polymath-society 0.2.19 → 0.2.20

package/dist/cli.js

@@ -110,6 +110,103 @@ var init_injected = __esm({
   }
 });
 
+// ../coding-core/dist/codex.js
+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;
+}
+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;
+}
+var SYSTEM_REMINDER_RE2, CODEX_INJECTED_TEXT_RE, EXIT_CODE_RE;
+var init_codex = __esm({
+  "../coding-core/dist/codex.js"() {
+    "use strict";
+    SYSTEM_REMINDER_RE2 = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
+    CODEX_INJECTED_TEXT_RE = /^\s*(<(environment_context|user_instructions|ENVIRONMENT_CONTEXT|turn_aborted)\b|#\s*AGENTS\.md instructions\b)/;
+    EXIT_CODE_RE = /(?:Process exited with code|Exit code:)\s+(\d+)/;
+  }
+});
+
 // ../coding-core/dist/raw.js
 import { promises as fs } from "fs";
 import path2 from "path";
@@ -382,7 +479,7 @@ function parseCodex(raw, file2) {
         const role = p.role;
         const text2 = codexText(p.content).replace(SYSTEM_REMINDER_RE, "").trim();
         if (role === "user") {
-          if (text2 && !CODEX_INJECTED_RE.test(text2)) {
+          if (text2 && !isCodexInjectedUserText(text2)) {
             s.humanTurns++;
             s.humanChars += text2.length;
             if (!s.firstHuman)
@@ -774,11 +871,12 @@ async function readAllSessions(opts = {}) {
   demoteTemplateFanouts(all);
   return all;
 }
-var SOURCE_ROOT_ENV_VARS, anyRootOverridden, claudeProjectsRoot, codexSessionsRoot, cursorWorkspaceRoot, cursorProjectsRoot, SKIP_DIR_RE, AGENT_CWD_RE, PERSONA_RE, OUTPUT_CONTRACT_RES, CAPS_HEADER_RE, PING_RE, ms, CODEX_INJECTED_RE, CURSOR_USER_QUERY_RE, NODE_SQLITE, FANOUT_PREFIX, FANOUT_MIN_SESSIONS;
+var SOURCE_ROOT_ENV_VARS, anyRootOverridden, claudeProjectsRoot, codexSessionsRoot, cursorWorkspaceRoot, cursorProjectsRoot, SKIP_DIR_RE, AGENT_CWD_RE, PERSONA_RE, OUTPUT_CONTRACT_RES, CAPS_HEADER_RE, PING_RE, ms, CURSOR_USER_QUERY_RE, NODE_SQLITE, FANOUT_PREFIX, FANOUT_MIN_SESSIONS;
 var init_raw = __esm({
   "../coding-core/dist/raw.js"() {
     "use strict";
     init_injected();
+    init_codex();
     SOURCE_ROOT_ENV_VARS = [
       "CLAUDE_PROJECTS_DIR",
       "CODEX_SESSIONS_DIR",
@@ -807,7 +905,6 @@ var init_raw = __esm({
       const n = Date.parse(ts);
       return Number.isFinite(n) ? n : null;
     };
-    CODEX_INJECTED_RE = /^\s*<(environment_context|user_instructions)>/;
     CURSOR_USER_QUERY_RE = /<user_query>\s*([\s\S]*?)\s*<\/user_query>/;
     NODE_SQLITE = "node:sqlite";
     FANOUT_PREFIX = 80;
@@ -1486,6 +1583,11 @@ async function readManifest(dataDir) {
     exports: { included: arr(ex.included), excluded: arr(ex.excluded) }
   };
 }
+function agentFlagsFromManifest(m) {
+  if (!m?.agents) return { claudeCode: true, codex: true };
+  const on = (id) => m.agents.included.includes(id);
+  return { claudeCode: on("claude-code"), codex: on("codex") };
+}
 async function writeManifest(dataDir, m) {
   await fs22.mkdir(dataDir, { recursive: true });
   await fs22.writeFile(manifestPath(dataDir), JSON.stringify(m, null, 2));
@@ -2182,17 +2284,18 @@ var init_plan = __esm({
 var wizard_exports = {};
 __export(wizard_exports, {
   runWizard: () => runWizard,
-  sayExportLinks: () => sayExportLinks
+  sayExportLinks: () => sayExportLinks,
+  scanJsonlDir: () => scanJsonlDir
 });
 import { promises as fs35 } from "fs";
 import os11 from "os";
 import path40 from "path";
 import * as readline from "node:readline/promises";
-async function scanJsonlDir(label, dir) {
+async function scanJsonlDir(label, dir, maxDepth = 1) {
   let files = 0;
   let newest = 0;
   async function walk(d, depth) {
-    if (depth > 1) return;
+    if (depth > maxDepth) return;
     let names = [];
     try {
       names = await fs35.readdir(d);
@@ -2317,8 +2420,8 @@ async function runWizard() {
     });
