claude-usage-limits 1.18.0 → 1.23.0

package/skills/usage-limits/scripts/mode.js

@@ -0,0 +1,1637 @@
+#!/usr/bin/env node
+'use strict';
+
+// Budget modes: how hard the plugin leans, and what it costs to lean.
+//
+// The plugin is not free. It puts a line into the context before every prompt,
+// refreshes readings after tool calls, and keeps a status line alive. A mode
+// called "save tokens" that still injects four hundred tokens of advice per
+// turn is not saving anything - it is charging for the advice about saving.
+//
+// So a mode changes two things, never one:
+//
+//   1. What the plugin TELLS the agent to do (the directive).
+//   2. What the plugin COSTS to say it (verbosity, cadence, silence).
+//
+// Both halves are in the table below, and there is a test for the second one:
+// the `max` line must never be longer than the `standard` line for the same
+// reading.
+//
+// What a mode must NEVER do is lower the quality of the work. When things are
+// tight you change the ORDER of the work, not the amount or the quality. The
+// efficient modes buy their savings from ceremony - speculative reads,
+// re-reads, preamble, subagents nobody needed, workflows that cost more
+// context than they save - and never from doing the job worse. The directives
+// say that outright, because a model reading "use fewer tokens" will otherwise
+// quietly decide to skip the hard part.
+//
+// It also never writes settings.json. See "two planes" below.
+
+const fs = require('fs');
+const os = require('os');
+const path = require('path');
+
+const host = require('./host.js');
+const codex = require('./codex.js');
+
+// ---------------------------------------------------------------------------
+// The policy table
+//
+// Each mode is a record other scripts READ. Nothing downstream hardcodes a
+// mode name in a conditional beyond reading these fields, so a fifth mode is
+// one entry here and nothing else.
+//
+// Two kinds of field live here and --explain keeps them apart, because a
+// record that lists both as though they were the same thing is a saving that
+// exists only where it is described. WIRED names the fields code actually
+// reads; the rest are the mode's stance, and the only way they reach anything
+// is by being restated in the directive prose below, which means they say
+// nothing at all in the two modes that have no directive.
+//
+// Two fields have been deleted rather than sorted: `statusline` and
+// `returnToBaseline`. Nothing read either of them, nothing said either of
+// them, and `statusline: compact` was not merely inert but backwards - the
+// status line is fourteen characters LONGER in `max`, because feed.js adds a
+// "budget max" token there and nothing consumes a compact flag.
+
+const MODES = {
+  max: {
+    name: 'max',
+    summary: 'fewest tokens that can still finish the job',
+    // One line, no table, no per-model rows.
+    briefStyle: 'terse',
+    // Say nothing when nothing a decision depends on has moved. The pressure
+    // is part of that digest, so the wall always speaks.
+    briefWhenUnchanged: false,
+    // The reading itself costs IO and tokens, so it is taken less often.
+    refreshSeconds: 600,
+    // The mid-turn cadence is a different number from the prompt-time one and
+    // always has been: the brief refreshes a reading it is about to print,
+    // while the pulse interrupts work in progress. They are listed separately
+    // so `standard` can keep BOTH of today's numbers rather than accidentally
+    // slowing the pulse to the brief's cadence.
+    pulseSeconds: 600,
+    recheckSeconds: 0,
+    directive: 'max',
+    // Claude runs the cheapest tier that can still do the job.
+    selfSwitch: 'active',
+    subagents: 'avoid',
+    workflows: 'off',
+    // How much emptier another window has to be before a switch is worth
+    // naming. Aggressive modes act on a smaller improvement; see escapeRoute.
+    switchGapPoints: 5,
+  },
+  high: {
+    name: 'high',
+    summary: 'full capability, continuously re-costed',
+    briefStyle: 'normal',
+    briefWhenUnchanged: true,
+    // The explicit two minutes.
+    refreshSeconds: 120,
+    pulseSeconds: 120,
+    // The mid-turn re-cost. See pulse.js.
+    recheckSeconds: 120,
+    directive: 'high',
+    // DOWN when the work is mechanical, back UP when it is not.
+    selfSwitch: 'balanced',
+    subagents: 'sized',
+    workflows: 'when-cheaper',
+    switchGapPoints: 5,
+  },
+  standard: {
+    name: 'standard',
+    summary: 'what the plugin does today',
+    briefStyle: 'normal',
+    briefWhenUnchanged: true,
+    // Today's default, unchanged.
+    refreshSeconds: 180,
+    // Today's pulse interval, also unchanged.
+    pulseSeconds: 120,
+    recheckSeconds: 0,
+    directive: null,
+    // Today's behaviour: switch rather than stop, at the wall.
+    selfSwitch: 'at-pressure',
+    subagents: 'allowed',
+    workflows: 'allowed',
+    switchGapPoints: 10,
+  },
+  off: {
+    name: 'off',
+    summary: 'injects nothing, hooks return immediately',
+    // Inject nothing at all.
+    briefStyle: 'none',
+    briefWhenUnchanged: true,
+    // Hooks short-circuit before any reading.
+    refreshSeconds: 0,
+    pulseSeconds: 0,
+    recheckSeconds: 0,
+    directive: null,
+    // Whatever the user set stands, untouched.
+    selfSwitch: 'never',
+    subagents: 'untouched',
+    workflows: 'untouched',
+    switchGapPoints: 10,
+  },
+};
+
+const ORDER = ['max', 'high', 'standard', 'off'];
+const DEFAULT_MODE = 'standard';
+
+// The fields some other script reads, and where. Kept beside the table so a
+// new field cannot be added and quietly reported as behaviour: if it is not
+// here, --explain says outright that nothing reads it.
+const WIRED = {
+  briefStyle: 'brief.js: how much of the line is said, and "none" for silence',
+  briefWhenUnchanged: 'brief.js: whether to repeat a line nothing has moved',
+  refreshSeconds: 'brief.js: how old a reading may be before it is retaken; 0 short-circuits the hook',
+  pulseSeconds: 'pulse.js: how often the mid-turn hook does real work',
+  recheckSeconds: 'pulse.js: how often the mid-turn re-cost may speak',
+  switchGapPoints: 'usage.js: how much emptier another window must be before a switch is named',
+};
+
+// ---------------------------------------------------------------------------
+// Names
+//
+// "normal" is the trap. In the vocabulary this was asked in, "normal" means
+// the FOURTH mode - the plugin ignored. To almost everyone else it means the
+// THIRD - the plugin working as usual. A silent wrong guess picks the opposite
+// of what was asked, so `normal` is an alias for neither: it is accepted and
+// answered with a disambiguation.
+const ALIASES = {
+  ultra: 'max',
+  ultraefficient: 'max',
+  'ultra-efficient': 'max',
+  maxtoken: 'max',
+  'max-token': 'max',
+  maxefficient: 'max',
+  maxefficiency: 'max',
+  'max-efficient': 'max',
+  highefficient: 'high',
+  'high-efficient': 'high',
+  'high-efficiency': 'high',
+  highefficiency: 'high',
+  smart: 'high',
+  tokenefficient: 'standard',
+  'token-efficient': 'standard',
+  efficient: 'standard',
+  default: 'standard',
+  on: 'standard',
+  none: 'off',
+  ignore: 'off',
+  quiet: 'off',
+  silent: 'off',
+};
+
+const AMBIGUOUS = {
+  normal: ['standard', 'off'],
+};
+
+function ambiguityText(word) {
+  const choices = AMBIGUOUS[word];
+  if (!choices) return null;
+  return (
+    word + ' is ambiguous here. Did you mean:\n' +
+    '  standard  the plugin working as usual\n' +
+    '  off       the plugin stays out of the way'
+  );
+}
+
+// Returns { mode } for a name that resolves, { ambiguous, message } for one
+// that deliberately does not, and null for a word that is not a mode at all.
+// Three outcomes rather than two, because "I will not guess" is an answer.
+function normalise(value) {
+  const word = String(value === undefined || value === null ? '' : value).trim().toLowerCase();
+  if (!word) return null;
+  if (AMBIGUOUS[word]) return { ambiguous: AMBIGUOUS[word].slice(), message: ambiguityText(word) };
+  if (MODES[word]) return { mode: word };
+  if (ALIASES[word]) return { mode: ALIASES[word], alias: word };
+  if (word === 'auto') return { auto: true };
+  return null;
+}
+
+// ---------------------------------------------------------------------------
+// Tiers, for the user's bounds
+//
+// The canonical effort ladder the host accepts is low|medium|high|xhigh|max.
+// "ultracode" is not a sixth level: it is xhigh plus standing dynamic workflow
+// orchestration, set through its own settings key, so it ranks with xhigh.
+//
+// Two guards the modes have to respect and do not get to argue with:
+//   - xhigh and max are refused outright when thinking is disabled.
