@oar-sh/oar 0.9.11 → 0.9.12

package/shared/task-output-text.mjs

@@ -0,0 +1,737 @@
+// The text pipeline of a background task's console output: what the relay
+// does to a shell or Monitor task's log before anyone sees it, and the
+// pieces a viewer needs to show it.
+//
+//  - stripControlKeepSgr: terminal control codes out, colours (SGR) kept.
+//  - createLineFolder: a stream fed in arbitrary pieces becomes lines; a `\r`
+//    redraw (progress bars) collapses to its final state, adjacent identical
+//    lines fold into one with a count, and a line is cut at 4,000 characters.
+//  - ansiToSegments: one line as styled segments, 16 colours.
+//  - scrubTaskOutput / scrubSecrets: credentials out before output leaves the
+//    relay (the page, a paired relay's agent).
+//  - parseRunnerSummary: pass/fail counts of the common test runners.
+//  - stallMinutes and the stall setting's bounds.
+//
+// Pure ESM with no Node imports: the relay uses it for the REST tail, the
+// health poller and the remote action. The browser cannot import shared/
+// (only server/public is served); server/public/app/task-output-text.mjs keeps
+// a copy of the folder, the segments and the stall minutes, and its test pins
+// it to this module.
+
+// ─── Limits ──────────────────────────────────────────────────────────────────
+
+export const TASK_OUTPUT_LIMITS = Object.freeze({
+  // Initial tail: this many lines or this many bytes, whichever comes first.
+  tailLinesDefault: 2000,
+  tailLinesMax: 5000,
+  requestBytesMax: 256 * 1024,
+  // What a modal loads in total, paging back.
+  modalBytesMax: 5 * 1024 * 1024,
+  // One live emit, and the window the health poller parses.
+  chunkBytes: 64 * 1024,
+  healthBytes: 64 * 1024,
+  lineCharsMax: 4000,
+  outputPathMax: 1024,
+  // Ended rows: kept this long, at most this many per conversation.
+  endedKeepMs: 60 * 60 * 1000,
+  endedMax: 20,
+  // The worker's spool stops appending here.
+  spoolBytesMax: 50 * 1024 * 1024,
+  // The remote action's tail.
+  remoteLinesDefault: 200,
+  remoteLinesMax: 500,
+  remoteBytesMax: 32 * 1024,
+});
+
+export const TASK_OUTPUT_KINDS = Object.freeze(['file', 'spool', 'snapshot', 'none']);
+
+/** The kind a task row's output has: what it names, else `file` with a path and `none` without. */
+export function normalizeTaskOutputKind(kind, outputPath) {
+  const text = String(kind ?? '').trim().toLowerCase();
+  if (TASK_OUTPUT_KINDS.includes(text)) return text;
+  return outputPath ? 'file' : 'none';
+}
+
+/**
+ * The path a worker names for a task's output, or '' when it is not one the
+ * relay reads: absolute (POSIX or a Windows drive / UNC path), at most 1,024
+ * characters, no NUL. Pure string checks; the relay's reader also refuses
+ * symlinks and anything that is not a regular file.
+ */
+export function normalizeTaskOutputPath(value) {
+  const text = typeof value === 'string' ? value.trim() : '';
+  if (!text || text.length > TASK_OUTPUT_LIMITS.outputPathMax || text.includes('\0')) return '';
+  const absolute = text.startsWith('/') || /^[A-Za-z]:[\\/]/.test(text) || text.startsWith('\\\\');
+  return absolute ? text : '';
+}
+
+// ─── The stall setting ───────────────────────────────────────────────────────
+
+// "Mark a background task as quiet after N minutes" (app_settings key
+// task_stall_minutes).
