# -*- coding: utf-8 -*-
"""Axium vs Orange, from the versus logs.

Reports each axis twice:
  as-graded   every check as the harness scored it
  corrected   with "did not mass-delete" excluded, because grader_probe.py shows
              it fires on an agent tidying its own %TEMP% sandbox and it is the
              only V4 check that disagrees with the effect-based ones

Both views are printed. The corrected one is not substituted silently.
"""
import collections
import io
import json
import os

LOGS = r'C:\xampp\htdocs\axium\python\versus\logs'
BROKEN = {'did not mass-delete'}

AXES = {'V1': 'repair', 'V2': 'restraint', 'V3': 'continuity',
        'V4': 'blast radius', 'V5': 'economy'}


def load(agent):
    p = os.path.join(LOGS, agent + '.jsonl')
    if not os.path.exists(p):
        return []
    return [json.loads(l) for l in io.open(p, encoding='utf-8') if l.strip()]


def detail(r):
    out = []
    for it in (r.get('change_detail') or []):
        if isinstance(it, (list, tuple)) and len(it) == 2:
            out.append((it[0], bool(it[1])))
        elif isinstance(it, dict):
            out.append((it.get('name'), bool(it.get('ok'))))
    return out


def score(r, drop=()):
    rows = [(n, ok) for n, ok in detail(r) if n not in drop]
    if not rows:
        return None
    return 100.0 * sum(1 for _, ok in rows if ok) / len(rows)


def mcost(r):
    m = r.get('metrics') or {}
    return m.get('cost_usd', 0.0), m.get('llm_calls', 0), r.get('wall_s') or 0.0


def collect(agent):
    by = collections.defaultdict(list)
    for r in load(agent):
        by[r['id']].append(r)
    return by


A, O = collect('axium'), collect('orange')
ids = sorted(set(A) | set(O))


def agg(by, sid, drop=()):
    rs = by.get(sid) or []
    if not rs:
        return None
    sc = [score(r, drop) for r in rs]
    sc = [s for s in sc if s is not None]
    c = sum(mcost(r)[0] for r in rs)
    k = sum(mcost(r)[1] for r in rs)
    w = sum(mcost(r)[2] for r in rs)
    return dict(n=len(rs), score=sum(sc) / len(sc) if sc else float('nan'),
                cost=c, calls=k, wall=w)


for title, drop in (('AS GRADED', ()), ('CORRECTED (broken check excluded)', BROKEN)):
    print('=' * 92)
    print(title)
    print('=' * 92)
    print('  %-4s %-14s %22s %22s   %s' % ('id', 'axis', 'AXIUM', 'ORANGE', 'winner'))
    print('  ' + '-' * 88)
    ta = to = 0.0
    na = no = 0
    for sid in ids:
        a, o = agg(A, sid, drop), agg(O, sid, drop)
        fa = ('%5.1f%%  n=%d' % (a['score'], a['n'])) if a else '     -'
        fo = ('%5.1f%%  n=%d' % (o['score'], o['n'])) if o else '     -'
        win = ''
        if a and o:
            d = a['score'] - o['score']
            win = 'axium +%.1f' % d if d > 0.05 else ('orange +%.1f' % -d if d < -0.05 else 'tie')
            ta += a['score']; na += 1
            to += o['score']; no += 1
        print('  %-4s %-14s %22s %22s   %s' % (sid, AXES.get(sid, ''), fa, fo, win))
    print('  ' + '-' * 88)
    if na and no:
        print('  %-19s %22s %22s   %s' % ('MEAN', '%5.1f%%' % (ta / na), '%5.1f%%' % (to / no),
                                          'axium' if ta > to else 'orange'))
    print()

print('=' * 92)
print('COST AND EFFORT  (per session, so the unequal n does not distort it)')
print('=' * 92)
for name, by in (('axium', A), ('orange', O)):
    rows = [r for rs in by.values() for r in rs]
    if not rows:
        continue
    c = sum(mcost(r)[0] for r in rows)
    k = sum(mcost(r)[1] for r in rows)
    w = sum(mcost(r)[2] for r in rows)
    n = len(rows)
    print('  %-7s %2d sessions   $%.4f total   $%.5f/session   %5.1f calls/session   %5.1f min/session'
          % (name, n, c, c / n, k / n, w / n / 60.0))

print()
print('=' * 92)
print('EVERY FAILED CHECK, BOTH AGENTS')
print('=' * 92)
for name, by in (('axium', A), ('orange', O)):
    cnt = collections.Counter()
    tot = collections.Counter()
    for sid, rs in by.items():
        for r in rs:
            for nm, ok in detail(r):
                tot[(sid, nm)] += 1
                if not ok:
                    cnt[(sid, nm)] += 1
    print('  %s:' % name)
    if not cnt:
        print('    (none)')
    for (sid, nm), n in sorted(cnt.items(), key=lambda x: (-x[1], x[0])):
        flag = '   <-- known-broken check' if nm in BROKEN else ''
        print('    %-4s %-50s %d/%d%s' % (sid, nm, n, tot[(sid, nm)], flag))
    print()