Avoid env-server trace recopy

[xeophon]

Overview

Keep completed V1 traces concrete inside env-server response wrappers and defer WireTask conversion to the existing client boundary. This removes a redundant full-trace dump and validation pass before the unchanged wire serializer.

Mechanism

episode.run() already returns typed, trusted Trace objects. The rollout path previously dumped each concrete trace to a dictionary, then response validation rebuilt it as Trace[WireTask]; the response field serializer subsequently dumped that rebuilt trace again for msgpack. Group responses repeated the same work for every trace.

The response wrappers now use model_construct around those trusted concrete traces. Their existing field serializers still produce the wire payload once, and the client continues validating decoded responses into Trace[WireTask]. Request handling, response schemas, msgpack encoding, ZMQ transport, task extras, excluded state, and client-side validation remain unchanged.

Performance

A PEP 723 response-path benchmark used one trace with one node and 1,825,000 token IDs, masks, and log probabilities (5,475,000 list entries total), producing a 27,244,201-byte msgpack response. Values below are medians of three fresh processes after warm-up on macOS.

Metric	Previous	Current	Saved
Wrapper construction	79.289 ms	0.018 ms	79.271 ms (99.98%)
Total synchronous response work	231.703 ms	150.555 ms	81.148 ms (35.02%)
Peak process RSS	399.8 MiB	358.3 MiB	41.5 MiB (10.38%)
Live response RSS growth	135.8 MiB	94.2 MiB	41.6 MiB (30.63%)
Post-cleanup RSS growth	135.8 MiB	94.2 MiB	41.6 MiB (30.63%)

The synchronous time reduction directly shortens the env-server event-loop stall for large responses. Group savings scale with the number and size of completed traces.

---

Note

Low Risk
Server-only response construction optimization on trusted rollout output; wire encoding and client validation paths are unchanged.

Overview
Rollout and group handlers no longer turn completed traces into dicts before building RunRolloutResponse / RunGroupResponse. They wrap the concrete episode.run() traces with model_construct instead of going through model_dump() plus full Pydantic validation on the way in.

That removes an extra full-trace copy and rebuild on the server while the existing trace / traces field serializers still run once when _handle calls model_dump for msgpack. WireTask re-typing stays on the client via model_validate on the decoded payload.

Group responses get the same treatment per trace, so savings grow with group size and trace payload size.

^{Reviewed by Cursor Bugbot for commit dd1ea56. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Avoid redundant trace serialization in EnvServer rollout and group responses

In server.py, _run_rollout and _run_group now use model_construct to build responses directly from existing Trace objects instead of calling model_dump() on each trace and re-validating. This skips an unnecessary serialize-then-validate round-trip per trace.

^{Macroscope summarized dd1ea56.}

[macroscopeapp]

Approvability

Verdict: Approved

Performance optimization using Pydantic's model_construct() to avoid redundant serialization and validation of trusted internal trace objects. Change is small, self-contained, and follows a standard Pydantic pattern.

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