+//   - a model bound cannot be enforced against a model nobody has ranked, so
+//     an unknown name is reported rather than silently treated as the floor.
+const EFFORT_ORDER = ['low', 'medium', 'high', 'xhigh', 'max'];
+const THINKING_ONLY_EFFORTS = ['xhigh', 'max'];
+
+// Cheapest first. The order is the reverse of usage.js's FAMILIES, which is
+// the account's own ordering of these families, so the two cannot drift into
+// disagreeing about which way is "down".
+const MODEL_ORDER = ['haiku', 'sonnet', 'opus', 'mythos', 'fable'];
+
+function effortRank(value) {
+  const name = String(value || '').trim().toLowerCase();
+  if (name === 'ultracode') return EFFORT_ORDER.indexOf('xhigh');
+  const at = EFFORT_ORDER.indexOf(name);
+  return at === -1 ? null : at;
+}
+
+function modelRank(value) {
+  const name = String(value || '').trim().toLowerCase();
+  for (let i = 0; i < MODEL_ORDER.length; i += 1) {
+    if (name.indexOf(MODEL_ORDER[i]) !== -1) return i;
+  }
+  return null;
+}
+
+// "sonnet/medium", "sonnet", "medium" - whichever half the user gave.
+function parseTier(value) {
+  const text = String(value || '').trim().toLowerCase();
+  if (!text) return null;
+  const parts = text.split('/').map((part) => part.trim()).filter(Boolean);
+  const tier = { model: null, effort: null };
+  for (const part of parts) {
+    if (effortRank(part) !== null) tier.effort = part === 'ultracode' ? 'xhigh' : part;
+    else if (modelRank(part) !== null) tier.model = MODEL_ORDER[modelRank(part)];
+    else return { error: 'not a model or effort: ' + part };
+  }
+  if (!tier.model && !tier.effort) return { error: 'nothing recognised in "' + value + '"' };
+  return tier;
+}
+
+function tierText(tier) {
+  if (!tier) return null;
+  return [tier.model, tier.effort].filter(Boolean).join('/');
+}
+
+// Does a suggestion respect the bounds the user set?
+//
+// A floor says "never below this, even in max". A ceiling says "never above
+// this, even on the hard part". Anything the plugin would SAY that points
+// outside them is dropped at the rendering boundary rather than argued with
+// downstream, because the invariant is about what reaches the reader.
+function allows(bounds, suggestion) {
+  if (!bounds || !suggestion) return true;
+  const check = (kind, rank) => {
+    const floor = bounds.floor && bounds.floor[kind] ? rank(bounds.floor[kind]) : null;
+    const ceiling = bounds.ceiling && bounds.ceiling[kind] ? rank(bounds.ceiling[kind]) : null;
+    const mine = suggestion[kind] ? rank(suggestion[kind]) : null;
+    // An unranked name is not evidence of a breach. Reporting it as one would
+    // suppress a true statement over a spelling.
+    if (mine === null) return true;
+    if (floor !== null && mine < floor) return false;
+    if (ceiling !== null && mine > ceiling) return false;
+    return true;
+  };
+  return check('effort', effortRank) && check('model', modelRank);
+}
+
+function boundsNote(bounds) {
+  if (!bounds) return null;
+  const bits = [];
+  if (bounds.floor) bits.push('never below ' + tierText(bounds.floor));
+  if (bounds.ceiling) bits.push('never above ' + tierText(bounds.ceiling));
+  if (bounds.pin) bits.push('no self-switching at all: report the gap and leave it');
+  if (!bits.length) return null;
+  return 'Bounds the user set: ' + bits.join('; ') + '.';
+}
+
+function thinkingCaveat(tier) {
+  if (!tier || !tier.effort) return null;
+  if (THINKING_ONLY_EFFORTS.indexOf(tier.effort) === -1) return null;
+  return (
+    tier.effort + ' is only accepted while thinking is on; with thinking disabled the request is ' +
+    'refused outright, so check that before pointing anything at it.'
+  );
+}
+
+// ---------------------------------------------------------------------------
+// The directives
+//
+// Injected verbatim. The wording is the feature.
+//
+// One correction against what the host actually allows, because a directive
+// that promises a lever nobody has is worse than no directive. Nothing a hook
+// emits can change the running session's own model or effort: the whole hook
+// output contract carries no such field, and PreModelSwitch is a veto on a
+// switch someone else started, not a way to start one. What IS the agent's is
+// the tier of what it spawns - the model on an Agent call, and model AND
+// effort inside a Workflow script - so that is what the wording points at.
+//
+// The max directive is also the shortest, and that is not a coincidence: it is
+// the mode whose whole promise is that the plugin costs little, and a
+// four-hundred-token lecture about saving tokens spends what it is asking to
+// save. Every clause of the longer draft is still here; the padding is not.
+const DIRECTIVES = {
+  max:
+    'Budget mode: max efficiency. Fewest tokens that still get it right: search before you read, ' +
+    'read ranges not whole files, never re-read what you just wrote, batch edits, no preamble. ' +
+    'No subagent or workflow unless this loop truly cannot do it. Quality is not negotiable - ' +
+    'ceremony is.',
+  high:
+    'Budget mode: high efficiency. Work at full capability, and every couple of minutes check ' +
+    'whether the model, effort and approach you are running are bigger than this task needs - ' +
+    'step down when they are, and step back up when the work turns hard again. Your own tier is ' +
+    'not yours to set mid-turn, so when it is the thing that is too big, say so in one line with ' +
+    'the exact command and carry on. What is yours is the tier of what you spawn: the model on an ' +
+    'Agent call, and the model and effort inside a Workflow script. Size those to the stage - low ' +
+    'effort for mechanical ones - and reach for a workflow only when it saves more context than ' +
+    'it costs.',
+  standard: null,
+  off: null,
+};
+
+// The bounds are deliberately NOT glued on here. They are the user's own
+// limits on what may be suggested, they apply in every mode including the two
+// with no directive at all, and a renderer that wants them asks boundsNote()
+// for them. Appending them here would have made them invisible in `standard`,
+// which is the mode most people are in.
+function directive(name) {
+  return DIRECTIVES[name] || null;
+}
+
+// ---------------------------------------------------------------------------
+// State
+//
+// One small file in the host-aware config directory, so a Codex session reads
+// its own mode and not the Claude Code one. Atomic write, same beside-and-
+// rename as the other stores: the prompt hook and a pulse can land here in the
+// same second.
+
+function configDir() {
+  return host.detect(process.argv.slice(2), process.env) === host.CODEX
+    ? codex.homeDir()
+    : process.env.CLAUDE_CONFIG_DIR || path.join(os.homedir(), '.claude');
+}
+
+function modeFile() {
+  return path.join(configDir(), 'usage-limits-mode.json');
+}
+
+function changesFile() {
+  return path.join(configDir(), 'usage-limits-changes.json');
+}
+
+// How many declines and how many change-log entries are worth keeping. Both
+// are bounded for the same reason as the drift ledger: the file must be the
+// same size after a year as after a day.
+const KEEP_DECLINED = 40;
+const KEEP_OFFERS = 10;
+const KEEP_CHANGES = 100;
+
+function empty() {
+  return {
+    version: 1,
+    mode: DEFAULT_MODE,
+    auto: false,
+    guardPercent: null,
+    // The ceiling is the only setting here that is ENFORCED rather than
+    // reported: past it, fan-out calls are refused at the hook. Null means no
+    // ceiling, and a ceiling nobody set never refuses anything. See ceiling.js.
+    ceilingPercent: null,
+    setAt: null,
+    setBy: null,
+    session: null,
+    floor: null,
+    ceiling: null,
+    pin: false,
+    advice: { off: false, declined: {}, offered: {} },
+  };
+}
+
+// A corrupt file falls back to the default rather than throwing. This is read
+// from inside hooks, and a hook that fails over its own state file would cost
+// more than the setting it was trying to honour.
+function read() {
+  const base = empty();
+  let raw = null;
+  try {
+    raw = fs.readFileSync(modeFile(), 'utf8');
+  } catch (err) {
+    // Never written is the ordinary case. Anything else is a file that exists
+    // and could not be read, which is a different thing from "no mode set" and
+    // is flagged so describe() can say so instead of reporting the default as
+    // though the user had chosen it.
+    if (err.code !== 'ENOENT') base.unreadable = true;
+    return base;
+  }
+  let parsed = null;
+  try {
+    parsed = JSON.parse(raw);
+  } catch (err) {
+    base.unreadable = true;
+    return base;
+  }
+  if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) {
+    base.unreadable = true;
+    return base;
+  }
+  const named = normalise(parsed.mode);
+  if (named && named.mode) base.mode = named.mode;
+  base.auto = parsed.auto === true;
+  base.guardPercent = Number.isFinite(parsed.guardPercent) ? parsed.guardPercent : null;
+  // Anything outside 1-100 is not a ceiling, and enforcing a number that was
+  // never a percentage would refuse work over a typo.
