jinzijian/EvoTrace

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

EvoTrace – Turning Your Claude Code / Codex Sessions into Re‑usable AI Training Assets

What it is – EvoTrace is a local‑first tool that reads the coding‑assistant history you already have on your machine (Claude Code or Codex), extracts the useful “episodes” and turns them into well‑structured assets that can be used for model training, evaluation, or reinforcement‑learning pipelines. It does not replace your coding agent; it simply compiles what the agent has already done into evidence‑backed tasks, verifiers, and Docker‑ready environments.

Core idea

  1. Import existing Claude Code or Codex session logs.
  2. Mine the logs to find coherent coding episodes that contain enough signal (intent, actions, failures, fixes).
  3. Review each episode with a deterministic four‑agent pipeline that decides whether the episode should become:
    • preference/recovery data (for DPO, SFT, QA),
    • an executable task bundle (repo snapshot + spec),
    • a verifier + reward candidate, or
    • be rejected with a clear reason.
  4. Validate the generated task in an isolated Docker container, ensuring the base state is rejected and the reference state is accepted.
  5. Optional hardening, calibration, and “evolution” steps let you make the task harder or test transfer to a held‑out task.

What you get out of it

Asset What’s produced Typical downstream use
Candidate catalog Ranked list of episodes with intent, repo, corrections, provenance gaps Quickly spot the few high‑value histories in a large chat dump
Preference / recovery data Chosen vs. rejected attempts, human corrections, successful recoveries DPO, supervised fine‑tuning, failure‑recovery training
Executable task bundle Repo snapshot, initial state, dependency list, task spec Coding‑agent benchmarks, regression suites, RL environments
Verifier & reward candidate Test commands, behavioral checks, policy, provenance Scoring roll‑outs, generating RL reward signals
Difficulty evidence Fresh solver attempts + verifier outcomes Curriculum construction, difficulty‑aware sampling
Execution experience Compressed runtime facts from exploration trajectories Training examples for experience‑transfer experiments

How it works (high‑level pipeline)

Claude/Codex history → /init (import + Git archaeology) → /candidates (rank) →
/review (four‑agent review: miner → gate → builder/hardener → verifier) →
[optional] /build → /validate (Docker two‑state run) → verified asset

The four agents are least‑privilege roles that run sequentially, so no single component can both build and approve an asset. Validation runs in a disposable Docker container with no host mounts or privileged access, keeping your source code safe.

Getting started (quick steps)

  1. Install – one‑liner script works on macOS, Linux, or WSL; Windows has a PowerShell variant.
  2. Run – evotrace launches the DeepSeek Harness web UI.
  3. Configure – Choose a provider (DeepSeek, OpenAI, Anthropic) in Settings; mining and import need no model.
  4. Create your first asset – In the UI type the slash commands:
    /init
/candidates
/show 1
/review 1
/assets
    The review step will either produce a verified task or give a concrete rejection reason.

When it’s useful

  • Research labs that want to harvest real‑world coding‑assistant interactions for DPO or RL‑from‑human‑feedback datasets.
  • Tool builders needing reproducible coding‑task benchmarks (e.g., “fix this bug” with exact repo state).
  • Developers who want to keep a personal archive of high‑quality coding sessions without sending raw logs to a cloud model.
  • Curriculum designers who need difficulty‑graded tasks measured by independent solver attempts.

Limitations & current status

  • Alpha / preview – built on the developer‑preview DeepSeek Harness; many features are still experimental.
  • Docker‑only validation – you need Docker installed to generate or verify executable bundles.
  • Supported languages / environments – currently focused on the languages and dependency setups that DeepSeek Harness can reconstruct; cross‑language or complex system‑level builds are on the roadmap.
  • Marketplace & fine‑tuning export – planned future products; the open‑source release only provides local asset generation.

Installation from source (for developers)

git clone https://github.com/jinzijian/EvoTrace.git
cd EvoTrace
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
pnpm install   # Node 22.19+ or 24+
pnpm dev       # runs the Harness UI locally

The Python CLI (et --help) offers the same deterministic compilation functions for scripting.

License

Apache‑2.0 – free for commercial and academic use.

Bottom line – EvoTrace is a specialized pipeline that bridges the gap between raw coding‑assistant chat logs and the structured, reproducible assets needed for modern AI‑driven code‑generation research. It does so locally, keeps your data under your control, and provides a clear audit trail from raw session to verified Docker task.

Related

  • Project
    oleksiijko/pmb

    PMB is an open‑source Python library that gives LLM‑based coding assistants a local, SQLite‑backed memory. It automatically indexes code, PDFs, git history, and user‑recorded facts, then injects the most relevant pieces back into the agent via the Model‑Context‑Protocol. The system works offline, offers a hybrid BM25‑+‑vector search (≈35 ms warm latency), provides a visual dashboard, and includes hooks that both recall memory before the model thinks and auto‑write observed actions. Install with `pip install pmb-ai`, run `pmb setup` to wire an agent, and the agent will remember decisions, lessons, and project context across restarts without any API keys or cloud calls.

  • Project
    sv-number/skills

    A minimal, MIT‑licensed *skill* that lets AI agents obtain disposable phone numbers and read SMS verification codes via a commercial API, enabling automated sign‑ups and 2FA without human interaction.

  • Project
    FonaTech/Clouds-Coder

    Clouds Coder is a Python runtime that separates CLI execution from a browser‑based IDE, adds AI‑agent orchestration, and provides a LAN‑only collaborative workspace with versioned files, conflict‑resolution, and admin‑approved tool policies.

  • Project
    NaomiProject/Naomi

    Naomi is an open‑source, always‑on voice‑assistant platform (Python, Linux/Raspberry Pi) that lets you control home devices, query information, and add custom “skills” via simple Python modules.