Captured payloads, re-verified 8 Aug 2026.

## 1 · The research question

*"Summarise what changed between these two releases."* The GitHub compare
endpoint has the answer — wrapped in 52 commits with author objects, trees
and verification blocks, then 183 per-file patches. The source itself is
the problem.

## 2 · One request, two costs

Fetched whole: **1.26 MB, roughly 315k tokens**
([full response](https://api.github.com/repos/ggml-org/llama.cpp/compare/b10256...b10308)).
With `select`, the trim happens server-side and the nested shape is
preserved, so rows stay aligned:

```
fetch(".../compare/b10256...b10308",
      select: ["ahead_by", "total_commits",
               "commits[].sha", "commits[].commit.message"])

{ "ahead_by": 52, "total_commits": 52,
  "commits": [
    { "sha": "22dc605c4ead...",
      "commit": { "message": "ci: fix vulkan llvmpipe runs (#26533)" } },
    { "sha": "935cad6497e8...",
      "commit": { "message": "llama : move n_vocab from llama_sampler..." } }
  ] }
```

## 3 · Measured improvement, and where this came from

**28 KB (~7k tokens)** reached the agent instead of **1.26 MB (~315k)** —
45× smaller, every commit message intact. The source stays cached for 24
hours, so follow-ups are free. The finished summary built from exactly this
payload is [the nightly release digest](/search/examples/release-digest.md).

- **Client:** Claude Code over MCP
- **Re-verified 8 Aug 2026:** `schema: true` → `commits[] $items: 52`,
  `files[] $items: 183`; select rerun → identical `ahead_by: 52` payload
- **Source:** the full 1.26 MB response is linked above — check the trim
  yourself
- **Retention:** body cached server-side, 24 h

## Run it yourself

Both endpoints are public — [install Search](/search#install) and issue the
call above.