[dmchoiboi]

Why

collect_artifacts reads each file by streaming base64 < file through sandbox.exec(). Sandboxes that only implement the buffered exec_with_output, such as smol_vm and the ARP provider in current agentenvhub-sdk pins, fail with an AttributeError. This blocks tasks moving off the retired beta sandbox onto ARP (FDE-2526, P0).

What

collect_artifacts.py _collect_file:

Sandbox has	Before	After
exec (streaming)	stream + incremental decode	unchanged
only exec_with_output	AttributeError: ... no attribute 'exec'	one exec_with_output("bash", "-c", "base64 < f"), decoded in one piece

The size integrity check and the upload path are unchanged for both branches.

For reviewers

The streaming branch is the same code, indented one level. The fallback holds the whole base64 text in memory, which is fine at artifact sizes. A streaming exec for the ARP provider is in scaleapi/scaleapi#169759; this PR unblocks collection on the SDK versions already pinned.

Verification

• New test_a_sandbox_without_streaming_exec_collects_over_buffered_exec (binary payload, 76-column wrapped base64). Before the change it fails with the same AttributeError users hit; after the change it passes.

• tst/unit/task_step/test_collect_* and test_container_sandbox_routing: 107 passed. ruff clean.

• Live on ARP staging (one 1-CPU VM, deleted afterwards), on a 3 MB random file:

  Check	Result
  the ARP provider's streaming exec("bash","-c","base64 < f; echo to-stderr >&2; exit 3") streamed	990 chunks, decoded SHA-256 matches the source, stderr = to-stderr, wait() = 3
  collect_artifacts._collect_file via streaming exec	uploaded bytes match the source
  collect_artifacts._collect_file via exec_with_output only (this PR)	uploaded bytes match the source

🤖 Generated with Claude Code

[greptile-apps]

smol_vm collection still fails

SmolVmSandbox inherits Sandbox.exec, which raises NotImplementedError. That makes hasattr(sandbox, "exec") true, so collection still takes the streaming branch and fails instead of using exec_with_output.

Detect unsupported streaming execution, not just a missing attribute. Add a test using a sandbox that inherits the base exec method; the current fake has no such method and misses this case.

Knowledge Base Used: Provider integrations and sandbox deployment

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agent_env/task_step/task_steps/collect_artifacts.py
Line: 482

Comment:
**`smol_vm` collection still fails**

`SmolVmSandbox` inherits `Sandbox.exec`, which raises `NotImplementedError`. That makes `hasattr(sandbox, "exec")` true, so collection still takes the streaming branch and fails instead of using `exec_with_output`.

Detect unsupported streaming execution, not just a missing attribute. Add a test using a sandbox that inherits the base `exec` method; the current fake has no such method and misses this case.

**Knowledge Base Used:** [Provider integrations and sandbox deployment](https://app.greptile.com/scale-ai/-/custom-context/knowledge-base/scaleapi/agentenv-framework/-/docs/provider-integrations.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.