polymath-society 0.2.22 → 0.2.23

package/dist/cli.js

@@ -22857,6 +22857,7 @@ function apportionHeavy(budget) {
 function stages(o) {
   const model = o.model ?? "sonnet";
   const gradeModel = o.model ?? "haiku";
+  const narrativeModel = o.model ?? "haiku";
   const budget = machineBudget();
   const h = apportionHeavy(budget);
   const conc = String(o.conc ?? h.grade);
@@ -22882,8 +22883,8 @@ function stages(o) {
     // finish inside its shadow). Budget: grade 8 + digest 3 + walkthrough 3 +
     // two single-call stages ≈ 16 peak subprocesses (measured-safe envelope).
     { script: "coding-grade.mts", label: "AI-grade every substantial session", llm: true, args: gradeArgs, batch: "heavy" },
-    { script: "coding-day-digest.mts", label: "Day-by-day digests", llm: true, args: ["--model", model, "--conc", String(h.digest)], out: "coding/day-digest.json", batch: "heavy" },
-    { script: "coding-walkthrough.mts", label: "Per-session walkthroughs", llm: true, args: ["--model", model, "--conc", String(h.walkthrough)], batch: "heavy" },
+    { script: "coding-day-digest.mts", label: "Day-by-day digests", llm: true, args: ["--model", narrativeModel, "--conc", String(h.digest)], out: "coding/day-digest.json", batch: "heavy" },
+    { script: "coding-walkthrough.mts", label: "Per-session walkthroughs", llm: true, args: ["--model", narrativeModel, "--conc", String(h.walkthrough)], batch: "heavy" },
     { script: "coding-frontier-detail.mts", label: "What you're doing at the frontier", llm: true, args: [], out: "coding/frontier-detail.json", batch: "heavy" },
     { script: "coding-delegation.mts", label: "Delegation patterns", llm: true, args: ["--model", model, "--conc", String(h.delegation)], out: "coding/delegation-rollup.json", batch: "heavy" },
     // The FULL focus stage: flow-rule judgment + the paste-into-CLAUDE.md block.
@@ -22901,7 +22902,13 @@ function stages(o) {
     // (one call per run; the axis verdicts are the most judgment-heavy output).
     // An explicit user model choice still wins.
     { script: "coding-agglomerate.mts", label: "Agglomerated criteria + verdicts + signature moves", llm: true, args: o.model ? ["--model", o.model] : [], out: "coding/agglomerate.json", batch: "grade-readers" },
-    { script: "coding-expertise.mts", label: "Domain-expertise map", llm: true, args: ["--conc", String(Math.max(2, budget - 1))], out: "coding/expertise.json", batch: "grade-readers" },
+    // Domain-expertise map DISABLED (owner call 2026-07-22): the single most
+    // expensive stage of the whole run ($13.88 of $49.34 measured — 129 sonnet
+    // calls) and for most users the hiring-bar filter correctly returns ZERO
+    // domains — $14 for an empty section. The machinery (coding-expertise.mts,
+    // expertise-floor.ts) stays for the planned cheap re-enable (haiku scan or
+    // fold into the grade call); the report renders nothing when the artifact
+    // is absent.
     // ---- final synthesis (reads the finished aggregates + batch outputs) ----
     { script: "coding-coaching.mts", label: "Coaching + feel-seen copy", llm: true, args: [], out: "coding/coaching.json", batch: "synthesis" },
     { script: "coding-gap.mts", label: "Gap-to-close vs best practice", llm: true, args: [], out: "coding/gap.json", batch: "synthesis" },
@@ -24039,7 +24046,7 @@ async function detectPlan(home = os9.homedir()) {
   return { tier, label: LABEL[tier], windowUsd: WINDOW_USD[tier] };
 }
 function estimateLaneCosts(input) {
-  const codingUsd = Math.max(5, Math.min(60, Math.round(input.agentLogFiles * 17e-4)));
+  const codingUsd = Math.max(5, Math.min(32, Math.round(input.agentLogFiles * 9e-4)));
   const chatUsd = input.chatIncluded ? Math.round(40 + input.exportTotalMB * 0.8) : 0;
   return { codingUsd, chatUsd };
 }
@@ -24074,19 +24081,16 @@ async function remainingFractions(dataRoot, agentLogFiles) {
   return { coding, chat: clamp(1 - chatDone, 0.1, 1) };
 }
 function scheduleLanes(plan, est) {
-  const dayShare = plan.tier === "max_5x" || plan.tier === "gpt_team" ? 0.2 : 0.5;
-  const fits = (usd) => usd <= plan.windowUsd * dayShare;
-  const coding = fits(est.codingUsd) ? "now" : "overnight";
-  const chat = est.chatUsd === 0 || fits(est.chatUsd) ? "now" : "overnight";
+  const chat = est.chatUsd === 0 || est.chatUsd <= plan.windowUsd * 0.9 ? "now" : "overnight";
   return {
-    coding,
+    coding: "now",
     chat,
     // No plan name here — the schedule header names the plan once, and
     // repeating it doubled the clumsy unknown-plan label (2026-07-22
     // walkthrough). Phrased for the two-column "Proposed schedule" rows:
     // "<lane>   <verdict> — <reason>".
-    codingReason: coding === "now" ? `runs now \u2014 fits comfortably inside one 5-hour window` : `starts at 2:00 am tonight \u2014 a full grade would eat most of your daytime window`,
-    chatReason: est.chatUsd === 0 ? "not planned \u2014 no chat sources included" : chat === "now" ? `runs now, alongside coding \u2014 your chat corpus is small enough` : `starts at 2:00 am tonight \u2014 a large chat corpus; this protects your daytime quota`
+    codingReason: `runs now \u2014 the report link appears right away and fills in as it goes`,
+    chatReason: est.chatUsd === 0 ? "not planned \u2014 no chat sources included" : chat === "overnight" ? `starts at 2:00 am tonight \u2014 a chat corpus this size is an overnight job; it runs while you sleep` : `runs now, alongside coding`
   };
 }
 var UNSUPPORTED_FULL_TIERS, WINDOW_USD, LABEL;

