polymath-society 0.2.5 → 0.2.6

package/dist/pipeline/coding-coaching.js

@@ -134,9 +134,7 @@ var require_secure_json_parse = __commonJS({
 });
 
 // ../../scripts/coding-coaching.mts
-import { promises as fs8 } from "fs";
-import os2 from "os";
-import path12 from "path";
+import { promises as fs11 } from "fs";
 
 // ../../lib/agents/shared/agent.ts
 import { readFileSync } from "fs";
@@ -16271,7 +16269,7 @@ function computeWorkstyle(inp) {
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-12.2",
+  version: "2026-07-14.2",
   // = RUBRIC_FINGERPRINT of the shipped npm rubric (packages/coding-analyzer/
   // src/grade/criteria.ts). A unit test fails loud when the rubric changes
   // without this being updated (and re-pushed to the server).
@@ -16360,7 +16358,7 @@ var DEFAULT_CALIBRATION = {
     }
   ],
   codingTiers: {
-    push: { median: 300, sigma: 1.121, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail shape from its actions-per-prompt distribution. AVERAGE over substantial days, cumulative not spiky." },
+    push: { median: 300, sigma: 0.91, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail: a 1-in-1,000 user AVERAGES ~5k directed words/day (near-daily heavy dictation, clearly below the ~10k single-day record). AVERAGE over substantial days, cumulative not spiky." },
     focus: { median: 0.08, sigma: 0.559, tag: "estimated \u2014 proxy-anchored", source: "median: Anthropic 20h/week engaged-user figure implies ~1h/day of true rapid exchange against an 8h working day; ceiling anchored to the ~4h/day deep-work limit (45% share = 1-in-1,000). AVERAGE share, cumulative not spiky." },
     orchestration: { tag: "estimated \u2014 behavior bands", source: "no published concurrency distribution exists; ceiling anchored to the Claude Code lead's documented 10-15 parallel sessions" }
   },
@@ -16393,12 +16391,18 @@ var DEFAULT_CALIBRATION = {
       line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
     },
     throughput: {
-      // Top-0.1% words/day reference lines, anchored to the throughput ladder
-      // (THROUGHPUT_LADDER in throughput.ts): score 10 "world-class sustained" =
-      // 9k/day, score 11 "superhuman ceiling" = 12k/day. Calibrated to the person's
-      // own anchors, not a wild guess — but still a target, not a measured population.
-      topAvgWordsPerDay: 9e3,
-      topPeakWordsPerDay: 12e3,
+      // Two DIFFERENT kinds of line (2026-07-14.2): topAvg is a PERCENTILE
+      // claim — what a 1-in-1,000 user AVERAGES over substantial days (5k/day:
+      // near-daily heavy dictation; a measured top dictating power user
+      // averages ~3.9k). topPeak is NOT a percentile — it is the CEILING
+      // reference, the wildest single days on record (~10k: a 14-16h
+      // launch-day marathon), rendered as "wildest single days", never
+      // "top X% peak". An avg-percentile and a peak-percentile can't both be
+      // shown at a same-ish value (avg≈peak reads broken — user call), and a
+      // peak PERCENTILE line is unknowable anyway; the record framing is
+      // honest and keeps the visual gap.
+      topAvgWordsPerDay: 5e3,
+      topPeakWordsPerDay: 1e4,
       benchLabel: "top 0.1%",
       // No published "words typed/day" for heavy users — the real story is OUTPUT.
       loopsHeadline: ">1,000,000 lines of Rust at 99.8% tests passing",
@@ -16733,7 +16737,7 @@ async function compileCodingProfile() {
     sRow(
       "How hard you push",
       aggregate?.throughput?.score ?? tCeiling,
-      `you peak at ~${peakWords.toLocaleString()} words of pure direction in a single day \u2014 the very top sustain ~${BEST.throughput.topAvgWordsPerDay.toLocaleString()} every day, week after week, and crest past ${BEST.throughput.topPeakWordsPerDay.toLocaleString()}`
+      `you peak at ~${peakWords.toLocaleString()} words of pure direction in a single day \u2014 the very top sustain ~${BEST.throughput.topAvgWordsPerDay.toLocaleString()} every day, week after week; the wildest single days on record push ~${BEST.throughput.topPeakWordsPerDay.toLocaleString()}`
     ),
     sRow(
       "How well you use AI",
@@ -16809,6 +16813,233 @@ async function readCodingAsset(name17) {
   );
 }
 
