Codeman/src/web/response-viewer-transcript.ts

export type ResponseViewerTranscriptKind = 'prompt' | 'response' | 'status' | 'tool';

export interface ResponseViewerTranscriptBlock {
  kind: ResponseViewerTranscriptKind;
  label: 'Prompt' | 'Response' | 'Status' | 'Tool';
  /** What the frontend renders by: 'user' gets the "You" badge, everything else the agent badge. */
  role: 'user' | 'assistant';
  text: string;
}

// Keep in lockstep with isExternalCliMode() in src/session.ts. Importing it here
// would drag node-pty and the whole session layer into this pure module, so the
// list is duplicated and test/response-viewer-transcript.test.ts pins the parity.
const EXTERNAL_CLI_MODES = new Set(['codex', 'gemini', 'opencode', 'antigravity', 'pi', 'grok', 'deepseek', 'copilot']);

/**
 * A pane that is not Codex-shaped (`›` prompt, `•` bullets) describes its own dialect as data, so
 * a new CLI adds a row here instead of another branch in the Codex-oriented loop below.
 */
interface PaneDialect {
  /** Group 1 is the prompt text. A bare composer glyph with nothing after it must NOT match. */
  prompt: RegExp;
  /** A right-aligned timestamp the TUI appends to a submitted prompt. */
  promptTrailer?: RegExp;
  /** The glyph that opens an answer. */
  responseBullet: RegExp;
  /** Chrome that can sit between turns: banner notes, the cwd/credits row, the footer, spinners. */
  status: RegExp[];
  /** Rows that carry nothing (the empty composer), dropped and treated as a block boundary. */
  ignore: RegExp[];
  /**
   * The composer row, the one between two rules at the bottom of the screen. Only the LAST rule,
   * composer row, rule triple counts: an answer can hold a rule-delimited paragraph or a quoted
   * prompt of its own, and that is answer text.
   */
  composer: RegExp;
  /** Furniture drawn at the end of a row (a scrollbar gutter), removed before anything else reads it. */
  lineTrailer?: RegExp;
  /**
   * Text that is on screen only before the first prompt has scrolled the screen: the banner, a
   * trust dialog. Its presence is how "nothing has been asked yet" is told apart from "the prompt
   * row has scrolled off the top".
   */
  startup: RegExp[];
}

const PANE_DIALECTS: Record<string, PaneDialect> = {
  // GitHub Copilot CLI, measured on 1.0.95 inside tmux: ` ❯ <prompt>   HH:MM`, then ` ● <answer>`
  // with 3-space continuation lines, a cwd + `Session: N AIC used` row, the composer between two
  // rules, and a `← open sidebar · … · / commands · ? help` footer. Tool-call rows are not
  // modelled: they open with the same bullet and read as part of the answer.
  copilot: {
    // Copilot draws a submitted prompt at column 0 or 1; the continuation rows of an answer are
    // indented three spaces, so a shell prompt inside a code block (`   ❯ npm test`) is not a prompt.
    prompt: /^ ?❯\s+(\S.*)$/,
    promptTrailer: /\s{2,}\d{1,2}:\d{2}\s*$/,
    responseBullet: /^\s*●\s+/,
    status: [
      /^\s*(?:!|●)\s+(?:Tip:|MCP Servers?\b|Model\b)/,
      /^\s*!\s+/,
      /\bSession: .*(?:AIC used|credits unavailable)/,
      /^\s*←\s+open sidebar/,
      /\/ commands · \? help/,
      /^\s*[◉◎]\s+Working\b/,
    ],
    ignore: [/^ ?❯\s*$/],
    composer: /^ ?❯/,
    // The right-hand scrollbar gutter, `┃` on a row (or alone on a blank one).
    lineTrailer: /\s*[┃▐]\s*$/,
    startup: [/Copilot v\d+\.\d+\.\d+ uses AI/, /Confirm folder trust/, /Do you trust the files in this folder/],
  },
};

function isPromptLine(line: string): boolean {
  return /^\s*›\s*/.test(line);
}

function isDividerDashChar(ch: string): boolean {
  return ch === '─' || ch === '-';
}

function isWhitespaceChar(ch: string): boolean {
  return /\s/.test(ch);
}

// Linear-time equivalent of the old /^[─-]+\s*(.+?)\s*[─-]{3,}$/. The lazy
// middle of that pattern backtracked catastrophically on a long dash run that
// does NOT end in 3+ dashes (measured >2min at 8,000 chars) — and pane text is
// agent-controlled with buffers up to 32MB, so this ran on hostile input. Same
// accept set and same captured content, computed with counters; equivalence is
// pinned char-for-char against the old regex by the brute-force corpus test in
// test/response-viewer-transcript.test.ts.
function normalizeDividerStatusLine(line: string): string | null {
  const s = line.trim();
  const n = s.length;
  // Minimum match: 1 leading dash + 1 content char + 3 trailing dashes.
  if (n < 5) return null;

