First in-container run of the new RecursivePi harness
(harnesses.recursive_pi:RecursivePi), the read-only-scout adaptation of
pi-recursive's rlm_query recursive-delegation tool. See AGENT_TODOS.md
("pi-recursive (rlm_query) recursive-delegation harness") and
harnesses/recursive_pi.py for the design.
Purpose: validate that the mechanism installs and runs inside a real task container (per-task base image), and that adding the package + provider patch + prompt does not regress a known-good task.
Result: regex-log PASS (reward 1.0), no errors, 12m27s.
pi-recursive package installed, pi-recursive child-spawn provider-inherit
patch active, pi-recursive bounds set (depth 2, calls 6, timeout 900s, jj
off, read-only children).recursive delegation prompt active at root.rlm_query calls — the model solved regex-log directly without
delegating. Expected: a trivial regex fix doesn't invite decomposition, and
the read-only-scout prompt doesn't force a first call (same delegation-decay
shape seen across the 18 SubagentsPi trials).What this run does NOT show: whether the child path works in-container (the
model never delegated). That was proven on the HOST instead — a forced
rlm_query child loaded llama-local/qwen3.6-35b-a3b, read a file with the
read-only tool, and returned it (twice, incl. under the final extension). The
provider-inheritance sed on pi-recursive's native-tool.ts (2nd -e for our
extension) is the load-bearing fix.
Next: run on a multi-part task where delegation should pay off (fast2:
mailman / git-multibranch) and, if the model still under-delegates, add a
first-action rlm_query scout nudge. Config knobs are not yet forwarded through
lib.sh, so config.json records recursive: null (activation is via the
RecursivePi class default) — see AGENT_TODOS for the promotion checklist.
llama-local/qwen3.6-35b-a3bagentharnesses.recursive_pi:RecursivePithinkingonreasoning budgetdirect (server/none)* — see journal for the authoritative mechanismmaxTokens / contextWindow65536 / 196608agent timeout ×2.0trials1 of 1 — 1 pass · 0 failmean reward1.00tokens (job total)4,281,691 in / 91,003 outstarted / finished2026-07-09T23:39 / 2026-07-09T23:52wall clock12m27s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | PASS | 12m27s | 11m42s | 4281691/91003 | 🔍 view |