EvoTrace

What it does

EvoTrace compiles the missing layer between a raw coding agent transcript and usable training data: stable task boundaries, reconstructable repository state, a replayable environment, a validated verifier, a privacy review and provenance. It imports Claude Code and Codex sessions, mines evidence backed trajectories, and produces preference data, RL environments and execution rewards that can be trained locally and evaluated. The framing is that a trajectory is not yet training data until that layer exists.

Architecture and agents

EvoTrace is now a specialized distribution of DeepSeek Harness, supplying the domain composition such as history import, trajectory mining, an Orchestrator that opens four least privilege subagents in sequence, and allowlisted compiler tools, while the Harness supplies the browser or agent shell, sessions, streaming, slash commands, credentials, approvals and plugin runtime. A second loop has an Explorer generate and execute repository questions, an Experience Compressor distill the grounded trajectory, and held out tasks test whether the packet improves a solver, judged by downstream execution success rather than by an LLM's opinion of the summary.

Posture and roadmap

The open source release works locally and the README states nothing is uploaded, sold or shared by default. A future opt in Marketplace and managed fine tuning integrations are roadmap products that would let users license reviewed, rights cleared assets instead of surrendering raw history to a data intermediary. The project is Apache-2.0 licensed, has a CI workflow, provides a Chinese README, and installs via a one line script on macOS, Linux and WSL.

Editorial conclusion

EvoTrace is Apache-2.0 licensed and reported eighty two stars at indexing, with the README stating the Marketplace is not yet available and the current release runs entirely locally.