File size: 33,008 Bytes
8c10cf2 | 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 | #!/usr/bin/env python3
import csv
import re
from pathlib import Path
from typing import Dict, List, Optional
from collections import defaultdict
class TraceNode:
def __init__(
self,
node_type: str,
name: str,
time: Optional[float] = None,
tokens: Optional[Dict[str, int]] = None,
raw_line: str = "",
):
self.type = node_type
self.name = name
self.time = time
self.tokens = tokens or {}
self.raw_line = raw_line
self.children: List["TraceNode"] = []
self.parent: Optional["TraceNode"] = None
self.depth: int = 0
self.in_mcp_subtree: bool = False
def add_child(self, child: "TraceNode") -> None:
child.parent = self
self.children.append(child)
class ExecutionTreeParser:
def __init__(self, md_file_path: str):
self.file_path = Path(md_file_path)
self.model: Optional[str] = None
self.project: Optional[str] = None
self.session_id: str = self.file_path.parent.name
self.root: Optional[TraceNode] = None
def _extract_metadata_from_path(self) -> None:
parts = self.file_path.parts
if "RESULTS" in parts:
idx = parts.index("RESULTS")
if idx + 2 < len(parts):
self.model = parts[idx + 1]
self.project = parts[idx + 2]
@staticmethod
def _parse_tokens(line: str) -> Optional[Dict[str, int]]:
agg_pattern = r"\[∑ tokens: \((\d+)→(\d+) \[REASONING:(\d+), OUTPUT:(\d+)\], total: (\d+)\)"
m = re.search(agg_pattern, line)
if not m:
llm_pattern = (
r"\((\d+)→(\d+) \[REASONING:(\d+), OUTPUT:(\d+)\], total: (\d+)\)"
)
m = re.search(llm_pattern, line)
if not m:
return None
return {
"input": int(m.group(1)),
"output": int(m.group(2)),
"reasoning": int(m.group(3)),
"result": int(m.group(4)),
"total": int(m.group(5)),
}
@staticmethod
def _parse_time(line: str) -> Optional[float]:
m = re.search(r"time:\s*([\d.]+)s", line)
if m:
return float(m.group(1))
m = re.search(r"∑\s*time:\s*([\d.]+)(ms|s)", line)
if m:
val = float(m.group(1))
return val / 1000.0 if m.group(2) == "ms" else val
m = re.search(r"\[([\d.]+)(ms|s)\]", line)
if m:
val = float(m.group(1))
return val / 1000.0 if m.group(2) == "ms" else val
return None
@staticmethod
def _clean_content_line(line: str) -> str:
clean = re.sub(r"^[│├└─\s]+", "", line).strip()
if not clean:
return ""
clean = re.sub(r"^❌\s+", "", clean)
clean = re.sub(r"\s*\(retry\s+\d+\)", "", clean)
clean = re.sub(r"\s*\[RETRY\d+\]", "", clean)
clean = re.sub(r"\s*\[ERROR:[^\]]*\]", "", clean)
return clean.strip()
@staticmethod
def _parse_node_from_content(line: str, raw_line: str) -> Optional[TraceNode]:
if not line:
return None
if line.startswith("[Task Created]"):
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode(
"Task Created", "Task Created", time=time_val, raw_line=raw_line
)
if line.startswith("[Crew Created]"):
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode(
"Crew Created", "Crew Created", time=time_val, raw_line=raw_line
)
if line.startswith("[SPAN]"):
m = re.match(r"\[SPAN\]\s+([^\[]+)", line)
name = m.group(1).strip() if m else "SPAN"
tokens = ExecutionTreeParser._parse_tokens(line)
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode(
"SPAN", name, time=time_val, tokens=tokens, raw_line=raw_line
)
if line.startswith("[Chain]"):
m = re.match(r"\[Chain\]\s+([^\[]+)", line)
name = m.group(1).strip() if m else "Chain"
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode("Chain", name, time=time_val, raw_line=raw_line)
if line.startswith("[AGENT]"):
m = re.match(r"\[AGENT\]\s+(.+?)(?:\s+\[|$)", line)
name = m.group(1).strip() if m else "AGENT"
tokens = ExecutionTreeParser._parse_tokens(line)
