File size: 41,715 Bytes
7f611c5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 | """
WebSocket + HTTP server for the live monitor dashboard.
Single port, zero extra dependencies beyond Python stdlib + optional websockets.
Uses raw asyncio TCP so it works regardless of websockets version.
HTTP GET / → serves dashboard.html
WS upgrade → real-time event stream
"""
import asyncio
import base64
import hashlib
import json
import logging
import os
import queue
import struct
import threading
import time
import urllib.error
import urllib.request
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
from typing import Any, Dict, List, Optional, Set
logger = logging.getLogger(__name__)
DASHBOARD_PATH = Path(__file__).parent / "dashboard.html"
WS_GUID = "258EAFA5-E914-47DA-95CA-C5AB0DC85B11"
def _ws_accept_key(client_key: str) -> str:
digest = hashlib.sha1((client_key + WS_GUID).encode()).digest()
return base64.b64encode(digest).decode()
def _ws_encode_text(text: str) -> bytes:
"""Encode a text frame (server→client, unmasked)."""
payload = text.encode("utf-8")
length = len(payload)
if length < 126:
header = struct.pack("BB", 0x81, length)
elif length < 65536:
header = struct.pack("!BBH", 0x81, 126, length)
else:
header = struct.pack("!BBQ", 0x81, 127, length)
return header + payload
async def _ws_read_frame(reader: asyncio.StreamReader) -> Optional[str]:
"""Read one WebSocket frame; return text payload or None on close/error."""
try:
header = await reader.readexactly(2)
except Exception:
return None
opcode = header[0] & 0x0F
masked = (header[1] & 0x80) != 0
length = header[1] & 0x7F
if opcode == 0x8: # Close
return None
if opcode == 0x9: # Ping — we could reply with pong but we ignore it here
return None
if length == 126:
ext = await reader.readexactly(2)
length = struct.unpack("!H", ext)[0]
elif length == 127:
ext = await reader.readexactly(8)
length = struct.unpack("!Q", ext)[0]
if masked:
mask = await reader.readexactly(4)
data = bytearray(await reader.readexactly(length))
for i in range(length):
data[i] ^= mask[i % 4]
payload = bytes(data)
else:
payload = await reader.readexactly(length)
if opcode == 0x1: # Text
return payload.decode("utf-8", errors="replace")
return None # Binary / continuation frames ignored
class MonitorServer:
"""
Single-port HTTP+WebSocket server for live solution discovery monitoring.
- GET / → dashboard.html
- WS upgrade → event broadcast
Runs in a daemon thread with its own asyncio event loop.
"""
def __init__(self, host: str = "127.0.0.1", port: int = 8765, max_solution_length: int = 10000):
self.host = host
self.port = port
self.max_solution_length = max_solution_length
self._queue: queue.Queue = queue.Queue()
# In-memory state for reconnecting clients
self._programs: List[Dict[str, Any]] = []
self._program_solutions: Dict[str, str] = {}
self._parent_solutions: Dict[str, str] = {}
self._best_program_id: Optional[str] = None
self._best_score: float = -float("inf")
self._stats: Dict[str, Any] = {}
self._config_summary: str = ""
# Per-program summary cache
self._program_summary_cache: Dict[str, str] = {}
# Human feedback reader (set via set_feedback_reader)
self._feedback_reader: Optional[Any] = None
# AI summary state
self._summary_model: str = ""
self._summary_api_key: str = ""
self._summary_api_base: str = "https://api.openai.com/v1"
self._summary_top_k: int = 3
self._summary_interval: int = 0 # 0 = manual only
self._summary_text: str = ""
self._summary_generating: bool = False
self._summary_last_program_count: int = 0
self._summary_executor: Optional[ThreadPoolExecutor] = None
self._loop: Optional[asyncio.AbstractEventLoop] = None
self._thread: Optional[threading.Thread] = None
self._clients: Set[asyncio.StreamWriter] = set()
self._stop_event = threading.Event()
self._ready_event = threading.Event() # set when TCP port is bound
self._dashboard_html: Optional[bytes] = None
def start(self) -> None:
"""Load the dashboard and start the server in a daemon thread."""
