[ananthsub]

What does this PR do?

This PR records the response envelope ID on each captured model call. The ID is a lightweight receipt that an agent can retain while token IDs and log probabilities remain in the configured token sink.

• Adds the optional TokenEntry.response_id field to the existing version 1 record schema.
• Records the top-level ID returned by the Responses, Chat Completions, or Messages endpoint.
• Reuses the underlying Responses ID in the Anthropic conversion so the client-visible ID matches the captured ID.

Response ID flow

sequenceDiagram
    participant Agent as Agent harness
    participant Capture as Capture middleware
    participant Route as Model API route
    participant Model as Model server
    participant Sink as Token sink

    Agent->>Capture: Request /ng-rollout/{rollout_id}/training-token-capture/v1/{dialect}
    Capture->>Capture: Strip prefix and set rollout_id + model_call_id
    Capture->>Route: Forward dialect request with capture context
    Route->>Model: Generate response
    Model-->>Route: NeMoGymResponse(id, output, token metadata)
    Route->>Sink: TokenEntry(model_call_id, response_id=id, token arrays, output items)
    alt Responses or Chat Completions
        Route-->>Agent: Client response carries the recorded id
    else Anthropic Messages
        Route->>Route: Convert response and preserve response.id as message.id
        Route-->>Agent: Message response carries the recorded id
    end

Loading

The response ID and token arrays are written to the same TokenEntry. The agent only needs to retain the response ID; later code can use it to locate the corresponding captured model call.

This is the first PR in the terminal-attribution stack. #2676 uses the recorded ID as one witness when resolving the model call that produced the verified response.

Checklist

• I have read the contributing guidelines.
• The change is focused; unrelated "drive-by" edits are tracked as separate issues/PRs.
• Tests added or updated and pass locally, or N/A for docs-only / non-code changes (so CI unit/server checks pass when applicable).
• Pre-commit checks pass locally (pre-commit run --all-files) (so CI lint/format/copyright pass).
• All commits have DCO sign-off (git commit -s) (so the DCO check passes).

[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.

[ananthsub]

/ok to test 6a91461

[ananthsub]

/claude review

[claude]

SHIP — LGTM, no reliability concerns.

Reviewed all 4 changed files against base. This is a small, additive change that threads the served envelope id through token capture:

• TokenEntry.response_id is a new optional field (= None, default), fully backward-compatible under extra="allow": old records deserialize to None, and keeping TOKEN_ENTRY_RECORD_SCHEMA_VERSION at 1 is correct since the addition is optional and non-semantic.
• capture_tokens records payload.get("id") and never mints — consistent with the stated invariant. Capture runs on the inner response (_invoke_responses/_invoke_chat_completions) before the Anthropic conversion drops token fields, so ordering is right.
• responses_to_anthropic_response reuses the inner Responses id on the outer Messages envelope, with the prior msg_<uuid> mint retained as a fallback when response.id is empty. The Anthropic SDK's Message.id is a plain str (no prefix validation), so reusing a resp_/chatcmpl_ id validates cleanly — the join holds across the responses, chat, and messages dialects.
• No async/HTTP concerns; no other TokenEntry consumer needs updating. Tests exercise all three routes plus the converter and the missing-id / newer-record cases, asserting real behavior.