[faresobeid]

renderers.client.generate() passes two more optional fields of prime-rl's /inference/v1/generate response through, next to sampling_mask. Pure pass-through: no request-side change, no parsing. Servers that do not send them (stock vLLM, current prime-rl main) give None.

• sampling_mask_logprobs: the sampler's logprob at each sampling-mask id, packed base64 float32, parallel to the packed sampling_mask ids.
• payload: by-handle segments {field, file, offset, pos, rows, dtype, shape} of a shared file that the inference server writes instead of returning routed_experts / sampling_mask inline.

Validation

uv run pytest tests/test_client.py (104 passed, 6 skipped), ruff check, ruff format --check, ty check renderers (warnings only, as on main); CI green.

Dependencies

• feat(v1): by-handle payload segments on traces verifiers#2765 reads payload onto traces; lands with this PR.
• No producer is on prime-rl main yet: the payload writer was feat: pass router-replay ids and sampling masks by handle on multi-node runs prime-rl#3902 and the sampling_mask_logprobs writer feat: exact score centering over the sampling mask (arXiv:2609.20807) prime-rl#3794 (both closed unmerged).

Note

Expose sampling_mask_logprobs and payload in client generate response

• generate in renderers/client.py now reads sampling_mask_logprobs from the inference response choice and returns it alongside sampling_mask.
• Also returns the choice's payload, which holds by-handle side-array metadata for data stored by the inference server instead of inline.
• Documents the packed float32 log probabilities emitted when score centering is enabled.
• Risk: callers of generate see two new keys in the returned dictionary; existing fields are unchanged.

^{Macroscope summarized 424ba10.}

---

Note

Low Risk
Additive response pass-through only; no request, validation, or behavior changes when engines omit the new fields.

Overview
Extends generate() so callers get two more optional fields from prime-rl’s /inference/v1/generate response, with no request or parsing changes.

sampling_mask_logprobs is forwarded from the choice (packed float32 logprobs parallel to sampling_mask ids when score centering is enabled). payload is forwarded for by-handle segments pointing at a shared file when the server offloads routed_experts / sampling_mask instead of returning them inline. Both appear in the returned dict and docstring; absent keys become None, same as existing optional pass-throughs like sampling_mask.

^{Reviewed by Cursor Bugbot for commit 424ba10. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs Human Approval

Macroscope's review found this PR requires human approval — The PR is a small, backwards-compatible pass-through of optional server metadata and does not materially change existing request or response behavior. Human review is still warranted because two unresolved comments identify concrete reasoning-text correctness issues elsewhere in the PR branch.

Notes:

• Macroscope's correctness review did not run, so approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

Dismissing prior approval to re-evaluate d3a1af0

Dismissing prior approval to re-evaluate 030b208

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 030b208. Configure here.}

[cursor]

Reasoning text drops think markers

Medium Severity

scan_reasoning already skips the leading <think> opener, then deletes every later open_id and every <think> substring before decoding what remains. Reasoning that mentions the tag, or a truncated </think>, loses those characters, and neighboring tokens can merge into different text.

 

^{Reviewed by Cursor Bugbot for commit 030b208. Configure here.}

[cursor]

DeepSeek reasoning boundary whitespace stripped

Medium Severity

parse_deepseek_v4 always rstrips reasoning and turns whitespace-only content into "". The newline the model emits before </think>, and a content channel that is only that separator, disappear from the parsed message.

 

^{Reviewed by Cursor Bugbot for commit 030b208. Configure here.}