self._load_dashboard()
self._thread = threading.Thread(target=self._run_loop, daemon=True)
self._thread.start()
# Wait until TCP port is actually bound (up to 5s)
self._ready_event.wait(timeout=5)
logger.info(f"Monitor server started → http://localhost:{self.port}/")
def stop(self) -> None:
"""Signal the server to stop and wait for the thread to finish."""
self._stop_event.set()
loop = self._loop
if loop is not None and not loop.is_closed():
# Schedule cancellation of all tasks, then stop the loop
try:
loop.call_soon_threadsafe(self._cancel_all_tasks)
except RuntimeError:
pass # Loop already closed
if self._thread:
self._thread.join(timeout=5)
def _cancel_all_tasks(self) -> None:
"""Cancel every pending task on the server's event loop, then stop it."""
loop = self._loop
if loop is None or loop.is_closed():
return
for task in asyncio.all_tasks(loop):
task.cancel()
loop.stop()
def push_event(self, event: Dict[str, Any]) -> None:
"""Enqueue an event for broadcast to all connected WebSocket clients."""
self._queue.put_nowait(event)
def set_config_summary(self, summary: str) -> None:
"""Set a human-readable config summary sent to new dashboard clients."""
self._config_summary = summary
def set_feedback_reader(self, reader: Any) -> None:
"""Attach a HumanFeedbackReader for dashboard human feedback controls."""
self._feedback_reader = reader
def configure_summary(
self,
model: str = "gpt-5-mini",
api_key: str = "",
api_base: str = "https://api.openai.com/v1",
top_k: int = 3,
interval: int = 0,
) -> None:
"""Configure the AI summary generator.
Args:
model: OpenAI model name (default gpt-5-mini).
api_key: API key. Falls back to OPENAI_API_KEY env var.
api_base: API base URL.
top_k: Number of top programs to include in summary prompt.
interval: Auto-generate every N new programs (0 = manual only).
"""
self._summary_model = model
self._summary_api_key = api_key or os.environ.get("OPENAI_API_KEY", "")
self._summary_api_base = api_base
self._summary_top_k = top_k
self._summary_interval = interval
self._summary_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="summary")
logger.info(
f"AI summary configured: model={model}, top_k={top_k}, "
f"interval={interval or 'manual'}, api_key={'set' if self._summary_api_key else 'MISSING'}"
)
def _get_feedback_state(self) -> Dict[str, Any]:
"""Return current human feedback state."""
if not self._feedback_reader:
return {
"human_feedback_enabled": False,
"feedback_text": "",
"feedback_active": False,
"human_feedback_mode": "append",
"human_feedback_current_prompt": "",
"human_feedback_history": [],
}
text = self._feedback_reader.read()
return {
"human_feedback_enabled": True,
"feedback_text": text,
"feedback_active": bool(text),
"human_feedback_mode": self._feedback_reader.mode,
"human_feedback_current_prompt": self._feedback_reader.get_current_prompt(),
"human_feedback_history": self._feedback_reader.get_history(),
}
def _build_init_state(self) -> Dict[str, Any]:
"""Build the full init_state payload for new/reconnecting WS clients."""
