atris 3.74.23 → 3.74.25

package/lib/atris-rating.js

@@ -0,0 +1,179 @@
+'use strict';
+
+// The Atris Rating: one number, 25 to 99, for how much real work a teammate
+// finishes. 99 means 99 of every 100. Each time a teammate halves its misses
+// it gains about 6 points, so the top is the hardest climb: going from 96 to
+// 99 percent finished is worth about 12.
+//
+// One scale rates two things. An engine gets its number from the combine
+// below: real tasks graded by the tests that shipped with them. A member
+// gets its number from the jobs it lands in its own workspace (gm-season).
+
+const { modelLabel } = require('./roster-models');
+
+const TOP = 99;
+const FLOOR = 25;
+const STEP = 6;
+
+// Every rating starts from half a finish and half a miss, so a short perfect
+// run cannot reach the top: 3 of 3 rates 76, and it takes about 60 in a row
+// to rate 99. null when nothing has been tried.
+function atrisRating(finished, tries) {
+  const all = Number(tries);
+  const done = Number(finished);
+  if (!Number.isFinite(all) || all <= 0 || !Number.isFinite(done)) return null;
+  const share = (Math.min(Math.max(done, 0), all) + 0.5) / (all + 1);
+  const raw = TOP - STEP * (Math.log2(TOP) - Math.log2(share / (1 - share)));
+  return Math.max(FLOOR, Math.min(TOP, Math.round(raw)));
+}
+
+// The combine: Loop Bench, tasks from the atris repos (javascript from the
+// CLI, python from the backend), each graded by the hidden tests that shipped
+// with its fix, run in Claude Code on the builder prompt Atris kept. unseen
+// tasks were held out of every prompt change, so the prompt was never tuned
+// on them. finished adds each task's pass rate over its tries; time and price
+// count each task once too. broke counts tries that passed their own tests
+// but broke another one. The file is what evals/loopbench/combine.py in
+// atrisos-backend prints, copied as is, so re-rating is one copy.
+function freeze(value) {
+  if (value && typeof value === 'object') {
+    for (const inner of Object.values(value)) freeze(inner);
+    Object.freeze(value);
+  }
+  return value;
+}
+
+const COMBINE = freeze(require('./atris-rating-combine.json'));
+
+const MONTHS = ['January', 'February', 'March', 'April', 'May', 'June', 'July',
+  'August', 'September', 'October', 'November', 'December'];
+
+// "2026-10-29" -> "October 2026"
+function monthOf(date) {
+  const [year, month] = String(date || '').split('-').map(Number);
+  return year && month >= 1 && month <= 12 ? `${MONTHS[month - 1]} ${year}` : '';
+}
+
+const RATED = monthOf(COMBINE.rated);
+
+// One engine's card: the overall and four parts, each on the same scale.
+function engineCard(row) {
+  return {
+    model: row.model,
+    label: modelLabel(row.model),
+    ovr: atrisRating(row.finished, row.tasks),
+    finished: row.finished,
+    tasks: row.tasks,
+    python: atrisRating(row.python.finished, row.python.tasks),
+    javascript: atrisRating(row.javascript.finished, row.javascript.tasks),
+    unseen: atrisRating(row.unseen.finished, row.unseen.tasks),
+    clean: atrisRating(row.tries - row.broke, row.tries),
+    parts: { python: row.python, javascript: row.javascript, unseen: row.unseen },
+    tries: row.tries,
+    broke: row.broke,
+    minutes_cap: row.minutes_cap,
+    median_minutes: row.median_minutes,
+    cost_per_finish: row.cost_per_finish,
+  };
+}
+
+// Every rated engine, best first.
+function combineCards() {
+  return COMBINE.engines.map(engineCard).sort((a, b) => b.ovr - a.ovr || a.label.localeCompare(b.label));
+}
+
+// The card for a model id, a label, or a family word that names one engine;
+// null for a model the combine has not rated.
+function combineCardFor(name) {
+  const text = String(name || '').trim().toLowerCase().replace(/\s+/g, ' ');
+  if (!text) return null;
+  const cards = combineCards();
+  const exact = cards.find((card) => card.model === text || card.label === text);
+  if (exact) return exact;