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode(
"AGENT", name, time=time_val, tokens=tokens, raw_line=raw_line
)
if line.startswith("[Tool]"):
m = re.match(r"\[Tool\]\s+([^\[]+?)(?:\s+\[|\s+@@@|$)", line)
name = m.group(1).strip() if m else "Tool"
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode("Tool", name, time=time_val, raw_line=raw_line)
if line.startswith("[LLM]"):
m = re.match(r"\[LLM\]\s+([^\(\[]+)", line)
name = m.group(1).strip() if m else "LLM"
tokens = ExecutionTreeParser._parse_tokens(line)
time_val = ExecutionTreeParser._parse_time(line)
return TraceNode(
"LLM", name, time=time_val, tokens=tokens, raw_line=raw_line
)
return None
def _mark_mcp_subtrees(self) -> None:
if not self.root:
return
def dfs(node: TraceNode, in_mcp: bool) -> None:
if node.type == "SPAN" and "mcp" in node.name:
in_mcp = True
node.in_mcp_subtree = in_mcp
for ch in node.children:
dfs(ch, in_mcp)
dfs(self.root, False)
def parse(self) -> Optional[TraceNode]:
if not self.file_path.exists():
return None
text = self.file_path.read_text(encoding="utf-8")
m = re.search(r"## Execution Path Tree.*?```\n(.*?)```", text, re.DOTALL)
if not m:
return None
block = m.group(1)
stack: List[TraceNode] = []
self.root = None
for raw in block.splitlines():
if not raw.strip():
continue
pm = re.match(r"^([│├└─\s]*)", raw)
prefix = pm.group(1) if pm else ""
depth = len(prefix)
clean = self._clean_content_line(raw)
node = self._parse_node_from_content(clean, raw)
if node is None:
continue
node.depth = depth
while stack and stack[-1].depth >= depth:
stack.pop()
if stack:
stack[-1].add_child(node)
else:
if self.root is None:
self.root = node
stack.append(node)
self._extract_metadata_from_path()
self._mark_mcp_subtrees()
return self.root
def iter_nodes(root: TraceNode):
stack = [root]
while stack:
node = stack.pop()
yield node
for ch in reversed(node.children):
stack.append(ch)
def compute_retry_time(root: TraceNode) -> float:
"""Sum time (seconds) of nodes marked as RETRY.
Rules:
- A node is considered a RETRY attempt if its raw_line contains "(retry N)" or "[RETRYN]".
- Only nodes with time are counted; MCP subtrees are skipped.
- Use the node's own time as the cost of that RETRY attempt (do not additionally sum its children).
"""
total = 0.0
retry_pattern = re.compile(r"\(retry\s+\d+\)|\[RETRY\d+\]")
for node in iter_nodes(root):
if node.in_mcp_subtree:
continue
if node.time is None:
continue
if retry_pattern.search(node.raw_line):
total += node.time
return total
def iter_subtree(root: TraceNode):
"""Iterate the subtree rooted at `root` (including `root`)."""
stack = [root]
while stack:
node = stack.pop()
yield node
for ch in reversed(node.children):
stack.append(ch)
def compute_llm_overhead_for_subtree(root: TraceNode) -> float:
"""Compute LLM time (seconds) within the given subtree, reusing the global LLM rules."""
total = 0.0
for node in iter_subtree(root):
if node.in_mcp_subtree:
continue
if node.type == "LLM":
parent = node.parent
if (
parent
and parent.type == "LLM"
and len(parent.children) == 1
and parent.children[0] is node
and parent.tokens
and node.tokens
and parent.tokens.get("total") == node.tokens.get("total")
and parent.time is not None
and node.time is not None
and abs(parent.time - node.time) < 1e-6
):
# AutoGen nested LLM dedup: keep the parent node only
continue
if node.time is not None:
total += node.time
return total
def compute_tool_overhead_for_subtree(root: TraceNode) -> float:
"""Compute Tool time (seconds) within the given subtree, reusing the global Tool rules."""