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

@@ -16193,7 +16193,7 @@ ${lines.join("\n")}`);
   }
   return blocks.join("\n\n");
 }
-var SYSTEM = "You read all the messages a person sent on ONE day across several coding-agent conversations, and say \u2014 per conversation \u2014 what they actually DID that day, semantically and specifically. Their words only; ignore mechanical scaffolding. One tight line per conversation (e.g. 'built the full prototype of the apprenticeship app', 'tuned the intelligence rubric with examples from their own read on people', 'set up an EC2 box to run the analysis overnight'). No flattery, no filler.";
+var SYSTEM = "You read all the messages a person sent on ONE day across several coding-agent conversations, and say \u2014 per conversation \u2014 what they actually DID that day, semantically and specifically. Their words only; ignore mechanical scaffolding. One tight line per conversation. HARD RULES: (1) NAME the concrete feature, bug, decision, or artifact in play ('added retry handling to the export queue', 'cut the sidebar's self-analysis tab but kept the public report') \u2014 never a category ('reviewed features', 'UI changes'). (2) A line must be so specific it could only describe THAT conversation \u2014 if it would fit another day's work, rewrite it. (3) BANNED verbs: 'worked on', 'iterated on', 'refined', 'reviewed', 'continued', 'various'. (4) Even a 1-2 message conversation gets a real read of what the person said or asked \u2014 never a meta-comment like 'repeated earlier context'. No flattery, no filler.";
 async function digestDay(dc, opts = {}) {
   const sandbox = await fs6.mkdtemp(path10.join(os3.tmpdir(), "ps-coding-day-"));
   try {
@@ -16205,7 +16205,7 @@ Output a SINGLE fenced \`\`\`json block:
 \`\`\`json
 { "overall": "", "conversations": [ { "sessionId": "<8-char id from a header>", "summary": "" } ] }
 \`\`\``;
-    const run2 = await invokeAgent({ prompt, systemPrompt: SYSTEM, addDir: sandbox, allowedTools: ["Read"], maxTurns: 4, timeoutMs: 3e5, model: opts.model ?? "sonnet" });
+    const run2 = await invokeAgent({ prompt, systemPrompt: SYSTEM, addDir: sandbox, allowedTools: ["Read"], maxTurns: 4, timeoutMs: 3e5, model: opts.model ?? "haiku" });
     const raw = run2.json ?? {};
     const summaries = new Map((raw.conversations ?? []).map((c) => [String(c.sessionId ?? "").slice(0, 8), String(c.summary ?? "")]));
     return {
@@ -16570,7 +16570,7 @@ var GRADES = path11.join(CODING_DIR, "grades");
 
