polymath-society 0.2.19 → 0.2.20

package/dist/pipeline/coding-nutshell.js

@@ -15687,6 +15687,160 @@ function selfImageReference(si, opts = {}) {
 ${parts.join("\n\n")}`;
 }
 
+// ../../lib/calibration/index.ts
+var DEFAULT_CALIBRATION = {
+  version: "2026-07-16.2",
+  // = 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).
+  rubricVersion: "7983c01d53",
+  // Canonical scale = the one the product owner set for the public report
+  // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
+  // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
+  // intent.
+  scoreBands: [
+    { score: 11, label: "Top 0.01%" },
+    { score: 10, label: "Top 0.1%" },
+    { score: 9, label: "Top 0.5%" },
+    { score: 8, label: "Top 2%" },
+    { score: 7, label: "Top 5%" },
+    { score: 6, label: "Top 15%" },
+    { score: 5, label: "Top 50%" },
+    { score: 4, label: "Top 65%" },
+    { score: 3, label: "Top 80%" },
+    { score: 2, label: "Top 90%" },
+    { score: 1, label: "Top 97%" }
+  ],
+  minRankedScore: 6,
+  benchmarkBands: [
+    {
+      level: "11",
+      name: "Superhuman",
+      rank: "Top 0.01%",
+      oneIn: "1 in 10,000+",
+      meaning: "Beyond the normal human ceiling for the trait \u2014 the genuine power-law outlier. Almost no one earns this; when the evidence shows it, it is scored, not rounded down.",
+      foundIn: [
+        "Fields Medal and Nobel-track researchers",
+        "Founders of generational companies",
+        "The handful of people a frontier field is named after"
+      ],
+      maxTopPct: 0.01
+    },
+    {
+      level: "10",
+      name: "World-class",
+      rank: "Top 0.1%",
+      oneIn: "1 in 1,000",
+      meaning: "Among the best alive at this trait \u2014 the level entire institutions are built to find.",
+      foundIn: [
+        "Researchers at frontier AI labs",
+        "Faculty at top-5 research universities",
+        "IMO / IOI medalists",
+        "Founders backed by the top decile of venture firms"
+      ],
+      maxTopPct: 0.1
+    },
+    {
+      level: "9",
+      name: "Exceptional",
+      rank: "Top 1%",
+      oneIn: "1 in 100",
+      meaning: "Clearly exceptional \u2014 the strongest person on most strong teams.",
+      foundIn: [
+        "PhD students at top programs",
+        "Early engineers at breakout startups",
+        "YC-class founders",
+        "National olympiad finalists"
+      ],
+      maxTopPct: 1
+    },
+    {
+      level: "8",
+      name: "Strong",
+      rank: "Top 10%",
+      oneIn: "1 in 10",
+      meaning: "Strong against the whole population \u2014 the best person in most rooms, not every room.",
+      foundIn: [
+        "Senior engineers at selective tech companies",
+        "Graduates of demanding technical programs",
+        "Operators who get promoted everywhere they go"
+      ],
+      maxTopPct: 10
+    },
+    {
+      level: "5",
+      name: "Median",
+      rank: "50th percentile",
+      oneIn: "1 in 2",
+      meaning: 'The middle of the general population. The reference class is all ~8 billion humans \u2014 so even "strong" above already means top 10% of everyone.',
+      foundIn: ["The general population \u2014 most people, most places"],
+      maxTopPct: 100
+    }
+  ],
+  codingTiers: {
+    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." },
+    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
+  },
+  codingBenchmarks: {
+    flow: {
+      // typical knowledge worker ≈ 35% of an 8h day focused; ~4h/day is the
+      // recognized deep-work ceiling (Newport) ≈ 50% — most sit far below it.