+// ../../lib/agents/coding/docsDir.ts
+import { promises as fs9 } from "fs";
+import path12 from "path";
+
+// ../../lib/agents/coding/materialize.ts
+import { promises as fs8 } from "fs";
+var SYSTEM_REMINDER_RE2 = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
+var INJECT_GIST = 120;
+var INJECTED = [
+  { re: /<task-notification>[\s\S]*?<\/task-notification>/g, label: "bg task", gistRe: /<summary>([\s\S]*?)<\/summary>/ },
+  { re: /<local-command-stdout>[\s\S]*?<\/local-command-stdout>/g, label: "cmd output" },
+  { re: /<local-command-caveat>[\s\S]*?<\/local-command-caveat>/g, label: null },
+  { re: /<ide_opened_file>[\s\S]*?<\/ide_opened_file>/g, label: null }
+];
+function collapseInjected(input) {
+  const markers = [];
+  let text2 = input;
+  for (const { re: re2, label, gistRe } of INJECTED) {
+    text2 = text2.replace(re2, (block) => {
+      if (label) {
+        const gist = (gistRe ? block.match(gistRe)?.[1] ?? "" : block.replace(/<[^>]*>/g, " ")).replace(/\s+/g, " ").trim().slice(0, INJECT_GIST);
+        markers.push(gist ? `[${label}: ${gist}]` : `[${label}]`);
+      }
+      return "";
+    });
+  }
+  return { text: text2.replace(SYSTEM_REMINDER_RE2, "").trim(), markers };
+}
+function extractText2(content) {
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    const parts = [];
+    for (const item of content) {
+      if (item && typeof item === "object" && item.type === "text") parts.push(item.text || "");
+    }
+    return parts.join("\n");
+  }
+  return "";
+}
+function isToolResultOnly2(content) {
+  return Array.isArray(content) && content.length > 0 && content.every((c) => c && typeof c === "object" && c.type === "tool_result");
+}
+var AI_GIST = 160;
+async function materializeSession(file2) {
+  let raw;
+  try {
+    raw = await fs8.readFile(file2, "utf-8");
+  } catch {
+    return null;
+  }
+  const lines = [];
+  let humanTurns = 0, humanChars = 0;
+  const pendingQueued = [];
+  let lastMode = "";
+  const pushMarkers = (markers) => {
+    for (const mk of markers) if (mk !== lines[lines.length - 1]) lines.push(mk);
+  };
+  let aiLastProse = "";
+  const aiTools = /* @__PURE__ */ new Map();
+  const flushAI = () => {
+    if (!aiLastProse && aiTools.size === 0) return;
+    const tools = [...aiTools.entries()].map(([n, c]) => c > 1 ? `${n}\xD7${c}` : n).join(", ");
+    const gist = aiLastProse ? aiLastProse.length > AI_GIST ? aiLastProse.slice(0, AI_GIST) + "\u2026" : aiLastProse : "";
+    const bits = [];
+    if (gist) bits.push(gist);
+    if (tools) bits.push(`ran ${tools}`);
+    lines.push(`[AI: ${bits.join(" \xB7 ")}]`);
+    aiLastProse = "";
+    aiTools.clear();
+  };
+  for (const lineRaw of raw.split("\n")) {
+    const line = lineRaw.trim();
+    if (!line) continue;
+    let e;
+    try {
+      e = JSON.parse(line);
+    } catch {
+      continue;
+    }
+    const type = e.type;
+    if (type === "ai-title") continue;
+    if (type === "mode") {
+      flushAI();
+      const m = e.mode;
+      if (m && String(m) !== lastMode) {
+        lastMode = String(m);
+        lines.push(`[mode \u2192 ${m}]`);
+      }
+      continue;
+    }
+    if (type === "queue-operation") {
+      if (e.operation === "enqueue" && typeof e.content === "string") {
+        const { text: t, markers } = collapseInjected(e.content);
+        if (t || markers.length) flushAI();
+        pushMarkers(markers);
+        if (t) {
+          pendingQueued.push(t);
+          lines.push(`>> (queued while AI working) ${t}`);
+        }
+      }
+      continue;
+    }
+    if (type === "user") {
+      if (isHarnessEntry(e)) continue;
+      const content = e.message?.content;
+      if (isToolResultOnly2(content)) continue;
+      const { text: t, markers } = collapseInjected(extractText2(content));
+      if (!t && !markers.length) continue;
+      flushAI();
+      pushMarkers(markers);
+      if (!t) continue;
+      humanTurns++;
+      humanChars += t.length;
+      const qi = pendingQueued.indexOf(t);
+      if (qi >= 0) {
+        pendingQueued.splice(qi, 1);
+        continue;
+      }
+      lines.push(`>> ${t}`);
+      continue;
+    }
+    if (type === "assistant") {
+      const msg = e.message ?? {};
+      let prose = "";
+      if (Array.isArray(msg.content)) {
+        for (const item of msg.content) {
+          const it = item;
+          if (it.type === "text" && it.text) prose += it.text;
+          else if (it.type === "tool_use" && it.name) aiTools.set(it.name, (aiTools.get(it.name) ?? 0) + 1);
+        }
+      } else prose = extractText2(msg.content);
+      prose = prose.replace(/\s+/g, " ").trim();
+      if (prose) aiLastProse = prose;
+      continue;
+    }
+  }
+  flushAI();
+  return { text: lines.join("\n"), humanTurns, humanChars };
+}
+
+// ../../lib/agents/coding/docsDir.ts
+async function writeScoreboard(codingDir, docs) {
+  const gradesDir = path12.join(codingDir, "grades");
+  let files = [];
+  try {
+    files = (await fs9.readdir(gradesDir)).filter((f) => f.endsWith(".json")).sort();
+  } catch {
+    return;
+  }
+  const byCriterion = /* @__PURE__ */ new Map();
+  for (const f of files) {
+    let g;
+    try {
+      g = JSON.parse(await fs9.readFile(path12.join(gradesDir, f), "utf8"));
+    } catch {
+      continue;
+    }
+    for (const c of g.criteria ?? []) {
+      const s = c.score ?? c.bestEstimate;
+      if (s == null) continue;
+      const line = `L${s}${c.score == null ? "~" : ""} \xB7 ${(g.date ?? "").slice(0, 10)} \xB7 ${(g.title ?? g.sessionId ?? "").replace(/\n.*/s, "").slice(0, 60)} \xB7 ${g.sessionId} \u2014 ${(c.instance ?? "").slice(0, 110)}`;
+      (byCriterion.get(c.key) ?? byCriterion.set(c.key, []).get(c.key)).push({ s, line });
+    }
+  }
+  if (!byCriterion.size) return;
+  const blocks = [...byCriterion.entries()].sort(([a], [b]) => a < b ? -1 : 1).map(
+    ([key, rows]) => `## ${key}
+${rows.sort((a, b) => b.s - a.s).map((r) => r.line).join("\n")}`
+  );
+  await fs9.writeFile(
+    path12.join(docs, "_scoreboard.md"),
+    `# Scoreboard \u2014 every graded session per criterion, best first.
+# Lx~ = estimate. Format: level \xB7 date \xB7 title \xB7 sessionId \u2014 graded instance.
+
+${blocks.join("\n\n")}
+`
+  );
+}
+async function ensureCodingDocs(codingDir, sessions) {
+  const docs = path12.join(codingDir, "docs");
+  await fs9.mkdir(docs, { recursive: true });
+  const user = sessions.filter((s) => s.sessionId && !s.duplicateOf && (!("klass" in s) || s.klass === "interactive"));
+  let written = 0;
+  for (const s of user) {
+    const out = path12.join(docs, `${s.sessionId}.txt`);
+    try {
+      await fs9.access(out);
+      continue;
+    } catch {
+    }
+    const mat = await materializeSession(s.file).catch(() => null);
+    if (!mat?.text) continue;