+  base.ceilingPercent =
+    Number.isFinite(parsed.ceilingPercent) && parsed.ceilingPercent > 0 && parsed.ceilingPercent <= 100
+      ? parsed.ceilingPercent
+      : null;
+  base.setAt = Number.isFinite(parsed.setAt) ? parsed.setAt : null;
+  base.setBy = typeof parsed.setBy === 'string' ? parsed.setBy : null;
+  if (parsed.session && typeof parsed.session === 'object' && parsed.session.id) {
+    const sessionMode = normalise(parsed.session.mode);
+    if (sessionMode && sessionMode.mode) {
+      base.session = { id: String(parsed.session.id), mode: sessionMode.mode, at: Number.isFinite(parsed.session.at) ? parsed.session.at : null };
+    }
+  }
+  for (const key of ['floor', 'ceiling']) {
+    const value = parsed[key];
+    if (value && typeof value === 'object' && (value.model || value.effort)) {
+      base[key] = {
+        model: modelRank(value.model) === null ? null : String(value.model).toLowerCase(),
+        effort: effortRank(value.effort) === null ? null : String(value.effort).toLowerCase(),
+      };
+      if (!base[key].model && !base[key].effort) base[key] = null;
+    }
+  }
+  base.pin = parsed.pin === true;
+  if (parsed.advice && typeof parsed.advice === 'object') {
+    base.advice.off = parsed.advice.off === true;
+    if (parsed.advice.declined && typeof parsed.advice.declined === 'object') {
+      base.advice.declined = Object.assign({}, parsed.advice.declined);
+    }
+    if (parsed.advice.offered && typeof parsed.advice.offered === 'object') {
+      base.advice.offered = Object.assign({}, parsed.advice.offered);
+    }
+  }
+  return base;
+}
+
+function writeAtomic(file, value) {
+  fs.mkdirSync(path.dirname(file), { recursive: true });
+  const tmp = file + '.' + process.pid + '.tmp';
+  fs.writeFileSync(tmp, JSON.stringify(value, null, 2) + '\n');
+  fs.renameSync(tmp, file);
+}
+
+// Never throws for the same reason read() does not.
+function write(state) {
+  try {
+    const next = Object.assign({}, state);
+    next.advice = Object.assign({ off: false, declined: {}, offered: {} }, state.advice || {});
+    next.advice.declined = trimStamps(next.advice.declined, KEEP_DECLINED);
+    next.advice.offered = trimStamps(next.advice.offered, KEEP_OFFERS, (entry) => entry && entry.at);
+    writeAtomic(modeFile(), next);
+    return true;
+  } catch (err) {
+    return false;
+  }
+}
+
+// Keeps the newest N by timestamp. `at` is either the value itself (declined)
+// or a field on it (offered), so one trimmer serves both.
+function trimStamps(table, keep, pick) {
+  const entries = Object.keys(table || {}).map((key) => ({
+    key,
+    at: pick ? Number(pick(table[key])) || 0 : Number(table[key]) || 0,
+  }));
+  entries.sort((a, b) => b.at - a.at);
+  const kept = {};
+  for (const entry of entries.slice(0, keep)) kept[entry.key] = table[entry.key];
+  return kept;
+}
+
+// ---------------------------------------------------------------------------
+// auto
+//
+// Not a fifth mode: a switch that picks one from pressure, so it is always
+// reported as the mode it resolved to ("auto -> max"), never as a mystery.
+//
+// It never resolves to `off`. Turning the plugin off is a decision a person
+// makes, not one a threshold makes.
+function autoPick(reading) {
+  const percent = reading && Number.isFinite(reading.percentUsed) ? reading.percentUsed : null;
+  const pressure = reading ? reading.pressure : null;
+  if (pressure === 'tight' || pressure === 'gone') return 'max';
+  if (percent === null) return DEFAULT_MODE;
+  if (percent >= 80) return 'max';
+  if (percent >= 50) return 'high';
+  return DEFAULT_MODE;
+}
+
+// Where autoPick's reading comes from when the caller has none.
+//
+// Every production caller had none. brief.js, pulse.js, feed.js, stop.js and
+// the report all ask forSession({ sessionId }) and nothing else, because they
+// need the mode BEFORE they can afford a reading - that is the whole point of
+// settling it first. autoPick's no-measurement branch then returned
+// `standard`, so an account could sit at 95 per cent used with `auto` on and
+// every hook, the status line and the ledger would read and behave as
+// standard. The headline behaviour of the mode was inert.
+//
+// So it takes the cheap reading itself: the snapshot already on disk plus
+// whatever correction an earlier turn has already paid for. No scan, no
+// request, no wait - the same source the status line redraws from many times a
+// second. It is only ever reached while `auto` is actually on.
+//
+// Required lazily because usage.js requires this module back; at module scope
+// that is a half-built export table.
+function readingNow(now) {
+  try {
+    const usage = require('./usage.js');
+    // The host this hook is running for, settled the same way every caller
+    // settles it, so a Codex session reads Codex's meter.
+    usage.setHost(host.detect(process.argv.slice(2), process.env));
+    const collected = usage.collect(Number.isFinite(now) ? now : Date.now());
+    if (!collected || !collected.utilization) return null;
+    const codexHome = usage.isCodex() ? codex.homeDir() : null;
+    const windows = usage.snapshotWindows(collected, now, codexHome);
+    // The same suppression every other reader makes: a per-model weekly for a
+    // model this session is not running cannot be the window that stops it, so
+    // it must not be the thing that picks the mode either.
+    const usable = windows.filter((w) => w.applies !== false && !w.stale && Number.isFinite(w.percentUsed));
+    if (!usable.length) return null;
+    return { percentUsed: usable.reduce((worst, w) => (w.percentUsed > worst.percentUsed ? w : worst)).percentUsed };
+  } catch (err) {
+    // No reading is the one thing autoPick already handles: it falls to the
+    // default rather than guessing.
+    return null;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Resolution and precedence
+//
+//   USAGE_LIMITS_MODE env  ->  --session override  ->  persisted file  ->  standard
+//
+// Each level reports its own source, the way the effort chain does, because
+// "which mode am I in" is uninteresting next to "and who said so".
+function resolve(options) {
+  const opts = options || {};
+  const env = opts.env || process.env;
+  const state = opts.state || read();
+  const sessionId = opts.sessionId || null;
+  const reading = opts.reading || null;
+
+  let name = null;
+  let source = null;
+  let auto = false;
+
+  const fromEnv = normalise(env.USAGE_LIMITS_MODE);
+  if (fromEnv && fromEnv.mode) {
+    name = fromEnv.mode;
+    source = 'environment';
+  } else if (fromEnv && fromEnv.auto) {
+    auto = true;
+    source = 'environment';
+  } else if (state.session && sessionId && state.session.id === sessionId) {
+    name = state.session.mode;
+    source = 'this session';
+  } else if (state.auto) {
+    auto = true;
+    source = 'auto, from the file';
+  } else {
+    name = state.mode || DEFAULT_MODE;
+    source = state.setAt ? 'the file' : 'the default';
+  }
+
+  let label = name;
+  if (auto) {
+    // A caller that has a reading passes it; one that has none gets the cheap
+    // one rather than silently resolving to the default. See readingNow().
+    name = autoPick(reading || readingNow(opts.now));
+    label = 'auto -> ' + name;
+  }
+
+  const policy = MODES[name] || MODES[DEFAULT_MODE];
+  return {
+    name: policy.name,
+    label,
+    source,
+    auto,
+    policy,
+    guardPercent: state.guardPercent,
+    ceilingPercent: state.ceilingPercent,
+    bounds: { floor: state.floor, ceiling: state.ceiling, pin: state.pin === true },
+    advice: state.advice,
+    state,
+    // The one field every caller needs and nobody should re-derive.
+    directive: directive(policy.name),
+  };
+}
+
+// What a hook wants: one call, never throws, and the reading is optional
+// because a hook that has not scanned yet still has to know whether to bother.
+function forSession(options) {
+  try {
+    return resolve(options);
+  } catch (err) {
+    const policy = MODES[DEFAULT_MODE];
+    return {
+      name: policy.name,
+      label: policy.name,
+      source: 'the default',
+      auto: false,
+      policy,
+      guardPercent: null,
+      ceilingPercent: null,
+      bounds: { floor: null, ceiling: null, pin: false },
+      advice: { off: false, declined: {}, offered: {} },
+      state: empty(),
+      directive: null,
+    };
+  }
+}
+
+// ---------------------------------------------------------------------------
+// The change log
+//
+// "Can you change it back?" needs a referent. Without one the agent is
+// guessing at the user's own settings, and a guess written into their file is
+// exactly the surprise the read-only rule exists to prevent. So every change
+// the plugin knows about is written down, and "back" means the last entry.