     say(bold("  Step 1 \xB7 Your coding agents"));
     const roots = sourceRoots();
-    const scanRoot = (label, dir) => dir ? scanJsonlDir(label, dir) : Promise.resolve({ label, dir: "(disabled \u2014 log roots redirected via env)", files: 0, newest: null });
-    const [cc, cx] = await Promise.all([scanRoot("Claude Code", roots.claudeProjects), scanRoot("Codex", roots.codexSessions)]);
+    const scanRoot = (label, dir, maxDepth) => dir ? scanJsonlDir(label, dir, maxDepth) : Promise.resolve({ label, dir: "(disabled \u2014 log roots redirected via env)", files: 0, newest: null });
+    const [cc, cx] = await Promise.all([scanRoot("Claude Code", roots.claudeProjects, 1), scanRoot("Codex", roots.codexSessions, 3)]);
     const agentIds = ["claude-code", "codex"];
     const agentScans = [cc, cx];
     const present = agentIds.filter((_, i) => agentScans[i].files > 0);
@@ -2392,8 +2495,11 @@ async function runWizard() {
     for (const f of await scanForExports(path40.join(os11.homedir(), "Downloads"))) add(f);
     for (const f of await scanForObsidianVaults(os11.homedir())) add(f);
     for (const f of [...previous?.exports.included ?? [], ...previous?.exports.excluded ?? []]) add(f);
+    let linksShown = false;
     if (!found.length) {
       say(dim("    Nothing found in the usual places."));
+      await sayExportLinks(say, /* @__PURE__ */ new Set());
+      linksShown = true;
       if (await yes(`    Search deeper with your own local agent? ${dim("(capped at ~90s)")}`, false)) {
         try {
           const { invokeAgent: invokeAgent3 } = await Promise.resolve().then(() => (init_agent(), agent_exports));
@@ -2418,7 +2524,7 @@ async function runWizard() {
       const defaults = found.map(() => true);
       say(`    Found ${dim("(identified locally from the files themselves)")}:`);
       found.forEach((f, i) => say(describeExport(f, i, defaults[i], hadPrevious)));
-      await sayExportLinks(say, new Set(found.map((f) => f.kind)));
+      if (!linksShown) await sayExportLinks(say, new Set(found.map((f) => f.kind)));
       const include = await pick2(defaults, hadPrevious);
       const manifest = {
         approvedAt: (/* @__PURE__ */ new Date()).toISOString(),
@@ -2438,7 +2544,7 @@ async function runWizard() {
       }
     } else {
       await writeManifest(dataDir, { approvedAt: (/* @__PURE__ */ new Date()).toISOString(), agents, exports: { included: [], excluded: [] } });
-      await sayExportLinks(say, /* @__PURE__ */ new Set());
+      if (!linksShown) await sayExportLinks(say, /* @__PURE__ */ new Set());
     }
     {
       const { EXPORT_LINKS: EXPORT_LINKS2 } = await Promise.resolve().then(() => (init_exportLinks(), exportLinks_exports));
@@ -2953,6 +3059,7 @@ init_gate2();
 
 // ../coding-core/dist/messages.js
 init_injected();
+init_codex();
 import { promises as fs2 } from "fs";
 function extractText2(content) {
   if (typeof content === "string")
@@ -2964,6 +3071,22 @@ function extractText2(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 {
@@ -2971,6 +3094,8 @@ async function extractUserMessages(file2) {
   } catch {
     return [];
   }
+  if (sniffCodexRaw(raw))
+    return extractUserMessagesCodex(raw);
   const out = [];
   const userTexts = /* @__PURE__ */ new Set();
   const seenUuids = /* @__PURE__ */ new Set();
@@ -3025,6 +3150,49 @@ function createMessageDeduper() {
     return true;
   };
 }
+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 {
@@ -3032,6 +3200,8 @@ async function extractMessages(file2) {
   } catch {
     return [];
   }
+  if (sniffCodexRaw(raw))
+    return extractMessagesCodex(raw);
   const out = [];
   const userTexts = /* @__PURE__ */ new Set();
   const seenUuids = /* @__PURE__ */ new Set();
@@ -3767,7 +3937,7 @@ var LABEL_BY_UIKEY = Object.fromEntries(GRADED_CRITERIA2.map((c) => [c.uiKey, c.