state = {
"type": "init_state",
"programs": self._programs,
"best_program_id": self._best_program_id,
"stats": self._stats,
"config_summary": self._config_summary,
"summary_enabled": bool(self._summary_model),
"summary_model": self._summary_model or "",
"summary_text": self._summary_text,
"summary_generating": self._summary_generating,
}
state.update(self._get_feedback_state())
return state
def _load_dashboard(self) -> None:
try:
raw = DASHBOARD_PATH.read_text(encoding="utf-8")
except FileNotFoundError:
logger.warning(f"Dashboard HTML not found at {DASHBOARD_PATH}")
raw = "<html><body><h1>Dashboard not found</h1></body></html>"
# No port injection needed — WS connects to the same host:port
self._dashboard_html = raw.encode("utf-8")
def _run_loop(self) -> None:
self._loop = asyncio.new_event_loop()
asyncio.set_event_loop(self._loop)
try:
self._loop.run_until_complete(self._serve())
except (RuntimeError, asyncio.CancelledError):
pass # Normal on shutdown
except Exception:
logger.exception("Monitor server error")
finally:
# Drain any remaining cancelled tasks so they don't warn on GC
try:
pending = asyncio.all_tasks(self._loop)
if pending:
for t in pending:
t.cancel()
self._loop.run_until_complete(asyncio.gather(*pending, return_exceptions=True))
except Exception:
logger.debug("Error cancelling tasks during stop", exc_info=True)
try:
self._loop.close()
except Exception:
logger.debug("Error closing event loop", exc_info=True)
async def _serve(self) -> None:
# Try configured port, then auto-increment if already in use
port = self.port
for attempt in range(10):
try:
server = await asyncio.start_server(self._handle_connection, self.host, port)
break
except OSError:
if attempt == 9:
raise
port += 1
self.port = port
async with server:
self._ready_event.set() # signal that port is bound
logger.debug(f"Listening on {self.host}:{self.port}")
consumer = asyncio.create_task(self._consume_queue())
hb = asyncio.create_task(self._heartbeat())
try:
await asyncio.gather(consumer, hb)
except (asyncio.CancelledError, RuntimeError):
pass
finally:
try:
consumer.cancel()
hb.cancel()
except RuntimeError:
pass # Event loop already closed
async def _handle_connection(
self, reader: asyncio.StreamReader, writer: asyncio.StreamWriter
) -> None:
"""Route an incoming connection to HTTP or WebSocket handler."""
try:
# Read HTTP request line + headers
raw_headers: Dict[str, str] = {}
request_line = (await reader.readline()).decode("utf-8", errors="replace").strip()
if not request_line:
writer.close()
return
while True:
line = (await reader.readline()).decode("utf-8", errors="replace").strip()
if not line:
break
if ":" in line:
k, _, v = line.partition(":")
raw_headers[k.strip().lower()] = v.strip()
is_ws = raw_headers.get("upgrade", "").lower() == "websocket"
if is_ws:
await self._handle_ws(reader, writer, raw_headers)
else:
await self._handle_http(writer)
except Exception:
logger.debug("Connection handler error", exc_info=True)
finally:
try:
writer.close()
except Exception:
logger.debug("Error closing writer", exc_info=True)
async def _handle_http(self, writer: asyncio.StreamWriter) -> None:
"""Serve the dashboard HTML over a plain HTTP GET."""
html = self._dashboard_html or b""
resp = (
"HTTP/1.1 200 OK\r\n"
"Content-Type: text/html; charset=utf-8\r\n"
f"Content-Length: {len(html)}\r\n"
"Connection: close\r\n"
"\r\n"
).encode() + html
writer.write(resp)
await writer.drain()
async def _handle_ws(
self,
reader: asyncio.StreamReader,
writer: asyncio.StreamWriter,
headers: Dict[str, str],
) -> None:
"""Complete the WebSocket handshake and enter the read loop."""
key = headers.get("sec-websocket-key", "")
accept = _ws_accept_key(key)
handshake = (
"HTTP/1.1 101 Switching Protocols\r\n"
"Upgrade: websocket\r\n"
"Connection: Upgrade\r\n"
f"Sec-WebSocket-Accept: {accept}\r\n"
"\r\n"
).encode()
writer.write(handshake)
await writer.drain()
self._clients.add(writer)
logger.debug(f"WS client connected ({len(self._clients)} total)")
try:
await self._ws_send(writer, json.dumps(self._build_init_state()))
# Read loop
while True:
text = await _ws_read_frame(reader)
if text is None:
break
await self._handle_client_msg(writer, text)
except Exception:
logger.debug("WebSocket handler error", exc_info=True)
finally:
self._clients.discard(writer)
logger.debug(f"WS client disconnected ({len(self._clients)} total)")
async def _handle_client_msg(self, writer: asyncio.StreamWriter, raw: str) -> None:
"""Dispatch an incoming WebSocket JSON message from a dashboard client."""
