ucode-agent 1.19.0 → 1.20.0

package/src/core/loop.js

@@ -1,2277 +1,2289 @@
-/**
- * loop.js — the agent itself.
- *
- * One turn, end to end: take a line, fold the conversation if it has grown too
- * big, ask the model, and if it asked for tools, run them and ask again. Repeat
- * until it answers with prose instead of a tool call.
- *
- * Two histories are kept, deliberately:
- *   session.messages  the complete record, written to disk after every step
- *   working           what is actually sent, which may have its older turns
- *                     folded into a summary once the window gets tight
- */
-
-import path from 'node:path';
-import os from 'node:os';
-import { appendFileSync } from 'node:fs';
-import { readFile, access, mkdir } from 'node:fs/promises';
-import { testRunnerFor, relatedCommand, summariseFailures } from './tests.js';
-import { LogWatch } from './livelog.js';
-import { runningServers } from '../tools/shell.js';
-import { spawn } from 'node:child_process';
-
-import {
-  ask, model, setModel, modelName, modelList, contextLimit, rateLimits,
-  MODELS, DEFAULT_MODEL, PROVIDER, fallbackFor,
-} from './provider.js';
-import {
-  tools, runTool, describe, setRoot, setConfirm, PARALLEL_SAFE, WRITES, FILE_WRITES,
-} from '../tools/index.js';
-import { projectMap, loadMemory, remember, MEMORY_FILE } from './context.js';
-import { autoUpdate } from './updater.js';
-import { closeBrowser, forgetReviews } from '../tools/browser.js';
-import {
-  newSession, save, load, list, remove, removeAll, titleFrom,
-} from './history.js';
-import { fold, usage, tooBig, SUMMARY_PROMPT, forSummary } from './window.js';
-import { loadSkills, catalogue, findSkill, skillMessage, autoLoadFor } from './skills.js';
-import { Screen, isLabel } from '../ui/screen.js';
-import { Plain } from '../ui/plain.js';
-import { theme, blue, sky, dim, formatTokens, relativeTime, shortenPath, clip } from '../ui/theme.js';
-import { Failure, ToolFailure, Declined } from './failure.js';
-import { StuckWatch, eventFor, describeHit } from './stuck.js';
-import { serversReadySince } from '../tools/shell.js';
-import { formatDuration } from '../ui/activity.js';
-import { runDoctor } from './doctor.js';
-import { deploy } from '../tools/deploy.js';
-
-/**
- * Tool calls allowed in one turn.
- *
- * Scaffolding an app is dozens of writes before anything can even be run, so a
- * small ceiling stops a real job halfway through — which from the outside is
- * indistinguishable from the agent giving up for no reason. The runaway-loop
- * protection this exists for still works at 250; a loop burns through that
- * just as visibly, only later.
- */
-const MAX_STEPS = Number(process.env.UCODE_MAX_STEPS) || 250;
-const MAX_ARG_RETRIES = 2;
-
-/**
- * How many times a reply cut off at the output limit is asked to carry on.
- * Three covers any answer a terminal should be printing; past that the model
- * is rambling and stopping is the kinder outcome.
- */
-const MAX_CONTINUATIONS = 3;
-
-/** Read-only tools whose result line adds nothing — the user saw the output. */
-/**
- * How many rows a diff adds and removes.
- *
- * The rows come through as "+12| text" and "-12| text", with a "~" heading
- * for each file in a multi-file write and an undecorated note counting what
- * was elided. Only the signs are counted.
- */
-export function countDiff(rows = []) {
-  let added = 0;
-  let removed = 0;
-  for (const row of rows) {
-    const line = String(row ?? '');
-    if (line.startsWith('~')) continue;
-    // "… 218 more removed" / "… 508 more added" stand for rows not shown.
-    const more = /^\s*[….]+\s*(\d+)\s+more\s+(added|removed)/.exec(line);
-    if (more) {
-      if (more[2] === 'added') added += Number(more[1]);
-      else removed += Number(more[1]);
-      continue;
-    }
-    if (line.startsWith('+')) added++;
-    else if (line.startsWith('-')) removed++;
-  }
-  return { added, removed };
-}
-
-const QUIET = new Set(['read_file', 'read_files', 'list_dir', 'glob', 'grep', 'web_search', 'update_plan']);
-
-/** Tools that draw their own line, so they get no "● Doing X" line of their own. */
-const SILENT = new Set(['update_plan']);
-
-/** How many rounds of "the type check found errors, fix them" one turn may take. */
-const MAX_FIX_ROUNDS = 3;
-
-/** Files worth checking after they change. */
-const CHECKABLE = /\.(?:[cm]?[jt]sx?|py)$/i;
-
-/** Where TypeScript keeps what it learned, so the next check is a quick one. */
-export const TSBUILDINFO = 'node_modules/.cache/ucode/types.tsbuildinfo';
-
-/** TypeScript before 4.0 rejects --incremental together with --noEmit. */
-const NO_INCREMENTAL = /TS5074|TS6304|'--incremental'/;
-
-/**
- * The type check to run. Incremental by default: the first check pays the
- * full cost and writes a build info file, and every one after it reads that
- * and reports in about a second.
- */
-export function typeCheckCommand(incremental = true) {
-  const base = 'npx --no-install tsc --noEmit --pretty false';
-  return incremental ? `${base} --incremental --tsBuildInfoFile ${TSBUILDINFO}` : base;
-}
-
-/**
- * Failures that are the provider's and not the model's: busy, slow, down, or
- * unreachable. None of them should end a build — the turn moves to another
- * model and carries on from exactly where it was.
- */
-const TRANSIENT = new Set(['rate_limit', 'timeout', 'server', 'network', 'no_content']);
-/** Will another model, or a little patience, get past this? Not the daily cap: it covers them all. */
-const passing = (err) => TRANSIENT.has(err.kind) && !err.detail?.daily;
-const MAX_FAILOVERS = 8;
-const COOLDOWN = 5 * 60_000;
-
-const wait = (ms) => new Promise((r) => setTimeout(r, ms));
-
-/**
- * UCODE_TRACE=1 writes one JSON line per model call and per tool to
- * ~/.ucode/trace.jsonl (or to the path UCODE_TRACE names): how long it took,
- * tokens in and out, what was called. It is how "it feels slow" becomes a
- * number with a cause attached.
- */
-const TRACE_FILE = process.env.UCODE_TRACE
-  ? (process.env.UCODE_TRACE === '1' ? path.join(os.homedir(), '.ucode', 'trace.jsonl') : process.env.UCODE_TRACE)
-  : null;
-
-/** Tool results worth re-sending in full only while they are recent. */
-const THIN_RESULTS = new Set([
-  'read_file', 'read_files', 'grep', 'glob', 'list_dir', 'run_command', 'run_commands',
-  'look_at_app', 'web_search', 'edit_file', 'multi_edit', 'edit_files',
-]);
-
-/** Replace long strings in old tool arguments with a note of their size. */
-function thinArgs(value) {
-  if (typeof value === 'string') {
-    return value.length > 400
-      ? `[${value.length} characters, already applied — read the file if you need its current text]`
-      : value;
-  }
-  if (Array.isArray(value)) return value.map(thinArgs);
-  if (value && typeof value === 'object') {
-    return Object.fromEntries(Object.entries(value).map(([k, v]) => [k, thinArgs(v)]));
-  }
-  return value;
-}
-
-/**
- * The conversation as sent, with bulky history thinned.
- *
- * Every file the model writes travels inside its own tool call, so a
- * thirty-file app is re-sent in full on every later step — tens of thousands
- * of tokens the provider has to read before it can answer, growing with
- * each file. Beyond the last few steps that text is replaced by a note of its
- * size; the files are on disk, and the model re-reads one when it needs it.
- * Call ids and results stay paired, and the saved session keeps everything.
- */
-/** Tools whose result is the contents of one named thing, so re-reads repeat. */
-const RE_READ = new Set(['read_file', 'read_files', 'list_dir', 'grep', 'glob']);
-
-/** The thing a call was about, when two calls for it return the same text. */
-function subjectOf(call) {
-  if (!RE_READ.has(call?.name)) return null;
-  const a = call.args ?? {};
-  const what = a.path ?? (Array.isArray(a.paths) ? a.paths.join('|') : null) ?? a.pattern;
-  if (typeof what !== 'string' || !what) return null;
-  // grep and glob also depend on what was asked, not only where.
-  const extra = call.name === 'grep' || call.name === 'glob' ? `|${a.pattern ?? ''}|${a.glob ?? ''}` : '';
-  return `${call.name}:${what}${extra}`;
-}
-
-/**
- * Send each file once.
- *
- * Reading a file four times over a long task puts four copies of it in the
- * conversation, and the first three are worth nothing: the model reads the
- * newest and the older ones only cost tokens and invite it to answer from a
- * stale copy. Every superseded copy becomes a line saying where the current
- * one is. The newest is always kept whole, so nothing the model needs is
- * taken away, and the saved session still holds the lot.
- */
-export function dedupe(messages) {
-  const subject = new Map(); // toolCallId -> subject
-  for (const m of messages) {
-    if (m.role !== 'assistant' || !m.toolCalls?.length) continue;
-    for (const c of m.toolCalls) {
-      const s = subjectOf(c);
-      if (s) subject.set(c.id, s);
-    }
-  }
-  if (!subject.size) return messages;
-
-  const newest = new Map(); // subject -> index of the last result for it
-  messages.forEach((m, i) => {
-    if (m.role !== 'tool') return;
-    const s = subject.get(m.toolCallId);
-    if (s) newest.set(s, i);
-  });
-
-  return messages.map((m, i) => {
-    if (m.role !== 'tool') return m;
-    const s = subject.get(m.toolCallId);
-    if (!s || newest.get(s) === i) return m;
-    // Short results are not worth a note in place of themselves.
-    if ((m.content?.length ?? 0) < 400) return m;
-    const what = s.slice(s.indexOf(':') + 1).split('|')[0];
-    return {
-      ...m,
-      content:
-        `[${m.content.length} characters. This was read again later, and the current ` +
-        `contents of ${what} are further down this conversation — use those, not this.]`,
-    };
-  });
-}
-
-export function lean(messages, keep = 3) {
-  let seen = 0;
-  let cut = -1;
-  for (let i = messages.length - 1; i >= 0; i--) {
-    if (messages[i].role === 'assistant' && messages[i].toolCalls?.length && ++seen === keep) { cut = i; break; }
-  }
-  if (cut <= 0) return messages;
-  return messages.map((m, i) => {
-    if (i >= cut) return m;
-    if (m.role === 'assistant' && m.toolCalls?.length) {
-      return { ...m, toolCalls: m.toolCalls.map((c) => ({ ...c, args: thinArgs(c.args) })) };
-    }
-    if (m.role === 'tool' && THIN_RESULTS.has(m.name) && (m.content?.length ?? 0) > 1500) {
-      return {
-        ...m,
-        content: `${m.content.slice(0, 300)}
-… [${m.content.length} characters from an earlier step, trimmed ` +
-          'to keep the conversation fast — run the tool again if you need this now]',
-      };
-    }
-    return m;
-  });
-}
-
-function trace(event) {
-  if (!TRACE_FILE) return;
-  try { appendFileSync(TRACE_FILE, `${JSON.stringify({ at: Date.now(), ...event })}
-`); } catch { /* never fatal */ }
-}
-
-/** What /stats reports, counted as the session goes. */
-function newStats() {
-  return {
-    started: Date.now(), workMs: 0, turns: 0, steps: 0, tokensIn: 0, tokensOut: 0,
-    tools: {}, failed: 0, written: 0, edited: 0, commands: 0, builds: 0, stuck: 0,
-  };
-}
-
-const BUILD_COMMAND = /\b(?:npm|pnpm|yarn|bun)\s+(?:run\s+)?build\b|\bnext\s+build\b|\bvite\s+build\b/;
-
-function countTool(stats, call, out, err) {
-  stats.tools[call.name] = (stats.tools[call.name] ?? 0) + 1;
-  if (err) { stats.failed++; return; }
-  const a = call.args ?? {};
-  if (call.name === 'write_file') stats.written++;
-  if (call.name === 'batch_write') stats.written += a.files?.length ?? 0;
-  if (call.name === 'edit_file' || call.name === 'multi_edit') stats.edited++;
-  if (call.name === 'edit_files') stats.edited += new Set((a.edits ?? []).map((e) => e.path)).size || 1;
-  if (call.name === 'run_command') {
-    stats.commands++;
-    if (BUILD_COMMAND.test(a.command ?? '')) stats.builds++;
-  }
-  if (call.name === 'run_commands') stats.commands += a.commands?.length ?? 0;
-}
-
-/** The /stats box. */
-export function statsLines(st, messages) {
-  const n = (v) => Number(v).toLocaleString();
-  const top = Object.entries(st.tools).sort((x, y) => y[1] - x[1]).slice(0, 6)
-    .map(([k, v]) => `${k} ${v}`).join(' · ') || 'none yet';
-  const plural = (v, w) => `${v} ${w}${v === 1 ? '' : 's'}`;
-  const rows = [
-    ['Session', `${formatDuration(Date.now() - st.started)} open · ${formatDuration(st.workMs)} working · ${plural(st.turns, 'request')}`],
-    ['Model', `${n(st.steps)} steps · ${formatTokens(st.tokensIn)} in · ${formatTokens(st.tokensOut)} out`],
-    ['Tools', top],
-    ['Files', `${st.written} written · ${st.edited} edited`],
-    ['Commands', `${st.commands} run · ${plural(st.builds, 'build')}${st.failed ? ` · ${plural(st.failed, 'tool call')} failed` : ''}`],
-    ['Context', `${messages} messages${st.stuck ? ` · ${plural(st.stuck, 'loop')} caught` : ''}`],
-  ];
-  const w = Math.max(...rows.map(([k]) => k.length));
-  return ['', `  ${blue('Stats')}`, ...rows.map(([k, v]) => `  ${dim(k.padEnd(w))}  ${v}`), ''];
-}
-
-/** Parallel workers at once, and how many steps each may take. */
-const MAX_WORKERS = 3;
-const WORKER_STEPS = Number(process.env.UCODE_WORKER_STEPS) || 60;
-
-/** Workers build; they do not plan, delegate further, or load skills themselves. */
-const WORKER_EXCLUDED = new Set(['delegate', 'update_plan', 'load_skill']);
-
-const planTool = {
-  name: 'update_plan',
-  description:
-    'Keep a short checklist the user can see, for work with three or more steps. ' +
-    'Send the whole list every time: at most 6 items, a few words each, with done: true ' +
-    'on the finished ones. Update it as items finish. Skip it for small tasks.',
-  parameters: {
-    type: 'object',
-    properties: {
-      items: {
-        type: 'array',
-        description: 'The whole plan, in order. At most 6.',
-        items: {
-          type: 'object',
-          properties: {
-            text: { type: 'string', description: 'A few words: "Build the upload box".' },
-            done: { type: 'boolean', description: 'True once it is finished.' },
-          },
-          required: ['text'],
-        },
-      },
-    },
-    required: ['items'],
-  },
-};
-
-const delegateTool = {
-  name: 'delegate',
-  description:
-    'Build independent parts in parallel. Up to 3 workers run at once, each with the ' +
-    'same tools as you. Use it when the work splits cleanly into parts that touch ' +
-    'different files - e.g. the API route, the upload component and the results view. ' +
-    'A worker sees only its instructions, so make them complete: the files it owns, ' +
-    'what to build, the exact interfaces (props, types, request and response shapes) ' +
-    'it must match, and the design rules. Set up shared files (package.json, design ' +
-    'tokens, shared types) yourself first. You get back each worker\'s summary and ' +
-    'the files it changed; wire the parts together and check the whole afterwards.',
-  parameters: {
-    type: 'object',
-    properties: {
-      tasks: {
-        type: 'array',
-        description: 'Up to 3 independent pieces of work.',
-        items: {
-          type: 'object',
-          properties: {
-            name: { type: 'string', description: 'Two or three words: "api route", "upload ui".' },
-            instructions: { type: 'string', description: 'Everything the worker needs to do its part completely.' },
-          },
-          required: ['name', 'instructions'],
-        },
-      },
-    },
-    required: ['tasks'],
-  },
-};
-
-/** The instructions a parallel worker starts with. */
-function workerPrompt({ cwd, name, memory, skills, map }) {
-  return [
-    `You are a ucode worker called "${name}" - one of several building parts of the same project at the same time.`,
-    '',
-    `Working directory: ${cwd}`,
-    `Platform: ${process.platform}`,
-    '',
-    '- Do exactly the task you were given. Touch only the files it names, or new files in',
-    '  the area it owns - other workers are editing the rest of the project right now.',
-    '- Read before you edit. read_files for several files, batch_write for several new',
-    '  files, edit_files for changes across files.',
-    '- Nothing has a keyboard: pass non-interactive flags. Do not start dev servers and do',
-    '  not install packages unless the task says to - say what you need instead.',
-    '- Finish what you build: real content, every state handled, no TODOs.',
-    '- When done, reply with two or three sentences: what you built, in which files, and',
-    '  anything the lead has to wire up.',
-    '- The files you own may not exist yet - create them. Do not go looking for them',
-    '  first. The project map below shows what does exist; read only the files whose',
-    '  interfaces you must match, then start writing within two or three steps.',
-    ...(memory ? ['', '## Project memory', '', memory] : []),
-    ...(map ? ['', '## Project map', '', map] : []),
-    ...(skills ? ['', '## Instructions in force', '', skills] : []),
-  ].join('\n');
-}
-
-/** The files a writing tool call touches. */
-function pathsOf(call) {
-  const a = call.args ?? {};
-  if (call.name === 'batch_write' || call.name === 'edit_files') return (a.files ?? []).map((f) => f?.path).filter(Boolean);
-  return a.path ? [a.path] : [];
-}
-
-const exists = (p) => access(p).then(() => true, () => false);
-
-/** Skills reach the model as one extra tool, so bodies load only when wanted. */
-const loadSkillTool = {
-  name: 'load_skill',
-  description:
-    'Load the full instructions for one of the skills listed in your system prompt. ' +
-    'Call it the moment a task matches one — before planning, before writing anything ' +
-    '— then follow what it says.',
-  parameters: {
-    type: 'object',
-    properties: { name: { type: 'string', description: 'The skill name, exactly as listed.' } },
-    required: ['name'],
-  },
-};
-
-function systemPrompt({ cwd, skills, mode, check, map, memory }) {
-  const list = catalogue(skills);
-
-  return [
-    'You are ucode, a coding agent working directly in the user\'s terminal.',
-    '',
-    `Working directory: ${cwd}`,
-    `Platform: ${process.platform}`,
-    ...(memory ? [
-      '',
-      '## Project memory',
-      '',
-      'Standing instructions from the user. They outrank your defaults.',
-      '',
-      memory,
-    ] : []),
-    '',
-    '## How to work',
-    '',
-    'Before building anything, turn the request into a list of what it must do — every',
-    'feature named, and the ones any user would expect whether or not they were named',
-    '(an empty state, an error state, the keyboard doing the obvious thing, working on',
-    'a phone). Keep it as the plan. Build against it, then go through it one item at a',
-    'time before you say a word about being finished. Most of what gets missed was',
-    'never written down.',
-    '',
-    'Do not put code in your reply. Not a snippet, not "here is the key part", not a',
-    'summary of the file. It is already in the file and the user can open it; pasting',
-    'it again buries the one or two sentences that actually matter. Say what it does',
-    'and what to try.',
-    '',
-    'Do not claim it is done while anything is still running or unchecked. "I have',
-    'built it" said before the build finishes is worse than saying nothing: the user',
-    'believes you, looks, and finds it broken. Finish, check, then say so — and if',
-    'something is incomplete, say which part and why.',
-    '',
-    'BE FAST. Every tool call is a round trip, and round trips are nearly all of the',
-    'time a build takes. So: write a whole app in ONE batch_write rather than a',
-    'write_file per file. Read every file you need in ONE read_files. Never read a',
-    'file you just wrote, and never read one back after edit_file — the result',
-    'already contains it. Do not re-check work the checks have already reported on.',
-    'Fast is not sloppy: it is the same work with the waiting taken out.',
-    '',
-    'SPEAK AS "I", NEVER "WE". You are doing this, not a committee: "I will build',
-    'Tide as a single HTML file", not "we have created the file".',
-    '',
-    'Say three or four things over a whole build, not one per step:',
-    '  - Open with what you are going to make and how, in one line, before anything',
-    '    else: "I will build Tide as a single HTML file - markup, one stylesheet, one',
-    '    module - with the tasks kept in localStorage."',
-    '  - One line when you move between the big pieces of work: "The layout is done,',
-    '    now the animations."',
-    '  - One line at the end saying what it does and how to try it.',
-    'That is all. A line before every tool call is not narration, it is noise: the',
-    'steps already show on screen, and repeating them in words buries the few',
-    'sentences worth reading.',
-    '',
-    'FIRST, EVERY TIME: write one short line saying what you are about to do, then',
-    'make the tool calls. Never open a turn with a tool call and no words. Examples:',
-    '"Right, the HTML structure first." / "Now the state and the render loop." /',
-    '"That is the layout done - onto the animations." / "Let me see what is there."',
-    'One sentence, your own voice, before the actions - not after them, not instead',
-    'of them, and not a restatement of what was asked. The user is watching this',
-    'scroll past; without those lines it is a list of file operations and they cannot',
-    'tell what you are building. This matters as much as the code.',
-    '',
-
-    '',
-    'Before you guess at an API, ask: type_of gives the exact signature from the',
-    'TypeScript this project has installed, and find_symbol says where something is declared without',
-    'reading five files to find it. Rename with rename_symbol rather than edit_file — a',
-    'find-and-replace that matches too much is the most common broken edit. Reach for',
-    'add_block before writing a table, an empty state or a dashboard by hand.',
-    '',
-    ...(mode === 'plan' ? [
-      'You are in PLAN MODE. Reading, searching and research are available; every tool',
-      'that writes a file or runs a command has been withheld. Investigate, then set out',
-      'what you would change: which files, which functions, in what order. Never imply',
-      'you have made a change you are not able to make.',
-      '',
-    ] : []),
-    ...(check ? [
-      `This project checks itself with: ${check}`,
-      'After changing code, run that and report what actually happened. "It should work"',
-      'is not a result.',
-      '',
-    ] : []),
-    '- Read before you write. Never edit a file you have not read this session.',
-    '- Prefer edit_file to write_file. Rewrite a whole file only when creating it, or',
-    '  when the change genuinely touches most of it.',
-    '- old_string must be copied out of what read_file showed you, character for',
-    '  character, without the line-number gutter, and must appear exactly once. Add',
-    '  surrounding lines until it does.',
-    '- read_file returns up to 600 lines. Read the whole file before editing it rather',
-    '  than editing from a fragment; pass offset to continue a long one.',
-    '',
-    '## Going fast',
-    '',
-    'Every tool call is a round trip to you, and the round trip — not the disk, not',
-    'the shell — is where the time goes. So:',
-    '',
-    '- Need more than one file? read_files, all of them in one call. Never read files',
-    '  one at a time when you already know which ones you want.',
-    '- An edit result shows the file as it now stands. Do not read a file again after',
-    '  editing it - you already have its current text.',
-    '- Put independent calls in the same message — several greps, a glob and a read.',
-    '  Read-only calls in one message run at the same time.',
-    '- New Next.js app? create_app - one step, never create-next-app or shadcn init. It',
-    '  copies a starter that already builds and installs it in the background.',
-    '- batch_write to lay out several new files at once - but at most about four per',
-    '  call: one slip in a huge call throws the whole call away.',
-    '- multi_edit for several changes',
-    '  to one file, edit_files for a change that spans several files.',
-    '- For work with three or more steps, keep a short plan with update_plan - at most',
-    '  six items of a few words - and tick items off as they finish. Skip it for small jobs.',
-    '- When a build splits into parts that touch different files (the API route, the',
-    '  upload component, the results view), set up the shared files yourself, then hand',
-    '  the parts to delegate so they are built in parallel.',
-    '- A package.json you write starts installing in the background immediately; keep',
-    '  writing files. Running the install yourself afterwards just waits for that one.',
-    '- When you finish, ucode type-checks what you changed and hands you the errors, so',
-    '  there is no need to run tsc yourself.',
-    '- Do not open or drive a browser. Checking the page in one is something the user',
-    '  asks for with /look; your job is to leave the app in a state worth looking at.',
-    '  reports - errors, layout that overflows a phone, the review points worth fixing -',
-    '  in one pass, then look once more. A clean second look means it is done: report',
-    '  back instead of polishing in circles. Never call an interface finished unlooked at.',
-    '  name, handles keys and returns the live link. Build locally first.',
-    '- Nothing you run has a keyboard. Pass the non-interactive flag to anything that',
-    '  would ask a question, or it fails instead of waiting: create-next-app --yes,',
-    '  npx shadcn@latest init -d -y, npx shadcn@latest add <names> -y, npm init -y.',
-    '- Dev servers start in the background by themselves, and the result tells you the',
-    '  URL once the server says it is ready. Do not start one twice, do not sleep while',
-    '  waiting for it, and do not curl it before that result comes back.',
-    '',
-    '## When something fails',
-    '',
-    '- A failed build names the problem. Fix exactly that, then build again. Never go',
-    '  exploring inside node_modules: a missing component or package is one install away.',
-    '- A build takes most of a minute. Fix every error it lists in one pass - multi_edit,',
-    '  edit_files - before building again, never one error per build.',
-    '- Never delete an app folder to start over. Fix it where it is - starting again throws',
-    '  away the install and everything already written.',
-    '- Run an app\'s commands with cwd set to its folder, and keep paths inside those',
-    '  commands relative to that folder.',
-    '',
-    '## Safety',
-    '',
-    '- run_command runs without asking. That is trust rather than licence: never run',
-    '  anything destructive the user did not ask for.',
-    '- Paths are relative to the working directory. Anything outside it needs the user',
-    '  to approve it first.',
-    '- Verify. After changing code, run the tests or a quick check with run_command.',
-    '',
-    '## Saying what you are doing',
-    '',
-    '- Before every tool call, write ONE short line naming the file or command:',
-    '  "Reading tui.js", "Fixing the spinner in loop.js", "Running npm test".',
-    '- Present tense, under ten words, and no full stop at the end. It is a label on',
-    '  work happening right now, not a sentence about work that is finished.',
-    '- That line and nothing else in the message. No preamble, no plan, no bullets —',
-    '  the user reads it live while the tool runs.',
-    '- Say the next one when you take the next step, not all of them up front.',
-    '',
-    '## Answering',
-    '',
-    '- Be short. Two or three sentences is usually the entire answer. This is a',
-    '  terminal, not a document.',
-    '- No preamble, no restating the question, no "I will now...". Just answer.',
-    '- Do not narrate what the tool output already showed. The user watched the diff',
-    '  and the command output; cover only what is not obvious from them.',
-    '- Skip closing summaries of work the user just watched you do — but never end',
-    '  a turn silently. If there is genuinely nothing to add, one short line saying',
-    '  what changed is the whole answer.',
-    '- Markdown. Fenced blocks with a language tag get highlighted.',
-    '- Point at code as path:line so the user can jump straight to it.',
-    '- Report honestly. If a command failed or you skipped something, say so.',
-    '- Length tracks the question: a one-line question gets a one-line answer.',
-    '- Brevity is about your prose and never about your work. What you build is',
-    '  finished: every control wired, every state handled, no TODO left behind.',
-    ...(list ? [
-      '',
-      '## Skills',
-      '',
-      'These instruction packs are available. When a task matches one, load it with',
-      'load_skill as your first step — before planning, before writing anything — and',
-      'then follow it. A skill already in this conversation outranks your own defaults',
-      'and is not advisory.',
-      '',
-      list,
-    ] : []),
-    '',
-    '## Project map',
-    '',
-    'Every file in the project at the start of this turn, with the names each code file',
-    'exports. Go straight to the files you need instead of searching for them.',
-    '',
-    map || '(not available)',
-  ].join('\n');
-}
-
-export class Agent {
-  constructor({ cwd, debug = false }) {
-    this.cwd = cwd;
-    this.debug = debug;
-    // A full-screen layout only makes sense on a real terminal. Piped input,
-    // CI and `echo ... | ucode` get the line-based interface instead.
