EvoTrace

Turn real-world Claude Code and Codex sessions into reusable training, evaluation, and verification assets.

A local-first trajectory compiler built on DeepSeek Harness.

Get started · What you get · Workflow · Architecture · 中文

EvoTrace imports the coding-agent work already stored on your machine, finds the sessions worth keeping, and compiles them into evidence-backed preference data, replayable coding tasks, RL environments, and execution-reward candidates. You keep the source data and the resulting assets under your control.

It is not another coding agent and does not require you to change how you use Claude Code or Codex.

[!WARNING] EvoTrace is early alpha, and DeepSeek Harness is a developer preview. Generated tasks and verifiers remain candidates until they pass the documented evidence and Docker validation gates.

Get started

1. Install

macOS, Linux, or WSL:

curl -LsSf https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.sh | sh

Windows PowerShell:

irm https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.ps1 | iex

2. Launch

evotrace

EvoTrace opens the DeepSeek Harness Web app. In Settings, choose any provider supported by your Harness setup, such as DeepSeek, OpenAI, or Anthropic. Local import and deterministic mining do not require a model; agent review, hardening, calibration, and evolution do.

New sessions start with Harness Full access by default. To use the narrower workspace sandbox:

DSH_PERMISSION_MODE=workspace-write evotrace

Docker is required only when you build or validate executable environments.

3. Build your first asset

Type / in the app and run:

/init              import existing Claude Code and Codex history
/candidates        show the strongest evidence-backed sessions
/show 1            inspect candidate 1 and its missing evidence
/review 1          run the sequential four-agent review
/assets            inspect anything that was compiled or verified

That is the main product loop. A review may build an asset, route the session to preference data, keep it only as a hardening seed, or reject it with explicit reasons. Rejection is a useful result: it prevents a long but weak trajectory from being mislabeled as training-ready.

What you get

Output	Recovered or generated from your sessions	Useful for
Candidate catalog	task intent, repo, corrections, failures, effective actions, provenance gaps	finding the small fraction of history worth keeping
Preference and recovery data	rejected/chosen attempts, human corrections, successful recoveries	DPO, SFT, QA, failure-recovery training
Executable task bundle	repository base, initial state, dependency evidence, task specification	coding-agent evals, regression tasks, RL environments
Verifier and reward candidate	test commands, behavioral checks, policy, provenance	rollout scoring and execution rewards after validation
Difficulty evidence	fresh independent solver attempts and verifier outcomes	curriculum construction instead of guessing from patch size
Execution experience	grounded runtime facts compressed from exploration trajectories	training examples and held-out experience-transfer experiments