try:
msg = json.loads(raw)
except Exception:
return
t = msg.get("type")
if t == "request_full_state":
await self._ws_send(writer, json.dumps(self._build_init_state()))
elif t == "request_program_solution":
pid = msg.get("program_id", "")
await self._ws_send(
writer,
json.dumps(
{
"type": "program_solution",
"program_id": pid,
"solution": self._program_solutions.get(pid, "")[
: self.max_solution_length
],
"parent_solution": self._parent_solutions.get(pid, "")[
: self.max_solution_length
],
}
),
)
elif t == "set_feedback":
text = msg.get("text", "").strip()
if self._feedback_reader:
self._feedback_reader.write_from_dashboard(text)
ack = {
"type": "feedback_ack",
"feedback_text": text,
"feedback_active": bool(text),
"human_feedback_mode": self._feedback_reader.mode,
}
await self._broadcast(json.dumps(ack))
logger.info(f"Human feedback set from dashboard ({len(text)} chars)")
else:
await self._ws_send(
writer,
json.dumps(
{
"type": "feedback_ack",
"feedback_text": "",
"feedback_active": False,
"error": "Human feedback not enabled",
}
),
)
elif t == "clear_feedback":
if self._feedback_reader:
self._feedback_reader.write_from_dashboard("")
ack = {
"type": "feedback_ack",
"feedback_text": "",
"feedback_active": False,
"human_feedback_mode": self._feedback_reader.mode,
}
await self._broadcast(json.dumps(ack))
logger.info("Human feedback cleared from dashboard")
elif t == "request_feedback_state":
await self._ws_send(
writer,
json.dumps(
{
"type": "feedback_ack",
**self._get_feedback_state(),
}
),
)
elif t == "set_human_feedback_mode":
mode = msg.get("mode", "append")
if self._feedback_reader:
self._feedback_reader.set_mode(mode)
ack = {
"type": "human_feedback_mode_ack",
"human_feedback_mode": mode,
}
await self._broadcast(json.dumps(ack))
logger.info(f"Human feedback mode set to: {mode}")
elif t == "request_system_prompt":
prompt_text = ""
if self._feedback_reader:
prompt_text = self._feedback_reader.get_current_prompt()
await self._ws_send(
writer,
json.dumps(
{
"type": "system_prompt",
"prompt_text": prompt_text,
}
),
)
elif t == "request_human_feedback_history":
history = []
if self._feedback_reader:
history = self._feedback_reader.get_history()
await self._ws_send(
writer,
json.dumps(
{
"type": "human_feedback_history",
"history": history,
}
),
)
elif t == "request_image":
image_path = msg.get("image_path", "")
program_id = msg.get("program_id", "")
if image_path and os.path.exists(image_path):
try:
import base64 as _b64
with open(image_path, "rb") as _f:
img_data = _b64.b64encode(_f.read()).decode()
ext = os.path.splitext(image_path)[1].lstrip(".").lower()
mime = {
"png": "image/png",
"jpg": "image/jpeg",
"jpeg": "image/jpeg",
"webp": "image/webp",
"gif": "image/gif",
}.get(ext, "image/png")
await self._ws_send(
writer,
json.dumps(
{
"type": "image_data",
"program_id": program_id,
"data_url": f"data:{mime};base64,{img_data}",
}
),
)
except Exception as e:
logger.warning(f"Failed to serve image {image_path}: {e}")
elif t == "request_program_summary":
pid = msg.get("program_id", "")
await self._generate_program_summary(writer, pid)
elif t == "request_summary":
await self._trigger_summary()
# ─── Queue consumer & broadcast ──────────────────────────
async def _consume_queue(self) -> None:
while not self._stop_event.is_set():