+    await fs9.writeFile(out, mat.text);
+    written++;
+  }
+  const rows = user.sort((a, b) => (a.start ?? "") < (b.start ?? "") ? -1 : 1).map((s) => `${(s.start ?? "").slice(0, 10)} \xB7 ${(s.title ?? "").replace(/\n.*/s, "").slice(0, 80)} \xB7 ${s.sessionId}`);
+  await fs9.writeFile(path12.join(docs, "_map.md"), rows.join("\n") + "\n");
+  await writeScoreboard(codingDir, docs);
+  if (written) console.log(`[docs] materialized ${written} new transcript(s) \u2192 ${docs}`);
+  return docs;
+}
+function sourceLookupNote() {
+  return `
+
+SOURCE LOOKUP \u2014 you have Read/Grep/Glob over the working directory, which holds EVERYTHING known about this person's coding work. MAPS FIRST, grep second (the engine rule): docs/_scoreboard.md ranks every graded session per criterion, best first \u2014 one glance shows where the spikes/flair are, so start there when hunting peaks or checking whether a high mark is consistent or a one-off; day-digest.json is the day-by-day story of what they did in order \u2014 read the relevant days to get the arc before searching; docs/_map.md lists every session (date \xB7 title \xB7 sessionId). Then drill down: walkthroughs/<sessionId>.json is that session's phase-by-phase account (cheap, already distilled), grades/<sessionId>.json its graded findings with quotes, and docs/<sessionId>.txt the FULL transcript (">> " = the person, "[AI]" = the model) \u2014 the ground truth, read it when the distilled layers can't settle it. Other artifacts (flow.json, delegation/, expertise/, aggregate.json) are there too if a claim depends on them. The digest you were given is the spine \u2014 use lookups when a CLAIM-MOVING question cannot be settled from it (a moment you are about to headline, a promotion/depth call, an ambiguous engagement, whether detail was vision or boilerplate). VERIFY every verbatim quote you cite by finding it in the transcript first; if you cannot find it, do not use it. Look up what changes the output \u2014 do not re-read the whole corpus.`;
+}
+
+// ../../lib/agents/coding/promptCache.ts
+import { createHash as createHash2 } from "crypto";
+import { promises as fs10 } from "fs";
+function promptSha(parts) {
+  const h = createHash2("sha256");
+  for (const p of parts) {
+    h.update(p ?? "");
+    h.update("\0");
+  }
+  return h.digest("hex").slice(0, 16);
+}
+async function reusableOutput(outPath, sha, valid, key = "promptSha") {
+  try {
+    const o = JSON.parse(await fs10.readFile(outPath, "utf8"));
+    if (o[key] === sha && valid(o)) return o;
+  } catch {
+  }
+  return null;
+}
+
 // ../../scripts/coding-coaching.mts
 var CODING = CODING_DIR;
 async function main() {
@@ -16847,13 +17078,27 @@ async function main() {
     whatYouAre: {
       gradedSessions: p.gradedSessions,
       qualities: (p.stackUp ?? []).map((r) => ({ quality: r.label, score: r.score })),
+      // Person-level judgments (coding-person, when it has run): one traversing
+      // agent per criterion already ranked the PERSON with source-verified
+      // receipts — the placement runs AFTER those judgments and builds on them.
+      personLevelJudgments: await fs11.readFile(`${CODING}/person.json`, "utf8").then((t) => {
+        const j = JSON.parse(t);
+        return Object.values(j.criteria ?? {}).map((g) => ({
+          quality: g.label,
+          score: g.score,
+          confidence: g.confidence,
+          headline: g.headline,
+          receipts: (g.receipts ?? []).slice(0, 2).map((r) => ({ date: r.date, quote: r.quote, what: r.what }))
+        }));
+      }).catch(() => null),
+      personNote: "When personLevelJudgments is present it is the AUTHORITATIVE per-quality read (a per-criterion judge already verified these receipts at the source) \u2014 pick the spikes for the 'line' from ITS highest scores and sharpest receipts; 'qualities' is the fallback when it is null.",
       receipts: p.criteria.filter((c) => c.graded).map((c) => ({
         criterion: c.label,
         takes: c.takes.slice(0, 2).map((t) => ({ date: t.date, title: t.title, score: t.score ?? t.best, what: t.instance, quote: t.quotes?.[0] ?? "" }))
       }))
     }
   };
-  const SYS = "You write the COACHING and FEEL-SEEN copy for a coding-skills report, strictly following this writing rubric:\n\n" + writing + "\n\nYou are given structured facts (numbers + concrete instances) from one person's Claude Code sessions. Three jobs:\n1) COACHING \u2014 for each section, write ONE sentence: the single most useful thing for THIS person, grounded in their actual numbers \u2014 how to improve, where they're weak, or the best thing to do differently. If a section is genuinely strong, say what to push further. Cite the number. No hedging, no fluff, no generic best-practice. Compare to the top where it helps (top ENGAGED time ~50% of day; the best sustain focus ~4h and run many agents in parallel). Vocabulary law: 'flow' means ONLY the strict metric (strictFlow facts); the loose metric is 'engaged time' \u2014 never call it flow, and never write a line that contradicts the strict flow map shown above this card. Deliberate-choice guard: NEVER recommend git worktrees (this person runs parallel agents on main by design and has never hit clobbering); do not recommend a frontier capability unless it appears in pendingFrontier.\n2) FEEL-SEEN \u2014 for abstraction, agency, and taste, write ONE concrete sentence that makes the person feel SEEN: name a SPECIFIC thing they did (from the instances) so they think 'it actually saw that', not generic praise. Second person, present tense.\n3) WHAT-YOU-ARE \u2014 the report's concluding identity read: a PLACEMENT verdict. Judge, from the receipts and scores, where this person would actually land on a real team, and say it flatteringly AND accurately. 'evidence' = one or two sentences pointing at the 2-3 sharpest receipts (named project, session, date, or number) that prove the line, chosen so the person thinks 'it actually saw that'.\n   'line' has EXACTLY this three-beat shape, hard-hitting and pithy:\n   BEAT 1 \u2014 THE SPIKES, first words: name their 1-2 standout qualities flat and confident, plain words ('You have <spike> and you <spike>') \u2014 chosen from the highest scores and sharpest receipts, so the reader's first glance already says what to expect from this person.\n   BEAT 2 \u2014 THE PLACEMENT those spikes buy: 'drop you into <team> and you'd be <placement>' \u2014 a coined identity (the register of 'high-agency outcomes machine'), NEVER a job title ('founding engineer', 'senior engineer' = flat, cringe), one clause ending on a punch word, sized to the evidence: star-of-any-startup for truly exceptional receipts, 'top third of engineers at any big tech company \u2014 ships without being managed' for a notch less. A size too big is patronizing, a size too small is a snub.