gstack/test/helpers/session-runner.ts

/**
 * Claude CLI subprocess runner for skill E2E testing.
 *
 * Spawns `claude -p` as a completely independent process (not via Agent SDK),
 * so it works inside Claude Code sessions. Pipes prompt via stdin, streams
 * NDJSON output for real-time progress, scans for browse errors.
 */

import * as fs from 'fs';
import * as path from 'path';
import * as os from 'os';
import { spawn } from 'child_process';
import { Readable } from 'node:stream';
import { createHash, type Hash } from 'node:crypto';
import { getProjectEvalDir } from './eval-store';
import { hermeticChildEnv, isHermeticEnabled } from './hermetic-env';
import { killProcessGroup } from '../../scripts/test-strict-output';
import { resolveEvalModel } from '../../lib/eval-model';

const GSTACK_DEV_DIR = path.join(os.homedir(), '.gstack-dev');
const HEARTBEAT_PATH = path.join(GSTACK_DEV_DIR, 'e2e-live.json'); // heartbeat stays global
const PROJECT_DIR = path.dirname(getProjectEvalDir()); // ~/.gstack/projects/$SLUG/

/** Sanitize test name for use as filename: strip leading slashes, replace / with - */
export function sanitizeTestName(name: string): string {
  return name.replace(/^\/+/, '').replace(/\//g, '-');
}

/** Atomic write: write to .tmp then rename. Non-fatal on error. */
function atomicWriteSync(filePath: string, data: string): void {
  const tmp = filePath + '.tmp';
  fs.writeFileSync(tmp, data);
  fs.renameSync(tmp, filePath);
}

export interface CostEstimate {
  inputChars: number;
  outputChars: number;
  estimatedTokens: number;
  estimatedCost: number;  // USD
  turnsUsed: number;
}

export interface SkillTestResult {
  toolCalls: Array<{ tool: string; input: any; output: string }>;
  browseErrors: string[];
  exitReason: string;
  duration: number;
  output: string;
  costEstimate: CostEstimate;
  transcript: any[];
  /** Which model was used for this test (added for Sonnet/Opus split diagnostics) */
  model: string;
  /** Time from spawn to first NDJSON line, in ms (added for rate-limit diagnostics) */
  firstResponseMs: number;
  /** Peak latency between consecutive tool calls, in ms */
  maxInterTurnMs: number;
}

/** Local default startup grace: 90s covers observed API queue latency
 *  (60-90s receipts) without letting a dead API burn a 600s budget. */
export const STARTUP_GRACE_MS = 90_000;
/** CI floor (TODOS-filed): shared runners queue harder; killing startup
 *  before 300s in CI converts ordinary queueing into false failures.
 *  Pinned by test/session-runner-startup-grace.test.ts. */
export const STARTUP_GRACE_CI_FLOOR_MS = 300_000;
/** Existing pipe-drain allowance; never adds model work time. */
export const SESSION_DRAIN_GRACE_MS = 5_000;

const BROWSE_ERROR_PATTERNS = [
  /Unknown command: \w+/,
  /Unknown snapshot flag: .+/,
  /ERROR: browse binary not found/,
  /Server failed to start/,
  /no such file or directory.*browse/i,
];

// --- Testable NDJSON parser ---

export interface ParsedNDJSON {
  transcript: any[];
  resultLine: any | null;
  turnCount: number;
  toolCallCount: number;
  toolCalls: Array<{ tool: string; input: any; output: string }>;
}

function toolResultText(content: unknown): string | null {
  if (typeof content === 'string') return content;
  if (!Array.isArray(content)) return null;
  return content.flatMap(block => block?.type === 'text' && typeof block.text === 'string'
    ? [block.text] : []).join('\n');
}

/**
 * Parse an array of NDJSON lines into structured transcript data.
 * Pure function — no I/O, no side effects. Used by both the streaming
 * reader and unit tests.
 */
export function parseNDJSON(lines: string[]): ParsedNDJSON {
  const transcript: any[] = [];
  let resultLine: any = null;
  let turnCount = 0;
  let toolCallCount = 0;
  const toolCalls: ParsedNDJSON['toolCalls'] = [];
  const callsByParent = new Map<string | null, Map<string, ParsedNDJSON['toolCalls'][number]>>();

  for (const line of lines) {
    if (!line.trim()) continue;
    try {
      const event = JSON.parse(line);
      transcript.push(event);

