polymath-society 0.2.18 → 0.2.20

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

@@ -85,6 +85,97 @@ 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(text) {
+  return CODEX_INJECTED_TEXT_RE.test(text);
+}
+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 text = codexContentText(p.content).replace(SYSTEM_REMINDER_RE2, "").trim();
+      if (!text)
+        continue;
+      if (role === "user") {
+        if (isCodexInjectedUserText(text))
+          continue;
+        out.push({ kind: "user", t, text });
+      } else if (role === "assistant") {
+        out.push({ kind: "assistant", t, text });
+      }
+      continue;
+    }
+    if (ptype === "function_call" || ptype === "custom_tool_call" || ptype === "local_shell_call") {
+      const name = 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, 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")
@@ -96,6 +187,49 @@ 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 text = prose;
+    if (toolStr)
+      text = (text ? text + " " : "") + `[ran ${toolStr}]`;
+    if (text)
+      out.push({ t: aiStart, role: "ai", text });
+    aiProse = [];
+    aiTools = {};
+  };
+  for (const ev of extractCodexEvents(raw)) {
+    if (!Number.isFinite(ev.t))
+      continue;
+    if (ev.kind === "user") {
+      const text = humanTypedText(ev.text);
+      if (text) {
+        flushAI();
+        out.push({ t: ev.t, role: "user", text });
+      }
+    } 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(file) {
   let raw;
   try {
@@ -103,6 +237,8 @@ async function extractMessages(file) {
   } catch {
     return [];
   }
+  if (sniffCodexRaw(raw))
+    return extractMessagesCodex(raw);
   const out = [];
   const userTexts = /* @__PURE__ */ new Set();
   const seenUuids = /* @__PURE__ */ new Set();

