EvoTrace

作者 jinzijian已验证

Compile real-world Claude Code and Codex trajectories into verified, tradable post-training assets.

127
Stars
4
Forks
Python
语言
2026/8/23
添加时间

⚠️ 第三方软件声明

本 Skill 为第三方开源软件,独立托管于 GitHub。SkillTip 仅为信息目录,不控制或维护底层仓库。所显示的安全检查为自动化且范围有限,安装前请自行审查源码。

阅读服务条款

安装

添加到你的 Claude Code skills 目录:

# Add to your Claude Code skills
git clone https://github.com/jinzijian/EvoTrace

快速入门

使用 EvoTrace 等 Skills 的指南。

安全报告

已验证

上次扫描:—

{
  "status": "PASSED",
  "issues": []
}

README.md

EvoTrace logo

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

CI DeepSeek Harness License: Apache-2.0

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.

Import, mine, review, and build with EvoTrace

What you get

OutputRecovered or generated from your sessionsUseful for
Candidate catalogtask intent, repo, corrections, failures, effective actions, provenance gapsfinding the small fraction of history worth keeping
Preference and recovery datarejected/chosen attempts, human corrections, successful recoveriesDPO, SFT, QA, failure-recovery training
Executable task bundlerepository base, initial state, dependency evidence, task specificationcoding-agent evals, regression tasks, RL environments
Verifier and reward candidatetest commands, behavioral checks, policy, provenancerollout scoring and execution rewards after validation
Difficulty evidencefresh independent solver attempts and verifier outcomescurriculum construction instead of guessing from patch size
Execution experiencegrounded runtime facts compressed from exploration trajectoriestraining examples and held-out experience-transfer experiments

The same validated task can evaluate an agent today, score newly sampled rollouts tomorrow, and produce verifier-grounded RL data later. The future opt-in EvoTrace Marketplace and fine-tuning integrations are intended to let users license reviewed assets on terms they control; they are roadmap products, not part of the current local release.

Why raw trajectories are not enough

A transcript may contain a prompt, messages, commands, and diffs, but post-training needs more:

  • one coherent task boundary rather than an entire chat;
  • the repository state from before the task began;
  • a reproducible dependency and execution environment;
  • an independent verifier that rejects the base state and accepts a known-good state;
  • provenance tying every task, patch, verifier, and run together;
  • difficulty measured by fresh attempts rather than token count or patch size.

EvoTrace automates that compilation gap with session import, Git/repository archaeology, deterministic gates, specialized agents, and isolated execution.

The core workflow

Claude Code / Codex history
           │
           ▼
        /init         import + normalize + Git archaeology
           │
           ▼
     /candidates      rank evidence, hide nested subagent duplicates
           │
           ▼
       /review        mine episode → gate route → build/harden → criticize
           │
      ┌────┴───────────────┐
      ▼                    ▼
preference/recovery   executable candidate
                           │
                           ▼
                    /validate in Docker
                           │
                           ▼
                 verified reward environment

EvoTrace trajectory-to-post-training pipeline

Status means evidence

StatusWhat it actually means
MinedThe session has useful signals. Nothing executable is implied.
BuildableTask, repo base, reconstruction confidence, reference patch, verifier commands, and environment gates pass.
Bundle generatedA Docker-ready candidate exists. Its verifier is not yet trusted.
VerifiedA conforming Docker run rejected the base, accepted the reference, and was recorded against the exact bundle digest.
CalibratedFresh solver attempts measured the task; the default target is two verifier passes in five attempts.

EvoTrace fails closed on empty prompt wrappers, low-confidence reconstruction, missing reference patches, missing verification commands, unsupported environments, candidate switching, and mismatched asset lineage.

Common recipes

Mine useful history without sending it to a model

/init
/candidates
/show 1

Compile and independently validate an executable task

/review 1
/build 1
/validate 1
/runs

Make an easy verified task meaningfully harder

/harden 1
/calibrate 2

Hardening must add testable behavior, compatibility, edge cases, or failure constraints. Making a patch longer is not treated as making a task harder.

Test whether execution experience transfers

/evolve 1 2

Asset 1 is explored and compressed; asset 2 must be an independently built held-out task from the same repository. Baseline and conditioned solver attempts are then compared using Docker rewards. Running /evolve 1 without a held-out asset is only a wiring smoke test and cannot certify transfer.

Command reference

CommandPurpose
/init [all|codex|claude]import history and refresh mining
/candidatesbrowse ranked candidates
/search payment retrysearch tasks, repositories, and evidence
/show 1inspect provenance and readiness gaps
/review 1run the sequential review pipeline
/build 1compile an execution candidate
/validate 1run two-state Docker validation
/harden 1derive and test a harder child task
/calibrate 1measure and adapt difficulty with self-play
/evolve 1 2test compressed experience on a held-out task
/assetslist compiled assets and their states
/runsinspect saved validation evidence
/doctorcheck local integrations

Built on DeepSeek Harness

EvoTrace is a specialized distribution of DeepSeek Harness. Harness supplies the Web UI, sessions, streaming, provider settings, permission surface, slash commands, and plugin runtime. EvoTrace adds the trajectory compiler and one managed Orchestrator with four foreground, least-privilege roles:

StageResponsibilityCannot do
Episode Minerisolate one coherent episode and count effective actionsbuild or approve an asset
Candidate Gatejudge value, complexity, reconstructability, and record one immutable routemutate data or change the route later
Task Builder / Hardenerbuild the exact routed candidate or derive a harder childedit the source checkout or approve itself
Verifier Criticaudit Docker runs, verifier evidence, lineage, and difficultycertify missing evidence

The children run sequentially, never in parallel. Review-bound tools enforce the exact review token, candidate ID, route, and produced-asset lineage in code rather than relying only on prompts.

