[JeffPengCoder]

prepare.py always emitted 1500 output tokens and temperature 1.0, overriding the Nano Omni profile's 4096 / 0.6 recipe. This change preserves profile and task-row sampling unless the caller passes explicit flags, pins the tested history-template argument, and documents the Omni v3 serving recipe.

Base and dependencies

This PR includes main at bd6f5f0d8f3a56cbf3f7e1d2d4a1d2220b7ff6fe. Updated head: 9de7ffcf8e75498a6ed2c97cec3efb3a71e642f4. The main-relative diff remains limited to the original seven OSWorld files.

• fix(sandbox): reconcile Docker creates and resolve E2B service endpoints #4822 and fix(core): close shared aiohttp clients on their owner loop #4823 are merged. The sandbox fields and E2B endpoint contract are present on main; there is no remaining stacked dependency on either PR.
• feat(osworld): add taskgen plugins, reward modes, and capture-aligned history #4896 is merged. Its taskgen, reward-mode and capture-aligned accepted-history behavior is retained unchanged.
• build(env): AUT-2574 lock managed environments and use uv sync #4837 moved managed components to pyproject.toml and uv.lock, with a separately locked opt-in OSWorld runtime layer. The recipe documentation now follows that installation contract.
• refactor(vllm_model): let callers own transport-log context fields #4886 remains independent and open; it is not a merge prerequisite.

What does this PR do?

It preserves the Nano Omni profile's sampling and documents the serving and agent configuration behind the reported Omni v3 scores, together with the current managed-install and SDK compatibility requirements.

1. prepare.py no longer overrides the profile's sampling.

   • Before, it always wrote responses_create_params with max_output_tokens: 1500 and temperature: 1.0 into env.yaml.
   • Rollout collection merges that block over every row (row_params | responses_create_params_overrides in nemo_gym/rollouts/rollout_collection.py). The OSWorld agent then uses the request's values instead of its config.
   • So --profile nano_omni without flags ran at 1500 tokens and temperature 1.0 instead of the profile's 4096 tokens, temperature 0.6 and top-p 0.95.
   • Now --max-output-tokens, --temperature and --top-p default to None, and only the values that are passed are written.
   • test_nano_omni_env_keeps_the_profile_sampling_recipe merges the base config, the nano_omni profile and the generated env.yaml, and checks the effective sampling, step limit and history policy.
   • test_main_writes_complete_nano_omni_profile now also checks that no responses_create_params is written when the flags are omitted.

2. The Omni v3 recipe is documented in benchmarks/osworld/README.md, with pointers from the profile config and the agent README:

   • Serve the checkpoint:
     • The checkpoint is nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 at revision 24e67ea0.
     • The chat template must be passed with --chat-template. The correct one is SHA-256 ab7813c3…, the template embedded in the checkpoint's tokenizer_config.json. By default vLLM serves the checkpoint's chat_template.jinja (41428e0c…) instead.
     • The section also covers the nano_v3 reasoning parser, where to get it, and the vllm serve command.
   • Scope: Omni v3 only. Other Nemotron generations, such as 3.5 Super and 4, ship their own tokenizer, chat template and parser.
   • Prepare and run on the 361 tasks of OSWorld's test_nogdrive manifest (test_all without the 8 tasks that need Google Drive), including the command that builds that set.
   • Tune the output limit: the limit is tunable, so the section gives a method rather than one value.
   • Reference scores for Omni v3, with the run-to-run spread.
   • Infeasible tasks: an infeasible task scores only when the model itself ends the episode with FAIL. The adapter never appends FAIL at the step limit.
   • AgentENV backend:
     • the E2B SDK version;
     • how the managed lock selects the SDK independently of the root environment and historical recipe;
     • certificates from a private CA;
     • proxies.

3. The profile pins truncate_history_thinking=true.

   • The profile sent no chat_template_kwargs, so the served template's own default decided whether earlier assistant turns keep their reasoning.
   • Every reference score used that default, true: image-backed history turns are rendered as an empty <think></think> plus the Action, and turns folded into the text history keep their Thought.
   • The profile now sets it explicitly, the serving table lists it, and test_nano_omni_profile_pins_truncated_history_reasoning checks it.
   • Sending false, as NVIDIA's computer-use cookbook for this model does, keeps the reasoning in those turns. That changes the model input, so the README treats it as a separate recipe to validate.

No issue: this documents the recipe behind the OSWorld scores reported in #3591. It also fixes a prepare.py default that was found while checking those scores.

Validation

Validated source: 9de7ffcf8e75498a6ed2c97cec3efb3a71e642f4, merging the original PR with main bd6f5f0d8f3a56cbf3f7e1d2d4a1d2220b7ff6fe and correcting README references to locked installation and SDK compatibility. The committed tree matches the audited candidate exactly.

