juno-code 1.0.47 → 1.0.49

package/dist/templates/scripts/parallel_runner.sh

@@ -0,0 +1,2242 @@
+#!/usr/bin/env python3
+"""
+Run juno-code tasks in parallel with queue management, output extraction, and aggregation.
+
+Modes:
+  Headless (default): ThreadPoolExecutor, output to log files only.
+  Tmux windows:       Each worker = tmux window, coordinator = window 0.
+  Tmux panes:         Workers as split panes, coordinator = top pane.
+
+Usage:
+  # Kanban mode (task IDs as input)
+  ./parallel_runner.sh --kanban TASK1,TASK2,TASK3 [--parallel 3]
+  ./parallel_runner.sh --kanban TASK1 TASK2 TASK3 --parallel 2
+  ./parallel_runner.sh --kanban "TASK1 TASK2 TASK3"
+  ./parallel_runner.sh --kanban T1 T2 --prompt-file instructions.md
+
+  # Generic items mode (any list, requires --prompt-file with {{item}})
+  ./parallel_runner.sh --items "url1,url2,url3" --prompt-file crawl.md
+  ./parallel_runner.sh --items shop1 shop2 shop3 --prompt-file analyze.md
+
+  # File mode (structured data)
+  ./parallel_runner.sh --items-file data.jsonl --prompt-file analyze.md
+  ./parallel_runner.sh --items-file data.csv --strict --file-format csv
+
+  # Common options
+  ./parallel_runner.sh --tmux --kanban T1 T2 T3 --parallel 2
+  ./parallel_runner.sh --tmux panes --kanban T1 T2 --parallel 2
+  ./parallel_runner.sh --tmux --kanban T1 T2 --name my-batch
+  ./parallel_runner.sh -s codex --kanban T1 T2
+  ./parallel_runner.sh -s pi -m gpt-5 --kanban T1 T2
+  ./parallel_runner.sh --stop                    # stop only running session
+  ./parallel_runner.sh --stop --name my-batch    # stop specific session
+  ./parallel_runner.sh --stop-all                # stop all sessions
+
+Input modes (exactly one required, unless --stop/--stop-all):
+  --kanban       Kanban task IDs. {{task_id}} = {{item}} = the ID.
+  --kanban-filter Filter string passed to kanban.sh list. Internally runs
+                 kanban.sh list {filters} -f json --raw and extracts IDs.
+  --items        Generic item list (comma/space separated). Auto-generates item-001 IDs.
+  --items-file   Path to file (JSONL, CSV, TSV, XLSX). Format auto-detected by extension.
+
+File options (for --items-file):
+  --format       Force file format: jsonl, csv, tsv, xlsx (default: auto-detect).
+  --no-header    CSV/TSV/XLSX: treat first row as data, not column headers.
+  --chunk-size   Records per item (default: 1). >1 groups records into a JSON array.
+  --start        First record to process, 1-indexed after header (default: 1).
+  --end          Last record to process, inclusive (default: end of file).
+
+Output extraction:
+  --file-format  Expected output format (e.g., json, csv, md). Sets {{file_format}} placeholder.
+  --strict       Extract response from fenced code block in output.
+                 Writes extracted content to {output_dir}/{task_id}.{file_format}.
+                 Marks task as ERROR if code block not found. Requires --file-format.
+
+Arguments:
+  --parallel     Max concurrent subprocesses (default: 3)
+  -s, --service  Backend service: claude, codex, pi (default: claude). Env: JUNO_SERVICE.
+  -m, --model    Model override. Env: JUNO_MODEL.
+  --env          Environment overrides. KEY=VALUE pairs or path to .env file.
+  --prompt-file  Path to a file whose content is appended to the prompt.
+                 Re-read per task. Placeholders: {{task_id}}, {{item}}, {{file_format}}.
+  --tmux         Run in tmux mode. 'windows' (default) or 'panes' (side-by-side).
+  --name         Session name (default: auto-generated batch-N). Tmux session = pc-{name}.
+  --output-dir   Structured output directory. Default: /tmp/juno-code-sessions/{date}/{run_id}.
+  --stop         Stop a session. Uses --name if provided, otherwise auto-detects.
+  --stop-all     Stop ALL running sessions.
+"""
+
+import argparse
+import csv
+import io
+import json
+import os
+import random
+import re
+import shlex
+import shutil
+import signal
+import string
+import subprocess
+import sys
+import textwrap
+import threading
+import time
+import uuid
+from collections import deque
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from dataclasses import dataclass
+from datetime import datetime
+from enum import Enum
+from pathlib import Path
+
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+# _log_base is the flat root for discoverable session files (PID, pause, dashboard).
+# LOG_DIR (computed at startup in main()) nests under _log_base/{date}/{run_id}/
+# for per-task logs and the combined log — isolating concurrent/repeated runs.
+_log_base = SCRIPT_DIR / "logs"
+LOG_DIR = _log_base  # overwritten in main() with run-ID path
+COMBINED_LOG = LOG_DIR / "parallel_runner.log"  # overwritten in main()
+
+# 5-char alphanumeric run ID, generated once at startup in main()
+_run_id = ""
+
+# Thread-safe lock for writing to the shared combined log
+_log_lock = threading.Lock()
+
+# Shared counters for remaining-task tracking
+_completed_count = 0
+_completed_lock = threading.Lock()
+_total_tasks = 0
+
+# Per-task elapsed times
+_task_times = {}
+_task_times_lock = threading.Lock()
+
+# Shutdown flag for graceful exit (set by signal handlers)
+_shutdown_event = threading.Event()
+
+# --- Color system for task identification ---
+# ANSI 256-color codes picked for high contrast between consecutive colors,
+# visibility on both dark and light terminal backgrounds.
+_TASK_COLORS = [
+    196,  # red
+    39,   # dodger blue
+    208,  # orange
+    35,   # cyan-green
+    201,  # magenta/pink
+    220,  # gold
+    27,   # blue
+    118,  # bright green
+    163,  # rose
+    45,   # turquoise
+    214,  # dark orange
+    99,   # purple
+    82,   # lime
+    197,  # hot pink
+    33,   # royal blue
+    215,  # sandy orange
+    48,   # sea green
+    135,  # medium purple
+    226,  # yellow
+    69,   # cornflower blue
+]
+_RESET = "\033[0m"
+
+# --- Service / model defaults ---
+_VALID_SERVICES = ("claude", "codex", "pi")
+_DEFAULT_SERVICE = "claude"
+_SERVICE_DEFAULT_MODEL = {
+    "claude": ":sonnet",
+    "codex": ":codex",
+    "pi": "openai-codex/gpt-5.3-codex",
+}
+
+# Resolved environment overrides from --env args (populated at parse time)
+_env_overrides = {}
+
+# Map task_id -> color (assigned at startup)
+_task_color_map = {}
+
+# Map task_id -> item value (the full input data for each task)
+# For --kanban: item == task_id. For --items/--items-file: item is the real data.
+_item_map = {}
+
+# Python helper script piped via tmux pipe-pane to format output into log files.
+# argv: task_id, task_log_path, combined_log_path, ansi_color_code
+# Combined log gets colored dot; task log stays plain.
+_LOG_PIPE_SCRIPT = r"""#!/usr/bin/env python3
+import re, sys
+from datetime import datetime
+
+task_id = sys.argv[1]
+task_log_path = sys.argv[2]
+combined_log_path = sys.argv[3]
+color_code = sys.argv[4] if len(sys.argv) > 4 else '7'
+ansi_re = re.compile(r'\x1b\[[0-9;]*[a-zA-Z]')
+dot = '\033[38;5;%sm\u25cf\033[0m' % color_code  # colored dot
+
+with open(task_log_path, 'w') as tl, open(combined_log_path, 'a') as cl:
+    while True:
+        line = sys.stdin.readline()
+        if not line:
+            break
+        line = line.rstrip('\n\r')
+        if not line:
+            continue
+        ts = datetime.now().strftime('%H:%M:%S')
+        clean = ansi_re.sub('', line)
+        tl.write('[%s] [%s] %s\n' % (ts, task_id, clean))
+        tl.flush()
+        cl.write('[%s] %s [%s] %s\n' % (ts, dot, task_id, line))
+        cl.flush()
+"""
+
+
+# ---------------------------------------------------------------------------
+# Per-session file helpers
+# ---------------------------------------------------------------------------
+
+def _session_name_to_tmux(name):
+    """Convert session name to tmux session name."""
+    return f"pc-{name}"
+
+
+def _dashboard_file(name):
+    return _log_base / f".dashboard_{name}"
+
+
+def _pause_file(name):
+    return _log_base / f".pause_{name}"
+
+
+def _pid_file(name):
+    return _log_base / f".orchestrator_pid_{name}"
+
+
+def _orchestrator_log(name):
+    return LOG_DIR / f"orchestrator_{name}.log"
+
+
+def _tmp_dir(name):
+    return Path(f"/tmp/pc-{name}")
+
+
+def _write_log_pipe_helper(name):
+    """Write the Python log-pipe helper to /tmp (once per session)."""
+    tmp = _tmp_dir(name)
+    tmp.mkdir(parents=True, exist_ok=True)
+    helper_path = tmp / "log_pipe.py"
+    helper_path.write_text(_LOG_PIPE_SCRIPT)
+    return str(helper_path)
+
+
+def _resolve_service_model(args):
+    """Resolve service and model from CLI args > env vars > defaults."""
+    service = getattr(args, "service", None) or os.environ.get("JUNO_SERVICE") or _DEFAULT_SERVICE
+    service = service.lower()
+    if service not in _VALID_SERVICES:
+        print(f"ERROR: Invalid service '{service}'. Must be one of: {', '.join(_VALID_SERVICES)}",
+              file=sys.stderr)
+        sys.exit(2)
+    model = getattr(args, "model", None) or os.environ.get("JUNO_MODEL") or _SERVICE_DEFAULT_MODEL[service]
+    return service, model
+
+
+def _parse_env_file(path):
+    """Parse a .env file into a dict. Supports KEY=VALUE, comments (#), empty lines."""
