polymath-society 0.2.5 → 0.2.6

package/dist/pipeline/coding-frontier-detail.js

@@ -134,9 +134,7 @@ var require_secure_json_parse = __commonJS({
 });
 
 // ../../scripts/coding-frontier-detail.mts
-import { promises as fs5 } from "fs";
-import os2 from "os";
-import path10 from "path";
+import { promises as fs7 } from "fs";
 
 // ../../lib/agents/shared/agent.ts
 import { readFileSync } from "fs";
@@ -15608,6 +15606,10 @@ function invokeAgent(inv, opts = {}) {
   return withLimitRetry(once, { label });
 }
 
+// ../../lib/agents/coding/docsDir.ts
+import { promises as fs5 } from "fs";
+import path9 from "path";
+
 // ../../lib/agents/coding/materialize.ts
 import { promises as fs4 } from "fs";
 
@@ -15751,13 +15753,100 @@ async function materializeSession(file2) {
   return { text: lines.join("\n"), humanTurns, humanChars };
 }
 
+// ../../lib/agents/coding/docsDir.ts
+async function writeScoreboard(codingDir, docs) {
+  const gradesDir = path9.join(codingDir, "grades");
+  let files = [];
+  try {
+    files = (await fs5.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 fs5.readFile(path9.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 fs5.writeFile(
+    path9.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.
+
+${blocks.join("\n\n")}
+`
+  );
+}
+async function ensureCodingDocs(codingDir, sessions) {
+  const docs = path9.join(codingDir, "docs");
+  await fs5.mkdir(docs, { recursive: true });
+  const user = sessions.filter((s) => s.sessionId && !s.duplicateOf && (!("klass" in s) || s.klass === "interactive"));
+  let written = 0;
+  for (const s of user) {
+    const out = path9.join(docs, `${s.sessionId}.txt`);
+    try {
+      await fs5.access(out);
+      continue;
+    } catch {
+    }
+    const mat = await materializeSession(s.file).catch(() => null);
+    if (!mat?.text) continue;
+    await fs5.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 fs5.writeFile(path9.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 } from "crypto";
+import { promises as fs6 } 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 fs6.readFile(outPath, "utf8"));
+    if (o[key] === sha && valid(o)) return o;
+  } catch {
+  }
+  return null;
+}
+
 // ../../lib/agents/coding/profile.ts
-import path9 from "path";
+import path10 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
@@ -15906,7 +15995,7 @@ 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).
@@ -15995,7 +16084,7 @@ 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" }
   },
@@ -16028,12 +16117,18 @@ 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",
@@ -16050,8 +16145,8 @@ var BEST = DEFAULT_CALIBRATION.codingBenchmarks;
 
 // ../../lib/agents/coding/profile.ts
 var ROOT = process.cwd();
-var CODING_DIR = process.env.POLYMATH_DATA_DIR ? path9.join(process.env.POLYMATH_DATA_DIR, "coding") : path9.join(ROOT, ".data", "coding");
-var GRADES = path9.join(CODING_DIR, "grades");
+var CODING_DIR = process.env.POLYMATH_DATA_DIR ? path10.join(process.env.POLYMATH_DATA_DIR, "coding") : path10.join(ROOT, ".data", "coding");
+var GRADES = path10.join(CODING_DIR, "grades");
 