 // ../../scripts/coding-day-digest.mts
 var argv = process.argv.slice(2);
-var model = argv.indexOf("--model") >= 0 ? argv[argv.indexOf("--model") + 1] : "sonnet";
+var model = argv.indexOf("--model") >= 0 ? argv[argv.indexOf("--model") + 1] : "haiku";
 var REDO = argv.includes("--redo");
 var CONC = Number(argv.indexOf("--conc") >= 0 ? argv[argv.indexOf("--conc") + 1] : "8");
 var CODING = CODING_DIR;

package/dist/pipeline/coding-walkthrough.js

@@ -16237,7 +16237,7 @@ function buildTranscript(msgs, tz) {
   if (text2.length > MAXC) text2 = text2.slice(0, MAXC) + "\n\u2026[transcript truncated]";
   return { text: text2, turnT, userTurns: n };
 }
-var SYSTEM = "You reconstruct, IN ORDER, what a person actually tried in one coding-agent session \u2014 the arc of attempts, realizations, dead-ends, and course-corrections, told as their story. You read the back-and-forth (YOU: = the person, AI: = the model, collapsed). You write specific, concrete phases grounded in what really happened \u2014 never generic ('iterated on the feature'). Capture the THINKING: what they tried, why it didn't satisfy them, what they realized, what they switched to. Honest about messiness and dead-ends.";
+var SYSTEM = "You reconstruct, IN ORDER, what a person actually tried in one coding-agent session \u2014 the arc of attempts, realizations, dead-ends, and course-corrections, told as their story. You read the back-and-forth (YOU: = the person, AI: = the model, collapsed). Every phase must be so specific that it could ONLY describe this session \u2014 a reader who saw ten sessions could pick this one out from the phase alone. Capture the THINKING: what they tried, why it didn't satisfy them, what they realized, what they switched to. Honest about messiness and dead-ends.";
 function buildPrompt(rec, transcript) {
   return `Below is a coding session, with the person's turns numbered [#N HH:MM].
 
@@ -16245,11 +16245,18 @@ SESSION: "${rec.title}"
 
 ${transcript}
 
-Reconstruct what happened as an ORDERED list of 5\u201314 phases \u2014 the person's actual arc. Each phase:
-- "what": 1\u20132 sentences, concrete and specific, in narrative voice (e.g. "Started by trying to show percentiles instead of raw scores, but nothing anchored what a percentile meant, so pivoted to comparing against big-tech engineers." / "Realized the output was AI slop \u2014 they'd only given a couple of examples, not a real spec, and asked the AI to generate more from that.").
-- "startTurn": the turn number [#N] where this phase begins.
+Reconstruct what happened as an ORDERED list of phases \u2014 the person's actual arc. ONE phase per distinct move the person made: a new demand, a pushback, a scope change, a realization, a decision, a dead-end. If a turn contains two separate moves, that is two phases. Short sessions may have 5; long many-turn sessions often need 12\u201320. Never merge distinct moves to stay short.
 
-Order phases by startTurn (ascending). Capture realizations and dead-ends, not just actions.
+Each phase:
+- "what": 1\u20132 sentences. VOICE: narrative, past tense, the person as the implied subject \u2014 start with the verb of their move ("Opened by demanding\u2026", "Pushed back when\u2026", "Caught a real gap\u2026", "Dropped the integration after\u2026"). NEVER start with "User" or "The user".
+- HARD RULES for "what":
+  1. NAME the concrete thing \u2014 the specific feature, file, bug, number, or behavior in play ("the login redirect that looped back to /home", "the CSV export that dropped its header row"), never a category ("an issue", "the feature", "some bugs").
+  2. CARRY the person's own key words when they reveal intent or feeling \u2014 short quoted fragments ('ship it as is', 'why is this still broken') are the highest-signal content.
+  3. STATE the why or the turn-of-thought when the transcript shows it: what dissatisfied them, what they realized, what they traded away. An action with no why is only half a phase.
+  4. BANNED: 'worked on', 'iterated on', 'continued', 'discussed', 'made progress', 'addressed issues', 'various fixes' \u2014 if a sentence would fit any other session, rewrite it until it wouldn't.
+- "startTurn": the turn number [#N] where this phase begins. Use the number printed in the transcript \u2014 never invent one.
+
+Order phases by startTurn (ascending). Cover the WHOLE session \u2014 later turns matter as much as early ones. Capture realizations and dead-ends, not just actions.
 
