acuvo-code 0.5.4 → 0.6.0

package/lib/chat.mjs

@@ -1,536 +1,545 @@
-/**
- * ── ⭐ THE INTERACTIVE SESSION ───────────────────────────────────────────────
- *
- * Every invocation of `acuvo "task"` started COLD: it re-gathered the workspace,
- * rebuilt the prompt, and knew nothing about the last thing you asked. So the
- * second instruction cost as much as the first, and "now do the same for the
- * other file" was not a sentence you could say.
- *
- * ⭐ AND THE ECONOMICS ARE THE ARGUMENT, NOT JUST THE ERGONOMICS. Measured
- * 2026-08-09: an identical prompt prefix cached at **97.2%**, dropping the call
- * cost **4.3x** ($0.000836 → $0.000195). A session that APPENDS keeps that
- * prefix intact, so every turn after the first is nearly free. A tool that
- * rebuilds its prompt each time throws that away and looks identical from the
- * outside — which is exactly why this is worth building rather than assuming.
- *
- * ── ⚠️ WHY NOT A FULL TUI ────────────────────────────────────────────────────
- * `readline` and plain writes, no alternate screen buffer, no cursor addressing.
- * A TUI that redraws breaks `>` redirection, breaks piping into a file, breaks
- * `tee`, and breaks every terminal that is not the one it was tested in. The
- * output here is append-only text, so a session transcript is a file you can
- * keep. That is a deliberate trade of polish for portability.
- */
-
-import { createInterface } from 'node:readline';
-import { readBoxedLine, pinRegion } from './input-box.mjs';
-import { EXIT_INTERRUPTED } from './interrupt.mjs';
-import { parseSlash, runSlashCommand } from './slash.mjs';
-import { estimateMessagesTokens } from './compact.mjs';
-
-/** What ends a session. `exit`/`quit` because both are muscle memory. */
-const QUIT = new Set(['exit', 'quit', ':q', 'bye']);
-
-/**
- * ⚠️ CONTEXT GROWS UNBOUNDED AND A CODING SESSION IS THE WORST CASE — tool
- * results carry whole files. Left alone, turn 30 sends everything from turns
- * 1-29 and eventually 400s on a context-length error mid-thought.
- *
- * The trim keeps the SYSTEM message and the FIRST user message (the workspace
- * context — the cacheable prefix, dropping it would cost more than it saves) and
- * discards the oldest middle turns.
- */
-export const MAX_HISTORY_MESSAGES = 40;
-
-/**
- * ── 💰⭐⭐⭐ THE TRIM WAS A CONTINUOUS SLIDE, AND IT COST ~90% OF THE CACHE ──
- *
- * `messages.slice(-keep)` re-slices on EVERY turn once a session passes the
- * cap, so the third message of the prompt is different every time. Prefix
- * caching matches from the first token and stops at the first difference — so
- * everything after the 2-message head was re-bought at full price, every turn,
- * for the rest of the session.
- *
- * ⭐ MEASURED (no credits, no network — a simulated 60-turn session counting how
- * much of each prompt is byte-identical to the previous one):
- *
- *     CURRENT slice(-38)        cache  9.2%    history 38-40
- *     stepped high=48 low=40    cache 74.5%    history 38-48
- *     stepped high=56 low=40    cache 85.3%    history 38-56
- *     stepped high=64 low=40    cache 89.7%    history 38-64   <- shipped
- *     stepped high=72 low=40    cache 92.0%    history 38-72
- *
- * ⚠️ THE HEAD BEING STABLE IS NOT THE SAME AS THE PROMPT BEING CACHED, and
- * measuring position 0 alone would have said this code was fine. The 2-message
- * head never moved; the cacheable prefix was still 2 messages out of 40.
- *
- * ⭐ `LOW` IS TODAY'S CAP ON PURPOSE, so this can only ever keep MORE history
- * than the rule it replaces — never less. That makes it a pure win rather than
- * a trade of memory for money, and it is why HIGH was raised instead of LOW
- * being lowered (console shipped 16/8, halving its window; the CLI holds whole
- * files in tool results and cannot afford to forget more).
- *
- * ⚠️ RAISING `HIGH` COSTS TOKENS PER REQUEST, and the ceiling that matters is
- * not here: `turn.mjs` compacts above CONTEXT_BUDGET_TOKENS (96k), which
- * `compact.mjs` warns is the moment the cache discount dies for good. This is a
- * MESSAGE-COUNT backstop, not the token bound — 64 messages of ordinary turns
- * sit far under 96k, but a session whose tool results carry whole files can
- * approach it, and then compaction (with its own hysteresis) takes over. 72 and
- * 80 buy 2 more points of cache for materially more of that risk, which is why
- * 64 is the shipped number and not the best one in the table.
- */
-export const HISTORY_HIGH_WATER = 64;
-export const HISTORY_LOW_WATER = MAX_HISTORY_MESSAGES;
-
-/**
- * ⚠️ WELL UNDER `turn.mjs`'s CONTEXT_BUDGET_TOKENS (96,000), not near it. The
- * request also carries the TOOL OFFER — measured at ~15,200 tokens for the
- * CLI's 63 verbs — plus the system prompt, so history must leave room for both
- * and still clear the line where compaction fires. 55k + ~15k offer + prompt
- * sits comfortably inside 96k.
- *
- * ⭐ This is a CEILING, not a target: an ordinary session (~551 tokens/message,
- * measured) reaches the 64-message cap at ~35k and never comes near it. It
- * exists for the file-heavy session, which is exactly the one that would
- * otherwise be pushed into permanent compaction by the higher message cap.
- */
-export const HISTORY_TOKEN_CEILING = 55_000;
-
-export function trimHistory(messages, max = HISTORY_HIGH_WATER) {
-  if (!Array.isArray(messages)) return messages;
-  /**
-   * ⚠️⚠️ THE GATE MUST ASK BOTH QUESTIONS, AND MY FIRST VERSION ASKED ONLY ONE.
-   * It read `messages.length <= max` and returned early — so the token ceiling
-   * below was unreachable for exactly the session it was written for: 64 heavy
-   * messages are under the COUNT cap and ~137,000 tokens, and the function
-   * handed them straight back. The test reported 133,491 tokens and I had to
-   * measure to find that the ceiling was never consulted at all rather than
-   * being wrong.
-   */
-  const overBudget = estimateMessagesTokens(messages) > HISTORY_TOKEN_CEILING;
-  if (messages.length <= max && !overBudget) return messages;
-  const head = messages.slice(0, 2);          // system + the context-bearing user turn
-  /**
-   * ⭐ THE HEAD MOVES IN STEPS, NOT EVERY TURN. `dropped` is a multiple of STEP,
-   * so it changes only when the conversation crosses the next boundary — and
-   * between boundaries the prompt is byte-identical to the previous turn's.