 // src/grade/materialize.ts
 init_injected();
 import { promises as fs3 } 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>/ },
@@ -3787,7 +3957,7 @@ function collapseInjected(input) {
       return "";
     });
   }
-  return { text: text2.replace(SYSTEM_REMINDER_RE2, "").trim(), markers };
+  return { text: text2.replace(SYSTEM_REMINDER_RE3, "").trim(), markers };
 }
 function extractText3(content) {
   if (typeof content === "string") return content;
@@ -21259,8 +21429,9 @@ import path17 from "node:path";
 
 // ../../lib/agents/coding/materialize.ts
 init_injected();
+init_codex();
 import { promises as fs11 } from "fs";
-var SYSTEM_REMINDER_RE3 = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
+var SYSTEM_REMINDER_RE4 = /<system-reminder>[\s\S]*?<\/system-reminder>/g;
 var INJECT_GIST2 = 120;
 var INJECTED2 = [
   { re: /<task-notification>[\s\S]*?<\/task-notification>/g, label: "bg task", gistRe: /<summary>([\s\S]*?)<\/summary>/ },
@@ -21280,7 +21451,7 @@ function collapseInjected2(input) {
       return "";
     });
   }
-  return { text: text2.replace(SYSTEM_REMINDER_RE3, "").trim(), markers };
+  return { text: text2.replace(SYSTEM_REMINDER_RE4, "").trim(), markers };
 }
 function extractText4(content) {
   if (typeof content === "string") return content;
@@ -21297,6 +21468,38 @@ function isToolResultOnly3(content) {
   return Array.isArray(content) && content.length > 0 && content.every((c) => c && typeof c === "object" && c.type === "tool_result");
 }
 var AI_GIST2 = 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_GIST2 ? aiLastProse.slice(0, AI_GIST2) + "\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 materializeSession2(file2) {
   let raw;
   try {
@@ -21304,6 +21507,7 @@ async function materializeSession2(file2) {
   } catch {
     return null;
   }
+  if (sniffCodexRaw(raw)) return materializeCodex(raw);
   const lines = [];
   let humanTurns = 0, humanChars = 0;
   const pendingQueued = [];
@@ -21839,124 +22043,6 @@ function handleDiagnostic(method, body, key2) {
 // src/codingChat.ts
 init_agent();
 
-// ../../lib/agents/coding/chat-context.ts
-var RUNGS_BY_KEY = new Map(CODING_CRITERIA.map((c) => [c.key, c.rungs ?? []]));
-var clip = (s, n) => {
-  const t = (s || "").replace(/\s+/g, " ").trim();
-  return t.length > n ? t.slice(0, n - 1).replace(/\s+\S*$/, "") + "\u2026" : t;
-};
-function criterionBlock(c) {
-  if (!c.graded || c.dist.n <= 0) return null;
-  const lines = [];
-  const mean2 = c.dist.mean != null ? `${c.dist.mean}` : "\u2014";
-  lines.push(`### ${c.label}  (your mean ${mean2} across ${c.dist.n} graded session${c.dist.n === 1 ? "" : "s"})`);
-  lines.push(`What it measures: ${clip(c.definition, 400)}`);
-  const rungs = RUNGS_BY_KEY.get(c.key) ?? [];
-  if (rungs.length) {
-    lines.push(`The ladder I grade against:`);
-    for (const r of rungs) lines.push(`  L${r.level} \u2014 ${r.marker}`);
-  }
-  const takes = (c.takes || []).slice(0, 5);
-  if (takes.length) {
-    lines.push(`How I scored your sessions here:`);
-    for (const t of takes) {
-      const sc = t.score != null ? `${t.score}` : t.best != null ? `~${t.best}` : "\u2014";
-      const parts = [`  \u2022 "${clip(t.title, 80)}" (${(t.date || "").slice(0, 10)}) \u2014 scored ${sc}${t.confidence ? ` (${t.confidence})` : ""}.`];
-      if (t.instance) parts.push(`What showed it: ${clip(t.instance, 240)}`);
-      if (t.why) parts.push(`Why that score: ${clip(t.why, 320)}`);
-      const qs = (t.quotes || []).filter(Boolean).slice(0, 2).map((q) => `"${clip(q, 160)}"`);
-      if (qs.length) parts.push(`You said: ${qs.join("  /  ")}`);
-      lines.push(parts.join(" "));
-    }
-  }
-  return lines.join("\n");
-}
-function buildCodingContext(profile) {
-  const out = [];
-  const t = profile.throughput;
-  const p = profile.parallelism;
-  const meta = [];
-  if (t?.sessions != null) meta.push(`${t.sessions} interactive coding sessions`);
-  if (profile.gradedSessions != null) meta.push(`${profile.gradedSessions} graded`);
-  if (p?.maxConcurrency != null) meta.push(`peak ${p.maxConcurrency} sessions in parallel`);
-  if (meta.length) out.push(`HOW THEY WORK: ${meta.join(" \xB7 ")}.`);
-  const blocks = (profile.criteria || []).map(criterionBlock).filter((b) => !!b);
-  if (blocks.length) {
-    out.push(`
-YOUR GRADED CRITERIA \u2014 the rubric ladder and how I scored you on each, with your own words as the receipts:`);
-    out.push(blocks.join("\n\n"));
-  } else {
-    out.push(`
-(No criteria have been graded yet, so I can only speak to how you work, not the qualitative scores.)`);
-  }
-  return out.join("\n");
-}
-var CODING_CHAT_SYSTEM = `You are the analyst who graded this person's coding-agent sessions (their Claude Code / Codex logs) against a FIXED rubric, and you are now talking WITH them about their report. Below you are given that rubric \u2014 the definition and the level ladder for each criterion \u2014 AND your own per-session grades: the exact moment that keyed each score, WHY you placed it on that rung, and their verbatim quotes. Everything you claim comes from that material.
