SkillFoxx

EvoTrace CLI

A local trajectory compiler that turns Claude Code and Codex session history into training data, tasks and verifiers

  • 281 ★ on the repository
  • Last commit
  • Apache-2.0
CLI

Broad access

Why this level

  • Reads the whole Claude Code and Codex session history, which can contain code, paths and secrets from conversations
  • New sessions start with full access by default: the app's agents run commands and change files
  • Session text goes to the chosen model provider, and building and validating start Docker
All reasons and checks

jinzijian/evotraceAuthor:jinzijian

Install

Manual install

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

For macOS, Linux and WSL. On Windows use install.ps1 from the repository. Read the script first.

Access and checks

How we review

Broad accessThe tool can change data, run commands or handle secrets.

  • Reads the whole Claude Code and Codex session history, which can contain code, paths and secrets from conversations
  • New sessions start with full access by default: the app's agents run commands and change files
  • Session text goes to the chosen model provider, and building and validating start Docker

This is third-party code. Review the repository files before installing.

What it does

EvoTrace reads the agent work history already stored on your machine and picks the sessions worth keeping for model training. It installs with one script, and the `evotrace` command opens a web app built on DeepSeek Harness. In settings you choose any supported model provider: DeepSeek, OpenAI, Anthropic. Import and deterministic mining work without a model, while agent review, hardening and calibration need one. You type commands starting with a slash in the app. `/init` imports Claude Code and Codex history, `/candidates` lists the best-evidenced sessions, `/show 1` opens a candidate with its missing evidence, `/review 1` runs a sequential four-agent review, `/assets` shows what was built. A session ends up as preference and failure-recovery data, as an executable task bundle with a repository, dependencies and a verifier, or is rejected with reasons. Statuses mark the stage of evidence: mined, buildable, bundle generated, verified in Docker, calibrated with fresh solver attempts. EvoTrace fails closed on empty prompt wrappers, low reconstruction confidence, a missing reference patch or missing verification commands. Docker is needed only for building and validating executable environments.

Who it is for. Researchers and teams with many Claude Code and Codex sessions who want to build fine-tuning sets, agent evaluations and RL environments from them.

Good fit when

  • Months of work left hundreds of sessions, and you need the few with a clear task, human corrections and a verifiable result
  • You need pairs of accepted and rejected agent attempts for DPO or SFT
  • You want to turn a successful session into a regression task for evaluating agents, with Docker and a verifier
  • Task difficulty should be measured with fresh solver attempts instead of patch size
  • The history must not leave your machine: import and mining run locally without a model

Not a fit when

  • You want a new coding agent: EvoTrace processes the history of agents that already worked and does not change how you use Claude Code or Codex
  • You have few or short sessions, which will not yield a usable set
  • You expect a ready data marketplace: it is on the roadmap and not part of the current local release
  • You need a stable product: the project is early alpha and DeepSeek Harness is a developer preview

Example request

/init, then /candidates: show the five sessions with the strongest evidence and tell me what the first one lacks to become a task

Limitations

  • The project is early alpha and its base, DeepSeek Harness, is a developer preview. Generated tasks and verifiers stay candidates until they pass the evidence and Docker gates.
  • New sessions start in Full access mode by default. You can limit writing to the workspace by launching `DSH_PERMISSION_MODE=workspace-write evotrace`.
  • Review, hardening and calibration need a model provider with your key and spend tokens.
  • Building and validating executable environments requires Docker.
  • The marketplace and fine-tuning integrations are described as plans and are not in the current version.

How to disable. Delete the `evotrace` executable the install script placed and the app's data in its working folder. EvoTrace only reads the original Claude Code and Codex sessions, so they do not need to be removed. For the exact install path see the output of install.sh or install.ps1.

FAQ

How do I install EvoTrace?

On macOS, Linux and WSL run `curl -LsSf https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.sh | sh`, on Windows the command from install.ps1 in PowerShell. Then run `evotrace`; the web app opens and in settings you pick a model provider.

Are my sessions sent to the cloud?

Import and deterministic mining run locally and need no model. Text goes to the provider you chose only during review, hardening and calibration. The source data and resulting sets stay under your control.

What does the Verified status mean?

A Docker run recorded against the exact bundle digest rejected the base state and accepted the reference. Until then the bundle is a candidate, and Calibrated adds fresh solver attempts: the default target is two verifier passes in five attempts.

Do I have to change how I work with Claude Code?

No. EvoTrace does not replace the agent and only reads the saved Claude Code and Codex history.

How do I limit the app's permissions?

Run `DSH_PERMISSION_MODE=workspace-write evotrace`. By default new sessions start with full access.

Discussion

No questions or reviews yet.

CLI tools

opentraces records Claude Code, Codex and Pi agent work into a local log, anchors edits to commits and prepares datasets.

CLIExtended access104

A skill and web viewer that turn Claude Code, Codex and Pi logs into a decision map of the agent run with links to originals.

SkillExtended access90

A CLI that reads local Claude Code, Codex and Hermes sessions, finds repeated failures and stages fixes for manual review

CLIExtended access73

Agent memory over MCP: past work trajectories become guidelines for new tasks, with a lite version for Claude Code and Codex

MCP serverExtended access137

A Claude Code plugin with hooks and 26 skills: measures what the agent knows, blocks edits until code is read, keeps memory

PluginExtended access255

Local analytics for Claude Code, Codex and five more agents: tokens, cost, time and log search in one archive

CLIExtended access75

README in short

The page explains how a raw transcript differs from a training asset: it needs task boundaries, the repository state before the work, a reproducible environment, an independent verifier and provenance. Next come installation and first launch, a table of outputs, a pipeline diagram, a decoding of statuses and recipes for working without a model and for validating in Docker. The early alpha status is stated separately.

README badge

Report a problem