try:
event = self._queue.get_nowait()
except queue.Empty:
await asyncio.sleep(0.05)
continue
etype = event.get("type")
if etype == "new_program":
p = event.get("program", {})
# Annotate with human feedback state for replay on reconnect
if self._feedback_reader:
fb = self._feedback_reader.read()
p["human_feedback_active"] = bool(fb)
else:
p["human_feedback_active"] = False
self._programs.append(p)
pid = p.get("id", "")
if "full_solution" in event:
self._program_solutions[pid] = event["full_solution"]
if "parent_full_solution" in event:
self._parent_solutions[pid] = event["parent_full_solution"]
# Independent best tracking: compare scores directly
new_score = p.get("score", 0)
if not isinstance(new_score, (int, float)):
new_score = 0
if new_score > self._best_score:
self._best_score = new_score
self._best_program_id = pid
event["is_best"] = True
elif event.get("is_best"):
self._best_program_id = pid
self._best_score = max(self._best_score, new_score)
self._stats = event.get("stats", self._stats)
# Strip full_solution from broadcast (clients request on demand)
broadcast = {
k: v for k, v in event.items() if k not in ("full_solution", "parent_full_solution")
}
# Include current human feedback status in program events
if etype == "new_program" and self._feedback_reader:
fb = self._feedback_reader.read()
broadcast["feedback_active"] = bool(fb)
broadcast["feedback_text"] = fb if fb else ""
broadcast["human_feedback_mode"] = self._feedback_reader.mode
await self._broadcast(json.dumps(broadcast))
# Auto-trigger AI summary every N new programs
if (
etype == "new_program"
and self._summary_interval > 0
and self._summary_model
and not self._summary_generating
):
count = len(self._programs)
if count - self._summary_last_program_count >= self._summary_interval:
await self._trigger_summary()
async def _broadcast(self, message: str) -> None:
if not self._clients:
return
dead = set()
for writer in list(self._clients):
try:
await self._ws_send(writer, message)
except Exception:
dead.add(writer)
self._clients -= dead
async def _ws_send(self, writer: asyncio.StreamWriter, text: str) -> None:
writer.write(_ws_encode_text(text))
await writer.drain()
async def _heartbeat(self) -> None:
while not self._stop_event.is_set():
await asyncio.sleep(5)
if self._clients:
await self._broadcast(json.dumps({"type": "heartbeat", "timestamp": time.time()}))
async def _generate_program_summary(self, writer: asyncio.StreamWriter, pid: str) -> None:
"""Generate a crisp LLM summary of what changed in a single program."""
# Return cached if available
if pid in self._program_summary_cache:
await self._ws_send(
writer,
json.dumps(
{
"type": "program_summary",
"program_id": pid,
"summary": self._program_summary_cache[pid],
}
),
)
return
# Need API key + model
if not self._summary_model or not self._summary_api_key:
await self._ws_send(
writer,
json.dumps(
{
"type": "program_summary",
"program_id": pid,
"summary": "AI summary not configured.",
}
),
)
return
# Find program data
prog = None
for p in self._programs:
if p.get("id") == pid:
prog = p
break
if not prog:
return
# Build prompt
code = self._program_solutions.get(pid, prog.get("solution_snippet", ""))
parent_solution = self._parent_solutions.get(pid, "")
score = prog.get("score", "?")
