[hallerite]

Summary

• verifiers 185b67d → 4fd5921
  • feat: run the agent as a task-declared user verifiers#2748: run the agent as a task-declared user (Harbor's [agent] user), e.g. ProgramBench's unprivileged agent next to an execute-only reference. Needs prime-sandboxes 0.4.1.
  • fix: accept loops in ACP semantic edges; bump nano-rlm to c38e20b verifiers#2744: nano-rlm default bumped to c38e20b; ACP semantic edges accept loops.
  • fix(responses): preserve inline reasoning under network restrictions verifiers#2736: preserve inline reasoning under network restrictions.
  • feat(v1): add E2B runtime provider verifiers#2294: E2B runtime provider.
• prime-envs e82de8b → d4fb80a
  • pi-programbench 0.1.2 (pi-programbench 0.1.2: run the solver as the images' unprivileged agent user prime-envs#960): pi-programbench@2026-10-02, with the solver as the images' agent user.
  • pi-programbench 0.1.1 (feat(pi-programbench): end the rollout on the submit command prime-envs#879): the submit command ends the rollout.
  • swe-sweep (Add SWE-sweep with isolated Harbor grading and QC prime-envs#873) and pi-swe-sweep (Add PI-SWESweep repaired-subset taskset prime-envs#878); OfficeQA updates (Require offline execution for PI-OfficeQA Pro V2 prime-envs#958, Use merged OfficeQA tasks and task-level network policy prime-envs#959).
• uv.lock: prime-sandboxes 0.3.2 → 0.4.1, pi-programbench 0.1.2, and the new env packages.

Validation

Run from this branch's stack (verifiers 4fd5921, prime-sandboxes 0.4.1):

• pi-programbench: loading with its defaults resolves pi-programbench@2026-10-02 with agent_user == "agent". Before the prime-envs bump, two uv run eval pi-programbench rollouts on Prime (bash harness, GLM-5.3) ran as agent: they saw the reference as ---x--x--x root, and the separate verifier graded both with no errors.
• programbench-synth 1.1.0 (PrimeIntellect-ai/prime-tasks-private#369, also [agent] user = "agent"), three rollouts on Prime:
  • the root healthcheck ran;
  • the agent ran as agent;
  • two episodes were graded in the separate offline verifier, and the third was still running.

🤖 Generated with Claude Code

---

Note

Medium Risk
Lockfile-only, but it pins sandbox and verifier runtime versions and new SWE env packages that affect how rollouts execute in containers.

Overview
Refreshes uv.lock after bumping verifiers and prime-envs, wiring in new SWE evaluation packages swe-sweep and pi-swe-sweep and bumping several editable envs (pi-programbench 0.1.2, pi-officeqa-pro-v2 0.1.4 with huggingface-hub dropped from that package’s lock metadata).

The lock also moves prime-sandboxes 0.3.2 → 0.4.1 (reflected in verifiers’ minimum sandbox version) and records verifiers’ new optional e2b extra. Together this tracks upstream Harbor/agent-user and sandbox API changes described in the linked verifiers and prime-envs PRs.

^{Reviewed by Cursor Bugbot for commit 3f59069. Bugbot is set up for automated code reviews on this repo. Configure here.}