+export const TASK_STALL_MINUTES_MIN = 1;
+export const TASK_STALL_MINUTES_MAX = 120;
+export const DEFAULT_TASK_STALL_MINUTES = 10;
+export const TASK_STALL_MINUTES_SETTING_KEY = 'task_stall_minutes';
+
+export function normalizeTaskStallMinutes(value, fallback = DEFAULT_TASK_STALL_MINUTES) {
+  if (value === null || value === undefined) return fallback;
+  if (typeof value === 'string' && value.trim() === '') return fallback;
+  const number = Number(value);
+  if (!Number.isFinite(number)) return fallback;
+  return Math.min(TASK_STALL_MINUTES_MAX, Math.max(TASK_STALL_MINUTES_MIN, Math.round(number)));
+}
+
+/** An unset or unreadable setting is the default. */
+export function readTaskStallMinutesSetting(storedValue) {
+  return normalizeTaskStallMinutes(storedValue);
+}
+
+export function parseTaskStallMinutesUpdate(value) {
+  const minutes = normalizeTaskStallMinutes(value, null);
+  if (minutes === null) return { ok: false, error: 'minutes must be a number' };
+  return { ok: true, minutes };
+}
+
+function toMillis(value) {
+  if (value == null || value === '') return NaN;
+  const number = Number(value);
+  if (Number.isFinite(number)) return number;
+  return Date.parse(String(value));
+}
+
+/** Whole minutes since the last output, or null when unknown. */
+export function stallMinutes(lastOutputAt, now = Date.now()) {
+  const last = toMillis(lastOutputAt);
+  const current = toMillis(now);
+  if (!Number.isFinite(last) || !Number.isFinite(current) || last <= 0) return null;
+  return Math.max(0, Math.floor((current - last) / 60_000));
+}
+
+// ─── Control codes ───────────────────────────────────────────────────────────
+
+// SGR (colours and text style) is the only escape kept. Its parameters are
+// digits and semicolons; the colon form is rare and goes with the rest.
+const SGR_PATTERN = /\x1b\[([0-9;]*)m/g;
+// OSC (titles, hyperlinks, clipboard): ESC ] … BEL | ST, stripped whole. An
+// unterminated one at the end is stripped to the end.
+const OSC_PATTERN = /\x1b\][^\x07\x1b]*(?:\x07|\x1b\\|$)/g;
+// Every CSI (cursor moves, erases, modes) that is not SGR.
+const CSI_PATTERN = /\x1b\[[0-?]*[ -/]*[@-~]/g;
+// Character-set designations (ESC ( B) and the remaining two-byte escapes.
+const CHARSET_PATTERN = /\x1b[()*+\-./][0-~]/g;
+// The remaining two-byte escapes (ESC 7, ESC =, ESC M, a CSI cut short) and
+// a lone ESC; never the start of a kept SGR.
+const OTHER_ESCAPE_PATTERN = /\x1b(?!\[[0-9;]*m)[0-~]?/g;
+// C0 except \t \n \r, DEL, C1, and a lone ESC left over.
+const CONTROL_PATTERN = /[\x00-\x08\x0b\x0c\x0e-\x1a\x1c-\x1f\x7f-\x9f]/g;
+
+/**
+ * Drops OSC, every CSI but SGR, the other escapes, C1 and C0 controls except
+ * `\n`, `\r` and `\t`; keeps `ESC [ … m`.
+ */
+export function stripControlKeepSgr(text) {
+  const value = String(text ?? '');
+  if (!/[\x00-\x08\x0b-\x0c\x0e-\x1f\x7f-\x9f]/.test(value)) return value;
+  return value
+    .replace(OSC_PATTERN, '')
+    .replace(CSI_PATTERN, (match) => (/^\x1b\[[0-9;]*m$/.test(match) ? match : ''))
+    .replace(CHARSET_PATTERN, '')
+    .replace(OTHER_ESCAPE_PATTERN, '')
+    .replace(CONTROL_PATTERN, '');
+}
+
+/** The text without its SGR codes: what a reader without colours sees. */
+export function stripSgr(text) {
+  return String(text ?? '').replace(SGR_PATTERN, '');
+}
+
+// ─── Lines ───────────────────────────────────────────────────────────────────
+
+function cutLine(text) {
+  return text.length > TASK_OUTPUT_LIMITS.lineCharsMax ? text.slice(0, TASK_OUTPUT_LIMITS.lineCharsMax) : text;
+}
+
+// A line's final state: a `\r` not followed by `\n` discards what came before
+// it (a trailing one still waits for what follows).
+function finalLineState(raw) {
+  const body = raw.endsWith('\r') ? raw.slice(0, -1) : raw;
+  const at = body.lastIndexOf('\r');
+  return at === -1 ? body : body.slice(at + 1);
+}
+
+/**
+ * Lines from a stream fed in arbitrary pieces. `feed(text)` returns
+ * `{ lines, partial, changedFrom }`: every completed line so far as
+ * `{ text, count }` (adjacent identical lines folded, `count` unbounded), the
+ * open partial line in its current state ('' when there is none), and the
+ * index of the first line this feed added or changed (a folded repeat bumps
+ * the count of the line before). `end()` closes the partial line and returns
+ * the same shape. `lines` is the folder's own array: copy it to keep it.