+      typicalPctOfDay: 0.35,
+      topPctOfDay: 0.5,
+      tag: "measured",
+      source: "RescueTime 2019 (~2h48m focused/day) \xB7 Cal Newport, Deep Work (~4h/day deep-work ceiling)",
+      line: "the best spend ~half the workday in true deep flow; ~4h/day is the human ceiling"
+    },
+    longestRun: {
+      typicalHours: 0.67,
+      // ~40 min before the average worker is interrupted
+      topHours: 4,
+      tag: "measured",
+      source: "RescueTime (avg max ~40 min focus before interruption) \xB7 deep-work single-block ceiling ~3\u20134h",
+      line: "a top performer can hold a single ~3\u20134h uninterrupted block; the average worker breaks at ~40 min"
+    },
+    parallelism: {
+      topMaxConcurrent: 12,
+      // Boris Cherny ~10–15 interactive sessions
+      topAvgConcurrent: 6,
+      // time-weighted average while active (anecdotal, same source)
+      worktreesTip: "3\u20135 git worktrees at once",
+      tag: "anecdotal",
+      source: "Boris Cherny (Claude Code lead, Anthropic): ~10\u201315 concurrent sessions, 'dozens of Claudes running at all times'; 3\u20135 worktrees = 'the single biggest productivity unlock'",
+      line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
+    },
+    throughput: {
+      // 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",
+      loopsDetail: "the Bun runtime was ported Zig\u2192Rust largely autonomously \u2014 hundreds of parallel subagents, two AI reviewers per file",
+      tag: "measured",
+      source: "The Register, May 2026 (Bun Zig\u2192Rust rewrite)",
+      line: "autonomous fleets ported >1M lines of Rust at 99.8% tests passing; a while-loop agent delivered a $50k contract for $297 of compute"
+    }
+  }
+};
+function bandLabel(score, doc = DEFAULT_CALIBRATION) {
+  const s = Math.max(1, Math.min(11, Math.round(score)));
+  let best = null;
+  for (const b of doc.scoreBands) if (b.score <= s && (!best || b.score > best.score)) best = b;
+  return (best ?? doc.scoreBands[0]).label;
+}
+
 // ../../lib/agents/coding/profile.ts
 import path11 from "path";
 
@@ -15840,154 +15994,6 @@ var IDLE_CAP_MS = 12 * 6e4;
 var NOTE = "((ran|created|edited|read|wrote|searched) (a|an|\\d+) [a-z]+|updated todos)";
 var TRANSCRIPT_LINE_RE = new RegExp(`^\\s*(${NOTE})((,| and) ${NOTE})*\\s*$|^\\s*thought for .{1,30}$`, "im");
 
-// ../../lib/calibration/index.ts
-var DEFAULT_CALIBRATION = {
-  version: "2026-07-16.2",
-  // = 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).
-  rubricVersion: "7983c01d53",
-  // Canonical scale = the one the product owner set for the public report
-  // (11=0.01% … 8=2%), extended downward from the old report ladder. The old
-  // CodeAnalysisView / npm-viewer maps (7→10%/15%, 6→25%/30%) were drift, not
-  // intent.
-  scoreBands: [
-    { score: 11, label: "Top 0.01%" },
-    { score: 10, label: "Top 0.1%" },
-    { score: 9, label: "Top 0.5%" },
-    { score: 8, label: "Top 2%" },
-    { score: 7, label: "Top 5%" },
-    { score: 6, label: "Top 15%" },
-    { score: 5, label: "Top 50%" },
-    { score: 4, label: "Top 65%" },
-    { score: 3, label: "Top 80%" },
-    { score: 2, label: "Top 90%" },
-    { score: 1, label: "Top 97%" }
-  ],
-  minRankedScore: 6,
-  benchmarkBands: [
-    {
-      level: "11",
-      name: "Superhuman",
-      rank: "Top 0.01%",
-      oneIn: "1 in 10,000+",
-      meaning: "Beyond the normal human ceiling for the trait \u2014 the genuine power-law outlier. Almost no one earns this; when the evidence shows it, it is scored, not rounded down.",
-      foundIn: [
-        "Fields Medal and Nobel-track researchers",
-        "Founders of generational companies",
-        "The handful of people a frontier field is named after"
-      ],
-      maxTopPct: 0.01
-    },
-    {
-      level: "10",
-      name: "World-class",
-      rank: "Top 0.1%",
-      oneIn: "1 in 1,000",