-    this.full = Boolean(process.stdout.isTTY && process.stdin.isTTY);
-    this.ui = this.full ? new Screen({ cwd }) : new Plain({ cwd });
-    this.stats = newStats();
-    this.skills = [];
-    this.session = newSession(cwd, model());
-    this.working = [];
-    this.loaded = new Set();
-    this.abort = null;
-    this.busy = false;
-    this.check = null;
-  }
-
-  // -- history -------------------------------------------------------------
-
-  push(message) {
-    this.session.messages.push(message);
-    this.working.push(message);
-  }
-
-  /**
-   * Answer every tool call a stopped turn never got to.
-   *
-   * An assistant message ends by asking for tools, and each of those asks
-   * needs an answer. Abandon them and the conversation is left mid-sentence,
-   * so the next time the model reads it the only sensible thing to do is
-   * carry on where it left off — which is exactly what the user pressed stop
-   * to prevent. Saying "this did not happen" for each one ends the sentence,
-   * and a line from the user ends the task.
-   */
-  closeInterrupted() {
-    const answered = new Set(this.working.filter((m) => m.role === 'tool').map((m) => m.toolCallId));
-    const missing = [];
-    for (const m of this.working) {
-      if (m.role !== 'assistant' || !m.toolCalls?.length) continue;
-      for (const call of m.toolCalls) {
-        if (!answered.has(call.id)) missing.push(call);
-      }
-    }
-    for (const call of missing) {
-      this.push({
-        role: 'tool',
-        toolCallId: call.id,
-        name: call.name,
-        content: 'The user stopped the turn before this ran. It did not happen, and it must not be retried.',
-      });
-    }
-    if (missing.length) {
-      this.push({
-        role: 'user',
-        content: 'I stopped that. Drop it and wait for what I ask next — do not pick it back up.',
-      });
-    }
-    return missing.length;
-  }
-
-  async persist() {
-    try {
-      this.session.model = model();
-      await save(this.session);
-    } catch (err) {
-      // Losing the save must not lose the turn.
-      this.ui.error(err, { debug: this.debug });
-    }
-  }
-
-  // -- startup -------------------------------------------------------------
-
-  /**
-   * Everything a turn needs, minus the terminal.
-   *
-   * Split out of start() so another front end could prepare an agent and drive
-   * turn() itself.
-   */
-  async bootstrap() {
-    setRoot(this.cwd);
-    // Parallel workers can ask at the same moment; the questions queue up and
-    // are put to the user one at a time, never on top of each other.
-    let asking = Promise.resolve();
-    setConfirm((request) => {
-      const next = asking.then(() => this.ui.confirm(request));
-      asking = next.catch(() => {});
-      return next;
-    });
-    this.skills = await loadSkills({ cwd: this.cwd });
-    await this.detectCheck();
-  }
-
-  async start() {
-    await this.bootstrap();
-
-    if (this.full) {
-      await this.ui.start();
-      this.ui.onInterrupt = () => {
-        if (this.busy && this.abort) {
-          this.abort.abort();
-          this.ui.stopSpinner();
-          this.ui.stopTimer?.();
-          this.ui.note('interrupted');
-        }
-      };
-      this.ui.onModeChange = () => this.showHeader({ clear: false });
-    }
-
-    for (const problem of this.skills.problems ?? []) {
-      this.ui.write(theme.warn(`  skill not loaded: ${problem}`));
-    }
-
-    this.showHeader();
-    this.installSignals();
-
-    // Checked in the background; nothing here waits on it.
-    autoUpdate({
-      onUpdated: (version) => {
-        this.ui.setFacts?.({ update: version });
-        if (!this.ui.welcoming?.()) this.ui.note(`updated to v${version} — it takes over the next time you start ucode`);
-      },
-    });
-    await this.repl();
-  }
-
-  showHeader({ clear = true } = {}) {
-    if (clear && this.full) this.ui.clearScreen();
-    const stats = usage(this.working, contextLimit());
-    this.ui.header({
-      cwd: this.cwd,
-      model: modelName(),
-      used: stats.used,
-      limit: stats.limit,
-      title: this.session.title === 'Untitled' ? 'new session' : this.session.title,
-    });
-  }
-
-  installSignals() {
-    const flush = async () => {
-      await save(this.session).catch(() => {});
-      process.exit(0);
-    };
-    process.on('SIGTERM', flush);
-
-    // The full-screen UI reads keys itself, so it owns ctrl+c and esc.
-    if (this.full) return;
-
-    this.ui.rl.on('SIGINT', () => {
-      if (this.busy && this.abort) {
-        this.abort.abort();
-        this.ui.stopSpinner();
-        this.ui.write(dim('  interrupted'));
-        return;
-      }
-      this.ui.write(dim('  (ctrl+d or /exit to quit)'));
-      this.ui.rl.prompt();
-    });
-  }
-
-  /**
-   * How this project verifies itself, worked out once at startup. Null when
-   * there is genuinely nothing to run — verification is only insisted on where
-   * there is something to insist on.
-   */
-  async detectCheck() {
-    const has = (f) => readFile(path.join(this.cwd, f)).then(() => true, () => false);
-
-    if (await has('package.json')) {
-      try {
-        const pkg = JSON.parse(await readFile(path.join(this.cwd, 'package.json'), 'utf8'));
-        if (pkg.scripts?.test && !/no test specified/i.test(pkg.scripts.test)) {
-          this.check = 'npm test';
-          return;
-        }
-      } catch { /* an unreadable package.json is not worth failing over */ }
-    }
-    if (await has('Cargo.toml')) { this.check = 'cargo test'; return; }
-    if (await has('go.mod')) { this.check = 'go test ./...'; return; }
-    if (await has('pyproject.toml') || await has('pytest.ini')) { this.check = 'pytest'; return; }
-    if (await has('Makefile')) { this.check = 'make test'; return; }
-    this.check = null;
-  }
-
-  // -- REPL ----------------------------------------------------------------
-
-  async repl() {
-    let sawInput = false;
-
-    for (;;) {
-      const line = await this.ui.ask();
-      if (line === null) {
-        // End of input before anything was typed. On Windows this is almost
-        // always npm's PowerShell shim, which runs the CLI as `$input | node`.
-        // The pipe makes stdin a non-TTY, readline hits EOF at once, and the
-        // banner flashes up and vanishes — which looks like a crash rather
-        // than like a program that was never given a keyboard. So say which.
-        if (!sawInput && !process.stdin.isTTY) this.explainNoKeyboard();
-        break;
-      }
-
-      const input = line.trim();
-      if (input) sawInput = true;
-      if (!input) continue;
-
-      if (input.startsWith('/')) {
-        if (await this.command(input) === 'exit') break;
-        continue;
-      }
-
-      try {
-        await this.turn(input);
-      } catch (err) {
-        this.ui.error(err, { debug: this.debug });
-      }
-    }
-
-    await this.shutdown();
-  }
-
-  explainNoKeyboard() {
-    this.ui.blank();
-    this.ui.write(theme.warn('  ucode could not reach the keyboard, so it stopped.'));
-    this.ui.blank();
-    this.ui.write('  That happens when input is piped rather than typed. On Windows it is');
-    this.ui.write("  usually npm's PowerShell wrapper, which pipes stdin.");
-    this.ui.blank();
-    this.ui.write(`  ${blue('Any of these work:')}`);
-    this.ui.write(`    ${sky('ucode.cmd')}          the cmd shim, which keeps the keyboard`);
-    this.ui.write(`    ${sky('npx ucode-agent')}    runs it directly`);
-    this.ui.write('    or start it from Command Prompt or Windows Terminal');
-    this.ui.blank();
-  }
-
-  async shutdown() {
-    await closeBrowser().catch(() => {});
-    this.ui.stopSpinner();
-    if (this.session.messages.length) {
-      await this.persist();
-      this.ui.write(dim(`\n  saved · ${this.session.title}`));
-    }
-    this.ui.close();
-  }
-
-  // -- one turn ------------------------------------------------------------
-
-  /**
-   * Pull image paths out of the message and load them, so "what is wrong in
-   * screenshot.png" works without a separate command for it.
-   */
-  async attachImages(input) {
-    const mentioned = input.match(/[^\s"']+\.(?:png|jpe?g|gif|webp)\b/gi) ?? [];
-    const images = [];
-
-    for (const name of mentioned) {
-      const file = path.resolve(this.cwd, name);
-      try {
-        const buf = await readFile(file);
-        if (buf.length > 4 * 1024 * 1024) {
-          this.ui.note(`${name} is ${(buf.length / 1024 / 1024).toFixed(1)}MB — too big to send, skipped`);
-          continue;
-        }
-        const ext = path.extname(file).toLowerCase().slice(1);
-        images.push(`data:image/${ext === 'jpg' ? 'jpeg' : ext};base64,${buf.toString('base64')}`);
-        this.ui.note(`attached ${name}`);
-      } catch {
-        // Just a filename mentioned in passing, not a file on disk.
-      }
-    }
-
-    return images;
-  }
-
-  /**
-   * Skills that this request should arrive with, already loaded.
-   *
-   * The load_skill tool asks the model to notice that a task needs a skill,
-   * and a model in a hurry to be helpful does not always notice. For work
-   * where the skill *is* the quality bar — anything with a user interface in
-   * it — that is not a discovery to make after the app has been built. So the
-   * request is matched against each skill's trigger words and the body goes in
-   * before the model takes its first step.
-   */
-  autoLoad(input) {
-    for (const skill of autoLoadFor(this.skills, input)) {
-      if (this.loaded.has(skill.name)) continue;
-      this.loaded.add(skill.name);
-      this.push(skillMessage(skill, { automatic: true }));
-      this.ui.note(`${skill.name} skill loaded for this`);
-    }
-  }
-
-  async turn(input) {
-    forgetReviews(); // a new request: its apps get a fresh design review
-    const images = await this.attachImages(input);
-    this.push(images.length
-      ? { role: 'user', content: input, images }
-      : { role: 'user', content: input });
-
-    if (!this.session.title || this.session.title === 'Untitled') {
-      this.session.title = titleFrom(input);
-    }
-
-    this.autoLoad(input);
-    // What the model is told about the project, fresh for this turn.
-    [this.map, this.memory] = await Promise.all([
-      projectMap(this.cwd).catch(() => ''),
-      loadMemory(this.cwd).catch(() => ''),
-    ]);
-    await this.persist();
-
-    // A busy model was swapped for a fallback earlier; after a few minutes the
-    // one the user chose gets another go.
-    this.preferred ??= model();
-    if (model() !== this.preferred && Date.now() > (this.cooldownUntil ?? 0)) {
-      setModel(this.preferred);
-      if (this.full) this.showHeader({ clear: false });
-    }
-
-    this.busy = true;
-    this.abort = new AbortController();
-    this.endedSilently = false;
-
-    const turnStarted = Date.now();
-    let finished = false;
-    this.ui.turnStart?.();
-    try {
-      for (;;) {
-        try {
-          await this.run();
-          break;
-        } catch (err) {
-          // The daily free cap mid-build: wait for the reset and carry on,
-          // rather than leaving a half-built app for the user to restart.
-          if (err?.detail?.daily && !this.abort.signal.aborted && (await this.waitForReset(err))) continue;
-          throw err;
-        }
-      }
-      finished = true;
-    } catch (err) {
-      if (err?.kind === 'aborted' || this.abort.signal.aborted) this.ui.write(dim('  turn cancelled'));
-      else throw err;
-    } finally {
-      trace({ kind: 'turn', ms: Date.now() - turnStarted });
-      this.stats.workMs += Date.now() - turnStarted;
-      this.stats.turns++;
-      if (!finished) this.closeInterrupted();
-      const ok = finished && !this.endedSilently;
-      this.busy = false;
-      this.abort = null;
-      this.ui.stopSpinner();
-      this.ui.stopTimer?.();
-      this.activity = null;
-      await this.persist();
-      if (this.full) this.showHeader({ clear: false });
-      if (finished) this.openWhenReady(turnStarted);
-      // Last, so "Done" is the last thing that happens rather than the last
-      // thing said before several more things happen.
-      this.ui.turnEnd?.({ ok });
-    }
-  }
-
-  /** The tools the model may see, given the mode. */
-  toolsNow() {
-    const all = [...tools, loadSkillTool, planTool, delegateTool];
-    if (this.ui.mode !== 'plan') return all;
-    return all.filter((t) => !WRITES.has(t.name));
-  }
-
-  /** Model, tools, model, until it answers with prose. */
-  async run() {
-    const available = this.toolsNow();
-    let argRetries = 0;
-    let continuations = 0;
-    let askedToVerify = false;
-    let askedToSpeak = false;
-    let fixRounds = 0;
-    this.failovers = 0;
-    this.tried = new Set([model()]);
-
-    this.stuck = new StuckWatch();
-    this.touched = new Set();
-    this.sinceCheck = new Set();
-    this.logWatch = new LogWatch();
-    this.ranSomething = false;
-
-    for (let step = 0; step < MAX_STEPS; step++) {
-      await this.maybeFold();
-      this.ui.startSpinner(step === 0 ? 'thinking' : 'working');
-
-      let reply;
-      let streaming = false;
-      this.early = new Map();
-
-      try {
-        const opts = {
-          signal: this.abort.signal,
-          onWait: (text) => this.ui.updateSpinner(text),
-        };
-        // Only a real terminal has somewhere to stream into.
-        if (this.full) {
-          opts.onThinking = (delta) => this.ui.thinkingDelta(delta);
-          opts.onText = (delta) => {
-            if (!streaming) {
-              streaming = true;
-              this.ui.thinkingEnd();
-              this.ui.streamBegin();
-            }
-            this.ui.streamDelta(delta);
-          };
-          // Read-only calls start the moment they are fully written, while the
-          // rest of the reply is still arriving. Nothing that writes or runs is
-          // started early: a reply that fails halfway must leave no side effects.
-          opts.onToolCall = (call) => {
-            if (PARALLEL_SAFE.has(call.name) && !call.parseError && !this.early.has(call.id)) {
-              this.early.set(call.id, this.execute(call));
-            }
-          };
-        }
-
-        var asked = Date.now();
-        reply = await ask(
-          [
-            {
-              role: 'system',
-              content: systemPrompt({
-                cwd: this.cwd,
-                skills: this.skills,
-                mode: this.ui.mode,
-                check: this.check,
-                map: this.map,
-                memory: this.memory,
-              }),
-            },
-            ...dedupe(lean(this.working)),
-          ],
-          available,
-          opts
-        );
-      } catch (err) {
-        this.ui.thinkingEnd();
-        if (streaming) this.ui.streamEnd();
-
-        // The model invented a tool and the provider rejected the request
-        // outright. Tell it what it did and let it try again.
-        if (err.kind === 'bad_tool_call' && argRetries < MAX_ARG_RETRIES) {
-          argRetries++;
-          this.ui.stopSpinner();
-          this.ui.toolFailed(
-            `${err.detail?.attemptedName ?? 'invalid tool call'} — retrying (${argRetries}/${MAX_ARG_RETRIES})`
-          );
-          this.push({
-            role: 'user',
-            content:
-              `Your last tool call was rejected. ${err.failed} The only tools that exist ` +
-              `are: ${available.map((t) => t.name).join(', ')}. Try again with one of them.`,
-          });
-          continue;
-        }
-
-        // Busy, slow or down: move to the next model and carry on, rather
-        // than ending a half-built app with an error.
-        if (passing(err) && !this.abort.signal.aborted && (await this.failover(err))) continue;
-        throw err;
-      }
-
-      // A reply that is nothing but tool calls never starts a text stream, so
-      // the thinking timer has to be closed out here as well.
-      this.ui.thinkingEnd();
-      this.ui.stopSpinner();
-      this.record(reply.usage);
-      this.ui.step?.();
-      this.stats.steps++;
-      this.stats.tokensIn += reply.usage.promptTokens ?? 0;
-      this.stats.tokensOut += reply.usage.outputTokens ?? 0;
-      trace({
-        kind: 'model', who: 'lead', model: model(), ms: Date.now() - asked,
-        in: reply.usage.promptTokens, out: reply.usage.outputTokens,
-        calls: reply.toolCalls.map((c) => c.name),
-      });
-
-      // Streamed text is already on screen; turn it into rendered markdown.
-      // Text that turns out to be narration ahead of a tool call folds into a
-      // status line instead — that is where the live commentary comes from.
-      const narrating = reply.toolCalls.length > 0;
-      if (streaming) this.ui.streamEnd({ asNarration: narrating });
-      else if (reply.text && narrating && isLabel(reply.text)) this.ui.narrate(reply.text);
-      else if (reply.text) this.ui.assistant(reply.text);
-
-      if (reply.toolCalls.length === 0) {
-        // The answer stopped at the provider's output cap rather than at the
-        // end of a thought, so it is cut mid-word. Ask for the rest instead of
-        // handing over half an answer with no sign there was more.
-        if (reply.finishReason === 'length' && reply.text && continuations < MAX_CONTINUATIONS) {
-          continuations++;
-          this.push({ role: 'assistant', content: reply.text });
-          this.push({
-            role: 'user',
-            content:
-              'Your reply stopped at the output limit, mid-sentence. Carry on from exactly ' +
-              'where it broke off. Do not repeat any of it, do not start again, and do not ' +
-              'introduce it — just continue.',
-          });
-          this.ui.note('hit the output limit — asking for the rest');
-          continue;
-        }
-
-        // Type-check what changed and hand back the errors, a few rounds at most.
