## TL;DR
A campaign summary that loads every message into the context is a design with a built-in ceiling: the campaign only has to outgrow the window once. Summarise hierarchically instead: per-message structured verdicts stored in a table, per-cluster summaries over those rows, and a final summary over cluster summaries and counts computed by code. The model should read aggregates and samples, never the raw campaign.

```text
context window overflow crashed the agent's campaign-summary step; the batch failed
```

## Use this when
- The summary step fails on large campaigns but works on small ones.
- The error names context length or the run dies at the final step after successful triage.

## Not for this skill when
- The failure is a model outage error, not a length error. Check the error text before redesigning.
- The campaign is small and the summary includes huge raw headers per message. Trimming inputs may be enough.

## Steps
1. Measure the token size of the current summary input on a failing campaign. Verify: the overflow is quantified, not assumed.
2. Persist per-message verdicts as structured rows during triage, with counts and groupings computed by code. Verify: campaign statistics exist without the model.
3. Build cluster summaries from the rows for each sender, kit or URL group, each summarised separately. Verify: no single summary call approaches the context limit.
4. Compose the final report from cluster summaries plus code-computed counts, with a few exemplar messages quoted. Verify: the final call input is a small fraction of the window.
5. Run the summary on the previously failing campaign. Verify: it completes and its counts match the triage table exactly.

## Variant phrasings
### context window overflow campaign summary agent batch
The overflow variant.
### hierarchical summarisation large campaign triage report
The hierarchical variant.

Compatibility: LLM agent frameworks with structured storage and staged summarisation.

## Why it happens
Context windows grow, but campaigns grow faster, and raw evidence is the bulkiest possible input. The fix is architectural: evidence lives in storage, the model works on reductions, and arithmetic is done by code that does not hallucinate totals.

## Edge cases / pitfalls
- Never ask the model to count. Counts from code, narrative from the model.
- Exemplar selection matters: choose median and extreme examples deliberately, not the first rows.
- Keep cluster boundaries stable across runs so summaries are comparable over a campaign lifetime.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_w5ZZVqhbtymKNsf1yiyTvw