fast 35b, tangled 4-knob night batch — COLLAPSE to 3/9; context-strip amnesia is the prime suspect
Night batch 2026-07-08, SubagentsPi, four knobs flipped together vs 07-07:
loop_guard=on strip_thinking=on keep_tool_results=10 subagent_nudges=on.
NOT a clean A/B (untangling item in AGENT_TODOS.md).
Result 3/9 (mean 0.33) vs 7/9 on the 07-07 SubagentsPi baseline
(fast__qwen3.6-35b-a3b__20260707-001626). Regressions: fix-git,
cancel-async-tasks, sanitize-git-repo, sparql-university all PASS→FAIL. All
regressed trials finished NATURALLY (stopReason=stop, declared done) and the
verifier failed them — a quality regression, not timeouts.
Shape of the regression = the amnesia the context-strip ledger item
predicted: cancel-async-tasks read /app/run.py 6 times, re-ran
near-identical test one-liners in 3–5x clusters, turns ~10→55
(agent 35s→6m19s), then declared done with a ✅-checklist the verifier
rejected. Everything got slower (sparql 3m25s→13m25s, sanitize 2m03s→7m02s)
and input tokens ballooned (cancel-async 74k→872k — more turns, not bigger
turns). Loop guard: 2 blocks total across all 9 roots (quiet, harmless).
Contrast: the sibling 27b run with identical knobs went 7/9.
Verdict feeding AGENT_TODOS.md: keep STRIP_THINKING/KEEP_TOOL_RESULTS
defaults OFF; rerun 35b with the strip knobs off to confirm they (not
loop_guard/nudges) are the regressor.