-   *
-   * ⚠️ NOT `length > HIGH ? slice(-LOW)`. That is still a continuous slide past
-   * the mark and is the exact mistake console's own guard records catching
-   * before it shipped.
-   */
-  /**
-   * ⚠️⚠️ THE LOW WATER MUST BE DERIVED FROM `max`, NOT PINNED TO A CONSTANT,
-   * AND I ALMOST SHIPPED IT PINNED. My first version was
-   * `Math.min(HISTORY_LOW_WATER, max)`. `runChat`'s default `maxHistory` was
-   * MAX_HISTORY_MESSAGES (40), which equals HISTORY_LOW_WATER — so `step`
-   * became `max(1, 0)` = 1, and one-message steps ARE the continuous slide this
-   * whole change exists to remove. The fix would have measured 89.6% in a
-   * simulation and done NOTHING on the only path that calls it.
-   *
-   * ⭐ 0.625 is 40/64 — the shipped ratio, so the default keeps exactly the
-   * numbers that were measured, and any other `max` still steps properly
-   * instead of silently degrading to a slide.
-   */
-  const low = Math.min(HISTORY_LOW_WATER, Math.floor(max * 0.625));
-  const step = Math.max(1, max - low);
-  const body = messages.slice(head.length);
-  const overflow = Math.max(0, messages.length - max);
-  let dropped = Math.ceil(overflow / step) * step;
-  /**
-   * ── ⚠️⚠️ AND A TOKEN CEILING, BECAUSE MESSAGE COUNT IS THE WRONG UNIT ──────
-   *
-   * MEASURED against the one real recorded session on disk (93,036 prompt
-   * tokens, 18 messages, `.acuvo/sessions/20260814-095636-99d4.json`):
-   *
-   *     avg message  ~551 tokens        largest observed  ~2,145 tokens
-   *
-   *     at 40 msgs: typical ~22,000   worst case (all like the largest) ~85,800
-   *     at 64 msgs: typical ~35,300   worst case                       ~137,300
-   *
-   * ⚠️ SO RAISING THE COUNT ALONE MOVES THE WORST CASE FROM JUST UNDER THE
-   * 96k CONTEXT BUDGET TO WELL OVER IT. Past that line `turn.mjs` compacts, and
-   * `compact.mjs` is explicit that compaction voids the cache discount and
-   * "once it starts, it never stops" — so a change made ENTIRELY to protect the
-   * cache would, in a file-heavy session, destroy it. A tool result carrying a
-   * whole file is not the exception in a coding session; it is the normal case.
-   *
-   * ⭐ THE CEILING DROPS IN THE SAME `step` BLOCKS. Any multiple of `step`
-   * keeps block alignment, so the prefix still changes only at boundaries and
-   * the caching win is untouched — this bounds the worst case without
-   * reintroducing a slide.
-   */
-  const budget = Math.max(1, HISTORY_TOKEN_CEILING - estimateMessagesTokens(head));
-  while (dropped < body.length && estimateMessagesTokens(body.slice(dropped)) > budget) {
-    dropped += step;
-  }
-  let tail = body.slice(dropped);
-  /**
-   * ⚠️ A `tool` MESSAGE WITHOUT ITS `assistant` TOOL CALL IS A HARD 400 from
-   * every OpenAI-shaped provider — "tool_call_id did not have a preceding
-   * message with tool_calls". Slicing mid-exchange produces exactly that, and it
-   * would surface as a mysterious API error thirty turns into a session.
-   */
-  while (tail.length > 0 && tail[0].role === 'tool') tail = tail.slice(1);
-  return [...head, ...tail];
-}
-
-/**
- * One prompt line. Returns null on EOF (Ctrl-D, or a pipe that ran out).
- *
- * ⚠️ THE CLOSED CHECK IS NOT DEFENSIVE, IT IS THE PIPED CASE. Found by piping a
- * list of prompts in: readline emits 'close' when the stream ends, and the NEXT
- * `rl.question()` throws ERR_USE_AFTER_CLOSE. The first turn had already
- * succeeded and written a real file, so the session crashed AFTER doing its job
- * — the worst shape of failure, because the work looks lost.
- *
- * Scripted input matters beyond tests: piping a prompt list is how anyone would
- * automate this.
- */
-function ask(rl, prompt, state) {
-  if (state.closed) return Promise.resolve(null);
-  return new Promise((resolve) => {
-    let answered = false;
-    const onClose = () => { if (!answered) resolve(null); };
-    rl.once('close', onClose);
-    rl.question(prompt, (line) => {
-      answered = true;
-      rl.removeListener('close', onClose);
-      resolve(line);
-    });
-  });
-}
-
-/**
- * Run an interactive session.
- *
- * `runOne(task, priorMessages)` performs one turn and returns the session
- * outcome — injected rather than imported so this loop is testable with a stub
- * and never needs a model or a terminal in a test.
- */
-/**
- * ── ⚠️ PIPED INPUT IS A DIFFERENT PROBLEM AND NEEDED A DIFFERENT ANSWER ──────
- * `readline` on a non-TTY DRAINS the stream as fast as it can and emits 'close'
- * the moment it ends. The model call for turn 1 takes seconds, by which point
- * the interface is already closed and every later prompt is lost — measured:
- * a three-line pipe ran exactly ONE turn and exited quietly, which is worse than
- * crashing because it looks like it worked.
- *
- * So a pipe is read WHOLE and replayed from a queue. A TTY keeps the real
- * readline loop, where a human types the next line after seeing the last answer.
- * Two input shapes, two mechanisms — pretending they are the same is what broke.
- */
-async function readAllLines(input) {
-  const chunks = [];
-  for await (const chunk of input) chunks.push(chunk);
-  return Buffer.concat(chunks.map((c) => (typeof c === 'string' ? Buffer.from(c) : c)))
-    .toString('utf8')
-    .split(String.fromCharCode(10))
-    .map((l) => l.replace(String.fromCharCode(13), ''));
-}
-
-/**
- * ── ⚠️⚠️⭐ READLINE EATS CTRL-C. MEASURED IN NODE'S OWN SOURCE ──────────────
- *
- * This is the finding that made an interrupt handler in `bin/acuvo.mjs`
- * necessary-but-not-sufficient. Read out of `process.binding('natives')` on
- * node v22.17.0, `internal/readline/interface.js`, the ttyWrite ctrl-key
- * switch, verbatim:
- *
- *     case 'c':
- *       if (this.listenerCount('SIGINT') > 0) {
- *         this.emit('SIGINT');
- *       } else {
- *         // This readline instance is finished
- *         this.close();
- *         this[kQuestionReject]?.(new AbortError('Aborted with Ctrl+C'));
- *       }
- *
- * ⚠️ So with a TTY readline open and NO `'SIGINT'` listener on the interface,
- * Ctrl-C never reaches `process.on('SIGINT')` at all — readline just closes
- * itself. The run in flight would have carried on to completion, for minutes,
- * with the user's Ctrl-C having produced nothing on screen. That is strictly
- * worse than the bug we set out to fix, and no amount of correct handling in
- * `bin/acuvo.mjs` would have been reached.