+    env = {}
+    for line_num, raw_line in enumerate(
+        Path(path).read_text(encoding="utf-8").splitlines(), 1
+    ):
+        line = raw_line.strip()
+        if not line or line.startswith("#"):
+            continue
+        if "=" not in line:
+            print(f"WARNING: Skipping invalid line {line_num} in {path}: {line}",
+                  file=sys.stderr)
+            continue
+        key, _, value = line.partition("=")
+        key = key.strip()
+        value = value.strip()
+        if len(value) >= 2 and value[0] == value[-1] and value[0] in ('"', "'"):
+            value = value[1:-1]
+        env[key] = value
+    return env
+
+
+def _resolve_env_overrides(env_args):
+    """Resolve --env arguments into a merged dict."""
+    overrides = {}
+    if not env_args:
+        return overrides
+    for arg in env_args:
+        if "=" in arg:
+            key_part = arg.split("=", 1)[0]
+            if Path(key_part).exists() and not key_part.replace("_", "").replace("-", "").isalnum():
+                overrides.update(_parse_env_file(arg))
+            else:
+                key, _, value = arg.partition("=")
+                overrides[key.strip()] = value.strip()
+        else:
+            path_obj = Path(arg)
+            if not path_obj.exists():
+                print(f"ERROR: Env file not found: {arg}", file=sys.stderr)
+                sys.exit(2)
+            overrides.update(_parse_env_file(arg))
+    return overrides
+
+
+def _build_process_env(extra_capture_env=None):
+    """Build the environment dict for a child process."""
+    env = os.environ.copy()
+    env.update(_env_overrides)
+    if extra_capture_env:
+        env.update(extra_capture_env)
+    return env
+
+
+def _generate_env_exports():
+    """Generate shell export lines for the full process environment."""
+    merged = os.environ.copy()
+    merged.update(_env_overrides)
+    lines = []
+    for key, value in sorted(merged.items()):
+        if key in ("TERM_SESSION_ID", "TMUX", "TMUX_PANE", "STY", "WINDOW",
+                    "SHLVL", "OLDPWD", "_"):
+            continue
+        lines.append(f"export {key}={shlex.quote(value)}")
+    return "\n".join(lines)
+
+
+# ---------------------------------------------------------------------------
+# File parsing — multi-format pipeline
+# ---------------------------------------------------------------------------
+
+_FORMAT_EXTENSIONS = {
+    ".jsonl": "jsonl",
+    ".ndjson": "jsonl",
+    ".csv": "csv",
+    ".tsv": "tsv",
+    ".xlsx": "xlsx",
+}
+_VALID_FORMATS = ("jsonl", "csv", "tsv", "xlsx")
+
+
+def _detect_format(path, format_override=None):
+    """Detect file format from extension or --format override."""
+    if format_override:
+        fmt = format_override.lower()
+        if fmt not in _VALID_FORMATS:
+            print(f"ERROR: Unknown format '{fmt}'. Must be one of: {', '.join(_VALID_FORMATS)}",
+                  file=sys.stderr)
+            sys.exit(2)
+        return fmt
+    ext = Path(path).suffix.lower()
+    fmt = _FORMAT_EXTENSIONS.get(ext)
+    if not fmt:
+        print(f"ERROR: Cannot detect format from extension '{ext}'. "
+              f"Use --format to specify one of: {', '.join(_VALID_FORMATS)}",
+              file=sys.stderr)
+        sys.exit(2)
+    return fmt
+
+
+def _record_to_item(record):
+    """Convert a parsed record to its {{item}} string representation."""
+    if isinstance(record, str):
+        return record
+    return json.dumps(record, ensure_ascii=False)
+
+
+def _parse_jsonl(path):
+    """Parse JSONL/NDJSON into a list of parsed values."""
+    records = []
+    for line_num, raw_line in enumerate(
+        Path(path).read_text(encoding="utf-8").splitlines(), 1
+    ):
+        line = raw_line.strip()
+        if not line or line.startswith("#"):
+            continue
+        try:
+            records.append(json.loads(line))
+        except json.JSONDecodeError as e:
+            print(f"ERROR: Invalid JSON at {path}:{line_num}: {e}", file=sys.stderr)
+            sys.exit(2)
+    return records
+
+
+def _parse_csv(path, has_header=True, delimiter=","):
+    """Parse CSV/TSV into a list of records (dicts if header, lists otherwise)."""
+    text = Path(path).read_text(encoding="utf-8")
+    reader = csv.reader(io.StringIO(text), delimiter=delimiter)
+    rows = list(reader)
+    if not rows:
+        return []
+    if has_header:
+        headers = rows[0]
+        return [dict(zip(headers, row)) for row in rows[1:]]
+    return [row for row in rows]
+
+
+def _parse_xlsx(path, has_header=True):
+    """Parse XLSX into a list of records (dicts if header, lists otherwise)."""
+    try:
+        import openpyxl
+    except ImportError:
+        print("ERROR: openpyxl is required for XLSX files. Install with:\n"
+              "  pip install openpyxl", file=sys.stderr)
+        sys.exit(2)
+    wb = openpyxl.load_workbook(path, read_only=True, data_only=True)
+    ws = wb.active
+    rows = []
+    for row in ws.iter_rows(values_only=True):
+        rows.append([cell if cell is not None else "" for cell in row])
+    wb.close()
+    if not rows:
+        return []
+    if has_header:
+        headers = [str(h) for h in rows[0]]
+        return [dict(zip(headers, row)) for row in rows[1:]]
+    return rows
+
+
+def _parse_items_file(path, format_override=None, has_header=True,
+                      start=None, end=None, chunk_size=1):
+    """Full pipeline: parse -> slice -> chunk -> return list of item strings."""
+    file_path = Path(path)
+    if not file_path.exists():
+        print(f"ERROR: Items file not found: {path}", file=sys.stderr)
+        sys.exit(2)
+
+    fmt = _detect_format(path, format_override)
+
+    if fmt == "jsonl":
+        records = _parse_jsonl(path)
+    elif fmt == "csv":
+        records = _parse_csv(path, has_header=has_header, delimiter=",")
+    elif fmt == "tsv":
+        records = _parse_csv(path, has_header=has_header, delimiter="\t")
+    elif fmt == "xlsx":
+        records = _parse_xlsx(path, has_header=has_header)
+    else:
+        print(f"ERROR: Unsupported format '{fmt}'", file=sys.stderr)
+        sys.exit(2)
+
+    if not records:
+        print(f"ERROR: No records found in {path}", file=sys.stderr)
+        sys.exit(2)
+
+    # Slice: --start and --end (1-indexed, inclusive, after header)
+    start_idx = (start - 1) if start and start >= 1 else 0
+    end_idx = end if end else len(records)
+    records = records[start_idx:end_idx]
+
+    if not records:
+        print(f"ERROR: No records in range --start {start} --end {end} "
+              f"(file has {end_idx} data rows)", file=sys.stderr)
+        sys.exit(2)
+
+    # Chunk: group records into chunks of --chunk-size
+    items = []
+    for i in range(0, len(records), chunk_size):
+        chunk = records[i:i + chunk_size]
+        if chunk_size == 1:
+            items.append(_record_to_item(chunk[0]))
+        else:
+            items.append(json.dumps(chunk, ensure_ascii=False))
+
+    return items
+
+
+# ---------------------------------------------------------------------------
+# Input resolution
+# ---------------------------------------------------------------------------
+
+def _resolve_input(args):
+    """Resolve input mode and return task_ids list. Populates _item_map."""
+    global _item_map
+
+    modes = sum([
+        bool(args.kanban),
+        bool(args.items),
+        bool(args.items_file),
+    ])
+    if modes == 0:
+        print("ERROR: One of --kanban, --items, or --items-file is required "
+              "(unless using --stop or --stop-all)", file=sys.stderr)
+        sys.exit(2)
+    if modes > 1:
+        print("ERROR: Only one of --kanban, --items, or --items-file can be used",
+              file=sys.stderr)
+        sys.exit(2)
+
+    if args.kanban:
+        task_ids = args.kanban
+        _item_map = {tid: tid for tid in task_ids}
+        return task_ids
+
+    # Generic items — auto-generate IDs
+    if args.items:
+        raw_items = []
+        for val in args.items:
+            for part in val.replace(",", " ").split():
+                part = part.strip()
+                if part:
+                    raw_items.append(part)
+    else:
+        raw_items = _parse_items_file(
+            args.items_file,
+            format_override=getattr(args, "format", None),
+            has_header=not getattr(args, "no_header", False),
+            start=getattr(args, "start", None),
+            end=getattr(args, "end", None),
+            chunk_size=getattr(args, "chunk_size", 1) or 1,
+        )
+
+    if not raw_items:
+        print("ERROR: No items found in input", file=sys.stderr)
+        sys.exit(2)
+
+    width = max(len(str(len(raw_items))), 3)
+    task_ids = []
+    for i, item_value in enumerate(raw_items, 1):
+        task_id = f"item-{i:0{width}d}"
+        task_ids.append(task_id)
+        _item_map[task_id] = item_value
+
+    return task_ids
+
+
+def _color_for(task_id):
+    """Get the ANSI colored dot prefix for a task_id."""
+    code = _task_color_map.get(task_id, 7)
+    return f"\033[38;5;{code}m\u25cf{_RESET}"
+
+
+def _colored_tag(task_id):
+    """Return a colored dot + [task_id] string for log lines."""
+    return f"{_color_for(task_id)} [{task_id}]"
+
+
+# ---------------------------------------------------------------------------
+# Structured output capture
+# ---------------------------------------------------------------------------
+
+def _resolve_output_dir(args):
+    """Resolve the output directory for structured task output."""
+    if args.output_dir:
+        output_dir = Path(args.output_dir)
+    elif os.environ.get("JUNO_OUTPUT_DIR"):
+        output_dir = Path(os.environ["JUNO_OUTPUT_DIR"])
+    else:
+        today = datetime.now().strftime("%Y-%m-%d")
+        output_dir = Path(f"/tmp/juno-code-sessions/{today}")
+    if _run_id:
+        output_dir = output_dir / _run_id
+    output_dir.mkdir(parents=True, exist_ok=True)
+    return output_dir
+
+
+def _task_output_path(output_dir, task_id):
+    """Path for the per-task structured output JSON."""
+    return output_dir / f"{task_id}.json"
+
+
+def _parse_result_from_log(task_log_path):
+    """Parse the juno-code result event from a task log file.
+
+    juno-code prints a JSON line with {"type":"result",...} to stdout.
+    We walk backwards to find it near the end.