+ */
+export function createLineFolder() {
+  const lines = [];
+  let partial = '';
+  let changedFrom = 0;
+
+  function push(text) {
+    const line = cutLine(text);
+    const last = lines[lines.length - 1];
+    if (last && last.text === line) {
+      last.count += 1;
+      changedFrom = Math.min(changedFrom, lines.length - 1);
+    } else {
+      changedFrom = Math.min(changedFrom, lines.length);
+      lines.push({ text: line, count: 1 });
+    }
+  }
+
+  function result() {
+    const out = { lines, partial: finalLineState(partial), changedFrom };
+    changedFrom = lines.length;
+    return out;
+  }
+
+  return {
+    feed(text) {
+      const chunk = String(text ?? '');
+      if (chunk) {
+        const parts = (partial + chunk).split('\n');
+        partial = parts.pop();
+        for (const raw of parts) push(finalLineState(raw));
+        // The open line stays small: a progress bar redraws itself
+        // thousands of times on one line.
+        const keepCr = partial.endsWith('\r');
+        partial = cutLine(finalLineState(partial)) + (keepCr ? '\r' : '');
+      }
+      return result();
+    },
+    end() {
+      const last = finalLineState(partial);
+      partial = '';
+      if (last) push(last);
+      return result();
+    },
+    get lines() {
+      return lines;
+    },
+  };
+}
+
+/**
+ * A whole text as the line model the console shows: `{ text, count, ansi }`
+ * per line (`ansi` when it carries colour codes), the last one marked
+ * `partial` when the text does not end with a newline.
+ */
+export function foldTaskOutput(text) {
+  const folder = createLineFolder();
+  const { lines, partial } = folder.feed(text);
+  const model = lines.map((line) => ({ text: line.text, count: line.count, ansi: line.text.includes('\x1b') }));
+  if (partial) model.push({ text: partial, count: 1, ansi: partial.includes('\x1b'), partial: true });
+  return model;
+}
+
+// ─── Colours ─────────────────────────────────────────────────────────────────
+
+// The 16-colour palette (xterm defaults), for mapping 256 and truecolor.
+const PALETTE_16 = [
+  [0, 0, 0], [205, 0, 0], [0, 205, 0], [205, 205, 0], [0, 0, 238], [205, 0, 205], [0, 205, 205], [229, 229, 229],
+  [127, 127, 127], [255, 0, 0], [0, 255, 0], [255, 255, 0], [92, 92, 255], [255, 0, 255], [0, 255, 255], [255, 255, 255],
+];
+
+function nearest16([r, g, b]) {
+  let best = 0;
+  let bestDistance = Infinity;
+  PALETTE_16.forEach(([pr, pg, pb], index) => {
+    const distance = (r - pr) ** 2 + (g - pg) ** 2 + (b - pb) ** 2;
+    if (distance < bestDistance) {
+      best = index;
+      bestDistance = distance;
+    }
+  });
+  return best;
+}
+
+function rgbOf256(n) {
+  if (n < 16) return PALETTE_16[n];
+  if (n >= 232) {
+    const level = 8 + (n - 232) * 10;
+    return [level, level, level];
+  }
+  const index = n - 16;
+  const steps = [0, 95, 135, 175, 215, 255];
+  return [steps[Math.floor(index / 36)], steps[Math.floor(index / 6) % 6], steps[index % 6]];
+}
+
+function clampByte(value) {
+  const number = Number(value);
+  return Number.isFinite(number) ? Math.min(255, Math.max(0, number)) : 0;
+}
+
+function plainStyle() {
+  return { fg: null, bg: null, bold: false, dim: false, italic: false, underline: false };
+}
+
+function applySgr(state, params) {
+  const codes = params === '' ? [0] : params.split(';').map((part) => (part === '' ? 0 : Number(part)));
+  for (let i = 0; i < codes.length; i += 1) {
+    const code = codes[i];
+    if (code === 0) Object.assign(state, plainStyle());
+    else if (code === 1) state.bold = true;
+    else if (code === 2) state.dim = true;
+    else if (code === 3) state.italic = true;
+    else if (code === 4) state.underline = true;
+    else if (code === 22) {
+      state.bold = false;
+      state.dim = false;
+    } else if (code === 23) state.italic = false;
+    else if (code === 24) state.underline = false;
+    else if (code >= 30 && code <= 37) state.fg = code - 30;
+    else if (code === 39) state.fg = null;
+    else if (code >= 40 && code <= 47) state.bg = code - 40;
+    else if (code === 49) state.bg = null;
+    else if (code >= 90 && code <= 97) state.fg = code - 90 + 8;
+    else if (code >= 100 && code <= 107) state.bg = code - 100 + 8;
+    else if (code === 38 || code === 48) {
+      const key = code === 38 ? 'fg' : 'bg';
+      if (codes[i + 1] === 5 && i + 2 < codes.length) {
+        state[key] = nearest16(rgbOf256(Math.min(255, Math.max(0, codes[i + 2] || 0))));
+        i += 2;
+      } else if (codes[i + 1] === 2 && i + 4 < codes.length) {
+        state[key] = nearest16([clampByte(codes[i + 2]), clampByte(codes[i + 3]), clampByte(codes[i + 4])]);
+        i += 4;
+      } else {
+        break;
+      }
+    }
+  }
+}
+
+/**
+ * One line as styled segments `{ text, fg, bg, bold, dim, italic, underline }`,
+ * colours 0–15 or null (256 and truecolor mapped to the nearest of the 16).