- *
- * ⭐ So the interface takes a listener whose whole job is to hand the signal
- * back to the process, where the one policy in `interrupt.mjs` decides between
- * "stop after this round" and "quit now".
- *
- * ── ⚠️⚠️ AND A SYNTHETIC EMIT INTO AN EMPTY EMITTER DOES NOTHING ────────────
- *
- * `process.emit('SIGINT')` is plain `EventEmitter.emit` — it does NOT invoke
- * the OS default action. `turn.mjs` installs the process signal handlers inside
- * `runSession`, so before the first turn of a session there are ZERO listeners
- * and the emit would return `false` having done absolutely nothing. Ctrl-C at
- * the very first prompt would be inert.
- *
- * ⭐ Hence the count check and the explicit exit: **every path out of this
- * function either aborts a run or ends the process.** That is the rule this
- * whole feature is built on, and it is the one that is easy to break here.
- */
-export function deliverInterrupt({
-  emit = (sig) => process.emit(sig),
-  listenerCount = (sig) => process.listenerCount(sig),
-  exit = (code) => process.exit(code),
-} = {}) {
-  if (listenerCount('SIGINT') > 0) {
-    emit('SIGINT');
-    return 'delegated';
-  }
-  exit(EXIT_INTERRUPTED);
-  return 'exited';
-}
-
-export async function runChat({
-  runOne,
-  render,
-  input = process.stdin,
-  output = process.stdout,
-  banner = '',
-  /**
-   * ⚠️ THE HIGH WATER MARK, NOT THE OLD FLAT CAP. This default is what makes
-   * the stepped trim REACH the live session — see the note in `trimHistory`
-   * about the version of this fix that measured 89.6% and changed nothing.
-   */
-  maxHistory = HISTORY_HIGH_WATER,
-  /**
-   * ⚠️ INJECTED SO A TEST CAN SEE IT. The real one exits the process, and a
-   * test that could not substitute it could only assert this feature by killing
-   * its own runner. Default is production behaviour, so no caller changes.
-   */
-  onInterrupt = deliverInterrupt,
-  /**
-   * ── ⭐ THE `/` SURFACE'S ONE SEAM ─────────────────────────────────────────
-   *
-   * Providers for the things a command reports on — skills, MCP servers, spend,
-   * the model. `bin/` owns where those facts come from; this loop only asks, for
-   * the same reason `workspace.mjs` takes `claimPath` and `journal` injected
-   * rather than importing them.
-   *
-   * ⚠️ `{}` BY DEFAULT, NOT `null`. Every command degrades to "not available in
-   * this session" on a missing provider (see `slash.mjs`), so an embedder that
-   * wires nothing still gets a working `/help` instead of a crash.
-   */
-  slashContext = {},
-}) {
-  const interactive = input.isTTY === true;
-
-  // ⚠️ A pipe is drained up front — see readAllLines. Doing this lazily is what
-  // silently lost every prompt after the first.
-  const queued = interactive ? null : await readAllLines(input);
-  let queueIndex = 0;
-
-  /**
-   * ── ⚠️⚠️ READLINE IS GONE FROM THE INTERACTIVE PATH, AND IT HAD TO GO ──────
-   *
-   * `input-box.mjs` now owns the keyboard. Leaving the readline interface
-   * attached to the SAME stream was not merely redundant — both it and the box
-   * saw every Ctrl-C, so `onInterrupt` fired TWICE for one keypress. Measured,
-   * not theorised: a single `\x03` produced `['interrupt', 'interrupt']`.
-   *
-   * ⭐ A DOUBLE INTERRUPT IS NOT A COSMETIC BUG. The second Ctrl-C is the one
-   * that QUITS — so one press would have armed and fired the escape hatch in the
-   * same instant, ending a session the user meant only to nudge.
-   *
-   * The Node-internals note below about `question()` swallowing Ctrl-C is kept
-   * because it is why the box reads raw keys itself rather than asking readline
-   * for a line. It is history now, not a live constraint.
-   */
-  const rl = null;
-  const state = { closed: false };
-  /**
-   * ⚠️ ATTACHED ONCE, NOT PER QUESTION — and the per-question version is why the
-   * first fix did not work. The stream can end WHILE the model call is in
-   * flight, when no question is pending and therefore no listener is attached;
-   * `close` fires into nothing, the flag stays false, and the next question
-   * throws anyway. A session-lifetime listener sees it whenever it happens.
-   */
-  if (rl) rl.once('close', () => { state.closed = true; });
-  /**
-   * ⭐ THE ONE LINE THAT MAKES CTRL-C REACH THE RUN — see `deliverInterrupt`
-   * above for the Node source that proves it is needed. `on`, not `once`: an
-   * interactive session runs many turns and the SECOND Ctrl-C (the one that
-   * quits) has to arrive here too, or the escape hatch is a single-use one.
-   */
-  if (rl) rl.on('SIGINT', () => { onInterrupt(); });
-  /**
-   * ⚠️⚠️ THE BANNER IS DEFERRED UNTIL AFTER THE PIN, AND THE ORDER IS NOT
-   * COSMETIC. `pinRegion` CLEARS the screen so the transcript starts at the top
-   * of a clean one — printed before that, the banner is erased by the very
-   * thing meant to sit beneath it, which is exactly what a screenshot showed:
-   * a banner cut in half with the conversation crammed at the bottom.
-   */
-  const writeBanner = () => { if (banner) output.write(`${banner}\n`); };
-  // ⚠️ `/help` IS ADVERTISED IN THE ONE LINE EVERY SESSION PRINTS. A command
-  // surface nobody is told about is the same defect item 14 closed for `--help`:
-  // the feature worked and nothing a stranger would read mentioned it.
-  /**
-   * ⚠️ A BLANK LINE BEFORE IT, not just after. Roman: *"move it one line down."*
-   * Pressed against the banner it read as a fifth detail row rather than as an
-   * instruction addressed to the person.
-   */
-  output.write('\nType what you want done. "/help" for commands, "exit" to leave.\n\n');
-
-  let history = null;
-  let turns = 0;
-  /**
-   * ⭐ SET BY `/skills <name>`, CONSUMED BY THE NEXT REAL TURN AND THEN CLEARED.
-   * A skill that was printed to the terminal would look loaded and be invisible
-   * to the model; this is the variable that makes the verb real.
-   */
-  let pendingInject = null;
-  /** Lines the user has submitted this session — the box's Up/Down history. */
-  const typed = [];
-
-  /**
-   * ── ⭐⭐⭐ THE BOX IS PINNED TO THE BOTTOM OF THE SCREEN ────────────────────
-   *
-   * Roman: *"we need that prompt box stuck down the bottom, it is professional."*
-   *
-   * `pinRegion` reserves the last three rows, so output scrolls ABOVE the box
-   * and the box itself never moves. Off a TTY, in CI, on a short terminal, or
-   * with ACUVO_NO_PIN=1 it does nothing at all.