-        // A turn that ends on a broken build is the most common way an app
-        // gets handed over as done when it is not.
-        if (fixRounds < MAX_FIX_ROUNDS) {
-          const problems = await this.autoCheck();
-          if (problems) {
-            fixRounds++;
-            if (reply.text) this.push({ role: 'assistant', content: reply.text });
-            this.push({
-              role: 'user',
-              content:
-                `ucode checked the files you changed and found errors (round ${fixRounds} of ` +
-                `${MAX_FIX_ROUNDS}). Fix all of them, then finish.\n\n${problems}`,
-            });
-            continue;
-          }
-        }
-
-        // It changed code and never ran anything. Send it back once.
-        if (this.touched.size && !this.ranSomething && this.check && !askedToVerify) {
-          askedToVerify = true;
-          if (reply.text) this.push({ role: 'assistant', content: reply.text });
-          this.push({
-            role: 'user',
-            content:
-              `You changed ${[...this.touched].join(', ')} and did not check it. Run ` +
-              `\`${this.check}\` with run_command now, then say what actually happened — ` +
-              'if it failed, show the output rather than claiming it worked. If that is ' +
-              'the wrong way to check this project, run the right one and say which.',
-          });
-          this.ui.note('verifying the change');
-          continue;
-        }
-
-        /**
-         * It did the work and then said nothing at all.
-         *
-         * Reasoning models do this, and the instruction to skip closing
-         * summaries makes it more likely. Silence is fine as a style; it is
-         * not fine as an answer, because the user cannot tell it apart from a
-         * crash — and if they asked what happened, they asked. One nudge,
-         * once per turn, and only when there was actually work to report.
-         */
-        if (!reply.text?.trim() && this.touched.size + (this.ranSomething ? 1 : 0) > 0 && !askedToSpeak) {
-          askedToSpeak = true;
-          this.push({
-            role: 'user',
-            content:
-              'You stopped without saying anything. If the thing I asked for is not ' +
-              'built yet, carry on and build it. If it is, tell me in one or two ' +
-              'sentences what it does and how to try it. No preamble, no diffs.',
-          });
-          continue;
-        }
-
-        this.push({ role: 'assistant', content: reply.text });
-        if (!reply.text?.trim()) {
-          // Silence after being asked to speak is not a finished turn. Saying
-          // "Done" here is the worst thing available: the user believes it,
-          // looks, and finds the thing they asked for was never built.
-          this.ui.note('the model stopped without saying anything — the work may be unfinished');
-          this.endedSilently = true;
-        }
-        return;
-      }
-
-      this.push({ role: 'assistant', content: reply.text || '', toolCalls: reply.toolCalls });
-      await this.persist();
-
-      const badArgs = await this.runCalls(reply.toolCalls);
-      if (this.abort.signal.aborted) return;
-
-      // Malformed arguments go back to the model, but only so many times.
-      if (badArgs) {
-        argRetries++;
-        if (argRetries > MAX_ARG_RETRIES) {
-          throw new Failure({
-            kind: 'bad_tool_args',
-            attempted: 'running the tools the model asked for',
-            failed:
-              `${modelName()} produced invalid tool arguments ${argRetries} times running ` +
-              'and could not correct itself.',
-            fix:
-              'Say what you want more concretely, or /model to another one — North Mini ' +
-              'Code and Nemotron 3.5 Lightning are both steadier with tool arguments.',
-          });
-        }
-      } else {
-        argRetries = 0;
-      }
-
-      await this.persist();
-    }
-
-    throw new Failure({
-      kind: 'step_limit',
-      attempted: 'finishing your request',
-      failed: `The model was still calling tools after ${MAX_STEPS} steps.`,
-      fix:
-        'Nothing is lost — everything so far is on disk. Say "carry on where you left ' +
-        'off" to continue. If it was repeating one step, it is looping: break the task ' +
-        'up, or /new to reset.',
-    });
-  }
-
-  /**
-   * Run one round of tool calls.
-   *
-   * Consecutive read-only calls go out together — four files read at once
-   * rather than four round trips — while anything that writes or executes runs
-   * on its own, in order. Returns whether any call had unusable arguments.
-   */
-  async runCalls(calls) {
-    const groups = [];
-    let batch = [];
-
-    for (const call of calls) {
-      if (PARALLEL_SAFE.has(call.name)) {
-        batch.push(call);
-      } else {
-        if (batch.length) { groups.push(batch); batch = []; }
-        groups.push([call]);
-      }
-    }
-    if (batch.length) groups.push(batch);
-
-    let badArgs = false;
-
-    for (const group of groups) {
-      if (this.abort.signal.aborted) return badArgs;
-
-      const noted = (call) => {
-        if (FILE_WRITES.has(call.name)) {
-          for (const p of pathsOf(call)) { this.touched.add(p); this.sinceCheck.add(p); }
-        }
-        if (call.name === 'run_command' || call.name === 'run_commands') this.ranSomething = true;
-      };
-
-      if (group.length > 1) {
-        for (const call of group) {
-          this.ui.toolCall(describe(call.name, call.args));
-          noted(call);
-        }
-        this.ui.startSpinner(`${group.length} lookups at once`);
-
-        const settled = await Promise.all(
-          group.map((call) => (this.early.get(call.id) ?? this.execute(call)).then((r) => ({ call, ...r })))
-        );
-
-        this.ui.stopSpinner();
-        for (const { call, out, err } of settled) {
-          if (err) badArgs = this.reportFailure(call, err) || badArgs;
-          else this.reportResult(call, out);
-        }
-        continue;
-      }
-
-      for (const call of group) {
-        if (this.abort.signal.aborted) return badArgs;
-
-        const label = describe(call.name, call.args);
-        if (!SILENT.has(call.name)) this.ui.toolCall(label);
-        this.ui.startSpinner(label);
-        noted(call);
-
-        const { out, err } = await (this.early.get(call.id) ?? this.execute(call));
-        this.ui.stopSpinner();
-        if (err) badArgs = this.reportFailure(call, err) || badArgs;
-        else this.reportResult(call, out);
-      }
-    }
-
-    return badArgs;
-  }
-
-  /**
-   * Check a finished call against the stuck patterns (see stuck.js) and return
-   * the note to add to its result. A nudge that did not work hands the turn to
-   * another model.
-   */
-  stuckNote(call, outcome) {
-    if (!this.stuck) return '';
-    const verdict = this.stuck.observe(eventFor(call, outcome));
-    if (!verdict) return '';
-    this.stats.stuck++;
-    if (verdict.action === 'switch') {
-      const next = fallbackFor(model(), this.tried ?? new Set([model()]));
-      if (next) {
-        this.tried?.add(next);
-        this.ui.note(`${modelName(model())} kept ${describeHit(verdict.hit)} — handing over to ${modelName(next)}`);
-        setModel(next);
-        this.cooldownUntil = Date.now() + COOLDOWN;
-        if (this.full) this.showHeader({ clear: false });
-      }
-    }
-    return verdict.text ? `\n\n${verdict.text}` : '';
-  }
-
-  /**
-   * The daily free cap was hit mid-turn: keep the session, count down to the
-   * reset, and carry on by itself. Esc stops the wait like any other turn.
-   */
-  async waitForReset(err) {
-    const resetAt = err.detail?.resetAt;
-    if (!Number.isFinite(resetAt)) return false;
-    const at = resetAt + 30_000;
-    const clock = new Date(at).toLocaleTimeString([], { hour: '2-digit', minute: '2-digit' });
-    this.ui.stopSpinner();
-    this.ui.note(`Daily limit reached — ucode carries on by itself at ${clock}. Esc to stop`);
-    await this.persist();
-    let lastNote = Date.now();
-    while (Date.now() < at) {
-      if (this.abort?.signal.aborted) return false;
-      const left = formatDuration(at - Date.now());
-      if (this.full) this.ui.startSpinner(`daily limit · carrying on at ${clock} (in ${left})`);
-      else if (Date.now() - lastNote > 30 * 60_000) { this.ui.note(`still waiting for the daily limit — ${left} to go`); lastNote = Date.now(); }
-      await new Promise((r) => setTimeout(r, Math.min(1000, Math.max(0, at - Date.now()))));
-    }
-    this.ui.stopSpinner();
-    this.ui.note('Daily limit has reset — carrying on');
-    return true;
-  }
-
-  /** A dev server came up during this turn: open it in the browser, once. */
-  /**
-   * Open the running app in a browser — only when asked.
-   *
-   * This used to happen on its own whenever a dev server came up. Something
-   * seizing the screen mid-thought is startling at the best of times, and
-   * during a demo it is worse. UCODE_OPEN=1 brings the old behaviour back for
-   * anyone who liked it; otherwise the URL is on screen to click.
-   */
-  openWhenReady(since) {
-    if (!this.full || process.env.UCODE_OPEN !== '1') return;
-    const server = serversReadySince(since).at(-1);
-    if (!server || (this.opened ??= new Set()).has(server.url)) return;
-    this.opened.add(server.url);
-    const [cmd, args] = process.platform === 'win32'
-      ? ['cmd', ['/c', 'start', '', server.url]]
-      : [process.platform === 'darwin' ? 'open' : 'xdg-open', [server.url]];
-    try {
-      spawn(cmd, args, { stdio: 'ignore', detached: true, windowsHide: true }).unref();
-      this.ui.note(`Opened ${server.url} in your browser`);
-    } catch { /* no browser to open — the link is in the answer */ }
-  }
-
-  cmdStats() {
-    for (const line of statsLines(this.stats, this.working?.length ?? 0)) this.ui.write(line);
-  }
-
-  async cmdDoctor() {
-    this.ui.startSpinner('checking your setup');
-    try {
-      const lines = await runDoctor();
-      this.ui.stopSpinner();
-      for (const line of lines) this.ui.write(line);
-    } finally {
-      this.ui.stopSpinner();
-    }
-  }
-
-  /** /deploy [folder] — the same tool the model calls, run directly. */
-  async cmdDeploy(arg) {
-    const folder = (arg ?? '').trim() || '.';
-    this.ui.toolCall(`Deploying ${folder}`);
-    this.ui.startSpinner('getting ready to deploy');
-    try {
-      const out = await deploy({ folder }, { onOutput: (lines) => this.ui.updateSpinner(lines.at(-1)) });
-      this.ui.stopSpinner();
-      this.ui.toolResult(out.summary);
-      this.ui.write(out.content.split('\n').map((l) => `  ${l}`).join('\n'));
-    } catch (err) {
-      this.ui.stopSpinner();
-      if (err instanceof ToolFailure) this.ui.error(err);
-      else throw err;
-    }
-  }
-
-  reportResult(call, out) {
-    // A build or type check that just passed already verified everything
-    // changed so far; the automatic check at the end would only repeat it.
-    if (call.name === 'run_command' && out.exitCode === 0 &&
-        /(?:next build|npm run build|pnpm (?:run )?build|tsc)/.test(call.args?.command ?? '')) {
-      this.sinceCheck?.clear();
-    }
-    if (!QUIET.has(call.name)) this.ui.toolResult(out.summary);
-    // The change as its two numbers, not as a copy of the file. The diff rows
-    // are still built by the tool — the model reads them in the result — they
-    // simply do not go on screen.
-    if (out.diff?.length) this.ui.diffStat?.(countDiff(out.diff));
-    this.push({ role: 'tool', toolCallId: call.id, name: call.name, content: out.content + this.stuckNote(call, { out }) });
-  }
-
-  /** Show a tool failure, hand it to the model, and say if it was bad arguments. */
-  reportFailure(call, err) {
-    if (!(err instanceof ToolFailure)) throw err;
-
-    // Bad arguments are the model talking to itself. "old_string and new_string
-    // are identical" is a correction it will make on the next step, and it means
-    // nothing to whoever is watching except that something went wrong. It goes
-    // to the model, which can act on it, and not to the screen. A refusal the
-    // user made, and anything that actually failed, still shows.
-    if (err instanceof Declined) this.ui.toolFailed('declined');
-    else if (err.kind !== 'bad_args') this.ui.toolFailed(`${err.kind}: ${err.failed}`);
-    this.push({
-      role: 'tool',
-      toolCallId: call.id,
-      name: call.name,
-      content: err.forModel() + (err instanceof Declined ? '' : this.stuckNote(call, { err })),
-    });
-    return err.kind === 'bad_args';
-  }
-
-  async dispatch(call) {
-    // The model emitted arguments that were not valid JSON. Hand the parser's
-    // own complaint straight back so it can correct itself next step.
-    if (call.parseError) {
-      throw new ToolFailure({
-        kind: 'bad_args',
-        attempted: `calling ${call.name}`,
-        failed: `The arguments were not valid JSON: ${call.parseError}`,
-        fix: `Call ${call.name} again with the arguments as one well-formed JSON object.`,
-      });
-    }
-
-    if (call.name === 'load_skill') return this.loadSkill(call.args?.name);
-    if (call.name === 'update_plan') return this.updatePlan(call.args?.items);
-    if (call.name === 'delegate') return this.delegate(call.args?.tasks);
-
-    // Output reaches the screen as the command produces it, so a slow build is
-    // something you watch rather than something you sit out in silence.
-    return runTool(call.name, call.args ?? {}, {
-      onOutput: (lines) => this.ui.progress(lines),
-    });
-  }
-
-  /**
-   * Switch to the next model after a provider failure. Returns false once
-   * there is nothing sensible left to try. When every model is busy at once,
-   * it waits a minute and goes round again rather than giving up.
-   */
-  async failover(err) {
-    if (++this.failovers > MAX_FAILOVERS) return false;
-    const from = model();
-    let next = fallbackFor(from, this.tried);
-
-    if (!next) {
-      const until = Date.now() + 60_000;
-      this.ui.startSpinner('every model is busy');
-      while (Date.now() < until && !this.abort?.signal.aborted) {
-        this.ui.updateSpinner(`every model is busy — trying again in ${Math.ceil((until - Date.now()) / 1000)}s`);
-        await wait(1000);
-      }
-      this.ui.stopSpinner();
-      if (this.abort?.signal.aborted) return false;
-      this.tried = new Set();
-      next = this.preferred && this.preferred !== from ? this.preferred : fallbackFor(from, this.tried) ?? from;
-    }
-
-    this.tried.add(next);
-    setModel(next);
-    this.cooldownUntil = Date.now() + COOLDOWN;
-    const why = err.kind === 'rate_limit' ? 'busy' : err.kind === 'timeout' ? 'too slow to answer' : 'not answering';
-    this.ui.note(`${modelName(from)} is ${why} — carrying on with ${modelName(next)}`);
-    if (this.full) this.showHeader({ clear: false });
-    return true;
-  }
-
-  /** Run a call and settle to { out } or { err } — never throws. */
-  execute(call) {
-    const started = Date.now();
-    return this.dispatch(call).then(
-      (out) => { trace({ kind: 'tool', name: call.name, ms: Date.now() - started }); countTool(this.stats, call, out); return { out }; },
-      (err) => { trace({ kind: 'tool', name: call.name, ms: Date.now() - started, err: err?.kind }); countTool(this.stats, call, null, err); return { err }; }
-    );
-  }
-
-  updatePlan(items) {
-    const list = (Array.isArray(items) ? items : [])
-      .filter((i) => i && String(i.text ?? '').trim())
-      .slice(0, 6);
-    this.ui.plan(list);
-    const done = list.filter((i) => i.done).length;
-    return { content: `Plan updated: ${done} of ${list.length} done.`, summary: `${done}/${list.length}` };
-  }
-
-  /** File writes from parallel workers take turns, so two never interleave. */
-  fileLock(fn) {
-    const run = (this.lockChain ?? Promise.resolve()).then(fn, fn);
-    this.lockChain = run.catch(() => {});
-    return run;
-  }
-
-  /**
-   * Several workers at once, each its own small agent loop with its own
-   * conversation, sharing the tools, the project, and whatever skills are
-   * already in force. Their lines in the transcript carry their name.
-   */
-  async delegate(tasks) {
-    const list = (Array.isArray(tasks) ? tasks : [])
-      .filter((t) => t && String(t.instructions ?? '').trim())
-      .slice(0, MAX_WORKERS);
-    if (!list.length) {
-      throw new ToolFailure({
-        kind: 'bad_args',
-        attempted: 'starting workers',
-        failed: 'No tasks with instructions were given.',
-        fix: 'Pass tasks as [{ name, instructions }, ...], up to 3.',
-      });
-    }
-
-    const results = await Promise.all(list.map((task, i) => this.runWorker(task, i).catch((err) => ({
-      name: task.name || `worker ${i + 1}`,
-      summary: `Failed: ${err?.failed ?? err?.message ?? err}`,
-      touched: [],
-    }))));
-
-    for (const r of results) for (const f of r.touched) { this.touched.add(f); this.sinceCheck.add(f); }
-
-    // A worker that wrote nothing has not done its part, whatever it said.
-    // The lead builds those itself rather than leaving holes in the app.
-    const empty = results.filter((r) => !r.touched.length).map((r) => r.name);
-
-    return {
-      content: results
-        .map((r) => `## ${r.name}\n${r.summary}\nFiles changed: ${r.touched.join(', ') || 'none'}`)
-        .join('\n\n') +
-        (empty.length
-          ? `\n\n${empty.join(', ')} wrote no files. Build ${empty.length === 1 ? 'that part' : 'those parts'} ` +
-            'yourself now, directly - do not delegate them again.'
-          : ''),
-      summary: results.map((r) => `${r.name} · ${r.touched.length} file${r.touched.length === 1 ? '' : 's'}`).join('  '),
-    };
-  }
-
-  async runWorker(task, index) {
-    const name = clip(String(task.name || `worker ${index + 1}`).trim(), 16);
-    const touched = new Set();
-    const skills = this.skills
-      .filter((s) => this.loaded.has(s.name))
-      .map((s) => `--- ${s.name} ---\n${s.body}`)
-      .join('\n\n');
-    const messages = [
-      { role: 'system', content: workerPrompt({ cwd: this.cwd, name, memory: this.memory, skills, map: this.map }) },
-      { role: 'user', content: String(task.instructions) },
-    ];
-    const available = this.toolsNow().filter((t) => !WORKER_EXCLUDED.has(t.name));
-    const wanted = process.env.UCODE_WORKER_MODEL;
-    let workerModel = wanted && MODELS[wanted] ? wanted : model();
-    const tried = new Set([workerModel]);
-    let failovers = 0;
-
-    // Start a moment apart. Three requests in the same instant is exactly what
-    // trips a free endpoint's rate limit, and the stagger costs a second or two.
-    if (index) await wait(index * 1500);
-
-    for (let step = 0; step < WORKER_STEPS; step++) {
-      if (this.abort?.signal.aborted) break;
-
-      let reply;
-      const asked = Date.now();
-      try {
-        reply = await ask(messages, available, { signal: this.abort?.signal, model: workerModel });
-      } catch (err) {
-        // Same rule as the lead: a busy model is swapped, not a reason to stop.
-        if (passing(err) && failovers < 6 && !this.abort?.signal.aborted) {
-          failovers++;
-          let next = fallbackFor(workerModel, tried);
-          if (!next) { tried.clear(); await wait(20_000); next = fallbackFor(workerModel, tried) ?? workerModel; }
-          tried.add(next);
-          this.ui.note(`${name}: ${modelName(workerModel)} is busy — switching to ${modelName(next)}`);
-          workerModel = next;
-          step--;
-          continue;
-        }
-        throw err;
-      }
-      this.record(reply.usage);
-      trace({
-        kind: 'model', who: name, model: workerModel, ms: Date.now() - asked,
-        in: reply.usage.promptTokens, out: reply.usage.outputTokens,
-        calls: reply.toolCalls.map((c) => c.name),
-      });
-
-      if (!reply.toolCalls.length) {
-        this.ui.toolResult(`${name} finished`);
-        return { name, summary: reply.text?.trim() || 'Finished without a summary.', touched: [...touched] };
-      }
-
-      messages.push({ role: 'assistant', content: reply.text || '', toolCalls: reply.toolCalls });
-      for (const call of reply.toolCalls) {
-        this.ui.toolCall(`${name} › ${describe(call.name, call.args)}`);
-        if (FILE_WRITES.has(call.name)) for (const p of pathsOf(call)) touched.add(p);
-        const { out, err } = FILE_WRITES.has(call.name)
-          ? await this.fileLock(() => this.execute(call))
-          : await this.execute(call);
-        if (err) {
-          if (!(err instanceof ToolFailure)) throw err;
-          this.ui.toolFailed(`${name}: ${err.kind}: ${err.failed}`);
-        }
-        messages.push({
-          role: 'tool', toolCallId: call.id, name: call.name,
-          content: err ? err.forModel() : out.content,
-        });
-      }
-    }
-
-    return { name, summary: `Stopped after ${WORKER_STEPS} steps without finishing.`, touched: [...touched] };
-  }
-
-  /**
-   * Check the code files changed since the last check, and return the
-   * errors as text for the model — or null when everything is clean.
-   *
-   * The check is incremental: TypeScript writes what it learned to a build
-   * info file, so the second check onwards reads that instead of retyping
-   * every dependency — seconds rather than the best part of a minute. The
-   * file sits in node_modules/.cache, which is already ignored by git and is
-   * deliberately left out of the starter package cache.
-   *
-   * TypeScript projects get one `tsc --noEmit` per project that owns a
-   * changed file (an app scaffolded into a subfolder is its own project).
-   * Plain JavaScript gets a syntax check, Python a compile check. Nothing
-   * runs that is not already installed.
-   */
-  async autoCheck() {
-    const changed = [...this.sinceCheck].filter((f) => CHECKABLE.test(f));
-    this.sinceCheck.clear();
-    if (!changed.length) return null;
-
-    const root = path.resolve(this.cwd);
-    const tsRoots = new Set();
-    const singles = [];
-
-    for (const rel of changed) {
-      const abs = path.resolve(root, rel);
-      if (!(await exists(abs))) continue;
-      if (/\.py$/i.test(rel)) { singles.push({ abs, rel, command: `python -m py_compile "${abs}"` }); continue; }
-      let dir = path.dirname(abs);
-      let owner = null;
-      while (dir.startsWith(root)) {
-        if (await exists(path.join(dir, 'tsconfig.json'))) { owner = dir; break; }
-        const up = path.dirname(dir);
-        if (up === dir) break;
-        dir = up;
-      }
-      if (owner && (await exists(path.join(owner, 'node_modules', 'typescript')))) tsRoots.add(owner);
-      else if (/\.[cm]?js$/i.test(rel)) singles.push({ abs, rel, command: `node --check "${abs}"` });
-    }
-
-    const problems = [];
-    const check = async (label, command, cwd) => {
-      this.ui.toolCall(label);
-      this.ui.startSpinner(label);
-      const { out, err } = await this.execute({
-        id: 'check', name: 'run_command',
-        args: { command, cwd: path.relative(root, cwd) || '.', timeout_ms: 180_000 },
-      });
-      this.ui.stopSpinner();
-      return err ? { exitCode: -1, content: String(err.failed ?? err.message) } : out;
-    };
-
-    for (const dir of tsRoots) {
-      const show = path.relative(root, dir) || '.';
-      await mkdir(path.join(dir, path.dirname(TSBUILDINFO)), { recursive: true }).catch(() => {});
-      let out = await check(`Checking types in ${show}`, typeCheckCommand(true), dir);
-      // Older TypeScript refuses --incremental alongside --noEmit. Say so once
-      // by simply checking again the slow way, rather than failing the edit.