-      meaning: "Among the best alive at this trait \u2014 the level entire institutions are built to find.",
-      foundIn: [
-        "Researchers at frontier AI labs",
-        "Faculty at top-5 research universities",
-        "IMO / IOI medalists",
-        "Founders backed by the top decile of venture firms"
-      ],
-      maxTopPct: 0.1
-    },
-    {
-      level: "9",
-      name: "Exceptional",
-      rank: "Top 1%",
-      oneIn: "1 in 100",
-      meaning: "Clearly exceptional \u2014 the strongest person on most strong teams.",
-      foundIn: [
-        "PhD students at top programs",
-        "Early engineers at breakout startups",
-        "YC-class founders",
-        "National olympiad finalists"
-      ],
-      maxTopPct: 1
-    },
-    {
-      level: "8",
-      name: "Strong",
-      rank: "Top 10%",
-      oneIn: "1 in 10",
-      meaning: "Strong against the whole population \u2014 the best person in most rooms, not every room.",
-      foundIn: [
-        "Senior engineers at selective tech companies",
-        "Graduates of demanding technical programs",
-        "Operators who get promoted everywhere they go"
-      ],
-      maxTopPct: 10
-    },
-    {
-      level: "5",
-      name: "Median",
-      rank: "50th percentile",
-      oneIn: "1 in 2",
-      meaning: 'The middle of the general population. The reference class is all ~8 billion humans \u2014 so even "strong" above already means top 10% of everyone.',
-      foundIn: ["The general population \u2014 most people, most places"],
-      maxTopPct: 100
-    }
-  ],
-  codingTiers: {
-    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." },
-    hours: { median: 3, sigma: 0.4, tag: "estimated \u2014 proxy-anchored", source: "median: RescueTime-measured ~2h48m of real productive time/day (same source as the flow benchmark) \u2248 3h of active building; shape: sustaining ~7.5h/day \u2248 top 1%, a ~12h/day average \u2248 1-in-3,000. AVERAGE work-day length, 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" }
-  },
-  codingBenchmarks: {
-    flow: {
-      // typical knowledge worker ≈ 35% of an 8h day focused; ~4h/day is the
-      // recognized deep-work ceiling (Newport) ≈ 50% — most sit far below it.
-      typicalPctOfDay: 0.35,
-      topPctOfDay: 0.5,
-      tag: "measured",
-      source: "RescueTime 2019 (~2h48m focused/day) \xB7 Cal Newport, Deep Work (~4h/day deep-work ceiling)",
-      line: "the best spend ~half the workday in true deep flow; ~4h/day is the human ceiling"
-    },
-    longestRun: {
-      typicalHours: 0.67,
-      // ~40 min before the average worker is interrupted
-      topHours: 4,
-      tag: "measured",
-      source: "RescueTime (avg max ~40 min focus before interruption) \xB7 deep-work single-block ceiling ~3\u20134h",
-      line: "a top performer can hold a single ~3\u20134h uninterrupted block; the average worker breaks at ~40 min"
-    },
-    parallelism: {
-      topMaxConcurrent: 12,
-      // Boris Cherny ~10–15 interactive sessions
-      topAvgConcurrent: 6,
-      // time-weighted average while active (anecdotal, same source)
-      worktreesTip: "3\u20135 git worktrees at once",
-      tag: "anecdotal",
-      source: "Boris Cherny (Claude Code lead, Anthropic): ~10\u201315 concurrent sessions, 'dozens of Claudes running at all times'; 3\u20135 worktrees = 'the single biggest productivity unlock'",
-      line: "the Claude Code lead runs 10\u201315 sessions at once; his #1 tip is 3\u20135 git worktrees in parallel"
-    },
-    throughput: {
-      // 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",
-      loopsDetail: "the Bun runtime was ported Zig\u2192Rust largely autonomously \u2014 hundreds of parallel subagents, two AI reviewers per file",
-      tag: "measured",
-      source: "The Register, May 2026 (Bun Zig\u2192Rust rewrite)",
-      line: "autonomous fleets ported >1M lines of Rust at 99.8% tests passing; a while-loop agent delivered a $50k contract for $297 of compute"