  let lead = 0;
  while (lead < n && isDividerDashChar(s.charAt(lead))) lead += 1;
  if (lead === 0) return null;

  let trail = 0;
  while (trail < n && isDividerDashChar(s.charAt(n - 1 - trail))) trail += 1;
  if (trail < 3) return null;

  // The regex was greedy on the leading run but gave dashes back until at least
  // one content char plus the 3-dash tail fit (an all-dash line matched with a
  // single leftover dash as its "content"), so the content window starts at the
  // end of the leading run, clamped to leave 4 chars.
  const contentStart = Math.min(lead, n - 4);
  let ws = 0;
  while (contentStart + ws < n - 4 && isWhitespaceChar(s.charAt(contentStart + ws))) ws += 1;
  const from = contentStart + ws;

  // The lazy middle stopped at the first position from which "optional
  // whitespace, then dashes to end-of-line" matches: the start of the
  // whitespace padding in front of the trailing dash run (never before the
  // first content char).
  const tailStart = n - trail;
  let padded = tailStart;
  while (padded > 0 && isWhitespaceChar(s.charAt(padded - 1))) padded -= 1;
  const end = Math.max(from + 1, padded);

  return s.slice(from, end).trim() || null;
}

function isDividerOnlyLine(line: string): boolean {
  return /^[\s─-]{8,}$/.test(line.trim());
}

// COD-227: unambiguous Codex tool-call markers. These only ever appear as internal
// activity, never as ordinary assistant prose, so they are always a Tool block.
function isToolActivityMarker(line: string): boolean {
  return /^\s*[•*-]\s+(Calling|Called)\b/.test(line.trim());
}

// COD-227: action verbs that ALSO occur in ordinary assistant prose (e.g.
// "• Created COD-226: …"). These are a Tool header only when corroborated by a
// box-drawing result tree on the next non-blank line (see the caller); the verb
// alone is not sufficient.
function isToolVerbBullet(line: string): boolean {
  return /^\s*[•*-]\s+(Explored|Viewed|Read|Edited|Updated|Created|Deleted|Ran|Searched|Opened|Listed|Found|Applied|Patched|Used|Wrote|Executed|Modified|Analyzed|Compared|Fetched|Installed)\b/.test(
    line.trim()
  );
}

// A genuine Codex tool block renders its result as a box-drawing tree (└ │ ├).
function isBoxDrawingLine(line: string): boolean {
  return /^[│├└]/.test(line.trim());
}

// Look past blank lines from `fromIndex + 1` for the next non-blank line and report
// whether it is a box-drawing tool-result line — the signal that a verb bullet is a
// real tool block rather than assistant prose that happens to start with a verb.
function nextNonBlankIsBoxDrawing(lines: string[], fromIndex: number): boolean {
  for (let j = fromIndex + 1; j < lines.length; j += 1) {
    const trimmed = (lines[j] || '').trim();
    if (!trimmed) continue;
    return isBoxDrawingLine(trimmed);
  }
  return false;
}

function isToolContinuationLine(line: string, currentKind: ResponseViewerTranscriptKind | null): boolean {
  if (currentKind !== 'tool') return false;
  const trimmed = line.trimEnd();
  if (!trimmed) return true;
  return /^\s*[│├└]/.test(trimmed) || /^\s{2,}\S/.test(line);
}

function isStatusLine(line: string, mode: string): boolean {
  if (!EXTERNAL_CLI_MODES.has(mode)) return false;
  const trimmed = line.trim();
  if (!trimmed) return false;
  if (normalizeDividerStatusLine(trimmed)) return true;
  if (/^(model|directory):\s+/i.test(trimmed)) return true;
  if (/\bContext\b.*\bleft\b/i.test(trimmed)) return true;
  if (/\b\/model to change\b/i.test(trimmed)) return true;
  if (/\bReady\b/i.test(trimmed) && /·/.test(trimmed)) return true;
  if (/^(gpt|o\d|claude|gemini)\b/i.test(trimmed) && /·/.test(trimmed)) return true;
  if (/^Tip:/i.test(trimmed)) return true;
  if (/^\s*[•*-]\s+(Working|Thinking|Loading|Starting\b|Waiting\b)/i.test(trimmed)) return true;
  return false;
}

