update-llama-cpp

name	update-llama-cpp
description	Update both llama-cpp conda recipes (source + binary) to the latest upstream release. Updates BOTH the active fork pin (beellama) and the commented-out mainline variant pin in each recipe.yaml. Use when the user wants to bump llama-cpp to a newer version.
compatibility	Uses `scripts/llama-cpp-changelog.py` for changelog/merge detection. No `gh` CLI or GitHub token needed — the script works from a local commits-only git clone (cached per-fork at `~/.cache/llama-cpp-changelog/<repo>.git`). The PR section is skipped without GitHub auth; tags/commits/dates come from git.
allowed-tools	Bash Read Edit

Context

Both pixi-recipes/llama-cpp-source/recipe.yaml and pixi-recipes/llama-cpp-binary/recipe.yaml pin two forks in their context: blocks:

• Active beellama fork — fork: + version: (stable vX.Y.Z OR preview preview-vX.Y.Z, depending on which recipe and which version you are targeting)
• Mainline variant — # fork: ggml-org/llama.cpp, # version: bNNNN
• Other experimental forks (source only)

CRITICAL GUARDRAILS:

⚠ Never change fork:. The user's fork (e.g. crusaderky/llama.cpp) is a deliberate choice and must be preserved. Only update version: strings. Upgrading fork: is a bug — it discards custom fork state and was a failure in the last update everything run.

⚠ Never conflate stable and preview tags. Source builds require stable git tags (vX.Y.Z). Binary builds may use preview tags (preview-vX.Y.Z) since pre-built binaries are published from preview releases. Fetching "the latest beellama stable tag" must return a vX.Y.Z tag, not a preview-vX.Y.Z. Using a preview tag in a source recipe is incorrect — source builds clone at the tag, and preview tags are pre-release.

⚠ Always verify the fetched tag exists on the fork before applying it. A tag name from a release page title or preview.json may not be a valid git tag. Query the API/refs to confirm the tag exists and resolves.

The uncommented block may differ between binary and source recipes. The update bumps the Anbeeld/beellama.cpp pin and ggml-org/llama.cpp (four version strings total across two recipes), so the commented variants stay current. It does NOT touch other forks.

The source recipe's source: block uses ${{ fork }} and ${{ version }} — there is NO separate source.rev / commit SHA field. The binary recipe additionally pins asset_prefix (beellama vs llama); it never changes on a version bump.

Steps

Phase 1 — Fetch latest versions (verify existence!)

Do NOT use the GitHub REST API for tag discovery (rate-limited without auth). Use git ls-remote instead, and verify each tag exists by checking the repo's tags endpoint or the fetched commit data.

1. Latest beellama stable tag (ignore preview-*):

   git ls-remote --tags --refs https://github.com/Anbeeld/beellama.cpp.git \
     | grep -oE 'refs/tags/v[0-9]+\.[0-9]+\.[0-9]+$' | sed 's|refs/tags/||' | sort -V | tail -1
   

   → LATEST_BEELLAMA (this is a stable vX.Y.Z tag, NOT preview-vX.Y.Z)

   Verification: Confirm LATEST_BEELLAMA exists as a real tag by checking https://api.github.com/repos/Anbeeld/beellama.cpp/tags?per_page=100 or git ls-remote --tags https://github.com/Anbeeld/beellama.cpp.git | grep refs/tags/$LATEST_BEELLAMA.

2. Latest mainline llama.cpp tag:

   git ls-remote --tags --refs https://github.com/ggml-org/llama.cpp.git \
     | grep -oE 'refs/tags/b[0-9]+$' | sed 's|refs/tags/||' | sort -V | tail -1
   

   → LATEST_MAINLINE

   Verification: Same existence check as above for ggml-org/llama.cpp.

Phase 2 — Update the active (beellama) pins

3. Read pixi-recipes/llama-cpp-source/recipe.yaml; note the uncommented version: under fork: (whatever fork is currently active in the source recipe). Do NOT touch fork:. If it already equals LATEST_BEELLAMA, report "Source beellama version already up to date"; else update only version: to LATEST_BEELLAMA. Report Source beellama: vOLD → vNEW.

   If the active fork is crusaderky/llama.cpp (your fork): preserve it. Only update version:. Do not replace fork: with Anbeeld/beellama.cpp.

4. Same for the uncommented version: in pixi-recipes/llama-cpp-binary/recipe.yaml. Report Binary beellama: vOLD → vNEW.

   Stable vs preview distinction for binary recipe:

   • If the binary recipe's active version: is a stable vX.Y.Z, compare against LATEST_BEELLAMA. If equal, no change.
   • If the binary recipe's active version: is a preview preview-vX.Y.Z, keep it as-is UNLESS the latest stable vX.Y.Z is strictly greater than the preview's base version — then migrate to the stable tag. Otherwise, keep the preview (binary builds often use preview pre-built assets).

   Never use preview-vX.Y.Z as a replacement for a stable vX.Y.Z in a context where only stable tags are valid (e.g. source recipe's git: tag field). Preview tags are pre-release and not suitable for source cloning.

5. Handle git-hash source versions (source recipe only): If the source recipe's version: is a raw commit SHA with a # vX.Y.Z comment (e.g. 4a834dd7... # v0.4.2 preview), then: a. Is there a stable version newer than vX.Y.Z? If so, replace the hash and version with LATEST_BEELLAMA. b. Otherwise, search the same branch for a more recent commit SHA that still corresponds to vX.Y.Z. Update the hash only.

Phase 3 — Update the commented (mainline) variant pins

6. In each recipe, update the commented # version: bNNNN line under # fork: ggml-org/llama.cpp to LATEST_MAINLINE (keep the # prefix). If already current, say so. Report Source mainline variant: bOLD → bNEW and Binary mainline variant: bOLD → bNEW.

Phase 4 — Verify & report

7. Show both changelogs for the ranges being updated:

   pixi r llama-cpp-changelog --repo Anbeeld/beellama.cpp <old_v> <LATEST_BEELLAMA>
   pixi r llama-cpp-changelog --repo ggml-org/llama.cpp <old_b> <LATEST_MAINLINE>
   

   Present the script output (or a themed summary, citing PR numbers + URLs). Do not hand-roll curl/compare calls — the script is the canonical dumper. Skip a changelog when that fork did not change.

8. Final verification checklist (run before reporting completion):

   • fork: unchanged in all recipe context blocks (user's fork preserved)
   • No preview-v tag was inserted where only stable v tags are valid
   • All fetched tags verified as existing on their respective repos
   • pixi lock run and pixi r lint clean

9. Report all changes (source/binary × beellama/mainline). Include a line like: "llama.cpp: no effective version change — latest stable is v0.4.2, binary already at preview-v0.4.3; fork pins untouched." when no versions actually moved.