+    """
+    try:
+        lines = Path(task_log_path).read_text(encoding="utf-8").splitlines()
+    except (OSError, UnicodeDecodeError):
+        return None
+
+    for line in reversed(lines):
+        idx = line.find('{"type":')
+        if idx == -1:
+            continue
+        candidate = line[idx:]
+        try:
+            parsed = json.loads(candidate)
+            if isinstance(parsed, dict) and parsed.get("type") == "result":
+                return parsed
+        except json.JSONDecodeError:
+            continue
+    return None
+
+
+def _extract_response(backend_result, file_format):
+    """Extract the clean response from a backend result.
+
+    Tries fenced code blocks first, falls back to raw result.
+    """
+    if not isinstance(backend_result, dict):
+        return None, None
+
+    raw_result = backend_result.get("result")
+    if not raw_result:
+        return None, None
+
+    if file_format:
+        content, _ = _extract_from_fenced_block(raw_result, file_format)
+        if content is not None:
+            return content, None
+
+    any_block = re.compile(r"```\w*\s*\n(.*?)```", re.DOTALL)
+    matches = list(any_block.finditer(raw_result))
+    if matches:
+        return matches[-1].group(1).rstrip(), None
+
+    error_msg = (f"No ```{file_format} code block found (used raw result)"
+                 if file_format else None)
+    return raw_result, error_msg
+
+
+def _write_task_output(output_dir, task_id, exit_code, wall_time, start_time,
+                       end_time, worker_id=-1,
+                       extracted_response=None, extraction_error=None,
+                       backend_result=None, file_format=""):
+    """Write per-task structured JSON output."""
+    output_path = _task_output_path(output_dir, task_id)
+
+    if extracted_response is None and backend_result is not None:
+        extracted_response, extraction_error = _extract_response(
+            backend_result, file_format,
+        )
+
+    session_id = None
+    if isinstance(backend_result, dict):
+        session_id = backend_result.get("session_id")
+
+    task_output = {
+        "task_id": task_id,
+        "session_id": session_id,
+        "exit_code": exit_code,
+        "wall_time_seconds": round(wall_time, 2),
+        "start_time": start_time,
+        "end_time": end_time,
+        "worker_id": worker_id,
+        "backend_result": backend_result,
+        "extracted_response": extracted_response,
+        "extraction_error": extraction_error,
+    }
+
+    output_path.write_text(
+        json.dumps(task_output, indent=2, ensure_ascii=False),
+        encoding="utf-8",
+    )
+    return task_output
+
+
+def _extract_from_fenced_block(log_text, file_format):
+    """Extract content from the last fenced code block matching the format."""
+    pattern = re.compile(
+        r"```" + re.escape(file_format) + r"\s*\n(.*?)```",
+        re.DOTALL,
+    )
+    matches = list(pattern.finditer(log_text))
+    if not matches:
+        return None, f"No ```{file_format} code block found in task output"
+    content = matches[-1].group(1)
+    content = content.rstrip()
+    return content, None
+
+
+def _extract_strict_output(task_id, task_log_path, output_dir, file_format, exit_code):
+    """Run strict extraction on a completed task's log."""
+    try:
+        log_text = Path(task_log_path).read_text(encoding="utf-8")
+    except (OSError, UnicodeDecodeError) as e:
+        _update_task_json(output_dir, task_id, None, f"Could not read task log: {e}")
+        return 1
+
+    content, error = _extract_from_fenced_block(log_text, file_format)
+
+    if error:
+        _update_task_json(output_dir, task_id, None, error)
+        return 1
+
+    extracted_path = output_dir / f"{task_id}.{file_format}"
+    extracted_path.write_text(content, encoding="utf-8")
+    _update_task_json(output_dir, task_id, content, None)
+
+    return exit_code
+
+
+def _update_task_json(output_dir, task_id, extracted_response, extraction_error):
+    """Update an existing per-task JSON file with extraction results."""
+    output_path = _task_output_path(output_dir, task_id)
+    if not output_path.exists():
+        return
+    try:
+        data = json.loads(output_path.read_text(encoding="utf-8"))
+    except (json.JSONDecodeError, OSError):
+        return
+    data["extracted_response"] = extracted_response
+    data["extraction_error"] = extraction_error
+    if extraction_error:
+        data["exit_code"] = 1
+    output_path.write_text(
+        json.dumps(data, indent=2, ensure_ascii=False),
+        encoding="utf-8",
+    )
+
+
+def _write_aggregation(output_dir, task_outputs, wall_time, parallelism,
+                       mode="headless", session_name=None, file_format=""):
+    """Build and write the aggregation file."""
+    succeeded = sum(1 for t in task_outputs.values() if t["exit_code"] == 0)
+    failed = sum(1 for t in task_outputs.values() if t["exit_code"] != 0)
+
+    merged_parts = []
+    failed_ids = []
+    failed_sessions = {}
+    for tid in sorted(task_outputs.keys()):
+        t = task_outputs[tid]
+        er = t.get("extracted_response")
+        br = t.get("backend_result") or {}
+        backend_ok = isinstance(br, dict) and br.get("exit_code", -1) == 0
+        if er and (t.get("exit_code") == 0 or backend_ok):
+            if file_format == "csv" and merged_parts:
+                lines = er.split("\n")
+                if len(lines) > 1:
+                    merged_parts.append("\n".join(lines[1:]))
+            else:
+                merged_parts.append(er)
+        else:
+            failed_ids.append(tid)
+            sid = t.get("session_id")
+            if sid:
+                failed_sessions[tid] = sid
+
+    separator = "\n" if file_format in ("csv", "tsv") else "\n\n"
+    merged_extracted = separator.join(merged_parts) if merged_parts else None
+
+    aggregation = {
+        "meta": {
+            "created_at": datetime.now().isoformat(),
+            "run_id": _run_id,
+            "total_tasks": len(task_outputs),
+            "succeeded": succeeded,
+            "failed": failed,
+            "wall_time_seconds": round(wall_time, 2),
+            "parallelism": parallelism,
+            "mode": mode,
+            "session_name": session_name,
+        },
+        "merged_extracted": merged_extracted,
+        "tasks": task_outputs,
+    }
+
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    agg_path = output_dir / f"aggregation_{timestamp}.json"
+    agg_path.write_text(
+        json.dumps(aggregation, indent=2, ensure_ascii=False),
+        encoding="utf-8",
+    )
+
+    merged_path = None
+    if merged_extracted and file_format:
+        merged_path = output_dir / f"merged_{timestamp}.{file_format}"
+        merged_path.write_text(merged_extracted, encoding="utf-8")
+
+    return {
+        "agg_path": str(agg_path),
+        "merged_path": str(merged_path) if merged_path else None,
+        "failed_ids": failed_ids,
+        "failed_sessions": failed_sessions,
+        "error_count": len(failed_ids),
+    }
+
+
+def _format_output_summary(agg_result):
+    """Format output paths and errors into a multi-line summary string."""
+    lines = []
+    lines.append("=" * 60)
+    lines.append("OUTPUT FILES")
+    lines.append("=" * 60)
+    lines.append(f"  Run ID:       {_run_id}")
+    lines.append(f"  Aggregation:  {agg_result['agg_path']}")
+    if agg_result["merged_path"]:
+        lines.append(f"  Merged file:  {agg_result['merged_path']}")
+    if agg_result["error_count"] > 0:
+        lines.append(f"  Errors:       {agg_result['error_count']} chunks failed extraction")
+        lines.append(f"  Failed IDs:   {', '.join(agg_result['failed_ids'])}")
+        failed_sessions = agg_result.get("failed_sessions", {})
+        if failed_sessions:
+            lines.append("  Sessions to investigate:")
+            for tid, sid in failed_sessions.items():
+                lines.append(f"    {tid}: {sid}")
+        no_session = [tid for tid in agg_result["failed_ids"] if tid not in failed_sessions]
+        if no_session:
+            lines.append(f"  No session ID: {', '.join(no_session)}")
+    else:
+        lines.append("  Errors:       0")
+    lines.append("=" * 60)
+    return "\n".join(lines)
+
+
+def _print_output_summary(agg_result):
+    """Print output summary to combined log."""
+    for line in _format_output_summary(agg_result).splitlines():
+        log_combined(line)
+
+
+# ---------------------------------------------------------------------------
+# Auto-naming
+# ---------------------------------------------------------------------------
+
+def _next_batch_name():
+    """Find the next available batch-N name."""
+    existing = set()
+    try:
+        result = subprocess.run(
+            ["tmux", "list-sessions", "-F", "#{session_name}"],
+            capture_output=True, text=True,
+        )
+        for line in result.stdout.strip().splitlines():
+            line = line.strip()
+            if line.startswith("pc-batch-"):
+                try:
+                    n = int(line[len("pc-batch-"):])
+                    existing.add(n)
+                except ValueError:
+                    pass
+    except Exception:
+        pass
+    try:
+        for f in _log_base.glob(".orchestrator_pid_batch-*"):
+            suffix = f.name[len(".orchestrator_pid_batch-"):]
+            try:
+                existing.add(int(suffix))
+            except ValueError:
+                pass
+    except Exception:
+        pass
+    counter = 1
+    while counter in existing:
+        counter += 1
+    return f"batch-{counter}"
+
+
+# ---------------------------------------------------------------------------
+# Stop commands
+# ---------------------------------------------------------------------------
+
+def _list_running_sessions():
+    """Return list of (name, pid) tuples for running sessions."""
+    sessions = []
+    try:
+        for f in _log_base.glob(".orchestrator_pid_*"):
+            name = f.name[len(".orchestrator_pid_"):]
+            try:
+                pid = int(f.read_text().strip())
+                os.kill(pid, 0)
+                sessions.append((name, pid))
+            except (ValueError, ProcessLookupError, PermissionError):
+                pass
+    except Exception:
+        pass
+    return sessions
+
+
+def _stop_session(name):
+    """Stop a single session by name."""
+    stopped = False
+    pid_path = _pid_file(name)
+    tmux_session = _session_name_to_tmux(name)
+
+    if pid_path.exists():
+        try:
+            pid = int(pid_path.read_text().strip())
+            os.kill(pid, signal.SIGTERM)
+            print(f"  Sent SIGTERM to orchestrator (PID {pid})")
+            stopped = True
+        except (ValueError, ProcessLookupError, PermissionError):
+            pass
+
+    result = subprocess.run(
+        ["tmux", "kill-session", "-t", tmux_session],
+        capture_output=True, text=True,
+    )
+    if result.returncode == 0:
+        print(f"  Killed tmux session '{tmux_session}'")
+        stopped = True
+
+    for f in [pid_path, _dashboard_file(name), _pause_file(name)]:
+        try:
+            f.unlink(missing_ok=True)
+        except OSError:
+            pass
+
+    tmp = _tmp_dir(name)
+    if tmp.exists():
+        shutil.rmtree(str(tmp), ignore_errors=True)
+
+    return stopped
+
+
+def run_stop(args):
+    """Handle --stop command."""