-      if (out.exitCode !== 0 && NO_INCREMENTAL.test(out.content)) {
-        out = await check(`Checking types in ${show}`, typeCheckCommand(false), dir);
-      }
-      if (out.exitCode === 0) { this.ui.toolResult('types check out'); continue; }
-      const errors = out.content.split('\n').filter((l) => /error TS\d+/.test(l));
-      this.ui.toolFailed(`${errors.length || 'some'} type error${errors.length === 1 ? '' : 's'}`);
-      problems.push(`In ${show} (tsc --noEmit):\n${(errors.length ? errors : out.content.split('\n')).slice(0, 40).join('\n')}`);
-    }
-
-    for (const f of singles) {
-      const out = await check(`Checking ${f.rel}`, f.command, root);
-      if (out.exitCode === 0) { this.ui.toolResult('ok'); continue; }
-      this.ui.toolFailed('does not compile');
-      problems.push(`${f.rel}:\n${out.content.split('\n').slice(0, 20).join('\n')}`);
-    }
-
-    // Only once it compiles: a failing test on code that does not build tells
-    // the model nothing it does not already know from the errors above.
-    if (!problems.length) {
-      const failed = await this.runRelatedTests(root, changed);
-      if (failed) problems.push(failed);
-    }
-
-    const live = await this.liveErrors();
-    if (live) problems.push(live);
-
-    return problems.length ? problems.join('\n\n') : null;
-  }
-
-  /**
-   * Anything the running app has complained about since the last look. A dev
-   * server knows about a broken import the moment it happens; without this
-   * nobody reads that until a build, or until the user says the page is blank.
-   */
-  async liveErrors() {
-    try {
-      const found = await this.logWatch.since(runningServers());
-      if (found) this.ui.toolFailed('the running app reported an error');
-      return found;
-    } catch {
-      return null; // reading a log must never be what breaks a turn
-    }
-  }
-
-  /**
-   * Run the tests that reach the files just changed, and return their
-   * failures as text — or null when they pass, or when this project has no
-   * runner that can be asked which tests matter.
-   */
-  async runRelatedTests(root, changed) {
-    const runner = await testRunnerFor(root);
-    if (!runner) return null;
-
-    const existing = [];
-    for (const rel of changed) if (await exists(path.resolve(root, rel))) existing.push(rel);
-    const command = relatedCommand(runner, existing);
-    if (!command) return null;
-
-    const label = `Running the ${runner} tests that cover this`;
-    this.ui.toolCall(label);
-    this.ui.startSpinner(label);
-    const { out, err } = await this.execute({
-      id: 'tests', name: 'run_command',
-      args: { command, cwd: '.', timeout_ms: 180_000 },
-    });
-    this.ui.stopSpinner();
-
-    if (err) { this.ui.toolResult('tests skipped'); return null; }
-    if (out.exitCode === 0) { this.ui.toolResult('tests pass'); return null; }
-
-    // A runner that is not installed is not a failing test; npx says so.
-    if (/could not determine executable|not found|Cannot find module/i.test(out.content)) {
-      this.ui.toolResult('no test runner installed');
-      return null;
-    }
-
-    this.ui.toolFailed('tests fail');
-    return `The tests covering your change fail (${runner}):\n${summariseFailures(runner, out.content)}`;
-  }
-
-  loadSkill(name) {
-    const skill = findSkill(this.skills, name);
-    if (!skill) {
-      // The available names go in `failed` rather than only in `fix`: the
-      // transcript shows the failure line, and a bare "no such skill" leaves
-      // the user guessing at what this session actually has.
-      const available = this.skills.map((s) => s.name).join(', ') || '(none)';
-      throw new ToolFailure({
-        kind: 'no_such_skill',
-        attempted: `loading the "${name}" skill`,
-        failed: `There is no skill called "${name}". This session has: ${available}.`,
-        fix: 'Use one of those names, or carry on without one.',
-      });
-    }
-
-    if (this.loaded.has(skill.name)) {
-      return { content: `The "${skill.name}" skill is already loaded above. Follow it.`, summary: 'already loaded' };
-    }
-
-    this.loaded.add(skill.name);
-    this.push(skillMessage(skill));
-    return {
-      content: `Loaded "${skill.name}". Its instructions are in your context now — follow them.`,
-      summary: `${skill.name} · ${skill.body.split('\n').length} lines`,
-    };
-  }
-
-  record(u) {
-    const total = this.session.usage;
-    total.promptTokens += u.promptTokens || 0;
-    total.outputTokens += u.outputTokens || 0;
-    total.totalTokens += u.totalTokens || 0;
-    total.turns += 1;
-  }
-
-  /** Fold older turns into a summary when the window gets tight. */
-  async maybeFold() {
-    const limit = contextLimit();
-    if (!tooBig(this.working, limit)) return;
-
-    this.ui.startSpinner('context is filling up — summarizing earlier turns');
-    try {
-      const result = await fold(this.working, {
-        limit,
-        summarize: async (older) => {
-          const reply = await ask(
-            [
-              { role: 'system', content: SUMMARY_PROMPT },
-              { role: 'user', content: forSummary(older) },
-            ],
-            [],
-            { signal: this.abort?.signal, temperature: 0 }
-          );
-          return reply.text;
-        },
-      });
-
-      this.ui.stopSpinner();
-      if (result.folded) {
-        this.working = result.messages;
-        this.ui.note(
-          `folded ${result.droppedCount} earlier messages into a summary ` +
-          '(the full history is still saved in this session)'
-        );
-      }
-    } catch (err) {
-      // If summarizing fails, carry on with the full history and let the API
-      // complain — better than silently throwing away the conversation.
-      this.ui.stopSpinner();
-      this.ui.note(`could not summarize older turns (${err.kind ?? 'error'}); carrying on uncompacted`);
-    }
-  }
-
-  // -- slash commands ------------------------------------------------------
-
-  async command(input) {
-    const [name, ...rest] = input.split(/\s+/);
-    const arg = rest.join(' ').trim();
-
-    switch (name) {
-      case '/help':
-        return this.cmdHelp();
-
-      // One thing, one command, however you happen to spell it.
-      case '/model':
-      case '/models':
-        return this.cmdModel(arg);
-
-      case '/session':
-      case '/sessions':
-      case '/resume':
-        return this.cmdSessions(arg);
-
-      case '/new':      return this.cmdNew();
-      case '/remember': return this.cmdRemember(arg);
-      case '/skills':   return this.cmdSkills();
-      case '/clear':    this.showHeader(); return;
-      case '/search':   return this.cmdSearch(arg);
-      case '/copy':     return this.cmdCopy();
-      case '/stats':    return this.cmdStats();
-      case '/doctor':   return this.cmdDoctor();
-      case '/look':     return this.cmdLook(arg);
-      case '/deploy':   return this.cmdDeploy(arg);
-      case '/exit':
-      case '/quit':     return 'exit';
-
-      default:
-        this.ui.write(theme.warn(`  no such command: ${name}`));
-        this.ui.note('/help lists them.');
-    }
-  }
-
-  /**
-   * Look at the running app, because the user asked to.
-   *
-   * This used to happen on its own, which meant a browser being driven while
-   * someone was reading, and a window taking the screen mid-thought. It is
-   * the same check as before; the difference is who starts it.
-   */
-  async cmdLook(url) {
-    const { lookAtApp } = await import('../tools/browser.js');
-    const server = runningServers().at(-1);
-    const at = (url ?? '').trim() || server?.url;
-    if (!at) {
-      this.ui.write(theme.warn('  nothing is running to look at.'));
-      this.ui.note('start the app first, or pass a URL: /look http://localhost:3000');
-      return;
-    }
-    this.ui.toolCall(`Looking at ${at}`);
-    try {
-      const out = await lookAtApp({ url: at });
-      this.ui.write(out.content);
-      // The model gets it too, so the next thing it says is about what is
-      // actually on the page rather than what it believes it built.
-      this.push({ role: 'user', content: `I looked at ${at}. This is what is there:
-
-${out.content}` });
-    } catch (err) {
-      this.ui.write(theme.error(`  ${err.failed ?? err.message}`));
-    }
-  }
-
-  cmdHelp() {
-    const rows = [
-      ['/help', 'this list'],
-      ['/stats', 'time, steps and tokens this session'],
-      ['/doctor', 'check that everything ucode needs is working'],
-      ['/look [url]', 'open the running app and report what is on the page'],
-      ['/deploy [folder]', 'put the app online and get its link'],
-      ['/model', 'show the models and switch between them'],
-      ['/resume', 'pick up an earlier conversation'],
-      ['/new', 'save this one and start fresh'],
-      ['/remember <note>', `add a standing note to ${MEMORY_FILE}`],
-      ['/skills', 'what ucode knows how to do'],
-      ['/search <query>', 'look something up on the web'],
-      ['/copy', 'copy the last reply to the clipboard'],
-      ['/clear', 'clear the screen, keep the conversation'],
-      ['/exit', 'save and quit'],
-    ];
-
-    this.ui.blank();
-    for (const [command, what] of rows) {
-      this.ui.write(`  ${blue(command.padEnd(18))} ${dim(what)}`);
-    }
-    this.ui.blank();
-    this.ui.write(dim('  /models, /session and /sessions do the same as /model and /resume.'));
-    this.ui.write(dim('  ctrl+b swaps plan and build · esc stops a running turn · ctrl+d quits'));
-    this.ui.blank();
-  }
-
-  /** The five models, and this session's spend. */
-  async cmdModel(arg) {
-    if (arg) {
-      try {
-        setModel(arg);
-        this.preferred = model();
-      } catch (err) {
-        this.ui.error(err, { debug: this.debug });
-        return;
-      }
-      this.session.model = model();
-      this.ui.note(`now using ${modelName()}`);
-      this.showHeader({ clear: false });
-      return;
-    }
-
-    const all = modelList();
-    const width = Math.max(...all.map((m) => m.name.length));
-
-    if (this.ui.pick) {
-      const items = all.map((m) => ({
-        label:
-          `${m.active ? blue('●') : dim('○')} ${m.star ? blue('★') : ' '} ` +
-          `${m.name.padEnd(width)}  ${dim(`${formatTokens(m.context)} · ${m.note}`)}`,
-      }));
-
-      const chosen = await this.ui.pick(items, {
-        active: Math.max(0, all.findIndex((m) => m.active)),
-        hint: '↑↓ move · enter to switch · esc to cancel',
-      });
-      if (chosen === null) return;
-
-      setModel(all[chosen].id);
-      this.preferred = model();
-      this.session.model = model();
-      this.ui.note(`now using ${modelName()}`);
-      this.showHeader({ clear: false });
-      return;
-    }
-
-    this.ui.blank();
-    for (const m of all) {
-      this.ui.write(
-        `  ${m.active ? blue('●') : dim('○')} ${m.star ? blue('★') : ' '} ` +
-        `${(m.active ? blue : dim)(m.name.padEnd(width))}  ${dim(`${formatTokens(m.context)} · ${m.note}`)}`
-      );
-      this.ui.write(`      ${dim(m.id)}`);
-    }
-
-    const u = this.session.usage;
-    const live = rateLimits();
-    this.ui.blank();
-    this.ui.write(
-      `  ${dim('this session')}  ${u.turns} turns · ${formatTokens(u.totalTokens)} tokens ` +
-      `(${formatTokens(u.promptTokens)} in, ${formatTokens(u.outputTokens)} out)`
-    );
-    if (live?.requestsRemaining != null && live?.requestsLimit) {
-      this.ui.write(`  ${dim('requests')}      ${live.requestsRemaining} of ${live.requestsLimit} left`);
-    }
-    this.ui.blank();
-    this.ui.write(dim('  /model <id> switches without the picker.'));
-    this.ui.blank();
-  }
-
-  /**
-   * One row per saved conversation.
-   *
-   * A list of titles and timestamps is not enough to recognise your own work
-   * by — half of them start "Fix the". So each row carries what it was about
-   * and how far it got, and the ones from this folder are marked, because that
-   * is nearly always the one being looked for.
-   */
-  describeSession(s, width, i) {
-    const room = Math.max(24, Math.min(46, width - 34));
-    const mark = s.mine ? blue('●') : dim('○');
-    const when = relativeTime(s.updatedAt).padEnd(9);
-    const turns = `${s.turns} turn${s.turns === 1 ? '' : 's'}`.padEnd(9);
-    const where = s.mine ? 'here' : shortenPath(s.cwd, 26);
-
-    return {
-      // Numbered in the picker, so /session delete 3 has something to point at.
-      label: `${i === undefined ? '' : `${dim(String(i + 1).padStart(2))} `}${mark} ${clip(s.title, room).padEnd(room)}  ${dim(when)}${dim(turns)}${dim(where)}`,
-      sub: s.preview ? dim(`     ${clip(s.preview, width - 10)}`) : '',
-    };
-  }
-
-  async cmdSessions(arg) {
-    if (arg === '--clear' || arg === 'clear') {
-      const yes = await this.ui.confirm({
-        action: 'delete every saved conversation',
-        detail: 'This cannot be undone.',
-        risk: 'write',
-      });
-      if (!yes) {
-        this.ui.note('cancelled');
-        return;
-      }
-      await removeAll();
-      this.ui.note('all sessions deleted');
-      return;
-    }
-
-    // /session delete 3   or   /session delete 2,5,7
-    const del = /^(?:delete|del|rm|remove)\b\s*(.*)$/i.exec(arg ?? '');
-    if (del) return this.deleteSessions(del[1]);
-
-    const sessions = await list({ cwd: this.cwd });
-    if (!sessions.length) {
-      this.ui.note('no saved conversations yet');
-      return;
-    }
-
-    for (const bad of sessions.unreadable ?? []) {
-      this.ui.write(theme.warn(`  could not read session file: ${bad}`));
-    }
-
-    const shown = sessions.slice(0, 25);
-    const width = this.ui.width ? this.ui.width() : 80;
-    let index;
-
-    if (arg) {
-      const n = Number(arg);
-      if (!Number.isInteger(n) || n < 1 || n > shown.length) {
-        this.ui.write(theme.warn(`  "${arg}" is not one of 1-${shown.length}`));
-        return;
-      }
-      index = n - 1;
-    } else if (this.ui.pick) {
-      // The picker stays open while you delete, so clearing out several old
-      // conversations is d d, d d, d d — then Enter on the one you want.
-      let active = 0;
-      for (;;) {
-        const here = shown.filter((s) => s.mine).length;
-        const picked = await this.ui.pick(
-          shown.map((s, i) => this.describeSession(s, width, i)),
-          {
-            active,
-            deletable: true,
-            hint:
-              `↑↓ move · enter to continue · d twice to delete · esc to cancel` +
-              (here ? ` — ${here} from this folder` : ''),
-          }
-        );
-        if (picked === null) return;
-        if (typeof picked === 'object' && picked.delete !== undefined) {
-          const doomed = shown[picked.delete];
-          active = picked.delete;
-          if (doomed.id === this.session.id) {
-            this.ui.flash?.('that is the conversation you are in — /new first, then delete it');
-            continue;
-          }
-          await remove(doomed.id);
-          shown.splice(picked.delete, 1);
-          this.ui.flash?.(`deleted · ${clip(doomed.title, 50)}`);
-          if (!shown.length) {
-            this.ui.note('no saved conversations left');
-            return;
-          }
-          active = Math.min(active, shown.length - 1);
-          continue;
-        }
-        index = picked;
-        break;
-      }
-    } else {
-      this.ui.blank();
-      this.ui.note('/session delete <number> removes one, or several: /session delete 2,5');
-      index = await this.ui.choose(
-        'continue which?',
-        shown.map((s, i) => this.describeSession(s, width, i).label)
-      );
-      if (index === null) return;
-    }
-
-    if (this.session.messages.length) await this.persist();
-    if (await this.resume(shown[index].id)) {
-      this.showHeader();
-      this.replayTail();
-    }
-  }
-
-  /** /session delete 3, or 2,5,7 — numbers as the session list shows them. */
-  async deleteSessions(spec) {
-    const sessions = (await list({ cwd: this.cwd })).slice(0, 25);
-    const numbers = [...new Set(String(spec).split(/[\s,]+/).filter(Boolean).map(Number))];
-    const bad = numbers.filter((n) => !Number.isInteger(n) || n < 1 || n > sessions.length);
-    if (!numbers.length || bad.length) {
-      this.ui.write(theme.warn(`  usage: /session delete <number>[,<number>…] — numbers from 1 to ${sessions.length}`));
-      return;
-    }
-    for (const n of numbers) {
-      const s = sessions[n - 1];
-      if (s.id === this.session.id) {
-        this.ui.note(`skipped ${n} — that is the conversation you are in`);
-        continue;
-      }
-      await remove(s.id);
-      this.ui.note(`deleted ${n} · ${s.title}`);
-    }
-  }
-
-  async resume(id) {
-    try {
-      const loaded = await load(id);
-      this.session = loaded;
-      this.working = [...loaded.messages];
-      this.loaded = new Set(loaded.messages.filter((m) => m.skill).map((m) => m.skill));
-      if (loaded.model && MODELS[loaded.model]) setModel(loaded.model);
-      return true;
-    } catch (err) {
-      this.ui.error(err, { debug: this.debug });
-      this.ui.note('Starting a fresh one instead.');
-      return false;
-    }
-  }
-
-  /** The last few exchanges, so a resumed conversation has visible context. */
-  replayTail(count = 4) {
-    const tail = this.session.messages
-      .filter((m) => (m.role === 'user' || m.role === 'assistant') && m.content)
-      .slice(-count);
-
-    for (const m of tail) {
-      if (m.role !== 'user') this.ui.assistant(m.content);
-      else if (this.ui.userMessage) this.ui.userMessage(m.content);
-      else this.ui.write(`${blue('›')} ${dim(m.content.split('\n')[0])}`);
-    }
-    if (tail.length) this.ui.write(dim('  ── picking up here ──\n'));
-  }
-
-  /** Add a line to this project's UCODE.md, read at the start of every turn. */
-  async cmdRemember(note) {
-    if (!note) {
-      this.ui.note(`usage: /remember <something ucode should always know here> — saved to ${MEMORY_FILE}`);
-      return;
-    }
-    try {
-      const file = await remember(this.cwd, note);
-      this.ui.note(`remembered · ${path.relative(this.cwd, file) || MEMORY_FILE}`);
-    } catch (err) {
-      this.ui.error(new Failure({
-        kind: 'memory_unwritable',
-        attempted: `saving to ${MEMORY_FILE}`,
-        failed: err.message,
-        fix: 'Check that this folder is writable.',
-      }), { debug: this.debug });
-    }
-  }
-
-  async cmdNew() {
-    if (this.session.messages.length) {
-      await this.persist();
-      this.ui.note(`saved · ${this.session.title}`);
-    }
-    this.session = newSession(this.cwd, model());
-    this.working = [];
-    this.loaded = new Set();
-    this.showHeader();
-  }
-
-  async cmdSkills() {
-    // Re-read from disk. Skills load once at startup, so one written during
-    // this session would otherwise stay invisible — and load_skill would fail
-    // on a name the user can see in the folder.
-    this.skills = await loadSkills({ cwd: this.cwd });
-    for (const problem of this.skills.problems ?? []) {
-      this.ui.write(theme.warn(`  skill not loaded: ${problem}`));
-    }
-
-    if (!this.skills.length) {
-      this.ui.note('no skills found — add a folder with a SKILL.md under .ucode/skills');
-      return;
-    }
-
-    this.ui.blank();
-    for (const s of this.skills) {
-      const live = this.loaded.has(s.name);
-      const auto = s.triggers.length ? dim('  · loads itself') : '';
-      this.ui.write(`  ${live ? blue('●') : dim('○')} ${blue(s.name)}${auto}`);
-      this.ui.write(`    ${dim(s.description)}`);
-    }
-    this.ui.blank();
-    this.ui.write(dim('  ● already loaded here · ucode pulls one in when the task matches'));
-    this.ui.blank();
-  }
-
-  async cmdSearch(query) {
-    if (!query) {
-      this.ui.note('usage: /search <what you want to look up>');
-      return;
-    }
-    await this.turn(
-      `Search the web for: ${query}\n\nUse web_search, then summarise what you found and cite the URLs.`
-    );
-  }
-
-  /** Copy the last reply. Every platform ships a clipboard pipe. */
-  async cmdCopy() {
-    const last = [...this.session.messages]
-      .reverse()
-      .find((m) => m.role === 'assistant' && m.content?.trim());
-
-    if (!last) {
-      this.ui.note('nothing to copy yet');
-      return;
-    }
-
-    const tool = process.platform === 'win32' ? 'clip'
-      : process.platform === 'darwin' ? 'pbcopy'
-      : 'xclip -selection clipboard';
-
-    try {
-      await new Promise((resolve, reject) => {
-        const child = spawn(tool, { shell: true, windowsHide: true });
-        child.on('error', reject);
-        child.on('close', (code) => (code === 0 ? resolve() : reject(new Error(`exit ${code}`))));
-        child.stdin.end(last.content);
-      });
-      const lines = last.content.split('\n').length;
-      this.ui.note(`copied ${lines} line${lines === 1 ? '' : 's'}`);
-    } catch (err) {
-      this.ui.error(new Failure({
-        kind: 'clipboard_failed',
-        attempted: 'copying the last reply',
-        failed: `${tool} could not run: ${err.message}`,
-        fix: process.platform === 'linux'
-          ? 'Install xclip (apt install xclip), or select the text with the mouse.'
-          : 'Select the text with the mouse instead.',
-      }), { debug: this.debug });
-    }
-  }
-}
-
-export { DEFAULT_MODEL, PROVIDER };
+/**
+ * loop.js — the agent itself.
+ *
+ * One turn, end to end: take a line, fold the conversation if it has grown too
+ * big, ask the model, and if it asked for tools, run them and ask again. Repeat
+ * until it answers with prose instead of a tool call.