      // Track turns and tool calls from assistant events
      if (event.type === 'assistant') {
        turnCount++;
        const content = event.message?.content || [];
        for (const item of content) {
          if (item.type === 'tool_use') {
            toolCallCount++;
            const call = {
              tool: item.name || 'unknown',
              input: item.input || {},
              output: '',
            };
            toolCalls.push(call);
            if (typeof item.id === 'string') {
              // Forwarded subagent events may reuse a parent's tool-use ID.
              const parent = event.parent_tool_use_id ?? null;
              let calls = callsByParent.get(parent);
              if (!calls) callsByParent.set(parent, calls = new Map());
              calls.set(item.id, call);
            }
          }
        }
      }

      if (event.type === 'user' && Array.isArray(event.message?.content)) {
        const results = event.message.content.filter((item: any) => item?.type === 'tool_result');
        const calls = callsByParent.get(event.parent_tool_use_id ?? null);
        for (const result of results) {
          const call = calls?.get(result.tool_use_id);
          if (!call) continue;
          // A sole Agent/Task result also carries the clean verdict separately
          // from the message's agentId/usage wrapper. Keep only public text.
          const verdict = results.length === 1 && ['Agent', 'Task'].includes(call.tool)
            ? toolResultText(event.tool_use_result?.content) : null;
          call.output = verdict ?? toolResultText(result.content) ?? '';
        }
      }

      if (event.type === 'result') resultLine = event;
    } catch { /* skip malformed lines */ }
  }

  return { transcript, resultLine, turnCount, toolCallCount, toolCalls };
}

function truncate(s: string, max: number): string {
  return s.length > max ? s.slice(0, max) + '…' : s;
}

/** Diagnostic-only projection. Partial input never becomes a complete tool call. */
function publicStreamProjection(startTime: number): (line: string) => string {
  let messageId: string | undefined;
  const blocks = new Map<number, { type: string; tool?: string; bytes: number; hash: Hash }>();
  return (line) => {
    let row: any;
    try { row = JSON.parse(line); } catch {
      // A truncated line may contain private reasoning or unfinished tool input.
      return JSON.stringify({ type: 'public_stream_diagnostic', kind: 'unparseable_line',
        elapsedMs: Date.now() - startTime, bytes: Buffer.byteLength(line) });
    }
    if (row === null || typeof row !== 'object' || Array.isArray(row)) {
      return JSON.stringify({ type: 'public_stream_diagnostic', kind: 'non_object_line',
        elapsedMs: Date.now() - startTime, bytes: Buffer.byteLength(line) });
    }
    if (row.type === 'stream_event') {
      const event = row.event ?? {};
      if (event.type === 'message_start') {
        messageId = event.message?.id;
        blocks.clear();
      }
      const index = event.index;
      if (event.type === 'content_block_start' && Number.isInteger(index)) {
        blocks.set(index, { type: event.content_block?.type, tool: event.content_block?.name,
          bytes: 0, hash: createHash('sha256') });
      }
      const block = blocks.get(index);
      if (event.type === 'content_block_delta' && block?.type === 'tool_use'
        && event.delta?.type === 'input_json_delta' && typeof event.delta.partial_json === 'string') {
        const chunk = Buffer.from(event.delta.partial_json);
        block.bytes += chunk.length;
        block.hash.update(chunk);
      }
      const diagnostic = { type: 'public_stream_diagnostic', kind: event.type,
        session_id: row.session_id, messageId, index, elapsedMs: Date.now() - startTime,
        blockType: block?.type, toolName: block?.tool, deltaType: event.delta?.type,
        ...(block?.type === 'tool_use' ? { inputBytes: block.bytes,
          inputSha256: block.hash.copy().digest('hex') } : {}),
        ...(event.type === 'message_delta' ? { stopReason: event.delta?.stop_reason } : {}),
      };
      if (event.type === 'content_block_stop') blocks.delete(index);
      return JSON.stringify(diagnostic);
    }
    if (Array.isArray(row.message?.content)) {
      row.message.content = row.message.content.map((block: any) =>
        block.type === 'thinking' || block.type === 'redacted_thinking'
          ? { type: block.type, omitted: true } : block);
    }
    return JSON.stringify(row);
  };
}

