Datasets:
File size: 13,601 Bytes
7bdf36e | 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 | """Stage 1: stream one rollout file, split it into turns, keep what later stages need.
Everything is read line by line (never the whole file). Every line is JSON-parsed so that the
malformed / scalar counts are exact. Raw content is only *parsed*, never executed or followed.
"""
import json
import os
import re
from collections import Counter
DROP_RESPONSE_ITEMS = ("reasoning", "web_search_call", "tool_search_call", "tool_search_output")
_CTX_PREFIXES = (
("# AGENTS.md instructions", "agents_md"),
("<environment_context>", "environment_context"),
("<turn_aborted>", "turn_aborted_note"),
("<subagent_notification>", "subagent_notification"),
("<permissions instructions>", "permissions"),
("<model_switch>", "model_switch"),
("<collaboration_mode>", "collaboration_mode"),
("<skill>", "skill"),
("<user_shell_command>", "user_shell_command"),
("<INSTRUCTIONS>", "instructions"),
)
def classify_injected(role, text):
t = (text or "").lstrip()[:60]
for pre, name in _CTX_PREFIXES:
if t.startswith(pre):
return f"{role}:{name}"
if role == "user":
return "user:copy_or_other"
return f"{role}:other"
def message_text(content):
"""Join the text parts of a response_item message content list."""
if isinstance(content, str):
return content
parts = []
if isinstance(content, list):
for c in content:
if isinstance(c, dict):
tx = c.get("text")
if isinstance(tx, str):
parts.append(tx)
return "\n".join(parts) if len(parts) != 1 else parts[0]
def uuid7_ms(u):
"""Milliseconds embedded in a UUIDv7 string, or None."""
try:
h = u.replace("-", "")
if len(h) == 32 and h[12] == "7":
return int(h[:12], 16)
except Exception:
pass
return None
def new_instance(turn_id, ts, started_at, lineno, goal):
return {
"turn_id": turn_id, "ts": ts, "started_at": started_at, "line": lineno,
"ctx": None, "events": [], "call_ids": set(), "complete": False, "last_msg": None,
"aborted": False, "n_compact": 0, "goal": goal, "goal_in_turn": None, "drop": Counter(),
"n_user_events": 0, "user_images": 0, "dup_calls": 0,
}
def scan_file(args):
path, rel = args
st = Counter()
size = os.path.getsize(path)
file_session = None
sm_count = 0
sm_ids = Counter()
instances = []
cur = None
last_by_tid = {}
last_goal = None
orphan_calls = 0
orphan_buf = [] # events seen before any task_started (head of the file destroyed): kept for manual recovery
orphan_events = Counter()
orphan_turns = [] # task_complete seen for a turn whose start was lost
ctx_models = Counter()
if size == 0:
st["empty_file"] += 1
return {"rel": rel, "size": size, "stats": st, "session": None, "sm_count": 0, "foreign_sm": 0, "instances": [],
"orphan_turns": [], "orphan_events": orphan_events, "pairs": []}
def open_cur():
return cur is not None and not cur["complete"]
with open(path, encoding="utf-8-sig", errors="replace", newline="\n") as f:
for lineno, line in enumerate(f):
st["lines"] += 1
s = line.strip()
if not s:
st["blank_lines"] += 1
continue
try:
rec = json.loads(s)
except (ValueError, RecursionError):
st["malformed_lines"] += 1
if "\x00" in s:
st["malformed:contains_NUL_bytes"] += 1
elif "\ufffd" in s:
st["malformed:contains_U+FFFD_(undecodable_bytes)"] += 1
elif s[:1] in "{[":
st["malformed:truncated_json"] += 1
else:
st["malformed:non_json_text_fragment"] += 1
continue
if not isinstance(rec, dict):
st["scalar_lines"] += 1
continue
t = rec.get("type")
p = rec.get("payload")
ts = rec.get("timestamp")
if not isinstance(p, dict):
st[f"non_dict_payload:{t}"] += 1
continue
pt = p.get("type")
st["records"] += 1
if t == "session_meta":
sm_count += 1
sm_ids[p.get("id")] += 1
if file_session is None:
file_session = {k: p.get(k) for k in
("id", "cwd", "thread_source", "agent_role", "agent_nickname", "timestamp")}
src = p.get("source")
file_session["source_kind"] = list(src.keys())[0] if isinstance(src, dict) and src else src
continue
if t == "turn_context":
tid = p.get("turn_id")
st["turn_context"] += 1
if tid is None:
continue
if cur is None or cur["turn_id"] != tid:
inst = last_by_tid.get(tid)
if inst is None:
inst = new_instance(tid, ts, None, lineno, last_goal)
instances.append(inst)
last_by_tid[tid] = inst
st["ctx_without_start"] += 1
cur = inst
if cur["ctx"] is None:
cur["ctx"] = {"model": p.get("model"), "effort": p.get("effort"), "cwd": p.get("cwd")}
ctx_models[p.get("model")] += 1
continue
if t == "compacted":
st["compacted_records"] += 1
if open_cur():
cur["events"].append(("K",))
cur["n_compact"] += 1
else:
st["compacted_outside_turn"] += 1
continue
if t == "event_msg":
if pt == "task_started":
tid = p.get("turn_id")
inst = new_instance(tid, ts, p.get("started_at"), lineno, last_goal)
if tid in last_by_tid:
st["restarted_turn_id_in_file"] += 1
instances.append(inst)
last_by_tid[tid] = inst
cur = inst
orphan_calls = 0
elif pt == "task_complete":
tid = p.get("turn_id")