total = 0.0
for node in iter_subtree(root):
if node.in_mcp_subtree:
continue
if node.type == "Chain" and node.name == "tools":
# LangGraph: use the [Chain] tools container time
p = node.parent
while p is not None and not (p.type == "Chain" and p.name == "LangGraph"):
p = p.parent
if p is not None and node.time is not None:
total += node.time
elif node.type == "Tool":
# CrewAI / AutoGen: sum Tool nodes; Tools under LangGraph are handled by the container above
if _is_under_langgraph_tools(node):
continue
if node.time is not None:
total += node.time
return total
def compute_langgraph_format_output_time_for_subtree(root: TraceNode) -> float:
"""Compute LangGraph [Chain] format_output time (seconds) within the given subtree."""
total = 0.0
for node in iter_subtree(root):
if node.in_mcp_subtree:
continue
if (
node.type == "Chain"
and node.name == "format_output"
and node.time is not None
):
total += node.time
return total
def find_orchestrator(root: TraceNode) -> TraceNode:
for node in iter_nodes(root):
if node.type == "SPAN" and "orchestrator" in node.name:
return node
return root
def compute_llm_overhead(root: TraceNode) -> float:
total = 0.0
for node in iter_nodes(root):
if node.in_mcp_subtree:
continue
if node.type == "LLM":
parent = node.parent
if (
parent
and parent.type == "LLM"
and len(parent.children) == 1
and parent.children[0] is node
and parent.tokens
and node.tokens
and parent.tokens.get("total") == node.tokens.get("total")
and parent.time is not None
and node.time is not None
and abs(parent.time - node.time) < 1e-6
):
continue
if node.time is not None:
total += node.time
return total
def _is_under_langgraph_tools(node: TraceNode) -> bool:
p = node.parent
seen_tools = False
while p is not None:
if p.type == "Chain" and p.name == "tools":
seen_tools = True
if seen_tools and p.type == "Chain" and p.name == "LangGraph":
return True
p = p.parent
return False
def compute_tool_overhead(root: TraceNode) -> float:
total = 0.0
for node in iter_nodes(root):
if node.in_mcp_subtree:
continue
if node.type == "Chain" and node.name == "tools":
p = node.parent
while p is not None and not (p.type == "Chain" and p.name == "LangGraph"):
p = p.parent
if p is not None and node.time is not None:
total += node.time
elif node.type == "Tool":
if _is_under_langgraph_tools(node):
continue
if node.time is not None:
total += node.time
return total
def compute_a2a_overhead(root: TraceNode) -> float:
total = 0.0
for node in iter_nodes(root):
if node.type == "SPAN" and node.name.startswith("a2a_call_"):
if node.time is None:
continue
server = None
for ch in node.children:
if ch.type == "SPAN" and ("server_execution" in ch.name):
server = ch
break
if server is not None and server.time is not None:
diff = node.time - server.time
if diff > 0:
total += diff
return total
def _sum_mcp_time(node: TraceNode) -> float:
total = 0.0
stack = [node]
while stack:
n = stack.pop()
if n is not node and n.in_mcp_subtree and n.time is not None:
total += n.time
for ch in n.children:
stack.append(ch)
return total
def _find_framework_child(server_node: TraceNode) -> Optional[TraceNode]:
for ch in server_node.children:
if ch.type == "Chain" and ch.name == "LangGraph":
return ch
if ch.type == "Chain" and re.match(r"Crew_.*\.kickoff", ch.name):
return ch
if ch.type == "AGENT" and ch.name.startswith("invoke_agent "):
return ch
return None
def compute_server_overhead(root: TraceNode) -> float:
total = 0.0
for node in iter_nodes(root):
if node.type == "SPAN" and "server_execution" in node.name:
if node.time is None:
continue
framework = _find_framework_child(node)
framework_time = (
framework.time if framework and framework.time is not None else 0.0
)
mcp_time = _sum_mcp_time(node)
diff = node.time - framework_time - mcp_time
if diff > 0:
total += diff
return total
def compute_framework_breakdown(root: TraceNode) -> (float, float, float):
"""Compute orchestration overhead for three frameworks: LangGraph / CrewAI kickoff / AutoGen invoke_agent."""
