## TL;DR
Summarise ablations from a results table built by code: extract each ablation arm's final metrics into rows first, then write the narrative from the table, which is small by construction. Ablations multiply runs by design, so any approach that reads run histories directly scales with the arm count and overflows exactly when the study is thorough.

```text
context overflow error when agent tried to include all ablation results and the summary failed
```

## Use this when
- The run or call fails with this symptom: context overflow error when an agent tried to include all ablation results in a summary.
- You are working on an agent summarising an ablation study across many arms and seeds and the failure repeats on the same run or configuration.
- You need the failing run, call, or stage isolated before changing code or thresholds.

## Not for this skill when
- One arm needing a deep dive into its training behaviour, where the fix is reading that arm alone after the table-level summary.
- The error text names a different tool, stage, or root cause. Match the verbatim error first and use that skill.

## Steps
1. Build the ablation table with code: one row per arm and seed with the final metrics and status. Verify: the whole study becomes one compact table
2. Compute the comparisons the summary needs, deltas against the full model per arm, in code. Verify: narrative inputs are precomputed numbers
3. Write the summary from the table and deltas, citing arm names and values, without opening run histories. Verify: the summary is produced within budget
4. Note any arm whose runs failed or are missing in the table so the narrative accounts for them. Verify: gaps in the study are stated, not smoothed over

## Variant phrasings
### agent context overflow all ablation results summary failed
The verbatim phrasing.
### ablation summary agent too many arms table first fix
The table-first phrasing.
### agent summarize ablation study overflow extract metrics code
The extraction phrasing.

Compatibility: Agent summarisation of ablation study results under context limits

## Why it happens
Ablation studies are combinatorial, so their raw record volume is the point of the design and the enemy of wholesale reading. The table projection is lossless for the questions ablations answer, which are comparisons of final numbers across arms.

## Edge cases / pitfalls
- Interaction effects between ablated components need specific arm pairs compared; precompute those pairs rather than narrating around them.
- Keep the table as the study's durable output; future questions should start from it, not from histories.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_qlu0h4TOOF2gbHTCkCrGvA