-    }
-  }
-};
-
 // ../../lib/agents/coding/benchmarks.ts
 var BEST = DEFAULT_CALIBRATION.codingBenchmarks;
 
@@ -16005,7 +16011,7 @@ var readJson = async (f, d) => {
     return d;
   }
 };
-var SYS = "Write a SHORT, punchy 'In a nutshell' \u2014 TWO LINES, 2-3 sentences MAX, no more. Structure: (1) what they BUILT in plain terms and why it is genuinely RARE, with a CONCRETE percentile ('top X% of builders/people'); (2) the single clearest place to improve. BE GENEROUS AND EMPHATIC ABOUT REAL ACHIEVEMENT \u2014 do not undersell or hedge a top-tier result. When a score is genuinely top-tier (\u2248 top 2% or rarer), say so plainly and make them FEEL it: name that this is NOT ordinary, that very few people can do it, that it is exceptional. A top-2% abstraction ceiling, for instance, means they can hold an entire system in their head without a blueprint \u2014 that is rare; say it like it is rare. Earned praise should land HARD. Voice: confident, warm, human, like 'you built X end-to-end and barely needed help, which very few people can do', then 'the one place to level up: Y'. NO clinical hedging, NO run-on clauses, NO laundry lists of features. The praise must be EARNED and SPECIFIC, tied to the concrete thing they actually pulled off, never generic horoscope or empty flattery. STYLE (hard rules): NO EM DASHES (\u2014), they are the clearest AI-slop tell; use a period, comma, or colon instead (target zero). No jargon, no fluff. Never invent a feature, number, or outcome that isn't in the evidence. This is the FIRST thing they read; make them feel genuinely seen AND genuinely impressed in two clean lines.";
+var SYS = "Write a SHORT, punchy 'In a nutshell' \u2014 TWO LINES, 2-3 sentences MAX, no more. Structure: (1) what they BUILT in plain terms and why it is genuinely RARE, with a CONCRETE percentile ('top X% of builders/people') when the KEY READS provide one; (2) the single clearest place to improve. PLACEMENT LAW (hard rule): a percentile or 'top X%' placement may ONLY RESTATE a placement already present in the KEY READS input \u2014 never invent, estimate, or extrapolate one. If the input says 0 sessions have been graded, quality is UNMEASURED: make NO quality or placement claim at all (no 'top X%', no rarity claim about their skill), speak only to the measured activity (hours, output volume, cadence), and say plainly that the quality read fills in once grading has run. BE GENEROUS AND EMPHATIC ABOUT REAL ACHIEVEMENT \u2014 do not undersell or hedge a top-tier result. When a score is genuinely top-tier (\u2248 top 2% or rarer), say so plainly and make them FEEL it: name that this is NOT ordinary, that very few people can do it, that it is exceptional. A top-2% abstraction ceiling, for instance, means they can hold an entire system in their head without a blueprint \u2014 that is rare; say it like it is rare. Earned praise should land HARD. Voice: confident, warm, human, like 'you built X end-to-end and barely needed help, which very few people can do', then 'the one place to level up: Y'. NO clinical hedging, NO run-on clauses, NO laundry lists of features. The praise must be EARNED and SPECIFIC, tied to the concrete thing they actually pulled off, never generic horoscope or empty flattery. STYLE (hard rules): NO EM DASHES (\u2014), they are the clearest AI-slop tell; use a period, comma, or colon instead (target zero). No jargon, no fluff. Never invent a feature, number, or outcome that isn't in the evidence. This is the FIRST thing they read; make them feel genuinely seen AND genuinely impressed in two clean lines.";
 function prompt(days, metrics, siRef) {
   return `WHAT THEY DID, DAY BY DAY (newest first):
 ${days}
@@ -16027,8 +16033,11 @@ async function main() {
   const a = agg.ability, t = agg.throughput, f = agg.flow;
   const exp = await readJson("expertise.json", {});
   const del = await readJson("delegation-rollup.json", {});
