[celve]

Lets Harbor's computer-1 (as Qwen3VLComputer1) run on a desktop its sandbox image already runs, such as CUA-Lite's GNOME template, and lets a launch pick the agent and verifier classes. This is what a run needs to train computer-1 on CUA-Lite's Lite.OSWorld tasks through the session server.

What changes

• examples/experimental/harbor_cua/desktop_computer.py: DesktopQwen3VLComputer1 uses the image's display on :1 as it is.
  • It starts no Xvfb, XFCE, VNC or Chromium, judges liveness from the X socket alone, and never relaunches a browser after a failed action.
  • It waits settle_sec after each screen-changing action and can shrink screenshots in the sandbox (screenshot_width/height).
  • It caps the screenshots a request carries (keep_screenshots), counting what each prompt actually holds, so a JSON-parse retry keeps the current screen and the turn-cap final answer stays within the cap.
  • It refuses zoom (crop-relative clicks are mis-scaled for Qwen3-VL's grid) and the browser actions, which also leave its prompt.
  • An image without ImageMagick on its PATH supplies its own capture through _capture_script and session_class.
• harbor_agent_function.py:
  • HARBOR_AGENT_IMPORT_PATH and HARBOR_AGENT_KWARGS pick and configure the agent class; a kwarg the binding sets is refused.
  • HARBOR_VERIFIER_IMPORT_PATH and HARBOR_VERIFIER_KWARGS run a Harbor verifier class in place of the task's tests/.
  • launch_common.harbor_env_vars forwards all four.

Verification

At 63279ff273, with the merge base 4d17c365df unchanged on the target:

• Unit tests (Python 3.12, Harbor ac63e51a8f, no sglang):
  • test_desktop_computer.py + test_qwen_computer.py: 28 passed.
  • tests/fast/examples/experimental/harbor/: 104 passed, 1 failed. The failure is test_agentic_wiring_is_the_shipped_wiring, which imports sglang and fails without this change too.
• Live, on 2×H200 with SGLang 3a93821 plus Keep message images on pretokenized chat requests sglang#5, CUA-Lite 2af6926 and the template lin2139-liteosworld:
  • Untouched desktops: 30 of 30 trials set up and scored 0 at 1280×720.
  • Oracle replays: 10 of 10 scored 1.
  • Three GRPO rounds with CUA-Lite's recipe at small scale, 30 turns:
    • engine weight versions 0, 2, 4;
    • trainer–engine gap 0.0064–0.0144 nats on all 6 steps;
    • TITO mismatch 0 on every round;
    • three evaluations ran.
  • Screening of the base model on an earlier commit: 16 tasks × 8, 4 of 16 groups mixed. CUA-Lite's own harness had 3 of 16.
  • Record: https://github.com/softlash/LIN-2149/blob/main/plans/harbor-liteosworld-run.md

Review

sol, two rounds; the second passed with no open P1, P2 or P3.

• Round 1 found two P2s, both fixed here with tests:
  • a JSON-parse retry dropped the current screen;
  • the turn-cap fallback bypassed the screenshot cap.
• Its P1 (my stage-4 gate) was met by a narrowed gate that sol accepted. The 16×8 probe is screening only, and readiness rests on the 3-round run on this commit.

Still owed to the human

The full 20-round run needs an 8-GPU node, which is Linyu's call. This PR does not certify that run.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-22T11:21:27.103649Z	63279ff	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.