Fourth job of the 122b fast2 one-task-per-job sequence, and a harness
lesson: a fast-fail cap must sit STRICTLY BELOW the task's declared
timeout × AGENT_TIMEOUT_MULT. headless-terminal declares 900s; ×2.0 =
1800s = exactly the cap this run used. Both timers fired at 30m00s and
Harbor's own AgentTimeoutError won the race, so the trial ERRORED
(exception recorded, NEVER VERIFIED) instead of taking the fast-timeout
path (exit 0 → verifier scores the on-disk state). The
agent/fast-timeout.json marker (exit_code 124) landed anyway — both
mechanisms triggered, Harbor's classification took precedence.
The agent worked the full 30 min (1.72M in / 28.7k out) without declaring
done. Rerun with cap 1700 (100s under the ceiling) as sequence job 7 →
fast2__qwen3.5-122b-a10b__20260712-180346, which got a scored verdict.
Ledger item added: validate cap maps against declared×mult at run start.
llama-local/qwen3.5-122b-a10bagentharnesses.minimal_pi:MinimalPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 0 pass · 0 fail · 1 erroredmean reward0.00tokens (job total)1,720,041 in / 28,659 outstarted / finished2026-07-12T16:24 / 2026-07-12T16:58wall clock34m00sErrored trials: AgentTimeoutError (headless-terminal)
| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | ERR | 34m00s | 30m00s | 1720041/28659 | 🔍 view |