 Output a SINGLE fenced \`\`\`json block:
 \`\`\`json
@@ -16287,7 +16294,9 @@ async function buildWalkthrough(rec, opts) {
       allowedTools: ["Read"],
       maxTurns: 6,
       timeoutMs: 3e5,
-      model: opts.model ?? "sonnet"
+      // haiku default (2026-07-22): with the explicit-rules prompt above it
+      // matches sonnet on real sessions at ~1/3 the cost; opts.model overrides.
+      model: opts.model ?? "haiku"
     });
     const phases = validate(run2.json, turnT, times, sessionEnd);
     if (phases.length === 0) return null;
@@ -16679,7 +16688,7 @@ var flag = (n, d) => {
   const i = argv.indexOf(`--${n}`);
   return i >= 0 ? argv[i + 1] : d;
 };
-var MODEL = flag("model", "sonnet");
+var MODEL = flag("model", "haiku");
 var CONC = Number(flag("conc", "8"));
 var REGEN = argv.includes("--regen");
 var CODING = CODING_DIR;

package/dist/web/app.js

@@ -9886,58 +9886,6 @@ function CodingAccountabilityCard({ a }) {
     ] })
   ] });
 }
-function Pill({ label, tone = "faint" }) {
-  const c = { good: "bg-accentSoft text-accent", accent: "bg-accentSoft text-accent", faint: "border border-line bg-canvas text-faint" }[tone];
-  return /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { className: `rounded px-1.5 py-0.5 text-[11px] font-semibold ${c}`, children: label });
-}
-function ExpertiseMapPanel({ x, coaching }) {
-  const tone = (d) => d === "deep" ? "good" : d === "working" ? "accent" : "faint";
-  const domainLi = (dm, i) => /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("li", { children: /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("details", { className: "group", children: [
-    /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("summary", { className: "grid cursor-pointer list-none grid-cols-1 items-baseline gap-x-4 gap-y-0.5 text-[13px] [&::-webkit-details-marker]:hidden sm:grid-cols-[1.05fr_1fr]", children: [
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("span", { className: "flex min-w-0 items-baseline gap-2", children: [
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { className: "text-accent", children: "\u25B8" }),
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(Pill, { label: dm.depth, tone: tone(dm.depth) }),
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { className: "font-semibold text-ink", children: dm.domain })
-      ] }),
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("span", { className: "text-[12px] leading-relaxed text-faint", children: [
-        dm.dateRange && /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { className: "font-medium", children: dm.dateRange }),
-        dm.whatTheyKnow && /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("span", { className: "text-muted", children: [
-          " \u2014 ",
-          dm.whatTheyKnow
-        ] })
-      ] })
-    ] }),
-    dm.why && /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("div", { className: "ml-5 mt-2 text-[12.5px] leading-relaxed text-ink", children: [
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { className: "font-semibold text-accent", children: "Why \u2014 " }),
-      dm.why
-    ] }),
-    dm.bullets?.length ? /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("ul", { className: "ml-5 mt-2 space-y-1", children: dm.bullets.map((b, j) => /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("li", { className: "flex gap-2 text-[12.5px] leading-relaxed text-ink", children: [
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { className: "mt-[7px] h-1 w-1 flex-none rounded-full bg-accentSoft" }),
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("span", { children: b })
-    ] }, j)) }) : dm.evidence?.slice(0, 3).map((e, j) => /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("div", { className: "ml-5 mt-2 border-l-2 border-line pl-2.5", children: [
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)("div", { className: "text-[11px] uppercase tracking-wide text-faint", children: [
-        e.kind,
-        " \xB7 ",
-        e.date
-      ] }),
-      /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("div", { className: "text-[12.5px] text-muted", children: e.knowledge })
-    ] }, j))
-  ] }) }, i);
-  const tech = x.domains.filter((d) => d.technical !== false);
-  const built = tech.filter((d) => d.kind !== "directs");
-  const directs = tech.filter((d) => d.kind === "directs");
