---
title: Track past runs
description: Follow an evaluation's outcomes and flaky cells across recorded runs.
---
> ## Documentation Index
> Fetch the complete documentation index at: https://docs.reeval.ai/llms.txt
> Use this file to discover all available pages before exploring further.

You follow each evaluation's outcomes across the runs you keep in Git, and you find cells that both passed and failed. For runnable projects, see [`with-history`](https://github.com/invertase/reeval/tree/main/examples/with-history) and [`with-query`](https://github.com/invertase/reeval/tree/main/examples/with-query). For more information about the `record` and `history` options, see [Recording](/evaluations/runs/recording).

## Follow outcomes across recorded runs

<Steps>
<Step title="Add a recording profile">

In the `reeval.config.ts` file, add a profile named `track`. Leave every other key in the `reeval.config.ts` file in place.

```ts title="reeval.config.ts"
import { defineConfig } from '{{ pkg }}'

export default defineConfig({
  profiles: {
    track: {
      record: 'always',
    },
  },
})
```

A run that uses the `track` profile is copied into the history folder when its status is `complete` or `budget`. Evaluations that fail still leave the run `complete`. An interrupted run stays in the `.reeval/runs` directory.

With the default include patterns, the history folder is the `evals/.history` directory. For more information about profiles, see [Profiles](/evaluations/runs/profiles).

</Step>
<Step title="Run the profile">

Run the evaluations with the `track` profile:

```bash
npx reeval run --profile track
```

The final output line confirms the copied record:

```text
recorded evals/.history/<run-id>.json + <n> artifact files …
```

</Step>
<Step title="Commit the history folder">

Commit the history folder beside the evaluations:

```bash
git add evals/.history
git commit -m "Record the evaluation run"
```

</Step>
<Step title="List recorded runs">

List up to the 20 newest recorded runs, oldest to newest:

```bash
npx reeval history
```

For more information about the strips, see [Browse history](/results/inspection/queries#browse-history). For more information about the flags, see [`reeval history`](/reference/cli#reeval-history).

</Step>
<Step title="Read the evaluation trend">

Read each evaluation's outcome on each run:

```bash
npx reeval query eval_trend
```

Each row is one evaluation on one run, with its outcome and its pass rate, `pass_at_1`. The outcome is `pass`, `fail`, `infra_error`, or `skipped`.

A run in the history folder has `recorded` set to 1. A run that exists only in the `.reeval/runs` directory has `recorded` set to 0. For more information about choosing columns, see [Query past runs](/results/inspection/queries#write-sql).

</Step>
<Step title="Find flaky cells">

Find cells that both passed and failed:

```bash
npx reeval query flaky_cells
```

The `flaky_cells` view prints one row per cell and comparable fingerprint that both passed and failed, or `(no rows)`. The view reads the same runs as the `eval_trend` view. `verdict` is `flaky` or `insufficient`.

reeval sets `cause` to the first match in the order `infra`, `judge`, `check`, and `agent`. `infra` means the pooled cells include an infrastructure error. `judge` means a failing attempt has a failing judgment whose score is at most 0.1 from its threshold.

`check` means the same change both passed and failed its assertions. `agent` means none of the earlier causes matched.

For more information about when a cell is flaky, see [Flaky cells](/reference/statistics#flaky-cells). For more information about the columns, see [Query views](/reference/query-views#views).

</Step>
</Steps>

Repeat the run and the commit so each later run is in the trend.

## Next

For more information about diffing two runs, see [Compare with past runs](/results/compare/past-runs).