// --- Main runner ---

export async function runSkillTest(options: {
  prompt: string;
  workingDirectory: string;
  maxTurns?: number;
  /** Optional harness contract appended to the native system prompt. */
  appendSystemPrompt?: string;
  /** Opt-in section completion reserve; describes the existing entry deadline. */
  completionReserveMs?: number;
  /** Approval allowlist; does not restrict which tools the model can see. */
  allowedTools?: string[];
  /** Optional built-in tool availability. Omit to preserve the CLI defaults. */
  tools?: string[];
  /** Opt-in public block timing/input-size diagnostics; never completion evidence. */
  publicStreamDiagnostics?: boolean;
  timeout?: number;
  testName?: string;
  runId?: string;
  /** Model to use. Defaults to the frontier eval model (overridable via EVALS_MODEL env). */
  model?: string;
  /** Extra env vars merged into the spawned claude -p process. Useful for
   *  per-test GSTACK_HOME overrides so the test doesn't have to spell out
   *  env setup in the prompt itself. */
  env?: Record<string, string>;
  /** Startup-phase deadline: if NO NDJSON byte arrives within this window,
   *  the run is killed EARLY with exitReason 'timeout_startup' instead of
   *  burning the whole work budget waiting on an API that is not answering
   *  (the recurring '0 turns / $0.00' class — four budget-bump receipts).
   *  Defaults to min(STARTUP_GRACE_MS, timeout); the CI floor is higher
   *  because CI queueing is real. Startup and model work share timeout;
   *  only pipe cleanup may use the separate SESSION_DRAIN_GRACE_MS. */
  startupGraceMs?: number;
  /** Cancel the owned process group when an enclosing attempt expires. */
  signal?: AbortSignal;
}): Promise<SkillTestResult> {
  const startTime = Date.now();
  options.signal?.throwIfAborted();
  const {
    prompt,
    workingDirectory,
    maxTurns = 15,
    allowedTools = ['Bash', 'Read', 'Write'],
    timeout = 120_000,
    testName,
    runId,
    env: extraEnv,
    signal,
  } = options;
  // The CI floor is a FLOOR, not a default: an explicit startupGraceMs below
  // 300s in CI would re-open the queueing-becomes-false-red hole the floor
  // exists for (review finding — the name promised a clamp the code lacked).
  // Local runs honor the caller verbatim; timeout still caps everything.
  const requestedGrace = options.startupGraceMs ?? (process.env.CI ? STARTUP_GRACE_CI_FLOOR_MS : STARTUP_GRACE_MS);
  const startupGraceMs = Math.min(
    process.env.CI ? Math.max(requestedGrace, STARTUP_GRACE_CI_FLOOR_MS) : requestedGrace,
    timeout,
  );
  const model = options.model ?? process.env.EVALS_MODEL ?? resolveEvalModel('capture');

  const deadline = startTime + timeout;
  const startedAt = new Date().toISOString();
  let systemPrompt = options.appendSystemPrompt;
  if (options.completionReserveMs !== undefined) {
    const reserve = options.completionReserveMs;
    if (!Number.isFinite(timeout) || timeout <= 0 || !Number.isFinite(reserve) || reserve <= 0 || reserve >= timeout) {
      throw new Error('Section completion reserve must be positive and smaller than the existing work timeout');
    }
    if (!allowedTools.includes('Bash') || (options.tools !== undefined && !options.tools.includes('Bash'))) {
      throw new Error('Section completion clock requires Bash in the declared tools and approval allowlist');
    }
    const notice = `Section completion clock (fixture contract):
Runner entry UTC: ${new Date(startTime).toISOString()}
Hard deadline UTC: ${new Date(deadline).toISOString()}
Completion reserve starts UTC: ${new Date(deadline - reserve).toISOString()}
Setup, CLI startup and API queueing consume this same window; it never resets.
Before source Reads and after each saved checkpoint, use Bash to run exactly \`date -u +%Y-%m-%dT%H:%M:%SZ\`. Compare that observed UTC time with the times above. When remaining time is at most ${reserve / 1000} seconds, prioritize the remaining required completion outputs and verification. No required content or gate may be skipped. If the clock read fails, report timing unavailable; do not invent remaining time or restart the deadline.`;
    systemPrompt = systemPrompt ? `${systemPrompt}\n\n${notice}` : notice;
  }

  // Set up per-run log directory if runId is provided
  let runDir: string | null = null;
  const safeName = testName ? sanitizeTestName(testName) : null;
  if (runId) {
    try {
      runDir = path.join(PROJECT_DIR, 'e2e-runs', runId);
      fs.mkdirSync(runDir, { recursive: true });
    } catch { /* non-fatal */ }
  }

