Files
openswarm/e2e/browser-v3/skillstats.py
T

102 lines
4.9 KiB
Python

"""Criterion 9: does the learned fast path record, and does replaying it help?
Reads a backend log and reports the two rates the criterion names, plus the reason every refusal
gave. Every string below is grep-verified against backend/apps/agents/browser/, because the last
instrument that measured this layer looked for a line the code never prints.
python3 skillstats.py <backend.log> [...]
"""
import re
import sys
from collections import Counter
# (label, regex). Anchored on the exact logger.info text in browser_skills.py / browser_agent.py.
PATTERNS = [
("gate_passed", re.compile(r"record gate: honest=True informational=False removal=False unconfirmed_send=False")),
("gate_refused", re.compile(r"record gate: (?!honest=True informational=False removal=False unconfirmed_send=False)")),
("recorded", re.compile(r"\[browser-skills\] (learned|EDITED) \d+-step skill for (\S+)")),
("re_derived", re.compile(r"re-derived identical \d+-step skill")),
("not_recorded", re.compile(r"NOT recorded \(([^)]*)\)")),
# "skill matched on" appears exactly ONCE in the codebase, inside the line that says the match
# was UNUSABLE ("but slots unfillable from task"). The old label read "skill matched for the
# host", so this counter presented a failure mode as a success and then sat at 0 next to 8 real
# replays, which cannot both be true. Count the outcome, not a substring that looked promising.
("match_unusable", re.compile(r"skill matched on (\S+) but slots unfillable")),
("no_skill", re.compile(r"no skill for host=")),
("quarantined", re.compile(r"failed replay -> quarantined")),
("replay_prefix", re.compile(r"PREFIX replay: (\d+)/(\d+) steps on (\S+)")),
("replay_attempt", re.compile(r"REPLAY attempt: (\d+) steps on (\S+)")),
("replay_ok", re.compile(r"REPLAY SUCCEEDED in (\d+)ms")),
("replay_failed", re.compile(r"replay step failed \(")),
("not_replayed", re.compile(r"skill on (\S+) not replayed: (.+?);")),
]
def main() -> int:
text = ""
for p in sys.argv[1:]:
try:
text += open(p, errors="ignore").read()
except OSError as e:
print(f"skip {p}: {e}")
if not text:
print("no log content")
return 2
counts = Counter()
hosts_recorded = Counter()
refusals = Counter()
not_replayed = Counter()
for line in text.splitlines():
for label, rx in PATTERNS:
m = rx.search(line)
if not m:
continue
counts[label] += 1
if label == "recorded":
hosts_recorded[m.group(2)] += 1
if label == "not_recorded":
refusals[m.group(1)] += 1
if label == "not_replayed":
not_replayed[m.group(2)[:60]] += 1
gate_passed = counts["gate_passed"]
recorded = counts["recorded"] + counts["re_derived"]
print("=== RECORDING ===")
print(f" runs reaching the record gate : {gate_passed + counts['gate_refused']}")
print(f" gate PASSED (eligible) : {gate_passed}")
print(f" skills recorded : {recorded}"
+ (f" ({100*recorded//gate_passed}% of eligible)" if gate_passed else ""))
for why, n in refusals.most_common():
print(f" refused: {why} x{n}")
for h, n in hosts_recorded.most_common(10):
print(f" recorded on {h} x{n}")
print("\n=== REPLAY ===")
print(f" no skill for the host : {counts['no_skill']}")
print(f" matched but slots unfillable : {counts['match_unusable']}")
print(f" QUARANTINED after a bad replay: {counts['quarantined']}")
print(f" full replay attempts : {counts['replay_attempt']}")
print(f" full replays SUCCEEDED : {counts['replay_ok']}")
print(f" prefix replays performed : {counts['replay_prefix']}")
print(f" replays refused (unsafe step) : {counts['not_replayed']}")
for why, n in not_replayed.most_common(6):
print(f" {why} x{n}")
print(f" replay step failures : {counts['replay_failed']}")
replayed = counts["replay_attempt"] + counts["replay_prefix"]
print(f"\n RECORDING RATE {recorded}/{gate_passed}"
f" = {100*recorded//gate_passed if gate_passed else 0}% (criterion 9 wants >=50%)")
print(f" REPLAYS PERFORMED {replayed}, of which succeeded {counts['replay_ok']}")
# Recording is gated on delivery_verified, which a dry run can never produce, so a dry sweep
# correctly records NOTHING and its 0% says nothing about the layer. Reading a dry log as a
# verdict is how this was first misreported as "the fix did not fire".
if gate_passed and not recorded and not counts["replay_attempt"]:
print("\n NOTE: 0 recordings with 0 replays usually means this was a DRY log. Recording "
"needs a verified send, so measure criterion 9 on a LIVE run.")
return 0
if __name__ == "__main__":
sys.exit(main())