EvoTrace on DeepSeek Harness

Install from source

Requires Git, Node.js 22.19+ or 24+, Python 3.9+, and optionally Docker.

git clone https://github.com/jinzijian/EvoTrace.git
cd EvoTrace
python3 -m venv .venv
.venv/bin/python -m pip install -e .
pnpm install
pnpm dev

The Python CLI remains available as a deterministic compiler and automation sidecar. Run et --help for its machine-oriented commands; the DeepSeek Harness app is the primary interface.

Execution and trust boundaries

  • Import and mining read local Claude Code/Codex history and Git evidence without changing source repositories.
  • Codex subagent and fork trajectories keep parent lineage but are hidden from the default candidate list.
  • The Orchestrator exposes fixed domain tools rather than arbitrary host shell or filesystem tools.
  • Validation runs in disposable Docker worlds without source bind mounts, the Docker socket, host networking, privileged mode, or host credentials.
  • Builder and Verifier Critic are separate child sessions; a builder cannot approve its own verifier.
  • Self-play and evolution are explicit operations because they send selected task context to the configured model.

Read the normative sandbox contract, task quality standard, schema, and design.

Current release and roadmap

The current open-source release includes local Claude Code/Codex import, evidence mining, sequential agent review, environment reconstruction, Docker bundle generation, two-state validation, self-play calibration, semantic task hardening, experience compression, and held-out transfer measurement.

Still in progress:

  • broader cross-language dependency repair and autonomous environment construction;
  • stronger hidden behavioral verifiers and adversarial task mutation;
  • validated DPO, SFT, and RL dataset exporters;
  • opt-in Marketplace and managed fine-tuning integrations.

Acknowledgements

  • DeepSeek Harness is the application and agent foundation.
  • Microsoft RepoLaunch is a primary inspiration for reproducible repository-to-environment construction. EvoTrace begins earlier by mining tasks and learning signals from lived coding-agent work.

License

Apache-2.0. See LICENSE and NOTICE.

常见问题

What is EvoTrace?

EvoTrace is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by jinzijian. Compile real-world Claude Code and Codex trajectories into verified, tradable post-training assets. It has 127 GitHub stars.

Is EvoTrace safe to use?

Yes. EvoTrace passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.

How do I install EvoTrace?

Clone the repository with "git clone https://github.com/jinzijian/EvoTrace" and add it to your Claude Code skills directory (see the Installation section above).

What programming language is EvoTrace written in?

EvoTrace is primarily written in Python. It is open-source under jinzijian on GitHub, so you can review or fork the full source.

Are there alternatives to EvoTrace?

Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh EvoTrace against similar tools.

评论 (0)

暂无评论,成为第一个分享想法的人!

ECC

by affaan-m

10

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

242,21936,702JavaScript
AI 智能体ai-agentsanthropicclaude-code
查看详情
15

An agentic skills framework & software development methodology that works.

234,96620,863Shell
AI 智能体ai-agentsbrainstorming
查看详情

hermes-agent

by NousResearch

10

The agent that grows with you

234,43747,175Python
AI 智能体ai-agentsagent-orchestration
查看详情

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

185,94028,768JavaScript
AI 智能体ai-agentsanthropicclaude-code
查看详情

cc-switch

by farion1231

3

A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Grok Build & Hermes Agent. Only official website: ccswitch.io

128,8688,826Rust
AI 智能体claude-codeai-tools
查看详情

claude-code

by anthropics

Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.

120,03119,897Shell
AI 智能体
查看详情

开发者还喜欢

基于喜欢此 Skill 的开发者投票和收藏

ECC

by affaan-m

10

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

242,21936,702JavaScript
AI 智能体ai-agentsanthropicclaude-code
查看详情
15

An agentic skills framework & software development methodology that works.

234,96620,863Shell
AI 智能体ai-agentsbrainstorming
查看详情

hermes-agent

by NousResearch

10

The agent that grows with you

234,43747,175Python
AI 智能体ai-agentsagent-orchestration
查看详情

n8n

by n8n-io

12

Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.

201,88160,308TypeScript
MCP 服务器apisai-tools
查看详情

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

185,94028,768JavaScript
AI 智能体ai-agentsanthropicclaude-code
查看详情

cc-switch

by farion1231

3

A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Grok Build & Hermes Agent. Only official website: ccswitch.io

128,8688,826Rust
AI 智能体claude-codeai-tools
查看详情