+//
+//   { at, plane: user|agent|mode, key, from, to, by: user|claude, reason }
+
+// "Nothing has been changed yet" and "the record of what changed is gone" are
+// different facts, and a file that cannot be parsed must not be reported as
+// the first. A user who has been making changes and is told nothing was ever
+// changed has no reason to suspect the file. ENOENT is one err.code check
+// away, so the distinction costs a line.
+function readChanges() {
+  let raw = null;
+  try {
+    raw = fs.readFileSync(changesFile(), 'utf8');
+  } catch (err) {
+    // Never written is the ordinary case, and it is not a fault.
+    return err.code === 'ENOENT' ? { entries: [] } : { entries: [], unreadable: true };
+  }
+  try {
+    const parsed = JSON.parse(raw);
+    if (!parsed || !Array.isArray(parsed.entries)) return { entries: [], unreadable: true };
+    // A truncated file can leave entries that are not objects; they are
+    // dropped rather than allowed to reach a renderer.
+    return { entries: parsed.entries.filter((e) => e && typeof e === 'object' && Number.isFinite(e.at)) };
+  } catch (err) {
+    return { entries: [], unreadable: true };
+  }
+}
+
+const UNREADABLE_LOG =
+  'The change log exists but could not be read, so what changed is not known ' +
+  'rather than empty. The file is ';
+
+function logChange(entry, now) {
+  if (!entry || !entry.key) return false;
+  try {
+    const state = readChanges();
+    state.entries.push({
+      at: Number.isFinite(now) ? now : Date.now(),
+      plane: entry.plane || 'mode',
+      key: String(entry.key),
+      from: entry.from === undefined ? null : entry.from,
+      to: entry.to === undefined ? null : entry.to,
+      by: entry.by === 'claude' ? 'claude' : 'user',
+      reason: entry.reason ? String(entry.reason).slice(0, 200) : null,
+      // How far the change reached, as a field rather than as English inside
+      // `to`. A session override used to be recorded as the string
+      // "max (this session)", which reads correctly and is useless to undo:
+      // it parsed back to the persisted mode, rewrote that, left the override
+      // in place, and reported a revert that had not happened.
+      scope: entry.scope === 'session' ? 'session' : 'global',
+      sessionId: entry.sessionId ? String(entry.sessionId) : null,
+    });
+    if (state.entries.length > KEEP_CHANGES) state.entries = state.entries.slice(-KEEP_CHANGES);
+    writeAtomic(changesFile(), state);
+    return true;
+  } catch (err) {
+    return false;
+  }
+}
+
+// The change undo will act on: the newest one that is still standing.
+//
+// Two kinds of entry are skipped. An undo's own reversal is a record of the
+// undo, not a change to reverse - taking it as the target made a second `undo`
+// redo the first, oscillating between two states forever instead of stepping
+// back through the log. And an entry already reversed is done with, so undo
+// twice reaches the change before it.
+function lastUndoable(entries) {
+  for (let i = entries.length - 1; i >= 0; i -= 1) {
+    const entry = entries[i];
+    if (!entry || entry.undone || entry.reason === 'undo') continue;
+    return { entry, index: i };
+  }
+  return null;
+}
+
+// Stamps the entry undo just reversed, so it is not offered again.
+function markUndone(index, now) {
+  try {
+    const state = readChanges();
+    if (!state.entries[index]) return false;
+    state.entries[index] = Object.assign({}, state.entries[index], { undone: Number.isFinite(now) ? now : Date.now() });
+    writeAtomic(changesFile(), state);
+    return true;
+  } catch (err) {
+    return false;
+  }
+}
+
+function ago(ms) {
+  if (!Number.isFinite(ms) || ms < 0) return 'just now';
+  const minutes = Math.round(ms / 60000);
+  if (minutes < 1) return 'just now';
+  if (minutes < 60) return minutes + 'm ago';
+  const hours = Math.round(minutes / 60);
+  if (hours < 48) return hours + 'h ago';
+  return Math.round(hours / 24) + 'd ago';
+}
+
+function describeChange(entry, now) {
+  return (
+    '  ' + ago(now - entry.at).padEnd(9) + entry.plane.padEnd(9) + entry.key + ': ' +
+    (entry.from === null ? '(unset)' : entry.from) + ' -> ' + (entry.to === null ? '(unset)' : entry.to) +
+    (entry.scope === 'session' ? ' (this session only)' : '') +
+    ' (by ' + entry.by + (entry.reason ? ', ' + entry.reason : '') + (entry.undone ? ', since undone' : '') + ')'
+  );
+}
+
+function history(now, limit) {
+  const log = readChanges();
+  const entries = log.entries.slice(-(limit || 20)).reverse();
+  if (!entries.length) {
+    return log.unreadable ? UNREADABLE_LOG + changesFile() + '.' : 'Nothing has been changed through this plugin yet.';
+  }
+  const at = Number.isFinite(now) ? now : Date.now();
+  return ['What changed, newest first:'].concat(entries.map((e) => describeChange(e, at))).join('\n');
+}
+
+// Reverses the last logged change, naming it first.
+//
+// It only reverses what the plugin itself owns - the mode plane and the agent
+// plane. A user-plane entry is settings.json, and this file never writes that:
+// it names the entry and the one command that undoes it, and leaves the file
+// alone. That is not timidity, it is the two-planes rule - the user's baseline
+// is theirs - and it is what keeps "no mode writes settings.json" true by
+// construction rather than by remembering.
+function undo(now) {
+  const state = readChanges();
+  const found = lastUndoable(state.entries);
+  if (!found) {
+    return {
+      ok: false,
+      text: state.unreadable
+        ? UNREADABLE_LOG + changesFile() + '. Nothing was changed.'
+        : 'There is nothing in the change log to undo.',
+    };
+  }
+  const last = found.entry;
+  const at = Number.isFinite(now) ? now : Date.now();
+  if (last.plane === 'user') {
+    // In Codex the command needs the host on it, or it edits the other agent's
+    // baseline: lowpower.js falls back to host.detect(argv), where Claude wins
+    // ties, and a machine with both installed has both.
+    const codexHere = host.detect(process.argv.slice(2), process.env) === host.CODEX;
+    // Which way the change went decides which command reverses it. Naming
+    // "lowpower off" after a restore would be telling the user to run the
+    // thing they just ran, which is a no-op dressed up as an undo.
+    const had = last.from === null || last.from === '(unset)' ? null : last.from;
+    const reverse = last.reason === 'lowpower off'
+      ? 'node scripts/lowpower.js on' +
+        (last.key === 'effortLevel' && had ? ' --effort ' + had : '') +
+        (last.key === 'model' && had ? ' --model ' + had : '')
+      : 'node scripts/lowpower.js off';
+    return {
+      ok: false,
+      entry: last,
+      text:
+        'The last change was to your own settings (' + last.key + ': ' +
+        (last.from === null ? '(unset)' : last.from) + ' -> ' + (last.to === null ? '(unset)' : last.to) +
+        '), ' + ago(at - last.at) + '. That plane is yours and this ' +
+        'script does not write it. To put it back: ' + reverse +
+        (codexHere ? ' --host codex' : '') + '. Note that it applies to ' +
+        'NEW sessions - the one you are in keeps ' +
+        (codexHere ? 'the tier it started with; Codex changes a running session through its own controls.' : 'the tier it started with.'),
+    };
+  }
+  if (last.plane !== 'mode') {
+    return {
+      ok: false,
+      entry: last,
+      text:
+        'The last change was on the agent plane (' + last.key + ': ' + last.from + ' -> ' + last.to +
+        '), which lives in the turn that made it and cannot be rewritten from here. Nothing was changed.',
+    };
+  }
+  const target = normalise(last.from);
+  const current = read();
+  const before = current.auto ? 'auto' : current.mode;
+  // A reversal is only true if the EFFECTIVE mode moves, and the session
+  // override outranks the file. Rewriting the file under a live override and
+  // reporting "reverted, nothing else was touched" was a revert that had not
+  // happened: mode --session-id S1 still answered with the override.
+  if (last.key === 'mode' && last.scope === 'session') {
+    const had = current.session;
+    current.session = null;
+    write(current);
+    markUndone(found.index, now);
+    logChange({ plane: 'mode', key: 'mode', from: last.to, to: current.mode, by: 'user', reason: 'undo' }, now);
+    return {
+      ok: true,
+      entry: last,
+      text:
+        'Reverting the session override' + (had && had.mode ? ' (' + had.mode + ')' : '') +
+        '. This session is back on the persisted mode, ' + current.mode + '. Nothing else was touched.',
+    };
+  }
+  if (last.key === 'mode' && target && (target.mode || target.auto)) {
+    // "turn auto on, try a mode, put it back" is the likeliest undo there is,
+    // and it was the one that failed: normalise('auto') returns { auto: true }
+    // with no `.mode`, so the branch fell through to "not one this script can
+    // reverse" while the log held exactly what was needed.