\n   BEAT 3 \u2014 a short hard verdict: why that pair is the rare, load-bearing thing, and that it's real.\n   Worked examples across the calibration range \u2014 generic people, FORM and SIZE only; never reuse their wording, never treat them as a menu (the spikes must come from THIS person's scores):\n   - exceptional agency + push: 'You push harder than almost anyone and you own outcomes end-to-end \u2014 drop you into any small startup and you'd be the engine of the whole thing, relentlessly shipping what's actually good. That pair is the rarest, most load-bearing thing here, and it's real.'\n   - exceptional taste + judgement: 'Your taste is the real thing and your calls under ambiguity keep being right \u2014 drop you into any product team and you'd be the bar: the one who knows which version is right before anyone can say why. Now that AI makes output infinite, that's the scarcest seat in the room.'\n   - strong delegation + expertise, a notch less: 'You run AI like a staff and your depth shows exactly where the tool runs out \u2014 you'd be in the top third of engineers at any big tech company, the one who ships without being managed. That combination is rarer than it sounds, and yours is proven.'\n   - real focus + follow-through, modest spikes: 'You hold deep focus for hours and you finish what you start \u2014 drop you onto any team and you'd be the steady core, the one whose thread always converges. Every team needs exactly one of those; few have one.'\n   - Pithy: two sentences max. No receipts crammed into the line (evidence carries those). No qualifier stacks, no consultant words (force multiplier, leverage, north star), no hedging.\n   - TESTABLE placement, not adjectives. No population percentages.\n   - If the receipts support NO honest flattering placement, output empty strings for both fields \u2014 nothing is better than an inflated or hedged read.\n   - Never invent; every claim in 'evidence' must trace to a given receipt.";
+  const SYS = "You write the COACHING and FEEL-SEEN copy for a coding-skills report, strictly following this writing rubric:\n\n" + writing + "\n\nYou are given structured facts (numbers + concrete instances) from one person's Claude Code sessions. Three jobs:\n1) COACHING \u2014 for each section, write ONE sentence: the single most useful thing for THIS person, grounded in their actual numbers \u2014 how to improve, where they're weak, or the best thing to do differently. If a section is genuinely strong, say what to push further. Cite the number. No hedging, no fluff, no generic best-practice. Compare to the top where it helps (top ENGAGED time ~50% of day; the best sustain focus ~4h and run many agents in parallel). Vocabulary law: 'flow' means ONLY the strict metric (strictFlow facts); the loose metric is 'engaged time' \u2014 never call it flow, and never write a line that contradicts the strict flow map shown above this card. Deliberate-choice guard: NEVER recommend git worktrees (this person runs parallel agents on main by design and has never hit clobbering); do not recommend a frontier capability unless it appears in pendingFrontier.\n2) FEEL-SEEN \u2014 for abstraction, agency, and taste, write ONE concrete sentence that makes the person feel SEEN: name a SPECIFIC thing they did (from the instances) so they think 'it actually saw that', not generic praise. Second person, present tense.\n3) WHAT-YOU-ARE \u2014 the report's concluding identity read: a PLACEMENT verdict. Judge, from the receipts and scores, where this person would actually land on a real team, and say it flatteringly AND accurately. 'evidence' = one or two sentences pointing at the 2-3 sharpest receipts (named project, session, date, or number) that prove the line, chosen so the person thinks 'it actually saw that'.\n   'line' has EXACTLY this three-beat shape, hard-hitting and pithy:\n   BEAT 1 \u2014 THE SPIKES, first words: name their 1-2 standout qualities flat and confident, plain words ('You have <spike> and you <spike>') \u2014 chosen from the highest scores and sharpest receipts, so the reader's first glance already says what to expect from this person.\n   BEAT 2 \u2014 THE PLACEMENT those spikes buy: 'drop you into <team> and you'd be <placement>' \u2014 a coined identity (the register of 'high-agency outcomes machine'), NEVER a job title ('founding engineer', 'senior engineer' = flat, cringe), one clause ending on a punch word, sized to the evidence: star-of-any-startup for truly exceptional receipts, 'top third of engineers at any big tech company \u2014 ships without being managed' for a notch less. A size too big is patronizing, a size too small is a snub.\n   BEAT 3 \u2014 a short hard verdict: why that pair is the rare, load-bearing thing, and that it's real.\n   Worked examples across the calibration range \u2014 generic people, FORM and SIZE only; never reuse their wording, never treat them as a menu (the spikes must come from THIS person's scores):\n   - exceptional agency + push: 'You push harder than almost anyone and you own outcomes end-to-end \u2014 drop you into any small startup and you'd be the engine of the whole thing, relentlessly shipping what's actually good. That pair is the rarest, most load-bearing thing here, and it's real.'\n   - exceptional taste + judgement: 'Your taste is the real thing and your calls under ambiguity keep being right \u2014 drop you into any product team and you'd be the bar: the one who knows which version is right before anyone can say why. Now that AI makes output infinite, that's the scarcest seat in the room.'\n   - strong delegation + expertise, a notch less: 'You run AI like a staff and your depth shows exactly where the tool runs out \u2014 you'd be in the top third of engineers at any big tech company, the one who ships without being managed. That combination is rarer than it sounds, and yours is proven.'\n   - real focus + follow-through, modest spikes: 'You hold deep focus for hours and you finish what you start \u2014 drop you onto any team and you'd be the steady core, the one whose thread always converges. Every team needs exactly one of those; few have one.'\n   - Pithy: two sentences max. No receipts crammed into the line (evidence carries those). No qualifier stacks, no consultant words (force multiplier, leverage, north star), no hedging.\n   - TESTABLE placement, not adjectives. No population percentages.\n   - If the receipts support NO honest flattering placement, output empty strings for both fields \u2014 nothing is better than an inflated or hedged read.\n   - Never invent; every claim in 'evidence' must trace to a given receipt." + sourceLookupNote();
   const prompt = `FACTS (numbers + concrete instances per section):
 ${JSON.stringify(facts, null, 1).slice(0, 6e4)}
 