+ *
+ * Two histories are kept, deliberately:
+ *   session.messages  the complete record, written to disk after every step
+ *   working           what is actually sent, which may have its older turns
+ *                     folded into a summary once the window gets tight
+ */
+
+import path from 'node:path';
+import os from 'node:os';
+import { appendFileSync } from 'node:fs';
+import { readFile, access, mkdir } from 'node:fs/promises';
+import { testRunnerFor, relatedCommand, summariseFailures } from './tests.js';
+import { LogWatch } from './livelog.js';
+import { checkHtml } from './htmlcheck.js';
+import { runningServers } from '../tools/shell.js';
+import { spawn } from 'node:child_process';
+
+import {
+  ask, model, setModel, modelName, modelList, contextLimit, rateLimits,
+  MODELS, DEFAULT_MODEL, PROVIDER, fallbackFor,
+} from './provider.js';
+import {
+  tools, runTool, describe, setRoot, setConfirm, PARALLEL_SAFE, WRITES, FILE_WRITES,
+} from '../tools/index.js';
+import { projectMap, loadMemory, remember, MEMORY_FILE } from './context.js';
+import { autoUpdate } from './updater.js';
+import { closeBrowser, forgetReviews } from '../tools/browser.js';
+import {
+  newSession, save, load, list, remove, removeAll, titleFrom,
+} from './history.js';
+import { fold, usage, tooBig, SUMMARY_PROMPT, forSummary } from './window.js';
+import { loadSkills, catalogue, findSkill, skillMessage, autoLoadFor } from './skills.js';
+import { Screen, isLabel } from '../ui/screen.js';
+import { Plain } from '../ui/plain.js';
+import { theme, blue, sky, dim, formatTokens, relativeTime, shortenPath, clip } from '../ui/theme.js';
+import { Failure, ToolFailure, Declined } from './failure.js';
+import { StuckWatch, eventFor, describeHit } from './stuck.js';
+import { serversReadySince } from '../tools/shell.js';
+import { formatDuration } from '../ui/activity.js';
+import { runDoctor } from './doctor.js';
+import { deploy } from '../tools/deploy.js';
+
+/**
+ * Tool calls allowed in one turn.
+ *
+ * Scaffolding an app is dozens of writes before anything can even be run, so a
+ * small ceiling stops a real job halfway through — which from the outside is
+ * indistinguishable from the agent giving up for no reason. The runaway-loop
+ * protection this exists for still works at 250; a loop burns through that
+ * just as visibly, only later.
+ */
+const MAX_STEPS = Number(process.env.UCODE_MAX_STEPS) || 250;
+const MAX_ARG_RETRIES = 2;
+
+/**
+ * How many times a reply cut off at the output limit is asked to carry on.
+ * Three covers any answer a terminal should be printing; past that the model
+ * is rambling and stopping is the kinder outcome.
+ */
+const MAX_CONTINUATIONS = 3;
+
+/** Read-only tools whose result line adds nothing — the user saw the output. */
+/**
+ * How many rows a diff adds and removes.
+ *
+ * The rows come through as "+12| text" and "-12| text", with a "~" heading
+ * for each file in a multi-file write and an undecorated note counting what
+ * was elided. Only the signs are counted.
+ */
+export function countDiff(rows = []) {
+  let added = 0;
+  let removed = 0;
+  for (const row of rows) {
+    const line = String(row ?? '');
+    if (line.startsWith('~')) continue;
+    // "… 218 more removed" / "… 508 more added" stand for rows not shown.
+    const more = /^\s*[….]+\s*(\d+)\s+more\s+(added|removed)/.exec(line);
+    if (more) {
+      if (more[2] === 'added') added += Number(more[1]);
+      else removed += Number(more[1]);
+      continue;
+    }
+    if (line.startsWith('+')) added++;
+    else if (line.startsWith('-')) removed++;
+  }
+  return { added, removed };
+}
+
+const QUIET = new Set(['read_file', 'read_files', 'list_dir', 'glob', 'grep', 'web_search', 'update_plan']);
+
+/** Tools that draw their own line, so they get no "● Doing X" line of their own. */
+const SILENT = new Set(['update_plan']);
+
+/** How many rounds of "the type check found errors, fix them" one turn may take. */
+const MAX_FIX_ROUNDS = 3;
+
+/** Files worth checking after they change. */
+const CHECKABLE = /\.(?:[cm]?[jt]sx?|py|html?)$/i;
+
+/** Where TypeScript keeps what it learned, so the next check is a quick one. */
+export const TSBUILDINFO = 'node_modules/.cache/ucode/types.tsbuildinfo';
+
+/** TypeScript before 4.0 rejects --incremental together with --noEmit. */
+const NO_INCREMENTAL = /TS5074|TS6304|'--incremental'/;
+
+/**
+ * The type check to run. Incremental by default: the first check pays the
+ * full cost and writes a build info file, and every one after it reads that
+ * and reports in about a second.
+ */
+export function typeCheckCommand(incremental = true) {
+  const base = 'npx --no-install tsc --noEmit --pretty false';
+  return incremental ? `${base} --incremental --tsBuildInfoFile ${TSBUILDINFO}` : base;
+}
+
+/**
+ * Failures that are the provider's and not the model's: busy, slow, down, or
+ * unreachable. None of them should end a build — the turn moves to another
+ * model and carries on from exactly where it was.
+ */
+const TRANSIENT = new Set(['rate_limit', 'timeout', 'server', 'network', 'no_content']);
+/** Will another model, or a little patience, get past this? Not the daily cap: it covers them all. */
+const passing = (err) => TRANSIENT.has(err.kind) && !err.detail?.daily;
+const MAX_FAILOVERS = 8;
+const COOLDOWN = 5 * 60_000;
+
+const wait = (ms) => new Promise((r) => setTimeout(r, ms));
+
+/**
+ * UCODE_TRACE=1 writes one JSON line per model call and per tool to
+ * ~/.ucode/trace.jsonl (or to the path UCODE_TRACE names): how long it took,
+ * tokens in and out, what was called. It is how "it feels slow" becomes a
+ * number with a cause attached.
+ */
+const TRACE_FILE = process.env.UCODE_TRACE
+  ? (process.env.UCODE_TRACE === '1' ? path.join(os.homedir(), '.ucode', 'trace.jsonl') : process.env.UCODE_TRACE)
+  : null;
+
+/** Tool results worth re-sending in full only while they are recent. */
+const THIN_RESULTS = new Set([
+  'read_file', 'read_files', 'grep', 'glob', 'list_dir', 'run_command', 'run_commands',
+  'look_at_app', 'web_search', 'edit_file', 'multi_edit', 'edit_files',
+]);
+
+/** Replace long strings in old tool arguments with a note of their size. */
+function thinArgs(value) {
+  if (typeof value === 'string') {
+    return value.length > 400
+      ? `[${value.length} characters, already applied — read the file if you need its current text]`
+      : value;
+  }
+  if (Array.isArray(value)) return value.map(thinArgs);
+  if (value && typeof value === 'object') {
+    return Object.fromEntries(Object.entries(value).map(([k, v]) => [k, thinArgs(v)]));
+  }
+  return value;
+}
+
+/**
+ * The conversation as sent, with bulky history thinned.
+ *
+ * Every file the model writes travels inside its own tool call, so a
+ * thirty-file app is re-sent in full on every later step — tens of thousands
+ * of tokens the provider has to read before it can answer, growing with
+ * each file. Beyond the last few steps that text is replaced by a note of its
+ * size; the files are on disk, and the model re-reads one when it needs it.
+ * Call ids and results stay paired, and the saved session keeps everything.
+ */
+/** Tools whose result is the contents of one named thing, so re-reads repeat. */
+const RE_READ = new Set(['read_file', 'read_files', 'list_dir', 'grep', 'glob']);
+
+/** The thing a call was about, when two calls for it return the same text. */
+function subjectOf(call) {
+  if (!RE_READ.has(call?.name)) return null;
+  const a = call.args ?? {};
+  const what = a.path ?? (Array.isArray(a.paths) ? a.paths.join('|') : null) ?? a.pattern;
+  if (typeof what !== 'string' || !what) return null;
+  // grep and glob also depend on what was asked, not only where.
+  const extra = call.name === 'grep' || call.name === 'glob' ? `|${a.pattern ?? ''}|${a.glob ?? ''}` : '';
+  return `${call.name}:${what}${extra}`;
+}
+
+/**
+ * Send each file once.
+ *
+ * Reading a file four times over a long task puts four copies of it in the
+ * conversation, and the first three are worth nothing: the model reads the
+ * newest and the older ones only cost tokens and invite it to answer from a
+ * stale copy. Every superseded copy becomes a line saying where the current
+ * one is. The newest is always kept whole, so nothing the model needs is
+ * taken away, and the saved session still holds the lot.
+ */
+export function dedupe(messages) {
+  const subject = new Map(); // toolCallId -> subject
+  for (const m of messages) {
+    if (m.role !== 'assistant' || !m.toolCalls?.length) continue;
+    for (const c of m.toolCalls) {
+      const s = subjectOf(c);
+      if (s) subject.set(c.id, s);
+    }
+  }
+  if (!subject.size) return messages;
+
+  const newest = new Map(); // subject -> index of the last result for it
+  messages.forEach((m, i) => {
+    if (m.role !== 'tool') return;
+    const s = subject.get(m.toolCallId);
+    if (s) newest.set(s, i);
+  });
+
+  return messages.map((m, i) => {
+    if (m.role !== 'tool') return m;
+    const s = subject.get(m.toolCallId);
+    if (!s || newest.get(s) === i) return m;
+    // Short results are not worth a note in place of themselves.
+    if ((m.content?.length ?? 0) < 400) return m;
+    const what = s.slice(s.indexOf(':') + 1).split('|')[0];
+    return {
+      ...m,
+      content:
+        `[${m.content.length} characters. This was read again later, and the current ` +
+        `contents of ${what} are further down this conversation — use those, not this.]`,
+    };
+  });
+}
+
+export function lean(messages, keep = 3) {
+  let seen = 0;
+  let cut = -1;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === 'assistant' && messages[i].toolCalls?.length && ++seen === keep) { cut = i; break; }
+  }
+  if (cut <= 0) return messages;
+  return messages.map((m, i) => {
+    if (i >= cut) return m;
+    if (m.role === 'assistant' && m.toolCalls?.length) {
+      return { ...m, toolCalls: m.toolCalls.map((c) => ({ ...c, args: thinArgs(c.args) })) };
+    }
+    if (m.role === 'tool' && THIN_RESULTS.has(m.name) && (m.content?.length ?? 0) > 1500) {
+      return {
+        ...m,
+        content: `${m.content.slice(0, 300)}
+… [${m.content.length} characters from an earlier step, trimmed ` +
+          'to keep the conversation fast — run the tool again if you need this now]',
+      };
+    }
+    return m;
+  });
+}
+
+function trace(event) {
+  if (!TRACE_FILE) return;
+  try { appendFileSync(TRACE_FILE, `${JSON.stringify({ at: Date.now(), ...event })}
+`); } catch { /* never fatal */ }
+}
+
+/** What /stats reports, counted as the session goes. */
+function newStats() {
+  return {
+    started: Date.now(), workMs: 0, turns: 0, steps: 0, tokensIn: 0, tokensOut: 0,
+    tools: {}, failed: 0, written: 0, edited: 0, commands: 0, builds: 0, stuck: 0,
+  };
+}
+
+const BUILD_COMMAND = /\b(?:npm|pnpm|yarn|bun)\s+(?:run\s+)?build\b|\bnext\s+build\b|\bvite\s+build\b/;
+
+function countTool(stats, call, out, err) {
+  stats.tools[call.name] = (stats.tools[call.name] ?? 0) + 1;
+  if (err) { stats.failed++; return; }
+  const a = call.args ?? {};
+  if (call.name === 'write_file') stats.written++;
+  if (call.name === 'batch_write') stats.written += a.files?.length ?? 0;
+  if (call.name === 'edit_file' || call.name === 'multi_edit') stats.edited++;
+  if (call.name === 'edit_files') stats.edited += new Set((a.edits ?? []).map((e) => e.path)).size || 1;
+  if (call.name === 'run_command') {
+    stats.commands++;
+    if (BUILD_COMMAND.test(a.command ?? '')) stats.builds++;
+  }
+  if (call.name === 'run_commands') stats.commands += a.commands?.length ?? 0;
+}
+
+/** The /stats box. */
+export function statsLines(st, messages) {
+  const n = (v) => Number(v).toLocaleString();
+  const top = Object.entries(st.tools).sort((x, y) => y[1] - x[1]).slice(0, 6)
+    .map(([k, v]) => `${k} ${v}`).join(' · ') || 'none yet';
+  const plural = (v, w) => `${v} ${w}${v === 1 ? '' : 's'}`;
+  const rows = [
+    ['Session', `${formatDuration(Date.now() - st.started)} open · ${formatDuration(st.workMs)} working · ${plural(st.turns, 'request')}`],
+    ['Model', `${n(st.steps)} steps · ${formatTokens(st.tokensIn)} in · ${formatTokens(st.tokensOut)} out`],
+    ['Tools', top],
+    ['Files', `${st.written} written · ${st.edited} edited`],
+    ['Commands', `${st.commands} run · ${plural(st.builds, 'build')}${st.failed ? ` · ${plural(st.failed, 'tool call')} failed` : ''}`],
+    ['Context', `${messages} messages${st.stuck ? ` · ${plural(st.stuck, 'loop')} caught` : ''}`],
+  ];
+  const w = Math.max(...rows.map(([k]) => k.length));
+  return ['', `  ${blue('Stats')}`, ...rows.map(([k, v]) => `  ${dim(k.padEnd(w))}  ${v}`), ''];
+}
+
+/** Parallel workers at once, and how many steps each may take. */
+const MAX_WORKERS = 3;
+const WORKER_STEPS = Number(process.env.UCODE_WORKER_STEPS) || 60;
+
+/** Workers build; they do not plan, delegate further, or load skills themselves. */
+const WORKER_EXCLUDED = new Set(['delegate', 'update_plan', 'load_skill']);
+
+const planTool = {
+  name: 'update_plan',
+  description:
+    'Keep a short checklist the user can see, for work with three or more steps. ' +
+    'Send the whole list every time: at most 6 items, a few words each, with done: true ' +
+    'on the finished ones. Update it as items finish. Skip it for small tasks.',
+  parameters: {
+    type: 'object',
+    properties: {
+      items: {
+        type: 'array',
+        description: 'The whole plan, in order. At most 6.',
+        items: {
+          type: 'object',
+          properties: {
+            text: { type: 'string', description: 'A few words: "Build the upload box".' },
+            done: { type: 'boolean', description: 'True once it is finished.' },
+          },
+          required: ['text'],
+        },
+      },
+    },
+    required: ['items'],
+  },
+};
+
+const delegateTool = {
+  name: 'delegate',
+  description:
+    'Build independent parts in parallel. Up to 3 workers run at once, each with the ' +
+    'same tools as you. Use it when the work splits cleanly into parts that touch ' +
+    'different files - e.g. the API route, the upload component and the results view. ' +
+    'A worker sees only its instructions, so make them complete: the files it owns, ' +
+    'what to build, the exact interfaces (props, types, request and response shapes) ' +
+    'it must match, and the design rules. Set up shared files (package.json, design ' +
+    'tokens, shared types) yourself first. You get back each worker\'s summary and ' +
+    'the files it changed; wire the parts together and check the whole afterwards.',
+  parameters: {
+    type: 'object',
+    properties: {
+      tasks: {
+        type: 'array',
+        description: 'Up to 3 independent pieces of work.',
+        items: {
+          type: 'object',
+          properties: {
+            name: { type: 'string', description: 'Two or three words: "api route", "upload ui".' },
+            instructions: { type: 'string', description: 'Everything the worker needs to do its part completely.' },
+          },
+          required: ['name', 'instructions'],
+        },
+      },
+    },
+    required: ['tasks'],
+  },
+};
+
+/** The instructions a parallel worker starts with. */
+function workerPrompt({ cwd, name, memory, skills, map }) {
+  return [
+    `You are a ucode worker called "${name}" - one of several building parts of the same project at the same time.`,
+    '',
+    `Working directory: ${cwd}`,
+    `Platform: ${process.platform}`,
+    '',
+    '- Do exactly the task you were given. Touch only the files it names, or new files in',
+    '  the area it owns - other workers are editing the rest of the project right now.',
+    '- Read before you edit. read_files for several files, batch_write for several new',
+    '  files, edit_files for changes across files.',
+    '- Nothing has a keyboard: pass non-interactive flags. Do not start dev servers and do',
+    '  not install packages unless the task says to - say what you need instead.',
+    '- Finish what you build: real content, every state handled, no TODOs.',
+    '- When done, reply with two or three sentences: what you built, in which files, and',
+    '  anything the lead has to wire up.',
+    '- The files you own may not exist yet - create them. Do not go looking for them',
+    '  first. The project map below shows what does exist; read only the files whose',
+    '  interfaces you must match, then start writing within two or three steps.',
+    ...(memory ? ['', '## Project memory', '', memory] : []),
+    ...(map ? ['', '## Project map', '', map] : []),
+    ...(skills ? ['', '## Instructions in force', '', skills] : []),
+  ].join('\n');
+}
+
+/** The files a writing tool call touches. */
+function pathsOf(call) {
+  const a = call.args ?? {};
+  if (call.name === 'batch_write' || call.name === 'edit_files') return (a.files ?? []).map((f) => f?.path).filter(Boolean);
+  return a.path ? [a.path] : [];
+}
+
+const exists = (p) => access(p).then(() => true, () => false);
+
+/** Skills reach the model as one extra tool, so bodies load only when wanted. */
+const loadSkillTool = {
+  name: 'load_skill',
+  description:
+    'Load the full instructions for one of the skills listed in your system prompt. ' +
+    'Call it the moment a task matches one — before planning, before writing anything ' +
+    '— then follow what it says.',
+  parameters: {
+    type: 'object',
+    properties: { name: { type: 'string', description: 'The skill name, exactly as listed.' } },
+    required: ['name'],
+  },
+};
+
+function systemPrompt({ cwd, skills, mode, check, map, memory }) {
+  const list = catalogue(skills);
+
+  return [
+    'You are ucode, a coding agent working directly in the user\'s terminal.',
+    '',
+    `Working directory: ${cwd}`,
+    `Platform: ${process.platform}`,
+    ...(memory ? [
+      '',
+      '## Project memory',
+      '',
+      'Standing instructions from the user. They outrank your defaults.',
+      '',
+      memory,
+    ] : []),
+    '',
+    '## How to work',
+    '',
+    'Before building anything, turn the request into a list of what it must do — every',
+    'feature named, and the ones any user would expect whether or not they were named',
+    '(an empty state, an error state, the keyboard doing the obvious thing, working on',
+    'a phone). Keep it as the plan. Build against it, then go through it one item at a',
+    'time before you say a word about being finished. Most of what gets missed was',
+    'never written down.',
+    '',
+    'Do not put code in your reply. Not a snippet, not "here is the key part", not a',
+    'summary of the file. It is already in the file and the user can open it; pasting',
+    'it again buries the one or two sentences that actually matter. Say what it does',
+    'and what to try.',
+    '',
+    'Do not claim it is done while anything is still running or unchecked. "I have',
+    'built it" said before the build finishes is worse than saying nothing: the user',
+    'believes you, looks, and finds it broken. Finish, check, then say so — and if',
+    'something is incomplete, say which part and why.',
+    '',
+    'BE FAST. Every tool call is a round trip, and round trips are nearly all of the',
+    'time a build takes. So: write a whole app in ONE batch_write rather than a',
+    'write_file per file. Read every file you need in ONE read_files. Never read a',
+    'file you just wrote, and never read one back after edit_file — the result',
+    'already contains it. Do not re-check work the checks have already reported on.',
+    'Fast is not sloppy: it is the same work with the waiting taken out.',
+    '',
+    'SPEAK AS "I", NEVER "WE". You are doing this, not a committee: "I will build',
+    'Tide as a single HTML file", not "we have created the file".',
+    '',
+    'Say three or four things over a whole build, not one per step:',
+    '  - Open with what you are going to make and how, in one line, before anything',
+    '    else: "I will build Tide as a single HTML file - markup, one stylesheet, one',
+    '    module - with the tasks kept in localStorage."',
+    '  - One line when you move between the big pieces of work: "The layout is done,',
+    '    now the animations."',
+    '  - One line at the end saying what it does and how to try it.',
+    'That is all. A line before every tool call is not narration, it is noise: the',
+    'steps already show on screen, and repeating them in words buries the few',
+    'sentences worth reading.',
+    '',
+    'FIRST, EVERY TIME: write one short line saying what you are about to do, then',
+    'make the tool calls. Never open a turn with a tool call and no words. Examples:',
+    '"Right, the HTML structure first." / "Now the state and the render loop." /',
+    '"That is the layout done - onto the animations." / "Let me see what is there."',
+    'One sentence, your own voice, before the actions - not after them, not instead',
+    'of them, and not a restatement of what was asked. The user is watching this',
+    'scroll past; without those lines it is a list of file operations and they cannot',
+    'tell what you are building. This matters as much as the code.',
+    '',
+
+    '',
+    'Before you guess at an API, ask: type_of gives the exact signature from the',
+    'TypeScript this project has installed, and find_symbol says where something is declared without',
+    'reading five files to find it. Rename with rename_symbol rather than edit_file — a',
+    'find-and-replace that matches too much is the most common broken edit. Reach for',
+    'add_block before writing a table, an empty state or a dashboard by hand.',
+    '',
+    ...(mode === 'plan' ? [
+      'You are in PLAN MODE. Reading, searching and research are available; every tool',
+      'that writes a file or runs a command has been withheld. Investigate, then set out',
+      'what you would change: which files, which functions, in what order. Never imply',
+      'you have made a change you are not able to make.',
+      '',
+    ] : []),
+    ...(check ? [
+      `This project checks itself with: ${check}`,
+      'After changing code, run that and report what actually happened. "It should work"',
+      'is not a result.',
+      '',
+    ] : []),
+    '- Read before you write. Never edit a file you have not read this session.',
+    '- Prefer edit_file to write_file. Rewrite a whole file only when creating it, or',
+    '  when the change genuinely touches most of it.',
+    '- old_string must be copied out of what read_file showed you, character for',
+    '  character, without the line-number gutter, and must appear exactly once. Add',
+    '  surrounding lines until it does.',
+    '- read_file returns up to 600 lines. Read the whole file before editing it rather',
+    '  than editing from a fragment; pass offset to continue a long one.',
+    '',
+    '## Going fast',
+    '',
+    'Every tool call is a round trip to you, and the round trip — not the disk, not',
+    'the shell — is where the time goes. So:',
+    '',
+    '- Need more than one file? read_files, all of them in one call. Never read files',
+    '  one at a time when you already know which ones you want.',
+    '- An edit result shows the file as it now stands. Do not read a file again after',
+    '  editing it - you already have its current text.',
+    '- Put independent calls in the same message — several greps, a glob and a read.',
+    '  Read-only calls in one message run at the same time.',
+    '- New Next.js app? create_app - one step, never create-next-app or shadcn init. It',
+    '  copies a starter that already builds and installs it in the background.',
+    '- batch_write to lay out several new files at once - but at most about four per',
+    '  call: one slip in a huge call throws the whole call away.',
+    '- multi_edit for several changes',
+    '  to one file, edit_files for a change that spans several files.',
+    '- For work with three or more steps, keep a short plan with update_plan - at most',
+    '  six items of a few words - and tick items off as they finish. Skip it for small jobs.',
+    '- When a build splits into parts that touch different files (the API route, the',
+    '  upload component, the results view), set up the shared files yourself, then hand',
+    '  the parts to delegate so they are built in parallel.',
+    '- A package.json you write starts installing in the background immediately; keep',
+    '  writing files. Running the install yourself afterwards just waits for that one.',
+    '- When you finish, ucode type-checks what you changed and hands you the errors, so',
+    '  there is no need to run tsc yourself.',
+    '- Do not open or drive a browser. Checking the page in one is something the user',
+    '  asks for with /look; your job is to leave the app in a state worth looking at.',
+    '  reports - errors, layout that overflows a phone, the review points worth fixing -',
+    '  in one pass, then look once more. A clean second look means it is done: report',
+    '  back instead of polishing in circles. Never call an interface finished unlooked at.',
+    '  name, handles keys and returns the live link. Build locally first.',
+    '- Nothing you run has a keyboard. Pass the non-interactive flag to anything that',
+    '  would ask a question, or it fails instead of waiting: create-next-app --yes,',
+    '  npx shadcn@latest init -d -y, npx shadcn@latest add <names> -y, npm init -y.',
+    '- Dev servers start in the background by themselves, and the result tells you the',
+    '  URL once the server says it is ready. Do not start one twice, do not sleep while',
+    '  waiting for it, and do not curl it before that result comes back.',
+    '',
+    '## When something fails',
+    '',
+    '- A failed build names the problem. Fix exactly that, then build again. Never go',
+    '  exploring inside node_modules: a missing component or package is one install away.',
+    '- A build takes most of a minute. Fix every error it lists in one pass - multi_edit,',
+    '  edit_files - before building again, never one error per build.',
+    '- Never delete an app folder to start over. Fix it where it is - starting again throws',
+    '  away the install and everything already written.',
+    '- Run an app\'s commands with cwd set to its folder, and keep paths inside those',
+    '  commands relative to that folder.',
+    '',
+    '## Safety',
+    '',
+    '- run_command runs without asking. That is trust rather than licence: never run',
+    '  anything destructive the user did not ask for.',
+    '- Paths are relative to the working directory. Anything outside it needs the user',
+    '  to approve it first.',
+    '- Verify. After changing code, run the tests or a quick check with run_command.',
+    '',
+    '## Saying what you are doing',
+    '',
+    '- Before every tool call, write ONE short line naming the file or command:',
+    '  "Reading tui.js", "Fixing the spinner in loop.js", "Running npm test".',
+    '- Present tense, under ten words, and no full stop at the end. It is a label on',
+    '  work happening right now, not a sentence about work that is finished.',
+    '- That line and nothing else in the message. No preamble, no plan, no bullets —',
+    '  the user reads it live while the tool runs.',
+    '- Say the next one when you take the next step, not all of them up front.',
+    '',
+    '## Answering',
+    '',
+    '- Be short. Two or three sentences is usually the entire answer. This is a',
+    '  terminal, not a document.',
+    '- No preamble, no restating the question, no "I will now...". Just answer.',
+    '- Do not narrate what the tool output already showed. The user watched the diff',
+    '  and the command output; cover only what is not obvious from them.',
+    '- Skip closing summaries of work the user just watched you do — but never end',
+    '  a turn silently. If there is genuinely nothing to add, one short line saying',
+    '  what changed is the whole answer.',
+    '- Markdown. Fenced blocks with a language tag get highlighted.',
+    '- Point at code as path:line so the user can jump straight to it.',
+    '- Report honestly. If a command failed or you skipped something, say so.',
+    '- Length tracks the question: a one-line question gets a one-line answer.',
+    '- Brevity is about your prose and never about your work. What you build is',
+    '  finished: every control wired, every state handled, no TODO left behind.',
+    ...(list ? [
+      '',
+      '## Skills',
+      '',
+      'These instruction packs are available. When a task matches one, load it with',
+      'load_skill as your first step — before planning, before writing anything — and',
+      'then follow it. A skill already in this conversation outranks your own defaults',
+      'and is not advisory.',
+      '',
+      list,
+    ] : []),
+    '',
+    '## Project map',
+    '',
+    'Every file in the project at the start of this turn, with the names each code file',
+    'exports. Go straight to the files you need instead of searching for them.',
+    '',
+    map || '(not available)',
+  ].join('\n');
+}
+
+export class Agent {
+  constructor({ cwd, debug = false }) {
+    this.cwd = cwd;
+    this.debug = debug;
+    // A full-screen layout only makes sense on a real terminal. Piped input,
+    // CI and `echo ... | ucode` get the line-based interface instead.