+  const family = cards.filter((card) => card.label.split(' ')[0] === text);
+  return family.length === 1 ? family[0] : null;
+}
+
+function finishedText(value) {
+  return Number.isInteger(value) ? String(value) : String(Math.round(value * 10) / 10);
+}
+
+function priceText(dollars) {
+  if (typeof dollars !== 'number') return '-';
+  return dollars < 1 ? `${Math.round(dollars * 100)}¢` : `$${dollars.toFixed(2)}`;
+}
+
+function minutesText(minutes) {
+  return typeof minutes === 'number' ? `${minutes.toFixed(1)} min` : '-';
+}
+
+const SCALE_LINES = Object.freeze([
+  'how much real work a teammate finishes, 25 to 99. 99 means 99 of every 100.',
+  'each halving of misses is worth about 6 points, so the top is the hardest climb.',
+]);
+
+// The league table: every rated engine, best first.
+function renderCombine(cards = combineCards()) {
+  const caps = [...new Set(cards.map((card) => card.minutes_cap))];
+  const capLine = caps.length > 1
+    ? cards.filter((card) => card.minutes_cap !== Math.min(...caps)).map((card) => `${card.label} gets ${card.minutes_cap} minutes a task`).join(', ') + `, the others ${Math.min(...caps)}.`
+    : `every engine gets ${caps[0]} minutes a task.`;
+  const lines = [
+    'atris rating',
+    ...SCALE_LINES,
+    '',
+    `the combine: ${COMBINE.tasks} atris tasks, each graded by the tests that shipped with it`,
+    '',
+    `  ${'ovr'.padEnd(5)}${'engine'.padEnd(12)}${'finished'.padEnd(12)}${'py'.padEnd(4)}${'js'.padEnd(4)}${'unseen'.padEnd(8)}${'clean'.padEnd(7)}${'time'.padEnd(9)}price`,
+  ];
+  for (const card of cards) {
+    lines.push(`  ${String(card.ovr).padEnd(5)}${card.label.padEnd(12)}${`${finishedText(card.finished)} of ${card.tasks}`.padEnd(12)}${String(card.python).padEnd(4)}${String(card.javascript).padEnd(4)}${String(card.unseen).padEnd(8)}${String(card.clean).padEnd(7)}${minutesText(card.median_minutes).padEnd(9)}${priceText(card.cost_per_finish)}`);
+  }
+  const top = cards[0];
+  lines.push(
+    '',
+    top && top.ovr >= TOP ? `${top.label} is a 99.` : 'no engine is a 99 yet.',
+    'py, js: python and javascript tasks. unseen: ones the prompt was never tuned on.',
+    'clean: tries that broke no other test. time: median to finish.',
+    'price: per finished task, pay-per-use.',
+    capLine,
+    `${COMBINE.bench}, ${RATED}. one engine: atris rating <engine>`,
+  );
+  return lines.join('\n');
+}
+
+// One engine's card.
+function renderCard(card) {
+  const head = `${card.label}, builder`;
+  const part = (name, value, text) => `  ${name.padEnd(13)}${String(value).padEnd(5)}${text}`;
+  return [
+    `${head}${' '.repeat(Math.max(2, 44 - head.length))}${card.ovr} ovr`,
+    '',
+    part('python', card.python, `${finishedText(card.parts.python.finished)} of ${card.parts.python.tasks} tasks`),
+    part('javascript', card.javascript, `${finishedText(card.parts.javascript.finished)} of ${card.parts.javascript.tasks} tasks`),
+    part('unseen', card.unseen, `${finishedText(card.parts.unseen.finished)} of ${card.parts.unseen.tasks} tasks the Atris prompt was never tuned on`),
+    part('clean', card.clean, `broke another test on ${card.broke} of ${card.tries} tries`),
+    '',
+    `  finished ${finishedText(card.finished)} of ${card.tasks} tasks, ${minutesText(card.median_minutes)} median, ${card.minutes_cap} minute cap`,
+    `  ${priceText(card.cost_per_finish)} per finished task on pay-per-use`,
+    `  ${COMBINE.bench}, ${RATED}, ${card.tries} tries graded by each task's shipped tests`,
+  ].join('\n');
+}
+
+// What the boot team line adds after a build engine, like " (86 ovr)", or ''
+// for a model the combine has not rated.
+function bootRatingText(model) {
+  const card = model ? combineCards().find((c) => c.model === String(model).trim()) : null;
+  return card ? ` (${card.ovr} ovr)` : '';
+}
+
+module.exports = {
+  COMBINE,
+  SCALE_LINES,
+  atrisRating,
+  bootRatingText,
+  combineCardFor,
+  combineCards,
+  renderCard,
+  renderCombine,
+};

