A normal MinimalPi smoke run of cancel-async-tasks, launched with
MEDIC_CAPTURE=on so exec_as_agent tars the workdir end-state to
agent/workspace.tar.gz (gitignored) after pi exits. This is the phase-1 half of
the first medic recovery pass (P0).
cancel-async-tasks FAIL (agent 17s) — the base agent's asyncio.gather +
semaphore draft passes 5/6 but mishandles cancellation above max_concurrent. Its
broken run.py was captured and fed to the medic.
Phase 2 (the medic): runs/medic__qwen3.6-35b-a3b__20260710-190552 — see its
NOTES.md and medic_recovery.json. Baseline scoring here is UNTOUCHED and stays
the standard number; the medic result lives on a separate rule-relaxed axis.
llama-local/qwen3.6-35b-a3bagentharnesses.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)12,598 in / 2,540 outstarted / finished2026-07-10T19:01 / 2026-07-10T19:03wall clock1m04s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | FAIL | 1m03s | 17s | 12598/2540 | 🔍 view |