+    if args.name:
+        name = args.name
+        print(f"Stopping session '{name}'...")
+        if _stop_session(name):
+            print(f"Session '{name}' stopped.")
+        else:
+            print(f"No running session found with name '{name}'.")
+        return
+
+    sessions = _list_running_sessions()
+    if len(sessions) == 0:
+        print("No running sessions found.")
+        return
+    if len(sessions) == 1:
+        name = sessions[0][0]
+        print(f"Stopping session '{name}'...")
+        _stop_session(name)
+        print(f"Session '{name}' stopped.")
+        return
+
+    print(f"Multiple sessions running ({len(sessions)}). Specify --name:")
+    for name, pid in sessions:
+        print(f"  --name {name}  (PID {pid}, tmux: pc-{name})")
+    sys.exit(1)
+
+
+def run_stop_all():
+    """Handle --stop-all command."""
+    sessions = _list_running_sessions()
+    session_names = set(s[0] for s in sessions)
+    try:
+        result = subprocess.run(
+            ["tmux", "list-sessions", "-F", "#{session_name}"],
+            capture_output=True, text=True,
+        )
+        for line in result.stdout.strip().splitlines():
+            line = line.strip()
+            if line.startswith("pc-"):
+                name = line[3:]
+                session_names.add(name)
+    except Exception:
+        pass
+
+    if not session_names:
+        print("No running sessions found.")
+        return
+
+    print(f"Stopping {len(session_names)} session(s)...")
+    for name in sorted(session_names):
+        print(f"\n  [{name}]")
+        _stop_session(name)
+    print(f"\nAll sessions stopped.")
+
+
+# ---------------------------------------------------------------------------
+# CLI parsing
+# ---------------------------------------------------------------------------
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="Run juno-code tasks in parallel with queue management and output extraction")
+
+    # --- Input modes ---
+    input_group = parser.add_argument_group("input modes (exactly one required unless --stop/--stop-all)")
+    input_group.add_argument(
+        "--kanban", nargs="+",
+        help="Kanban task IDs (comma/space/quoted). {{task_id}} = {{item}} = the ID.",
+    )
+    input_group.add_argument(
+        "--kanban-filter", type=str, default=None,
+        help="Filter string passed to kanban.sh list. Internally runs "
+             "kanban.sh list {filters} -f json --raw and extracts IDs.",
+    )
+    input_group.add_argument(
+        "--items", nargs="+",
+        help="Generic item list (comma/space/quoted). Auto-generates item-NNN IDs.",
+    )
+    input_group.add_argument(
+        "--items-file", type=str, default=None,
+        help="Path to file (JSONL, CSV, TSV, XLSX). Format auto-detected by extension.",
+    )
+
+    # --- File options ---
+    file_group = parser.add_argument_group("file options (for --items-file)")
+    file_group.add_argument(
+        "--format", type=str, default=None, choices=["jsonl", "csv", "tsv", "xlsx"],
+        help="Force file format (default: auto-detect by extension).",
+    )
+    file_group.add_argument(
+        "--no-header", action="store_true", default=False,
+        help="CSV/TSV/XLSX: treat first row as data, not column headers.",
+    )
+    file_group.add_argument(
+        "--chunk-size", type=int, default=1,
+        help="Records per item (default: 1). >1 groups records into a JSON array.",
+    )
+    file_group.add_argument(
+        "--start", type=int, default=None,
+        help="First record to process, 1-indexed after header (default: 1).",
+    )
+    file_group.add_argument(
+        "--end", type=int, default=None,
+        help="Last record to process, inclusive (default: end of file).",
+    )
+
+    # --- Output extraction ---
+    extract_group = parser.add_argument_group("output extraction")
+    extract_group.add_argument(
+        "--file-format", type=str, default=None,
+        help="Expected output format (e.g., json, csv, md). Sets {{file_format}} placeholder.",
+    )
+    extract_group.add_argument(
+        "--strict", action="store_true", default=False,
+        help="Extract response from fenced code block. Requires --file-format.",
+    )
+
+    # --- Execution options ---
+    parser.add_argument(
+        "--parallel", type=int, default=3,
+        help="Max concurrent subprocesses (default: 3)",
+    )
+    parser.add_argument(
+        "-s", "--service", type=str, default=None, choices=["claude", "codex", "pi"],
+        help="Backend service (default: claude). Env: JUNO_SERVICE.",
+    )
+    parser.add_argument(
+        "-m", "--model", type=str, default=None,
+        help="Model override. Env: JUNO_MODEL.",
+    )
+    parser.add_argument(
+        "--env", nargs="+", default=None,
+        help="Environment overrides. KEY=VALUE or .env file path.",
+    )
+    parser.add_argument(
+        "--prompt-file", type=str, default=None,
+        help="Prompt template file. Re-read per task. Placeholders: {{task_id}}, {{item}}, {{file_format}}.",
+    )
+    parser.add_argument(
+        "--tmux", nargs="?", const="windows", default=None, choices=["windows", "panes"],
+        help="Run in tmux mode. 'windows' (default) or 'panes'.",
+    )
+    parser.add_argument(
+        "--name", type=str, default=None,
+        help="Session name (default: auto-generated batch-N). Tmux session = pc-{name}.",
+    )
+    parser.add_argument(
+        "--output-dir", type=str, default=None,
+        help="Structured output directory. Default: /tmp/juno-code-sessions/{date}/{run_id}.",
+    )
+    parser.add_argument(
+        "--stop", action="store_true", default=False,
+        help="Stop a session. Uses --name if provided, otherwise auto-detects.",
+    )
+    parser.add_argument(
+        "--stop-all", action="store_true", default=False,
+        help="Stop ALL running sessions.",
+    )
+    args = parser.parse_args()
+
+    # Handle stop commands first
+    if args.stop_all:
+        return args
+    if args.stop:
+        return args
+
+    # Resolve --kanban-filter -> --kanban
+    if args.kanban_filter:
+        if args.kanban:
+            parser.error("Cannot use --kanban-filter together with --kanban")
+        kanban_script = SCRIPT_DIR / "kanban.sh"
+        if not kanban_script.exists():
+            parser.error(f"Kanban script not found: {kanban_script}")
+        filter_args = shlex.split(args.kanban_filter)
+        cmd = [str(kanban_script), "list"] + filter_args + ["-f", "json", "--raw"]
+        print(f"Running kanban filter: {' '.join(cmd)}", file=sys.stderr)
+        result = subprocess.run(cmd, capture_output=True, text=True, timeout=30)
+        if result.returncode != 0:
+            parser.error(f"kanban.sh list failed (exit {result.returncode}):\n{result.stderr.strip()}")
+        first_line = result.stdout.strip().split("\n")[0] if result.stdout.strip() else "[]"
+        try:
+            tasks = json.loads(first_line)
+        except json.JSONDecodeError as e:
+            parser.error(f"Failed to parse kanban output as JSON: {e}\nOutput: {first_line[:200]}")
+        if not isinstance(tasks, list):
+            parser.error(f"Expected JSON array from kanban output, got {type(tasks).__name__}")
+        ids = [t["id"] for t in tasks if isinstance(t, dict) and "id" in t]
+        if not ids:
+            parser.error("kanban-filter returned 0 tasks. Check your filters.")
+        print(f"kanban-filter resolved {len(ids)} task(s): {', '.join(ids[:10])}"
+              f"{'...' if len(ids) > 10 else ''}", file=sys.stderr)
+        args.kanban = ids
+
+    # Auto-infer --file-format from --items-file extension
+    if not args.file_format and args.items_file:
+        ext = Path(args.items_file).suffix.lstrip(".").lower()
+        if ext in ("csv", "tsv", "json", "jsonl", "ndjson", "md", "txt", "xlsx"):
+            infer_map = {"xlsx": "csv", "ndjson": "jsonl"}
+            args.file_format = infer_map.get(ext, ext)
+            print(f"Auto-inferred --file-format={args.file_format} from input file extension",
+                  file=sys.stderr)
+
+    if args.strict and not args.file_format:
+        parser.error("--strict requires --file-format to be set")
+
+    global _env_overrides
+    _env_overrides = _resolve_env_overrides(args.env)
+
+    # Flatten --kanban
+    if args.kanban:
+        flat = []
+        for val in args.kanban:
+            for part in val.replace(",", " ").split():
+                part = part.strip()
+                if part:
+                    flat.append(part)
+        args.kanban = flat
+
+    task_ids = _resolve_input(args)
+    args.kanban = task_ids
+    return args
+
+
+# ---------------------------------------------------------------------------
+# Shared helpers
+# ---------------------------------------------------------------------------
+
+def log_combined(msg, task_id=None):
+    """Write a timestamped line to both stdout and the combined log (thread-safe)."""
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    plain_line = f"[{timestamp}] {msg}"
+    if task_id:
+        colored_line = f"[{timestamp}] {_colored_tag(task_id)} {msg}"
+    else:
+        colored_line = plain_line
+    with _log_lock:
+        print(colored_line, flush=True)
+        with open(COMBINED_LOG, "a") as f:
+            f.write(plain_line + "\n")
+
+
+def stream_to_log(pipe, task_id, task_log_path):
+    """Read lines from a subprocess pipe and write to both per-task and combined logs."""
+    tag_colored = _colored_tag(task_id)
+    with open(task_log_path, "a") as task_log:
+        for raw_line in iter(pipe.readline, b""):
+            decoded = raw_line.decode("utf-8", errors="replace").rstrip("\n")
+            timestamp = datetime.now().strftime("%H:%M:%S")
+            plain_entry = f"[{timestamp}] [{task_id}] {decoded}"
+            colored_entry = f"[{timestamp}] {tag_colored} {decoded}"
+
+            task_log.write(plain_entry + "\n")
+            task_log.flush()
+
+            with _log_lock:
+                with open(COMBINED_LOG, "a") as f:
+                    f.write(colored_entry + "\n")
+
+
+def format_duration(seconds):
+    """Format seconds into human-readable duration."""