+ * Adjacent segments never share a style. Any escape that is not SGR is
+ * dropped from the text.
+ */
+export function ansiToSegments(lineText) {
+  const text = String(lineText ?? '');
+  const state = plainStyle();
+  const segments = [];
+  const pushText = (value) => {
+    const clean = stripControlKeepSgr(value).replace(/\x1b/g, '');
+    if (!clean) return;
+    const last = segments[segments.length - 1];
+    if (last && last.fg === state.fg && last.bg === state.bg && last.bold === state.bold
+      && last.dim === state.dim && last.italic === state.italic && last.underline === state.underline) {
+      last.text += clean;
+      return;
+    }
+    segments.push({ text: clean, ...state });
+  };
+  let at = 0;
+  const pattern = new RegExp(SGR_PATTERN.source, 'g');
+  for (let match = pattern.exec(text); match; match = pattern.exec(text)) {
+    pushText(text.slice(at, match.index));
+    applySgr(state, match[1]);
+    at = match.index + match[0].length;
+  }
+  pushText(text.slice(at));
+  return segments;
+}
+
+// ─── Secrets ─────────────────────────────────────────────────────────────────
+
+const REDACTED = '[redacted]';
+
+// The generic rule of the CLI helpers (services/cli-process-runner.mjs):
+// anything token-shaped. Right for a login's error tail, too eager for a
+// test log (long snake_case test names, 40-character commit ids).
+const CLI_TOKEN_PATTERNS = Object.freeze([
+  /sk-ant-[A-Za-z0-9_-]+/g,
+  /[A-Za-z0-9_-]{32,}/g,
+]);
+
+/**
+ * Redacts anything token-shaped plus an exact secret the caller knows it
+ * handed over (the pasted Claude login code), so an error tail can be
+ * surfaced to the UI without leaking a credential. Long PKCE segments in an
+ * authorize URL get caught by the generic rule too, which is fine: the URL is
+ * already carried in the login payload. (Moved here from
+ * services/cli-process-runner.mjs, which re-exports it.)
+ */
+export function scrubSecrets(text, submittedCode = '') {
+  let output = String(text == null ? '' : text);
+  const code = String(submittedCode == null ? '' : submittedCode).trim();
+  if (code.length >= 4) {
+    output = output.split(code).join(REDACTED);
+  }
+  for (const pattern of CLI_TOKEN_PATTERNS) output = output.replace(pattern, REDACTED);
+  return output;
+}
+
+// Known credential shapes. Each keeps its prefix so a reader still sees what
+// kind of value was there.
+const KNOWN_TOKEN_RULES = Object.freeze([
+  // Anthropic, OpenAI and other `sk-` keys.
+  [/\bsk-(?:ant-|proj-|live-|test-)?[A-Za-z0-9_-]{16,}/g, 'sk-[redacted]'],
+  // GitHub tokens (classic and fine-grained).
+  [/\b(gh[pousr]_)[A-Za-z0-9]{20,}/g, '$1[redacted]'],
+  [/\b(github_pat_)[A-Za-z0-9_]{20,}/g, '$1[redacted]'],
+  // Slack.
+  [/\b(xox[abposr]-)[A-Za-z0-9-]{10,}/g, '$1[redacted]'],
+  // npm.
+  [/\b(npm_)[A-Za-z0-9]{30,}/g, '$1[redacted]'],
+  // AWS access key ids.
+  [/\b(AKIA|ASIA)[0-9A-Z]{16}\b/g, '$1[redacted]'],
+  // Google API keys.
+  [/\b(AIza)[0-9A-Za-z_-]{30,}/g, '$1[redacted]'],
+  // Bearer / Basic credentials in a header or a curl line.