-
-HOW TO ANSWER:
-- Ground every claim in the receipts you were given. When they ask "why did I get that score on X", name the specific session, the rung you landed them on, what that rung requires, and quote their actual words. Never invent a session, a score, a rung, or a criterion that isn't below.
-- The rubric is the authority. If they push on a score, explain the ladder honestly: what rung they hit and what the NEXT rung would concretely have required of them. Don't inflate a score to be nice and don't apologize for a low one \u2014 being read accurately is the point. If they're genuinely strong, say so plainly.
-- Be honest about the instrument's limits. These grades come only from their coding logs, which UNDER-observe real skill (a lot of thinking never reaches the transcript). When it's relevant, say what you can and can't actually see from the logs rather than overclaiming.
-- Voice: warm, plain, declarative, second person. Specific, never a horoscope, never flattery. Numbers are calibration, not a verdict \u2014 don't recite them at the person.
-- Keep replies tight and human unless they ask you to go deep. Output ONLY your reply \u2014 no preamble, no meta.
-
-Do NOT use em dashes. Use a comma or a period instead.`;
-function stripSlop(raw) {
-  return (raw || "").trim().replace(/\s*—\s*/g, ", ").replace(/\s+–\s+/g, ", ").replace(/\s+,/g, ",");
-}
-function buildCodingChatPrompt(profile, messages) {
-  const context = buildCodingContext(profile);
-  const history = messages.map((m) => `${m.role === "user" ? "Them" : "You"}: ${m.content}`).join("\n\n");
-  return `YOUR READ OF THIS PERSON (the rubric + your grades + their receipts):
-${context}
-
---- the conversation so far ---
-${history}
-
-Write your next reply to them now.`;
-}
-
-// src/codingChat.ts
-async function handleCodingChat(profile, dataDir, body, res) {
-  const raw = Array.isArray(body.messages) ? body.messages : [];
-  const messages = raw.filter((m) => (m.role === "user" || m.role === "assistant") && typeof m.content === "string").slice(-20);
-  if (!messages.length) {
-    res.writeHead(400, { "Content-Type": "application/json; charset=utf-8" });
-    res.end(JSON.stringify({ error: "no messages" }));
-    return;
-  }
-  const prompt = buildCodingChatPrompt(profile, messages);
-  res.writeHead(200, {
-    "Content-Type": "application/x-ndjson; charset=utf-8",
-    "Cache-Control": "no-store, no-transform",
-    "X-Accel-Buffering": "no"
-  });
-  const send = (o) => {
-    try {
-      res.write(JSON.stringify(o) + "\n");
-    } catch {
-    }
-  };
-  try {
-    const run4 = await invokeAgent({
-      prompt,
-      systemPrompt: CODING_CHAT_SYSTEM,
-      // Everything it needs is in the prompt; dataDir is a harmless read-only
-      // sandbox. maxTurns 2 → answer directly, never a tool-loop.