lg_total = 0.0
crew_total = 0.0
autogen_total = 0.0
for node in iter_nodes(root):
if node.in_mcp_subtree or node.time is None:
continue
if node.type == "Chain" and node.name == "LangGraph":
children_time = sum(
(ch.time or 0.0) for ch in node.children if not ch.in_mcp_subtree
)
diff = node.time - children_time
if diff > 0:
lg_total += diff
elif node.type == "Chain" and re.match(r"Crew_.*\.kickoff", node.name):
children_time = sum(
(ch.time or 0.0) for ch in node.children if not ch.in_mcp_subtree
)
diff = node.time - children_time
if diff > 0:
crew_total += diff
elif node.type == "AGENT" and node.name.startswith("invoke_agent "):
children_time = sum(
(ch.time or 0.0) for ch in node.children if not ch.in_mcp_subtree
)
diff = node.time - children_time
if diff > 0:
autogen_total += diff
return lg_total, crew_total, autogen_total
def compute_framework_overhead(root: TraceNode) -> float:
"""Kept for backward compatibility: return the sum of orchestration overhead across frameworks."""
lg_total, crew_total, autogen_total = compute_framework_breakdown(root)
return lg_total + crew_total + autogen_total
def analyze_file(path: Path) -> Optional[Dict[str, float]]:
parser = ExecutionTreeParser(str(path))
root = parser.parse()
if root is None:
return None
orch = find_orchestrator(root)
total_time_s = orch.time if orch.time is not None else None
if total_time_s is None or total_time_s <= 0:
return None
# All internal computations use seconds (s)
llm_s = compute_llm_overhead(root)
tool_s = compute_tool_overhead(root)
a2a_s = compute_a2a_overhead(root)
# Compute per-framework overheads
lg_fw_s, crew_fw_s, autogen_fw_s = compute_framework_breakdown(root)
framework_s = lg_fw_s + crew_fw_s + autogen_fw_s
server_s = compute_server_overhead(root)
retry_s = compute_retry_time(root)
classified_s = llm_s + tool_s + a2a_s + framework_s + server_s
residual_s = total_time_s - classified_s
# Compute ratios in seconds first (unit-independent)
llm_ratio = llm_s / total_time_s
tool_ratio = tool_s / total_time_s
a2a_ratio = a2a_s / total_time_s
framework_ratio = framework_s / total_time_s
server_ratio = server_s / total_time_s
residual_ratio = residual_s / total_time_s
retry_ratio = retry_s / total_time_s if total_time_s > 0 else 0.0
# Convert time to milliseconds (integer ms) for CSV output
def to_ms(x: float) -> int:
return int(round(x * 1000.0))
total_time = to_ms(total_time_s)
llm = to_ms(llm_s)
tool = to_ms(tool_s)
a2a = to_ms(a2a_s)
lg_fw = to_ms(lg_fw_s)
crew_fw = to_ms(crew_fw_s)
autogen_fw = to_ms(autogen_fw_s)
framework = lg_fw + crew_fw + autogen_fw
server = to_ms(server_s)
retry_time = to_ms(retry_s)
classified = llm + tool + a2a + framework + server
residual = total_time - classified
result: Dict[str, float] = {
"model": parser.model or "",
"project": parser.project or "",
"session_id": parser.session_id,
"orchestrator_time": total_time,
"LLM_OVERHEAD": llm,
"Tool_OVERHEAD": tool,
"A2A_OVERHEAD": a2a,
"Framework_OVERHEAD": framework,
"LangGraph_Framework_OVERHEAD": lg_fw,
"CrewAI_Framework_OVERHEAD": crew_fw,
"AutoGen_Framework_OVERHEAD": autogen_fw,
"Server_OVERHEAD": server,
"retry_time_ms": retry_time,
"total_classified": classified,
"residual": residual,
}
result.update(
{
"LLM_ratio": llm_ratio,
"Tool_ratio": tool_ratio,
"A2A_ratio": a2a_ratio,
"Framework_ratio": framework_ratio,
"Server_ratio": server_ratio,
"residual_ratio": residual_ratio,
"retry_ratio_vs_orch": retry_ratio,
}
)
return result
def find_results_root() -> Path:
p = Path(__file__).resolve()
for parent in p.parents:
if parent.name == "RESULTS":
return parent
return p.parent.parent.parent
def collect_execution_paths(results_dir: Path, project_name: str) -> List[Path]:
paths: List[Path] = []
for model_dir in results_dir.iterdir():
if not model_dir.is_dir():
continue
proj_dir = model_dir / project_name / "test_results"
if not proj_dir.exists():
continue
for session_dir in proj_dir.iterdir():
if not session_dir.is_dir():
continue
ep = session_dir / "execution_path.md"
if ep.exists():
paths.append(ep)
paths.sort()
return paths
def write_model_summary(rows: List[Dict[str, float]], out_path: Path) -> None:
"""Aggregate per-run time breakdown results by model and write a summary CSV.