parent_score = prog.get("parent_score")
label = prog.get("label_type", "unknown")
delta_str = ""
if isinstance(score, (int, float)) and isinstance(parent_score, (int, float)):
d = score - parent_score
delta_str = f" (delta: {'+' if d >= 0 else ''}{d:.4f})"
# Truncate code for prompt efficiency
if len(code) > 2000:
code = code[:2000] + "\n... (truncated)"
if len(parent_solution) > 2000:
parent_solution = parent_solution[:2000] + "\n... (truncated)"
is_image_mode = prog.get("image_path") is not None
if is_image_mode:
system = (
"You are analyzing one step in an image generation run. "
"Given the parent generation prompt and the child generation prompt, describe in 1-2 concise bullet points "
"what specifically changed in the prompt.\n\n"
"Rules:\n"
"- Be specific: name style changes, subject modifications, added details\n"
"- Each bullet under 25 words\n"
"- Start each bullet with `- `\n"
"- No headers, no sections — just 1-2 bullets"
)
else:
system = (
"You are analyzing one step in a solution discovery run. "
"Given the parent code and the child code, describe in 1-2 concise bullet points "
"what specifically changed.\n\n"
"Rules:\n"
"- Be specific: name algorithms, parameters, structural changes\n"
"- Each bullet under 25 words\n"
"- Start each bullet with `- `\n"
"- No headers, no sections — just 1-2 bullets\n"
"- Consider the evolution label: exploration = trying new ideas, "
"exploitation = refining current best, diverge = deliberately different strategy"
)
content_label = "prompt" if is_image_mode else "code"
user_parts = [f"Label: {label}{delta_str}"]
if parent_score is not None:
user_parts.append(f"Score: {parent_score} -> {score}")
else:
user_parts.append(f"Score: {score} (no parent)")
if parent_solution:
user_parts.append(f"\nParent {content_label}:\n```\n{parent_solution}\n```")
user_parts.append(f"\nNew {content_label}:\n```\n{code}\n```")
prompt_data = {"system": system, "user": "\n".join(user_parts)}
# Ensure executor exists
if not self._summary_executor:
self._summary_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="summary")
# Run LLM call in executor
result = ""
loop = asyncio.get_running_loop()
try:
result = await loop.run_in_executor(
self._summary_executor,
self._call_program_summary_api,
prompt_data,
)
self._program_summary_cache[pid] = result
except Exception as e:
logger.warning(f"Program summary failed for {pid[:8]}: {e}", exc_info=True)
result = f"Summary unavailable: {e}"
await self._ws_send(
writer,
json.dumps(
{
"type": "program_summary",
"program_id": pid,
"summary": result or "Summary unavailable (empty response).",
}
),
)
def _call_program_summary_api(self, prompt_data: Dict[str, str]) -> str:
"""Call LLM for per-program summary (blocking, runs in executor)."""
return self._call_llm_api(prompt_data, max_tokens=2048, timeout=120)
async def _trigger_summary(self) -> None:
"""Trigger async AI summary generation."""
if not self._summary_model:
await self._broadcast(
json.dumps(
{
"type": "summary_update",
"summary_text": "AI summary not configured (no model set).",
"summary_generating": False,
"summary_enabled": False,
}
)
)
return
if not self._summary_api_key:
await self._broadcast(
json.dumps(
{
"type": "summary_update",
"summary_text": "AI summary not configured. Set OPENAI_API_KEY environment variable or summary_api_key in config.",
"summary_generating": False,
"summary_enabled": False,
}
)
)
return
if self._summary_generating:
return # Already in progress
# Ensure executor exists
if not self._summary_executor:
self._summary_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="summary")
self._summary_generating = True
self._summary_last_program_count = len(self._programs)
# Notify clients that generation started
await self._broadcast(
json.dumps(
{
"type": "summary_update",
"summary_text": self._summary_text,
"summary_generating": True,
"summary_enabled": True,
}
)
)
try:
# Build the prompt data from current programs
top_programs = self._get_top_k_programs()
if not top_programs:
self._summary_text = "No scored programs yet. Run some iterations first."
logger.info("AI summary skipped: no scored programs")
else:
prompt_data = self._build_summary_prompt(top_programs)
logger.info(
f"AI summary: calling {self._summary_model} with {len(top_programs)} "
f"top programs, api_base={self._summary_api_base}"
)
# Run the blocking API call in a thread
loop = asyncio.get_running_loop()
result = await loop.run_in_executor(
self._summary_executor,
self._call_llm_api,
prompt_data,
)
self._summary_text = result or "AI returned empty response."