  // Spawn claude -p with streaming NDJSON output. Prompt piped via stdin to
  // avoid shell escaping issues. --verbose is required for stream-json mode.
  const args = [
    '-p',
    '--model', model,
    '--output-format', 'stream-json',
    '--verbose',
    '--dangerously-skip-permissions',
    '--max-turns', String(maxTurns),
    '--allowed-tools', ...allowedTools,
  ];
  if (systemPrompt) args.push('--append-system-prompt', systemPrompt);
  // --allowed-tools controls approval, including when permissions are skipped;
  // only --tools removes unrelated built-ins such as Agent, Bash, and Skill.
  // Keep this opt-in: existing workflow evals intentionally use CLI defaults.
  if (options.tools !== undefined) args.push('--tools', options.tools.join(','));
  if (options.publicStreamDiagnostics) args.push('--include-partial-messages');
  // Hermetic children get zero MCP servers (no --mcp-config is passed).
  // Gated on the same call-time check as the env scrub so EVALS_HERMETIC=0
  // restores operator MCP along with the operator env.
  if (isHermeticEnabled()) args.push('--strict-mcp-config');

  // Spawn claude directly with array-form args (no shell interpolation).
  // node:child_process spawn (not Bun.spawn): `detached` puts the child in
  // its OWN process group, so the timeout handler can killpg the whole tree.
  // Bun.spawn has no detached option, and its bare proc.kill() signalled only
  // claude itself — tool subprocesses claude spawned survived as orphans
  // burning shared API rate for the rest of the shard's lifetime.
  // Prompt is piped via stdin to avoid temp files and shell escaping.
  const childEnv = hermeticChildEnv({ GSTACK_HEADLESS: '1', ...extraEnv });
  signal?.throwIfAborted();
  // Entry, logging, and hermetic setup consume the same startup/work budget.
  if (Date.now() >= Math.min(deadline, startTime + startupGraceMs)) {
    return {
      toolCalls: [], browseErrors: [], exitReason: 'timeout_startup', duration: Date.now() - startTime,
      output: '', transcript: [], model, firstResponseMs: 0, maxInterTurnMs: 0,
      costEstimate: { inputChars: prompt.length, outputChars: 0, estimatedTokens: 0, estimatedCost: 0, turnsUsed: 0 },
    };
  }
  const proc = spawn('claude', args, {
    cwd: workingDirectory,
    // Hermetic by default (see test/helpers/hermetic-env.ts): operator
    // session context (CONDUCTOR_*, CLAUDECODE, ~/.claude config, ~/.gstack)
    // never reaches the child; EVALS_HERMETIC=0 restores the legacy env.
    // Default GSTACK_HEADLESS=1 so eval/E2E runs classify as headless (BLOCK on an
    // AskUserQuestion failure rather than emit a prose question no human reads). A
    // suite exercising the INTERACTIVE prose-fallback path opts out by passing
    // `env: { GSTACK_HEADLESS: '' }` — extraEnv wins because it spreads last.
    env: childEnv,
    stdio: ['pipe', 'pipe', 'pipe'],
    detached: process.platform !== 'win32',
  });
  const stdoutWeb = Readable.toWeb(proc.stdout!) as ReadableStream<Uint8Array>;
  const reader = stdoutWeb.getReader();
  // Exit is independent of descendant-held pipes. Both stream drains and a
  // failed kill that never emits exit have a bounded completion path.
  let exitCode: number | undefined;
  let stdoutDone = false;
  let stderrDone = false;
  let stderrEnded = false;
  let drainExpired = false;
  let streamError: Error | undefined;
  let processError: Error | undefined;
  let drainTimer: ReturnType<typeof setTimeout> | undefined;
  let drainDeadline = Infinity;
  let releaseDrain!: () => void;
  const forcedDrain = new Promise<void>(resolve => { releaseDrain = resolve; });
  let releaseExit!: () => void;
  const procExited = new Promise<void>(resolve => { releaseExit = resolve; });
  let releaseStderr!: () => void;
  const stderrClosed = new Promise<void>(resolve => { releaseStderr = resolve; });