-   *
-   * ⚠️⚠️ RELEASED ON EVERY PATH OUT, INCLUDING THE ONES NOBODY PLANS FOR. A
-   * process that exits with a scroll region still set leaves the user with a
-   * terminal that scrolls inside a box until they type `reset` blind — the same
-   * class of harm as leaving raw mode on. The `finally` covers a normal end and
-   * a throw; the signal handlers cover Ctrl-C and `kill`.
-   */
-  const pin = interactive ? pinRegion(output) : { enabled: false, release() {} };
-  const releasePin = () => pin.release();
-  if (pin.enabled) {
-    process.once('exit', releasePin);
-    process.once('SIGINT', releasePin);
-    process.once('SIGTERM', releasePin);
-  }
-
-  /**
-   * ⭐ NOW the banner, at the top of a screen the pin has just cleared, with the
-   * conversation free to grow downward beneath it — which is the direction every
-   * scrollback in existence runs.
-   */
-  writeBanner();
-
-  try {
-    for (;;) {
-      /**
-       * ── ⭐⭐ THE INPUT IS A BOX, BECAUSE A BARE `› ` IS NOT A PLACE TO TYPE ──
-       *
-       * Roman, comparing against a real Claude Code screenshot: *"you can see
-       * the box where you type, acuvo doesn't have that."* He is right — the
-       * prompt was two characters floating in the scrollback, which reads as
-       * output rather than as somewhere input goes.
-       *
-       * ⚠️ THE BOX OPENS BEFORE THE LINE AND CLOSES AFTER IT, deliberately.
-       * Drawing all four sides up front needs absolute cursor control, and
-       * readline owns the cursor once `question()` starts — every version of
-       * that fights the line editor the moment input wraps, a history entry is
-       * recalled, or the window is resized. Opening on entry and closing on
-       * submit needs no cursor math at all, survives all three, and leaves a
-       * transcript where each turn is visibly a closed unit.
-       *
-       * ⚠️ WIDTH IS READ FRESH EACH TURN — a terminal resized mid-session would
-       * otherwise draw rules at the old width for the rest of the run. Clamped,
-       * because `columns` is `undefined` when stdout is not a TTY and enormous
-       * when someone maximises on an ultrawide.
-       */
-      let line;
-      if (interactive) {
-        output.write('\n');
-        const got = await readBoxedLine({
-          atRow: pin.enabled ? pin.bottom + 1 : 0,
-          input,
-          output,
-          history: typed,
-          onInterrupt,
-        });
-        line = got.value;
-      } else {
-        line = queueIndex < queued.length ? queued[queueIndex++] : null;
-      }
-      // Echo a piped prompt so a scripted transcript reads like a session.
-      if (!interactive && line !== null && line.trim()) output.write(`› ${line.trim()}
-`);
-      // ⚠️ EOF is not an error. A closed pipe or Ctrl-D ends the session the
-      // same way "exit" does — treating it as a fault would print a stack trace
-      // at the end of every scripted run.
-      if (line === null) break;
-      const task = line.trim();
-      if (!task) continue;
-      // ⚠️ Deduped against the PREVIOUS entry only: pressing Up should walk
-      // distinct instructions, not scroll through five copies of `npm test`.
-      if (typed[typed.length - 1] !== task) typed.push(task);
-      if (QUIT.has(task.toLowerCase())) break;
-
-      /**
-       * ── ⭐ THE `/` SURFACE, BEFORE ANYTHING IS SENT TO A MODEL ────────────
-       *
-       * ⚠️ IT COSTS NOTHING AND MUST NOT COUNT AS A TURN. `/cost` is asked
-       * precisely by somebody watching their spend, and answering it by
-       * incrementing the turn counter and appending to the history would make
-       * the question change the answer.
-       *
-       * ⚠️ AN UNRECOGNISED COMMAND IS ANSWERED HERE AND NOT FORWARDED. Passing
-       * `/skil` to the model gets a confident essay about a typo; the one thing
-       * the person needed was the word `/skills`, which `slash.mjs` supplies.
-       * See its header for why `/etc/hosts` is NOT treated as a command.
-       */
-      const command = parseSlash(line.trim());
-      if (command) {
-        const result = runSlashCommand(command, slashContext);
-        for (const l of result.output ?? []) output.write(`${l}\n`);
-        output.write('\n');
-        if (result.effect === 'clear') history = null;
-        if (typeof result.inject === 'string' && result.inject) pendingInject = result.inject;
-        continue;
-      }
-
-      /**
-       * ⚠️ THE SKILL IS PREPENDED TO THE TASK, NOT SUBSTITUTED FOR IT. The user
-       * typed an instruction; the skill is context for it. And it is cleared
-       * BEFORE the call rather than after, so a turn that throws cannot leave it
-       * armed and silently attach it to an unrelated question later.
-       */
-      let sendTask = task;
-      if (pendingInject) {
-        sendTask = `${pendingInject}\n\n---\n\n${task}`;
-        pendingInject = null;
-      }
-
-      let outcome;
-      try {
-        outcome = await runOne(sendTask, history);
-      } catch (err) {
-        /**
-         * ⚠️ ONE BAD TURN MUST NOT END THE SESSION. A timeout or a provider blip
-         * after twenty minutes of context is infuriating if it drops everything;
-         * the history is still valid, so report and keep the prompt.
-         */
-        output.write(`\n  ✖ that turn failed: ${String(err?.message || err)}\n\n`);
-        continue;
-      }
-
-      turns += 1;
-      render(outcome, output);
-
-      if (outcome?.ok && Array.isArray(outcome.messages)) {
-        history = trimHistory(outcome.messages, maxHistory);
-      } else if (!outcome?.ok) {
-        // A failed turn leaves history UNTOUCHED. Appending a turn that produced
-        // nothing would poison the next one with a dead exchange.
-        output.write(`\n  (history unchanged — that turn did not complete)\n`);
-      }
-      output.write('\n');
-    }
-  } finally {
-    if (rl) rl.close();
-    pin.release();
-    // ⚠️ The listeners come off too — a long-lived process that started several
-    // sessions would otherwise accumulate them and warn about a leak.
-    process.off('exit', releasePin);
-    process.off('SIGINT', releasePin);
-    process.off('SIGTERM', releasePin);
-  }
-  return { turns };
-}
+/**
+ * ── ⭐ THE INTERACTIVE SESSION ───────────────────────────────────────────────
+ *
+ * Every invocation of `acuvo "task"` started COLD: it re-gathered the workspace,
+ * rebuilt the prompt, and knew nothing about the last thing you asked. So the
+ * second instruction cost as much as the first, and "now do the same for the
+ * other file" was not a sentence you could say.