Aggregation:
- For each model:
- Sum total time and each component time.
- Compute component shares as: component_share = total_component / total_orchestrator.
This matches the table style in the document (holistic share instead of averaging per-run shares).
"""
agg = defaultdict(
lambda: {
"count": 0,
"total_orchestrator_time": 0.0,
"total_LLM_OVERHEAD": 0.0,
"total_Tool_OVERHEAD": 0.0,
"total_A2A_OVERHEAD": 0.0,
"total_Framework_OVERHEAD": 0.0,
"total_LangGraph_Framework_OVERHEAD": 0.0,
"total_CrewAI_Framework_OVERHEAD": 0.0,
"total_AutoGen_Framework_OVERHEAD": 0.0,
"total_Server_OVERHEAD": 0.0,
"total_retry_time_ms": 0.0,
"total_classified": 0.0,
"total_residual": 0.0,
}
)
for row in rows:
model = str(row.get("model", ""))
m = agg[model]
m["count"] += 1
m["total_orchestrator_time"] += float(row.get("orchestrator_time", 0.0))
m["total_LLM_OVERHEAD"] += float(row.get("LLM_OVERHEAD", 0.0))
m["total_Tool_OVERHEAD"] += float(row.get("Tool_OVERHEAD", 0.0))
m["total_A2A_OVERHEAD"] += float(row.get("A2A_OVERHEAD", 0.0))
m["total_Framework_OVERHEAD"] += float(row.get("Framework_OVERHEAD", 0.0))
m["total_LangGraph_Framework_OVERHEAD"] += float(
row.get("LangGraph_Framework_OVERHEAD", 0.0)
)
m["total_CrewAI_Framework_OVERHEAD"] += float(
row.get("CrewAI_Framework_OVERHEAD", 0.0)
)
m["total_AutoGen_Framework_OVERHEAD"] += float(
row.get("AutoGen_Framework_OVERHEAD", 0.0)
)
m["total_Server_OVERHEAD"] += float(row.get("Server_OVERHEAD", 0.0))
m["total_retry_time_ms"] += float(row.get("retry_time_ms", 0.0))
m["total_classified"] += float(row.get("total_classified", 0.0))
m["total_residual"] += float(row.get("residual", 0.0))
summary_rows: List[Dict[str, float]] = []
retry_rows: List[Dict[str, float]] = []
for model, m in sorted(agg.items(), key=lambda kv: kv[0]):
total_time = m["total_orchestrator_time"] or 1e-9 # ms, kept for comparison
llm = m["total_LLM_OVERHEAD"]
tool = m["total_Tool_OVERHEAD"]
a2a = m["total_A2A_OVERHEAD"]
framework = m["total_Framework_OVERHEAD"]
lg_fw = m["total_LangGraph_Framework_OVERHEAD"]
crew_fw = m["total_CrewAI_Framework_OVERHEAD"]
autogen_fw = m["total_AutoGen_Framework_OVERHEAD"]
server = m["total_Server_OVERHEAD"]
residual = m["total_residual"]
retry_total = m["total_retry_time_ms"]
# Total component time (LLM + Tool + A2A + Framework + Server + residual), in ms
components_time = llm + tool + a2a + framework + server + residual
denom = components_time or 1e-9
# Component shares (use components_time as denominator so the sum is ~1)
llm_share = llm / denom
tool_share = tool / denom
a2a_share = a2a / denom
framework_share = framework / denom
lg_share = lg_fw / denom
crew_share = crew_fw / denom
autogen_share = autogen_fw / denom
server_share = server / denom
residual_share = residual / denom
# RETRY share relative to total orchestrator time (in ms)
retry_share_vs_orch = retry_total / (total_time or 1e-9)
# Sum of major component shares (sanity check, should be close to 1)
sum_component_shares = (
llm_share
+ tool_share
+ a2a_share