  // Two-phase timeout. Phase 1 (startup): no NDJSON byte yet — a shorter
  // deadline kills a non-answering API run EARLY and names it, instead of
  // the old single timer burning the full work budget to produce an opaque
  // '0 turns / $0.00' failure. Phase 2 (work): armed by the read loop when
  // the FIRST byte arrives, for the REMAINING budget — model work always
  // stays inside timeout; pipe cleanup has a separate bounded allowance.
  let stderr = '';
  let exitReason = 'unknown';
  let timedOut = false;
  let timedOutInStartup = false;
  let phaseTimer: ReturnType<typeof setTimeout>;

  const closePipes = () => {
    reader.cancel().catch(() => { /* already closed */ });
    proc.stdin!.destroy();
    proc.stdout!.destroy();
    proc.stderr!.destroy();
  };
  const expireDrain = () => {
    drainExpired ||= !stdoutDone || !stderrDone;
    killProcessGroup(proc, 'SIGKILL');
    closePipes();
    releaseDrain();
  };
  const armDrain = () => {
    // Exit after cancellation must not restart the five-second allowance.
    drainDeadline = Math.min(drainDeadline, Date.now() + SESSION_DRAIN_GRACE_MS);
    clearTimeout(drainTimer);
    drainTimer = setTimeout(expireDrain, Math.max(0, drainDeadline - Date.now()));
  };

  const killRun = (startupPhase: boolean): void => {
    // Labeling and unblocking are SEPARATE concerns: a timer firing after
    // the child already exited must not relabel a real exit (auth error,
    // crash) as a timeout — but it must STILL group-kill and cancel the
    // reader, or an orphan holding the pipes re-creates the exact
    // blocked-drain hang this runner fixed (an early `return` here was the
    // bug the adversarial pass caught in the first version of this guard).
    if (exitCode === undefined) {
      if (!timedOut) timedOutInStartup = startupPhase;
      timedOut = true;
    } else drainExpired ||= !stdoutDone || !stderrDone;
    // Group SIGKILL (mirrors runShardChild): claude AND every tool
    // subprocess it spawned die together — a bare proc.kill() left orphans
    // that inherited our stdout/stderr pipes and kept the API burning
    // (observed: a 600s timeout stretching past 1400s while an orphan held
    // the pipes open).
    killProcessGroup(proc, 'SIGKILL');
    // Belt and braces with the group kill: even if an orphan survives (EPERM
    // fallback path), cancel() unblocks the read loop below.
    closePipes();
    armDrain();
  };
  const onAbort = () => killRun(false);
  const onExit = (code: number | null, exitSignal: NodeJS.Signals | null) => {
    exitCode = code ?? (exitSignal ? 128 + (os.constants.signals[exitSignal] ?? 0) : 1);
    clearTimeout(phaseTimer);
    armDrain();
    releaseExit();
  };
  const onError = (error: Error) => {
    processError = error;
    stderr += error.message;
    exitCode = 1;
    closePipes();
    releaseExit();
    releaseDrain();
  };
  const onStderr = (chunk: string) => { stderr += chunk; };
  const onStderrDone = () => { stderrDone = true; releaseStderr(); };
  const onStderrEnd = () => { stderrEnded = true; onStderrDone(); };
  const onStderrClose = () => {
    // Close releases the drain, but only end proves stderr reached EOF.
    if (!stderrEnded) streamError ??= new Error('stderr closed before EOF');
    onStderrDone();
  };
  const onStreamError = (error: Error) => { streamError = error; };
  proc.on('exit', onExit);
  proc.on('error', onError);
  proc.stderr!.setEncoding('utf8');
  proc.stderr!.on('data', onStderr);
  proc.stderr!.on('end', onStderrEnd).on('close', onStderrClose).on('error', onStreamError);
  proc.stdout!.on('error', onStreamError);
  signal?.addEventListener('abort', onAbort, { once: true });
  phaseTimer = setTimeout(() => killRun(true), Math.max(0, startTime + startupGraceMs - Date.now()));
  proc.stdin!.on('error', () => { /* exit handling reports early child failure */ });
  if (signal?.aborted || Date.now() >= deadline) onAbort();
  else proc.stdin!.end(prompt);
  /** Called once by the read loop on the first NDJSON byte. */
  const armWorkPhase = (elapsedMs: number): void => {
    clearTimeout(phaseTimer);
    if (exitCode === undefined && !timedOut) {
      phaseTimer = setTimeout(() => killRun(false), Math.max(0, timeout - elapsedMs));
    }
  };