+  const gradedN = (await fs5.readdir(path12.join(CODING, "grades")).catch(() => [])).filter((f2) => f2.endsWith(".json")).length;
+  const band = (sc) => typeof sc === "number" ? ` (~${bandLabel(sc, DEFAULT_CALIBRATION)})` : "";
   const metrics = [
-    `Abstraction ceiling ${a?.abstraction?.score}/11 (~Top 2%), Taste ${a?.taste?.score}/11`,
+    `Graded sessions: ${gradedN}${gradedN === 0 ? " \u2014 ZERO sessions have been graded; every quality score is absent and quality is UNMEASURED" : ""}`,
+    `Abstraction ceiling ${a?.abstraction?.score != null ? `${a.abstraction.score}/11${band(a.abstraction.score)}` : "not graded"}, Taste ${a?.taste?.score != null ? `${a.taste.score}/11${band(a.taste.score)}` : "not graded"}`,
     `Throughput avg ${t?.score}/11 (~${t?.avgWordsPerDay} words/day on real coding days), peak ${t?.ceiling}`,
     `Flow ${f?.flow?.pctOfWorkDay != null ? Math.round(f.flow.pctOfWorkDay * 100) + "% of the work day" : "?"}, deepest solo`,
     `Genuine domain expertise (deep): ${(exp.domains || []).filter((d) => d.depth === "deep").map((d) => d.domain).join(", ") || "\u2014"}`,

package/dist/pipeline/coding-walkthrough.js

@@ -15694,6 +15694,97 @@ function isHarnessEntry(e) {
   return e.isMeta === true || e.isCompactSummary === true;
 }
 
+// ../coding-core/dist/codex.js
+var SYSTEM_REMINDER_RE2 = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
+var CODEX_INJECTED_TEXT_RE = /^\s*(<(environment_context|user_instructions|ENVIRONMENT_CONTEXT|turn_aborted)\b|#\s*AGENTS\.md instructions\b)/;
+function isCodexInjectedUserText(text2) {
+  return CODEX_INJECTED_TEXT_RE.test(text2);
+}
+function codexContentText(content) {
+  if (typeof content === "string")
+    return content;
+  if (!Array.isArray(content))
+    return "";
+  const parts = [];
+  for (const item of content) {
+    if (typeof item === "string") {
+      parts.push(item);
+      continue;
+    }
+    if (item && typeof item === "object") {
+      const it = item;
+      if (it.type === "input_text" || it.type === "output_text" || it.type === "text")
+        parts.push(it.text || "");
+    }
+  }
+  return parts.join("\n");
+}
+function sniffCodexRaw(raw) {
+  let checked = 0;
+  for (const lineRaw of raw.split("\n")) {
+    const line = lineRaw.trim();
+    if (!line)
+      continue;
+    if (checked++ >= 5)
+      break;
+    try {
+      const e = JSON.parse(line);
+      if (e.type === "session_meta" || e.type === "response_item" || e.type === "turn_context")
+        return true;
+      if (e.type === "user" || e.type === "assistant" || e.type === "summary" || e.type === "queue-operation")
+        return false;
+    } catch {
+    }
+  }
+  return false;
+}
+var EXIT_CODE_RE = /(?:Process exited with code|Exit code:)\s+(\d+)/;
+function extractCodexEvents(raw) {
+  const out = [];
+  for (const lineRaw of raw.split("\n")) {
+    const line = lineRaw.trim();
+    if (!line)
+      continue;
+    let e;
+    try {
+      e = JSON.parse(line);
+    } catch {
+      continue;
+    }
+    if (e.type !== "response_item")
+      continue;
+    const p = e.payload ?? {};
+    const t = typeof e.timestamp === "string" ? Date.parse(e.timestamp) : NaN;
+    const ptype = p.type;
+    if (ptype === "message") {
+      const role = p.role;
+      const text2 = codexContentText(p.content).replace(SYSTEM_REMINDER_RE2, "").trim();
+      if (!text2)
+        continue;
+      if (role === "user") {
+        if (isCodexInjectedUserText(text2))
+          continue;
+        out.push({ kind: "user", t, text: text2 });
+      } else if (role === "assistant") {
+        out.push({ kind: "assistant", t, text: text2 });
+      }
+      continue;
+    }
+    if (ptype === "function_call" || ptype === "custom_tool_call" || ptype === "local_shell_call") {
+      const name17 = typeof p.name === "string" && p.name ? p.name : ptype === "local_shell_call" ? "shell" : String(ptype);