+    if seconds < 60:
+        return f"{seconds:.1f}s"
+    minutes = int(seconds // 60)
+    secs = seconds % 60
+    if minutes < 60:
+        return f"{minutes}m {secs:.0f}s"
+    hours = int(minutes // 60)
+    mins = minutes % 60
+    return f"{hours}h {mins}m {secs:.0f}s"
+
+
+def resolve_prompt_file(args, pwd):
+    """Resolve the prompt file path, exit if not found."""
+    if not args.prompt_file:
+        return None
+    prompt_path = Path(args.prompt_file)
+    if not prompt_path.is_absolute():
+        prompt_path = Path(pwd) / prompt_path
+    if not prompt_path.exists():
+        print(f"ERROR: Prompt file not found: {prompt_path}", file=sys.stderr)
+        sys.exit(2)
+    return str(prompt_path)
+
+
+def print_summary(task_ids, results, task_times, wall_elapsed, total_tasks):
+    """Print final results and stats summary."""
+    log_combined("")
+    log_combined("=" * 60)
+    log_combined("RESULTS")
+    log_combined("=" * 60)
+
+    ok = 0
+    failed = 0
+    for task_id in task_ids:
+        rc = results.get(task_id, -1)
+        elapsed = task_times.get(task_id, 0)
+        status_str = "OK" if rc == 0 else f"FAILED (exit {rc})"
+        log_combined(f"  {status_str}  [{format_duration(elapsed)}]", task_id)
+        if rc == 0:
+            ok += 1
+        else:
+            failed += 1
+
+    log_combined("-" * 60)
+    log_combined("STATS")
+    log_combined(f"  Total tasks:    {total_tasks}")
+    log_combined(f"  Succeeded:      {ok}")
+    log_combined(f"  Failed:         {failed}")
+    log_combined(f"  Wall time:      {format_duration(wall_elapsed)}")
+    if task_times:
+        avg = sum(task_times.values()) / len(task_times)
+        fastest = min(task_times.values())
+        slowest = max(task_times.values())
+        log_combined(f"  Avg per task:   {format_duration(avg)}")
+        log_combined(f"  Fastest task:   {format_duration(fastest)}")
+        log_combined(f"  Slowest task:   {format_duration(slowest)}")
+    log_combined(f"  Run ID:         {_run_id}")
+    log_combined(f"  Per-task logs:  {LOG_DIR}/task_<TASK_ID>.log")
+    log_combined("=" * 60)
+
+    return failed
+
+
+# ---------------------------------------------------------------------------
+# Headless mode
+# ---------------------------------------------------------------------------
+
+def run_task(task_id, semaphore, pwd, prompt_file_path=None, output_dir=None,
+             service="claude", model=":sonnet", file_format="", strict=False):
+    """Run a single juno-code subprocess (called from its own thread)."""
+    global _completed_count
+
+    semaphore.acquire()
+    try:
+        log_combined(f"Starting juno-code (thread {threading.current_thread().name})", task_id)
+        start = time.monotonic()
+        start_iso = datetime.now().isoformat()
+
+        task_log_path = LOG_DIR / f"task_{task_id}.log"
+
+        prompt = ""
+
+        if prompt_file_path:
+            try:
+                extra = Path(prompt_file_path).read_text(encoding="utf-8")
+                extra = extra.replace("{{task_id}}", task_id)
+                extra = extra.replace("{{item}}", _item_map.get(task_id, task_id))
+                extra = extra.replace("{{file_format}}", file_format)
+                prompt += "\n\n---\n\n" + extra
+                log_combined(f"Loaded prompt file ({len(extra)} chars)", task_id)
+            except Exception as e:
+                log_combined(f"WARNING: Could not read prompt file: {e}", task_id)
+
+        tmp_prompt_dir = Path("/tmp/pc-headless")
+        tmp_prompt_dir.mkdir(parents=True, exist_ok=True)
+        tmp_prompt_path = tmp_prompt_dir / f"prompt_{task_id}.txt"
+        tmp_prompt_path.write_text(prompt, encoding="utf-8")
+
+        env = _build_process_env()
+
+        proc = subprocess.Popen(
+            [
+                "juno-code",
+                "-b", "shell",
+                "-s", service,
+                "-m", model,
+                "-i", "1",
+                "-v",
+                "--no-hooks",
+                "-f", str(tmp_prompt_path),
+            ],
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            cwd=pwd,
+            bufsize=0,
+            env=env,
+        )
+
+        reader = threading.Thread(
+            target=stream_to_log,
+            args=(proc.stdout, task_id, task_log_path),
+            daemon=True,
+        )
+        reader.start()
+
+        proc.wait()
+        reader.join()
+
+        try:
+            tmp_prompt_path.unlink(missing_ok=True)
+        except OSError:
+            pass
+
+        elapsed = time.monotonic() - start
+
+        with _completed_lock:
+            _completed_count += 1
+            done = _completed_count
+            remaining = _total_tasks - done
+
+        with _task_times_lock:
+            _task_times[task_id] = elapsed
+
+        actual_exit_code = proc.returncode
+        backend_result = _parse_result_from_log(task_log_path) if output_dir else None
+
+        if output_dir:
+            _write_task_output(
+                output_dir, task_id, actual_exit_code, elapsed,
+                start_iso, datetime.now().isoformat(),
+                backend_result=backend_result, file_format=file_format,
+            )
+
+        if strict and file_format and output_dir:
+            actual_exit_code = _extract_strict_output(
+                task_id, task_log_path, output_dir, file_format, actual_exit_code,
+            )
+
+        status = "OK" if actual_exit_code == 0 else f"FAILED (exit {actual_exit_code})"
+        log_combined(
+            f"Finished - {status} ({format_duration(elapsed)}) "
+            f"| Progress: {done}/{_total_tasks} done, {remaining} remaining",
+            task_id,
+        )
+        return task_id, actual_exit_code
+
+    finally:
+        semaphore.release()
+
+
+def run_headless_mode(args, pwd, prompt_file_path, output_dir, service, model):
+    """Run tasks in headless mode using ThreadPoolExecutor."""
+    global _total_tasks
+    _total_tasks = len(args.kanban)
+    wall_start = time.monotonic()
+
+    log_combined("=" * 60)
+    log_combined(f"Starting parallel task execution")
+    log_combined(f"Run ID: {_run_id}")
+    log_combined(f"PWD: {pwd}")
+    log_combined(f"Tasks ({_total_tasks}): {', '.join(args.kanban)}")
+    if any(tid.startswith("item-") for tid in args.kanban):
+        preview = [f"  {tid} -> {_item_map[tid][:80]}" for tid in args.kanban[:3]]
+        log_combined(f"Items preview:\n" + "\n".join(preview)
+                     + (f"\n  ... and {len(args.kanban) - 3} more" if len(args.kanban) > 3 else ""))
+    log_combined(f"Parallelism: {args.parallel}")
+    log_combined(f"Service: {service} | Model: {model}")
+    if prompt_file_path:
+        log_combined(f"Prompt file: {prompt_file_path} (re-read per task)")
+    if output_dir:
+        log_combined(f"Output dir: {output_dir}")
+    legend = "  ".join(f"{_color_for(tid)} {tid}" for tid in args.kanban)
+    with _log_lock:
+        print(f"[{datetime.now().strftime('%Y-%m-%d %H:%M:%S')}] Legend: {legend}", flush=True)
+        with open(COMBINED_LOG, "a") as f:
+            f.write(f"[{datetime.now().strftime('%Y-%m-%d %H:%M:%S')}] Legend: {', '.join(args.kanban)}\n")
+    log_combined("=" * 60)
+
+    semaphore = threading.Semaphore(args.parallel)
+
+    file_format = getattr(args, "file_format", "") or ""
+    strict = getattr(args, "strict", False)
+
+    with ThreadPoolExecutor(max_workers=len(args.kanban)) as pool:
+        futures = {
+            pool.submit(run_task, task_id, semaphore, pwd, prompt_file_path, output_dir,
+                        service, model, file_format, strict): task_id
+            for task_id in args.kanban
+        }
+
+        results = {}
+        for future in as_completed(futures):
+            task_id = futures[future]
+            try:
+                tid, returncode = future.result()
+                results[tid] = returncode
+            except Exception as exc:
+                log_combined(f"EXCEPTION: {exc}", task_id)
+                results[task_id] = -1
+
+    wall_elapsed = time.monotonic() - wall_start
+    failed = print_summary(args.kanban, results, _task_times, wall_elapsed, _total_tasks)
+
+    if output_dir:
+        task_outputs = {}
+        for tid in args.kanban:
+            cap = _task_output_path(output_dir, tid)
+            if cap.exists():
+                try:
+                    task_outputs[tid] = json.loads(cap.read_text(encoding="utf-8"))
+                except (json.JSONDecodeError, OSError):
+                    task_outputs[tid] = {"task_id": tid, "exit_code": results.get(tid, -1),
+                                         "backend_result": None}
+            else:
+                task_outputs[tid] = {"task_id": tid, "exit_code": results.get(tid, -1),
+                                     "backend_result": None}
+        agg_result = _write_aggregation(
+            output_dir, task_outputs, wall_elapsed, args.parallel,
+            mode="headless", file_format=args.file_format,
+        )
+        _print_output_summary(agg_result)
+
+    sys.exit(1 if failed > 0 else 0)
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — data structures
+# ---------------------------------------------------------------------------
+
+class TaskStatus(Enum):
+    PENDING = "pending"
+    RUNNING = "running"
+    DONE = "done"
+    FAILED = "failed"
+
+
+@dataclass
+class TaskState:
+    task_id: str
+    status: TaskStatus = TaskStatus.PENDING
+    worker_id: int = -1
+    start_time: float = 0.0
+    end_time: float = 0.0
+    exit_code: int = -1
+    sentinel_id: str = ""
+    start_time_iso: str = ""
+    end_time_iso: str = ""
+
+
+@dataclass
+class WorkerState:
+    worker_id: int
+    tmux_target: str
+    current_task: str = ""
+    busy: bool = False
+    sentinel_id: str = ""
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — session creation
+# ---------------------------------------------------------------------------
+
+def tmux_run(cmd, check=True):
+    """Run a tmux command, return stdout."""