  // Stream NDJSON from stdout for real-time progress
  const collectedLines: string[] = [];
  let liveTurnCount = 0;
  let liveToolCount = 0;
  let firstResponseMs = 0;
  let workPhaseArmed = false;
  let lastToolTime = 0;
  let maxInterTurnMs = 0;
  const decoder = new TextDecoder();
  let buf = '';
  const projectLine = options.publicStreamDiagnostics ? publicStreamProjection(startTime) : (line: string) => line;

  try {
    try {
      while (true) {
        const { done, value } = await reader.read();
        if (done) break;
        buf += decoder.decode(value, { stream: true });
        const lines = buf.split('\n');
        buf = lines.pop() || '';
        for (const rawLine of lines) {
          if (!rawLine.trim()) continue;
          const line = projectLine(rawLine);
          collectedLines.push(line);

          // Track time to first NDJSON line (measures latency from spawn to first Claude response)
          if (!workPhaseArmed) {
            // Flag, not `firstResponseMs === 0`: a first line landing in the
            // same millisecond as spawn would read as "not yet seen" and leave
            // the startup timer live for the whole run (claude adversarial).
            workPhaseArmed = true;
            firstResponseMs = Date.now() - startTime;
            // First byte: startup phase over — arm the work phase for the
            // REMAINING budget (total wall stays <= timeout).
            armWorkPhase(firstResponseMs);
          }

          // Real-time progress to stderr + persistent logs
          try {
            const event = JSON.parse(line);
            if (event.type === 'assistant') {
              liveTurnCount++;
              const content = event.message?.content || [];
              for (const item of content) {
                if (item.type === 'tool_use') {
                  liveToolCount++;
                  const now = Date.now();
                  const elapsed = Math.round((now - startTime) / 1000);
                  // Track inter-turn latency (tool call to tool call)
                  if (lastToolTime > 0) {
                    const interTurn = now - lastToolTime;
                    if (interTurn > maxInterTurnMs) maxInterTurnMs = interTurn;
                  }
                  lastToolTime = now;
                  const progressLine = `  [${elapsed}s] turn ${liveTurnCount} tool #${liveToolCount}: ${item.name}(${truncate(JSON.stringify(item.input || {}), 80)})\n`;
                  process.stderr.write(progressLine);

                  // Persist progress.log
                  if (runDir) {
                    try { fs.appendFileSync(path.join(runDir, 'progress.log'), progressLine); } catch { /* non-fatal */ }
                  }

                  // Write heartbeat (atomic)
                  if (runId && testName) {
                    try {
                      const toolDesc = `${item.name}(${truncate(JSON.stringify(item.input || {}), 60)})`;
                      atomicWriteSync(HEARTBEAT_PATH, JSON.stringify({
                        runId,
                        pid: proc.pid,
                        startedAt,
                        currentTest: testName,
                        status: 'running',
                        turn: liveTurnCount,
                        toolCount: liveToolCount,
                        lastTool: toolDesc,
                        lastToolAt: new Date().toISOString(),
                        elapsedSec: elapsed,
                      }, null, 2) + '\n');
                    } catch { /* non-fatal */ }
                  }
                }
              }
            }
          } catch { /* skip — parseNDJSON will handle it later */ }

          // Append raw NDJSON line to per-test transcript file
          if (runDir && safeName) {
            try { fs.appendFileSync(path.join(runDir, `${safeName}.ndjson`), line + '\n'); } catch { /* non-fatal */ }
          }
        }
      }
    } catch (error) { streamError = error as Error; }
    stdoutDone = true;

    // Flush remaining buffer
    if (buf.trim()) {
      const line = projectLine(buf);
      collectedLines.push(line);
      if (options.publicStreamDiagnostics && runDir && safeName) {
        try { fs.appendFileSync(path.join(runDir, `${safeName}.ndjson`), line + '\n'); } catch { /* non-fatal */ }
      }
    }

    await Promise.race([Promise.all([procExited, stderrClosed]), forcedDrain]);
  } finally {
    clearTimeout(phaseTimer);
    clearTimeout(drainTimer);
    signal?.removeEventListener('abort', onAbort);
    killProcessGroup(proc, 'SIGKILL');
    closePipes();
    proc.removeListener('exit', onExit);
    proc.stderr!.removeListener('data', onStderr);
  }