 // ../../scripts/coding-frontier-detail.mts
 var CODING = CODING_DIR;
@@ -16068,7 +16163,7 @@ var isUuid = (s) => /^[0-9a-f]{8}-?[0-9a-f]/.test(s);
 async function parseSkillCalls(file2) {
   let raw;
   try {
-    raw = await fs5.readFile(file2, "utf8");
+    raw = await fs7.readFile(file2, "utf8");
   } catch {
     return [];
   }
@@ -16091,7 +16186,7 @@ async function parseSkillCalls(file2) {
   return out;
 }
 async function main() {
-  const all = JSON.parse(await fs5.readFile(`${CODING}/sessions.json`, "utf8"));
+  const all = JSON.parse(await fs7.readFile(`${CODING}/sessions.json`, "utf8"));
   const sessions = all.filter((s) => s.klass === "interactive");
   const dateOf = (s) => (s.start || "").slice(0, 10);
   const byActive = (a, b) => (b.activeMin ?? 0) - (a.activeMin ?? 0);
@@ -16156,20 +16251,41 @@ Assess each. Output a SINGLE fenced json block (every string a complete punctuat
 } 
 \`\`\`
 Put anything with verdict "tried-failed" into BOTH its list (with the verdict) AND triedButFailed. Use only the exact skill/capability names given.`;
-  console.log(`frontier-detail \u2014 ${skillsUsed.length} skills, ${CAP_TARGETS.length} capability candidates, ${mcpServers.length} MCP \u2192 1 Sonnet call`);
-  const sandbox = await fs5.mkdtemp(path10.join(os2.tmpdir(), "frontdet-"));
+  const SYSX = SYS + sourceLookupNote();
+  const sha = promptSha([SYSX, prompt, "sonnet"]);
+  if (!process.argv.includes("--refresh")) {
+    const prev = await reusableOutput(`${CODING}/frontier-detail.json`, sha, (o) => Array.isArray(o.skills));
+    if (prev) {
+      console.log(`[frontier-detail] inputs unchanged \u2014 reusing output from ${prev.generatedAt} (--refresh to regenerate)`);
+      return;
+    }
+  }
+  await ensureCodingDocs(CODING, sessions);
+  console.log(`frontier-detail \u2014 ${skillsUsed.length} skills, ${CAP_TARGETS.length} capability candidates, ${mcpServers.length} MCP \u2192 1 Sonnet call (+source lookups)`);
   let detail = { skills: [], capabilities: [], triedButFailed: [] };
+  let ok = false;
   try {
-    const run2 = await invokeAgent({ prompt, systemPrompt: SYS, addDir: sandbox, allowedTools: ["Read"], maxTurns: 4, timeoutMs: 6e5, model: "sonnet" });
-    detail = run2.json || detail;
+    const run2 = await invokeAgent({ prompt, systemPrompt: SYSX, addDir: CODING, allowedTools: ["Read", "Grep", "Glob"], maxTurns: 10, timeoutMs: 6e5, model: "sonnet" });
+    if (run2.json) {
+      detail = run2.json;
+      ok = true;
+    }
   } catch (e) {
     console.log("ERR", e.message.slice(0, 160));
-  } finally {
-    await fs5.rm(sandbox, { recursive: true, force: true }).catch(() => {
-    });
   }
-  const out = { generatedAt: (/* @__PURE__ */ new Date()).toISOString(), skillsUsed, mcpServers, ...detail };
-  await fs5.writeFile(`${CODING}/frontier-detail.json`, JSON.stringify(out, null, 2));
+  if (!ok) {
+    const existing = await fs7.readFile(`${CODING}/frontier-detail.json`, "utf8").then(JSON.parse).catch(() => null);
+    if (existing && (existing.skills || []).length) {
+      console.error(`REFUSING to write ${CODING}/frontier-detail.json \u2014 the LLM pass failed and the existing file has ${existing.skills.length} assessed skill(s)`);
+      process.exit(1);
+    }
+  }
+  const out = { generatedAt: (/* @__PURE__ */ new Date()).toISOString(), ...ok ? { promptSha: sha } : {}, skillsUsed, mcpServers, ...detail };
+  await fs7.writeFile(`${CODING}/frontier-detail.json`, JSON.stringify(out, null, 2));
+  if (!ok) {
+    console.error(`frontier-detail LLM pass failed \u2014 wrote deterministic portion only (skillsUsed/mcpServers), exiting non-zero`);
+    process.exit(1);
+  }
   const sk = detail.skills || [];
   const surf = sk.filter((x) => x.verdict === "core" || x.verdict === "occasional").length;
   console.log(`WROTE ${CODING}/frontier-detail.json \u2014 skills assessed:${sk.length} (surfaced:${surf}) \xB7 caps:${(detail.capabilities || []).length} \xB7 triedFailed:${(detail.triedButFailed || []).length}`);

package/dist/pipeline/coding-frontier.js

@@ -203,7 +203,7 @@ 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).
@@ -292,7 +292,7 @@ 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" }
   },
@@ -325,12 +325,18 @@ 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",

package/dist/pipeline/coding-gap-dist.js

@@ -394,7 +394,7 @@ var IDLE_CAP_MS = 12 * 6e4;
 
 // ../../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).
@@ -483,7 +483,7 @@ 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" }
   },
@@ -516,12 +516,18 @@ 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",