# AI use cases for Evo Trace (jinzijian/EvoTrace)

Repo: [jinzijian/EvoTrace](https://github.com/jinzijian/EvoTrace) | Use cases: 4 | Page: https://dedupedai.com/use-cases/jinzijian-evotrace/

EvoTrace is a local-first compiler that turns real Claude Code and Codex sessions into evidence-backed training data, replayable coding tasks, and verified execution rewards, built on DeepSeek Harness.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- Find the worth-keeping sessions in your coding history: https://dedupedai.com/use-cases/jinzijian-evotrace/mine-coding-history-for-candidates/
- Compile a session into a Docker-validated coding task: https://dedupedai.com/use-cases/jinzijian-evotrace/build-and-validate-executable-task/
- Make a verified task harder and measure its difficulty: https://dedupedai.com/use-cases/jinzijian-evotrace/harden-and-calibrate-verified-task/
- Test whether compressed experience transfers to a new task: https://dedupedai.com/use-cases/jinzijian-evotrace/test-experience-transfer-held-out-task/

## How to find the worth-keeping sessions in your coding history with Evo Trace

Repo: [jinzijian/EvoTrace](https://github.com/jinzijian/EvoTrace) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Import local Claude Code and Codex history and rank the sessions that hold useful task evidence.

**Who it is for:** An ML engineer or developer with months of Claude Code or Codex sessions who wants to know which sessions are worth turning into training or eval data, without sending them to a model.

**The problem:** Raw agent transcripts are long and mostly weak. Finding the few coherent, verifiable sessions by hand is slow.

**The outcome:** A ranked candidate list with task intent, repository, corrections, failures, and the evidence gaps for each session.

### Steps

1. Run `curl -LsSf https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.sh | sh` on macOS, Linux, or WSL.
2. Run `evotrace` to open the DeepSeek Harness Web app.
3. In the app, run `/init` to import existing Claude Code and Codex history.
4. Run `/candidates` to show the strongest evidence-backed sessions.
5. Run `/show 1` to inspect candidate 1 and its missing evidence.

### Prompt to paste into your agent

```text
Using the EvoTrace repo at https://github.com/jinzijian/EvoTrace, install it and run /init then /candidates in the DeepSeek Harness app on my local Claude Code and Codex history. Return the top 5 candidates with task intent, repository, and missing evidence for each. Verify that the candidate list loads with no model provider configured.
```

### Caveat

Evo Trace is early alpha, and this use case only mines and ranks candidates; nothing is verified or trained yet.

Risk, remote script: The install command pipes a script from GitHub straight into sh, so read install.sh first if you do not trust the source.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/jinzijian-evotrace/mine-coding-history-for-candidates/

## How to compile a session into a Docker-validated coding task with Evo Trace

Repo: [jinzijian/EvoTrace](https://github.com/jinzijian/EvoTrace) | Difficulty: Advanced | Time: 2 hours | Basis: From the README

Review a candidate, build an executable bundle, and run two-state Docker validation on it.

**Who it is for:** A developer building regression tasks or RL environments from real agent work who needs each task to be independently checked before use.

**The problem:** A transcript is not a task. It needs a repo base, reference patch, verifier, and a Docker run that rejects the base and accepts the reference.

**The outcome:** A bundle with a recorded Docker validation run that moves from generated to verified, or a clear rejection with reasons.

### Steps

1. Install and launch Evo Trace with `curl -LsSf https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.sh | sh` and then `evotrace`.
2. Configure a model provider in Harness Settings, and make sure Docker is running.
3. Run `/review 1` to run the sequential four-agent review on candidate 1.
4. Run `/build 1` to compile an execution candidate, then `/validate 1` to run two-state Docker validation.
5. Run `/runs` to inspect the saved validation evidence and check the status.

### Prompt to paste into your agent

```text
Using the EvoTrace repo at https://github.com/jinzijian/EvoTrace, run /review 1, /build 1, and /validate 1 in the DeepSeek Harness app for the candidate <candidate-number>. Report the asset status and the Docker validation result from /runs. Verify that the asset is marked Verified only if the base was rejected and the reference was accepted.
```

### Caveat

Generated tasks and verifiers stay candidates until they pass the documented evidence and Docker validation gates, and Docker is required for this path.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/jinzijian-evotrace/build-and-validate-executable-task/

## How to make a verified task harder and measure its difficulty with Evo Trace

Repo: [jinzijian/EvoTrace](https://github.com/jinzijian/EvoTrace) | Difficulty: Advanced | Time: 2 hours | Basis: From the README

Derive a harder child task from a verified one, then measure it with fresh solver attempts.

**Who it is for:** A researcher building a difficulty curriculum for coding agents who needs measured difficulty rather than guesses from patch size.

**The problem:** Easy verified tasks saturate quickly, and making a patch longer does not make a task harder.

**The outcome:** A harder child task with a calibration result based on fresh solver attempts against the default target of two verifier passes in five attempts.

### Steps

1. Install and launch Evo Trace with `curl -LsSf https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.sh | sh` and then `evotrace`.
2. Make sure a verified asset exists and Docker is available.
3. Run `/harden 1` to derive and test a harder child task.
4. Run `/calibrate 2` on the child task to measure difficulty with self-play.

### Prompt to paste into your agent

```text
Using the EvoTrace repo at https://github.com/jinzijian/EvoTrace, run /harden 1 and then /calibrate on the resulting child asset in the DeepSeek Harness app. Report the calibration pass count out of attempts and the added behavior. Verify that the child task adds testable behavior, compatibility, edge cases, or failure constraints rather than only a longer patch.
```

### Caveat

Self-play and evolution send selected task context to the configured model, so only run them on material you are allowed to send.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/jinzijian-evotrace/harden-and-calibrate-verified-task/

## How to test whether compressed experience transfers to a new task with Evo Trace

Repo: [jinzijian/EvoTrace](https://github.com/jinzijian/EvoTrace) | Difficulty: Advanced | Time: 3 hours | Basis: From the README

Compress execution experience from one task and score it on an independent held-out task.

**Who it is for:** An ML researcher studying whether agent experience generalizes within a repository and needs a controlled held-out comparison.

**The problem:** Reusing lessons from one task proves little unless a separately built task from the same repository measures the effect.

**The outcome:** A comparison of baseline and experience-conditioned solver attempts on a held-out task, scored by Docker rewards.

### Steps

1. Install and launch Evo Trace with `curl -LsSf https://raw.githubusercontent.com/jinzijian/EvoTrace/main/install.sh | sh` and then `evotrace`.
2. Build two independent verified assets from the same repository using `/review`, `/build`, and `/validate`.
3. Run `/evolve 1 2` so asset 1 is explored and compressed and asset 2 is the held-out task.
4. Run `/runs` to compare the baseline and conditioned solver attempts and their Docker rewards.

### Prompt to paste into your agent

```text
Using the EvoTrace repo at https://github.com/jinzijian/EvoTrace, build two independent verified assets from the same repository, then run /evolve <source-asset> <held-out-asset> in the DeepSeek Harness app. Report baseline and conditioned Docker reward results from /runs. Verify that the held-out asset was built independently and that a run with only one asset is treated as a wiring smoke test, not transfer evidence.
```

### Caveat

Running /evolve with only one asset is only a wiring smoke test and cannot certify transfer.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/jinzijian-evotrace/test-experience-transfer-held-out-task/