package/dist/pipeline/coding-gap.js

@@ -15811,6 +15811,97 @@ 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")
@@ -15822,6 +15913,22 @@ 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 extractUserMessagesCodex(raw) {
+  const out = [];
+  let lastTs = null;
+  for (const ev of extractCodexEvents(raw)) {
+    const t = ev.t;
+    const gapMs = Number.isFinite(t) && lastTs != null ? t - lastTs : NaN;
+    if (Number.isFinite(t))
+      lastTs = t;
+    if (ev.kind !== "user")
+      continue;
+    const text2 = humanTypedText(ev.text);
+    if (text2 && Number.isFinite(t))
+      out.push({ t, text: text2, gapMs });
+  }
+  return out.sort((a, b) => a.t - b.t);
+}
 async function extractUserMessages(file2) {
   let raw;
   try {
@@ -15829,6 +15936,8 @@ async function extractUserMessages(file2) {
   } catch {
     return [];
   }
+  if (sniffCodexRaw(raw))
+    return extractUserMessagesCodex(raw);
   const out = [];
   const userTexts = /* @__PURE__ */ new Set();
   const seenUuids = /* @__PURE__ */ new Set();
@@ -16488,6 +16597,18 @@ function distOf(takes) {
   const hist = [...h.keys()].sort((a, b) => b - a).map((level) => ({ level, count: h.get(level) }));
   return { n, observed, mean, median: median2, hist };
 }
+var phi = (z) => {
+  const t = 1 / (1 + 0.2316419 * Math.abs(z));
+  const d = 0.3989423 * Math.exp(-z * z / 2);
+  let pr = d * t * (0.3193815 + t * (-0.3565638 + t * (1.781478 + t * (-1.821256 + t * 1.330274))));
+  if (z > 0) pr = 1 - pr;
+  return pr;
+};
+var lognPct = (x, c) => x == null || !c || x <= 0 ? null : Math.min(99.9, Math.max(0.1, Math.round(phi(Math.log(x / c.median) / c.sigma) * 1e3) / 10));
+var NO_GRADES_GAP = "no sessions graded for this yet \u2014 this fills in once grading has run";
+function estimateRowParts(score, nTakes, evidenceGap) {
+  return nTakes === 0 && score == null ? { score: null, gap: NO_GRADES_GAP } : { score, gap: evidenceGap };
+}
 async function compileCodingProfile() {
   const allRecords = await readJson(path10.join(CODING_DIR, "sessions.json"), []);
   const dupIds = new Set(allRecords.filter((s) => s.duplicateOf).map((s) => s.sessionId));
@@ -16742,15 +16863,11 @@ async function compileCodingProfile() {
   const unb = ceilOf("unblocking");
   const agency = out2 != null && unb != null ? Math.round((out2 + unb) / 2) : out2 ?? unb;
   const sRow = (label, score, gap2) => ({ label, score, percentile: pctOf(score), gap: gap2 });
-  const tiers = DEFAULT_CALIBRATION.codingTiers;
-  const phi = (z) => {
-    const t = 1 / (1 + 0.2316419 * Math.abs(z));
-    const d = 0.3989423 * Math.exp(-z * z / 2);
-    let pr = d * t * (0.3193815 + t * (-0.3565638 + t * (1.781478 + t * (-1.821256 + t * 1.330274))));
-    if (z > 0) pr = 1 - pr;
-    return pr;
+  const est = (label, score, key, evidenceGap) => {
+    const e = estimateRowParts(score, nTk(key), evidenceGap);
+    return sRow(label, e.score, e.gap);
   };
-  const lognPct = (x, c) => x == null || !c || x <= 0 ? null : Math.round(phi(Math.log(x / c.median) / c.sigma) * 1e3) / 10;
+  const tiers = DEFAULT_CALIBRATION.codingTiers;
   const avgWords = aggregate?.throughput?.avgWordsPerDay ?? null;
   const pushPctl = lognPct(avgWords, tiers?.push);
   const focusPctl = lognPct(flowPctDay, tiers?.focus);
@@ -16769,10 +16886,16 @@ async function compileCodingProfile() {
       null,
       workDayHrs != null ? `~${workDayHrs}h of active building on a typical work day \u2014 engaged builders average ~3h, and sustaining 7h+ day after day, week after week, is top-1% territory` : "not enough dated activity yet to measure a typical work day"
     ),
+    // The copy FOLLOWS the person's actual orchestration band (orchPctl above) —
+    // a hardcoded "single biggest speed-up still on the table" once shipped to a
+    // user at orchPctl 99.5 (18 agents peak, 48% of hours at 2+), flatly
+    // contradicting the gap section's "already strong" on the same page
+    // (2026-07-17). The still-on-the-table claim is reserved for genuinely low
+    // orchestration.
     sRow(
       "Your AI parallelism",
       ceilOf("delegation"),
-      soloPct != null ? `~${soloPct}% of your hours you drive a single agent (${Math.round(soloMin / 60)}h solo vs ${Math.round(parMin / 60)}h with 2+ live) \u2014 the Claude Code lead keeps 10\u201315 going at once, so this is the single biggest speed-up still on the table` : "you brief the AI sharply, but mostly drive one agent at a time \u2014 the top run 10\u201315 in parallel, a several-fold speed-up"
+      soloPct != null ? orchPctl >= 97 ? `you already run parallel at the frontier \u2014 ${Math.round(parMin / 60)}h of your hours with 2+ agents live vs ${Math.round(soloMin / 60)}h solo; the Claude Code lead keeps 10\u201315 going at once, and you work the same way` : `~${soloPct}% of your hours you drive a single agent (${Math.round(soloMin / 60)}h solo vs ${Math.round(parMin / 60)}h with 2+ live) \u2014 the Claude Code lead keeps 10\u201315 going at once, so this is the single biggest speed-up still on the table` : "you brief the AI sharply, but mostly drive one agent at a time \u2014 the top run 10\u201315 in parallel, a several-fold speed-up"
     ),
     // The evidence line cites the RANKING metric (flow as a share of the work
     // day), never absolute hours — a friend with MORE flow hours over a longer
@@ -16788,22 +16911,25 @@ async function compileCodingProfile() {
     // The four estimates LEAD with a concrete moment from their own sessions, then an
     // HONEST caveat — how many graded sessions back the read. No invented weakness, no
     // claim about what they "didn't" do (false specifics churn a real user); it's an estimate.
-    sRow(
+    est(
       "Agency",
       agency,
+      "outcomes",
       `${proud("outcomes", "you ship working systems and unblock yourself")} \u2014 one of ${nTk("outcomes")} sessions behind this read; a strong, consistent signal, still an estimate from your logs`
     ),
     // Judgement/Taste use the SAME numbers their ability cards prove (aggregate
     // peak-with-consistency over near-peak instances) — a chip saying "top 1%"
