fix: pass explicit 60s timeout to get_background_job in poll_job_completion

[rasdani]

Summary

CliAgentEnv.poll_job_completion polls background-job status via self.sandbox_client.get_background_job(...). Without an explicit timeout, that call inherits the prime-sandboxes APIClient default of 30s (httpx.Timeout(30.0, connect=10.0)). When the sandbox gateway is briefly slow on the status read, httpx.ReadTimeout fires and poll_job_completion wraps it as AgentError('Agent polling failed: Read file timed out after 30s: /tmp/job_XXX.exit').

Observed ~51 such failures on one 30B opencode-math run.

Change

One line, one kwarg:

status: BackgroundJobStatus = await self.sandbox_client.get_background_job(
    sandbox_id, background_job, timeout=60.0
)

Why 60s (and not higher)?

• 60s doubles the tolerance for a transient gateway blip (which is the actual observed failure mode — short stalls, not permanently-dead reads).
• A much larger value would let a truly stuck gateway wedge the rollout for a long time before the outer retry/timeout machinery gets a chance to intervene; 60s is long enough to absorb a blip, short enough to fail fast on a real hang.
• Raising the APIClient default globally would also affect unrelated fast lookups — this keeps the change scoped to the known-slow path.

Dependency

Requires prime-sandboxes PR PrimeIntellect-ai/prime#540 (merged), which adds the timeout kwarg to get_background_job. Consumed via prime-sandboxes release PR PrimeIntellect-ai/prime#541 (bumps to 0.2.21).

The pyproject.toml constraint has been bumped to prime-sandboxes>=0.2.21 (see commit 874faa8f). Lockfile update is deferred until the 0.2.21 release lands on the index — uv lock currently fails to resolve because 0.2.21 is not yet published. Once PR #541 merges and the package is published, a follow-up commit will run uv lock to pin the resolved version.

Test plan

• uv run ruff check verifiers/envs/experimental/cli_agent_env.py — clean.
• uv run ruff format --check verifiers/envs/experimental/cli_agent_env.py — clean.
• uv run pytest tests/test_cli_agent_env.py — 17 passed. (No existing coverage of poll_job_completion; this change is a kwarg passthrough, so no new tests added.)
• Pre-commit hooks (ruff check, ruff format, ty) passed on commit and push.

🤖 Generated with Claude Code

---

Note

Low Risk
Small, localized change to background-job polling behavior plus a minor dependency bump; main risk is compatibility/regression if the new prime-sandboxes timeout handling behaves unexpectedly.

Overview
Reduces transient ReadTimeout failures when polling sandbox background-job status by passing an explicit timeout=60.0 to sandbox_client.get_background_job() in CliAgentEnv.poll_job_completion.

Bumps prime-sandboxes dependency to >=0.2.21 to pick up the new timeout parameter support.

^{Reviewed by Cursor Bugbot for commit 9f807c5. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e136b0c. Configure here.}