@@ -16865,17 +17110,24 @@ Output a SINGLE fenced json block, every value a single tight sentence (whatYouA
   "whatYouAre": { "line": "", "evidence": "" }
 }
 \`\`\``;
-  console.log(`coaching \u2014 1 ${process.env.AGENT_MODEL || "opus"} pass over compiled profile`);
-  const sandbox = await fs8.mkdtemp(path12.join(os2.tmpdir(), "coach-"));
+  const model = process.env.AGENT_MODEL || "opus";
+  const sha = promptSha([SYS, prompt, model]);
+  if (!process.argv.includes("--refresh")) {
+    const prev = await reusableOutput(`${CODING}/coaching.json`, sha, (o) => Object.keys(o.coaching || {}).length >= 5 && Object.keys(o.feelSeen || {}).length >= 3);
+    if (prev) {
+      console.log(`[coaching] inputs unchanged \u2014 reusing output from ${prev.generatedAt} (--refresh to regenerate)`);
+      return;
+    }
+  }
+  const sessions = await fs11.readFile(`${CODING}/sessions.json`, "utf8").then(JSON.parse).catch(() => []);
+  await ensureCodingDocs(CODING, sessions);
+  console.log(`coaching \u2014 1 ${model} pass over compiled profile (+source lookups)`);
   let out = { coaching: {}, feelSeen: {} };
   try {
-    const run2 = await invokeAgent({ prompt, systemPrompt: SYS, addDir: sandbox, allowedTools: ["Read"], maxTurns: 4, timeoutMs: 54e4, model: process.env.AGENT_MODEL || "opus" });
+    const run2 = await invokeAgent({ prompt, systemPrompt: SYS, addDir: CODING, allowedTools: ["Read", "Grep", "Glob"], maxTurns: 10, timeoutMs: 54e4, model });
     out = run2.json || out;
   } catch (e) {
     console.log("ERR", e.message.slice(0, 120));
-  } finally {
-    await fs8.rm(sandbox, { recursive: true, force: true }).catch(() => {
-    });
   }
   const wya = out.whatYouAre;
   const complete = Object.keys(out.coaching || {}).length >= 5 && Object.keys(out.feelSeen || {}).length >= 3 && wya && typeof wya.line === "string" && typeof wya.evidence === "string";
@@ -16883,8 +17135,8 @@ Output a SINGLE fenced json block, every value a single tight sentence (whatYouA
     console.error(`REFUSING to write ${CODING}/coaching.json \u2014 incomplete output (coaching:${Object.keys(out.coaching || {}).length} feelSeen:${Object.keys(out.feelSeen || {}).length} whatYouAre:${wya ? "malformed" : "missing"})`);
     process.exit(1);
   }
-  const result = { generatedAt: (/* @__PURE__ */ new Date()).toISOString(), ...out };
-  await fs8.writeFile(`${CODING}/coaching.json`, JSON.stringify(result, null, 2));
+  const result = { generatedAt: (/* @__PURE__ */ new Date()).toISOString(), promptSha: sha, ...out };
+  await fs11.writeFile(`${CODING}/coaching.json`, JSON.stringify(result, null, 2));
   console.log(`WROTE ${CODING}/coaching.json \u2014 coaching:${Object.keys(out.coaching || {}).length} feelSeen:${Object.keys(out.feelSeen || {}).length} whatYouAre:${wya.line.trim() ? "ok" : "declined"}`);
 }
 main().catch((e) => {