     // above a card proving 10/Top 0.1% with ten receipts reads as a bug.
-    sRow(
+    est(
       "Judgement",
       aggregate?.ability?.abstraction?.score ?? ceilOf("generative"),
+      "generative",
       `${proud("generative", "you make sharp, non-obvious calls on what actually matters")} \u2014 one of ${nTk("generative")} sessions we've graded for this; clear so far, not yet a final verdict`
     ),
-    sRow(
+    est(
       "Taste",
       aggregate?.ability?.taste?.score ?? ceilOf("taste"),
+      "taste",
       `${proud("taste", "you hold a high, opinionated quality bar")} \u2014 one of ${nTk("taste")} sessions behind this read; a real signal, still an estimate`
     )
     // Expertise: NO chip — the scored criterion is deprecated; the real judge is
@@ -16860,7 +16986,7 @@ 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 SYSTEM_REMINDER_RE3 = /<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>/ },
@@ -16880,7 +17006,7 @@ function collapseInjected(input) {
       return "";
     });
   }
-  return { text: text2.replace(SYSTEM_REMINDER_RE2, "").trim(), markers };
+  return { text: text2.replace(SYSTEM_REMINDER_RE3, "").trim(), markers };
 }
 function extractText2(content) {
   if (typeof content === "string") return content;
@@ -16897,6 +17023,38 @@ 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;
+function materializeCodex(raw) {
+  const lines = [];
+  let humanTurns = 0, humanChars = 0;
+  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 ev of extractCodexEvents(raw)) {
+    if (ev.kind === "user") {
+      flushAI();
+      humanTurns++;
+      humanChars += ev.text.length;
+      lines.push(`>> ${ev.text}`);
+    } else if (ev.kind === "assistant") {
+      const prose = ev.text.replace(/\s+/g, " ").trim();
+      if (prose) aiLastProse = prose;
+    } else if (ev.kind === "tool") {
+      aiTools.set(ev.name, (aiTools.get(ev.name) ?? 0) + 1);
+    }
+  }
+  flushAI();
+  return { text: lines.join("\n"), humanTurns, humanChars };
+}
 async function materializeSession(file2) {
   let raw;
   try {
@@ -16904,6 +17062,7 @@ async function materializeSession(file2) {
   } catch {
     return null;
   }
+  if (sniffCodexRaw(raw)) return materializeCodex(raw);
   const lines = [];
   let humanTurns = 0, humanChars = 0;
   const pendingQueued = [];
@@ -17191,6 +17350,9 @@ async function main() {
     return m ? `--- "${s.title}" \xB7 ${(s.start || "").slice(0, 10)} ---
 ${m.text.slice(0, 3e3)}` : "";
   }))).filter(Boolean);
+  const digestDays = (await fs11.readFile(path13.join(CODING, "day-digest.json"), "utf-8").then(JSON.parse).catch(() => ({ days: [] }))).days ?? [];
+  const digestLines = digestDays.slice().sort((a, b) => a.date < b.date ? -1 : 1).slice(-21).map((d) => `${d.date}: ${d.overall ?? ""}
+${(d.conversations ?? []).slice(0, 4).map((c) => `  - ${c.title ?? ""}: ${c.summary ?? ""}`).join("\n")}`).join("\n").slice(0, 12e3);
   const SYS = "You are a Claude Code power-user coach. You judge ONE person against the COMPLETE catalog of best ways to use Claude Code (the rubric below, sourced from Boris Cherny + first-party Claude Code features) and write a comprehensive, honest gap analysis, ORDERED most-difference-making \u2192 least. Rules: (1) Go through the whole catalog and surface the techniques they are NOT doing (or under-using) that would genuinely help \u2014 respecting each tip's [EMPHASIZE]/[DE-EMPHASIZE]/[GATE] curation. (2) ORDER by differenceRank (1 = biggest unlock for THIS person). Follow the rubric's PRIME DIRECTIVE: if usage.parallelism.alreadyParallel is FALSE \u2192 'Do more in parallel' is rank 1; else \u2192 'Invest in your CLAUDE.md / tune your context' is rank 1. Then Plan mode, then 'Give Claude test cases so it can auto-verify and run longer' (Boris's 'most important thing' \u2014 rank it high, right after plan mode), then subagents, and the rest per the rubric order, skipping any that don't apply. (3) GATE the conditional tips on the STRUCTURAL usage facts, NOT on keywords \u2014 and a tip whose gate FAILS is OMITTED from recommendations entirely, never included as a softened/conditional card (recommending something their data says doesn't apply, or that they already do, reads as not having looked): loops require real PR/branch volume in usage.collaboration, else OMIT; the Chrome tip requires ~zero browser-driving in usage.chrome, else OMIT and credit in alreadyStrong; forking uses the RECENT count in usage.forking \u2014 if they fork in the recent window they already do it, OMIT and credit; remote/mobile/EC2 fires only when usage.surface.workstationBound is TRUE; plan mode by usage.planMode; subagents by usage.subagents. If they already do a thing, it goes in alreadyStrong, NOT in recommendations. (4) \u26D4 CRITICAL \u2014 this person BUILDS Claude Code tooling, so their transcripts are full of the words 'worktree', 'pull request', 'fork', 'plan mode', 'by the way', 'teleport'. A keyword in the transcript is NOT evidence they use the feature. Rely on the STRUCTURAL facts only. (5) Ground every recommendation in a RECEIPT \u2014 a structural fact from their usage or a concrete session pain point \u2014 and CONVINCE with a specific case for how it helps THIS person, not a generic benefit. (6) Tailor every action to the Claude Desktop Mac app; NEVER give terminal-only `claude -p` advice. (7) Honor config.deliberateChoices (they stay on `main` on purpose \u2014 do NOT tell them to hand-juggle worktrees; parallel SESSIONS instead) and config.claudeAutoTools (never tell them to 'use' a tool Claude drives itself; for 'explain with visuals' the user just ASKS for a visual \u2014 do not name the tool). (8) Also name what they ALREADY do well. (9) SELF-CONTAINED but not verbose \u2014 lead with the PROBLEM, then the SOLUTION, then HOW to use it in their context. Plain language, no jargon, no enumerating multiple dates. An essay-like rec is a failed rec; so is a cryptic one. Follow this writing rubric strictly:\n\n" + writing + sourceLookupNote();
   const prompt = `THE RUBRIC \u2014 the complete catalog of best ways to use Claude Code. Judge the person against ALL of it, and ORDER by differenceRank per its PRIME DIRECTIVE:
 
