## TL;DR
At 100k rows there is no prompt-based alternative: aggregate in code, stream the file row by row, and keep the model on the summaries. Extract filtered slices only for targeted drill-downs.

```text
agent hit context window limit on 100k-row usage export, run failed
```

## Use this when
- An agent fails with context overflow on a ~100k-row export.
- The file is tens of megabytes.
- Any attempt to read it into the prompt fails.

## Not for this skill when
- The file fits in context. Just read it.
- The overflow is from something else. Find the real large input.
- Only a few rows are needed. Filter first, then read.

## Steps
1. Write the aggregation as a standalone script. Verify: it runs without the model.
2. Stream the export row by row, never loading it whole. Verify: memory stays flat.
3. Emit per-app and per-status summaries. Verify: the summaries answer the audit question.
4. Bring only the summaries into context. Verify: the prompt stays small.
5. Confirm the run completes. Verify: no overflow.

## Variant phrasings
### context window exceeded
The error-first phrasing.
### 100k rows
The number-first search.
### export too large for prompt
The plain-English phrasing.

Compatibility: Any agent runtime with code execution and a context window.

## Why it happens
100k rows are tens of millions of tokens. Code is the only path; the model directs the analysis but never holds the data.

## Edge cases / pitfalls
- Needing row-level detail for a few outliers: extract by filter, not by loading all.
- The script needing the schema first; sample rows to learn it.
- Summaries that answer the wrong question; define the needed aggregates before coding.
- The agent falling back to reading the file when the script has a bug; fix the script.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_Cf0pxQePk2CuoPY3a2n7IA