+      const input = typeof p.arguments === "string" ? p.arguments : typeof p.input === "string" ? p.input : "";
+      out.push({ kind: "tool", t, name: name17, input });
+      continue;
+    }
+    if (ptype === "function_call_output" || ptype === "custom_tool_call_output") {
+      const output = typeof p.output === "string" ? p.output : "";
+      const m = EXIT_CODE_RE.exec(output.slice(0, 200));
+      out.push({ kind: "tool_output", t, output, isError: !!m && m[1] !== "0" });
+    }
+  }
+  return out;
+}
+
 // ../coding-core/dist/messages.js
 function extractText(content) {
   if (typeof content === "string")
@@ -15705,6 +15796,49 @@ function extractText(content) {
 function isToolResultOnly(content) {
   return Array.isArray(content) && content.length > 0 && content.every((c) => c && typeof c === "object" && c.type === "tool_result");
 }
+function extractMessagesCodex(raw) {
+  const out = [];
+  let aiProse = [];
+  let aiTools = {};
+  let aiStart = 0;
+  const flushAI = () => {
+    if (!aiProse.length && Object.keys(aiTools).length === 0)
+      return;
+    const toolStr = Object.entries(aiTools).map(([n, c]) => c > 1 ? `${n}\xD7${c}` : n).join(", ");
+    let prose = aiProse.join(" ").replace(/\s+/g, " ").trim();
+    if (prose.length > 2200)
+      prose = prose.slice(0, 2200) + "\u2026";
+    let text2 = prose;
+    if (toolStr)
+      text2 = (text2 ? text2 + " " : "") + `[ran ${toolStr}]`;
+    if (text2)
+      out.push({ t: aiStart, role: "ai", text: text2 });
+    aiProse = [];
+    aiTools = {};
+  };
+  for (const ev of extractCodexEvents(raw)) {
+    if (!Number.isFinite(ev.t))
+      continue;
+    if (ev.kind === "user") {
+      const text2 = humanTypedText(ev.text);
+      if (text2) {
+        flushAI();
+        out.push({ t: ev.t, role: "user", text: text2 });
+      }
+    } else if (ev.kind === "assistant") {
+      if (!aiProse.length && Object.keys(aiTools).length === 0)
+        aiStart = ev.t;
+      if (ev.text)
+        aiProse.push(ev.text);
+    } else if (ev.kind === "tool") {
+      if (!aiProse.length && Object.keys(aiTools).length === 0)
+        aiStart = ev.t;
+      aiTools[ev.name] = (aiTools[ev.name] ?? 0) + 1;
+    }
+  }
+  flushAI();
+  return out.sort((a, b) => a.t - b.t);
+}
 async function extractMessages(file2) {
   let raw;
   try {
@@ -15712,6 +15846,8 @@ async function extractMessages(file2) {
   } catch {
     return [];
   }
+  if (sniffCodexRaw(raw))
+    return extractMessagesCodex(raw);
   const out = [];
   const userTexts = /* @__PURE__ */ new Set();
   const seenUuids = /* @__PURE__ */ new Set();
@@ -16298,10 +16434,12 @@ async function main() {
   if (todo.length < eligible.length) console.log(`[walkthrough] cap: the ${todo.length} richest of ${eligible.length} eligible sessions (POLYMATH_MAX_GRADED; 0 = uncapped)`);
   const lim = limiter(CONC);
   startRun("walkthrough", pending.length);
+  let doneN = 0;
   await Promise.all(pending.map((s) => lim(async () => {
     try {
       const w = await buildWalkthrough(s, { model: MODEL, outDir: DIR });
       tick(!!w);
+      if (w) doneN++;
       console.log(`  \u2022 ${s.title.slice(0, 46)} \u2192 ${w ? w.phases.length + " phases" : "skip"}`);
     } catch (e) {
       tick(false);
@@ -16309,6 +16447,12 @@ async function main() {
     }
   })));
   endRun();
+  if (pending.length > 0 && doneN === 0) {
+    console.error(`
+[walkthrough] FATAL: 0 of ${pending.length} attempted sessions produced a walkthrough \u2014 every attempt failed.`);
+    console.error(`[walkthrough] Likely cause: unreadable or unsupported transcript files (a log format the reader cannot parse). Fix ingestion, then re-run this stage.`);
+    process.exit(1);
+  }
   console.log(`[walkthrough] done \u2192 ${path11.relative(process.cwd(), DIR)}/`);
 }
 main().catch((e) => {

package/dist/web/app.js

@@ -8561,6 +8561,10 @@ function bandLabel(score, doc = DEFAULT_CALIBRATION) {
