Rerun of the single-task SubagentsPi container validation after strengthening
the workflow append ("make your very FIRST action a subagent tool call with
agent scout" — still positively phrased, no prohibitions). Compare
fast__qwen3.6-35b-a3b__20260706-225730 (same task, advisory phrasing, zero
delegation).
PI_SUBAGENT_CHILD).rm -rf .pi-subagents cleanup ran; no .pi-subagents droppings in the
workdir artifacts.This closes the "live subagent tool exercised only by a container run" caveat: install → global extension → child provider inheritance → delegation → cleanup all proven in the real Harbor pipeline. Wall-clock cost of delegation on a trivial task: ~2× the no-delegation run (1m55s vs 1m04s) — expect much larger multipliers on real tasks (per-switch re-prefills on the single llama-swap slot).
llama-local/qwen3.6-35b-a3bagentharnesses.subagents_pi:SubagentsPithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 1 pass · 0 failmean reward1.00tokens (job total)102,458 in / 2,333 outstarted / finished2026-07-06T23:01 / 2026-07-06T23:03wall clock1m55s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | PASS | 1m55s | 1m15s | 102458/2333 | 🔍 view |