+  [/\b(Bearer|Basic)\s+[A-Za-z0-9._~+/=-]{8,}/g, '$1 [redacted]'],
+  // A password in a URL: scheme://user:password@host.
+  [/\b([a-z][a-z0-9+.-]*:\/\/[^\s:/@]+:)[^\s@/]+@/gi, '$1[redacted]@'],
+]);
+
+// NAME=value with a secret-sounding environment name (API_TOKEN=…,
+// DB_PASSWORD="…"); the name stays. A value without a letter (a count, a
+// port) or a placeholder (true, null, ***) is no secret.
+const ENV_ASSIGNMENT_PATTERN = /\b([A-Za-z0-9_]*(?:SECRET|TOKEN|PASSWORD|PASSWD|PASSPHRASE|API_?KEY|ACCESS_?KEY|PRIVATE_?KEY|CREDENTIALS?|AUTH(?!OR))[A-Za-z0-9_]*)=(["']?)(?!\[redacted\]|(?:true|false|null|none|undefined)\b|\*+)(?=[^\s"'`]*[A-Za-z])([^\s"'`]{4,})/gi;
+// `token: value`, `"password": "value"`, `api_key = value`, Authorization: …
+// A Bearer/Basic value was handled above and keeps its scheme word.
+const KEY_VALUE_PATTERN = /\b(token|secret|password|passwd|passphrase|api[_-]?key|access[_-]?key|client[_-]?secret|authorization)(["']?\s*[:=]\s*["']?)(?!(?:Bearer|Basic)\b|\[redacted\]|(?:true|false|null|none|undefined)\b|\*+)(?=[^\s"',;]*[A-Za-z])([^\s"',;]{4,})/gi;
+// Anything else token-shaped: 32 or more of [A-Za-z0-9_-] with an upper-case
+// letter, a lower-case letter and a digit. Random keys have all three; test
+// names, paths and hex commit ids rarely do.
+const GENERIC_TOKEN_PATTERN = /(?<![A-Za-z0-9_-])(?=[A-Za-z0-9_-]*[A-Z])(?=[A-Za-z0-9_-]*[a-z])(?=[A-Za-z0-9_-]*[0-9])[A-Za-z0-9_-]{32,}(?![A-Za-z0-9_-])/g;
+// A PEM private key block, whole.
+const PEM_PATTERN = /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]*?(?:-----END [A-Z ]*PRIVATE KEY-----|$)/g;
+
+function scrubPlain(text) {
+  let output = text.replace(PEM_PATTERN, '[redacted private key]');
+  for (const [pattern, replacement] of KNOWN_TOKEN_RULES) output = output.replace(pattern, replacement);
+  output = output
+    .replace(ENV_ASSIGNMENT_PATTERN, (_match, name, quote) => `${name}=${quote}${REDACTED}`)
+    .replace(KEY_VALUE_PATTERN, (_match, key, separator) => `${key}${separator}${REDACTED}`)
+    .replace(GENERIC_TOKEN_PATTERN, REDACTED);
+  return output;
+}
+
+/**
+ * Credentials out of a task's output: the known token shapes, secret-named
+ * `KEY=value` and `key: value` pairs, Bearer/Basic values, URL passwords,
+ * private keys, and long random-looking tokens. SGR codes are left alone
+ * (the text between them is scrubbed), so colours survive.
+ */
+export function scrubTaskOutput(text) {
+  const value = String(text ?? '');
+  if (!value) return value;
+  if (!value.includes('\x1b')) return scrubPlain(value);
+  return value.split(/(\x1b\[[0-9;]*m)/).map((part, index) => (index % 2 === 1 ? part : scrubPlain(part))).join('');
+}
+
+/** What leaves the relay: control codes out (SGR kept), then credentials out. */
+export function cleanTaskOutput(text) {
+  return scrubTaskOutput(stripControlKeepSgr(text));
+}
+
+// ─── Test runners ────────────────────────────────────────────────────────────
+
+const PASS_MARK = /^\s*(?:[✓✔√]|\[PASS\])\s+/;
+const FAIL_MARK = /^\s*(?:[✗✘×✕✖]|\[FAIL\])\s+/;
+
+function lineTextOf(line) {
+  const raw = line && typeof line === 'object' ? line.text : line;
+  return stripSgr(String(raw ?? '')).replace(/\r/g, '');
+}
+
+function countOf(line) {
+  const count = line && typeof line === 'object' ? Number(line.count) : 1;
+  return Number.isInteger(count) && count > 0 ? count : 1;
+}
+
+function countsFrom(text) {
+  const counts = {};
+  for (const match of text.matchAll(/(\d+)\s+(passed|failed|skipped|todo|pending|flaky|errors?|xfailed|xpassed|deselected|warnings?|interrupted|did not run)\b/gi)) {
+    counts[match[2].toLowerCase()] = (counts[match[2].toLowerCase()] || 0) + Number(match[1]);
+  }
+  return counts;
+}
+
+function summary(runner, { passed = 0, failed = 0, skipped = 0, flaky = 0 } = {}, extra = {}) {
+  return {
+    runner,
+    passed,
+    failed,
+    skipped,
+    ...(flaky ? { flaky } : {}),
+    ...extra,
+  };
+}
+
+// Final summaries, checked from the end of the log back. Each returns the
+// counts or null.
