Skip to content

v0.9.6 — Trajectory generation and isolated grading

Latest

Choose a tag to compare

@adithya-s-k adithya-s-k released this 10 Nov 23:43
· 1 commit to main since this release

Repo2RLEnv 0.9.6 adds experimental trajectory-to-environment generation and strengthens task grading.

  • Trajectory generation: convert ATIF, Claude Code, Codex and Pi sessions into repair, build, extend and QA Harbor tasks. Includes reference-solution calibration, safe task paths and isolated verifiers. Thanks to @younesbelkada for the original implementation (#237).
  • Grading isolation: native tasks use a fresh verifier; supported suites run contradictory assertions to detect forged passing results. PR-diff attempts without correctness evidence remain ungraded (#232, #233, #234).
  • Generation fixes: validate typed bootstrap options, remove unreachable fix-bearing objects from partial clones, recover pytest collection-error cases as fail-to-pass with audit logging, and reject SCALER string references that would otherwise be truncated (#246, #247, #248, #249).
  • Dependency maintenance: stable LiteLLM 1.108.3 and OpenAI 3.42.1, patched filelock in wheel metadata, and a scoped Effect update for CVE-2026-115779 (#250). Hub 2 remains deferred until stable LiteLLM compatibility.

Validation: 2,696 local tests passed; 25 real Pi/OpenCode mock-server contracts passed. Python, coding-runtime and website audits found no known vulnerabilities. CI additionally exercises Windows wheels, Docker/Harbor lifecycle controls, docs and packaging.

pip install --upgrade --upgrade-strategy eager repo2rlenv==0.9.6

Add the extras your pipeline needs. Rebuild remote worker wheels and rerun repo2rlenv tasksmith install-runtime to install the updated coding runtimes.

Existing Hub datasets are unchanged. Regenerate and revalidate affected bundles to adopt the fixes. Trajectory task construction checks do not replace task-quality review or blind model rollouts; generated tasks remain unverified until evaluated. Contradictory assertions detect supported forgeries, not arbitrary malicious code.