+ *
+ * ⭐ AND THE ECONOMICS ARE THE ARGUMENT, NOT JUST THE ERGONOMICS. Measured
+ * 2026-08-09: an identical prompt prefix cached at **97.2%**, dropping the call
+ * cost **4.3x** ($0.000836 → $0.000195). A session that APPENDS keeps that
+ * prefix intact, so every turn after the first is nearly free. A tool that
+ * rebuilds its prompt each time throws that away and looks identical from the
+ * outside — which is exactly why this is worth building rather than assuming.
+ *
+ * ── ⚠️ WHY NOT A FULL TUI ────────────────────────────────────────────────────
+ * `readline` and plain writes, no alternate screen buffer, no cursor addressing.
+ * A TUI that redraws breaks `>` redirection, breaks piping into a file, breaks
+ * `tee`, and breaks every terminal that is not the one it was tested in. The
+ * output here is append-only text, so a session transcript is a file you can
+ * keep. That is a deliberate trade of polish for portability.
+ */
+
+import { createInterface } from 'node:readline';
+import { readBoxedLine, pinRegion } from './input-box.mjs';
+import { createPainter, colourEnabled } from './colour.mjs';
+import { EXIT_INTERRUPTED } from './interrupt.mjs';
+import { parseSlash, runSlashCommand, SLASH_COMMANDS } from './slash.mjs';
+import { estimateMessagesTokens } from './compact.mjs';
+
+/** What ends a session. `exit`/`quit` because both are muscle memory. */
+const QUIT = new Set(['exit', 'quit', ':q', 'bye']);
+
+/**
+ * ⚠️ CONTEXT GROWS UNBOUNDED AND A CODING SESSION IS THE WORST CASE — tool
+ * results carry whole files. Left alone, turn 30 sends everything from turns
+ * 1-29 and eventually 400s on a context-length error mid-thought.
+ *
+ * The trim keeps the SYSTEM message and the FIRST user message (the workspace
+ * context — the cacheable prefix, dropping it would cost more than it saves) and
+ * discards the oldest middle turns.
+ */
+export const MAX_HISTORY_MESSAGES = 40;
+
+/**
+ * ── 💰⭐⭐⭐ THE TRIM WAS A CONTINUOUS SLIDE, AND IT COST ~90% OF THE CACHE ──
+ *
+ * `messages.slice(-keep)` re-slices on EVERY turn once a session passes the
+ * cap, so the third message of the prompt is different every time. Prefix
+ * caching matches from the first token and stops at the first difference — so
+ * everything after the 2-message head was re-bought at full price, every turn,
+ * for the rest of the session.
+ *
+ * ⭐ MEASURED (no credits, no network — a simulated 60-turn session counting how
+ * much of each prompt is byte-identical to the previous one):
+ *
+ *     CURRENT slice(-38)        cache  9.2%    history 38-40
+ *     stepped high=48 low=40    cache 74.5%    history 38-48
+ *     stepped high=56 low=40    cache 85.3%    history 38-56
+ *     stepped high=64 low=40    cache 89.7%    history 38-64   <- shipped
+ *     stepped high=72 low=40    cache 92.0%    history 38-72
+ *
+ * ⚠️ THE HEAD BEING STABLE IS NOT THE SAME AS THE PROMPT BEING CACHED, and
+ * measuring position 0 alone would have said this code was fine. The 2-message
+ * head never moved; the cacheable prefix was still 2 messages out of 40.
+ *
+ * ⭐ `LOW` IS TODAY'S CAP ON PURPOSE, so this can only ever keep MORE history
+ * than the rule it replaces — never less. That makes it a pure win rather than
+ * a trade of memory for money, and it is why HIGH was raised instead of LOW
+ * being lowered (console shipped 16/8, halving its window; the CLI holds whole
+ * files in tool results and cannot afford to forget more).
+ *
+ * ⚠️ RAISING `HIGH` COSTS TOKENS PER REQUEST, and the ceiling that matters is
+ * not here: `turn.mjs` compacts above CONTEXT_BUDGET_TOKENS (96k), which
+ * `compact.mjs` warns is the moment the cache discount dies for good. This is a
+ * MESSAGE-COUNT backstop, not the token bound — 64 messages of ordinary turns
+ * sit far under 96k, but a session whose tool results carry whole files can
+ * approach it, and then compaction (with its own hysteresis) takes over. 72 and
+ * 80 buy 2 more points of cache for materially more of that risk, which is why
+ * 64 is the shipped number and not the best one in the table.
+ */
+export const HISTORY_HIGH_WATER = 64;
+export const HISTORY_LOW_WATER = MAX_HISTORY_MESSAGES;
+
+/**
+ * ⚠️ WELL UNDER `turn.mjs`'s CONTEXT_BUDGET_TOKENS (96,000), not near it. The
+ * request also carries the TOOL OFFER — measured at ~15,200 tokens for the
+ * CLI's 63 verbs — plus the system prompt, so history must leave room for both
+ * and still clear the line where compaction fires. 55k + ~15k offer + prompt
+ * sits comfortably inside 96k.
+ *
+ * ⭐ This is a CEILING, not a target: an ordinary session (~551 tokens/message,
+ * measured) reaches the 64-message cap at ~35k and never comes near it. It
+ * exists for the file-heavy session, which is exactly the one that would
+ * otherwise be pushed into permanent compaction by the higher message cap.
+ */
+export const HISTORY_TOKEN_CEILING = 55_000;
+
+export function trimHistory(messages, max = HISTORY_HIGH_WATER) {
+  if (!Array.isArray(messages)) return messages;
+  /**
+   * ⚠️⚠️ THE GATE MUST ASK BOTH QUESTIONS, AND MY FIRST VERSION ASKED ONLY ONE.
+   * It read `messages.length <= max` and returned early — so the token ceiling
+   * below was unreachable for exactly the session it was written for: 64 heavy
+   * messages are under the COUNT cap and ~137,000 tokens, and the function
+   * handed them straight back. The test reported 133,491 tokens and I had to
+   * measure to find that the ceiling was never consulted at all rather than
+   * being wrong.
+   */
+  const overBudget = estimateMessagesTokens(messages) > HISTORY_TOKEN_CEILING;
+  if (messages.length <= max && !overBudget) return messages;
+  const head = messages.slice(0, 2);          // system + the context-bearing user turn
+  /**
+   * ⭐ THE HEAD MOVES IN STEPS, NOT EVERY TURN. `dropped` is a multiple of STEP,
+   * so it changes only when the conversation crosses the next boundary — and
+   * between boundaries the prompt is byte-identical to the previous turn's.
+   *
+   * ⚠️ NOT `length > HIGH ? slice(-LOW)`. That is still a continuous slide past
+   * the mark and is the exact mistake console's own guard records catching
+   * before it shipped.