+    current.auto = Boolean(target.auto);
+    if (target.mode) current.mode = target.mode;
+    current.setAt = at;
+    current.setBy = 'undo';
+    write(current);
+    markUndone(found.index, now);
+    const back = target.auto ? 'auto' : target.mode;
+    logChange({ plane: 'mode', key: 'mode', from: before, to: back, by: 'user', reason: 'undo' }, now);
+    return { ok: true, entry: last, text: 'Reverting mode ' + before + ' back to ' + back + '. Nothing else was touched.' };
+  }
+  if (last.key === 'auto') {
+    current.auto = last.from === true || last.from === 'on';
+    write(current);
+    markUndone(found.index, now);
+    logChange({ plane: 'mode', key: 'auto', from: last.to, to: current.auto, by: 'user', reason: 'undo' }, now);
+    return { ok: true, entry: last, text: 'Reverting auto back to ' + (current.auto ? 'on' : 'off') + '.' };
+  }
+  if (last.key === 'guard') {
+    current.guardPercent = Number.isFinite(last.from) ? last.from : null;
+    write(current);
+    markUndone(found.index, now);
+    logChange({ plane: 'mode', key: 'guard', from: last.to, to: current.guardPercent, by: 'user', reason: 'undo' }, now);
+    return { ok: true, entry: last, text: 'Reverting the guard back to ' + (current.guardPercent === null ? 'none' : current.guardPercent + '%') + '.' };
+  }
+  if (last.key === 'floor' || last.key === 'ceiling') {
+    current[last.key] = last.from ? parseTier(last.from) : null;
+    if (current[last.key] && current[last.key].error) current[last.key] = null;
+    write(current);
+    markUndone(found.index, now);
+    logChange({ plane: 'mode', key: last.key, from: last.to, to: tierText(current[last.key]), by: 'user', reason: 'undo' }, now);
+    return { ok: true, entry: last, text: 'Reverting the ' + last.key + ' back to ' + (tierText(current[last.key]) || 'none') + '.' };
+  }
+  if (last.key === 'pin') {
+    current.pin = last.from === true;
+    write(current);
+    markUndone(found.index, now);
+    logChange({ plane: 'mode', key: 'pin', from: last.to, to: current.pin, by: 'user', reason: 'undo' }, now);
+    return { ok: true, entry: last, text: 'Reverting pin back to ' + (current.pin ? 'on' : 'off') + '.' };
+  }
+  return { ok: false, entry: last, text: 'The last change (' + last.key + ') is not one this script can reverse. Nothing was changed.' };
+}
+
+// ---------------------------------------------------------------------------
+// The advice channel
+//
+// The two planes are not sealed off from each other: they talk, in both
+// directions, through the conversation. The whole rule in one line:
+//
+//   Claude may RECOMMEND a user-plane change. Claude may MAKE one when asked.
+//   Claude may never make one unasked.
+//
+// The failure mode of a feature like this is nagging, and a plugin that nags
+// gets turned off, at which point it protects nothing. So:
+//   - Evidence or silence. A recommendation cites a measurement or it is not
+//     made. "Recommended" with no number is nagging.
+//   - One per session, at most.
+//   - A declined recommendation is remembered and never raised again.
+//   - Never in off. In max, allowed but terse, and still capped at one.
+//   - Always names the exact command, the plane it changes, and when it takes
+//     effect. A recommendation the user cannot act on in one step is a
+//     complaint.
+
+// The id is what a decline is remembered by, so it describes the SUGGESTION
+// and not the moment: the same advice next week is the same advice.
+function adviceId(fit) {
+  if (!fit) return null;
+  return 'effort:' + fit.effort + '>' + fit.cheaper;
+}
+
+// `fit` is usage.settingFit()'s output: measured, or null. Passing it in
+// rather than reaching for usage.js keeps this module free of the cycle and
+// makes "evidence or silence" structural - with no measurement there is
+// nothing to build a recommendation out of.
+function advicePending(options) {
+  const opts = options || {};
+  const decided = opts.decided || resolve(opts);
+  const fit = opts.fit || null;
+  const sessionId = opts.sessionId || null;
+  const advice = decided.advice || { off: false, declined: {}, offered: {} };
+
+  if (decided.policy.briefStyle === 'none') return { ok: false, reason: 'off', text: null };
+  if (advice.off) return { ok: false, reason: 'muted', text: null };
+  if (!fit) return { ok: false, reason: 'no measurement', text: null };
+  const id = adviceId(fit);
+  if (advice.declined && advice.declined[id]) return { ok: false, reason: 'declined before', id, text: null };
+  // A bound the user set outranks the measurement: advice that points below
+  // their own floor is advice they already refused.
+  if (!allows(decided.bounds, { effort: fit.cheaper })) return { ok: false, reason: 'below the bounds set', id, text: null };
+  const offered = advice.offered && sessionId ? advice.offered[sessionId] : null;
+  const alreadyOffered = Boolean(offered && offered.id === id);
+
+  // Terse in max, and the same measurement either way.
+  const terse = decided.policy.briefStyle === 'terse';
+  const text = terse
+    ? 'Recommendation: ' + fit.effort + ' measured ' + fit.multiple + 'x ' + fit.cheaper +
+      ' a turn here (' + fit.sample + ' turns). Yours to make: ' + fit.command + ', from your next turn.'
+    : 'One recommendation, from this account\'s own record: ' + fit.effort + ' has measured ' +
+      fit.multiple + ' times the cost of ' + fit.cheaper + ' a turn (' + fit.sample + ' turns against ' +
+      fit.cheaperSample + '). If the stretch ahead is mechanical, ' + fit.command + ' is the change and it ' +
+      'is the user\'s own setting to make - offer it, do not make it. Say plainly that it applies from ' +
+      'the next turn onward, not retroactively.';
+
+  return { ok: true, id, text, alreadyOffered, terse };
+}
+
+// What was offered, and what it said.
+//
+// The text is stored, not only the id, because the CLI has to be able to
+// answer "what is pending" and "decline what you just offered" without a
+// transcript scan. The brief is the only thing that HAS the measurement - it
+// has already paid for the scan - so the offer it writes is where a later
+// `mode --advice` gets its evidence from.
+function adviceOffer(id, sessionId, now, text) {
+  if (!id || !sessionId) return false;
+  const state = read();
+  state.advice.offered[sessionId] = {
+    id,
+    at: Number.isFinite(now) ? now : Date.now(),
+    text: text ? String(text).slice(0, 600) : null,
+  };
+  return write(state);
+}
+
+// The most recent offer: this session's when there is one, otherwise the
+// newest across sessions. "The user just said no" is said in a session, but a
+// user typing the command in a fresh shell has no session id to give and still
+// means the recommendation they were just shown.
+function adviceLastOffer(sessionId, state) {
+  const current = state || read();
+  const offered = (current.advice && current.advice.offered) || {};
+  if (sessionId && offered[sessionId] && offered[sessionId].id) {
+    return Object.assign({ sessionId }, offered[sessionId]);
+  }
+  let best = null;
+  for (const key of Object.keys(offered)) {
+    const entry = offered[key];
+    if (!entry || !entry.id) continue;
+    if (!best || (Number(entry.at) || 0) > (Number(best.at) || 0)) best = Object.assign({ sessionId: key }, entry);
+  }
+  return best;
+}
+
+function adviceDecline(id, now) {
+  if (!id) return false;
+  const state = read();
+  state.advice.declined[id] = Number.isFinite(now) ? now : Date.now();
+  return write(state);
+}
+
+function adviceMute(off) {
+  const state = read();
+  state.advice.off = Boolean(off);
+  return write(state);
+}
+
+// ---------------------------------------------------------------------------
+// The two planes, side by side
+//
+// The user plane is settings.json, the pickers, lowpower.js: the user saying
+// what they want for themselves. It is read here and never written.
+//
+// The agent plane is the tier actually running this turn. That is what costs
+// money, and it is the agent's to move - for what it spawns. Its own tier it
+// can only report and name the command for.
+//
+// CLAUDE_EFFORT is read FIRST and deliberately not folded into usage.js's
+// effort chain. The host sets it per turn, after any silent downgrade for the
+// selected model, which makes it the most accurate reading there is - and it
+// is present in the environment of everything the host launches, so putting it
+// into the shared chain would change what every other caller sees.
