fast2 27b, tangled 4-knob night batch — 5/6; fix-ocaml-gc PASSES in 31m with all knobs on

Night batch 2026-07-08, SubagentsPi, loop_guard=on strip_thinking=on keep_tool_results=10 subagent_nudges=on. Baseline: 07-07 run (fast2__qwen3.6-27b__20260707-213441) went 6/6.

Result 5/6 (mean 0.83): only git-multibranch regressed PASS→FAIL, and it failed FAST (1m29s agent, 25 turns) — the known flaky task for this model (mixed history on the board), not an obvious knob effect. fix-ocaml-gc PASSED in 31m38s (vs 14m30s on 07-07 — slower but solved; same night the 35b sibling burned 1h28m on it and was cancelled, see its NOTES). reshard-c4-data PASSED at 53m23s wall (51m34s agent), the long pole of the run's 2h23m.

Zero loop-guard blocks in any root; 3 saturated-thinking turns only in reshard-c4-data. 27b keeps tolerating the context strip that collapsed 35b.

modelllama-local/qwen3.6-27bagentharnesses.subagents_pi:SubagentsPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials6 of 6 — 5 pass · 1 failmean reward0.83tokens (job total)4,812,162 in / 114,104 outstarted / finished2026-07-08T05:10 / 2026-07-08T07:34wall clock2h23m

fast2qwen3.6-27b20260708-051015

fast2 27b, tangled 4-knob night batch — 5/6; fix-ocaml-gc PASSES in 31m with all knobs on

Run details

Tasks

financial-document-processor — 1/1 passed

fix-ocaml-gc — 1/1 passed

git-multibranch — 0/1 passed

headless-terminal — 1/1 passed

mailman — 1/1 passed

reshard-c4-data — 1/1 passed

fast2__qwen3.6-27b__20260708-051015

fast2 27b, tangled 4-knob night batch — 5/6; fix-ocaml-gc PASSES in 31m with all knobs on

Run details

Tasks

financial-document-processor — 1/1 passed

fix-ocaml-gc — 1/1 passed

git-multibranch — 0/1 passed

headless-terminal — 1/1 passed

mailman — 1/1 passed

reshard-c4-data — 1/1 passed

fast2qwen3.6-27b20260708-051015