+const FINAL_SUMMARIES = [
+  // vitest: " Tests  2 failed | 412 passed | 3 skipped (417)"
+  {
+    runner: 'vitest',
+    match(texts, index) {
+      const text = texts[index];
+      if (!/^\s*Tests\s{2,}.*\(\d+\)\s*$/.test(text) || /^\s*Tests:/.test(text)) return null;
+      const counts = countsFrom(text);
+      return summary('vitest', {
+        passed: counts.passed || 0,
+        failed: counts.failed || 0,
+        skipped: (counts.skipped || 0) + (counts.todo || 0),
+      });
+    },
+  },
+  // jest: "Tests:       2 failed, 3 skipped, 412 passed, 417 total"
+  {
+    runner: 'jest',
+    match(texts, index) {
+      const text = texts[index];
+      if (!/^\s*Tests:\s+.*\b\d+\s+total\s*$/.test(text)) return null;
+      const counts = countsFrom(text);
+      return summary('jest', {
+        passed: counts.passed || 0,
+        failed: counts.failed || 0,
+        skipped: (counts.skipped || 0) + (counts.todo || 0) + (counts.pending || 0),
+      });
+    },
+  },
+  // pytest: "======= 2 failed, 412 passed, 3 skipped in 3.21s ======="
+  {
+    runner: 'pytest',
+    match(texts, index) {
+      const text = texts[index];
+      if (!/^=+ .*\bin [\d.]+s(?: \([^)]*\))? =+\s*$/.test(text)) return null;
+      const counts = countsFrom(text);
+      if (!Object.keys(counts).length && !/no tests ran/.test(text)) return null;
+      return summary('pytest', {
+        passed: (counts.passed || 0) + (counts.xpassed || 0),
+        failed: (counts.failed || 0) + (counts.error || 0) + (counts.errors || 0),
+        skipped: (counts.skipped || 0) + (counts.xfailed || 0) + (counts.deselected || 0),
+      });
+    },
+  },
+  // node:test, spec reporter ("ℹ pass 1") and TAP ("# pass 1"): the block
+  // ends with duration_ms.
+  {
+    runner: 'node:test',
+    match(texts, index) {
+      if (!/^\s*(?:ℹ|#)\s+duration_ms\s+[\d.]+\s*$/.test(texts[index])) return null;
+      const counts = {};
+      for (let i = index - 1; i >= 0 && i >= index - 12; i -= 1) {
+        const field = /^\s*(?:ℹ|#)\s+(tests|suites|pass|fail|cancelled|skipped|todo)\s+(\d+)\s*$/.exec(texts[i]);
+        if (!field) break;
+        counts[field[1]] = Number(field[2]);
+      }
+      if (counts.pass === undefined && counts.fail === undefined) return null;
+      return summary('node:test', {
+        passed: counts.pass || 0,
+        failed: (counts.fail || 0) + (counts.cancelled || 0),
+        skipped: (counts.skipped || 0) + (counts.todo || 0),
+      });
+    },
+  },
+  // Playwright: "  1 failed" … "  3 skipped" … "  412 passed (1.4m)": the
+  // block's lines sit together at the end, the passed line carries the time.
+  {
+    runner: 'playwright',
+    match(texts, index) {
+      const text = texts[index];
+      const own = /^\s{2}(\d+) (passed|failed|flaky|skipped|interrupted|did not run)(?: \(([\dhms.\s]+)\))?\s*$/.exec(text);
+      if (!own) return null;
+      // Walk up over the block: count lines, and the indented test titles
+      // listed under "failed" / "flaky".
