[yaoyu-33]

Note

Stacked on #3591. This branch starts from #3591's head (c004893a9), so until #3591 merges this diff also shows #3591's 11 files. Review only the last commit, 56c4518d5. It changes 5 files (+106 / −106). I will rebase onto main after #3591 lands, then mark this PR ready for review.

What does this PR do?

This moves the OSWorld-only transport-log fields out of the vLLM model server, the follow-up requested in the #3591 review.

responses_api_models/vllm_model/app.py hard-coded thirteen x-nemo-gym-log-* header names: seven on main, six more added by #3591. The only sender of any of them is osworld_agent, which caused four problems:

• the generic model server carried OSWorld/NeMo-RL field names such as domain, parse_attempt and sampling_event_id;
• every new OSWorld field needed a vllm_model change;
• four fields the agent already sends (rollout_purpose, agent_contract_id, history_policy_id, model_protocol_id) were silently dropped;
• the transport log showed a top-level rollout_id that is not Gym's _ng_rollout_id and is easy to mistake for it.

Changes:

• osworld_agent sends its normalized log context as one JSON object in the x-nemo-gym-log-context header. The field list, integer normalization and 1,024-character bound stay in _OSWORLD_LOG_CONTEXT_FIELDS / _normalize_log_context, where they already lived. The JSON is ASCII-only, so control characters are escaped instead of stripped.
• vllm_model reads that header as an opaque object and records it unchanged under one key, log_context, in its opt-in transport log (NEMO_GYM_VLLM_TRANSPORT_LOG). It no longer names any field. The caller's fields can no longer pass for the log's own (event, call_index, payloads, …). Transport events move to schema version 3. A missing, malformed or non-object header is ignored, and the context never reaches the model body or the engine.
• benchmarks/osworld/README.md names the header and the log_context key.

No issue was opened: this is the follow-up agreed in the #3591 review.

Compatibility

• Transport log format: changed. Correlation fields move from the top level of each model-io-transport.jsonl record to log_context, with the same names and types (integers stay integers), plus the four fields above that were previously dropped. Records now carry schema_version: 3, so analysis that reads task_id, step and similar fields at the top level must read them from log_context for version 3. The agent log (OSWORLD_MODEL_IO_LOG) is unchanged and still carries the same fields at the top level.
• Header contract: vllm_model no longer reads the per-field x-nemo-gym-log-* headers. The only in-repo sender (osworld_agent) is updated here. An out-of-tree client that set those headers would lose them from the transport log until it sends x-nemo-gym-log-context instead.
• Benchmark impact: none. This affects only an opt-in diagnostic log, not requests, responses, token IDs or rewards.

Validation

On macOS ARM64 with Python 3.13.14, using a venv built from this branch (uv sync --extra dev --extra telemetry --extra ray):

• responses_api_models/vllm_model/tests: 571 passed.

• responses_api_agents/osworld_agent/tests: 385 passed; the base commit has 384.

• The new and changed tests fail when only the two app.py files are reverted (the end-to-end test also fails against the earlier flat version of this PR):

  • test_transport_log_context_is_the_callers_json_object
  • test_transport_log_records_the_callers_context_under_one_key, an end-to-end /v1/responses call. It checks that the context lands under log_context (schema version 3) and not at the top level, that a caller-supplied event stays inside it, and that the context does not reach the engine.
  • test_log_context_headers_do_not_change_model_payload
  • test_log_context_headers_drop_unknown_fields_and_bound_values
  • test_messages_model_fn_propagates_task_context_in_headers_and_logs

  The malformed-header cases are guards and also pass on the old code.

• pre-commit run on the changed files: passed.

Rollouts

N/A. This changes only the opt-in transport log's context plumbing. It does not change the model request body, sampling, responses, token IDs, verification or rewards, so no GPU or OSWorld VM run was made.

🤖 Generated with Claude Code

[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.

[JeffPengCoder]

I re-checked this at 56c4518d5, the commit we reviewed in #3591 (comment). It hasn't changed since. No problems from our side:

• No OSWorld field names in vllm_model.
  • It stores the caller's context unchanged under log_context, with transport schema_version: 3. There is no longer a top-level rollout_id that could be confused with _ng_rollout_id.
  • _OSWORLD_LOG_CONTEXT_FIELDS is now the only field list.
  • OSWORLD_MODEL_IO_LOG keeps the same fields at its top level.
• The header is always valid.
  • It is one compact JSON object written with ensure_ascii=True, which is more reliable than stripping CR/LF.
  • A missing or malformed header becomes {} and does not fail the request.
• Nothing depends on the moved fields. No other sender or reader on main uses them. Our OSWorld validation scripts only count transport events by the top-level event, which stays where it is.
• The PR body is accurate. Its compatibility note matches the change.
• It merges cleanly. The commit applies cleanly on current main (aa53f3bdf). It doesn't conflict with fix(osworld): keep the Nano Omni profile's sampling and document its Omni v3 recipe #4897: both edit benchmarks/osworld/README.md, but in different sections.

The branch still carries #3591's pre-merge commits. Because #3591 was squash-merged, those commits now conflict with main, so the branch needs the planned rebase before it can leave draft. @Glorf, could you approve once it's rebased? Approvals are dismissed on push here, so an approval on the current head wouldn't carry over.