Transfer V1 completion logprobs without copying

[xeophon]

Overview

Transfer the transient TurnTokens.completion_logprobs list directly into the assistant MessageNode during graph commit instead of materializing a second list.

Reasoning

_commit_turn constructs the node with MessageNode.model_construct, so it can preserve the already-typed list's identity without another validation or copy. TurnTokens is a per-response carrier that is discarded after commit, while graph and training consumers treat node logprobs as read-only. The graph node can therefore become the long-lived owner without changing logprob values, ordering, serialization, branching, or partial-logprob behavior.

Performance

A PEP 723 microbenchmark measured 2,000,000 logprob values over nine repetitions using time.perf_counter and tracemalloc:

Path	Median wall time	Peak traced allocation
List copy	2.818 ms	15.26 MiB
Ownership transfer	0.005 ms	0.00 MiB
Saved	2.813 ms (99.83%)	15.26 MiB (100%)

This removes 2,000,000 copied references (about 16,000,000 bytes of pointer storage) and the transient duplicate list buffer. Long-lived trace storage remains unchanged at one logprob list.

---

Note

Low Risk
Single-line ownership change in graph commit with no intended change to logprob values or training semantics; only risk is if something mutates the list before discard (unchanged from prior copy behavior for readers).

Overview
When committing an assistant turn in _commit_turn, the graph node now takes ownership of TurnTokens.completion_logprobs instead of building a second list with list(...).

TurnTokens is only used for that commit path and is dropped afterward, so the assistant MessageNode becomes the long-lived holder of the same list. This matches the existing model_construct pattern for large token slices and avoids extra allocation on long completions.

^{Reviewed by Cursor Bugbot for commit 2daa0c3. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Transfer V1 completion logprobs without copying in _commit_turn

In graph.py, _commit_turn now assigns tokens.completion_logprobs directly to MessageNode.logprobs instead of wrapping it in list(). Since TurnTokens is discarded after commit, the copy was unnecessary.

^{Macroscope summarized 2daa0c3.}

[macroscopeapp]

Approvability

Verdict: Approved

Minor performance optimization removing an unnecessary list copy. The change is well-documented and doesn't alter runtime behavior beyond avoiding allocation overhead.

^{You can customize Macroscope's approvability policy. Learn more.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2daa0c3e06

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Detach logprobs before committing the trace

When a caller keeps or reuses the Response/TurnTokens object after commit(), this stores the same mutable completion_logprobs list on the long-lived trace. Any later normalization, clearing, or buffer reuse by a custom train client mutates trace.nodes[-1].logprobs, and Branch.logprobs/trace dumps will then report corrupted training logprobs; the previous list(...) copy isolated committed graph state from the transient response object.

Useful? React with 👍 / 👎.