+    this.full = Boolean(process.stdout.isTTY && process.stdin.isTTY);
+    this.ui = this.full ? new Screen({ cwd }) : new Plain({ cwd });
+    this.stats = newStats();
+    this.skills = [];
+    this.session = newSession(cwd, model());
+    this.working = [];
+    this.loaded = new Set();
+    this.abort = null;
+    this.busy = false;
+    this.check = null;
+  }
+
+  // -- history -------------------------------------------------------------
+
+  push(message) {
+    this.session.messages.push(message);
+    this.working.push(message);
+  }
+
+  /**
+   * Answer every tool call a stopped turn never got to.
+   *
+   * An assistant message ends by asking for tools, and each of those asks
+   * needs an answer. Abandon them and the conversation is left mid-sentence,
+   * so the next time the model reads it the only sensible thing to do is
+   * carry on where it left off — which is exactly what the user pressed stop
+   * to prevent. Saying "this did not happen" for each one ends the sentence,
+   * and a line from the user ends the task.
+   */
+  closeInterrupted() {
+    const answered = new Set(this.working.filter((m) => m.role === 'tool').map((m) => m.toolCallId));
+    const missing = [];
+    for (const m of this.working) {
+      if (m.role !== 'assistant' || !m.toolCalls?.length) continue;
+      for (const call of m.toolCalls) {
+        if (!answered.has(call.id)) missing.push(call);
+      }
+    }
+    for (const call of missing) {
+      this.push({
+        role: 'tool',
+        toolCallId: call.id,
+        name: call.name,
+        content: 'The user stopped the turn before this ran. It did not happen, and it must not be retried.',
+      });
+    }
+    if (missing.length) {
+      this.push({
+        role: 'user',
+        content: 'I stopped that. Drop it and wait for what I ask next — do not pick it back up.',
+      });
+    }
+    return missing.length;
+  }
+
+  async persist() {
+    try {
+      this.session.model = model();
+      await save(this.session);
+    } catch (err) {
+      // Losing the save must not lose the turn.
+      this.ui.error(err, { debug: this.debug });
+    }
+  }
+
+  // -- startup -------------------------------------------------------------
+
+  /**
+   * Everything a turn needs, minus the terminal.
+   *
+   * Split out of start() so another front end could prepare an agent and drive
+   * turn() itself.
+   */
+  async bootstrap() {
+    setRoot(this.cwd);
+    // Parallel workers can ask at the same moment; the questions queue up and
+    // are put to the user one at a time, never on top of each other.
+    let asking = Promise.resolve();
+    setConfirm((request) => {
+      const next = asking.then(() => this.ui.confirm(request));
+      asking = next.catch(() => {});
+      return next;
+    });
+    this.skills = await loadSkills({ cwd: this.cwd });
+    await this.detectCheck();
+  }
+
+  async start() {
+    await this.bootstrap();
+
+    if (this.full) {
+      await this.ui.start();
+      this.ui.onInterrupt = () => {
+        if (this.busy && this.abort) {
+          this.abort.abort();
+          this.ui.stopSpinner();
+          this.ui.stopTimer?.();
+          this.ui.note('interrupted');
+        }
+      };
+      this.ui.onModeChange = () => this.showHeader({ clear: false });
+    }
+
+    for (const problem of this.skills.problems ?? []) {
+      this.ui.write(theme.warn(`  skill not loaded: ${problem}`));
+    }
+
+    this.showHeader();
+    this.installSignals();
+
+    // Checked in the background; nothing here waits on it.
+    autoUpdate({
+      onUpdated: (version) => {
+        this.ui.setFacts?.({ update: version });
+        if (!this.ui.welcoming?.()) this.ui.note(`updated to v${version} — it takes over the next time you start ucode`);
+      },
+    });
+    await this.repl();
+  }
+
+  showHeader({ clear = true } = {}) {
+    if (clear && this.full) this.ui.clearScreen();
+    const stats = usage(this.working, contextLimit());
+    this.ui.header({
+      cwd: this.cwd,
+      model: modelName(),
+      used: stats.used,
+      limit: stats.limit,
+      title: this.session.title === 'Untitled' ? 'new session' : this.session.title,
+    });
+  }
+
+  installSignals() {
+    const flush = async () => {
+      await save(this.session).catch(() => {});
+      process.exit(0);
+    };
+    process.on('SIGTERM', flush);
+
+    // The full-screen UI reads keys itself, so it owns ctrl+c and esc.
+    if (this.full) return;
+
+    this.ui.rl.on('SIGINT', () => {
+      if (this.busy && this.abort) {
+        this.abort.abort();
+        this.ui.stopSpinner();
+        this.ui.write(dim('  interrupted'));
+        return;
+      }
+      this.ui.write(dim('  (ctrl+d or /exit to quit)'));
+      this.ui.rl.prompt();
+    });
+  }
+
+  /**
+   * How this project verifies itself, worked out once at startup. Null when
+   * there is genuinely nothing to run — verification is only insisted on where
+   * there is something to insist on.
+   */
+  async detectCheck() {
+    const has = (f) => readFile(path.join(this.cwd, f)).then(() => true, () => false);
+
+    if (await has('package.json')) {
+      try {
+        const pkg = JSON.parse(await readFile(path.join(this.cwd, 'package.json'), 'utf8'));
+        if (pkg.scripts?.test && !/no test specified/i.test(pkg.scripts.test)) {
+          this.check = 'npm test';
+          return;
+        }
+      } catch { /* an unreadable package.json is not worth failing over */ }
+    }
+    if (await has('Cargo.toml')) { this.check = 'cargo test'; return; }
+    if (await has('go.mod')) { this.check = 'go test ./...'; return; }
+    if (await has('pyproject.toml') || await has('pytest.ini')) { this.check = 'pytest'; return; }
+    if (await has('Makefile')) { this.check = 'make test'; return; }
+    this.check = null;
+  }
+
+  // -- REPL ----------------------------------------------------------------
+
+  async repl() {
+    let sawInput = false;
+
+    for (;;) {
+      const line = await this.ui.ask();
+      if (line === null) {
+        // End of input before anything was typed. On Windows this is almost
+        // always npm's PowerShell shim, which runs the CLI as `$input | node`.
+        // The pipe makes stdin a non-TTY, readline hits EOF at once, and the
+        // banner flashes up and vanishes — which looks like a crash rather
+        // than like a program that was never given a keyboard. So say which.
+        if (!sawInput && !process.stdin.isTTY) this.explainNoKeyboard();
+        break;
+      }
+
+      const input = line.trim();
+      if (input) sawInput = true;
+      if (!input) continue;
+
+      if (input.startsWith('/')) {
+        if (await this.command(input) === 'exit') break;
+        continue;
+      }
+
+      try {
+        await this.turn(input);
+      } catch (err) {
+        this.ui.error(err, { debug: this.debug });
+      }
+    }
+
+    await this.shutdown();
+  }
+
+  explainNoKeyboard() {
+    this.ui.blank();
+    this.ui.write(theme.warn('  ucode could not reach the keyboard, so it stopped.'));
+    this.ui.blank();
+    this.ui.write('  That happens when input is piped rather than typed. On Windows it is');
+    this.ui.write("  usually npm's PowerShell wrapper, which pipes stdin.");
+    this.ui.blank();
+    this.ui.write(`  ${blue('Any of these work:')}`);
+    this.ui.write(`    ${sky('ucode.cmd')}          the cmd shim, which keeps the keyboard`);
+    this.ui.write(`    ${sky('npx ucode-agent')}    runs it directly`);
+    this.ui.write('    or start it from Command Prompt or Windows Terminal');
+    this.ui.blank();
+  }
+
+  async shutdown() {
+    await closeBrowser().catch(() => {});
+    this.ui.stopSpinner();
+    if (this.session.messages.length) {
+      await this.persist();
+      this.ui.write(dim(`\n  saved · ${this.session.title}`));
+    }
+    this.ui.close();
+  }
+
+  // -- one turn ------------------------------------------------------------
+
+  /**
+   * Pull image paths out of the message and load them, so "what is wrong in
+   * screenshot.png" works without a separate command for it.
+   */
+  async attachImages(input) {
+    const mentioned = input.match(/[^\s"']+\.(?:png|jpe?g|gif|webp)\b/gi) ?? [];
+    const images = [];
+
+    for (const name of mentioned) {
+      const file = path.resolve(this.cwd, name);
+      try {
+        const buf = await readFile(file);
+        if (buf.length > 4 * 1024 * 1024) {
+          this.ui.note(`${name} is ${(buf.length / 1024 / 1024).toFixed(1)}MB — too big to send, skipped`);
+          continue;
+        }
+        const ext = path.extname(file).toLowerCase().slice(1);
+        images.push(`data:image/${ext === 'jpg' ? 'jpeg' : ext};base64,${buf.toString('base64')}`);
+        this.ui.note(`attached ${name}`);
+      } catch {
+        // Just a filename mentioned in passing, not a file on disk.
+      }
+    }
+
+    return images;
+  }
+
+  /**
+   * Skills that this request should arrive with, already loaded.
+   *
+   * The load_skill tool asks the model to notice that a task needs a skill,
+   * and a model in a hurry to be helpful does not always notice. For work
+   * where the skill *is* the quality bar — anything with a user interface in
+   * it — that is not a discovery to make after the app has been built. So the
+   * request is matched against each skill's trigger words and the body goes in
+   * before the model takes its first step.
+   */
+  autoLoad(input) {
+    for (const skill of autoLoadFor(this.skills, input)) {
+      if (this.loaded.has(skill.name)) continue;
+      this.loaded.add(skill.name);
+      this.push(skillMessage(skill, { automatic: true }));
+      this.ui.note(`${skill.name} skill loaded for this`);
+    }
+  }
+
+  async turn(input) {
+    forgetReviews(); // a new request: its apps get a fresh design review
+    const images = await this.attachImages(input);
+    this.push(images.length
+      ? { role: 'user', content: input, images }
+      : { role: 'user', content: input });
+
+    if (!this.session.title || this.session.title === 'Untitled') {
+      this.session.title = titleFrom(input);
+    }
+
+    this.autoLoad(input);
+    // What the model is told about the project, fresh for this turn.
+    [this.map, this.memory] = await Promise.all([
+      projectMap(this.cwd).catch(() => ''),
+      loadMemory(this.cwd).catch(() => ''),
+    ]);
+    await this.persist();
+
+    // A busy model was swapped for a fallback earlier; after a few minutes the
+    // one the user chose gets another go.
+    this.preferred ??= model();
+    if (model() !== this.preferred && Date.now() > (this.cooldownUntil ?? 0)) {
+      setModel(this.preferred);
+      if (this.full) this.showHeader({ clear: false });
+    }
+
+    this.busy = true;
+    this.abort = new AbortController();
+    this.endedSilently = false;
+
+    const turnStarted = Date.now();
+    let finished = false;
+    this.ui.turnStart?.();
+    try {
+      for (;;) {
+        try {
+          await this.run();
+          break;
+        } catch (err) {
+          // The daily free cap mid-build: wait for the reset and carry on,
+          // rather than leaving a half-built app for the user to restart.
+          if (err?.detail?.daily && !this.abort.signal.aborted && (await this.waitForReset(err))) continue;
+          throw err;
+        }
+      }
+      finished = true;
+    } catch (err) {
+      if (err?.kind === 'aborted' || this.abort.signal.aborted) this.ui.write(dim('  turn cancelled'));
+      else throw err;
+    } finally {
+      trace({ kind: 'turn', ms: Date.now() - turnStarted });
+      this.stats.workMs += Date.now() - turnStarted;
+      this.stats.turns++;
+      if (!finished) this.closeInterrupted();
+      const ok = finished && !this.endedSilently;
+      this.busy = false;
+      this.abort = null;
+      this.ui.stopSpinner();
+      this.ui.stopTimer?.();
+      this.activity = null;
+      await this.persist();
+      if (this.full) this.showHeader({ clear: false });
+      if (finished) this.openWhenReady(turnStarted);
+      // Last, so "Done" is the last thing that happens rather than the last
+      // thing said before several more things happen.
+      this.ui.turnEnd?.({ ok });
+    }
+  }
+
+  /** The tools the model may see, given the mode. */
+  toolsNow() {
+    const all = [...tools, loadSkillTool, planTool, delegateTool];
+    if (this.ui.mode !== 'plan') return all;
+    return all.filter((t) => !WRITES.has(t.name));
+  }
+
+  /** Model, tools, model, until it answers with prose. */
+  async run() {
+    const available = this.toolsNow();
+    let argRetries = 0;
+    let continuations = 0;
+    let askedToVerify = false;
+    let askedToSpeak = false;
+    let fixRounds = 0;
+    this.failovers = 0;
+    this.tried = new Set([model()]);
+
+    this.stuck = new StuckWatch();
+    this.touched = new Set();
+    this.sinceCheck = new Set();
+    this.logWatch = new LogWatch();
+    this.ranSomething = false;
+
+    for (let step = 0; step < MAX_STEPS; step++) {
+      await this.maybeFold();
+      this.ui.startSpinner(step === 0 ? 'thinking' : 'working');
+
+      let reply;
+      let streaming = false;
+      this.early = new Map();
+
+      try {
+        const opts = {
+          signal: this.abort.signal,
+          onWait: (text) => this.ui.updateSpinner(text),
+        };
+        // Only a real terminal has somewhere to stream into.
+        if (this.full) {
+          opts.onThinking = (delta) => this.ui.thinkingDelta(delta);
+          opts.onText = (delta) => {
+            if (!streaming) {
+              streaming = true;
+              this.ui.thinkingEnd();
+              this.ui.streamBegin();
+            }
+            this.ui.streamDelta(delta);
+          };
+          // Read-only calls start the moment they are fully written, while the
+          // rest of the reply is still arriving. Nothing that writes or runs is
+          // started early: a reply that fails halfway must leave no side effects.
+          opts.onToolCall = (call) => {
+            if (PARALLEL_SAFE.has(call.name) && !call.parseError && !this.early.has(call.id)) {
+              this.early.set(call.id, this.execute(call));
+            }
+          };
+        }
+
+        var asked = Date.now();
+        reply = await ask(
+          [
+            {
+              role: 'system',
+              content: systemPrompt({
+                cwd: this.cwd,
+                skills: this.skills,
+                mode: this.ui.mode,
+                check: this.check,
+                map: this.map,
+                memory: this.memory,
+              }),
+            },
+            ...dedupe(lean(this.working)),
+          ],
+          available,
+          opts
+        );
+      } catch (err) {
+        this.ui.thinkingEnd();
+        if (streaming) this.ui.streamEnd();
+
+        // The model invented a tool and the provider rejected the request
+        // outright. Tell it what it did and let it try again.
+        if (err.kind === 'bad_tool_call' && argRetries < MAX_ARG_RETRIES) {
+          argRetries++;
+          this.ui.stopSpinner();
+          this.ui.toolFailed(
+            `${err.detail?.attemptedName ?? 'invalid tool call'} — retrying (${argRetries}/${MAX_ARG_RETRIES})`
+          );
+          this.push({
+            role: 'user',
+            content:
+              `Your last tool call was rejected. ${err.failed} The only tools that exist ` +
+              `are: ${available.map((t) => t.name).join(', ')}. Try again with one of them.`,
+          });
+          continue;
+        }
+
+        // Busy, slow or down: move to the next model and carry on, rather
+        // than ending a half-built app with an error.
+        if (passing(err) && !this.abort.signal.aborted && (await this.failover(err))) continue;
+        throw err;
+      }
+
+      // A reply that is nothing but tool calls never starts a text stream, so
+      // the thinking timer has to be closed out here as well.
+      this.ui.thinkingEnd();
+      this.ui.stopSpinner();
+      this.record(reply.usage);
+      this.ui.step?.();
+      this.stats.steps++;
+      this.stats.tokensIn += reply.usage.promptTokens ?? 0;
+      this.stats.tokensOut += reply.usage.outputTokens ?? 0;
+      trace({
+        kind: 'model', who: 'lead', model: model(), ms: Date.now() - asked,
+        in: reply.usage.promptTokens, out: reply.usage.outputTokens,
+        calls: reply.toolCalls.map((c) => c.name),
+      });
+
+      // Streamed text is already on screen; turn it into rendered markdown.
+      // Text that turns out to be narration ahead of a tool call folds into a
+      // status line instead — that is where the live commentary comes from.
+      const narrating = reply.toolCalls.length > 0;
+      if (streaming) this.ui.streamEnd({ asNarration: narrating });
+      else if (reply.text && narrating && isLabel(reply.text)) this.ui.narrate(reply.text);
+      else if (reply.text) this.ui.assistant(reply.text);
+
+      if (reply.toolCalls.length === 0) {
+        // The answer stopped at the provider's output cap rather than at the
+        // end of a thought, so it is cut mid-word. Ask for the rest instead of
+        // handing over half an answer with no sign there was more.
+        if (reply.finishReason === 'length' && reply.text && continuations < MAX_CONTINUATIONS) {
+          continuations++;
+          this.push({ role: 'assistant', content: reply.text });
+          this.push({
+            role: 'user',
+            content:
+              'Your reply stopped at the output limit, mid-sentence. Carry on from exactly ' +
+              'where it broke off. Do not repeat any of it, do not start again, and do not ' +
+              'introduce it — just continue.',
+          });
+          this.ui.note('hit the output limit — asking for the rest');
+          continue;
+        }
+
+        // Type-check what changed and hand back the errors, a few rounds at most.