logger.info(f"AI summary generated ({len(self._summary_text)} chars)")
except Exception as e:
logger.warning(f"AI summary generation failed: {e}", exc_info=True)
self._summary_text = f"Summary generation failed: {e}"
finally:
self._summary_generating = False
# Broadcast the result
await self._broadcast(
json.dumps(
{
"type": "summary_update",
"summary_text": self._summary_text,
"summary_generating": False,
"summary_enabled": True,
}
)
)
def _get_top_k_programs(self) -> List[Dict[str, Any]]:
"""Get top-k programs by score across all islands."""
if not self._programs:
return []
scored = [p for p in self._programs if isinstance(p.get("score"), (int, float))]
scored.sort(key=lambda p: p["score"], reverse=True)
# Deduplicate by score (keep best per unique score to show diversity)
seen_scores = set()
unique = []
for p in scored:
key = round(p["score"], 6)
if key not in seen_scores:
seen_scores.add(key)
unique.append(p)
if len(unique) >= self._summary_top_k:
break
# Fall back to just top-k if not enough unique
if len(unique) < self._summary_top_k:
unique = scored[: self._summary_top_k]
return unique
def _compute_solution_discovery_analysis(self) -> str:
"""Compute evolution progress, improvement patterns, and stagnation analysis."""
programs = self._programs
if not programs:
return ""
scored = [p for p in programs if isinstance(p.get("score"), (int, float))]
if not scored:
return ""
lines = []
n = len(scored)
improvements = 0
regressions = 0
total_with_parent = 0
improvement_deltas = []
for p in scored:
parent_score = p.get("parent_score")
if isinstance(parent_score, (int, float)):
total_with_parent += 1
delta = p["score"] - parent_score
if delta > 0:
improvements += 1
improvement_deltas.append(delta)
elif delta < 0:
regressions += 1
if total_with_parent > 0:
hit_rate = improvements / total_with_parent * 100
avg_gain = (
sum(improvement_deltas) / len(improvement_deltas) if improvement_deltas else 0
)
lines.append("=== Improvement Rate ===")
lines.append(
f" {improvements}/{total_with_parent} programs improved over parent ({hit_rate:.0f}% hit rate)"
)
lines.append(f" Avg improvement when positive: {avg_gain:+.4f}")
if n >= 10:
quarter = max(n // 4, 1)
early_scores = [p["score"] for p in scored[:quarter]]
mid_scores = [p["score"] for p in scored[quarter : quarter * 2]]
recent_scores = [p["score"] for p in scored[-quarter:]]
early_avg = sum(early_scores) / len(early_scores)
mid_avg = sum(mid_scores) / len(mid_scores) if mid_scores else early_avg
recent_avg = sum(recent_scores) / len(recent_scores)
lines.append("\n=== Score Trend ===")
lines.append(
f" Early avg (first {quarter}): {early_avg:.4f} | "
f"Mid avg: {mid_avg:.4f} | "
f"Recent avg (last {quarter}): {recent_avg:.4f}"
)
if recent_avg > mid_avg + 0.001:
lines.append(" Trend: IMPROVING")
elif recent_avg < mid_avg - 0.005:
lines.append(" Trend: REGRESSING")
elif abs(recent_avg - mid_avg) < 0.001 and n > 30:
lines.append(" Trend: PLATEAUED")
else:
lines.append(" Trend: STABLE")
if n >= 5:
best_so_far = -float("inf")
streak = 0
longest_streak = 0
for p in scored:
if p["score"] > best_so_far:
best_so_far = p["score"]
streak = 0
else:
streak += 1
longest_streak = max(longest_streak, streak)
lines.append("\n=== Stagnation ===")
lines.append(
f" Current non-improving streak: {streak} iterations | "
f"Longest streak: {longest_streak}"
)
islands: Dict[Any, list] = {}
for p in scored:
isl = p.get("island")
if isl is not None:
islands.setdefault(isl, []).append(p["score"])
if len(islands) > 1:
lines.append(f"\n=== Island Diversity ({len(islands)} islands) ===")
for isl in sorted(islands.keys()):
scores = islands[isl]
lines.append(
f" Island {isl}: {len(scores)} programs, "
f"best={max(scores):.4f}, avg={sum(scores)/len(scores):.4f}"
)
return "\n".join(lines)
def _build_summary_prompt(self, top_programs: List[Dict[str, Any]]) -> Dict[str, str]:
"""Build the system + user prompt for the summary LLM call."""