+ framework_share
+ server_share
+ residual_share
)
summary_rows.append(
{
"model": model,
"count": m["count"],
"total_orchestrator_time": total_time,
"total_LLM_OVERHEAD": llm,
"total_Tool_OVERHEAD": tool,
"total_A2A_OVERHEAD": a2a,
"total_Framework_OVERHEAD": framework,
"total_LangGraph_Framework_OVERHEAD": lg_fw,
"total_CrewAI_Framework_OVERHEAD": crew_fw,
"total_AutoGen_Framework_OVERHEAD": autogen_fw,
"total_Server_OVERHEAD": server,
"total_retry_time_ms": retry_total,
"total_classified": m["total_classified"],
"total_residual": residual,
"total_components_time": components_time,
# Component shares relative to total time
"LLM_share": llm_share,
"Tool_share": tool_share,
"A2A_share": a2a_share,
"Framework_share": framework_share,
"LangGraph_Framework_share": lg_share,
"CrewAI_Framework_share": crew_share,
"AutoGen_Framework_share": autogen_share,
"Server_share": server_share,
"residual_share": residual_share,
"retry_share_vs_orch": retry_share_vs_orch,
"sum_component_shares": sum_component_shares,
}
)
# RETRY-focused compact row: written to retry_breakdown_summary_by_model.csv
retry_rows.append(
{
"model": model,
"count": m["count"],
"total_orchestrator_time_ms": total_time,
"total_retry_time_ms": retry_total,
"retry_share_vs_orch": retry_share_vs_orch,
}
)
if not summary_rows:
return
# Main per-model summary table
fieldnames = list(summary_rows[0].keys())
with out_path.open("w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=fieldnames)
writer.writeheader()
writer.writerows(summary_rows)
# Dedicated RETRY cost summary table (one row per model)
if retry_rows:
retry_out_path = out_path.with_name("retry_breakdown_summary_by_model.csv")
retry_fieldnames = list(retry_rows[0].keys())
with retry_out_path.open("w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=retry_fieldnames)
writer.writeheader()
writer.writerows(retry_rows)
def _infer_langgraph_agent_name(node: TraceNode) -> str:
"""Walk upwards from a LangGraph container to find the nearest SPAN as the business agent name."""
p = node.parent
while p is not None:
if p.type == "SPAN":
name = p.name
# Remove common server_execution suffix
name = re.sub(r"_server_execution$", "", name)
return name
p = p.parent
return "LangGraph"
def _normalize_crewai_agent_name(name: str) -> str:
"""Normalize CrewAI agent names.
Handle variants like "Senior Candidate Evaluator._execute_core" or
"Senior Candidate Evaluator._execute_core]" and normalize to
"Senior Candidate Evaluator".
"""
# Strip trailing _execute_core or _execute_core]
name = re.sub(r"\._execute_core\]?$", "", name)
return name.strip()
def collect_agent_llm_tool_breakdown(exec_paths: List[Path]) -> List[Dict[str, float]]:
"""Aggregate LLM/Tool time (ms) by (model, framework, agent_name).
- LangGraph: treat [Chain] LangGraph as the container; sum LLM/Tool and format_output in its subtree.
- CrewAI: use direct child [AGENT] xxx._execute_core under [Chain] Crew***.kickoff as the container.
- AutoGen: use [AGENT] invoke_agent xxx as the container.
