First job of the 122b fast2 sequence (2026-07-12/13, MinimalPi, caps ×3
stock clamped at 1800s). Run as SEVEN single-task jobs so the new
between-jobs llama-swap auto-restart (maybe_restart_llama_swap, shipped
the same day, LLAMA_RESTART_MODELS=qwen3.5-122b*) resets the 122b
llama-server RAM ratchet before every task — this job's log shows the
mechanism's first live firing.
reshard-c4-data FAIL at 12m37s agent (finished under its 900s cap — clean fail, not a cut), 2.02M input tokens. The 35b failed this at 48s; the 122b worked at it 15× longer with the same verdict.
Sequence verdict (see the …180346 headless rerun NOTES for the full table): 2/6 — git-multibranch and mailman are FIRST-EVER passes on this box (35b/27b never passed any fast2 task).
llama-local/qwen3.5-122b-a10bagentharnesses.minimal_pi:MinimalPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 0 pass · 1 failmean reward0.00tokens (job total)2,024,579 in / 10,781 outstarted / finished2026-07-12T15:35 / 2026-07-12T15:49wall clock14m33s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | FAIL | 14m33s | 12m37s | 2024579/10781 | 🔍 view |