@@ -17200,6 +17362,9 @@ ${techniques}
 WHAT THIS PERSON ACTUALLY DOES (STRUCTURAL usage facts + config \u2014 ground truth; trust these over anything in the transcripts):
 ${JSON.stringify(usage, null, 1)}
 
+THEIR RECENT DAYS, task by task (dated ground material \u2014 draw applyItHere moments from HERE and from the excerpts, never from imagination):
+${digestLines}
+
 SESSION EXCERPTS (for pain points / friction / grounding ONLY \u2014 never as proof they USE a feature; ">> " = the person, "[AI]" = the model):
 ${excerpts.join("\n\n").slice(0, 6e4)}
 
@@ -17207,11 +17372,12 @@ ${excerpts.join("\n\n").slice(0, 6e4)}
 Produce the gap analysis. Be thorough: cover every relevant catalog item they're missing, ordered most-difference-making first. Output a SINGLE fenced json block (every string a complete, punctuated, no-fluff sentence):
 \`\`\`json
 {
-  "recommendations": [ { "title": "Boris's OWN / the feature's general wording for the tip (e.g. 'Invest in your CLAUDE.md', 'Start every complex task in Plan mode', 'Give Claude test cases so it can auto-verify and run longer', 'Code from anywhere with a remote instance', 'Fork your session', 'Use /btw for side queries') \u2014 NOT a hyper-specific name. For the verification tip the title MUST contain 'verify its own work' or 'check its own work' so the report picks the right icon.", "category": "rubric tip", "status": "not-doing|under-using", "differenceRank": 1, "gap": "LEAD WITH THE PROBLEM in plain language \u2014 what it's currently costing them, grounded in a STRUCTURAL fact. One clear, self-contained sentence (~12-18 words), no jargon. If they already do it somewhat, frame it as 'you do X; the gap is pushing it further'.", "whyItHelps": "The SOLUTION, then HOW to use it in THEIR context with a concrete example. Plain language, self-contained, ~30-45 words. Two crisp beats: fix, then their-context application. HARD RULE: no jargon, and NEVER name a Claude-auto tool. EXAMPLE (plan mode): 'Skipping the plan makes Claude pause to ask and drift. Fix: iterate the whole approach in Plan mode first, and drop in concrete expected results as test cases so Claude self-verifies and runs autonomously far longer.'", "payoff": 5 } ],
+  "recommendations": [ { "title": "Boris's OWN / the feature's general wording for the tip (e.g. 'Invest in your CLAUDE.md', 'Start every complex task in Plan mode', 'Give Claude test cases so it can auto-verify and run longer', 'Code from anywhere with a remote instance', 'Fork your session', 'Use /btw for side queries') \u2014 NOT a hyper-specific name. For the verification tip the title MUST contain 'verify its own work' or 'check its own work' so the report picks the right icon.", "category": "rubric tip", "status": "not-doing|under-using", "differenceRank": 1, "gap": "LEAD WITH THE PROBLEM in plain language \u2014 what it's currently costing them, grounded in a STRUCTURAL fact. One clear, self-contained sentence (~12-18 words), no jargon. If they already do it somewhat, frame it as 'you do X; the gap is pushing it further'.", "whyItHelps": "The SOLUTION, then HOW to use it in THEIR context with a concrete example. Plain language, self-contained, ~30-45 words. Two crisp beats: fix, then their-context application. HARD RULE: no jargon, and NEVER name a Claude-auto tool. EXAMPLE (plan mode): 'Skipping the plan makes Claude pause to ask and drift. Fix: iterate the whole approach in Plan mode first, and drop in concrete expected results as test cases so Claude self-verifies and runs autonomously far longer.'", "payoff": 5, "applyItHere": [ { "moment": "A SPECIFIC dated moment from THEIR OWN days/excerpts \u2014 name the repo/project and the actual task in their vocabulary (~15-25 words).", "insteadDo": "How this exact technique changes that exact moment \u2014 concrete, imperative, their context (~15-25 words)." } ] } ],
   "alreadyStrong": [ { "title": "", "note": "What they already do well, concretely." } ]
 }
 \`\`\`
-\`differenceRank\` is a strict 1..N ordering (1 = highest). Assign it per the PRIME DIRECTIVE and the rubric's general order; the report renders recommendations sorted by it. \`payoff\` (1-5) is a secondary strength cue. Recommend ONLY from the catalog.`;
+\`differenceRank\` is a strict 1..N ordering (1 = highest). Assign it per the PRIME DIRECTIVE and the rubric's general order; the report renders recommendations sorted by it. \`payoff\` (1-5) is a secondary strength cue. Recommend ONLY from the catalog.
+\`applyItHere\` HARD RULES: 3 entries per recommendation when the material supports 3, minimum 2 \u2014 and every moment must be VISIBLY present in THEIR RECENT DAYS or the excerpts (a real dated task, named the way they name it). NEVER invent, generalize, or reuse the same moment across recommendations. If the material truly supports only one real moment, give that one \u2014 a fabricated example is worse than a short list.`;
   const sha = promptSha([SYS, prompt, "sonnet"]);
   if (!process.argv.includes("--refresh")) {
     const prev = await reusableOutput(path13.join(CODING, "gap.json"), sha, (o) => Array.isArray(o.recommendations) && o.recommendations.length > 0);