• OSWorld benchmark tests, all OSWorld agent tests, public-API/facade tests and token-capture tests: 1565 passed, 1 failed. The failure is the existing main test benchmarks/osworld/tests/test_run_scripts.py::test_managed_osworld_agent_installs_opensandbox_sdk, which still asserts the old managed-install contract, beginning with nemo-gym[dev] instead of nemo-gym[dev,ray].
• Pure main's benchmark suite: 42 passed, 1 failed, with that same test failing. The two PR sampling regression tests, copied verbatim and run against main, both fail as expected; both pass in the merged candidate.
• Sampling diagnostic: 60 complete cases, 30 on main and 30 on the candidate, covering default, pointer, nano_omni, omitted flags, single-field overrides, zero temperature, all flags, and task rows with their own sampling parameters. Every candidate emission matches the explicit flags. These are configuration/precedence checks, not live model rollouts.
• pre-commit run --all-files --show-diff-on-failure: passed. git diff --check passed. Only the intended README corrections differ from the initially tested merged tree; executable source is unchanged after the test run.
• All 12 Markdown relative links and in-page anchors across the two changed READMEs resolve.
• 55 runtime/config/lock files under the OSWorld agent, OSWorld resources server and vLLM transport match main byte for byte. The agent parser files also match the original PR head byte for byte. No parser, model-output normalization, or repair behavior was added by this integration.

Tests ran on macOS ARM64, Python 3.13.14, with a separate validation environment using root-lock aiohttp 3.14.4, Ray 2.58.0 and OpenAI 2.44.0. This is not a complete installation or smoke test of the managed component lock (which selects aiohttp 3.14.5 and different sandbox SDK versions).

The current CI classifier makes this a full-suite diff because benchmark Python/YAML files are uncategorized paths; the agent test file also prevents a docs-only interpretation. A trusted main CI run is still required for this updated head. An authorized /ok to test <full-sha> applies only to the specified commit; no CI authorization comment was posted as part of this update.

Rollouts

No new model/VM rollout was run during this main-sync review. The historical runs below validate the intended sampling on their original sources and SDKs; they do not qualify the new managed lock. A model-and-verifier smoke with the final frozen environment remains outstanding before claiming current-main runtime compatibility.

With this change, --profile nano_omni without sampling flags gives the values the reference runs passed explicitly:

• temperature 0.6 and top-p 0.95;
• 100 steps;
• the fixed history policy with three screenshots;
• 4096 output tokens, unless --max-output-tokens is given.

test_nano_omni_env_keeps_the_profile_sampling_recipe checks these values on the merged config.

The README's reference scores come from real OSWorld runs on the 361 tasks with the documented serving configuration:

Backend	Proxy-required tasks	Output limit	Raw reward rate	Binary success rate
OpenSandbox Pool	proxy enabled	4096	50.97%	48.48%
AgentENV	proxy disabled	16384	44.61–48.13% (4 runs)	42.66–45.98%

The most recent AgentENV run is the October 20 reproduction in #3591's description.

Compatibility and benchmark impact

• When --max-output-tokens, --temperature and --top-p are all omitted, prepare.py no longer writes responses_create_params, so the profile's agent config decides the sampling:
  • nano_omni: such runs change from 1500 output tokens at temperature 1.0 to the profile's 4096 tokens, temperature 0.6 and top-p 0.95, so their scores change.
  • default and pointer: with standard task rows that supply no sampling values, they still inherit 1500 tokens, temperature 1.0 and no top-p from osworld_agent.yaml.
  • All profiles with task-row sampling: without CLI flags, existing per-row sampling is now honored instead of having max-output tokens and temperature replaced by prepare's old defaults. For example, a row requesting 8192 tokens at temperature 0.2 now retains those values; main previously forced 1500 and 1.0. Explicit CLI flags still take priority, one field at a time.
• Each explicitly passed flag retains its value; omitted fields now follow task-row sampling and then profile defaults.
• The profile now sends chat_template_kwargs: {truncate_history_thinking: true} on every request. With the documented template this is the template's own default, so the model input and the scores do not change. With a server or template whose default differs, the rendering now stays the tested one.
• Everything else is documentation and config comments.

Managed-runtime compatibility

The reference AgentENV runs used E2B 2.46.0, whose sandbox-create call uses POST /sandboxes. The managed OSWorld agent lock now selects E2B 2.58.0, whose sync and async create calls use POST /v2/sandboxes. The root Gym lock still selects 2.46.0; that does not control the isolated agent runtime. SDK wheels were read and their hashes checked against the committed locks to verify this difference.

The README now makes that distinction explicit. A control plane must support the selected SDK's endpoint, or the run needs a separately recorded agent manifest and lock selecting the compatible SDK. An ad hoc post-install pin and UV_CONSTRAINT do not define a reproducible locked environment. No deployment compatibility or new score is claimed from unit tests, and no SDK, provider or parser fallback is introduced here.

Checklist

• I have read the contributing guidelines.
• The change is focused; unrelated "drive-by" edits are tracked as separate issues/PRs.
• Local targeted checks are recorded above; the pre-existing main installer assertion and a final-environment model/verifier smoke remain outstanding.
• Documentation added or updated, or N/A with justification.
• Pre-commit checks pass locally (pre-commit run --all-files) (so CI lint/format checks pass).
• New source files include the required Apache-2.0 SPDX header. (No new files.)
• All commits have DCO sign-off (git commit -s) (so the DCO check passes).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.