---
title: Analyze a run
description: Explain a run's results with an analysis model.
---
> ## 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.

Set an analysis model when you want a language model to explain a run. Insights print at the end of `reeval run`. `reeval analyze` writes a fuller analysis of one run.

For more information about what insights and Remarks are, see [Analysis](/reference/glossary#analysis).

## Set an analysis model

In the `reeval.config.ts` file, set `analysis.model` to a language model. `reeval init` writes this value. reeval refuses an evaluation model, which only scores yes-or-no questions.

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

export default defineConfig({
  harness: claudeCode({ model: 'claude-sonnet-5' }),
  analysis: { model: 'google/gemini-3.8-flash' },
})
```

In the `.env.local` file, add the key for the analysis model:

```bash title=".env.local"
GOOGLE_GENERATIVE_AI_API_KEY=replace-with-your-key
```

For more information about other model prefixes, see [`AI_GATEWAY_API_KEY`](/reference/environment-variables#aigatewayapikey).

When the key is missing, the run keeps its outcome and prints this line:

```text
insights skipped: set GOOGLE_GENERATIVE_AI_API_KEY, or analysis: { insights: false }
```

For more information about the other `analysis` fields, see [the `analysis` option](/reference/options#analysis).

## Analyze a run

`reeval analyze` makes one language model call over a digest of the most recent run's facts. When a recorded run on an ancestor commit exists, the digest includes that run.

1. In the `reeval.config.ts` file, set `analysis.instructions` to the question you want answered:

   ```ts title="reeval.config.ts"
   analysis: {
     model: 'google/gemini-3.8-flash',
     instructions: 'Fewer tool calls count as a win.',
   },
   ```

2. Run the analysis:

   ```bash
   npx reeval analyze
   ```

The analysis starts with `analysis` and the run id, then `vs` and the ancestor's run id when the digest includes one. `TL;DR`, `Suggested actions`, and `Confidence` always print. `What happened`, `Results`, `Why`, and `Anomalies` print when the model has something for them. reeval caches each analysis in the `.reeval/analysis` directory, so the same runs, model, and `analysis.instructions` print the cached analysis. To analyze another run, pass its id, and to choose the run it compares with, pass `--vs`.

## Read insights

Without `reeval analyze`, `reeval run` prints insights after the attempts finish: a sentence under a comparison row or a failure, and a headline above the tally. reeval fixes the outcome and exit code first. This is the block the scripted model in [`with-analysis`](https://github.com/invertase/reeval/tree/main/examples/with-analysis) printed:

```text
insights: 1 eval with fake (Ctrl-C skips the rest; --no-insights skips them all)

pass             adds a tested multiply function
  Yes: the agent added it and the tests pass.

Every eval passed.
```

To skip insights on one run, pass `--no-insights`. To skip them on every run, set `analysis.insights` to `false`. For more information about the `Remarks:` that can follow the tally, see [Read a run's output](/results/output).

## Reference

For more information about every flag, see [`reeval analyze`](/reference/cli#reeval-analyze).