+function tierNow(options) {
+  const opts = options || {};
+  const env = opts.env || process.env;
+  const usage = opts.usage || require('./usage.js');
+  const sessionId = opts.sessionId || null;
+
+  let effort = null;
+  const perTurn = String(env.CLAUDE_EFFORT || '').trim().toLowerCase();
+  if (perTurn && (effortRank(perTurn) !== null || perTurn === 'ultracode')) {
+    effort = { effort: perTurn, source: 'this turn', live: true };
+  }
+  if (!effort) {
+    try {
+      effort = usage.effortNow(sessionId, env);
+    } catch (err) {
+      effort = null;
+    }
+  }
+
+  let settings = null;
+  try {
+    settings = usage.collect(Number.isFinite(opts.now) ? opts.now : Date.now()).settings || null;
+  } catch (err) {
+    settings = null;
+  }
+
+  let running = null;
+  try {
+    const seen = sessionId ? usage.liveModel(sessionId) : null;
+    running = seen && seen.model ? seen.model : null;
+  } catch (err) {
+    running = null;
+  }
+
+  return {
+    baseline: {
+      model: settings && settings.model ? settings.model : null,
+      effort: settings && settings.effortLevel ? settings.effortLevel : null,
+    },
+    running: {
+      model: running,
+      effort: effort ? effort.effort : null,
+      source: effort ? effort.source : null,
+    },
+  };
+}
+
+function sameFamily(a, b) {
+  const left = modelRank(a);
+  const right = modelRank(b);
+  if (left === null || right === null) return String(a || '') === String(b || '');
+  return left === right;
+}
+
+// One clause for the brief, or two lines for `--baseline`.
+//
+// Where the baseline and the running tier agree there is nothing interesting
+// to say, so it says it once. Where they differ, THAT is the story, and it is
+// the whole reason the line exists: a turn that opened with the window and
+// never said what tier was producing it was hiding the number that decides
+// what the turn costs.
+function tierLine(tier, options) {
+  const opts = options || {};
+  if (!tier) return null;
+  const base = tier.baseline || {};
+  const run = tier.running || {};
+  const shortModel = (name) => {
+    const rank = modelRank(name);
+    return rank === null ? name : MODEL_ORDER[rank];
+  };
+  const runningText = [shortModel(run.model) || shortModel(base.model), run.effort || base.effort].filter(Boolean).join('/');
+  if (!runningText) return null;
+  const baseText = [shortModel(base.model), base.effort].filter(Boolean).join('/');
+  const differs =
+    baseText && runningText !== baseText &&
+    (!sameFamily(run.model || base.model, base.model) || (run.effort || base.effort) !== base.effort);
+  const source = run.source ? ' (' + run.source + ')' : '';
+  if (opts.terse) {
+    return differs ? runningText + source + ', yours ' + baseText : runningText + source;
+  }
+  return differs
+    ? 'Running ' + runningText + source + '; your baseline is ' + baseText + '. The gap is the ' +
+      'interesting part: your baseline is yours and is not being changed.'
+    : 'Running ' + runningText + source + '.';
+}
+
+// ---------------------------------------------------------------------------
+// The ledger
+//
+// Modes are evidence, not vibes. The measurement lives in drift.js, which
+// already has the bounded-append pattern and a file of its own; a second store
+// for the same kind of after-the-fact measurement would be a second thing to
+// keep correct.
+function ledger(now) {
+  const drift = require('./drift.js');
+  const codexHome = host.detect(process.argv.slice(2), process.env) === host.CODEX ? codex.homeDir() : null;
+  const rows = drift.modeSummary(codexHome);
+  if (!rows.length) return 'Mode ledger: nothing measured yet. It fills in as replies land in each mode.';
+  const lines = ['Measured cost per mode:'];
+  for (const row of rows) {
+    lines.push(
+      '  ' + row.mode.padEnd(9) + String(row.turns).padStart(4) + ' turns   ' +
+        (Number.isFinite(row.usdPerTurn) ? '$' + row.usdPerTurn.toFixed(2) + '/turn' : 'no price')
+    );
+  }
+  lines.push('');
+  lines.push('Observed, not predicted: what replies actually cost while each mode was on.');
+  // Said rather than left to be noticed: `off` returns before the Stop hook
+  // reads anything, so it has no rows here and never will. A mode that
+  // injects nothing has no injection to attribute a cost to.
+  if (!rows.some((row) => row.mode === 'off')) {
+    lines.push('`off` never appears: its hooks return before the reply is measured.');
+  }
+  return lines.join('\n');
+}
+
+// ---------------------------------------------------------------------------
+// Rendering
+
+// The mode's full record, in two halves that are not the same kind of fact.
+//
+// The first half is what the plugin DOES: every field here is read by a named
+// script, and changing it changes behaviour. The second is the mode's stance
+// towards the agent, which has no effect except through the directive - so
+// when there is no directive it is said plainly that nothing carries it. The
+// old version printed both under one heading, which reported five fields as
+// behaviour when nothing anywhere read them.
+function explain(name) {
+  const policy = MODES[name];
+  if (!policy) return 'No such mode: ' + name + '. Try one of: ' + ORDER.join(', ') + '.';
+  const text = DIRECTIVES[name];
+  const lines = [name + ' - ' + policy.summary, '', '  What the plugin does (read by the scripts named):'];
+  for (const key of Object.keys(policy)) {
+    if (!WIRED[key]) continue;
+    lines.push('    ' + key.padEnd(20) + String(policy[key]).padEnd(10) + WIRED[key]);
+  }
+  const stance = Object.keys(policy).filter((key) => key !== 'name' && key !== 'summary' && key !== 'directive' && !WIRED[key]);
+  if (stance.length) {
+    lines.push('');
+    lines.push(
+      text
+        ? '  What the mode asks of the agent. Nothing reads these fields; they reach'
+        : '  What the mode asks of the agent. Nothing reads these fields, and this mode'
+    );
+    lines.push(text ? '  the agent only by being restated in the directive below:' : '  has no directive, so nothing carries them - they describe intent only:');
+    for (const key of stance) lines.push('    ' + key.padEnd(20) + String(policy[key]));
+  }
+  lines.push('');
+  lines.push(text ? '  Directive, injected verbatim:' : '  Directive: none. ' + (name === 'off' ? 'Nothing is injected at all.' : 'The line is what it is today.'));
+  if (text) lines.push('    ' + text);
+  return lines.join('\n');
+}
+
+function list(decided) {
+  const lines = ['Budget modes:', ''];
+  for (const name of ORDER) {
+    const policy = MODES[name];
+    const here = decided && decided.name === name ? ' <- current' : '';
+    lines.push('  ' + name.padEnd(9) + policy.summary + here);
+  }
+  lines.push('');
+  lines.push('Aliases: ' + Object.keys(ALIASES).sort().join(', ') + '.');
+  lines.push('"normal" is deliberately not an alias: it means opposite things to different people.');
+  return lines.join('\n');
+}
+
+function describe(decided, now) {
+  const lines = [];
+  lines.push('Mode: ' + decided.label + ' (from ' + decided.source + ')');
+  lines.push('  ' + decided.policy.summary);
+  lines.push('');
+  lines.push('  brief         ' + decided.policy.briefStyle + (decided.policy.briefWhenUnchanged ? '' : ', silent when nothing moved'));
+  lines.push('  readings      ' + (decided.policy.refreshSeconds ? 'every ' + decided.policy.refreshSeconds + 's' : 'none: the hooks return before reading anything'));
+  if (decided.policy.recheckSeconds) lines.push('  recheck       every ' + decided.policy.recheckSeconds + 's, mid-turn');
+  lines.push('  subagents     ' + decided.policy.subagents);
+  lines.push('  workflows     ' + decided.policy.workflows);
+  if (decided.guardPercent !== null && decided.guardPercent !== undefined) {
+    lines.push('  guard         one line at ' + decided.guardPercent + '% used, and nothing else');
+  }
+  const note = boundsNote(decided.bounds);
+  if (note) lines.push('  bounds        ' + note.replace('Bounds the user set: ', ''));
+  if (decided.advice && decided.advice.off) lines.push('  advice        off');
+  lines.push('');
+  lines.push('  It governs the agent plane only. Your settings.json is never written by this.');
+  lines.push('  File: ' + modeFile());
+  if (decided.state && decided.state.unreadable) {
+    lines.push('');
+    lines.push('  That file exists but could not be read, so the mode above is the default');
+    lines.push('  rather than anything you chose. Setting it again rewrites it.');
+  }
+  return lines.join('\n');
+}
+
+// ---------------------------------------------------------------------------
+// CLI
+
+function setMode(name, opts, now) {
+  const state = read();
+  const before = state.auto ? 'auto' : state.mode;
+  const at = Number.isFinite(now) ? now : Date.now();
+  if (opts && opts.session) {
+    if (!opts.sessionId) {
+      return 'A session override needs the session id. Run it with --session-id <id>, or set it for good with: mode ' + name;
+    }
+    state.session = { id: opts.sessionId, mode: name, at };
+    write(state);
+    logChange({ plane: 'mode', key: 'mode', from: before, to: name, by: 'user', reason: 'session override', scope: 'session', sessionId: opts.sessionId }, at);
+    return 'Mode ' + name + ' for this session only. The persisted mode is still ' + state.mode + '.';
+  }
+  // A live session override outranks the file, so setting the mode globally
+  // while one is in force changed nothing at all for the session that typed
+  // the command - and the reply named a directive that would never appear.