-      addDir: dataDir,
-      allowedTools: ["Read"],
-      maxTurns: 2,
-      model: "sonnet",
-      onText: (delta) => send({ t: "delta", v: delta })
-    });
-    send({ t: "done", reply: stripSlop(run4.raw || "") || "(no response)" });
-  } catch {
-    send({ t: "error", message: "Couldn't reach your local Claude \u2014 is Claude Code (or Codex) installed and logged in?" });
-  } finally {
-    res.end();
-  }
-}
-
 // src/personApi.ts
 init_agent();
 import { promises as fs18 } from "fs";
@@ -22519,6 +22605,137 @@ ${dataMap}`,
   }
 }
 
+// ../../lib/agents/coding/chat-context.ts
+var RUNGS_BY_KEY = new Map(CODING_CRITERIA.map((c) => [c.key, c.rungs ?? []]));
+var clip = (s, n) => {
+  const t = (s || "").replace(/\s+/g, " ").trim();
+  return t.length > n ? t.slice(0, n - 1).replace(/\s+\S*$/, "") + "\u2026" : t;
+};
+function criterionBlock(c) {
+  if (!c.graded || c.dist.n <= 0) return null;
+  const lines = [];
+  const mean2 = c.dist.mean != null ? `${c.dist.mean}` : "\u2014";
+  lines.push(`### ${c.label}  (your mean ${mean2} across ${c.dist.n} graded session${c.dist.n === 1 ? "" : "s"})`);
+  lines.push(`What it measures: ${clip(c.definition, 400)}`);
+  const rungs = RUNGS_BY_KEY.get(c.key) ?? [];
+  if (rungs.length) {
+    lines.push(`The ladder I grade against:`);
+    for (const r of rungs) lines.push(`  L${r.level} \u2014 ${r.marker}`);
+  }
+  const takes = (c.takes || []).slice(0, 5);
+  if (takes.length) {
+    lines.push(`How I scored your sessions here:`);
+    for (const t of takes) {
+      const sc = t.score != null ? `${t.score}` : t.best != null ? `~${t.best}` : "\u2014";
+      const parts = [`  \u2022 "${clip(t.title, 80)}" (${(t.date || "").slice(0, 10)}) \u2014 scored ${sc}${t.confidence ? ` (${t.confidence})` : ""}.`];
+      if (t.instance) parts.push(`What showed it: ${clip(t.instance, 240)}`);
+      if (t.why) parts.push(`Why that score: ${clip(t.why, 320)}`);
+      const qs = (t.quotes || []).filter(Boolean).slice(0, 2).map((q) => `"${clip(q, 160)}"`);
+      if (qs.length) parts.push(`You said: ${qs.join("  /  ")}`);
+      lines.push(parts.join(" "));
+    }
+  }
+  return lines.join("\n");
+}
+var CODING_DATA_HOME_MAP = `THE DATA HOME YOU CAN READ (tools: Read, Grep, Glob; every path is relative to the extra read-only directory you were given):
+- coding/sessions.json \u2014 every session: title, times, activeMin, humanTurns, tool counts, klass (only "interactive" sessions are the person; the rest are machine runs), duplicateOf (fork copies \u2014 skip them or you double-count).
+- coding/focus.json \u2014 the flow analysis: per-day half-hour cells, steering runs, gaps, false starts, workstream features, causal findings. coding/flow.json \u2014 the older looser metric; the report calls it "engaged time", never flow. Do not conflate the two.
+- coding/day-digest.json and coding/day-chart.json \u2014 per-day narratives and per-conversation active intervals.
+- coding/projects.json \u2014 per-repo attribution. coding/delegation/ \u2014 delegation pattern reads.
+- coding/grades/*.json \u2014 per-session rubric grades (score on the 1-11 ladder, keyed instance, why, verbatim quotes). coding/walkthroughs/ \u2014 per-session walkthroughs.
+- coding/aggregate.json (ability ceilings + the bigPicture nutshell), coding/expertise.json, coding/gap.json (ranked technique recommendations), coding/coaching.json, coding/frontier.json, coding/frontier-detail.json, coding/concurrency.json \u2014 each self-describing; read the file before citing it.