+    result = subprocess.run(
+        ["tmux"] + cmd,
+        capture_output=True,
+        text=True,
+    )
+    if check and result.returncode != 0:
+        raise RuntimeError(f"tmux {' '.join(cmd)} failed: {result.stderr.strip()}")
+    return result.stdout.strip()
+
+
+def _ansi256_to_tmux_color(code):
+    """Convert ANSI 256-color code to tmux colour string."""
+    return f"colour{code}"
+
+
+def update_pane_border_color(worker_target, task_id):
+    """Set the pane/window border color to match the task's assigned color."""
+    color_code = _task_color_map.get(task_id, 7)
+    tmux_color = _ansi256_to_tmux_color(color_code)
+    tmux_run([
+        "select-pane", "-t", worker_target,
+        "-P", f"border-style=fg={tmux_color}",
+    ], check=False)
+    tmux_run([
+        "select-pane", "-t", worker_target,
+        "-T", f"{task_id}",
+    ], check=False)
+
+
+def create_tmux_session(session_name, mode, num_workers, pwd):
+    """Create tmux session with coordinator + worker windows/panes."""
+    subprocess.run(
+        ["tmux", "kill-session", "-t", session_name],
+        capture_output=True,
+    )
+
+    workers = []
+
+    def _setup_status_bar():
+        tmux_run(["set-option", "-t", session_name, "status", "on"], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-style", "bg=black,fg=white"], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-left", f"#[fg=cyan,bold] {session_name} #[default]"], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-left-length", "25"], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-right", "Starting..."], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-right-length", "120"], check=False)
+        tmux_run(["set-option", "-t", session_name, "pane-border-status", "top"], check=False)
+        tmux_run(["set-option", "-t", session_name, "pane-border-format",
+                   " #{pane_index}: #{pane_title} "], check=False)
+        tmux_run(["set-option", "-t", session_name, "pane-border-indicators", "colour"], check=False)
+
+    if mode == "windows":
+        tmux_run([
+            "new-session", "-d", "-s", session_name,
+            "-n", "coordinator", "-x", "200", "-y", "50",
+        ])
+        tmux_run(["set-option", "-t", session_name, "remain-on-exit", "off"])
+        _setup_status_bar()
+
+        for i in range(num_workers):
+            name = f"worker-{i}"
+            tmux_run(["new-window", "-t", session_name, "-n", name])
+            target = f"{session_name}:{name}"
+            tmux_run(["send-keys", "-t", target, f"cd {shlex.quote(pwd)}", "Enter"])
+            workers.append(WorkerState(worker_id=i, tmux_target=target))
+
+        tmux_run(["select-window", "-t", f"{session_name}:coordinator"])
+        coordinator_target = f"{session_name}:coordinator"
+
+    else:  # panes
+        tmux_run([
+            "new-session", "-d", "-s", session_name,
+            "-n", "main", "-x", "200", "-y", "50",
+        ])
+        tmux_run(["set-option", "-t", session_name, "remain-on-exit", "off"])
+        # FIX-005: Increase scrollback for panes mode to prevent sentinel eviction
+        tmux_run(["set-option", "-t", session_name, "history-limit", "50000"], check=False)
+        _setup_status_bar()
+
+        base = f"{session_name}:main"
+
+        for i in range(num_workers):
+            tmux_run(["split-window", "-t", base, "-v"])
+            tmux_run(["select-layout", "-t", base, "tiled"])
+
+        for i in range(num_workers):
+            pane_idx = i + 1
+            target = f"{base}.{pane_idx}"
+            tmux_run(["send-keys", "-t", target, f"cd {shlex.quote(pwd)}", "Enter"])
+            workers.append(WorkerState(worker_id=i, tmux_target=target))
+
+        coordinator_target = f"{base}.0"
+
+    return coordinator_target, workers
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — command building & dispatch
+# ---------------------------------------------------------------------------
+
+def write_runner_script(task_id, pwd, prompt_file_path, session_name_short,
+                        output_dir=None, service="claude", model=":sonnet",
+                        file_format=""):
+    """Write prompt file + bash runner script for a task."""
+    prompt = ""
+
+    if prompt_file_path:
+        try:
+            extra = Path(prompt_file_path).read_text(encoding="utf-8")
+            extra = extra.replace("{{task_id}}", task_id)
+            extra = extra.replace("{{item}}", _item_map.get(task_id, task_id))
+            extra = extra.replace("{{file_format}}", file_format)
+            prompt += "\n\n---\n\n" + extra
+        except Exception:
+            pass
+
+    sentinel_id = uuid.uuid4().hex[:12]
+
+    tmp = _tmp_dir(session_name_short)
+    tmp.mkdir(parents=True, exist_ok=True)
+
+    prompt_path = tmp / f"prompt_{task_id}.txt"
+    prompt_path.write_text(prompt)
+
+    env_exports = _generate_env_exports()
+
+    env_path = tmp / f"env_{task_id}.sh"
+    env_path.write_text(env_exports + "\n")
+
+    runner_path = tmp / f"run_{task_id}.sh"
+    runner_path.write_text(textwrap.dedent("""\
+        #!/bin/bash
+        source %(env_path)s
+        cd %(pwd)s
+        juno-code -b shell -s %(service)s -m %(model)s -i 1 -v --no-hooks \\
+            -f %(prompt_path)s
+        echo "___DONE_%(sentinel_id)s_${?}___"
+    """) % {
+        "env_path": shlex.quote(str(env_path)),
+        "pwd": shlex.quote(pwd),
+        "prompt_path": shlex.quote(str(prompt_path)),
+        "sentinel_id": sentinel_id,
+        "service": shlex.quote(service),
+        "model": shlex.quote(model),
+    })
+
+    return str(runner_path), sentinel_id
+
+
+def dispatch_task(worker, task_id, task_state, pwd, prompt_file_path,
+                  session_name_short, output_dir=None, service="claude",
+                  model=":sonnet", file_format=""):
+    """Send a task command to a worker's tmux pane/window."""
+    runner_path, sentinel_id = write_runner_script(
+        task_id, pwd, prompt_file_path, session_name_short, output_dir,
+        service, model, file_format)
+
+    # FIX-003: Stop old pipe-pane explicitly before starting new one
+    tmux_run(["pipe-pane", "-t", worker.tmux_target], check=False)
+    time.sleep(0.1)
+
+    task_log = str(LOG_DIR / f"task_{task_id}.log")
+    helper = str(_tmp_dir(session_name_short) / "log_pipe.py")
+    color_code = str(_task_color_map.get(task_id, 7))
+    tmux_run([
+        "pipe-pane", "-t", worker.tmux_target,
+        "-o", "python3 -u %s %s %s %s %s" % (
+            shlex.quote(helper),
+            shlex.quote(task_id),
+            shlex.quote(task_log),
+            shlex.quote(str(COMBINED_LOG)),
+            shlex.quote(color_code),
+        ),
+    ])
+
+    update_pane_border_color(worker.tmux_target, task_id)
+
+    tmux_run(["send-keys", "-t", worker.tmux_target, f"bash {shlex.quote(runner_path)}", "Enter"])
+
+    worker.busy = True
+    worker.current_task = task_id
+    worker.sentinel_id = sentinel_id
+    task_state.status = TaskStatus.RUNNING
+    task_state.worker_id = worker.worker_id
+    task_state.start_time = time.monotonic()
+    task_state.start_time_iso = datetime.now().isoformat()
+    task_state.sentinel_id = sentinel_id
+
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    tag = _colored_tag(task_id)
+    with open(COMBINED_LOG, "a") as f:
+        f.write(f"[{timestamp}] {tag} Dispatched to worker-{worker.worker_id}\n")
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — completion detection
+# ---------------------------------------------------------------------------
+
+_SENTINEL_RE = re.compile(r"___DONE_([a-f0-9]+)_(\d+)___")
+
+
+def check_worker_done(worker):
+    """Check if a worker's current task has finished by looking for the sentinel."""
+    if not worker.busy:
+        return False, -1
+
+    try:
+        # FIX-004: -J joins wrapped lines (prevents sentinel wrap in narrow panes)
+        output = tmux_run([
+            "capture-pane", "-t", worker.tmux_target,
+            "-p", "-J", "-S", "-20",
+        ], check=False)
+    except Exception:
+        return False, -1
+
+    for match in _SENTINEL_RE.finditer(output):
+        if match.group(1) == worker.sentinel_id:
+            exit_code = int(match.group(2))
+            return True, exit_code
+
+    return False, -1
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — dashboard
+# ---------------------------------------------------------------------------
+
+def update_tmux_status_bar(task_states, paused, wall_start, session_name):
+    """Update tmux bottom status bar with progress summary."""
+    now = time.monotonic()
+    done = sum(1 for t in task_states.values() if t.status == TaskStatus.DONE)
+    failed = sum(1 for t in task_states.values() if t.status == TaskStatus.FAILED)
+    running = sum(1 for t in task_states.values() if t.status == TaskStatus.RUNNING)
+    total = len(task_states)
+    completed = done + failed
+    pct = int(completed / total * 100) if total > 0 else 0
+    elapsed = format_duration(now - wall_start)
+
+    pause_tag = " #[fg=yellow,bold]PAUSED#[default]" if paused else ""
+    fail_tag = f" #[fg=red]{failed}F#[default]" if failed > 0 else ""
+
+    status = (
+        f"#[fg=cyan,bold]Progress:#[default] {completed}/{total} ({pct}%)"
+        f" | #[fg=green]{done}OK#[default]{fail_tag}"
+        f" | #[fg=blue]{running} running#[default]"
+        f" | {elapsed}{pause_tag}"
+    )
+
+    try:
+        tmux_run(["set-option", "-t", session_name, "status-right-length", "120"], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-right", status], check=False)
+        tmux_run(["set-option", "-t", session_name, "status-style", "bg=black,fg=white"], check=False)
+    except Exception:
+        pass
+
+
+def update_dashboard_file(task_states, workers, paused, wall_start, session_name_short, session_name):
+    """Write dashboard content to .dashboard_{name} file."""