+        // A turn that ends on a broken build is the most common way an app
+        // gets handed over as done when it is not.
+        if (fixRounds < MAX_FIX_ROUNDS) {
+          const problems = await this.autoCheck();
+          if (problems) {
+            fixRounds++;
+            if (reply.text) this.push({ role: 'assistant', content: reply.text });
+            this.push({
+              role: 'user',
+              content:
+                `ucode checked the files you changed and found errors (round ${fixRounds} of ` +
+                `${MAX_FIX_ROUNDS}). Fix all of them, then finish.\n\n${problems}`,
+            });
+            continue;
+          }
+        }
+
+        // It changed code and never ran anything. Send it back once.
+        if (this.touched.size && !this.ranSomething && this.check && !askedToVerify) {
+          askedToVerify = true;
+          if (reply.text) this.push({ role: 'assistant', content: reply.text });
+          this.push({
+            role: 'user',
+            content:
+              `You changed ${[...this.touched].join(', ')} and did not check it. Run ` +
+              `\`${this.check}\` with run_command now, then say what actually happened — ` +
+              'if it failed, show the output rather than claiming it worked. If that is ' +
+              'the wrong way to check this project, run the right one and say which.',
+          });
+          this.ui.note('verifying the change');
+          continue;
+        }
+
+        /**
+         * It did the work and then said nothing at all.
+         *
+         * Reasoning models do this, and the instruction to skip closing
+         * summaries makes it more likely. Silence is fine as a style; it is
+         * not fine as an answer, because the user cannot tell it apart from a
+         * crash — and if they asked what happened, they asked. One nudge,
+         * once per turn, and only when there was actually work to report.
+         */
+        if (!reply.text?.trim() && this.touched.size + (this.ranSomething ? 1 : 0) > 0 && !askedToSpeak) {
+          askedToSpeak = true;
+          this.push({
+            role: 'user',
+            content:
+              'You stopped without saying anything. If the thing I asked for is not ' +
+              'built yet, carry on and build it. If it is, tell me in one or two ' +
+              'sentences what it does and how to try it. No preamble, no diffs.',
+          });
+          continue;
+        }
+
+        this.push({ role: 'assistant', content: reply.text });
+        if (!reply.text?.trim()) {
+          // Silence after being asked to speak is not a finished turn. Saying
+          // "Done" here is the worst thing available: the user believes it,
+          // looks, and finds the thing they asked for was never built.
+          this.ui.note('the model stopped without saying anything — the work may be unfinished');
+          this.endedSilently = true;
+        }
+        return;
+      }
+
+      this.push({ role: 'assistant', content: reply.text || '', toolCalls: reply.toolCalls });
+      await this.persist();
+
+      const badArgs = await this.runCalls(reply.toolCalls);
+      if (this.abort.signal.aborted) return;
+
+      // Malformed arguments go back to the model, but only so many times.
+      if (badArgs) {
+        argRetries++;
+        if (argRetries > MAX_ARG_RETRIES) {
+          throw new Failure({
+            kind: 'bad_tool_args',
+            attempted: 'running the tools the model asked for',
+            failed:
+              `${modelName()} produced invalid tool arguments ${argRetries} times running ` +
+              'and could not correct itself.',
+            fix:
+              'Say what you want more concretely, or /model to another one — North Mini ' +
+              'Code and Nemotron 3.5 Lightning are both steadier with tool arguments.',
+          });
+        }
+      } else {
+        argRetries = 0;
+      }
+
+      await this.persist();
+    }
+
+    throw new Failure({
+      kind: 'step_limit',
+      attempted: 'finishing your request',
+      failed: `The model was still calling tools after ${MAX_STEPS} steps.`,
+      fix:
+        'Nothing is lost — everything so far is on disk. Say "carry on where you left ' +
+        'off" to continue. If it was repeating one step, it is looping: break the task ' +
+        'up, or /new to reset.',
+    });
+  }
+
+  /**
+   * Run one round of tool calls.
+   *
+   * Consecutive read-only calls go out together — four files read at once
+   * rather than four round trips — while anything that writes or executes runs
+   * on its own, in order. Returns whether any call had unusable arguments.
+   */
+  async runCalls(calls) {
+    const groups = [];
+    let batch = [];
+
+    for (const call of calls) {
+      if (PARALLEL_SAFE.has(call.name)) {
+        batch.push(call);
+      } else {
+        if (batch.length) { groups.push(batch); batch = []; }
+        groups.push([call]);
+      }
+    }
+    if (batch.length) groups.push(batch);
+
+    let badArgs = false;
+
+    for (const group of groups) {
+      if (this.abort.signal.aborted) return badArgs;
+
+      const noted = (call) => {
+        if (FILE_WRITES.has(call.name)) {
+          for (const p of pathsOf(call)) { this.touched.add(p); this.sinceCheck.add(p); }
+        }
+        if (call.name === 'run_command' || call.name === 'run_commands') this.ranSomething = true;
+      };
+
+      if (group.length > 1) {
+        for (const call of group) {
+          this.ui.toolCall(describe(call.name, call.args));
+          noted(call);
+        }
+        this.ui.startSpinner(`${group.length} lookups at once`);
+
+        const settled = await Promise.all(
+          group.map((call) => (this.early.get(call.id) ?? this.execute(call)).then((r) => ({ call, ...r })))
+        );
+
+        this.ui.stopSpinner();
+        for (const { call, out, err } of settled) {
+          if (err) badArgs = this.reportFailure(call, err) || badArgs;
+          else this.reportResult(call, out);
+        }
+        continue;
+      }
+
+      for (const call of group) {
+        if (this.abort.signal.aborted) return badArgs;
+
+        const label = describe(call.name, call.args);
+        if (!SILENT.has(call.name)) this.ui.toolCall(label);
+        this.ui.startSpinner(label);
+        noted(call);
+
+        const { out, err } = await (this.early.get(call.id) ?? this.execute(call));
+        this.ui.stopSpinner();
+        if (err) badArgs = this.reportFailure(call, err) || badArgs;
+        else this.reportResult(call, out);
+      }
+    }
+
+    return badArgs;
+  }
+
+  /**
+   * Check a finished call against the stuck patterns (see stuck.js) and return
+   * the note to add to its result. A nudge that did not work hands the turn to
+   * another model.
+   */
+  stuckNote(call, outcome) {
+    if (!this.stuck) return '';
+    const verdict = this.stuck.observe(eventFor(call, outcome));
+    if (!verdict) return '';
+    this.stats.stuck++;
+    if (verdict.action === 'switch') {
+      const next = fallbackFor(model(), this.tried ?? new Set([model()]));
+      if (next) {
+        this.tried?.add(next);
+        this.ui.note(`${modelName(model())} kept ${describeHit(verdict.hit)} — handing over to ${modelName(next)}`);
+        setModel(next);
+        this.cooldownUntil = Date.now() + COOLDOWN;
+        if (this.full) this.showHeader({ clear: false });
+      }
+    }
+    return verdict.text ? `\n\n${verdict.text}` : '';
+  }
+
+  /**
+   * The daily free cap was hit mid-turn: keep the session, count down to the
+   * reset, and carry on by itself. Esc stops the wait like any other turn.
+   */
+  async waitForReset(err) {
+    const resetAt = err.detail?.resetAt;
+    if (!Number.isFinite(resetAt)) return false;
+    const at = resetAt + 30_000;
+    const clock = new Date(at).toLocaleTimeString([], { hour: '2-digit', minute: '2-digit' });
+    this.ui.stopSpinner();
+    this.ui.note(`Daily limit reached — ucode carries on by itself at ${clock}. Esc to stop`);
+    await this.persist();
+    let lastNote = Date.now();
+    while (Date.now() < at) {
+      if (this.abort?.signal.aborted) return false;
+      const left = formatDuration(at - Date.now());
+      if (this.full) this.ui.startSpinner(`daily limit · carrying on at ${clock} (in ${left})`);
+      else if (Date.now() - lastNote > 30 * 60_000) { this.ui.note(`still waiting for the daily limit — ${left} to go`); lastNote = Date.now(); }
+      await new Promise((r) => setTimeout(r, Math.min(1000, Math.max(0, at - Date.now()))));
+    }
+    this.ui.stopSpinner();
+    this.ui.note('Daily limit has reset — carrying on');
+    return true;
+  }
+
+  /** A dev server came up during this turn: open it in the browser, once. */
+  /**
+   * Open the running app in a browser — only when asked.
+   *
+   * This used to happen on its own whenever a dev server came up. Something
+   * seizing the screen mid-thought is startling at the best of times, and
+   * during a demo it is worse. UCODE_OPEN=1 brings the old behaviour back for
+   * anyone who liked it; otherwise the URL is on screen to click.
+   */
+  openWhenReady(since) {
+    if (!this.full || process.env.UCODE_OPEN !== '1') return;
+    const server = serversReadySince(since).at(-1);
+    if (!server || (this.opened ??= new Set()).has(server.url)) return;
+    this.opened.add(server.url);
+    const [cmd, args] = process.platform === 'win32'
+      ? ['cmd', ['/c', 'start', '', server.url]]
+      : [process.platform === 'darwin' ? 'open' : 'xdg-open', [server.url]];
+    try {
+      spawn(cmd, args, { stdio: 'ignore', detached: true, windowsHide: true }).unref();
+      this.ui.note(`Opened ${server.url} in your browser`);
+    } catch { /* no browser to open — the link is in the answer */ }
+  }
+
+  cmdStats() {
+    for (const line of statsLines(this.stats, this.working?.length ?? 0)) this.ui.write(line);
+  }
+
+  async cmdDoctor() {
+    this.ui.startSpinner('checking your setup');
+    try {
+      const lines = await runDoctor();
+      this.ui.stopSpinner();
+      for (const line of lines) this.ui.write(line);
+    } finally {
+      this.ui.stopSpinner();
+    }
+  }
+
+  /** /deploy [folder] — the same tool the model calls, run directly. */
+  async cmdDeploy(arg) {
+    const folder = (arg ?? '').trim() || '.';
+    this.ui.toolCall(`Deploying ${folder}`);
+    this.ui.startSpinner('getting ready to deploy');
+    try {
+      const out = await deploy({ folder }, { onOutput: (lines) => this.ui.updateSpinner(lines.at(-1)) });
+      this.ui.stopSpinner();
+      this.ui.toolResult(out.summary);
+      this.ui.write(out.content.split('\n').map((l) => `  ${l}`).join('\n'));
+    } catch (err) {
+      this.ui.stopSpinner();
+      if (err instanceof ToolFailure) this.ui.error(err);
+      else throw err;
+    }
+  }
+
+  reportResult(call, out) {
+    // A build or type check that just passed already verified everything
+    // changed so far; the automatic check at the end would only repeat it.
+    if (call.name === 'run_command' && out.exitCode === 0 &&
+        /(?:next build|npm run build|pnpm (?:run )?build|tsc)/.test(call.args?.command ?? '')) {
+      this.sinceCheck?.clear();
+    }
+    if (!QUIET.has(call.name)) this.ui.toolResult(out.summary);
+    // The change as its two numbers, not as a copy of the file. The diff rows
+    // are still built by the tool — the model reads them in the result — they
+    // simply do not go on screen.
+    if (out.diff?.length) this.ui.diffStat?.(countDiff(out.diff));
+    this.push({ role: 'tool', toolCallId: call.id, name: call.name, content: out.content + this.stuckNote(call, { out }) });
+  }
+
+  /** Show a tool failure, hand it to the model, and say if it was bad arguments. */
+  reportFailure(call, err) {
+    if (!(err instanceof ToolFailure)) throw err;
+
+    // Bad arguments are the model talking to itself. "old_string and new_string
+    // are identical" is a correction it will make on the next step, and it means
+    // nothing to whoever is watching except that something went wrong. It goes
+    // to the model, which can act on it, and not to the screen. A refusal the
+    // user made, and anything that actually failed, still shows.
+    if (err instanceof Declined) this.ui.toolFailed('declined');
+    else if (err.kind !== 'bad_args') this.ui.toolFailed(`${err.kind}: ${err.failed}`);
+    this.push({
+      role: 'tool',
+      toolCallId: call.id,
+      name: call.name,
+      content: err.forModel() + (err instanceof Declined ? '' : this.stuckNote(call, { err })),
+    });
+    return err.kind === 'bad_args';
+  }
+
+  async dispatch(call) {
+    // The model emitted arguments that were not valid JSON. Hand the parser's
+    // own complaint straight back so it can correct itself next step.
+    if (call.parseError) {
+      throw new ToolFailure({
+        kind: 'bad_args',
+        attempted: `calling ${call.name}`,
+        failed: `The arguments were not valid JSON: ${call.parseError}`,
+        fix: `Call ${call.name} again with the arguments as one well-formed JSON object.`,
+      });
+    }
+
+    if (call.name === 'load_skill') return this.loadSkill(call.args?.name);
+    if (call.name === 'update_plan') return this.updatePlan(call.args?.items);
+    if (call.name === 'delegate') return this.delegate(call.args?.tasks);
+
+    // Output reaches the screen as the command produces it, so a slow build is
+    // something you watch rather than something you sit out in silence.
+    return runTool(call.name, call.args ?? {}, {
+      onOutput: (lines) => this.ui.progress(lines),
+    });
+  }
+
+  /**
+   * Switch to the next model after a provider failure. Returns false once
+   * there is nothing sensible left to try. When every model is busy at once,
+   * it waits a minute and goes round again rather than giving up.
+   */
+  async failover(err) {
+    if (++this.failovers > MAX_FAILOVERS) return false;
+    const from = model();
+    let next = fallbackFor(from, this.tried);
+
+    if (!next) {
+      const until = Date.now() + 60_000;
+      this.ui.startSpinner('every model is busy');
+      while (Date.now() < until && !this.abort?.signal.aborted) {
+        this.ui.updateSpinner(`every model is busy — trying again in ${Math.ceil((until - Date.now()) / 1000)}s`);
+        await wait(1000);
+      }
+      this.ui.stopSpinner();
+      if (this.abort?.signal.aborted) return false;
+      this.tried = new Set();
+      next = this.preferred && this.preferred !== from ? this.preferred : fallbackFor(from, this.tried) ?? from;
+    }
+
+    this.tried.add(next);
+    setModel(next);
+    this.cooldownUntil = Date.now() + COOLDOWN;
+    const why = err.kind === 'rate_limit' ? 'busy' : err.kind === 'timeout' ? 'too slow to answer' : 'not answering';
+    this.ui.note(`${modelName(from)} is ${why} — carrying on with ${modelName(next)}`);
+    if (this.full) this.showHeader({ clear: false });
+    return true;
+  }
+
+  /** Run a call and settle to { out } or { err } — never throws. */
+  execute(call) {
+    const started = Date.now();
+    return this.dispatch(call).then(
+      (out) => { trace({ kind: 'tool', name: call.name, ms: Date.now() - started }); countTool(this.stats, call, out); return { out }; },
+      (err) => { trace({ kind: 'tool', name: call.name, ms: Date.now() - started, err: err?.kind }); countTool(this.stats, call, null, err); return { err }; }
+    );
+  }
+
+  updatePlan(items) {
+    const list = (Array.isArray(items) ? items : [])
+      .filter((i) => i && String(i.text ?? '').trim())
+      .slice(0, 6);
+    this.ui.plan(list);
+    const done = list.filter((i) => i.done).length;
+    return { content: `Plan updated: ${done} of ${list.length} done.`, summary: `${done}/${list.length}` };
+  }
+
+  /** File writes from parallel workers take turns, so two never interleave. */
+  fileLock(fn) {
+    const run = (this.lockChain ?? Promise.resolve()).then(fn, fn);
+    this.lockChain = run.catch(() => {});
+    return run;
+  }
+
+  /**
+   * Several workers at once, each its own small agent loop with its own
+   * conversation, sharing the tools, the project, and whatever skills are
+   * already in force. Their lines in the transcript carry their name.
+   */
+  async delegate(tasks) {
+    const list = (Array.isArray(tasks) ? tasks : [])
+      .filter((t) => t && String(t.instructions ?? '').trim())
+      .slice(0, MAX_WORKERS);
+    if (!list.length) {
+      throw new ToolFailure({
+        kind: 'bad_args',
+        attempted: 'starting workers',
+        failed: 'No tasks with instructions were given.',
+        fix: 'Pass tasks as [{ name, instructions }, ...], up to 3.',
+      });
+    }
+
+    const results = await Promise.all(list.map((task, i) => this.runWorker(task, i).catch((err) => ({
+      name: task.name || `worker ${i + 1}`,
+      summary: `Failed: ${err?.failed ?? err?.message ?? err}`,
+      touched: [],
+    }))));
+
+    for (const r of results) for (const f of r.touched) { this.touched.add(f); this.sinceCheck.add(f); }
+
+    // A worker that wrote nothing has not done its part, whatever it said.
+    // The lead builds those itself rather than leaving holes in the app.
+    const empty = results.filter((r) => !r.touched.length).map((r) => r.name);
+
+    return {
+      content: results
+        .map((r) => `## ${r.name}\n${r.summary}\nFiles changed: ${r.touched.join(', ') || 'none'}`)
+        .join('\n\n') +
+        (empty.length
+          ? `\n\n${empty.join(', ')} wrote no files. Build ${empty.length === 1 ? 'that part' : 'those parts'} ` +
+            'yourself now, directly - do not delegate them again.'
+          : ''),
+      summary: results.map((r) => `${r.name} · ${r.touched.length} file${r.touched.length === 1 ? '' : 's'}`).join('  '),
+    };
+  }
+
+  async runWorker(task, index) {
+    const name = clip(String(task.name || `worker ${index + 1}`).trim(), 16);
+    const touched = new Set();
+    const skills = this.skills
+      .filter((s) => this.loaded.has(s.name))
+      .map((s) => `--- ${s.name} ---\n${s.body}`)
+      .join('\n\n');
+    const messages = [
+      { role: 'system', content: workerPrompt({ cwd: this.cwd, name, memory: this.memory, skills, map: this.map }) },
+      { role: 'user', content: String(task.instructions) },
+    ];
+    const available = this.toolsNow().filter((t) => !WORKER_EXCLUDED.has(t.name));
+    const wanted = process.env.UCODE_WORKER_MODEL;
+    let workerModel = wanted && MODELS[wanted] ? wanted : model();
+    const tried = new Set([workerModel]);
+    let failovers = 0;
+
+    // Start a moment apart. Three requests in the same instant is exactly what
+    // trips a free endpoint's rate limit, and the stagger costs a second or two.
+    if (index) await wait(index * 1500);
+
+    for (let step = 0; step < WORKER_STEPS; step++) {
+      if (this.abort?.signal.aborted) break;
+
+      let reply;
+      const asked = Date.now();
+      try {
+        reply = await ask(messages, available, { signal: this.abort?.signal, model: workerModel });
+      } catch (err) {
+        // Same rule as the lead: a busy model is swapped, not a reason to stop.
+        if (passing(err) && failovers < 6 && !this.abort?.signal.aborted) {
+          failovers++;
+          let next = fallbackFor(workerModel, tried);
+          if (!next) { tried.clear(); await wait(20_000); next = fallbackFor(workerModel, tried) ?? workerModel; }
+          tried.add(next);
+          this.ui.note(`${name}: ${modelName(workerModel)} is busy — switching to ${modelName(next)}`);
+          workerModel = next;
+          step--;
+          continue;
+        }
+        throw err;
+      }
+      this.record(reply.usage);
+      trace({
+        kind: 'model', who: name, model: workerModel, ms: Date.now() - asked,
+        in: reply.usage.promptTokens, out: reply.usage.outputTokens,
+        calls: reply.toolCalls.map((c) => c.name),
+      });
+
+      if (!reply.toolCalls.length) {
+        this.ui.toolResult(`${name} finished`);
+        return { name, summary: reply.text?.trim() || 'Finished without a summary.', touched: [...touched] };
+      }
+
+      messages.push({ role: 'assistant', content: reply.text || '', toolCalls: reply.toolCalls });
+      for (const call of reply.toolCalls) {
+        this.ui.toolCall(`${name} › ${describe(call.name, call.args)}`);
+        if (FILE_WRITES.has(call.name)) for (const p of pathsOf(call)) touched.add(p);
+        const { out, err } = FILE_WRITES.has(call.name)
+          ? await this.fileLock(() => this.execute(call))
+          : await this.execute(call);
+        if (err) {
+          if (!(err instanceof ToolFailure)) throw err;
+          this.ui.toolFailed(`${name}: ${err.kind}: ${err.failed}`);
+        }
+        messages.push({
+          role: 'tool', toolCallId: call.id, name: call.name,
+          content: err ? err.forModel() : out.content,
+        });
+      }
+    }
+
+    return { name, summary: `Stopped after ${WORKER_STEPS} steps without finishing.`, touched: [...touched] };
+  }
+
+  /**
+   * Check the code files changed since the last check, and return the
+   * errors as text for the model — or null when everything is clean.
+   *
+   * The check is incremental: TypeScript writes what it learned to a build
+   * info file, so the second check onwards reads that instead of retyping
+   * every dependency — seconds rather than the best part of a minute. The
+   * file sits in node_modules/.cache, which is already ignored by git and is
+   * deliberately left out of the starter package cache.
+   *
+   * TypeScript projects get one `tsc --noEmit` per project that owns a
+   * changed file (an app scaffolded into a subfolder is its own project).
+   * Plain JavaScript gets a syntax check, Python a compile check. Nothing
+   * runs that is not already installed.
+   */
+  async autoCheck() {
+    const changed = [...this.sinceCheck].filter((f) => CHECKABLE.test(f));
+    this.sinceCheck.clear();
+    if (!changed.length) return null;
+
+    const root = path.resolve(this.cwd);
+    const tsRoots = new Set();
+    const singles = [];
+    const problems = [];
+
+    for (const rel of changed) {
+      const abs = path.resolve(root, rel);
+      if (!(await exists(abs))) continue;
+      if (/\.py$/i.test(rel)) { singles.push({ abs, rel, command: `python -m py_compile "${abs}"` }); continue; }
+      // A single-file app keeps all its logic in an inline <script>, which no
+      // other check here ever looks at.
