Captured from a production run, re-verified 8 Aug 2026 — payloads verbatim,
long bodies excerpted.

## 1 · The research question

A scheduled job wakes nightly and asks: *"What changed in llama.cpp since
the tag we digested last night?"* — then posts a summary the team actually
reads.

## 2 · The ordinary tool — refused before it starts

The job must *construct* `compare/<last>...<tonight>` from two tags. Native
fetch tools only accept URLs that appeared verbatim in the conversation:

```
web_fetch("https://api.github.com/repos/ggml-org/llama.cpp/compare/b10256...b10308")

✗ URL not present in conversation — refused
✗ retried from a search hit: empty body, no JSON
```

## 3 · The trace, call one — tonight's tag, three fields

The release object carries 30+ fields; the job needs three. `select` trims
server-side:

```json
fetch("https://api.github.com/repos/ggml-org/llama.cpp/releases/latest",
      select: ["tag_name", "name", "published_at"])

{ "tag_name": "b10308", "name": "b10308",
  "published_at": "2026-08-07T14:05:22Z" }
```

## 4 · The trace, call two — the diff, without the megabyte

`compare/b10256...b10308` answers with 52 commits plus 183 per-file
patches — **1.26 MB, roughly 315k tokens**
([full response](https://api.github.com/repos/ggml-org/llama.cpp/compare/b10256...b10308)).
The jq-lite path reaches into every array element, shape intact:

```json
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_data..." } }
  ] }
```

What arrives is **28 KB, roughly 7k tokens** — 45× smaller.

## 5 · The digest the job posts

Generated from the 52 captured commit messages — every line traces to a PR
number in the payload above:

> **llama.cpp nightly — b10256 → b10308 · 52 commits**
>
> - **Breaking:** Qwen3-TTS support lands in mtmd; the `llama-tts` binary
>   changes (#26254)
> - **Models & backends:** Vulkan gains GATED_LINEAR_ATTN (#25601); SYCL
>   adds DSv4 ops (#26568) and full set_rows type coverage (#26515); AMD
>   ROCm gfx1151 joins CI (#26544)
> - **Hardening:** gguf-py validates n_dims and guards uint64 overflow on
>   crafted files (#25401); server file-search walk hardened (#26626);
>   UI dependency bumps clear vulnerable packages (#26607)
> - **Server/UI:** spec-decode counters on `/metrics` (#26389);
>   per-conversation working directory in the agent UI (#26518)
>
> Sources: `releases/latest` + `compare/b10256...b10308`, read via
> `select` — 28 KB of a 1.26 MB response reached the agent.

## 6 · Measured cost, and where this came from

- **Per run:** two selected reads ≈ **28 KB / ~7k tokens** into context,
  against **~1.29 MB / ~320k tokens** without `select` — and the compare
  call wouldn't have happened at all on a native fetch tool.
- **Provenance:** captured from a production run; re-verified 8 Aug 2026
  from Claude Code over MCP — `schema: true` reports `commits[] $items: 52`,
  `files[] $items: 183`, and the select rerun returned the identical
  `ahead_by: 52` payload. Bodies cached server-side for 24 h.

## 7 · Run it yourself

Both endpoints are public — [install Search](/search#install) and issue the
two `fetch` calls above, or start from the
[oversized source](/search/examples/oversized-source) example this digest
builds on.