+   */
+  /**
+   * ⚠️⚠️ THE LOW WATER MUST BE DERIVED FROM `max`, NOT PINNED TO A CONSTANT,
+   * AND I ALMOST SHIPPED IT PINNED. My first version was
+   * `Math.min(HISTORY_LOW_WATER, max)`. `runChat`'s default `maxHistory` was
+   * MAX_HISTORY_MESSAGES (40), which equals HISTORY_LOW_WATER — so `step`
+   * became `max(1, 0)` = 1, and one-message steps ARE the continuous slide this
+   * whole change exists to remove. The fix would have measured 89.6% in a
+   * simulation and done NOTHING on the only path that calls it.
+   *
+   * ⭐ 0.625 is 40/64 — the shipped ratio, so the default keeps exactly the
+   * numbers that were measured, and any other `max` still steps properly
+   * instead of silently degrading to a slide.
+   */
+  const low = Math.min(HISTORY_LOW_WATER, Math.floor(max * 0.625));
+  const step = Math.max(1, max - low);
+  const body = messages.slice(head.length);
+  const overflow = Math.max(0, messages.length - max);
+  let dropped = Math.ceil(overflow / step) * step;
+  /**
+   * ── ⚠️⚠️ AND A TOKEN CEILING, BECAUSE MESSAGE COUNT IS THE WRONG UNIT ──────
+   *
+   * MEASURED against the one real recorded session on disk (93,036 prompt
+   * tokens, 18 messages, `.acuvo/sessions/20260814-095636-99d4.json`):
+   *
+   *     avg message  ~551 tokens        largest observed  ~2,145 tokens
+   *
+   *     at 40 msgs: typical ~22,000   worst case (all like the largest) ~85,800
+   *     at 64 msgs: typical ~35,300   worst case                       ~137,300
+   *
+   * ⚠️ SO RAISING THE COUNT ALONE MOVES THE WORST CASE FROM JUST UNDER THE
+   * 96k CONTEXT BUDGET TO WELL OVER IT. Past that line `turn.mjs` compacts, and
+   * `compact.mjs` is explicit that compaction voids the cache discount and
+   * "once it starts, it never stops" — so a change made ENTIRELY to protect the
+   * cache would, in a file-heavy session, destroy it. A tool result carrying a
+   * whole file is not the exception in a coding session; it is the normal case.
+   *
+   * ⭐ THE CEILING DROPS IN THE SAME `step` BLOCKS. Any multiple of `step`
+   * keeps block alignment, so the prefix still changes only at boundaries and
+   * the caching win is untouched — this bounds the worst case without
+   * reintroducing a slide.
+   */
+  const budget = Math.max(1, HISTORY_TOKEN_CEILING - estimateMessagesTokens(head));
+  while (dropped < body.length && estimateMessagesTokens(body.slice(dropped)) > budget) {
+    dropped += step;
+  }
+  let tail = body.slice(dropped);
+  /**
+   * ⚠️ A `tool` MESSAGE WITHOUT ITS `assistant` TOOL CALL IS A HARD 400 from
+   * every OpenAI-shaped provider — "tool_call_id did not have a preceding
+   * message with tool_calls". Slicing mid-exchange produces exactly that, and it
+   * would surface as a mysterious API error thirty turns into a session.
+   */
+  while (tail.length > 0 && tail[0].role === 'tool') tail = tail.slice(1);
+  return [...head, ...tail];
+}
+
+/**
+ * One prompt line. Returns null on EOF (Ctrl-D, or a pipe that ran out).
+ *
+ * ⚠️ THE CLOSED CHECK IS NOT DEFENSIVE, IT IS THE PIPED CASE. Found by piping a
+ * list of prompts in: readline emits 'close' when the stream ends, and the NEXT
+ * `rl.question()` throws ERR_USE_AFTER_CLOSE. The first turn had already
+ * succeeded and written a real file, so the session crashed AFTER doing its job
+ * — the worst shape of failure, because the work looks lost.
+ *
+ * Scripted input matters beyond tests: piping a prompt list is how anyone would
+ * automate this.
+ */
+function ask(rl, prompt, state) {
+  if (state.closed) return Promise.resolve(null);
+  return new Promise((resolve) => {
+    let answered = false;
+    const onClose = () => { if (!answered) resolve(null); };
+    rl.once('close', onClose);
+    rl.question(prompt, (line) => {
+      answered = true;
+      rl.removeListener('close', onClose);
+      resolve(line);
+    });
+  });
+}
+
+/**
+ * Run an interactive session.
+ *
+ * `runOne(task, priorMessages)` performs one turn and returns the session
+ * outcome — injected rather than imported so this loop is testable with a stub
+ * and never needs a model or a terminal in a test.
+ */
+/**
+ * ── ⚠️ PIPED INPUT IS A DIFFERENT PROBLEM AND NEEDED A DIFFERENT ANSWER ──────
+ * `readline` on a non-TTY DRAINS the stream as fast as it can and emits 'close'
+ * the moment it ends. The model call for turn 1 takes seconds, by which point
+ * the interface is already closed and every later prompt is lost — measured:
+ * a three-line pipe ran exactly ONE turn and exited quietly, which is worse than
+ * crashing because it looks like it worked.
+ *
+ * So a pipe is read WHOLE and replayed from a queue. A TTY keeps the real
+ * readline loop, where a human types the next line after seeing the last answer.
+ * Two input shapes, two mechanisms — pretending they are the same is what broke.
+ */
+async function readAllLines(input) {
+  const chunks = [];
+  for await (const chunk of input) chunks.push(chunk);
+  return Buffer.concat(chunks.map((c) => (typeof c === 'string' ? Buffer.from(c) : c)))
+    .toString('utf8')
+    .split(String.fromCharCode(10))
+    .map((l) => l.replace(String.fromCharCode(13), ''));
+}
+
+/**
+ * ── ⚠️⚠️⭐ READLINE EATS CTRL-C. MEASURED IN NODE'S OWN SOURCE ──────────────
+ *
+ * This is the finding that made an interrupt handler in `bin/acuvo.mjs`
+ * necessary-but-not-sufficient. Read out of `process.binding('natives')` on
+ * node v22.17.0, `internal/readline/interface.js`, the ttyWrite ctrl-key
+ * switch, verbatim:
+ *
+ *     case 'c':
+ *       if (this.listenerCount('SIGINT') > 0) {
+ *         this.emit('SIGINT');
+ *       } else {
+ *         // This readline instance is finished
+ *         this.close();
+ *         this[kQuestionReject]?.(new AbortError('Aborted with Ctrl+C'));
+ *       }
+ *
+ * ⚠️ So with a TTY readline open and NO `'SIGINT'` listener on the interface,
+ * Ctrl-C never reaches `process.on('SIGINT')` at all — readline just closes
+ * itself. The run in flight would have carried on to completion, for minutes,
+ * with the user's Ctrl-C having produced nothing on screen. That is strictly
+ * worse than the bug we set out to fix, and no amount of correct handling in
+ * `bin/acuvo.mjs` would have been reached.
