[abhishekgahlot2]

Computer-use agents now run at the model's full capacity by default, never act on an unfinished response, and environments can mount provider volumes.

• No output caps. AnthropicComputerAgent sends max_tokens=128000 (the model maximum, ANTHROPIC_MAX_TOKENS) over a stream; OpenAIComputerAgent sends no max_output_tokens. Reasoning effort defaults to max for both.
• Unfinished responses stop the run. An OpenAI response whose status is not completed returns error with Response did not complete: <reason> and applies no action. An Anthropic stream must reach message_stop; a stream that ends early or disconnects returns error, keeps partial usage, and records the cost as unresolved.
• Unlimited budgets. Budget.max_turns and Budget.max_seconds accept None (no limit), as do MultiTurnEnvironment.max_steps and ComputerUseEnvironment.max_steps. Scripted QA controls still end after one round of each allowed action when an env sets no step limit.
• Volumes. MultiTurnEnvironment.volumes mounts provider volumes into Daytona sessions (VolumeMount(volume_id, mount_path, subpath)). The API rejects volumes on inline environments. A volume is shared across episodes, so the docstring says to mount it read-only behind a root-only path and check it against a pinned manifest.

Tests: ruff check clean; pytest -n 4: 2555 passed, 1779 skipped, 3 failed. The failures also fail on main: test_installed_cli_supports_adapter_flags[openai] (the local Codex CLI auto-updated to 0.163.0, pinned 0.162.1), and test_computer_use_suite_is_separate / test_load_computer_use_suite (the computer-use suite now holds 187 environments, the tests expect 5).

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}