  if (timedOut) {
    // 'timeout_startup' = the API never sent a byte inside the grace — an
    // availability problem, not a test failure worth reading transcripts
    // for. Distinct so triage (and WS10's inconclusive classification) can
    // key off it without receipts archaeology.
    exitReason = timedOutInStartup ? 'timeout_startup' : 'timeout';
  } else if (exitCode === 0) {
    exitReason = drainExpired ? 'error_output_drain' : streamError ? 'error_output_stream' : 'success';
    if (drainExpired) stderr += `\nOutput drain exceeded ${SESSION_DRAIN_GRACE_MS}ms after exit 0`;
    else if (streamError) stderr += `\nOutput stream failed: ${streamError.message}`;
  } else {
    exitReason = `exit_code_${exitCode ?? 1}`;
  }

  const duration = Date.now() - startTime;

  // Parse all collected NDJSON lines
  const parsed = parseNDJSON(collectedLines);
  const { transcript, resultLine, toolCalls } = parsed;
  const browseErrors: string[] = [];

  // Scan transcript + stderr for browse errors
  const allText = transcript.map(e => JSON.stringify(e)).join('\n') + '\n' + stderr;
  for (const pattern of BROWSE_ERROR_PATTERNS) {
    const match = allText.match(pattern);
    if (match) {
      browseErrors.push(match[0].slice(0, 200));
    }
  }

  // Native Claude uses exit 1 for its structured max-turns result. Preserve
  // that existing semantic outcome only after a complete, uncancelled drain;
  // success-shaped payloads must never override a process/stream failure.
  const maxTurnsExit = exitCode === 1 && resultLine?.subtype === 'error_max_turns'
    && resultLine.is_error === true && !timedOut && !signal?.aborted
    && stdoutDone && stderrDone && !drainExpired && !streamError && !processError;
  if (maxTurnsExit) {
    exitReason = 'error_max_turns';
  } else if (resultLine && exitReason === 'success') {
    if (resultLine.subtype === 'success' && resultLine.is_error) {
      // claude -p can return subtype=success with is_error=true (e.g. API connection failure)
      exitReason = 'error_api';
    } else if (resultLine.subtype === 'success' && exitCode === 0 && !timedOut) {
      exitReason = 'success';
    } else if (resultLine.subtype && resultLine.subtype !== 'success') {
      // Preserve known subtypes like error_max_turns even if is_error is set
      exitReason = resultLine.subtype;
    }
  }

  // Save failure transcript to persistent run directory (or fallback to workingDirectory)
  if (browseErrors.length > 0 || exitReason !== 'success') {
    try {
      const failureDir = runDir || path.join(workingDirectory, '.gstack', 'test-transcripts');
      fs.mkdirSync(failureDir, { recursive: true });
      const failureName = safeName
        ? `${safeName}-failure.json`
        : `e2e-${new Date().toISOString().replace(/[:.]/g, '-')}.json`;
      fs.writeFileSync(
        path.join(failureDir, failureName),
        JSON.stringify({
          prompt: prompt.slice(0, 500),
          testName: testName || 'unknown',
          exitReason,
          browseErrors,
          duration,
          turnAtTimeout: timedOut ? liveTurnCount : undefined,
          lastToolCall: liveToolCount > 0 ? `tool #${liveToolCount}` : undefined,
          stderr: stderr.slice(0, 2000),
          result: resultLine ? { type: resultLine.type, subtype: resultLine.subtype, result: resultLine.result?.slice?.(0, 500) } : null,
        }, null, 2),
      );
    } catch { /* non-fatal */ }
  }

  // Cost from result line (exact) or estimate from chars
  const turnsUsed = resultLine?.num_turns || 0;
  const estimatedCost = resultLine?.total_cost_usd || 0;
  const inputChars = prompt.length;
  const outputChars = (resultLine?.result || '').length;
  const estimatedTokens = (resultLine?.usage?.input_tokens || 0)
    + (resultLine?.usage?.output_tokens || 0)
    + (resultLine?.usage?.cache_read_input_tokens || 0);

  const costEstimate: CostEstimate = {
    inputChars,
    outputChars,
    estimatedTokens,
    estimatedCost: Math.round((estimatedCost) * 100) / 100,
    turnsUsed,
  };

  return { toolCalls, browseErrors, exitReason, duration, output: resultLine?.result || '', costEstimate, transcript, model, firstResponseMs, maxInterTurnMs };
}