+ *
+ * ⭐ So the interface takes a listener whose whole job is to hand the signal
+ * back to the process, where the one policy in `interrupt.mjs` decides between
+ * "stop after this round" and "quit now".
+ *
+ * ── ⚠️⚠️ AND A SYNTHETIC EMIT INTO AN EMPTY EMITTER DOES NOTHING ────────────
+ *
+ * `process.emit('SIGINT')` is plain `EventEmitter.emit` — it does NOT invoke
+ * the OS default action. `turn.mjs` installs the process signal handlers inside
+ * `runSession`, so before the first turn of a session there are ZERO listeners
+ * and the emit would return `false` having done absolutely nothing. Ctrl-C at
+ * the very first prompt would be inert.
+ *
+ * ⭐ Hence the count check and the explicit exit: **every path out of this
+ * function either aborts a run or ends the process.** That is the rule this
+ * whole feature is built on, and it is the one that is easy to break here.
+ */
+export function deliverInterrupt({
+  emit = (sig) => process.emit(sig),
+  listenerCount = (sig) => process.listenerCount(sig),
+  exit = (code) => process.exit(code),
+} = {}) {
+  if (listenerCount('SIGINT') > 0) {
+    emit('SIGINT');
+    return 'delegated';
+  }
+  exit(EXIT_INTERRUPTED);
+  return 'exited';
+}
+
+export async function runChat({
+  runOne,
+  render,
+  input = process.stdin,
+  output = process.stdout,
+  banner = '',
+  /**
+   * ⚠️ THE HIGH WATER MARK, NOT THE OLD FLAT CAP. This default is what makes
+   * the stepped trim REACH the live session — see the note in `trimHistory`
+   * about the version of this fix that measured 89.6% and changed nothing.
+   */
+  maxHistory = HISTORY_HIGH_WATER,
+  /**
+   * ⚠️ INJECTED SO A TEST CAN SEE IT. The real one exits the process, and a
+   * test that could not substitute it could only assert this feature by killing
+   * its own runner. Default is production behaviour, so no caller changes.
+   */
+  onInterrupt = deliverInterrupt,
+  /**
+   * ── ⭐ THE `/` SURFACE'S ONE SEAM ─────────────────────────────────────────
+   *
+   * Providers for the things a command reports on — skills, MCP servers, spend,
+   * the model. `bin/` owns where those facts come from; this loop only asks, for
+   * the same reason `workspace.mjs` takes `claimPath` and `journal` injected
+   * rather than importing them.
+   *
+   * ⚠️ `{}` BY DEFAULT, NOT `null`. Every command degrades to "not available in
+   * this session" on a missing provider (see `slash.mjs`), so an embedder that
+   * wires nothing still gets a working `/help` instead of a crash.
+   */
+  slashContext = {},
+}) {
+  const interactive = input.isTTY === true;
+
+  // ⚠️ A pipe is drained up front — see readAllLines. Doing this lazily is what
+  // silently lost every prompt after the first.
+  const queued = interactive ? null : await readAllLines(input);
+  let queueIndex = 0;
+
+  /**
+   * ── ⚠️⚠️ READLINE IS GONE FROM THE INTERACTIVE PATH, AND IT HAD TO GO ──────
+   *
+   * `input-box.mjs` now owns the keyboard. Leaving the readline interface
+   * attached to the SAME stream was not merely redundant — both it and the box
+   * saw every Ctrl-C, so `onInterrupt` fired TWICE for one keypress. Measured,
+   * not theorised: a single `\x03` produced `['interrupt', 'interrupt']`.
+   *
+   * ⭐ A DOUBLE INTERRUPT IS NOT A COSMETIC BUG. The second Ctrl-C is the one
+   * that QUITS — so one press would have armed and fired the escape hatch in the
+   * same instant, ending a session the user meant only to nudge.
+   *
+   * The Node-internals note below about `question()` swallowing Ctrl-C is kept
+   * because it is why the box reads raw keys itself rather than asking readline
+   * for a line. It is history now, not a live constraint.
+   */
+  const rl = null;
+  const state = { closed: false };
+  /**
+   * ⚠️ ATTACHED ONCE, NOT PER QUESTION — and the per-question version is why the
+   * first fix did not work. The stream can end WHILE the model call is in
+   * flight, when no question is pending and therefore no listener is attached;
+   * `close` fires into nothing, the flag stays false, and the next question
+   * throws anyway. A session-lifetime listener sees it whenever it happens.
+   */
+  if (rl) rl.once('close', () => { state.closed = true; });
+  /**
+   * ⭐ THE ONE LINE THAT MAKES CTRL-C REACH THE RUN — see `deliverInterrupt`
+   * above for the Node source that proves it is needed. `on`, not `once`: an
+   * interactive session runs many turns and the SECOND Ctrl-C (the one that
+   * quits) has to arrive here too, or the escape hatch is a single-use one.
+   */
+  if (rl) rl.on('SIGINT', () => { onInterrupt(); });
+  /**
+   * ⚠️⚠️ THE BANNER IS DEFERRED UNTIL AFTER THE PIN, AND THE ORDER IS NOT
+   * COSMETIC. `pinRegion` CLEARS the screen so the transcript starts at the top
+   * of a clean one — printed before that, the banner is erased by the very
+   * thing meant to sit beneath it, which is exactly what a screenshot showed:
+   * a banner cut in half with the conversation crammed at the bottom.
+   */
+  const writeBanner = () => { if (banner) output.write(`${banner}\n`); };
+  // ⚠️ `/help` IS ADVERTISED IN THE ONE LINE EVERY SESSION PRINTS. A command
+  // surface nobody is told about is the same defect item 14 closed for `--help`:
+  // the feature worked and nothing a stranger would read mentioned it.
+  /**
+   * ⚠️ A BLANK LINE BEFORE IT, not just after. Roman: *"move it one line down."*
+   * Pressed against the banner it read as a fifth detail row rather than as an
+   * instruction addressed to the person.
+   */
+  output.write('\nType what you want done. "/help" for commands, "exit" to leave.\n\n');
+
+  let history = null;
+  let turns = 0;
+  /**
+   * ⭐ SET BY `/skills <name>`, CONSUMED BY THE NEXT REAL TURN AND THEN CLEARED.
+   * A skill that was printed to the terminal would look loaded and be invisible
+   * to the model; this is the variable that makes the verb real.
+   */
+  let pendingInject = null;
+  /** Lines the user has submitted this session — the box's Up/Down history. */
+  const typed = [];
+
+  /**
+   * ── ⭐⭐⭐ THE BOX IS PINNED TO THE BOTTOM OF THE SCREEN ────────────────────
+   *
+   * Roman: *"we need that prompt box stuck down the bottom, it is professional."*
+   *
+   * `pinRegion` reserves the last three rows, so output scrolls ABOVE the box
+   * and the box itself never moves. Off a TTY, in CI, on a short terminal, or
+   * with ACUVO_NO_PIN=1 it does nothing at all.