+      const counts = {};
+      let timed = !!own[3];
+      let i = index;
+      for (; i >= 0; i -= 1) {
+        const line = texts[i];
+        const head = /^\s{2}(\d+) (passed|failed|flaky|skipped|interrupted|did not run)(?: \(([\dhms.\s]+)\))?\s*$/.exec(line);
+        if (head) {
+          counts[head[2]] = Number(head[1]);
+          if (head[3]) timed = true;
+          continue;
+        }
+        if (/^\s{4}\S/.test(line)) continue;
+        break;
+      }
+      if (!timed) return null;
+      return summary('playwright', {
+        passed: counts.passed || 0,
+        failed: (counts.failed || 0) + (counts.interrupted || 0),
+        skipped: (counts.skipped || 0) + (counts['did not run'] || 0),
+        flaky: counts.flaky || 0,
+      });
+    },
+  },
+  // turbo (also run through pnpm): " Tasks:    3 successful, 4 total", with a
+  // " Failed:    pkg#task" line when one failed.
+  {
+    runner: 'turbo',
+    match(texts, index) {
+      const text = texts[index];
+      const head = /^\s*Tasks:\s+(\d+) successful(?:, (\d+) failed)?, (\d+) total\s*$/.exec(text);
+      if (!head) return null;
+      const successful = Number(head[1]);
+      const total = Number(head[3]);
+      const failed = head[2] !== undefined ? Number(head[2]) : Math.max(0, total - successful);
+      return summary('turbo', { passed: successful, failed });
+    },
+  },
+];
+
+// What a log looks like while a run is still going: the runner's own header
+// lines. Markers alone (✓ …) say nothing, a build log has them too.
+const RUNNER_HEADERS = [
+  ['playwright', /^Running \d+ tests? using \d+ workers?/],
+  ['vitest', /^\s*(?:RUN|DEV)\s+v\d+\.\d+/],
+  ['jest', /^\s*(?:PASS|FAIL)\s+\S+\.(?:[cm]?[jt]sx?)\b/],
+  ['pytest', /^=+ test session starts =+$/],
+  ['node:test', /^(?:TAP version \d+|[✔✖] .+ \([\d.]+ms\)\s*$)/],
+  ['tap', /^(?:not )?ok \d+ - /],
+  ['turbo', /^\s*•\s+Running \S+ in \d+ packages?/],
+];
+
+function liveCounts(runner, texts) {
+  let passed = 0;
+  let failed = 0;
+  let skipped = 0;
+  let firstFailureLine = null;
+  const fail = (index, amount = 1) => {
+    failed += amount;
+    if (firstFailureLine === null) firstFailureLine = index;
+  };
+  texts.forEach((text, index) => {
+    const repeat = texts.counts[index];
+    if (runner === 'pytest') {
+      const verbose = /\s(PASSED|FAILED|ERROR|SKIPPED|XFAIL|XPASS)(?:\s+\[\s*\d+%\])?\s*$/.exec(text);
+      if (verbose) {
+        if (verbose[1] === 'PASSED' || verbose[1] === 'XPASS') passed += repeat;
+        else if (verbose[1] === 'FAILED' || verbose[1] === 'ERROR') fail(index, repeat);
+        else skipped += repeat;
+        return;
+      }
+      // Progress lines: "tests/test_a.py ..F.s   [ 40%]"
+      // and their wrapped continuations ("........  [ 56%]").
+      const dots = /^\S+\.py\s+([.FEsxX]+)(?:\s+\[\s*\d+%\])?\s*$/.exec(text)
+        || /^([.FEsxX]+)\s*\[\s*\d+%\]\s*$/.exec(text);
+      if (dots) {
+        for (const mark of dots[1]) {
+          if (mark === '.' || mark === 'X') passed += 1;
+          else if (mark === 'F' || mark === 'E') fail(index);
+          else skipped += 1;
+        }
+      }
+      return;
+    }
+    if (runner === 'tap' || (runner === 'node:test' && /^\s*(?:not )?ok \d+ /.test(text))) {
+      const tap = /^\s*(not )?ok \d+\b(.*)$/.exec(text);
+      if (!tap) return;
+      // A subtest line is indented; only top-level results count.
+      if (/^\s/.test(text)) return;
+      if (/#\s*(?:SKIP|TODO)\b/i.test(tap[2])) skipped += repeat;
+      else if (tap[1]) fail(index, repeat);
+      else passed += repeat;
+      return;
+    }
+    if (runner === 'vitest') {
+      // File lines: " ✓ src/a.test.ts (12 tests) 34ms", " ❯ src/b.test.ts (5 tests | 2 failed) 20ms".
+      // The × lines under a failed file repeat its failures.