package/dist/pipeline/coding-day-digest.js

@@ -15975,7 +15975,7 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-12.2",
+  version: "2026-07-14.2",
   // = RUBRIC_FINGERPRINT of the shipped npm rubric (packages/coding-analyzer/
   // src/grade/criteria.ts). A unit test fails loud when the rubric changes
   // without this being updated (and re-pushed to the server).
@@ -16064,7 +16064,7 @@ var DEFAULT_CALIBRATION = {
     }
   ],
   codingTiers: {
-    push: { median: 300, sigma: 1.121, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail shape from its actions-per-prompt distribution. AVERAGE over substantial days, cumulative not spiky." },
+    push: { median: 300, sigma: 0.91, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail: a 1-in-1,000 user AVERAGES ~5k directed words/day (near-daily heavy dictation, clearly below the ~10k single-day record). AVERAGE over substantial days, cumulative not spiky." },
     focus: { median: 0.08, sigma: 0.559, tag: "estimated \u2014 proxy-anchored", source: "median: Anthropic 20h/week engaged-user figure implies ~1h/day of true rapid exchange against an 8h working day; ceiling anchored to the ~4h/day deep-work limit (45% share = 1-in-1,000). AVERAGE share, cumulative not spiky." },
     orchestration: { tag: "estimated \u2014 behavior bands", source: "no published concurrency distribution exists; ceiling anchored to the Claude Code lead's documented 10-15 parallel sessions" }
   },
@@ -16097,12 +16097,18 @@ var DEFAULT_CALIBRATION = {
       line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
     },
     throughput: {
-      // Top-0.1% words/day reference lines, anchored to the throughput ladder
-      // (THROUGHPUT_LADDER in throughput.ts): score 10 "world-class sustained" =
-      // 9k/day, score 11 "superhuman ceiling" = 12k/day. Calibrated to the person's
-      // own anchors, not a wild guess — but still a target, not a measured population.
-      topAvgWordsPerDay: 9e3,
-      topPeakWordsPerDay: 12e3,
+      // Two DIFFERENT kinds of line (2026-07-14.2): topAvg is a PERCENTILE
+      // claim — what a 1-in-1,000 user AVERAGES over substantial days (5k/day:
+      // near-daily heavy dictation; a measured top dictating power user
+      // averages ~3.9k). topPeak is NOT a percentile — it is the CEILING
+      // reference, the wildest single days on record (~10k: a 14-16h
+      // launch-day marathon), rendered as "wildest single days", never
+      // "top X% peak". An avg-percentile and a peak-percentile can't both be
+      // shown at a same-ish value (avg≈peak reads broken — user call), and a
+      // peak PERCENTILE line is unknowable anyway; the record framing is
+      // honest and keeps the visual gap.
+      topAvgWordsPerDay: 5e3,
+      topPeakWordsPerDay: 1e4,
       benchLabel: "top 0.1%",
       // No published "words typed/day" for heavy users — the real story is OUTPUT.
       loopsHeadline: ">1,000,000 lines of Rust at 99.8% tests passing",
@@ -16131,8 +16137,9 @@ var CODING = CODING_DIR;
 var OUT = path11.join(CODING, "day-digest.json");
 var TZ = process.env.CODING_TZ || "America/Los_Angeles";
 async function main() {
-  const sessions = JSON.parse(await fs6.readFile(path11.join(CODING, "sessions.json"), "utf8"));
-  console.log(`[day-digest] grouping ${sessions.length} sessions' messages by day (tz ${TZ})\u2026`);
+  const all = JSON.parse(await fs6.readFile(path11.join(CODING, "sessions.json"), "utf8"));
+  const sessions = all.filter((s) => s.klass === "interactive");
+  console.log(`[day-digest] grouping ${sessions.length}/${all.length} interactive sessions' messages by day (tz ${TZ})\u2026`);
   const dayConvs = await buildDayConversations(sessions, TZ);
   console.log(`[day-digest] ${dayConvs.length} active days \xB7 ${dayConvs.reduce((a, d) => a + d.totalMessages, 0)} user messages total`);
   const prev = REDO ? { generatedAt: "", tz: TZ, days: [] } : await fs6.readFile(OUT, "utf8").then(JSON.parse).catch(() => ({ generatedAt: "", tz: TZ, days: [] }));

