122b SubagentsPi never-passed sequence 1/8 — cancel-async-tasks FAIL (same assertion as MinimalPi)

First of 8 single-task SubagentsPi jobs on the 122b's never-passed tasks (user hypothesis: better instruction-following → gains from plan/review). FAIL 6m31s; zero spontaneous delegation, only the forced reviewer; verifier 5/6 with the identical single SIGINT-cleanup assertion MinimalPi missed. Full analysis in COMMENTS.md. Sequence verdict: 3/8 flips, all fast2 — see fast2__qwen3.5-122b-a10b__20260712-205638/NOTES.md.

💬 1 analyst comment inline below (from runs/fast__qwen3.5-122b-a10b__20260712-183411/COMMENTS.md).

Run details

modelllama-local/qwen3.5-122b-a10bagentharnesses.subagents_pi:SubagentsPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 0 pass · 1 failmean reward0.00tokens (job total)93,368 in / 2,342 outstarted / finished2026-07-12T18:34 / 2026-07-12T18:41wall clock7m21s

Tasks

cancel-async-tasks — 0/1 passed

💬 analyst comment

Zero spontaneous delegation: the root wrote /app/run.py itself on turn 1, smoke-tested only the happy path (5 tasks complete), and tried to finish; the only child was the FORCED reviewer (force-review attempt 1/2, working as designed — the model's inline python3 -c test didn't match the self-check gate's checker patterns). The reviewer did touch the cancellation path (claimed to fix a re-run-after-cancel bug), but neither it nor the root ever tested SIGINT behavior — the root re-ran the same happy-path smoke test and declared done.

Verifier: 5/6 PASS; the single fail is test_tasks_cancel_above_max_concurrent — "Cleaned up." printed 0 times instead of 2. The final run.py is the textbook asyncio.gather gotcha the test docstring names: gather + re-raise on CancelledError does not cancel still-queued siblings and await the running tasks' finally blocks on SIGINT. The MinimalPi 122b trial (fast__…143738) failed the IDENTICAL single assertion — subagents moved nothing here. The criterion is functionally hidden (the SIGINT test lives in the verifier mount, /app starts empty), but the requirement WAS stated: the agent's own docstring says "Handles keyboard interrupt gracefully" — it just never self-tested that claim. The correctable behavior is "write a mini SIGINT test for the stated requirement", not anything subagents-shaped.