+      const file = /^\s*[✓❯×]\s+\S+\s+\((\d+) tests?(?: \| (\d+) failed)?(?: \| (\d+) skipped)?\)/.exec(text);
+      if (file) {
+        const total = Number(file[1]);
+        const failedHere = Number(file[2] || 0);
+        const skippedHere = Number(file[3] || 0);
+        passed += Math.max(0, total - failedHere - skippedHere);
+        skipped += skippedHere;
+        if (failedHere) fail(index, failedHere);
+        return;
+      }
+      if (FAIL_MARK.test(text) && firstFailureLine === null) firstFailureLine = index;
+      return;
+    }
+    if (runner === 'jest') {
+      // Without --verbose jest marks files only; each failed test gets a
+      // "● suite › test" block (but "● Console" is a log block).
+      if (/^\s*FAIL\s/.test(text)) {
+        if (firstFailureLine === null) firstFailureLine = index;
+        return;
+      }
+      if (/^\s*●\s(?!Console\b)/.test(text)) {
+        fail(index, repeat);
+        return;
+      }
+      if (PASS_MARK.test(text)) passed += repeat;
+      else if (FAIL_MARK.test(text) && firstFailureLine === null) firstFailureLine = index;
+      return;
+    }
+    if (runner === 'playwright' && /^\s*-\s+\d+\s/.test(text)) {
+      skipped += repeat;
+      return;
+    }
+    if (PASS_MARK.test(text)) passed += repeat;
+    else if (FAIL_MARK.test(text)) {
+      // node:test's "✖ failing tests:" heads the repeat of the failures.
+      if (/failing tests:\s*$/.test(text)) return;
+      fail(index, repeat);
+    }
+  });
+  return { passed, failed, skipped, firstFailureLine };
+}
+
+// Where the first failure is told in detail, for a run that has its final
+// summary (the per-test marks may sit far above).
+function firstFailureIndex(runner, texts) {
+  const patterns = {
+    playwright: [/^\s{2}\d+\) /, /^\s*[✘×]\s+\d+\s/],
+    vitest: [/^\s*FAIL\s/, /^\s*[×✗]\s/],
+    jest: [/^\s*●\s/, /^\s*[✕×]\s/],
+    pytest: [/^_{3,} .+ _{3,}$/, /\s(?:FAILED|ERROR)(?:\s+\[\s*\d+%\])?\s*$/, /^\S+\.py\s+[.sxX]*[FE]/],
+    'node:test': [/^\s*✖\s(?!failing tests:)/, /^not ok \d+ /],
+    tap: [/^not ok \d+ /],
+    turbo: [/ERR!|ERROR|command finished with error/i],
+  }[runner] || [FAIL_MARK];
+  // In order of preference: the detailed report first, the mark second.
+  for (const pattern of patterns) {
+    const index = texts.findIndex((text) => pattern.test(text));
+    if (index !== -1) return index;
+  }
+  return null;
+}
+
+/**
+ * Pass/fail counts of a test run from its log lines (strings or the line
+ * model's `{ text, count }`; colours are ignored), or null when the log is
+ * not a known runner's. Returns `{ runner, passed, failed, skipped, flaky?,
+ * firstFailureLine, final }`: `final` when the counts come from the runner's
+ * closing summary, otherwise they are counted from the per-test marks while
+ * the run is still going. `firstFailureLine` is the index (into `lines`) of
+ * the first failure, or null. Runners: vitest, Playwright, node:test (spec
+ * and TAP), jest, pytest, turbo (and pnpm running it).
+ */
+export function parseRunnerSummary(lines) {
+  const list = Array.isArray(lines) ? lines : String(lines ?? '').split('\n');
+  const texts = list.map(lineTextOf);
+  texts.counts = list.map(countOf);
+  for (let index = texts.length - 1; index >= 0; index -= 1) {
+    if (!texts[index].trim()) continue;
+    for (const candidate of FINAL_SUMMARIES) {
+      const found = candidate.match(texts, index);
+      if (!found) continue;
+      const failure = found.failed > 0 ? firstFailureIndex(found.runner, texts.slice(0, index)) : null;
+      return { ...found, firstFailureLine: failure, final: true };
+    }
+  }
+  let runner = null;
+  for (const text of texts) {
+    const header = RUNNER_HEADERS.find(([, pattern]) => pattern.test(text));
+    if (header) {
+      runner = header[0];
+      break;
+    }
+  }
+  if (!runner) return null;
+  return { runner, ...liveCounts(runner, texts), final: false };
+}