function isStandaloneMarkdownLine(line: string): boolean {
  const trimmed = line.trim();
  if (!trimmed) return false;
  if (/^#{1,6}\s/.test(trimmed)) return true;
  if (/^>\s/.test(trimmed)) return true;
  if (/^(```|~~~)/.test(trimmed)) return true;
  if (/^[-*+]\s/.test(trimmed)) return true;
  if (/^\d+[.)]\s/.test(trimmed)) return true;
  if (/^\|/.test(trimmed)) return true;
  if (/^\s{4,}\S/.test(line)) return true;
  return false;
}

function shouldJoinWrappedLine(previous: string, next: string): boolean {
  const prev = previous.trimEnd();
  const curr = next.trim();
  if (!prev || !curr) return false;
  if (/[.!?]$/.test(prev)) return false;
  if (/[:;]$/.test(prev)) return false;
  if (isStandaloneMarkdownLine(curr)) return false;
  if (/^[a-z(]/.test(curr)) return true;
  if (prev.length >= 72 && /^[A-Za-z0-9"'(]/.test(curr)) return true;
  return false;
}

function normalizeWrappedText(lines: string[]): string {
  const out: string[] = [];
  let paragraph = '';

  const flushParagraph = () => {
    if (!paragraph) return;
    out.push(paragraph);
    paragraph = '';
  };

  for (const rawLine of lines) {
    const line = rawLine.trimEnd();
    const trimmed = line.trim();
    if (!trimmed) {
      flushParagraph();
      if (out[out.length - 1] !== '') out.push('');
      continue;
    }
    if (isStandaloneMarkdownLine(line)) {
      flushParagraph();
      out.push(trimmed);
      continue;
    }
    if (!paragraph) {
      paragraph = trimmed;
      continue;
    }
    if (shouldJoinWrappedLine(paragraph, trimmed)) {
      paragraph += ` ${trimmed}`;
      continue;
    }
    flushParagraph();
    paragraph = trimmed;
  }

  flushParagraph();
  return out
    .join('\n')
    .replace(/\n{3,}/g, '\n\n')
    .trim();
}

function cleanTerminalTranscript(buffer: string): string {
  // Stripping ANSI/OSC/DCS escape sequences and stray C0/C1 control bytes
  // legitimately requires control characters in these patterns.
  /* eslint-disable no-control-regex */
  return String(buffer || '')
    .replace(/\x1b\[[\x30-\x3F]*[\x20-\x2F]*[\x40-\x7E]/g, '')
    .replace(/\x1b\][^\x07\x1b]*(?:\x07|\x1b\\)/g, '')
    .replace(/\x1b[PX^_][^\x1b]*\x1b\\/g, '')
    .replace(/\x1b[NO()][A-Z0-9]?/g, '')
    .replace(/\x1b[>=<78cDEHM]/g, '')
    .replace(/[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]/g, '')
    .replace(/\r\n/g, '\n')
    .replace(/\r/g, '\n')
    .replace(/[ \t]+$/gm, '')
    .trim();
  /* eslint-enable no-control-regex */
}

function trimLeadingStartup(lines: string[]): string[] {
  let index = 0;
  while (index < lines.length) {
    const line = lines[index] || '';
    const trimmed = line.trim();
    if (!trimmed) {
      index += 1;
      continue;
    }
    if (/^[╭╰│─].*[╮╯│]?$/.test(trimmed)) {
      index += 1;
      continue;
    }
    if (/^>_\s*OpenAI Codex/i.test(trimmed)) {
      index += 1;
      continue;
    }
    if (/^(model|directory):\s+/i.test(trimmed)) {
      index += 1;
      continue;
    }
    break;
  }
  return lines.slice(index);
}

function pushBlock(
  blocks: ResponseViewerTranscriptBlock[],
  kind: ResponseViewerTranscriptKind | null,
  lines: string[]
): void {
  if (!kind || lines.length === 0) return;
  const normalizedLines =
    kind === 'status'
      ? lines.map((line) => normalizeDividerStatusLine(line) || line.trim()).filter((line) => line.length > 0)
      : lines;
  const text =
    kind === 'tool' || kind === 'status'
      ? normalizedLines
          .join('\n')
          .replace(/\n{3,}/g, '\n\n')
          .trim()
      : normalizeWrappedText(normalizedLines);
  if (!text) return;
  const label = (kind.charAt(0).toUpperCase() + kind.slice(1)) as ResponseViewerTranscriptBlock['label'];
  // The frontend's loadFullContext() renders via msg.role — a block without it
  // lost the "You" badge on prompts and rendered every block as the agent.
  blocks.push({ kind, label, role: kind === 'prompt' ? 'user' : 'assistant', text });
}

export function isExternalCliTranscriptMode(mode: string | null | undefined): boolean {
  return EXTERNAL_CLI_MODES.has(String(mode || ''));
}