package/dist/pipeline/coding-delegation.js

@@ -134,7 +134,7 @@ var require_secure_json_parse = __commonJS({
 });
 
 // ../../scripts/coding-delegation.mts
-import { promises as fs8 } from "fs";
+import { promises as fs9 } from "fs";
 import path12 from "path";
 
 // ../../lib/agents/coding/delegation/detect.ts
@@ -16221,13 +16221,33 @@ function partition(sessions) {
   return { gradable, thin, trivial };
 }
 
+// ../../lib/agents/coding/promptCache.ts
+import { createHash } from "crypto";
+import { promises as fs8 } from "fs";
+function promptSha(parts) {
+  const h = createHash("sha256");
+  for (const p of parts) {
+    h.update(p ?? "");
+    h.update("\0");
+  }
+  return h.digest("hex").slice(0, 16);
+}
+async function reusableOutput(outPath, sha, valid, key = "promptSha") {
+  try {
+    const o = JSON.parse(await fs8.readFile(outPath, "utf8"));
+    if (o[key] === sha && valid(o)) return o;
+  } catch {
+  }
+  return null;
+}
+
 // ../../lib/agents/coding/profile.ts
 import path11 from "path";
 
 // ../../lib/calibration/fingerprint.ts
-import { createHash } from "node:crypto";
+import { createHash as createHash2 } from "node:crypto";
 function rubricFingerprint(criteria) {
-  return createHash("sha256").update(JSON.stringify(criteria)).digest("hex").slice(0, 10);
+  return createHash2("sha256").update(JSON.stringify(criteria)).digest("hex").slice(0, 10);
 }
 
 // ../../lib/agents/coding/criteria.ts
@@ -16376,7 +16396,7 @@ var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s
 