package/lib/gm-bet.js

@@ -4573,8 +4573,8 @@ function createBetDesk(deps) {
 
   // The team page: every member on the team, one row each. role is a few
   // words from its MEMBER.md. ovr is the board's own rating (atris gm
-  // --board): 60 to 99 from the share of finished jobs that landed in the
-  // last 30 days, "new" with 1 or 2 finished, and two dots with none
+  // --board): the Atris Rating, 25 to 99, of the finished jobs that landed
+  // in the last 30 days, "new" with 1 or 2 finished, and two dots with none
   // finished in that window (a first claim alone is not a rating) or no
   // task list to read. engine is the tool and the exact model the roster
   // gives the member (atris engine roster), two dots when no model is

package/lib/gm-season.js

@@ -10,6 +10,7 @@
 
 const fs = require('fs');
 const path = require('path');
+const { atrisRating } = require('./atris-rating');
 
 const HOUR_MS = 60 * 60 * 1000;
 const DAY_MS = 24 * HOUR_MS;
@@ -49,17 +50,18 @@ function tally(outcomes) {
   return out;
 }
 
-// round(60 + 39 * landed / finished), rookie under 3 finished jobs. This is
+// The Atris Rating of landed over finished (lib/atris-rating.js): 25 to 99,
+// where 99 means 99 of every 100, rookie under 3 finished jobs. This is
 // stricter than the website rule (kept / graded) on purpose: drops and
 // abandoned jobs count.
 function ovr(outcomes) {
   const { landed, finished } = tally(outcomes);
   if (finished < ROOKIE_FINISHED) return null;
-  return Math.round(60 + 39 * landed / finished);
+  return atrisRating(landed, finished);
 }
 
 // Each member's rating over one window of plays, the board's own numbers:
-// ovr is 60 to 99, or null for a rookie under 3 finished. A member with no
+// ovr is 25 to 99, or null for a rookie under 3 finished. A member with no
 // plays in the window is not in the map.
 function memberRatings(outcomes = []) {
   const groups = new Map();

package/lib/known-commands.js

@@ -3,7 +3,7 @@
 const knownCommands = ['init', 'guide', 'doctor', 'doc-health', 'install', 'workspace', 'log', 'logs', 'later', 'idea', 'done', 'wish', 'ask', 'approve', 'stop', 'ready', 'check', 'drill', 'dream', 'now', '99', 'goal', 'wtf', 'founder', 'orb', 'radar', 'who', 'stream', 'ctop', 'launchpad', 'status', 'analytics', 'revisions', 'visualize', 'brain', 'brainstorm', 'autopilot', 'run', 'plan', 'do', 'review', 'release',
                        'activate', '_activate', 'agent', 'team', 'chat', 'fast', 'ax', 'console', 'serve', 'login', 'logout', 'whoami', 'switch', 'use', 'accounts', '_resolve', '_profile-email', '_switch-session', 'shell-init', 'update', 'upgrade', 'version', 'help', 'next', 'atris',
                        'clean', 'close', 'harvest', 'verify', 'recover', 'search', 'scout', 'skill', 'member', 'codex-goal', 'app', 'apps', 'learn', 'lesson', 'taste', 'teach', 'plugin', 'experiments', 'bench', 'router', 'tree', 'receipt', 'proof', 'openclaw', 'pull', 'push', 'watch', 'cloud', 'live', 'align', 'terminal', 'computer', 'diff', 'business', 'sync', 'youtube', 'x-search',
-                       'ingest', 'query', 'lint', 'loop', 'pulse', 'task', 'mission', 'decide', 'campaign', 'agents', 'probe', 'worktree', 'land', 'caretaker', 'autoland', 'drive', 'aeo', 'slop', 'voice', 'strings', 'write', 'security-review', 'secure', 'deck', 'site', 'theme', 'card', 'reel', 'improve', 'study', 'rainmaker', 'xp', 'play', 'gm', 'game', 'x', 'recap', 'report', 'signup', 'clarity', 'interview', 'meet', 'moves', 'unknowns', 'avail', 'scoreboard', 'sync-checkout', 'zen',
+                       'ingest', 'query', 'lint', 'loop', 'pulse', 'task', 'mission', 'decide', 'campaign', 'agents', 'probe', 'worktree', 'land', 'caretaker', 'autoland', 'drive', 'aeo', 'slop', 'voice', 'strings', 'write', 'security-review', 'secure', 'deck', 'site', 'theme', 'card', 'reel', 'improve', 'study', 'rainmaker', 'xp', 'play', 'gm', 'game', 'x', 'recap', 'report', 'signup', 'clarity', 'interview', 'meet', 'moves', 'unknowns', 'avail', 'scoreboard', 'scorecard', 'rating', 'sync-checkout', 'zen',
                        'ci', 'github', 'vercel', 'supabase', 'linear', 'stripe', 'balance', 'usage', 'spend', 'api-key', 'agent-client', 'topup', 'buy', 'transactions', 'hire', 'hires', 'mail', 'design', 'mcp', 'gmail', 'calendar', 'twitter', 'slack', 'imessage', 'integrations', 'setup', 'clean-workspace', 'cw',
                        'fork', 'browse', 'publish', 'pack', 'sleep', 'wake', 'feedback', 'errors', 'wiki', 'code-review', 'cr', 'soul', 'fleet', 'fleet-report', 'loops', 'self-improve', 'compile', 'spaceship', 'truth', 'sign', 'engine', 'engines', 'link', 'playbook', 'feed', 'brief', 'social', 'host', 'people', 'follow', 'unfollow', 'friends', 'msg', 'inbox', 'invite', 'join', 'rsi', 'pay', 'trust'];