package/dist/pipeline/coding-grade.js

@@ -15863,8 +15863,99 @@ function isHarnessEntry(e) {
   return e.isMeta === true || e.isCompactSummary === true;
 }
 
-// ../../lib/agents/coding/materialize.ts
+// ../coding-core/dist/codex.js
 var SYSTEM_REMINDER_RE = /<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_RE, "").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;
+}
+
+// ../../lib/agents/coding/materialize.ts
+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>/ },
@@ -15884,7 +15975,7 @@ function collapseInjected(input) {
       return "";
     });
   }
-  return { text: text2.replace(SYSTEM_REMINDER_RE, "").trim(), markers };
+  return { text: text2.replace(SYSTEM_REMINDER_RE2, "").trim(), markers };
 }
 function extractText(content) {
   if (typeof content === "string") return content;
@@ -15901,6 +15992,38 @@ function isToolResultOnly(content) {
   return Array.isArray(content) && content.length > 0 && content.every((c) => c && typeof c === "object" && c.type === "tool_result");
 }
 var AI_GIST = 160;
+function materializeCodex(raw) {
+  const lines = [];
+  let humanTurns = 0, humanChars = 0;
+  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 ev of extractCodexEvents(raw)) {
+    if (ev.kind === "user") {
+      flushAI();
+      humanTurns++;
+      humanChars += ev.text.length;
+      lines.push(`>> ${ev.text}`);
+    } else if (ev.kind === "assistant") {
+      const prose = ev.text.replace(/\s+/g, " ").trim();
+      if (prose) aiLastProse = prose;
+    } else if (ev.kind === "tool") {
+      aiTools.set(ev.name, (aiTools.get(ev.name) ?? 0) + 1);
+    }
+  }
+  flushAI();
+  return { text: lines.join("\n"), humanTurns, humanChars };
+}
 async function materializeSession(file2) {
   let raw;
   try {
@@ -15908,6 +16031,7 @@ async function materializeSession(file2) {
   } catch {
     return null;
   }
+  if (sniffCodexRaw(raw)) return materializeCodex(raw);
   const lines = [];
   let humanTurns = 0, humanChars = 0;
   const pendingQueued = [];
@@ -16440,6 +16564,12 @@ async function main() {
     }
   })));
   endRun();
+  if (picked.length > 0 && doneN === 0) {
+    console.error(`
+[coding-grade] FATAL: 0 of ${picked.length} attempted sessions produced a grade \u2014 every attempt failed or returned no usable grade.`);
+    console.error(`[coding-grade] The report must NOT pretend to be graded. 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(`[coding-grade] wrote \u2192 .data/coding/grades/`);
 }
 main().catch((e) => {