   for (const b of doc.scoreBands) if (b.score <= s && (!best || b.score > best.score)) best = b;
   return (best ?? doc.scoreBands[0]).label;
 }
+function topShare(percentile) {
+  const t = Math.max(0.1, 100 - percentile);
+  return t < 1 ? String(+t.toFixed(2)) : String(Math.round(t));
+}
 function validateCalibration(raw) {
   const fail = (msg) => {
     throw new Error(`invalid calibration doc: ${msg}`);
@@ -10239,10 +10243,6 @@ function FeelSeenConclusionPanel({ data }) {
   };
   const rows = (data.stackUp ?? []).map((r) => ({ ...r, percentile: livePctl(r) })).filter((r) => r.percentile != null && r.score != null);
   if (!rows.length) return null;
-  const topPct2 = (p) => {
-    const t = 100 - p;
-    return t < 1 ? +t.toFixed(2) : Math.round(t);
-  };
   const ranked = rows.slice().sort((a, b) => b.percentile - a.percentile);
   const standout = ranked.filter((r) => r.percentile >= 90);
   const lead = (standout.length ? standout : ranked.slice(0, 2)).slice(0, 4);
@@ -10262,7 +10262,7 @@ function FeelSeenConclusionPanel({ data }) {
       " \u2014 drop you into any small startup and you\u2019d independently own the outcome and just get the thing done, relentlessly. That\u2019s the rarest, most load-bearing thing here, and it\u2019s real: ",
       /* @__PURE__ */ (0, import_jsx_runtime6.jsxs)("span", { className: "font-semibold text-accent", children: [
         "top ",
-        topPct2(agencyRow.percentile),
+        topShare(agencyRow.percentile),
         "%"
       ] }),
       " on agency",
@@ -10270,7 +10270,7 @@ function FeelSeenConclusionPanel({ data }) {
         ", and ",
         /* @__PURE__ */ (0, import_jsx_runtime6.jsxs)("span", { className: "font-semibold text-accent", children: [
           "top ",
-          topPct2(best.percentile),
+          topShare(best.percentile),
           "%"
         ] }),
         " on ",
@@ -10281,7 +10281,7 @@ function FeelSeenConclusionPanel({ data }) {
       "The honest, generous read: you\u2019re genuinely ",
       /* @__PURE__ */ (0, import_jsx_runtime6.jsxs)("span", { className: "font-semibold text-accent", children: [
         "top ",
-        topPct2(best.percentile),
+        topShare(best.percentile),
         "%"
       ] }),
       " at ",
@@ -10292,7 +10292,7 @@ function FeelSeenConclusionPanel({ data }) {
       /* @__PURE__ */ (0, import_jsx_runtime6.jsx)("span", { className: "font-semibold text-ink", children: r.label }),
       /* @__PURE__ */ (0, import_jsx_runtime6.jsxs)("span", { className: "font-semibold text-accent", children: [
         "top ",
-        topPct2(r.percentile),
+        topShare(r.percentile),
         "%"
       ] })
     ] }, r.label)) }),
@@ -10337,8 +10337,7 @@ function StackUpPanel({ rows }) {
     /* @__PURE__ */ (0, import_jsx_runtime6.jsx)("div", { className: "mt-0.5 text-[12.5px] text-faint", children: "where you land against everyone who works in coding agents" }),
     /* @__PURE__ */ (0, import_jsx_runtime6.jsx)("ul", { className: "mt-3 space-y-3.5", children: shown.map((r, i) => {
       const pct = r.percentile;
-      const top = pct == null ? null : 100 - pct;
-      const topStr = top == null ? null : top < 1 ? String(+top.toFixed(2)) : String(Math.round(top));
+      const topStr = pct == null ? null : topShare(pct);
       return /* @__PURE__ */ (0, import_jsx_runtime6.jsxs)("li", { className: "flex gap-2.5", children: [
         /* @__PURE__ */ (0, import_jsx_runtime6.jsx)("span", { className: "mt-[8px] h-1.5 w-1.5 flex-none rounded-full bg-ink" }),
         /* @__PURE__ */ (0, import_jsx_runtime6.jsxs)("div", { className: "min-w-0 flex-1", children: [

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polymath-society",
-  "version": "0.2.19",
+  "version": "0.2.20",
   "description": "Deterministic metrics from your local Claude Code / Codex agent logs — parallelism, flow, throughput, projects. Zero LLM, zero server, on-device.",
   "license": "UNLICENSED",
   "type": "module",