Issues

• --findings-budget accepts non-numeric input silently (NaN), unlike --n-ctx

  lloyal-ai/reasoning-run#15 · Filter by author kazimali07 opened on Jul 8, 2026

• defaults.maxTurns in harness.json is ignored — run uses a hard-coded limit

  lloyal-ai/reasoning-run#14 · Filter by author kazimali07 opened on Jul 8, 2026

• GLM-5.2 hangs GPU-idle at staggered recovery entry on 0.4.2 (worked on 0.3.1) — runaway recovery decode → KV exhaustion; multi-GPU llama_decode failure wedges silently

  lloyal-ai/reasoning-run#9 · Filter by author kazimali07 opened on Jul 6, 2026

• 0.4.x recovery silently drops the recovery system prompt — buildUserDelta ignores {system}; no published sdk supports it

  lloyal-ai/reasoning-run#8 · Filter by author kazimali07 opened on Jul 6, 2026

• --gpu cuda flag silently falls back to CPU; only LLOYAL_GPU=cuda env engages CUDA (0.4.2, Linux x64)

  lloyal-ai/reasoning-run#7 · Filter by author kazimali07 opened on Jul 5, 2026

• report.md includes the model's raw <think> block — RunDirSink writes the answer event verbatim (the TUI strips it)

  lloyal-ai/reasoning-run#5 · Filter by author kazimali07 opened on Jul 5, 2026

• lloyal.node 3.0.0 requires AVX-512 — crashes in createContext() on most consumer CPUs (regression from 2.1.1)

  lloyal-ai/reasoning-run#1 · Filter by author kazimali07 opened on Jun 17, 2026
  1