+Trust the filesystem over this list: Glob/ls a directory before assuming its contents, and a file may simply not exist yet on a fresh run \u2014 say so instead of guessing. Budget 2-4 lookups per question, then answer.`;
+function buildCodingContext(profile) {
+  const out = [];
+  const t = profile.throughput;
+  const p = profile.parallelism;
+  const meta = [];
+  if (t?.sessions != null) meta.push(`${t.sessions} interactive coding sessions`);
+  if (profile.gradedSessions != null) meta.push(`${profile.gradedSessions} graded`);
+  if (p?.maxConcurrency != null) meta.push(`peak ${p.maxConcurrency} sessions in parallel`);
+  if (meta.length) out.push(`HOW THEY WORK: ${meta.join(" \xB7 ")}.`);
+  const blocks = (profile.criteria || []).map(criterionBlock).filter((b) => !!b);
+  if (blocks.length) {
+    out.push(`
+YOUR GRADED CRITERIA \u2014 the rubric ladder and how I scored you on each, with your own words as the receipts:`);
+    out.push(blocks.join("\n\n"));
+  } else {
+    out.push(`
+NO QUALITATIVE GRADES YET: the rubric grading pass has not scored any sessions, so there are no criterion scores, rungs, or graded receipts to cite. If they ask about grades or scores, say that plainly and never fabricate one. Everything measurable is still answerable: read the deterministic artifacts in the data home (sessions, focus/flow, projects, day digests) and give them real numbers with dates instead of shrugging.`);
+  }
+  out.push(`
+${CODING_DATA_HOME_MAP}`);
+  return out.join("\n");
+}
+var CODING_CHAT_SYSTEM = `You are the analyst who graded this person's coding-agent sessions (their Claude Code / Codex logs) against a FIXED rubric, and you are now talking WITH them about their report. Below you are given that rubric \u2014 the definition and the level ladder for each criterion \u2014 AND your own per-session grades: the exact moment that keyed each score, WHY you placed it on that rung, and their verbatim quotes. Everything you claim comes from that material or from the served data-home artifacts you can read with your tools (the grounding block maps them) \u2014 never invent a number, score, session, or quote that no file backs. The person's raw session logs are in bounds only through the artifacts present in the data home; do not hunt outside the directory you were given.
+
+HOW TO ANSWER:
+- Ground every claim in the receipts you were given. When they ask "why did I get that score on X", name the specific session, the rung you landed them on, what that rung requires, and quote their actual words. Never invent a session, a score, a rung, or a criterion that isn't below.
+- The rubric is the authority. If they push on a score, explain the ladder honestly: what rung they hit and what the NEXT rung would concretely have required of them. Don't inflate a score to be nice and don't apologize for a low one \u2014 being read accurately is the point. If they're genuinely strong, say so plainly.
+- Be honest about the instrument's limits. These grades come only from their coding logs, which UNDER-observe real skill (a lot of thinking never reaches the transcript). When it's relevant, say what you can and can't actually see from the logs rather than overclaiming.
+- Voice: warm, plain, declarative, second person. Specific, never a horoscope, never flattery. Numbers are calibration, not a verdict \u2014 don't recite them at the person.
+- Keep replies tight and human unless they ask you to go deep. Output ONLY your reply \u2014 no preamble, no meta.
+
+Do NOT use em dashes. Use a comma or a period instead.`;
+function stripSlop(raw) {
+  return (raw || "").trim().replace(/\s*—\s*/g, ", ").replace(/\s+–\s+/g, ", ").replace(/\s+,/g, ",");
+}
+function buildCodingChatPrompt(profile, messages) {
+  const context = buildCodingContext(profile);
+  const history = messages.map((m) => `${m.role === "user" ? "Them" : "You"}: ${m.content}`).join("\n\n");
+  return `YOUR READ OF THIS PERSON (the rubric + your grades + their receipts):
+${context}
+
+--- the conversation so far ---
+${history}
+
+Write your next reply to them now.`;
+}
+
+// src/codingChat.ts
+async function handleCodingChat(profile, dataDir, body, res) {
+  const raw = Array.isArray(body.messages) ? body.messages : [];
+  const messages = raw.filter((m) => (m.role === "user" || m.role === "assistant") && typeof m.content === "string").slice(-20);
+  if (!messages.length) {
+    res.writeHead(400, { "Content-Type": "application/json; charset=utf-8" });
+    res.end(JSON.stringify({ error: "no messages" }));
+    return;
+  }
+  const prompt = buildCodingChatPrompt(profile, messages);
+  res.writeHead(200, {
+    "Content-Type": "application/x-ndjson; charset=utf-8",
+    "Cache-Control": "no-store, no-transform",
+    "X-Accel-Buffering": "no"
+  });
+  const send = (o) => {
+    try {
+      res.write(JSON.stringify(o) + "\n");
+    } catch {
+    }
+  };
+  try {
+    const dataMap = await readDataMap();
+    const run4 = await invokeAgent({
+      prompt,
+      systemPrompt: `${CODING_CHAT_SYSTEM}
+
+AGENT MODE: you also have read-only tools (Read/Grep/Glob) over the person's local record. Use them whenever the pre-baked context can't answer precisely \u2014 the map below tells you what lives where. Traverse first, then answer with receipts. Never claim you lack access to something the map covers.