"""
agg = defaultdict(
lambda: {
"llm_s": 0.0,
"tool_s": 0.0,
"format_output_s": 0.0,
"occurrences": 0,
}
)
for ep in exec_paths:
parser = ExecutionTreeParser(str(ep))
root = parser.parse()
if root is None:
continue
model = parser.model or ""
for node in iter_nodes(root):
if node.in_mcp_subtree:
continue
# Treat the LangGraph container as one agent
if node.type == "Chain" and node.name == "LangGraph":
framework = "LangGraph"
agent_name = _infer_langgraph_agent_name(node)
llm_s = compute_llm_overhead_for_subtree(node)
tool_s = compute_tool_overhead_for_subtree(node)
fmt_s = compute_langgraph_format_output_time_for_subtree(node)
if llm_s == 0.0 and tool_s == 0.0 and fmt_s == 0.0:
continue
key = (model, framework, agent_name)
m = agg[key]
m["llm_s"] += llm_s
m["tool_s"] += tool_s
m["format_output_s"] += fmt_s
m["occurrences"] += 1
# CrewAI: each direct child AGENT under kickoff is treated as an agent
elif node.type == "Chain" and re.match(r"Crew_.*\.kickoff", node.name):
for ch in node.children:
if ch.in_mcp_subtree or ch.type != "AGENT":
continue
framework = "CrewAI"
agent_name = _normalize_crewai_agent_name(ch.name)
llm_s = compute_llm_overhead_for_subtree(ch)
tool_s = compute_tool_overhead_for_subtree(ch)
fmt_s = 0.0
if llm_s == 0.0 and tool_s == 0.0:
continue
key = (model, framework, agent_name)
m = agg[key]
m["llm_s"] += llm_s
m["tool_s"] += tool_s
m["format_output_s"] += fmt_s
m["occurrences"] += 1
# AutoGen: invoke_agent is treated as an agent
elif node.type == "AGENT" and node.name.startswith("invoke_agent "):
framework = "AutoGen"
agent_name = node.name[len("invoke_agent ") :]
llm_s = compute_llm_overhead_for_subtree(node)
tool_s = compute_tool_overhead_for_subtree(node)
fmt_s = 0.0
if llm_s == 0.0 and tool_s == 0.0:
continue
key = (model, framework, agent_name)
m = agg[key]
m["llm_s"] += llm_s
m["tool_s"] += tool_s
m["format_output_s"] += fmt_s
m["occurrences"] += 1
rows: List[Dict[str, float]] = []
for (model, framework, agent_name), st in sorted(
agg.items(), key=lambda kv: (kv[0][0], kv[0][1], kv[0][2])
):
llm_ms = int(round(st["llm_s"] * 1000.0))
tool_ms = int(round(st["tool_s"] * 1000.0))
fmt_ms = int(round(st["format_output_s"] * 1000.0))
total_ms = llm_ms + tool_ms + fmt_ms
denom = total_ms or 1e-9
rows.append(
{
"model": model,
"framework": framework,
"agent_name": agent_name,
"occurrences": st["occurrences"],
"total_llm_time_ms": llm_ms,
"total_tool_time_ms": tool_ms,
"total_format_output_time_ms": fmt_ms,
"total_agent_llm_tool_time_ms": total_ms,
"llm_share_in_agent": llm_ms / denom,
"tool_share_in_agent": tool_ms / denom,
"format_output_share_in_agent": fmt_ms / denom,
}
)
return rows
def main() -> None:
results_dir = find_results_root()
project_name = "RecruitmentAssistant-H_A2A"
exec_paths = collect_execution_paths(results_dir, project_name)
rows: List[Dict[str, float]] = []
for ep in exec_paths:
metrics = analyze_file(ep)
if metrics is not None:
rows.append(metrics)
out_dir = Path(__file__).resolve().parent
per_run_path = out_dir / "performance_breakdown_summary.csv"
per_model_path = out_dir / "performance_breakdown_summary_by_model.csv"
agent_path = out_dir / "agent_llm_tool_breakdown_by_model.csv"
if rows:
# Per-run detailed table (one row per run)
fieldnames = list(rows[0].keys())
with per_run_path.open("w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=fieldnames)
writer.writeheader()
writer.writerows(rows)
print(f"written {len(rows)} rows to {per_run_path}")
# Per-model aggregated summary table
write_model_summary(rows, per_model_path)
print(f"written model summary to {per_model_path}")
# Agent-level LLM/Tool breakdown (grouped by model × agent)
agent_rows = collect_agent_llm_tool_breakdown(exec_paths)
if agent_rows:
agent_fieldnames = list(agent_rows[0].keys())
with agent_path.open("w", newline="", encoding="utf-8") as f:
writer = csv.DictWriter(f, fieldnames=agent_fieldnames)
writer.writeheader()
writer.writerows(agent_rows)
print(f"written agent LLM/Tool breakdown to {agent_path}")
else:
print("no valid execution_path.md found")
if __name__ == "__main__":
main()
|