EvoTrace

Turn every Claude Code and Codex session into reusable training, evaluation, and verification assets.

Quickstart · What you get · How it works · Security · 中文

Your agent sessions are not disposable chat logs. They are compounding data assets.

⚡ Quickstart

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

Then run one command:

evotrace init

EvoTrace discovers existing Claude Code and Codex sessions, indexes them locally, and mines useful candidates. The installer only installs the CLI—it does not read session data. You can inspect the Unix installer or the PowerShell installer before running it.

Already use uv?

uv tool install git+https://github.com/jinzijian/evotrace.git
evotrace init

Prerequisites: Git plus uv, pipx, or Python 3.9+. Docker is only needed when running sandboxed execution. et is the short command; the older CLI names remain compatibility aliases.

What you get

These are composable asset layers, not separate silos. The same executable task and verifier can benchmark an agent, generate new rollouts, score their outcomes, and feed the verified trajectories back into training. Quality comes from executable outcomes and preserved provenance—not from treating every raw transcript as training data.

[!NOTE] Product direction: EvoTrace starts local-first. A future opt-in data marketplace will let users publish or license reviewed assets, and fine-tuning integrations will train on datasets users explicitly select. Nothing is shared by default.

EvoTrace works with the agents developers already use. No proxy, hosted agent, or new editor is required. The V0.3 curator is deterministic and auditable: it does not call an LLM or upload session data.

[!WARNING] EvoTrace is an early alpha. Mining labels and generated verifiers are evidence, not proof of task quality or semantic correctness. Inspect every eval before relying on it or sharing it.

How it works

Relationship to RepoLaunch

Microsoft RepoLaunch is a primary technical inspiration for EvoTrace's executable-environment layer. RepoLaunch demonstrates that an agent can turn a repository and base commit into a Docker environment with reproducible build commands, test commands, test-output parsing, and per-test execution— infrastructure that can support both SWE benchmarking and agentic SFT/RL. See the RepoLaunch paper.

EvoTrace starts one layer earlier: before a curated task dataset exists, it mines lived Claude Code and Codex work to identify which tasks, corrections, recoveries, and execution signals are worth turning into learning assets.

The planned integration seam is an optional RepoLaunch-compatible environment backend for candidates whose build and test environment cannot be recovered conservatively. EvoTrace remains responsible for history import, task selection, trajectory curation, provenance, privacy, and user-controlled distribution. No RepoLaunch code is vendored in the current release.

Core workflow

evotrace init — import and mine in one command

evotrace init
evotrace init --source codex --last 50

Use init on day one. It combines automatic discovery, incremental import, and local mining into a single onboarding command. The commands below expose each stage when you want more control.

evotrace import — index history you already have

# Discover both sources and index all available sessions.
evotrace import

# Limit or select a source.
evotrace import codex
evotrace import claude --last 20

# Import exact files.
evotrace import codex ~/.codex/sessions/2026/08/19/rollout-*.jsonl
evotrace import codex ~/.codex/history.jsonl
evotrace import claude ~/.claude/projects/my-project/session.jsonl

The importer honors $CODEX_HOME and $CLAUDE_CONFIG_DIR. For Codex it recognizes rich session JSONL files and the lighter prompt history separately, keeping the richer copy when both identify the same session. Claude Code documents plaintext session transcripts under ~/.claude/projects/ and a default 30-day cleanup window; installing EvoTrace early preserves a normalized local index before old transcripts disappear. See the Claude Code session docs, Claude Code data-path docs, Codex configuration reference, and Codex CLI resume reference.

Imports are incremental: unchanged source files are skipped using size and modification-time fingerprints. Use --refresh to force re-indexing. Raw history files are read in place, never copied into the EvoTrace store.

evotrace mine — find valuable experience

evotrace mine
evotrace mine --source codex --min-score 4
evotrace mine --json

V0.3 scores only observable signals: non-trivial task text, code-edit calls, verification commands, failed then successful checks, human corrections after agent work, and repository reconstruction confidence. Every candidate contains its score, labels, signals, and human-readable evidence in ~/.evotrace/candidates/.

• preference_candidate: a likely rejected/chosen or correction pair for DPO, preference, QA, or SFT curation.
• execution_verifiable: code changes plus recovered verification and a reconstructable repository base.
• recovery_trajectory: a failure or correction followed by subsequent repair work.

This deliberately avoids model-judged labels in the first release. A model curator can later sit behind the same schema without weakening provenance.

evotrace build — compile executable eval assets

# Build the highest-ranked execution-verifiable candidates.
evotrace build --limit 10

# Build one session and optionally add trusted verifier commands.
evotrace build SESSION_ID \
  --verify "python -m pytest tests/integration -q" \
  --verify "python -m ruff check src"

A history transcript is not a complete environment. The builder combines session evidence with the local Git repository: it uses a commit captured by the session when available, otherwise tries a time-aligned Git commit, and records reconstruction confidence instead of pretending the result is exact. It then emits:

benchmark-id/
├── task.md
├── task.json
├── task.yaml
├── verifier.py
├── verifier.json
├── sandbox-policy.json
├── setup.sh
├── Dockerfile
├── environment/
│   ├── base.tar.gz
│   ├── environment.json
│   └── untracked-initial.tar.gz
└── patches/
    ├── initial.patch
    └── reference.patch

Verifier provenance is always visible: explicit user command, trajectory-recovered command, repository convention, or a warning that no behavioral verifier was found.

Day two and beyond

Keep the local index current without placing EvoTrace in front of either agent:

evotrace watch                 # poll every five minutes and re-run mining
evotrace watch --interval 60
evotrace watch --once          # useful in cron or a nightly job

The watcher reads changed history files only. It does not modify Clau