+      if (/\.html?$/i.test(rel)) {
+        const text = await readFile(abs, 'utf8').catch(() => null);
+        const bad = text === null ? [] : checkHtml(text);
+        if (bad.length) {
+          problems.push(`${rel} — the script in this page does not parse, so none of it runs:\n` +
+            bad.map((b) => `  line ${b.line}: ${b.message}`).join('\n'));
+        }
+        continue;
+      }
+      let dir = path.dirname(abs);
+      let owner = null;
+      while (dir.startsWith(root)) {
+        if (await exists(path.join(dir, 'tsconfig.json'))) { owner = dir; break; }
+        const up = path.dirname(dir);
+        if (up === dir) break;
+        dir = up;
+      }
+      if (owner && (await exists(path.join(owner, 'node_modules', 'typescript')))) tsRoots.add(owner);
+      else if (/\.[cm]?js$/i.test(rel)) singles.push({ abs, rel, command: `node --check "${abs}"` });
+    }
+
+    const check = async (label, command, cwd) => {
+      this.ui.toolCall(label);
+      this.ui.startSpinner(label);
+      const { out, err } = await this.execute({
+        id: 'check', name: 'run_command',
+        args: { command, cwd: path.relative(root, cwd) || '.', timeout_ms: 180_000 },
+      });
+      this.ui.stopSpinner();
+      return err ? { exitCode: -1, content: String(err.failed ?? err.message) } : out;
+    };
+
+    for (const dir of tsRoots) {
+      const show = path.relative(root, dir) || '.';
+      await mkdir(path.join(dir, path.dirname(TSBUILDINFO)), { recursive: true }).catch(() => {});
+      let out = await check(`Checking types in ${show}`, typeCheckCommand(true), dir);
+      // Older TypeScript refuses --incremental alongside --noEmit. Say so once
+      // by simply checking again the slow way, rather than failing the edit.
+      if (out.exitCode !== 0 && NO_INCREMENTAL.test(out.content)) {
+        out = await check(`Checking types in ${show}`, typeCheckCommand(false), dir);
+      }
+      if (out.exitCode === 0) { this.ui.toolResult('types check out'); continue; }
+      const errors = out.content.split('\n').filter((l) => /error TS\d+/.test(l));
+      this.ui.toolFailed(`${errors.length || 'some'} type error${errors.length === 1 ? '' : 's'}`);
+      problems.push(`In ${show} (tsc --noEmit):\n${(errors.length ? errors : out.content.split('\n')).slice(0, 40).join('\n')}`);
+    }
+
+    for (const f of singles) {
+      const out = await check(`Checking ${f.rel}`, f.command, root);
+      if (out.exitCode === 0) { this.ui.toolResult('ok'); continue; }
+      this.ui.toolFailed('does not compile');
+      problems.push(`${f.rel}:\n${out.content.split('\n').slice(0, 20).join('\n')}`);
+    }
+
+    // Only once it compiles: a failing test on code that does not build tells
+    // the model nothing it does not already know from the errors above.
+    if (!problems.length) {
+      const failed = await this.runRelatedTests(root, changed);
+      if (failed) problems.push(failed);
+    }
+
+    const live = await this.liveErrors();
+    if (live) problems.push(live);
+
+    return problems.length ? problems.join('\n\n') : null;
+  }
+
+  /**
+   * Anything the running app has complained about since the last look. A dev
+   * server knows about a broken import the moment it happens; without this
+   * nobody reads that until a build, or until the user says the page is blank.
+   */
+  async liveErrors() {
+    try {
+      const found = await this.logWatch.since(runningServers());
+      if (found) this.ui.toolFailed('the running app reported an error');
+      return found;
+    } catch {
+      return null; // reading a log must never be what breaks a turn
+    }
+  }
+
+  /**
+   * Run the tests that reach the files just changed, and return their
+   * failures as text — or null when they pass, or when this project has no
+   * runner that can be asked which tests matter.
+   */
+  async runRelatedTests(root, changed) {
+    const runner = await testRunnerFor(root);
+    if (!runner) return null;
+
+    const existing = [];
+    for (const rel of changed) if (await exists(path.resolve(root, rel))) existing.push(rel);
+    const command = relatedCommand(runner, existing);
+    if (!command) return null;
+
+    const label = `Running the ${runner} tests that cover this`;
+    this.ui.toolCall(label);
+    this.ui.startSpinner(label);
+    const { out, err } = await this.execute({
+      id: 'tests', name: 'run_command',
+      args: { command, cwd: '.', timeout_ms: 180_000 },
+    });
+    this.ui.stopSpinner();
+
+    if (err) { this.ui.toolResult('tests skipped'); return null; }
+    if (out.exitCode === 0) { this.ui.toolResult('tests pass'); return null; }
+
+    // A runner that is not installed is not a failing test; npx says so.
+    if (/could not determine executable|not found|Cannot find module/i.test(out.content)) {
+      this.ui.toolResult('no test runner installed');
+      return null;
+    }
+
+    this.ui.toolFailed('tests fail');
+    return `The tests covering your change fail (${runner}):\n${summariseFailures(runner, out.content)}`;
+  }
+
+  loadSkill(name) {
+    const skill = findSkill(this.skills, name);
+    if (!skill) {
+      // The available names go in `failed` rather than only in `fix`: the
+      // transcript shows the failure line, and a bare "no such skill" leaves
+      // the user guessing at what this session actually has.
+      const available = this.skills.map((s) => s.name).join(', ') || '(none)';
+      throw new ToolFailure({
+        kind: 'no_such_skill',
+        attempted: `loading the "${name}" skill`,
+        failed: `There is no skill called "${name}". This session has: ${available}.`,
+        fix: 'Use one of those names, or carry on without one.',
+      });
+    }
+
+    if (this.loaded.has(skill.name)) {
+      return { content: `The "${skill.name}" skill is already loaded above. Follow it.`, summary: 'already loaded' };
+    }
+
+    this.loaded.add(skill.name);
+    this.push(skillMessage(skill));
+    return {
+      content: `Loaded "${skill.name}". Its instructions are in your context now — follow them.`,
+      summary: `${skill.name} · ${skill.body.split('\n').length} lines`,
+    };
+  }
+
+  record(u) {
+    const total = this.session.usage;
+    total.promptTokens += u.promptTokens || 0;
+    total.outputTokens += u.outputTokens || 0;
+    total.totalTokens += u.totalTokens || 0;
+    total.turns += 1;
+  }
+
+  /** Fold older turns into a summary when the window gets tight. */
+  async maybeFold() {
+    const limit = contextLimit();
+    if (!tooBig(this.working, limit)) return;
+
+    this.ui.startSpinner('context is filling up — summarizing earlier turns');
+    try {
+      const result = await fold(this.working, {
+        limit,
+        summarize: async (older) => {
+          const reply = await ask(
+            [
+              { role: 'system', content: SUMMARY_PROMPT },
+              { role: 'user', content: forSummary(older) },
+            ],
+            [],
+            { signal: this.abort?.signal, temperature: 0 }
+          );
+          return reply.text;
+        },
+      });
+
+      this.ui.stopSpinner();
+      if (result.folded) {
+        this.working = result.messages;
+        this.ui.note(
+          `folded ${result.droppedCount} earlier messages into a summary ` +
+          '(the full history is still saved in this session)'
+        );
+      }
+    } catch (err) {
+      // If summarizing fails, carry on with the full history and let the API
+      // complain — better than silently throwing away the conversation.
+      this.ui.stopSpinner();
+      this.ui.note(`could not summarize older turns (${err.kind ?? 'error'}); carrying on uncompacted`);
+    }
+  }
+
+  // -- slash commands ------------------------------------------------------
+
+  async command(input) {
+    const [name, ...rest] = input.split(/\s+/);
+    const arg = rest.join(' ').trim();
+
+    switch (name) {
+      case '/help':
+        return this.cmdHelp();
+
+      // One thing, one command, however you happen to spell it.
+      case '/model':
+      case '/models':
+        return this.cmdModel(arg);
+
+      case '/session':
+      case '/sessions':
+      case '/resume':
+        return this.cmdSessions(arg);
+
+      case '/new':      return this.cmdNew();
+      case '/remember': return this.cmdRemember(arg);
+      case '/skills':   return this.cmdSkills();
+      case '/clear':    this.showHeader(); return;
+      case '/search':   return this.cmdSearch(arg);
+      case '/copy':     return this.cmdCopy();
+      case '/stats':    return this.cmdStats();
+      case '/doctor':   return this.cmdDoctor();
+      case '/look':     return this.cmdLook(arg);
+      case '/deploy':   return this.cmdDeploy(arg);
+      case '/exit':
+      case '/quit':     return 'exit';
+
+      default:
+        this.ui.write(theme.warn(`  no such command: ${name}`));
+        this.ui.note('/help lists them.');
+    }
+  }
+
+  /**
+   * Look at the running app, because the user asked to.
+   *
+   * This used to happen on its own, which meant a browser being driven while
+   * someone was reading, and a window taking the screen mid-thought. It is
+   * the same check as before; the difference is who starts it.
+   */
+  async cmdLook(url) {
+    const { lookAtApp } = await import('../tools/browser.js');
+    const server = runningServers().at(-1);
+    const at = (url ?? '').trim() || server?.url;
+    if (!at) {
+      this.ui.write(theme.warn('  nothing is running to look at.'));
+      this.ui.note('start the app first, or pass a URL: /look http://localhost:3000');
+      return;
+    }
+    this.ui.toolCall(`Looking at ${at}`);
+    try {
+      const out = await lookAtApp({ url: at });
+      this.ui.write(out.content);
+      // The model gets it too, so the next thing it says is about what is
+      // actually on the page rather than what it believes it built.
+      this.push({ role: 'user', content: `I looked at ${at}. This is what is there:
+
+${out.content}` });
+    } catch (err) {
+      this.ui.write(theme.error(`  ${err.failed ?? err.message}`));
+    }
+  }
+
+  cmdHelp() {
+    const rows = [
+      ['/help', 'this list'],
+      ['/stats', 'time, steps and tokens this session'],
+      ['/doctor', 'check that everything ucode needs is working'],
+      ['/look [url]', 'open the running app and report what is on the page'],
+      ['/deploy [folder]', 'put the app online and get its link'],
+      ['/model', 'show the models and switch between them'],
+      ['/resume', 'pick up an earlier conversation'],
+      ['/new', 'save this one and start fresh'],
+      ['/remember <note>', `add a standing note to ${MEMORY_FILE}`],
+      ['/skills', 'what ucode knows how to do'],
+      ['/search <query>', 'look something up on the web'],
+      ['/copy', 'copy the last reply to the clipboard'],
+      ['/clear', 'clear the screen, keep the conversation'],
+      ['/exit', 'save and quit'],
+    ];
+
+    this.ui.blank();
+    for (const [command, what] of rows) {
+      this.ui.write(`  ${blue(command.padEnd(18))} ${dim(what)}`);
+    }
+    this.ui.blank();
+    this.ui.write(dim('  /models, /session and /sessions do the same as /model and /resume.'));
+    this.ui.write(dim('  ctrl+b swaps plan and build · esc stops a running turn · ctrl+d quits'));
+    this.ui.blank();
+  }
+
+  /** The five models, and this session's spend. */
+  async cmdModel(arg) {
+    if (arg) {
+      try {
+        setModel(arg);
+        this.preferred = model();
+      } catch (err) {
+        this.ui.error(err, { debug: this.debug });
+        return;
+      }
+      this.session.model = model();
+      this.ui.note(`now using ${modelName()}`);
+      this.showHeader({ clear: false });
+      return;
+    }
+
+    const all = modelList();
+    const width = Math.max(...all.map((m) => m.name.length));
+
+    if (this.ui.pick) {
+      const items = all.map((m) => ({
+        label:
+          `${m.active ? blue('●') : dim('○')} ${m.star ? blue('★') : ' '} ` +
+          `${m.name.padEnd(width)}  ${dim(`${formatTokens(m.context)} · ${m.note}`)}`,
+      }));
+
+      const chosen = await this.ui.pick(items, {
+        active: Math.max(0, all.findIndex((m) => m.active)),
+        hint: '↑↓ move · enter to switch · esc to cancel',
+      });
+      if (chosen === null) return;
+
+      setModel(all[chosen].id);
+      this.preferred = model();
+      this.session.model = model();
+      this.ui.note(`now using ${modelName()}`);
+      this.showHeader({ clear: false });
+      return;
+    }
+
+    this.ui.blank();
+    for (const m of all) {
+      this.ui.write(
+        `  ${m.active ? blue('●') : dim('○')} ${m.star ? blue('★') : ' '} ` +
+        `${(m.active ? blue : dim)(m.name.padEnd(width))}  ${dim(`${formatTokens(m.context)} · ${m.note}`)}`
+      );
+      this.ui.write(`      ${dim(m.id)}`);
+    }
+
+    const u = this.session.usage;
+    const live = rateLimits();
+    this.ui.blank();
+    this.ui.write(
+      `  ${dim('this session')}  ${u.turns} turns · ${formatTokens(u.totalTokens)} tokens ` +
+      `(${formatTokens(u.promptTokens)} in, ${formatTokens(u.outputTokens)} out)`
+    );
+    if (live?.requestsRemaining != null && live?.requestsLimit) {
+      this.ui.write(`  ${dim('requests')}      ${live.requestsRemaining} of ${live.requestsLimit} left`);
+    }
+    this.ui.blank();
+    this.ui.write(dim('  /model <id> switches without the picker.'));
+    this.ui.blank();
+  }
+
+  /**
+   * One row per saved conversation.
+   *
+   * A list of titles and timestamps is not enough to recognise your own work
+   * by — half of them start "Fix the". So each row carries what it was about
+   * and how far it got, and the ones from this folder are marked, because that
+   * is nearly always the one being looked for.
+   */
+  describeSession(s, width, i) {
+    const room = Math.max(24, Math.min(46, width - 34));
+    const mark = s.mine ? blue('●') : dim('○');
+    const when = relativeTime(s.updatedAt).padEnd(9);
+    const turns = `${s.turns} turn${s.turns === 1 ? '' : 's'}`.padEnd(9);
+    const where = s.mine ? 'here' : shortenPath(s.cwd, 26);
+
+    return {
+      // Numbered in the picker, so /session delete 3 has something to point at.
+      label: `${i === undefined ? '' : `${dim(String(i + 1).padStart(2))} `}${mark} ${clip(s.title, room).padEnd(room)}  ${dim(when)}${dim(turns)}${dim(where)}`,
+      sub: s.preview ? dim(`     ${clip(s.preview, width - 10)}`) : '',
+    };
+  }
+
+  async cmdSessions(arg) {
+    if (arg === '--clear' || arg === 'clear') {
+      const yes = await this.ui.confirm({
+        action: 'delete every saved conversation',
+        detail: 'This cannot be undone.',
+        risk: 'write',
+      });
+      if (!yes) {
+        this.ui.note('cancelled');
+        return;
+      }
+      await removeAll();
+      this.ui.note('all sessions deleted');
+      return;
+    }
+
+    // /session delete 3   or   /session delete 2,5,7
+    const del = /^(?:delete|del|rm|remove)\b\s*(.*)$/i.exec(arg ?? '');
+    if (del) return this.deleteSessions(del[1]);
+
+    const sessions = await list({ cwd: this.cwd });
+    if (!sessions.length) {
+      this.ui.note('no saved conversations yet');
+      return;
+    }
+
+    for (const bad of sessions.unreadable ?? []) {
+      this.ui.write(theme.warn(`  could not read session file: ${bad}`));
+    }
+
+    const shown = sessions.slice(0, 25);
+    const width = this.ui.width ? this.ui.width() : 80;
+    let index;
+
+    if (arg) {
+      const n = Number(arg);
+      if (!Number.isInteger(n) || n < 1 || n > shown.length) {
+        this.ui.write(theme.warn(`  "${arg}" is not one of 1-${shown.length}`));
+        return;
+      }
+      index = n - 1;
+    } else if (this.ui.pick) {
+      // The picker stays open while you delete, so clearing out several old
+      // conversations is d d, d d, d d — then Enter on the one you want.
+      let active = 0;
+      for (;;) {
+        const here = shown.filter((s) => s.mine).length;
+        const picked = await this.ui.pick(
+          shown.map((s, i) => this.describeSession(s, width, i)),
+          {
+            active,
+            deletable: true,
+            hint:
+              `↑↓ move · enter to continue · d twice to delete · esc to cancel` +
+              (here ? ` — ${here} from this folder` : ''),
+          }
+        );
+        if (picked === null) return;
+        if (typeof picked === 'object' && picked.delete !== undefined) {
+          const doomed = shown[picked.delete];
+          active = picked.delete;
+          if (doomed.id === this.session.id) {
+            this.ui.flash?.('that is the conversation you are in — /new first, then delete it');
+            continue;
+          }
+          await remove(doomed.id);
+          shown.splice(picked.delete, 1);
+          this.ui.flash?.(`deleted · ${clip(doomed.title, 50)}`);
+          if (!shown.length) {
+            this.ui.note('no saved conversations left');
+            return;
+          }
+          active = Math.min(active, shown.length - 1);
+          continue;
+        }
+        index = picked;
+        break;
+      }
+    } else {
+      this.ui.blank();
+      this.ui.note('/session delete <number> removes one, or several: /session delete 2,5');
+      index = await this.ui.choose(
+        'continue which?',
+        shown.map((s, i) => this.describeSession(s, width, i).label)
+      );
+      if (index === null) return;
+    }
+
+    if (this.session.messages.length) await this.persist();
+    if (await this.resume(shown[index].id)) {
+      this.showHeader();
+      this.replayTail();
+    }
+  }
+
+  /** /session delete 3, or 2,5,7 — numbers as the session list shows them. */
+  async deleteSessions(spec) {
+    const sessions = (await list({ cwd: this.cwd })).slice(0, 25);
+    const numbers = [...new Set(String(spec).split(/[\s,]+/).filter(Boolean).map(Number))];
+    const bad = numbers.filter((n) => !Number.isInteger(n) || n < 1 || n > sessions.length);
+    if (!numbers.length || bad.length) {
+      this.ui.write(theme.warn(`  usage: /session delete <number>[,<number>…] — numbers from 1 to ${sessions.length}`));
+      return;
+    }
+    for (const n of numbers) {
+      const s = sessions[n - 1];
+      if (s.id === this.session.id) {
+        this.ui.note(`skipped ${n} — that is the conversation you are in`);
+        continue;
+      }
+      await remove(s.id);
+      this.ui.note(`deleted ${n} · ${s.title}`);
+    }
+  }
+
+  async resume(id) {
+    try {
+      const loaded = await load(id);
+      this.session = loaded;
+      this.working = [...loaded.messages];
+      this.loaded = new Set(loaded.messages.filter((m) => m.skill).map((m) => m.skill));
+      if (loaded.model && MODELS[loaded.model]) setModel(loaded.model);
+      return true;
+    } catch (err) {
+      this.ui.error(err, { debug: this.debug });
+      this.ui.note('Starting a fresh one instead.');
+      return false;
+    }
+  }
+
+  /** The last few exchanges, so a resumed conversation has visible context. */
+  replayTail(count = 4) {
+    const tail = this.session.messages
+      .filter((m) => (m.role === 'user' || m.role === 'assistant') && m.content)
+      .slice(-count);
+
+    for (const m of tail) {
+      if (m.role !== 'user') this.ui.assistant(m.content);
+      else if (this.ui.userMessage) this.ui.userMessage(m.content);
+      else this.ui.write(`${blue('›')} ${dim(m.content.split('\n')[0])}`);
+    }
+    if (tail.length) this.ui.write(dim('  ── picking up here ──\n'));
+  }
+
+  /** Add a line to this project's UCODE.md, read at the start of every turn. */
+  async cmdRemember(note) {
+    if (!note) {
+      this.ui.note(`usage: /remember <something ucode should always know here> — saved to ${MEMORY_FILE}`);
+      return;
+    }
+    try {
+      const file = await remember(this.cwd, note);
+      this.ui.note(`remembered · ${path.relative(this.cwd, file) || MEMORY_FILE}`);
+    } catch (err) {
+      this.ui.error(new Failure({
+        kind: 'memory_unwritable',
+        attempted: `saving to ${MEMORY_FILE}`,
+        failed: err.message,
+        fix: 'Check that this folder is writable.',
+      }), { debug: this.debug });
+    }
+  }
+
+  async cmdNew() {
+    if (this.session.messages.length) {
+      await this.persist();
+      this.ui.note(`saved · ${this.session.title}`);
+    }
+    this.session = newSession(this.cwd, model());
+    this.working = [];
+    this.loaded = new Set();
+    this.showHeader();
+  }
+
+  async cmdSkills() {
+    // Re-read from disk. Skills load once at startup, so one written during
+    // this session would otherwise stay invisible — and load_skill would fail
+    // on a name the user can see in the folder.
+    this.skills = await loadSkills({ cwd: this.cwd });
+    for (const problem of this.skills.problems ?? []) {
+      this.ui.write(theme.warn(`  skill not loaded: ${problem}`));
+    }
+
+    if (!this.skills.length) {
+      this.ui.note('no skills found — add a folder with a SKILL.md under .ucode/skills');
+      return;
+    }
+
+    this.ui.blank();
+    for (const s of this.skills) {
+      const live = this.loaded.has(s.name);
+      const auto = s.triggers.length ? dim('  · loads itself') : '';
+      this.ui.write(`  ${live ? blue('●') : dim('○')} ${blue(s.name)}${auto}`);
+      this.ui.write(`    ${dim(s.description)}`);
+    }
+    this.ui.blank();
+    this.ui.write(dim('  ● already loaded here · ucode pulls one in when the task matches'));
+    this.ui.blank();
+  }
+
+  async cmdSearch(query) {
+    if (!query) {
+      this.ui.note('usage: /search <what you want to look up>');
+      return;
+    }
+    await this.turn(
+      `Search the web for: ${query}\n\nUse web_search, then summarise what you found and cite the URLs.`
+    );
+  }
+
+  /** Copy the last reply. Every platform ships a clipboard pipe. */
+  async cmdCopy() {
+    const last = [...this.session.messages]
+      .reverse()
+      .find((m) => m.role === 'assistant' && m.content?.trim());
+
+    if (!last) {
+      this.ui.note('nothing to copy yet');
+      return;
+    }
+
+    const tool = process.platform === 'win32' ? 'clip'
+      : process.platform === 'darwin' ? 'pbcopy'
+      : 'xclip -selection clipboard';
+
+    try {
+      await new Promise((resolve, reject) => {
+        const child = spawn(tool, { shell: true, windowsHide: true });
+        child.on('error', reject);
+        child.on('close', (code) => (code === 0 ? resolve() : reject(new Error(`exit ${code}`))));
+        child.stdin.end(last.content);
+      });
+      const lines = last.content.split('\n').length;
+      this.ui.note(`copied ${lines} line${lines === 1 ? '' : 's'}`);
+    } catch (err) {
+      this.ui.error(new Failure({
+        kind: 'clipboard_failed',
+        attempted: 'copying the last reply',
+        failed: `${tool} could not run: ${err.message}`,
+        fix: process.platform === 'linux'
+          ? 'Install xclip (apt install xclip), or select the text with the mouse.'
+          : 'Select the text with the mouse instead.',
+      }), { debug: this.debug });
+    }
+  }
+}
+
+export { DEFAULT_MODEL, PROVIDER };