chore(deps): pin openai to 2.44.0

[ananthsub]

Closes #2452

This change pins openai==2.44.0 and updates the Responses and Chat schemas used by Gym.

Gym subclasses OpenAI SDK models and copies several request schemas into Pydantic models. The dependency is pinned to one audited version because a version range would allow schema changes that Gym has not handled. OpenAI 2.44.0 is the latest usable release in this update; 2.45.0 makes InputTokensDetails.cache_write_tokens required and breaks NeMoGymResponseInputTokensDetails construction.

Responses items

OpenAI 2.44.0 adds the Codex tool family and context-management items:

• shell_call and shell_call_output
• apply_patch_call and apply_patch_call_output
• tool_search_call and tool_search_output
• compaction and compaction_trigger
• additional_tools

The model-generated calls are generation boundaries. Client results, approvals, tool carriers, and compaction records are non-boundaries. Types without a Chat Completions representation remain explicitly inconvertible.

Output and replay input

OpenAI 2.44.0 gives some same-tag output and input models different value domains. NeMoGymResponseOutputItem and NeMoGymResponseInputItem use separate owners for those items.

flowchart LR
    PROVIDER["Provider response"] --> OUTPUT["NeMoGymResponseOutputItem"]
    OUTPUT --> PRESERVE["Preserve provider fields"]
    PRESERVE --> REPLAY["Next request input"]
    REPLAY --> NORMALIZE["Normalize output-only values"]
    NORMALIZE --> INPUT["NeMoGymResponseInputItem"]
    INPUT --> API["OpenAI 2.44 input schema"]

Loading

additional_tools.role is preserved on provider output. Replay changes non-developer roles to developer, which is the only value accepted by the SDK input model.

computer_call_output.status="failed" is preserved on provider output. Replay omits that optional status because the SDK input model does not accept failed.

Local agents may construct a response with the replay-input function_call_output model. The output union continues to accept that model alongside the SDK provider-output model.

Request schemas

The Responses request model includes the fields added through OpenAI 2.44.0, including context management, conversation, moderation, cache controls, and safety identifiers. Native Responses model servers forward these fields.

The Chat Completions request model includes moderation, prompt_cache_key, prompt_cache_retention, safety_identifier, and verbosity. Unknown fields are rejected instead of being silently removed. Deprecated function_call and functions remain disabled.

Shared request fields round-trip through the converter. text.verbosity maps to Chat verbosity. reasoning.effort maps to Chat reasoning_effort, including the 2.44 values none and xhigh. See the Responses reasoning configuration and Chat reasoning effort references.

Nested reasoning controls without a Chat equivalent are rejected. text.format is also rejected until its different Chat and Responses object shapes have an explicit conversion.

Function tools preserve strict. Deferred loading and tool types without an implemented conversion raise NotImplementedError instead of being partially converted. See the Responses tools reference.

flowchart TD
    REQUEST["Responses request"] --> ROUTE{"Model server API"}
    ROUTE -->|"Responses"| PASS["Forward supported Responses fields"]
    ROUTE -->|"Chat Completions"| CHECK{"Lossless Chat conversion?"}
    CHECK -->|"Yes"| CONVERT["Map shared fields"]
    CHECK -->|"No"| ERROR["Raise NotImplementedError"]

Loading

Replay safeguards

OpenAI 2.44.0 adds values that Chat Completions cannot represent. Downconversion now rejects them before information is lost:

• message.phase
• function-call namespace
• function-tool defer_loading=true
• input images referenced by file_id
• input-image detail original

Other Responses-only items and fields retain the same explicit rejection path.

Streaming and component compatibility

The streaming sanitizer removes stream_options when it removes stream, because the option is only valid for streaming requests. It consumes additional_tools after hoisting supported function tools into the top-level tools field.

Response comparisons in affected agent tests remove null values recursively. This accounts for optional fields added inside nested output items. Duplicate item fixtures were removed so each type tag has one payload.

The resources_servers/cvdp OpenAI pin was removed. The package-level pin is the source used when server environments are created.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[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.

[ananthsub]

/claude review

[claude]

SHIP — LGTM, no reliability concerns.

Reviewed the openai 2.7.2→2.44.0 bump and the Responses/Chat converter expansion end-to-end.

What I checked and why it's sound:

• Converter correctness (responses_converter.py): the new responses_to_chat_completion_create_params fails loudly (NotImplementedError) on every field/tool/tool-choice with no Chat representation (context_management, conversation, defer_loading, message phase, function namespace, custom-tool grammar, hosted tool choices, text.format) rather than silently dropping — the correct failure mode for a downconverter feeding a training/eval run. exclude_none=True on the dump means null Responses-only fields are treated as absent (tested), so callers passing explicit None aren't spuriously rejected.
• Behavior change worth flagging as intentional, not a bug: function-tool strict now round-trips to the model (previously pop("strict", None) dropped it) — the vLLM test asserts strict: True now reaches the backend. This is a lossy-drop fix, and it's covered.
• Item-union expansion: new Codex-family output items (apply_patch/shell/tool_search/compaction/additional_tools) are mirrored in both input and output unions, with normalize_output_items_for_replay correctly rewriting additional_tools→role=developer and dropping status=failed from computer_call_output on replay. test_gym_output_union_represents_every_sdk_output_type and test_request_model_carries_every_sdk_request_field pin the union to the SDK so future bumps fail fast.
• genrm_compare grouping key: exclude_none=True stabilizes the prompt-hash against SDK field additions; grouping stays consistent within a run, no scoring impact.
• Dependency hygiene: exact pin is deliberate and documented (2.45.0 breaks InputTokensDetails), overrides files reconciled (finance_agent floor now compatible, strands bumped, cvdp de-duplicated), uv.lock consistent.
• Test robustness: _drop_nulls on both sides of equality only masks None==None; a real-value vs None mismatch still fails, so assertions keep their teeth.

No async-path, public-API-break, or silent-corruption issues found. Merge.

[ananthsub]

/ok to test 9097819

[ananthsub]

/ok to test a9fc2b6