inst = last_by_tid.get(tid)
if inst is None:
st["task_complete_unknown_turn"] += 1
o = {"turn_id": tid, "orphan_calls": orphan_calls}
if orphan_calls > 0:
ca, dm = p.get("completed_at"), p.get("duration_ms")
o.update({"events": orphan_buf, "last_msg": p.get("last_agent_message"), "ts": ts,
"completed_at": ca, "duration_ms": dm})
orphan_turns.append(o)
orphan_buf = []
orphan_calls = 0
else:
inst["complete"] = True
inst["last_msg"] = p.get("last_agent_message")
elif pt == "user_message":
if open_cur():
cur["events"].append(("U", p.get("message")))
cur["n_user_events"] += 1
if p.get("images") or p.get("local_images"):
cur["user_images"] += 1
else:
orphan_events["event_msg/user_message"] += 1
if cur is None:
orphan_buf.append(("U", p.get("message")))
elif pt == "thread_goal_updated":
st["goal_events"] += 1
g = p.get("goal")
obj = g.get("objective") if isinstance(g, dict) else None
if isinstance(obj, str) and obj.strip():
last_goal = obj
if open_cur() and cur["goal_in_turn"] is None:
cur["goal_in_turn"] = obj
elif pt == "turn_aborted":
if cur is not None:
cur["aborted"] = True
else:
if open_cur():
cur["drop"][f"event_msg/{pt}"] += 1
continue
if t == "response_item":
if not open_cur():
if cur is None:
orphan_events[f"response_item/{pt}"] += 1
if pt in ("function_call", "custom_tool_call"):
orphan_calls += 1
orphan_buf.append(("C", p.get("call_id"), p.get("name"),
p.get("arguments") if pt == "function_call" else p.get("input"),
"f" if pt == "function_call" else "c"))
elif pt in ("function_call_output", "custom_tool_call_output"):
orphan_buf.append(("O", p.get("call_id"), p.get("output"), "f" if pt == "function_call_output" else "c"))
elif pt == "message" and p.get("role") == "assistant":
tx = message_text(p.get("content"))
if tx and tx.strip():
orphan_buf.append(("A", tx))
else:
st["response_item_after_complete"] += 1
continue
if pt == "message":
role = p.get("role")
text = message_text(p.get("content"))
if role == "assistant":
if text and text.strip():
cur["events"].append(("A", text))
cur["drop"]["assistant_empty"] += 0 if (text and text.strip()) else 1
else:
cur["drop"]["injected:" + classify_injected(role, text)] += 1
elif pt in ("function_call", "custom_tool_call"):
cid = p.get("call_id")
if cid in cur["call_ids"]:
cur["dup_calls"] += 1
continue
cur["call_ids"].add(cid)
if pt == "function_call":
cur["events"].append(("C", cid, p.get("name"), p.get("arguments"), "f"))
else:
cur["events"].append(("C", cid, p.get("name"), p.get("input"), "c"))
elif pt in ("function_call_output", "custom_tool_call_output"):
cur["events"].append(("O", p.get("call_id"), p.get("output"), "f" if pt == "function_call_output" else "c"))
elif pt in DROP_RESPONSE_ITEMS:
cur["drop"][pt] += 1
else:
cur["drop"][f"response_item/{pt}"] += 1
continue
# anything else (unknown top-level type)
st[f"other_top_type:{t}"] += 1
# ---- finalize instances -------------------------------------------------
# "Own" turn = belongs to the file's own session. A forked subagent file embeds a (thinned) copy of the parent
# history behind extra session_meta records that carry the PARENT id. Parent turns were created before the
# child session existed, so their UUIDv7 time is earlier than the child's session id; non-v7 ids only occur
# in the embedded history. Files whose extra session_meta records all repeat the own id (resumes) are all own.
own_id = (file_session or {}).get("id")
foreign_sm = sm_count - sm_ids.get(own_id, 0) if own_id else 0
sess_ms = uuid7_ms(own_id) if own_id else None
out_instances = []
for inst in instances:
if foreign_sm == 0:
own = True
elif sess_ms is not None and inst["turn_id"]:
tms = uuid7_ms(inst["turn_id"])
own = tms is not None and tms >= sess_ms - 1000
else:
own = inst["ctx"] is not None
passing = inst["complete"] and len(inst["call_ids"]) >= 1
first_user = None
for ev in inst["events"]:
if ev[0] == "U":
first_user = ev[1]
break
d = {
"turn_id": inst["turn_id"], "ts": inst["ts"], "started_at": inst["started_at"], "line": inst["line"],
"ctx": inst["ctx"], "complete": inst["complete"], "last_msg": inst["last_msg"],
"n_calls": len(inst["call_ids"]), "n_compact": inst["n_compact"], "own": own,
"goal": inst["goal"], "goal_in_turn": inst["goal_in_turn"], "first_user": first_user,
"n_user_events": inst["n_user_events"], "aborted": inst["aborted"],
"user_images": inst["user_images"], "dup_calls": inst["dup_calls"],
"drop": dict(inst["drop"]), "passing": passing,
"events": inst["events"] if passing else None,
}
out_instances.append(d)
# only the first_user text of own instances is needed for brief/report pairs; passing instances carry events
return {"rel": rel, "size": size, "stats": st, "session": file_session, "sm_count": sm_count,
"foreign_sm": foreign_sm, "instances": out_instances, "orphan_turns": orphan_turns, "orphan_events": orphan_events,
"ctx_models": ctx_models, "tail_orphan_calls": orphan_calls}
|