+  // Setting the mode outright is the user saying what the mode is now, so the
+  // override goes, and it is said rather than done quietly.
+  const cleared = state.session;
+  state.session = null;
+  state.mode = name;
+  state.auto = false;
+  state.setAt = at;
+  state.setBy = 'user';
+  if (opts && Number.isFinite(opts.guard)) state.guardPercent = opts.guard;
+  if (opts && Number.isFinite(opts.ceiling)) state.ceilingPercent = opts.ceiling;
+  write(state);
+  logChange({ plane: 'mode', key: 'mode', from: before, to: name, by: 'user', reason: null }, at);
+
+  const policy = MODES[name];
+  const lines = ['Mode ' + name + ': ' + policy.summary + '.'];
+  if (cleared && cleared.mode) {
+    lines.push(
+      'The session override (' + cleared.mode + ') was cleared, so this takes effect ' +
+        'in that session too.'
+    );
+  }
+  if (name === 'off') {
+    // off means off, including at 100 per cent. That is what was asked and it
+    // is honoured literally. But a silent cutoff at the wall is the exact
+    // failure this plugin exists to prevent, so the consequence is stated once
+    // - and the way to keep one line is offered rather than imposed.
+    lines.push('off: nothing will be injected, including at the wall.');
+    // Said out loud because it is the one consequence a user would otherwise
+    // discover by noticing something missing: the end-of-reply cost line and
+    // the closing line are hooks too, and off stops them reading the
+    // transcript at all. The panel stops animating for the same reason.
+    lines.push(
+      'That covers every hook: no end-of-reply cost line, no closing line, and ' +
+        'the live panel will not animate, because nothing runs to tell it anything.'
+    );
+    if (state.guardPercent === null || state.guardPercent === undefined) {
+      lines.push('Run "mode off --guard 95" if you want one short line when the window is nearly spent.');
+    } else {
+      lines.push('The guard is set: one short line at ' + state.guardPercent + '% used, and nothing else.');
+    }
+  } else {
+    const text = DIRECTIVES[name];
+    if (text) lines.push('From the next prompt, this goes in front of the work: ' + text);
+  }
+  return lines.join('\n');
+}
+
+function main(argv) {
+  const args = (argv || []).slice();
+  const now = Date.now();
+  const flag = (name) => args.indexOf(name) !== -1;
+  const value = (name) => {
+    const at = args.indexOf(name);
+    if (at === -1) return null;
+    const next = args[at + 1];
+    return next && next.indexOf('--') !== 0 ? next : null;
+  };
+  const sessionId = value('--session-id') || process.env.CLAUDE_SESSION_ID || null;
+  const decided = resolve({ sessionId });
+
+  if (flag('--list')) return list(decided);
+  if (flag('--explain')) return explain(String(value('--explain') || '').toLowerCase());
+  if (flag('--ledger')) return ledger(now);
+  if (flag('--history')) return history(now, Number(value('--history')) || 20);
+  if (flag('--baseline')) {
+    const tier = tierNow({ sessionId, now });
+    const base = [tier.baseline.model, tier.baseline.effort].filter(Boolean).join('/') || 'not set';
+    const run = [tier.running.model, tier.running.effort].filter(Boolean).join('/') || 'unknown';
+    // Codex has no settings.json and no /model or /effort, so naming them
+    // there is telling Codex to reach for controls it does not have. usage.js
+    // is already scrupulous about this for the commands it prints; these two
+    // lines were not.
+    const codexHere = host.detect(process.argv.slice(2), process.env) === host.CODEX;
+    return [
+      'baseline  ' + base + '   (yours: ' +
+        (codexHere ? 'config.toml and the /model picker' : 'settings.json and the pickers') +
+        ', never written by this plugin)',
+      'running   ' + run + (tier.running.source ? '   (' + tier.running.source + ')' : ''),
+      '',
+      'Changing the baseline applies to NEW sessions. For the one you are in, ' +
+        (codexHere
+          ? "Codex's own model and effort controls are the only lever."
+          : 'the picker is the only lever.'),
+    ].join('\n');
+  }
+  // What is pending, read off the record of what was actually offered.
+  //
+  // It used to call usage.settingFit(null, ...) - an event list of null, which
+  // that function turns into [] and returns null from before it looks at
+  // anything - so the answer was "nothing to recommend (no measurement)" on
+  // every machine, including one with four thousand measured turns on disk.
+  // The measurement is not this command's to take: it costs a transcript scan,
+  // the brief has already paid for one, and what the brief offered is written
+  // down. So this reads that.
+  if (flag('--advice')) {
+    const state = read();
+    if (state.advice && state.advice.off) {
+      return 'Recommendations are off (mode --advice-on turns them back on). Nothing is pending.';
+    }
+    const offer = adviceLastOffer(sessionId, state);
+    if (!offer) {
+      return 'Nothing has been offered yet. A recommendation is made from a measurement the ' +
+        'briefing takes, so there is nothing to show until one has been.';
+    }
+    if (state.advice.declined && state.advice.declined[offer.id]) {
+      return 'Nothing pending: "' + offer.id + '" was offered and declined, and will not be raised again.';
+    }
+    return (offer.text || 'Pending: ' + offer.id + '.') +
+      '\nTo say no to it, and never see it again: mode --decline';
+  }
+  // Somebody has to be able to say no, or "a declined recommendation is never
+  // raised again" is a promise with no way to keep it. This is the verb for
+  // the moment the user says "no, leave it": it records the id and that
+  // recommendation is never volunteered again, in this session or any later
+  // one. With no id it declines whatever is currently pending.
+  if (flag('--decline')) {
+    const given = value('--decline');
+    // Same fix as --advice, and it matters more here: this is the documented
+    // way for the user to say no, and with the id taken from a settingFit()
+    // call that could only ever return null, saying no recorded nothing and
+    // the same recommendation came back in the next session.
+    const id = given || (adviceLastOffer(sessionId) || {}).id || null;
+    if (!id) return 'There is nothing pending to decline.';
+    adviceDecline(id, now);
+    logChange({ plane: 'mode', key: 'advice', from: id, to: 'declined', by: 'user', reason: null }, now);
+    return 'Declined, and remembered: "' + id + '" will not be suggested again.';
+  }
+  if (flag('--no-advice')) {
+    adviceMute(true);
+    logChange({ plane: 'mode', key: 'advice', from: 'on', to: 'off', by: 'user', reason: null }, now);
+    return 'Recommendations are off. Nothing will be suggested about your own settings again until: mode --advice-on';
+  }
+  if (flag('--advice-on')) {
+    adviceMute(false);
+    logChange({ plane: 'mode', key: 'advice', from: 'off', to: 'on', by: 'user', reason: null }, now);
+    return 'Recommendations are back on, capped at one per session and never repeated once declined.';
+  }
+  if (flag('--pin') || flag('--no-pin')) {
+    const state = read();
+    const before = state.pin;
+    state.pin = flag('--pin');
+    write(state);
+    logChange({ plane: 'mode', key: 'pin', from: before, to: state.pin, by: 'user', reason: null }, now);
+    return state.pin
+      ? 'Pinned. Nothing will self-switch: the gap between your baseline and what is running is reported and left alone.'
+      : 'Unpinned. Self-switching is back to what the mode says.';
+  }
+  // The bounds, and the answer for each one asked for.
+  //
+  // Returned as a list rather than straight out of the loop, because this ran
+  // BEFORE the positional argument was looked at: `mode max --floor sonnet`
+  // set the floor, said so, and returned - leaving the mode untouched and
+  // unmentioned. The user asked for two things and was told about one.