system = (
"You are an expert analyst monitoring a solution discovery process. "
"You will be given run statistics, evolution progress data, and the source code "
"of the top-performing programs from the current run.\n\n"
"Respond using EXACTLY this markdown structure:\n\n"
"## Status\n"
"One sentence: is the search improving, stagnating, or plateauing? "
"Cite the score trend numbers.\n\n"
"## Key Techniques\n"
"Bullet list of the main algorithmic ideas found in the top programs' code. "
"Be specific — name the techniques (e.g. 'Kalman filter with adaptive Q', "
"'hexagonal lattice packing', 'exponential moving average').\n\n"
"## Diversity\n"
"Are the top programs converging on one approach or exploring different strategies? "
"One sentence.\n\n"
"## Recommendation\n"
"One specific, actionable suggestion grounded in the code. "
"For example: **try wavelet denoising** — the top programs all use simple "
"moving averages which limits frequency response.\n\n"
"Rules:\n"
"- Use markdown: **bold** for key terms, `- ` for bullets, `##` for sections\n"
"- Be concise — max 250 words total\n"
"- Every claim must reference what you see in the actual code"
)
# Build user message with stats + solution discovery analysis + top-k programs
parts = []
if self._stats:
parts.append(
f"Run: {self._config_summary}\n"
f"Total programs: {self._stats.get('total_programs', len(self._programs))}\n"
f"Current iteration: {self._stats.get('current_iteration', '?')}\n"
f"Best score: {self._stats.get('best_score', '?')}\n"
f"Programs/min: {self._stats.get('programs_per_min', '?')}\n"
f"Elapsed: {self._stats.get('elapsed_seconds', '?')}s\n"
f"Iterations since improvement: {self._stats.get('iterations_since_improvement', '?')}"
)
# Add solution discovery analysis
solution_discovery_analysis = self._compute_solution_discovery_analysis()
if solution_discovery_analysis:
parts.append(f"\n{solution_discovery_analysis}")
for i, p in enumerate(top_programs, 1):
pid = p.get("id", "?")
code = self._program_solutions.get(pid, p.get("solution_snippet", ""))
# Truncate code to keep prompt reasonable
if len(code) > 2000:
code = code[:2000] + "\n... (truncated)"
island_str = f", island={p.get('island')}" if p.get("island") is not None else ""
parts.append(
f"\n--- Top Program #{i} ---\n"
f"ID: {pid}\n"
f"Score: {p.get('score', '?')}\n"
f"Iteration: {p.get('iteration', '?')}{island_str}\n"
f"Metrics: {json.dumps(p.get('metrics', {}))}\n"
f"Code:\n{code}"
)
return {"system": system, "user": "\n".join(parts)}
def _call_llm_api(
self, prompt_data: Dict[str, str], max_tokens: int = 8192, timeout: int = 180
) -> str:
"""Call OpenAI-compatible API (blocking, runs in executor thread)."""
url = f"{self._summary_api_base}/chat/completions"
body = json.dumps(
{
"model": self._summary_model,
"messages": [
{"role": "system", "content": prompt_data["system"]},
{"role": "user", "content": prompt_data["user"]},
],
"max_completion_tokens": max_tokens,
}
).encode("utf-8")
req = urllib.request.Request(
url,
data=body,
headers={
"Content-Type": "application/json",
"Authorization": f"Bearer {self._summary_api_key}",
},
method="POST",
)
try:
with urllib.request.urlopen(req, timeout=timeout) as resp:
data = json.loads(resp.read().decode("utf-8"))
return data["choices"][0]["message"]["content"].strip()
except urllib.error.HTTPError as e:
error_body = e.read().decode("utf-8", errors="replace")[:500]
raise RuntimeError(f"API error {e.code}: {error_body}") from e
except Exception as e:
raise RuntimeError(f"API call failed: {e}") from e
|