+    now = time.monotonic()
+    lines = []
+    dashboard_path = _dashboard_file(session_name_short)
+    pause_path = _pause_file(session_name_short)
+
+    pending = sum(1 for t in task_states.values() if t.status == TaskStatus.PENDING)
+    running = sum(1 for t in task_states.values() if t.status == TaskStatus.RUNNING)
+    done = sum(1 for t in task_states.values() if t.status == TaskStatus.DONE)
+    failed = sum(1 for t in task_states.values() if t.status == TaskStatus.FAILED)
+    total = len(task_states)
+
+    other_sessions = _list_running_sessions()
+    total_sessions = len(other_sessions)
+
+    lines.append("=" * 58)
+    session_header = f"  PARALLEL RUNNER — {session_name_short.upper()}"
+    if total_sessions > 1:
+        session_header += f"  ({total_sessions} sessions active)"
+    lines.append(session_header)
+    lines.append("=" * 58)
+
+    # Recent completions
+    finished = sorted(
+        [t for t in task_states.values() if t.status in (TaskStatus.DONE, TaskStatus.FAILED)],
+        key=lambda t: t.end_time,
+        reverse=True,
+    )[:4]
+    lines.append("")
+    lines.append("  RECENT COMPLETIONS:")
+    if finished:
+        for t in finished:
+            elapsed_t = format_duration(t.end_time - t.start_time) if t.start_time > 0 else "?"
+            dot = _color_for(t.task_id)
+            if t.status == TaskStatus.DONE:
+                status = f"\033[32mOK\033[0m"
+            else:
+                status = f"\033[31mFAIL(exit {t.exit_code})\033[0m"
+            log_path = f"logs/task_{t.task_id}.log"
+            lines.append(f"    {dot} {t.task_id}: {status} [{elapsed_t}] -> {log_path}")
+    else:
+        lines.append("    (none yet)")
+
+    # Help
+    lines.append("")
+    if not paused:
+        lines.append(f"  Pause:    touch {pause_path}")
+    else:
+        lines.append(f"  Resume:   rm {pause_path}")
+    lines.append("  Logs:     .juno_task/scripts/logs/task_<TASK_ID>.log")
+    lines.append("  Detach:   Ctrl-b d  (orchestrator keeps running)")
+    lines.append(f"  Reattach: tmux attach -t {session_name}")
+    lines.append(f"  Stop:     Ctrl-c  (or --stop --name {session_name_short})")
+
+    if paused:
+        lines.append("")
+        lines.append("  \033[33;1m*** PAUSED ***\033[0m")
+
+    # Running workers
+    lines.append("")
+    lines.append("  RUNNING:")
+    running_tasks = [t for t in task_states.values() if t.status == TaskStatus.RUNNING]
+    if running_tasks:
+        for t in running_tasks:
+            elapsed_t = format_duration(now - t.start_time)
+            dot = _color_for(t.task_id)
+            lines.append(f"    {dot} worker-{t.worker_id}: {t.task_id} ({elapsed_t})")
+    else:
+        lines.append("    (none)")
+
+    # Progress bar
+    lines.append("")
+    lines.append("-" * 58)
+    elapsed = format_duration(now - wall_start)
+    lines.append(f"  Pending: {pending}  |  Running: {running}  |  Done: {done}  |  Failed: {failed}  |  Total: {total}")
+    if total > 0:
+        completed = done + failed
+        remaining = pending + running
+        pct = int(completed / total * 100)
+        bar_width = 40
+        filled = int(bar_width * completed / total)
+        bar = "\u2588" * filled + "\u2591" * (bar_width - filled)
+        eta_str = ""
+        if completed > 0 and remaining > 0:
+            finished_tasks = [t for t in task_states.values()
+                              if t.status in (TaskStatus.DONE, TaskStatus.FAILED)
+                              and t.start_time > 0 and t.end_time > 0]
+            if finished_tasks:
+                avg_task = sum(t.end_time - t.start_time for t in finished_tasks) / len(finished_tasks)
+                active_workers = max(len(workers), 1)
+                eta_secs = avg_task * remaining / active_workers
+                eta_str = f"  ETA: ~{format_duration(eta_secs)}"
+        lines.append(f"  [{bar}] {pct}% ({completed}/{total})  Wall: {elapsed}{eta_str}")
+    lines.append("-" * 58)
+
+    dashboard_path.write_text("\n".join(lines) + "\n")
+    update_tmux_status_bar(task_states, paused, wall_start, session_name)
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — orchestration loop
+# ---------------------------------------------------------------------------
+
+def orchestration_loop(task_states, workers, task_queue, pwd, prompt_file_path,
+                       wall_start, session_name_short, session_name,
+                       output_dir=None, service="claude", model=":sonnet",
+                       file_format="", strict=False):
+    """Main orchestration loop — polls workers, dispatches tasks, updates dashboard."""
+    all_task_ids = list(task_states.keys())
+    pause_path = _pause_file(session_name_short)
+
+    while True:
+        if _shutdown_event.is_set():
+            print(f"[{datetime.now()}] Shutdown signal received, stopping dispatch...",
+                  file=sys.stderr)
+            break
+
+        result = subprocess.run(
+            ["tmux", "has-session", "-t", session_name],
+            capture_output=True,
+        )
+        if result.returncode != 0:
+            print(f"[{datetime.now()}] Tmux session '{session_name}' gone, shutting down...",
+                  file=sys.stderr)
+            break
+
+        paused = pause_path.exists()
+
+        # Check each worker for completion
+        for worker in workers:
+            if not worker.busy:
+                continue
+
+            done, exit_code = check_worker_done(worker)
+            if not done:
+                continue
+
+            task_id = worker.current_task
+            ts = task_states[task_id]
+            ts.end_time = time.monotonic()
+            ts.end_time_iso = datetime.now().isoformat()
+            ts.exit_code = exit_code
+
+            if exit_code == 0:
+                ts.status = TaskStatus.DONE
+            else:
+                ts.status = TaskStatus.FAILED
+
+            worker.busy = False
+            worker.current_task = ""
+            worker.sentinel_id = ""
+
+            elapsed = ts.end_time - ts.start_time
+            status_str = "OK" if exit_code == 0 else f"FAILED (exit {exit_code})"
+
+            # FIX-002: Parse result from log, with capture-pane fallback
+            task_log_path = LOG_DIR / f"task_{task_id}.log"
+            backend_result = _parse_result_from_log(task_log_path) if output_dir else None
+
+            if output_dir and backend_result is None:
+                try:
+                    scrollback = tmux_run([
+                        "capture-pane", "-t", worker.tmux_target,
+                        "-p", "-S", "-",
+                    ], check=False)
+                    if scrollback:
+                        ansi_re = re.compile(r'\x1b\[[0-9;]*[a-zA-Z]')
+                        clean = ansi_re.sub('', scrollback)
+                        task_log_path.write_text(clean, encoding="utf-8")
+                        backend_result = _parse_result_from_log(task_log_path)
+                except Exception:
+                    pass
+
+            if strict and file_format and output_dir:
+                exit_code = _extract_strict_output(
+                    task_id, task_log_path, output_dir, file_format, exit_code,
+                )
+                ts.exit_code = exit_code
+                if exit_code != 0 and ts.status != TaskStatus.FAILED:
+                    ts.status = TaskStatus.FAILED
+
+            if output_dir:
+                _write_task_output(
+                    output_dir, task_id, exit_code, elapsed,
+                    ts.start_time_iso, ts.end_time_iso, ts.worker_id,
+                    backend_result=backend_result, file_format=file_format,
+                )
+
+            # Clean up temp files
+            tmp = _tmp_dir(session_name_short)
+            for tmp_f in [tmp / f"prompt_{task_id}.txt", tmp / f"run_{task_id}.sh",
+                         tmp / f"env_{task_id}.sh"]:
+                try:
+                    tmp_f.unlink(missing_ok=True)
+                except OSError:
+                    pass
+
+            timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            tag = _colored_tag(task_id)
+            with open(COMBINED_LOG, "a") as f:
+                f.write(
+                    f"[{timestamp}] {tag} Finished - {status_str} "
+                    f"({format_duration(elapsed)})\n"
+                )
+
+        # Dispatch new tasks to free workers
+        if not paused and not _shutdown_event.is_set():
+            for worker in workers:
+                if worker.busy:
+                    continue
+                if not task_queue:
+                    break
+                next_task_id = task_queue.popleft()
+                dispatch_task(
+                    worker, next_task_id, task_states[next_task_id],
+                    pwd, prompt_file_path, session_name_short, output_dir,
+                    service, model, file_format,
+                )
+
+        # Update dashboard
+        update_dashboard_file(task_states, workers, paused, wall_start,
+                              session_name_short, session_name)
+
+        # Check if all done
+        all_done = all(
+            ts.status in (TaskStatus.DONE, TaskStatus.FAILED)
+            for ts in task_states.values()
+        )
+        if all_done:
+            break
+
+        if _shutdown_event.is_set():
+            any_busy = any(w.busy for w in workers)
+            if not any_busy:
+                break
+
+        time.sleep(2)
+
+    # Close all pipe-panes
+    for worker in workers:
+        tmux_run(["pipe-pane", "-t", worker.tmux_target], check=False)
+
+    shutil.rmtree(str(_tmp_dir(session_name_short)), ignore_errors=True)
+
+    # Final dashboard update
+    update_dashboard_file(task_states, workers, False, wall_start,
+                          session_name_short, session_name)
+
+    # Build results
+    results = {}
+    task_times = {}
+    for tid, ts in task_states.items():
+        results[tid] = ts.exit_code
+        if ts.start_time > 0 and ts.end_time > 0:
+            task_times[tid] = ts.end_time - ts.start_time
+
+    wall_elapsed = time.monotonic() - wall_start
+
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    ok = sum(1 for ts in task_states.values() if ts.status == TaskStatus.DONE)
+    failed = sum(1 for ts in task_states.values() if ts.status == TaskStatus.FAILED)
+
+    summary_lines = [
+        "",
+        "=" * 60,
+        "RESULTS",
+        "=" * 60,
+    ]
+    for tid in all_task_ids:
+        rc = results.get(tid, -1)
+        el = task_times.get(tid, 0)
+        st = "OK" if rc == 0 else f"FAILED (exit {rc})"
+        tag = _colored_tag(tid)
+        summary_lines.append(f"  {tag} {st}  [{format_duration(el)}]")
+
+    summary_lines.extend([
+        "-" * 60,
+        "STATS",
+        f"  Total tasks:    {len(task_states)}",
+        f"  Succeeded:      {ok}",
+        f"  Failed:         {failed}",
+        f"  Wall time:      {format_duration(wall_elapsed)}",
+    ])
+    if task_times:
+        avg = sum(task_times.values()) / len(task_times)
+        fastest = min(task_times.values())
+        slowest = max(task_times.values())
+        summary_lines.extend([
+            f"  Avg per task:   {format_duration(avg)}",
+            f"  Fastest task:   {format_duration(fastest)}",
+            f"  Slowest task:   {format_duration(slowest)}",
+        ])
+    summary_lines.extend([
+        f"  Run ID:         {_run_id}",
+        f"  Per-task logs:  {LOG_DIR}/task_<TASK_ID>.log",
+        "=" * 60,
+    ])
+
+    with open(COMBINED_LOG, "a") as f:
+        for line in summary_lines:
+            f.write(f"[{timestamp}] {line}\n")
+
+    dashboard_path = _dashboard_file(session_name_short)
+    with open(dashboard_path, "a") as f:
+        f.write("\n")
+        for line in summary_lines:
+            f.write(f"  {line}\n")
+
+    # Write aggregation
+    if output_dir:
+        task_outputs = {}
+        for tid in all_task_ids:
+            cap = _task_output_path(output_dir, tid)
+            if cap.exists():
+                try:
+                    task_outputs[tid] = json.loads(cap.read_text(encoding="utf-8"))
+                except (json.JSONDecodeError, OSError):
+                    task_outputs[tid] = {"task_id": tid, "exit_code": results.get(tid, -1),
+                                         "backend_result": None}
+            else:
+                task_outputs[tid] = {"task_id": tid, "exit_code": results.get(tid, -1),
+                                     "backend_result": None}
+        agg_result = _write_aggregation(
+            output_dir, task_outputs, wall_elapsed, len(workers),
+            mode=f"tmux/{session_name}", session_name=session_name_short,
+            file_format=file_format,
+        )
+        summary_text = _format_output_summary(agg_result)
+        with open(COMBINED_LOG, "a") as f:
+            f.write(f"[{timestamp}] {summary_text}\n")
+        with open(dashboard_path, "a") as f:
+            f.write(f"\n{summary_text}\n")
+
+    # Tmux notification
+    try:
+        tmux_run([
+            "display-message", "-t", session_name,
+            f"All {len(task_states)} tasks complete! ({ok} OK, {failed} failed)",
+        ], check=False)
+    except Exception:
+        pass
+
+    return 1 if failed > 0 else 0
+
+
+# ---------------------------------------------------------------------------
+# Tmux mode — entry point
+# ---------------------------------------------------------------------------
+
+def run_tmux_mode(args, pwd, prompt_file_path, output_dir, service, model):
+    """Set up tmux session and run orchestrator."""
