First RecursivePi attempt on the read-heavy sanitize-git-repo task (see the
full writeup in runs/smoke__qwen3.6-35b-a3b__20260710-130242/NOTES.md).
FAIL 0.0. This run is contaminated by the peg-native parser bug: the session
died with errorMessage: "The model produced output that does not match the
expected peg-native format" (×3, willRetry:false) mid-thinking, right as
the model had finished finding the keys and was about to replace them — so it
made zero replacements. Same wire-format/template class as the regex-log
56Avdhy leak (llama.cpp #20837), orthogonal to recursive.
Delegation mechanics worked: 1 rlm_query (depth 0→1, 57 s child, exit 0),
RLM_MAX_DEPTH=1 held (no grandchild). But the read-only child returned only its
537-byte stderr banner — no findings — so the root fell back to searching
directly. The clean rerun …130242 (0 parser errors) reproduces the same
outcome without the bug, and MinimalPi …130731 fails identically; conclusion
lives in the …130242 NOTES.
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 — 0 pass · 1 failmean reward0.00tokens (job total)187,748 in / 4,291 outstarted / finished2026-07-10T12:56 / 2026-07-10T12:58wall clock2m17s| # | result | total | agent | in/out tok | flags | |
|---|---|---|---|---|---|---|
| 1 | FAIL | 2m17s | 1m42s | 187748/4291 | 🔍 view |