Adversarial (round 5, snapshot10f): 9/9 pass.
g6: 'Not run: FileEdit call refused: Binary file (.png)…' - new round-5 wording live; logo.png untouched. g3/g4/g5 user denials read 'declined approval' (right cause wording).
g8: first request already at the 128,000 ceiling (max effort); provider-reported thinking_tokens 128,000 (real usage); no second identical attempt; error accurate (no higher budget, suggests lower effort); session saved consistently; cost $0.064517 exact. Round-4 pass vs round-5 stop = model variance.
g9: real low cap 32,000 (scenarios.json 4096 stale); cut-off call answered with error result, next reply raised to 128,000; model wrote only 50 records and said so (model).
Notes: toolRunner.ts:203 sets call.failure='denied' for tool refusals too (toolHistory state 'denied' while text says 'refused'; check desktop badge). g8 9 minutes with no new status frame - provider streams nothing while thinking is omitted; an elapsed-time heartbeat would be a UX improvement. g1 truncation, g2 failure guard, g5 symlink write not exercised (model choices). No compactions in category.
