raw) => JSON.parse(raw),
+    () => null
+  );
+}
+var REPEATED_ON = 2;
+var REPEATED_NOTES = 3;
+function repeatedReasons(notes, reasons, slug) {
+  return reasons.flatMap((reason) => {
+    const others = notes.filter((n) => n.creative !== slug && n.reasons.includes(reason));
+    const creatives = new Set(others.map((n) => n.creative)).size;
+    return creatives >= REPEATED_ON ? [
+      {
+        label: CREATIVE_FEEDBACK_REASON_LABELS[reason],
+        creatives,
+        notes: [...new Set(others.map((n) => n.note))].slice(0, REPEATED_NOTES)
+      }
+    ] : [];
+  });
+}
+function ruleCandidateHints(drafted, repeated) {
+  const asks = drafted.map(
+    (d) => `Their words state a rule for every ad: "${d.text}". Ask them once, in their language: make this a rule for every ad from now on? Only on yes: \`baker creatives rules add --candidate ${d.id}\` (reword it with --text if they correct it). On no, never ask again in this chat.`
+  );
+  const patterns = repeated.map(
+    (r) => `"${r.label}" was also the reason on ${r.creatives} other ads (their words: ${r.notes.map((n) => `"${n}"`).join("; ")}). ${drafted.length > 0 ? "Say so when you ask." : 'Ask them once whether there is a rule for every ad behind it; on yes, `baker creatives rules add --text "<the rule, in their words>" \u2026`.'}`
+  );
+  return [...asks, ...patterns];
+}
+
+export {
+  modelRuleDrafter,
+  ruleCandidates,
+  saveCandidates,
+  loadCandidate,
+  repeatedReasons,
+  ruleCandidateHints
+};
+//# sourceMappingURL=chunk-2JFCDUZO.js.map