+
+${dataMap}`,
+      addDir: dataDir,
+      allowedTools: ["Read", "Grep", "Glob"],
+      maxTurns: 16,
+      model: "sonnet",
+      onText: (delta) => send({ t: "delta", v: delta })
+    });
+    send({ t: "done", reply: stripSlop(run4.raw || "") || "(no response)" });
+  } catch {
+    send({ t: "error", message: "Couldn't reach your local Claude \u2014 is Claude Code (or Codex) installed and logged in?" });
+  } finally {
+    res.end();
+  }
+}
+
 // src/publicReportApi.ts
 init_agent();
 
@@ -23555,6 +23772,18 @@ function distOf3(takes) {
   const hist = [...h.keys()].sort((a, b) => b - a).map((level) => ({ level, count: h.get(level) }));
   return { n, observed, mean: mean2, 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 readJson2(path30.join(CODING_DIR2, "sessions.json"), []);
   const dupIds = new Set(allRecords.filter((s) => s.duplicateOf).map((s) => s.sessionId));
@@ -23809,15 +24038,11 @@ 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, key2, evidenceGap) => {
+    const e = estimateRowParts(score, nTk(key2), 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);
@@ -23836,10 +24061,16 @@ 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 ${Math.round(parMin / 60)}h with 2+ live) \u2014 the Claude Code lead keeps 10\u201315 going at once, so this is the single biggest speed-up still on the table` : "you brief the AI sharply, but mostly drive one agent at a time \u2014 the top run 10\u201315 in parallel, a several-fold speed-up"
     ),
     // The evidence line cites the RANKING metric (flow as a share of the work
     // day), never absolute hours — a friend with MORE flow hours over a longer
@@ -23855,22 +24086,25 @@ async function compileCodingProfile() {
     // The four estimates LEAD with a concrete moment from their own sessions, then an
     // HONEST caveat — how many graded sessions back the read. No invented weakness, no
     // claim about what they "didn't" do (false specifics churn a real user); it's an estimate.
-    sRow(
+    est(
       "Agency",
       agency,
+      "outcomes",
       `${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
@@ -25112,6 +25346,7 @@ function stages(o) {
   const conc = String(o.conc ?? h.grade);
   const buildArgs = [];
   if (o.codex === false) buildArgs.push("--no-codex");
+  if (o.claudeCode === false) buildArgs.push("--no-claude");
   if (o.idleGapMin != null) buildArgs.push("--idle", String(o.idleGapMin));
   const gradeArgs = ["--all", "--model", gradeModel, "--conc", conc];
   if (o.regrade) gradeArgs.push("--regrade");
@@ -25341,6 +25576,7 @@ async function runPipeline(o) {
           `\u2713 report is up to date \u2014 ${gate2.pending} new session(s) since the last full run (below the ${threshold}-session threshold). Skipping the AI stages; free metrics still refresh. They'll be analyzed once ${threshold}+ accumulate (POLYMATH_MIN_NEW_SESSIONS=0 or --regrade to force now).`
         );
       } else if (gate2.hasPriorRun) {
+        o.onNewWork?.(gate2.pending);
         o.onLine?.(
           `\u2713 your finished report from the last run is safe and already serving at the link below \u2014 nothing restarts from scratch. ${gate2.pending >= 0 ? `${gate2.pending} pending session(s)` : "New work"} plus any stages added by an update run on top of it, and sections refresh in place as they finish.`
         );
@@ -25542,6 +25778,8 @@ async function runChatPipeline(o) {
         upToDate = true;
         o.onUpToDate?.();
         o.onLine?.(`\u2713 chat report is up to date \u2014 ${newDocs} new doc(s) since the last full run (below the ${minNewDocs}-doc threshold), so the AI stages are skipped and the existing report is served as-is. POLYMATH_MIN_NEW_DOCS=0 forces a full run.`);
+      } else {
+        o.onNewWork?.(newDocs);
       }
     }
   }
@@ -25781,6 +26019,11 @@ Graded ${detail.gradedSessions} sessions \u2192 ${gradeOut}/graded.json`);
   }
   if (opts.cmd === "serve") {
     let earlyHandle = null;
+    const consentManifest = await readManifest(DATA_ROOT);
+    const consent = agentFlagsFromManifest(consentManifest);
+    const consentWhy = (id) => consentManifest?.agents?.excluded.includes(id) ? "you excluded it during setup" : "setup never asked about it";
+    if (!consent.claudeCode) console.error(`  \u26D4 Claude Code logs will NOT be analyzed \u2014 ${consentWhy("claude-code")}. POLYMATH_WIZARD=1 polymath-society to change.`);
+    if (!consent.codex) console.error(`  \u26D4 Codex logs will NOT be analyzed \u2014 ${consentWhy("codex")}. POLYMATH_WIZARD=1 polymath-society to change.`);