+    num_workers = args.parallel
+    mode = args.tmux
+
+    session_name_short = args.name if args.name else _next_batch_name()
+    session_name = _session_name_to_tmux(session_name_short)
+
+    if mode == "panes" and num_workers > 5:
+        print(
+            f"WARNING: --parallel {num_workers} with panes mode may make panes too small. "
+            f"Consider using 'windows' mode or reducing --parallel to 5.",
+            file=sys.stderr,
+        )
+
+    _log_base.mkdir(parents=True, exist_ok=True)
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+    # Kill stale orchestrator daemon for THIS session name only
+    pid_path = _pid_file(session_name_short)
+    if pid_path.exists():
+        try:
+            old_pid = int(pid_path.read_text().strip())
+            os.kill(old_pid, signal.SIGTERM)
+            print(f"Killed stale orchestrator for '{session_name_short}' (PID {old_pid})")
+            time.sleep(0.3)
+        except (ValueError, ProcessLookupError, PermissionError):
+            pass
+
+    for f in [_dashboard_file(session_name_short), _pause_file(session_name_short), pid_path]:
+        if f.exists():
+            f.unlink()
+
+    # Log startup
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    with open(COMBINED_LOG, "a") as f:
+        f.write(f"[{timestamp}] {'=' * 60}\n")
+        f.write(f"[{timestamp}] Starting parallel task execution (tmux {mode} mode)\n")
+        f.write(f"[{timestamp}] Run ID: {_run_id}\n")
+        f.write(f"[{timestamp}] Session: {session_name} (name: {session_name_short})\n")
+        f.write(f"[{timestamp}] PWD: {pwd}\n")
+        f.write(f"[{timestamp}] Tasks ({len(args.kanban)}): {', '.join(args.kanban)}\n")
+        if any(tid.startswith("item-") for tid in args.kanban):
+            for tid in args.kanban[:3]:
+                f.write(f"[{timestamp}]   {tid} -> {_item_map[tid][:80]}\n")
+            if len(args.kanban) > 3:
+                f.write(f"[{timestamp}]   ... and {len(args.kanban) - 3} more\n")
+        f.write(f"[{timestamp}] Parallelism: {num_workers}\n")
+        f.write(f"[{timestamp}] Service: {service} | Model: {model}\n")
+        if prompt_file_path:
+            f.write(f"[{timestamp}] Prompt file: {prompt_file_path} (re-read per task)\n")
+        if output_dir:
+            f.write(f"[{timestamp}] Output dir: {output_dir}\n")
+        f.write(f"[{timestamp}] {'=' * 60}\n")
+
+    _write_log_pipe_helper(session_name_short)
+
+    print(f"Creating tmux session '{session_name}' ({mode} mode, {num_workers} workers)...")
+
+    coordinator_target, workers = create_tmux_session(session_name, mode, num_workers, pwd)
+
+    task_states = {}
+    for tid in args.kanban:
+        task_states[tid] = TaskState(task_id=tid)
+    task_queue = deque(args.kanban)
+
+    wall_start = time.monotonic()
+
+    # Start coordinator dashboard
+    pid_path_str = shlex.quote(str(pid_path))
+    dashboard_file_str = shlex.quote(str(_dashboard_file(session_name_short)))
+    dashboard_cmd = (
+        f"trap 'kill $(cat {pid_path_str}) 2>/dev/null; exit' INT; "
+        f"while true; do printf '\\033[H'; cat {dashboard_file_str} 2>/dev/null "
+        f"|| echo 'Waiting for dashboard...'; printf '\\033[J'; sleep 2; done"
+    )
+    tmux_run(["send-keys", "-t", coordinator_target, dashboard_cmd, "Enter"])
+
+    update_dashboard_file(task_states, workers, False, wall_start,
+                          session_name_short, session_name)
+
+    # Fork: child = orchestrator daemon, parent = tmux attach
+    pid = os.fork()
+
+    if pid == 0:
+        # Child process — orchestrator daemon
+        try:
+            os.setsid()
+        except OSError:
+            pass
+
+        pid_path.write_text(str(os.getpid()))
+
+        log_fd = os.open(
+            str(_orchestrator_log(session_name_short)),
+            os.O_WRONLY | os.O_CREAT | os.O_APPEND,
+        )
+        os.dup2(log_fd, 1)
+        os.dup2(log_fd, 2)
+        os.close(log_fd)
+
+        signal.signal(signal.SIGHUP, signal.SIG_IGN)
+
+        def _shutdown_handler(signum, frame):
+            sig_name = "SIGTERM" if signum == signal.SIGTERM else "SIGINT"
+            print(f"[{datetime.now()}] Received {sig_name}, initiating graceful shutdown...",
+                  file=sys.stderr)
+            _shutdown_event.set()
+
+        signal.signal(signal.SIGTERM, _shutdown_handler)
+        signal.signal(signal.SIGINT, _shutdown_handler)
+
+        try:
+            for worker in workers:
+                if not task_queue:
+                    break
+                next_task_id = task_queue.popleft()
+                file_format = getattr(args, "file_format", "") or ""
+                dispatch_task(
+                    worker, next_task_id, task_states[next_task_id],
+                    pwd, prompt_file_path, session_name_short, output_dir,
+                    service, model, file_format,
+                )
+
+            update_dashboard_file(task_states, workers, False, wall_start,
+                                  session_name_short, session_name)
+
+            file_format = getattr(args, "file_format", "") or ""
+            strict = getattr(args, "strict", False)
+            exit_code = orchestration_loop(
+                task_states, workers, task_queue,
+                pwd, prompt_file_path, wall_start,
+                session_name_short, session_name, output_dir,
+                service, model, file_format, strict,
+            )
+        except Exception:
+            import traceback
+            traceback.print_exc()
+            exit_code = 1
+
+        if pid_path.exists():
+            try:
+                pid_path.unlink()
+            except OSError:
+                pass
+
+        os._exit(exit_code)
+
+    else:
+        # Parent process — attach to tmux session
+        print(f"Orchestrator daemon started (PID {pid})")
+        print(f"Run ID: {_run_id}")
+        print(f"Session: {session_name} (name: {session_name_short})")
+        print(f"Tasks: {', '.join(args.kanban)}")
+        print(f"Workers: {num_workers}")
+        print(f"Logs: {LOG_DIR}/")
+        print(f"Pause: touch {_pause_file(session_name_short)}")
+        print(f"Stop:  --stop --name {session_name_short}")
+        print(f"Attaching to tmux session...")
+        print()
+
+        time.sleep(0.5)
+
+        os.execvp("tmux", ["tmux", "attach-session", "-t", session_name])
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+def main():
+    global LOG_DIR, COMBINED_LOG, _run_id
+
+    args = parse_args()
+
+    if args.stop_all:
+        run_stop_all()
+        return
+    if args.stop:
+        run_stop(args)
+        return
+
+    pwd = os.getcwd()
+
+    # Generate run ID and compute per-run LOG_DIR
+    _run_id = ''.join(random.choices(string.ascii_lowercase + string.digits, k=5))
+    date_str = datetime.now().strftime("%Y-%m-%d")
+    LOG_DIR = _log_base / date_str / _run_id
+    COMBINED_LOG = LOG_DIR / "parallel_runner.log"
+
+    # Assign colors
+    for i, tid in enumerate(args.kanban):
+        _task_color_map[tid] = _TASK_COLORS[i % len(_TASK_COLORS)]
+
+    service, model = _resolve_service_model(args)
+    prompt_file_path = resolve_prompt_file(args, pwd)
+    output_dir = _resolve_output_dir(args)
+
+    _log_base.mkdir(parents=True, exist_ok=True)
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+
+    if args.tmux:
+        run_tmux_mode(args, pwd, prompt_file_path, output_dir, service, model)
+    else:
+        run_headless_mode(args, pwd, prompt_file_path, output_dir, service, model)
+
+
+if __name__ == "__main__":
+    main()