• v1.6.0 6fb77f5263

    v1.6.0 — Autoresearch Loop Phase A automation operational
    CD / Lint / Test / Vet (push) Failing after 3s
    CD / Deploy via GitOps (push) Has been skipped
    CD / Build & Import (push) Has been skipped

    mathias released this 2026-06-29 21:36:06 +00:00 | 2 commits to main since this release

    The .gitea/workflows/autoresearch.yml workflow now runs end-to-end on the
    self-hosted koala runner: deps (histdata + torch cu130) → committed toy data
    fixture → scaffold run dir → GPU baseline (val_vol_r2=0.2177) → 3 agent
    iterations (berget/gemma4-31b proposes train.py edits, keep-if-improved else
    revert) → artifacts uploaded.

    Took the feature from never-green to green via six fixes:

    • deps: histdata>=1.3 (nonexistent) → >=1.1
    • ci: upload-artifact@v4 (unsupported on Gitea/GHES) → @v3
    • ci: install torch from cu130 wheel index in venv step
    • ci: commit 59KB eurusd_daily.parquet CI fixture (scoped #10 exception)
    • loop: project root on PYTHONPATH so copied train.py imports scripts/
    • ci: drop empty LOOP_MODEL env that blanked the model → LLM 400
      Plus repo Actions secrets LITELLM_KEY (scoped virtual key) + LITELLM_BASE.
    Downloads