Verifiers v0.1.15.dev12 Release Notes

Date: 05/29/2026

Highlights since v0.1.15.dev11

• Renderer trajectories carry richer token provenance. Renderer-backed clients now preserve per-token prompt attribution, serialize rendered-token payloads cleanly across the v1 boundary, route chat_template_kwargs through RendererConfig, and surface every parsed tool call from renderer responses.
• v1 sandbox execution is more production-ready. Sandbox-backed programs use the managed Python runtime, attach v1 sandbox labels, reuse sandbox clients correctly, report upload errors clearly, and handle routed-expert replay starts for bridged prompts.
• Reusable v1 packages and tasksets are staged for broader use. Standalone tasksets and harnesses packages are in place with hardened program config handling, and the Lean taskset now matches the final LAST EXIT_CODE marker to close a reward bypass.

Changes included in v0.1.15.dev12 (since v0.1.15.dev11)

Renderer and trajectory plumbing

• feat(renderer-client): give per-token prompt attribution to TrajectoryStep (#1414)
• fix(renderer-client): surface every parsed tool_call, drop status filter (#1463)
• Pipe chat_template_kwargs into typed RendererConfig (#1468)
• renderers <-> v1: serializable RenderedTokens (#1471)

v1 sandbox runtime

• Use managed sandbox Python runtime (#1450)
• Set routed experts replay start for bridged prompts (#1466)
• Add v1 sandbox labels (#1476)
• Fix v1 sandbox client reuse and upload errors (#1483)

Reusable v1 packages and tasksets

• fix(lean-taskset): match LAST EXIT_CODE marker to close reward bypass (#1480)
• Stage standalone tasksets and harnesses packages with hardened v1 program configs (#1475)

Full Changelog: v0.1.15.dev11...v0.1.15.dev12