/**
 * Whether this mode's pane is a repainting full-screen TUI whose output stream is absolute cursor
 * moves rather than lines. The segmenter needs the SCREEN for those (tmux capture-pane), because
 * stripping the escapes from the stream concatenates every row into one run of text.
 */
export function transcriptNeedsScreenCapture(mode: string | null | undefined): boolean {
  return Object.prototype.hasOwnProperty.call(PANE_DIALECTS, String(mode || ''));
}

/**
 * Segment a pane that has its own dialect. Everything before the first prompt is the banner (or a
 * trust dialog) and is dropped, so a session that has not been asked anything answers "nothing
 * yet" rather than serving its logo as a reply.
 *
 * The screen is all there is to read, so a long answer scrolls the prompt row off the top. When no
 * prompt row is on screen AND no startup text is either, the session has already been asked
 * something: the rows above the composer are the tail of the last answer, not "nothing yet".
 */
function parseDialectTranscript(cleaned: string, dialect: PaneDialect): ResponseViewerTranscriptBlock[] {
  const blocks: ResponseViewerTranscriptBlock[] = [];
  let kind: ResponseViewerTranscriptKind | null = null;
  let current: string[] = [];
  const rows = cleaned.split('\n').map((row) => (dialect.lineTrailer ? row.replace(dialect.lineTrailer, '') : row));
  // The composer sits between two rules and holds whatever is typed but not sent. That is not a
  // prompt row: otherwise typing under a tall answer would make the answer's tail read as banner.
  let composerIndex = -1;
  for (let i = rows.length - 2; i > 0; i -= 1) {
    if (
      dialect.composer.test(rows[i] ?? '') &&
      isDividerOnlyLine(rows[i - 1] ?? '') &&
      isDividerOnlyLine(rows[i + 1] ?? '')
    ) {
      composerIndex = i;
      break;
    }
  }
  const isComposerRow = (i: number) => i === composerIndex;
  const isComposerRule = (i: number) => composerIndex !== -1 && (i === composerIndex - 1 || i === composerIndex + 1);
  const hasPrompt = rows.some((row, i) => dialect.prompt.test(row) && !isComposerRow(i));
  const startupVisible = rows.some((row) => dialect.startup.some((re) => re.test(row)));
  let sawPrompt = !hasPrompt && !startupVisible;
  // A rule drawn INSIDE an answer (a markdown `---`, a quoted passage between two rules) is part
  // of the answer, so it must not end the block, or `agent read` returns only the text after the
  // last rule. It is held until more of the same answer follows, so a rule that only separates
  // the answer from the footer never trails it. The composer's own two rules stay boundaries.
  let pendingRule = false;

  const flush = () => {
    if (sawPrompt) pushBlock(blocks, kind, current);
    kind = null;
    current = [];
    pendingRule = false;
  };

  for (const [index, line] of rows.entries()) {
    if (kind === 'response' && isDividerOnlyLine(line) && !isComposerRule(index)) {
      pendingRule = true;
      continue;
    }
    if (isDividerOnlyLine(line) || isComposerRow(index) || dialect.ignore.some((re) => re.test(line))) {
      flush();
      continue;
    }
    const prompt = dialect.prompt.exec(line);
    if (prompt) {
      flush();
      sawPrompt = true;
      kind = 'prompt';
      current = [dialect.promptTrailer ? prompt[1].replace(dialect.promptTrailer, '').trim() : prompt[1].trim()];
      continue;
    }
    if (!sawPrompt) continue;
    if (dialect.status.some((re) => re.test(line))) {
      if (kind !== 'status') flush();
      kind = 'status';
      current.push(line.trim());
      continue;
    }
    // A wrapped warning or tip: indented under its status row.
    if (kind === 'status' && /^\s{3,}\S/.test(line)) {
      current.push(line.trim());
      continue;
    }
    if (!line.trim()) {
      if (kind) current.push('');
      continue;
    }
    const bullet = dialect.responseBullet.exec(line);
    if (bullet) {
      flush();
      kind = 'response';
      current = [line.slice(bullet[0].length)];
      continue;
    }
    if (!kind) kind = 'response';
    if (pendingRule) {
      current.push('', '---', '');
      pendingRule = false;
    }
    current.push(line);
  }
  flush();
  return blocks.filter((block) => block.text.trim().length > 0);
}