+  const boundLines = [];
+  for (const key of ['--floor', '--ceiling']) {
+    if (!flag(key)) continue;
+    const raw = value(key);
+    const field = key.slice(2);
+    const state = read();
+    const before = tierText(state[field]);
+    if (!raw || raw === 'none' || raw === 'off') {
+      state[field] = null;
+      write(state);
+      logChange({ plane: 'mode', key: field, from: before, to: null, by: 'user', reason: null }, now);
+      boundLines.push('The ' + field + ' is cleared.');
+      continue;
+    }
+    const tier = parseTier(raw);
+    if (!tier || tier.error) {
+      return 'Could not read a tier from "' + raw + '": ' + ((tier && tier.error) || 'nothing recognised') + '. Try sonnet/medium.';
+    }
+    state[field] = tier;
+    write(state);
+    logChange({ plane: 'mode', key: field, from: before, to: tierText(tier), by: 'user', reason: null }, now);
+    const caveat = thinkingCaveat(tier);
+    boundLines.push(
+      'The ' + field + ' is ' + tierText(tier) + '. Nothing the plugin says will point ' +
+        (field === 'floor' ? 'below' : 'above') + ' it.' + (caveat ? '\nNote: ' + caveat : '')
+    );
+  }
+
+  const first = String(args[0] || '').toLowerCase();
+  if (first === 'undo') return undo(now).text;
+  if (first === 'auto') {
+    const state = read();
+    const before = state.auto;
+    const off = String(args[1] || '').toLowerCase() === 'off';
+    state.auto = !off;
+    write(state);
+    logChange({ plane: 'mode', key: 'auto', from: before, to: state.auto, by: 'user', reason: null }, now);
+    if (off) return 'auto is off. The mode is ' + state.mode + ' until you change it.';
+    return [
+      'auto is on. It picks from pressure and always reports which one it picked:',
+      '  under 50% used and roomy   -> standard',
+      '  50 to 79                   -> high',
+      '  80+, or tight or gone      -> max',
+      'It never picks off. Turning the plugin off is a decision a person makes.',
+    ].join('\n');
+  }
+  if (first && first.indexOf('--') !== 0) {
+    const named = normalise(first);
+    if (!named) return 'No such mode: ' + first + '.\n\n' + list(decided);
+    if (named.ambiguous) return named.message;
+    if (named.auto) return main(['auto'].concat(args.slice(1)));
+    const guard = guardValue(args, flag, value);
+    if (guard.error) return guard.error;
+    const ceilingArg = ceilingValue(args, flag, value);
+    if (ceilingArg.error) return ceilingArg.error;
+    const set = setMode(named.mode, {
+      session: flag('--session'),
+      sessionId,
+      guard: guard.percent,
+      ceiling: ceilingArg.percent,
+    }, now);
+    if (ceilingArg.clear) clearCeiling(now);
+    const lines = [set];
+    if (ceilingArg.percent !== undefined || ceilingArg.clear) lines.push(ceilingLine());
+    return lines.concat(boundLines).join('\n');
+  }
+
+  // The usage cap on its own, without changing the mode.
+  //
+  // Spelled --cap, NOT --ceiling: this file already uses --ceiling for the
+  // model/effort bound ("never go above opus/xhigh"), and taking that spelling
+  // swallowed it - mode --ceiling opus/xhigh started answering "could not read a
+  // percentage". The two are unrelated
+  // settings: the mode says how loudly the plugin talks, the ceiling says where
+  // it stops the work.
+  if (flag('--cap')) {
+    const ceilingArg = ceilingValue(args, flag, value);
+    if (ceilingArg.error) return ceilingArg.error;
+    const state = read();
+    const previous = state.ceilingPercent;
+    state.ceilingPercent = ceilingArg.clear ? null : ceilingArg.percent;
+    write(state);
+    logChange(
+      { plane: 'mode', key: 'ceiling', from: previous, to: state.ceilingPercent, by: 'user', reason: null },
+      now
+    );
+    return ceilingLine();
+  }
+
+  if (boundLines.length) return boundLines.join('\n');
+  return describe(decided, now);
+}
+
+function clearCeiling(now) {
+  const state = read();
+  const previous = state.ceilingPercent;
+  state.ceilingPercent = null;
+  write(state);
+  logChange({ plane: 'mode', key: 'ceiling', from: previous, to: null, by: 'user', reason: null }, now);
+}
+
+// What the ceiling is now, and what it will actually do. Said in terms of the
+// refusal, because a percentage on its own does not tell anyone what changes.
+function ceilingLine() {
+  const state = read();
+  if (state.ceilingPercent === null || state.ceilingPercent === undefined) {
+    return 'Ceiling off. Nothing is refused; the plugin reports and does not intervene.';
+  }
+  return (
+    'Ceiling ' + state.ceilingPercent + '%. Past that, fan-out calls (Agent, Task, Workflow and ' +
+    'their equivalents) are refused at the hook, in every session on this machine. Nothing else ' +
+    'is blocked, so the work still finishes - sequentially, in one session, which is where the ' +
+    'saving comes from. "mode --cap off" removes it.'
+  );
+}
+
+// The ceiling percentage, or the reason it was refused. Same validation as the
+// guard, and for the same reason: a number accepted without being read is a
+// setting that silently does the opposite of what was asked.
+//
+// A ceiling differs from the guard in one way - "off" is a meaningful value,
+// because a ceiling is the one setting here that takes something away, and
+// taking it back has to be as easy as setting it.
+function ceilingValue(args, flag, value) {
+  if (!flag('--cap')) return { percent: undefined, clear: false };
+  const token = value('--cap');
+  if (token === null) {
+    return { error: '--cap needs a percentage or "off", for example: mode --cap 60' };
+  }
+  const text = String(token).trim().toLowerCase();
+  if (text === 'off' || text === 'none' || text === 'no') return { percent: undefined, clear: true };
+  const raw = Number(text.replace(/%$/, ''));
+  if (!Number.isFinite(raw)) {
+    return { error: 'Could not read a percentage from "' + token + '". Try: mode --cap 60' };
+  }
+  if (raw < 1 || raw > 100) {
+    return {
+      error:
+        'The ceiling is a percentage of a window, so it has to be between 1 and 100. ' +
+        (raw > 100
+          ? String(raw) + ' can never be reached, so nothing would ever be refused.'
+          : String(raw) + ' would refuse every fan-out from the start of the window.'),
+    };
+  }
+  return { percent: raw, clear: false };
+}
+
+// The guard percentage, or the reason it was refused.
+//
+// Three ways this went wrong and all three are the same shape - a number that
+// was accepted without being read.
+//
+//   `mode off --guard` with no value: value() returns null when the next token
+//   is missing, and Number(null) is 0, which is finite. That stored a guard of
+//   zero, which fires on every prompt at any usage - the exact opposite of
+//   what "off" was asked for, and described in the reply as "one short line at
+//   0% used".
+//
+//   `--guard 500`: never fires, so the user asked for a line at the wall and
+//   silently has none. That is the one scenario the guard exists for.
+//
+//   `--guard 0`: a line on every prompt, as above.
+//
+// A percentage of a window is a number between 1 and 100. Anything else is
+// refused with the reason, because a guard that silently does nothing is worse
+// than no guard at all.
+function guardValue(args, flag, value) {
+  if (!flag('--guard')) return { percent: undefined };
+  const token = value('--guard');
+  if (token === null) {
+    return { error: '--guard needs a percentage, for example: mode off --guard 95' };
+  }
+  const raw = Number(token);
+  if (!Number.isFinite(raw)) {
+    return { error: 'Could not read a percentage from "' + token + '". Try: mode off --guard 95' };
+  }
+  if (raw < 1 || raw > 100) {
+    return {
+      error:
+        'The guard is a percentage of a window, so it has to be between 1 and 100. ' +
+        (raw > 100
+          ? String(raw) + ' can never be reached, so the line would never appear.'
+          : String(raw) + ' is reached immediately, so the line would appear on every prompt.'),
+    };
+  }
+  return { percent: raw };
+}
+
+if (require.main === module) {
+  try {
+    process.stdout.write(main(process.argv.slice(2)) + '\n');
+  } catch (err) {
+    process.stdout.write('mode: ' + err.message + '\n');
+  }
+  process.exit(0);
+}
+
+module.exports = {
+  MODES,
+  ORDER,
+  WIRED,
+  ALIASES,
+  AMBIGUOUS,
+  DIRECTIVES,
+  DEFAULT_MODE,
+  EFFORT_ORDER,
+  MODEL_ORDER,
+  THINKING_ONLY_EFFORTS,
+  KEEP_CHANGES,
+  KEEP_DECLINED,
+  configDir,
+  modeFile,
+  changesFile,
+  empty,
+  read,
+  write,
+  normalise,
+  ambiguityText,
+  parseTier,
+  tierText,
+  effortRank,
+  modelRank,
+  allows,
+  boundsNote,
+  thinkingCaveat,
+  directive,
+  autoPick,
+  readingNow,
+  resolve,
+  forSession,
+  readChanges,
+  logChange,
+  lastUndoable,
+  history,
+  undo,
+  adviceId,
+  advicePending,
+  adviceOffer,
+  adviceLastOffer,
+  adviceDecline,
+  adviceMute,
+  tierNow,
+  tierLine,
+  ledger,
+  explain,
+  list,
+  describe,
+  setMode,
+  guardValue,
+  main,
+};