-  return /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)(Card, { className: "mb-5", children: [
-    /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("div", { className: "text-[12px] text-faint", children: "hard technical / systems domains where you have real, accumulated depth \u2014 not your eval/product work, not smart-generalist problem-solving \xB7 expand any line for the proof" }),
-    tech.length ? /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)(import_jsx_runtime8.Fragment, { children: [
-      built.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(SubSection, { label: "Built & studied deeply", children: /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("ul", { className: "space-y-1.5", children: built.map(domainLi) }) }),
-      directs.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)(SubSection, { label: "Directs the AI with authority", children: [
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("div", { className: "mb-2 text-[11.5px] leading-relaxed text-faint", children: "you haven\u2019t necessarily built these here, but you command the AI in them with mechanism-level authority it defers to" }),
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("ul", { className: "space-y-1.5", children: directs.map(domainLi) })
-      ] })
-    ] }) : /* @__PURE__ */ (0, import_jsx_runtime8.jsx)("p", { className: "mt-2 text-[13px] leading-relaxed text-muted", children: "No hard technical / systems domain cleared the bar \u2014 you\u2019re a very capable generalist builder, but deep specialized domains (vector DBs, distributed systems, kernels, RAG internals\u2026) aren\u2019t where your accumulated expertise lives. That\u2019s the honest picture." }),
-    tech.length > 0 && /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(Coaching, { text: coaching })
-  ] });
-}
 var DO_THIS = {
   worktrees: "Replace `claude` with `claude --worktree <name>` \u2014 each session gets its own isolated checkout. That's the whole change; there's no branching workflow to adopt.",
   subagents: "Run `/agents` in Claude Code to save a reusable specialist (e.g. a per-criterion grader with your rubric baked in) to `.claude/agents/` \u2014 or just tell Claude \u201Cuse 4 subagents, one per criterion\u201D to fan out on the spot.",
@@ -11022,10 +10970,6 @@ function CodeAnalysisView({ data: dataProp } = {}) {
         data.frontierArsenal && /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(FrontierPanel, { f: data.frontierArsenal, detail: data.frontierDetail, coaching: data.coaching?.coaching?.frontier, gap: data.gap, parallelism: data.parallelism }),
         data.delegationRollup && /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(DelegationPanel, { r: data.delegationRollup, takes: data.criteria?.find((c) => c.key === "delegation")?.takes })
       ] }), ["delegation", "frontier"]),
-      data.expertiseMap && fs("coding:expertise", "Expertise", /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)(import_jsx_runtime8.Fragment, { children: [
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(SectionDivider, { label: "Expertise" }),
-        /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(ExpertiseMapPanel, { x: data.expertiseMap, coaching: data.coaching?.coaching?.expertise })
-      ] }), ["expertise"]),
       (data.gradedSessions ?? 0) > 0 && (data.agglomeration?.verdicts || data.aggregate?.ability) && fs("coding:qualitatives", "The qualitatives", /* @__PURE__ */ (0, import_jsx_runtime8.jsxs)(import_jsx_runtime8.Fragment, { children: [
         /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(SectionDivider, { label: "The qualitatives", note: "Our best read, with the specific chats behind it. Coding logs are a narrow window, so thin evidence reads as a range: your real level can sit above or below what shows up here." }),
         /* @__PURE__ */ (0, import_jsx_runtime8.jsx)(AbilityPanel, { a: data.aggregate?.ability, verdicts: data.agglomeration?.verdicts, criteria: data.criteria, feelSeen: data.coaching?.feelSeen, agencyScore: (data.stackUp ?? []).find((r) => r.label === "Agency")?.score ?? null })

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "polymath-society",
-  "version": "0.2.22",
+  "version": "0.2.23",
   "description": "Deterministic metrics from your local Claude Code / Codex agent logs — parallelism, flow, throughput, projects. Zero LLM, zero server, on-device.",
   "license": "UNLICENSED",
   "type": "module",