export function parseExternalCliTranscript(
  buffer: string,
  mode: string | null | undefined
): ResponseViewerTranscriptBlock[] {
  const resolvedMode = String(mode || '');
  if (!isExternalCliTranscriptMode(resolvedMode)) return [];

  const cleaned = cleanTerminalTranscript(buffer);
  if (!cleaned) return [];

  const dialect = PANE_DIALECTS[resolvedMode];
  if (dialect) return parseDialectTranscript(cleaned, dialect);

  const lines = trimLeadingStartup(cleaned.split('\n'));
  const blocks: ResponseViewerTranscriptBlock[] = [];
  let currentKind: ResponseViewerTranscriptKind | null = null;
  let currentLines: string[] = [];

  const flush = () => {
    pushBlock(blocks, currentKind, currentLines);
    currentKind = null;
    currentLines = [];
  };

  for (let i = 0; i < lines.length; i += 1) {
    const line = lines[i] ?? '';
    // COD-226: within a prompt, the 2-space Codex gutter is authoritative. Blank
    // lines and gutter-indented (2+ leading spaces) continuation lines stay in the
    // Prompt block ahead of every structural detector below, so multiline prompts —
    // including bullets, indented dividers, and literal › lines — are not
    // misclassified as responses. Only a non-blank, non-gutter line ends the prompt
    // and falls through (a column-0 › then opens a NEW prompt).
    if (currentKind === 'prompt') {
      if (!line.trim()) {
        currentLines.push('');
        continue;
      }
      if (/^ {2,}\S/.test(line)) {
        currentLines.push(line);
        continue;
      }
    }

    if (isDividerOnlyLine(line)) {
      flush();
      continue;
    }

    if (isPromptLine(line)) {
      flush();
      currentKind = 'prompt';
      currentLines = [line.replace(/^\s*›\s*/, '').trim()];
      continue;
    }

    // COD-227: • Calling / • Called are always tool markers; the other action verbs
    // are a tool header only when a box-drawing result tree follows on the next
    // non-blank line — otherwise the verb bullet is ordinary assistant prose.
    if (isToolActivityMarker(line) || (isToolVerbBullet(line) && nextNonBlankIsBoxDrawing(lines, i))) {
      if (currentKind !== 'tool') flush();
      currentKind = 'tool';
      currentLines.push(line.trimEnd());
      continue;
    }

    if (isToolContinuationLine(line, currentKind)) {
      currentLines.push(line.trimEnd());
      continue;
    }

    if (isStatusLine(line, resolvedMode)) {
      if (currentKind !== 'status') flush();
      currentKind = 'status';
      currentLines.push(line.trim());
      continue;
    }

    if (!line.trim()) {
      currentLines.push('');
      continue;
    }

    if (currentKind !== 'response') flush();
    currentKind = 'response';
    currentLines.push(line);
  }

  flush();
  return blocks.filter((block) => block.text.trim().length > 0);
}

export function getLastTranscriptResponse(blocks: ResponseViewerTranscriptBlock[]): string {
  for (let i = blocks.length - 1; i >= 0; i -= 1) {
    if (blocks[i]?.kind === 'response') return blocks[i].text;
  }
  return '';
}

/**
 * The messages of the most recent turn that has an answer: every assistant
 * message whose `turn` matches the highest turn any assistant message carries.
 *
 * This is what the viewer's brief ("Last Response") view renders for Claude.
 * The brief `text` is one row — the last assistant row — and a Claude turn is
 * a median of 3 rows (p90 11), so that row alone was usually the tail of the
 * answer ("Done.") with the substance in the rows before it. Reading the whole
 * turn gives the same cards the full view shows for it, and no more.
 *
 * Deliberately NOT "everything after the last user message": a prompt queued
 * while the agent works opens a new, still-unanswered turn, and the honest
 * brief view is then the previous, answered one — exactly the row `text`
 * already points at. Messages without a numeric `turn` (Codex, the pane
 * parser, an older reader) yield an empty list so callers fall back to `text`.
 */
export function selectLastAnsweredTurn<T extends { role: string; turn?: number }>(messages: T[]): T[] {
  let latest = -1;
  for (const message of messages) {
    if (message.role === 'assistant' && typeof message.turn === 'number' && message.turn > latest) {
      latest = message.turn;
    }
  }
  if (latest < 0) return [];
  return messages.filter((message) => message.role === 'assistant' && message.turn === latest);
}