     const runMarker = path41.join(DATA_ROOT, "coding", ".grade-run.json");
     if (!opts.grade) {
       try {
@@ -25858,12 +26101,19 @@ Running the full analysis on ${backend.name === "cli" ? "Claude Code" : "Codex"}
       mbar.setLink(earlyHandle.url);
       const codingLane = mbar.lane("coding");
       let codingPinned = false;
+      if (await readPriorAnalysis()) {
+        codingPinned = true;
+        progress.coding = { i: 2, total: 1, label: "up to date \u2014 checking for anything new", done: false };
+        earlyHandle.setProgress(progress);
+        codingLane.finish("up to date \u2014 checking for anything new");
+      }
       const codingDone = runPipeline({
         repoRoot: REPO_ROOT,
         model: opts.model,
         conc: opts.conc,
         overnight: opts.overnight,
-        codex: opts.codex,
+        codex: opts.codex && consent.codex,
+        claudeCode: consent.claudeCode,
         idleGapMin: opts.idleGapMin,
         regrade: opts.regrade,
         onUpToDate: () => {
@@ -25872,14 +26122,22 @@ Running the full analysis on ${backend.name === "cli" ? "Claude Code" : "Codex"}
           earlyHandle.setProgress(progress);
           codingLane.finish("everything up to date");
         },
+        onNewWork: (pending2) => {
+          codingPinned = false;
+          codingLane.log(`  ${pending2} new session(s) to analyze \u2014 updating your report in place.`);
+        },
         onStage: (i, total, label, llm) => {
           if (codingPinned) return;
           progress.coding = { ...progress.coding, i, total, label, done: false };
           earlyHandle.setProgress(progress);
           codingLane.stage(i, total, `${label}${llm ? "  (AI)" : ""}`);
         },
+        // While pinned, the scrolling stage output stays quiet ("prints hella
+        // things when everything is cached", owner 2026-07-17) — warnings and
+        // failures always pass through.
         onLine: (line) => {
           harvestWithin("coding", line);
+          if (codingPinned && !/⚠|error|fail/i.test(line)) return;
           codingLane.log(`    ${line}`);
         }
       }).then(async (codingRes2) => {
@@ -25912,6 +26170,16 @@ Running the full analysis on ${backend.name === "cli" ? "Claude Code" : "Codex"}
         }
         chatLane.log(`  chat analysis starting over your ${manifest.exports.included.length} approved source(s) \u2014 same engine as the hosted app.`);
         let chatPinned = false;
+        const chatComplete = await fs36.readFile(path41.join(DATA_ROOT, "engine-pilot", "public-report.json"), "utf8").then((s) => {
+          const r = JSON.parse(s);
+          return !!(r.headline && (r.spikes?.length ?? 0) > 0);
+        }).catch(() => false);
+        if (chatComplete) {
+          chatPinned = true;
+          progress.chat = { i: 2, total: 1, label: "up to date \u2014 checking for anything new", done: false };
+          earlyHandle.setProgress(progress);
+          chatLane.finish("up to date \u2014 checking for anything new");
+        }
         const chat2 = await runChatPipeline({
           repoRoot: REPO_ROOT,
           model: opts.model,
@@ -25922,6 +26190,10 @@ Running the full analysis on ${backend.name === "cli" ? "Claude Code" : "Codex"}
             earlyHandle.setProgress(progress);
             chatLane.finish("everything up to date");
           },
+          onNewWork: (pending2) => {
+            chatPinned = false;
+            chatLane.log(`  ${pending2} new doc(s) to analyze \u2014 updating your chat report in place.`);
+          },
           onStage: (i, total, label, llm) => {
             if (chatPinned) return;
             progress.chat = { ...progress.chat, i, total, label, done: false };
@@ -25930,6 +26202,7 @@ Running the full analysis on ${backend.name === "cli" ? "Claude Code" : "Codex"}
           },
           onLine: (line) => {
             harvestWithin("chat", line);
+            if (chatPinned && !/⚠|error|fail/i.test(line)) return;
             chatLane.log(`    ${line}`);
           }
         });
@@ -25985,7 +26258,8 @@ No analysis here yet \u2014 computing the instant deterministic metrics now (fre
       const det = await runPipeline({
         repoRoot: REPO_ROOT,
         deterministicOnly: true,
-        codex: opts.codex,
+        codex: opts.codex && consent.codex,
+        claudeCode: consent.claudeCode,
         idleGapMin: opts.idleGapMin,
         onStage: (i, total, label) => dbar.stage(i, total, label),
         onLine: (line) => dbar.log(`    ${line}`)