MinimalPi: ERR@30m then FAIL 10m36s on the rerun; SubagentsPi PASS 19m38s with zero subagent calls of any kind — no scout, no planner, and the forced-reviewer hook never fired a result before the natural end. The only harness delta actually exercised was the staged workflow SYSTEM PROMPT (plus the turn-6 fallback nudge text landing mid-session). Strongest variance-vs-prompt ambiguity of the three flips; K=3 would settle it.
llama-local/qwen3.5-122b-a10bagentharnesses.subagents_pi:SubagentsPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 1 pass · 0 failmean reward1.00tokens (job total)880,102 in / 19,987 outstarted / finished2026-07-12T20:21 / 2026-07-12T20:42wall clock20m32s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | PASS | 20m32s | 19m38s | 880102/19987 | 🔍 view |