+   *
+   * ⚠️⚠️ RELEASED ON EVERY PATH OUT, INCLUDING THE ONES NOBODY PLANS FOR. A
+   * process that exits with a scroll region still set leaves the user with a
+   * terminal that scrolls inside a box until they type `reset` blind — the same
+   * class of harm as leaving raw mode on. The `finally` covers a normal end and
+   * a throw; the signal handlers cover Ctrl-C and `kill`.
+   */
+  const pin = interactive ? pinRegion(output) : { enabled: false, release() {} };
+  const releasePin = () => pin.release();
+  if (pin.enabled) {
+    process.once('exit', releasePin);
+    process.once('SIGINT', releasePin);
+    process.once('SIGTERM', releasePin);
+  }
+
+  /**
+   * ⭐ NOW the banner, at the top of a screen the pin has just cleared, with the
+   * conversation free to grow downward beneath it — which is the direction every
+   * scrollback in existence runs.
+   */
+  writeBanner();
+
+  try {
+    for (;;) {
+      /**
+       * ── ⭐⭐ THE INPUT IS A BOX, BECAUSE A BARE `› ` IS NOT A PLACE TO TYPE ──
+       *
+       * Roman, comparing against a real Claude Code screenshot: *"you can see
+       * the box where you type, acuvo doesn't have that."* He is right — the
+       * prompt was two characters floating in the scrollback, which reads as
+       * output rather than as somewhere input goes.
+       *
+       * ⚠️ THE BOX OPENS BEFORE THE LINE AND CLOSES AFTER IT, deliberately.
+       * Drawing all four sides up front needs absolute cursor control, and
+       * readline owns the cursor once `question()` starts — every version of
+       * that fights the line editor the moment input wraps, a history entry is
+       * recalled, or the window is resized. Opening on entry and closing on
+       * submit needs no cursor math at all, survives all three, and leaves a
+       * transcript where each turn is visibly a closed unit.
+       *
+       * ⚠️ WIDTH IS READ FRESH EACH TURN — a terminal resized mid-session would
+       * otherwise draw rules at the old width for the rest of the run. Clamped,
+       * because `columns` is `undefined` when stdout is not a TTY and enormous
+       * when someone maximises on an ultrawide.
+       */
+      let line;
+      if (interactive) {
+        output.write('\n');
+        const got = await readBoxedLine({
+          atRow: pin.enabled ? pin.bottom + 1 : 0,
+          /**
+           * ⚠️ DERIVED FROM `SLASH_COMMANDS`, NEVER RETYPED. A hardcoded list
+           * here would be right today and wrong the first time a command is
+           * added — the menu would quietly omit it, which is the same
+           * built-but-invisible failure the menu exists to fix.
+           */
+          commands: SLASH_COMMANDS.map((c) => c.name),
+          paintFn: createPainter(colourEnabled({ stream: output })).brand,
+          input,
+          output,
+          history: typed,
+          onInterrupt,
+        });
+        line = got.value;
+      } else {
+        line = queueIndex < queued.length ? queued[queueIndex++] : null;
+      }
+      // Echo a piped prompt so a scripted transcript reads like a session.
+      if (!interactive && line !== null && line.trim()) output.write(`› ${line.trim()}
+`);
+      // ⚠️ EOF is not an error. A closed pipe or Ctrl-D ends the session the
+      // same way "exit" does — treating it as a fault would print a stack trace
+      // at the end of every scripted run.
+      if (line === null) break;
+      const task = line.trim();
+      if (!task) continue;
+      // ⚠️ Deduped against the PREVIOUS entry only: pressing Up should walk
+      // distinct instructions, not scroll through five copies of `npm test`.
+      if (typed[typed.length - 1] !== task) typed.push(task);
+      if (QUIT.has(task.toLowerCase())) break;
+
+      /**
+       * ── ⭐ THE `/` SURFACE, BEFORE ANYTHING IS SENT TO A MODEL ────────────
+       *
+       * ⚠️ IT COSTS NOTHING AND MUST NOT COUNT AS A TURN. `/cost` is asked
+       * precisely by somebody watching their spend, and answering it by
+       * incrementing the turn counter and appending to the history would make
+       * the question change the answer.
+       *
+       * ⚠️ AN UNRECOGNISED COMMAND IS ANSWERED HERE AND NOT FORWARDED. Passing
+       * `/skil` to the model gets a confident essay about a typo; the one thing
+       * the person needed was the word `/skills`, which `slash.mjs` supplies.
+       * See its header for why `/etc/hosts` is NOT treated as a command.
+       */
+      const command = parseSlash(line.trim());
+      if (command) {
+        const result = runSlashCommand(command, slashContext);
+        for (const l of result.output ?? []) output.write(`${l}\n`);
+        output.write('\n');
+        if (result.effect === 'clear') history = null;
+        if (typeof result.inject === 'string' && result.inject) pendingInject = result.inject;
+        continue;
+      }
+
+      /**
+       * ⚠️ THE SKILL IS PREPENDED TO THE TASK, NOT SUBSTITUTED FOR IT. The user
+       * typed an instruction; the skill is context for it. And it is cleared
+       * BEFORE the call rather than after, so a turn that throws cannot leave it
+       * armed and silently attach it to an unrelated question later.
+       */
+      let sendTask = task;
+      if (pendingInject) {
+        sendTask = `${pendingInject}\n\n---\n\n${task}`;
+        pendingInject = null;
+      }
+
+      let outcome;
+      try {
+        outcome = await runOne(sendTask, history);
+      } catch (err) {
+        /**
+         * ⚠️ ONE BAD TURN MUST NOT END THE SESSION. A timeout or a provider blip
+         * after twenty minutes of context is infuriating if it drops everything;
+         * the history is still valid, so report and keep the prompt.
+         */
+        output.write(`\n  ✖ that turn failed: ${String(err?.message || err)}\n\n`);
+        continue;
+      }
+
+      turns += 1;
+      render(outcome, output);
+
+      if (outcome?.ok && Array.isArray(outcome.messages)) {
+        history = trimHistory(outcome.messages, maxHistory);
+      } else if (!outcome?.ok) {
+        // A failed turn leaves history UNTOUCHED. Appending a turn that produced
+        // nothing would poison the next one with a dead exchange.
+        output.write(`\n  (history unchanged — that turn did not complete)\n`);
+      }
+      output.write('\n');
+    }
+  } finally {
+    if (rl) rl.close();
+    pin.release();
+    // ⚠️ The listeners come off too — a long-lived process that started several
+    // sessions would otherwise accumulate them and warn about a leak.
+    process.off('exit', releasePin);
+    process.off('SIGINT', releasePin);
+    process.off('SIGTERM', releasePin);
+  }
+  return { turns };
+}