 // ../../lib/calibration/index.ts
 var DEFAULT_CALIBRATION = {
-  version: "2026-07-12.2",
+  version: "2026-07-14.2",
   // = RUBRIC_FINGERPRINT of the shipped npm rubric (packages/coding-analyzer/
   // src/grade/criteria.ts). A unit test fails loud when the rubric changes
   // without this being updated (and re-pushed to the server).
@@ -16465,7 +16485,7 @@ var DEFAULT_CALIBRATION = {
     }
   ],
   codingTiers: {
-    push: { median: 300, sigma: 1.121, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail shape from its actions-per-prompt distribution. AVERAGE over substantial days, cumulative not spiky." },
+    push: { median: 300, sigma: 0.91, tag: "estimated \u2014 log-normal fit", source: "Anthropic 2026 Claude Code study (400k sessions): median engaged user ~8 prompts/day at ~35 words \u2014 ~300 directed words/day; tail: a 1-in-1,000 user AVERAGES ~5k directed words/day (near-daily heavy dictation, clearly below the ~10k single-day record). AVERAGE over substantial days, cumulative not spiky." },
     focus: { median: 0.08, sigma: 0.559, tag: "estimated \u2014 proxy-anchored", source: "median: Anthropic 20h/week engaged-user figure implies ~1h/day of true rapid exchange against an 8h working day; ceiling anchored to the ~4h/day deep-work limit (45% share = 1-in-1,000). AVERAGE share, cumulative not spiky." },
     orchestration: { tag: "estimated \u2014 behavior bands", source: "no published concurrency distribution exists; ceiling anchored to the Claude Code lead's documented 10-15 parallel sessions" }
   },
@@ -16498,12 +16518,18 @@ var DEFAULT_CALIBRATION = {
       line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
     },
     throughput: {
-      // Top-0.1% words/day reference lines, anchored to the throughput ladder
-      // (THROUGHPUT_LADDER in throughput.ts): score 10 "world-class sustained" =
-      // 9k/day, score 11 "superhuman ceiling" = 12k/day. Calibrated to the person's
-      // own anchors, not a wild guess — but still a target, not a measured population.
-      topAvgWordsPerDay: 9e3,
-      topPeakWordsPerDay: 12e3,
+      // Two DIFFERENT kinds of line (2026-07-14.2): topAvg is a PERCENTILE
+      // claim — what a 1-in-1,000 user AVERAGES over substantial days (5k/day:
+      // near-daily heavy dictation; a measured top dictating power user
+      // averages ~3.9k). topPeak is NOT a percentile — it is the CEILING
+      // reference, the wildest single days on record (~10k: a 14-16h
+      // launch-day marathon), rendered as "wildest single days", never
+      // "top X% peak". An avg-percentile and a peak-percentile can't both be
+      // shown at a same-ish value (avg≈peak reads broken — user call), and a
+      // peak PERCENTILE line is unknowable anyway; the record framing is
+      // honest and keeps the visual gap.
+      topAvgWordsPerDay: 5e3,
+      topPeakWordsPerDay: 1e4,
       benchLabel: "top 0.1%",
       // No published "words typed/day" for heavy users — the real story is OUTPUT.
       loopsHeadline: ">1,000,000 lines of Rust at 99.8% tests passing",
@@ -16533,6 +16559,7 @@ var MODEL = flag("model", "sonnet");
 var CONC = Number(flag("conc", "4"));
 var IDS = flag("ids", "").split(",").map((s) => s.trim()).filter(Boolean);
 var N = argv.includes("--n") ? Number(flag("n", "0")) : 0;
+var REFRESH = argv.includes("--refresh");
 function makeLimiter(max) {
   let active = 0;
   const q = [];
@@ -16551,19 +16578,29 @@ var CODING = CODING_DIR;
 var OUT = path12.join(CODING, "delegation");
 var LOGS = path12.join(OUT, "logs");
 async function main() {
-  const sessions = JSON.parse(await fs8.readFile(path12.join(CODING, "sessions.json"), "utf8"));
-  const live = sessions.filter((s) => !s.duplicateOf);
+  const sessions = JSON.parse(await fs9.readFile(path12.join(CODING, "sessions.json"), "utf8"));
+  const live = sessions.filter((s) => s.klass === "interactive" && !s.duplicateOf);
   const { gradable } = partition(live);
   let picked = gradable;
   if (IDS.length) {
     const set = new Set(IDS);
     picked = live.filter((s) => set.has(s.sessionId));
   } else if (N > 0) picked = gradable.slice().sort((a, b) => b.humanChars * Math.log1p(b.humanTurns) - a.humanChars * Math.log1p(a.humanTurns)).slice(0, N);
-  await fs8.mkdir(LOGS, { recursive: true });
+  await fs9.mkdir(LOGS, { recursive: true });
   console.log(`[delegation] ${live.length} live \xB7 ${gradable.length} gradable \xB7 running detect on ${picked.length} (LLM only on stuck-loop candidates)\u2026`);
   const lim = makeLimiter(CONC);
+  let reused = 0;
   startRun("coding-delegation", picked.length);
   const reads = (await Promise.all(picked.map((s) => lim(async () => {
+    if (!REFRESH) {
+      try {
+        const prev = JSON.parse(await fs9.readFile(path12.join(OUT, `${s.sessionId}.json`), "utf8"));
+        reused++;
+        tick(true);
+        return prev;
+      } catch {
+      }
+    }
     try {
       const r = await detectDelegation(s, { model: MODEL, outDir: OUT, logDir: LOGS });
       tick(!!r);
@@ -16576,9 +16613,17 @@ async function main() {
     }
   })))).filter(Boolean);
   endRun();
-  console.log(`[delegation] corpus rollup (Case A count + Case B over-drive judgment)\u2026`);
-  const rollup = await rollupDelegation(reads, live, { model: MODEL });
-  await fs8.writeFile(path12.join(CODING, "delegation-rollup.json"), JSON.stringify(rollup, null, 2));
+  if (reused) console.log(`[delegation] reused ${reused} existing per-session read(s) (--refresh to re-detect)`);
+  const stable = reads.map((r) => ({ id: r.sessionId, loops: r.stuckLoops, od: r.overDrive })).sort((a, b) => a.id < b.id ? -1 : 1);
+  const sha = promptSha([JSON.stringify(stable), MODEL]);
+  const rollupPath = path12.join(CODING, "delegation-rollup.json");
+  let rollup = REFRESH ? null : await reusableOutput(rollupPath, sha, (o) => typeof o.sessions === "number");
+  if (rollup) console.log(`[delegation] rollup inputs unchanged \u2014 reusing (--refresh to regenerate)`);
+  else {
+    console.log(`[delegation] corpus rollup (Case A count + Case B over-drive judgment)\u2026`);
+    rollup = { ...await rollupDelegation(reads, live, { model: MODEL }), promptSha: sha };
+    await fs9.writeFile(rollupPath, JSON.stringify(rollup, null, 2));
+  }
   console.log(`
 \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 DELEGATION \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550`);
   console.log(`CASE A \u2014 stuck loops: ${rollup.stuckLoopInstances} across ${rollup.sessions} sessions`);