update-packages

name	update-packages
description	Run nix-update across all packages in this repo, fix build failures, and land each package's changes as its own separate jj commit. Use when asked to "update all packages", "bump everything", "run nix-update on all packages", or similar bulk-update requests.

Bulk-update every package in default.nix via nix-update, fix whatever breaks, verify builds, and end with one clean jj commit per changed package.

Rules (non-negotiable)

• One package change per commit. Never let fixes for package B leak into package A's commit. Use jj squash --from <rev> --into <rev> -- <path> to move stray file changes into the right commit, then jj rebase to keep the stack linear.
• Never regress a version. For every package whose version/rev changes, explicitly verify the new value is newer/forward from the old one before committing — don't just trust nix-update's pick. This is mandatory, not just for cases where you're unsure:

  git clone -q <homepage-url> /tmp/<pkg>-check
  cd /tmp/<pkg>-check
  # if both sides are tags/versions, compare dates:
  git log -1 --format=%ci <old-ref>
  git log -1 --format=%ci <new-ref>
  # if either side is a raw commit hash (not a tag), confirm ancestry:
  git merge-base --is-ancestor <old-rev> <new-rev> \
    && echo "old is ancestor of new (safe upgrade)" \
    || echo "STOP: old is NOT an ancestor of new — do not commit this"
  

  If merge-base --is-ancestor fails (exit non-zero) or the old ref's commit date is later than the new ref's, treat this as a bug in the proposed update — do not commit it. Re-check upstream tags/releases manually and either pick the actually-latest one or leave the package untouched.
• Every touched package must build before it's considered done (nix-build -A <pkg>, or nix build .#<pkg> if the local <nixpkgs> channel is too old — see Known Issues below).
• Attribute renames with no version bump (e.g. cargoSha256 -> cargoHash) still get their own commit — don't bundle with unrelated version bumps.

Steps

1. List packages for the current system:

   nix flake show --json 2>/dev/null | \
     jq -r '.inventory.packages.output.children["aarch64-darwin"].children | keys[]'
   

   (swap aarch64-darwin for the relevant system)

2. Run nix-update per package, non-interactively, logging output:

   <list-of-packages> | xargs -n1 nix-update --commit --flake 2>&1 | tee /tmp/nix-update-log.txt
   

   This auto-commits via jj/git as it goes — expect it to create messy working-copy commits you'll clean up in step 4. Some packages will error out (unstable/prerelease versions, hash mismatches from tag renames, flake-eval quirks) — note them, don't stop the batch.

3. Triage failures from the log:

   • "Found an unstable version ... ignored" → rerun that package alone with nix-update --commit --flake --version=unstable <pkg>.
   • Hash mismatch after a rev change → check if upstream tags have a v prefix (rev = version; vs rev = "v${version}";), fix, then re-run nix-update or manually recompute via lib.fakeSha256 swap + nix-build.
   • Go vendorHash / Rust cargoHash mismatches after any src change → always need to be recomputed; let the build error tell you the correct hash and paste it in.
   • Rust packages with a vendored Cargo.lock that's gitignored upstream (check pkgs/<pkg>/Cargo.lock existing in this repo) → regenerate with cargo generate-lockfile against a fresh clone of the new tag, copy it over pkgs/<pkg>/Cargo.lock.
   • Git dependencies in Cargo.lock (source = "git+https://...") need explicit hashes in cargoLock.outputHashes — get them via nix run nixpkgs#nix-prefetch-git -- --url <url> --rev <rev> --quiet.

4. Verify each version/rev bump is actually forward, before doing any jj cleanup. For every package where version or rev changed, run the ancestry/date check from the "Never regress a version" rule above. Do this for every single change — including ones picked automatically — not just the ones that look suspicious. If a package fails the check, revert its change ( or drop that package's commit) and investigate upstream manually before retrying.

Commit Message Format

Follow the repo convention from AGENTS.md: <package-name>: <description>, one line, no body needed for routine bumps.

• Version bump: <pkg>: <old-version> -> <new-version> e.g. esa: 0.2.0 -> 0.3.0, prr: 0.20.0 -> 0.21.0
• Commit-pinned package moving to a tagged release: use the short hash for the old side, plain version for the new side — <pkg>: <old-short-hash> -> <new-version> e.g. protodot: 87817c3 -> 1.2.0 (only do this after confirming the old commit is actually an ancestor of the new tag — see the "Never regress a version" rule above)
• Hash/attribute-only fixups with no version change: <pkg>: <old-attr> -> <new-attr> or a short imperative description — e.g. toffee: cargoSha256 -> cargoHash, gloc: recompute
• Non-version code fixes (e.g. migrating away from a removed nixpkgs attribute): a short imperative description, still scoped to one package — e.g. prr: migrate darwin frameworks to apple-sdk
• New package added to the repo (not a bulk-update scenario, but keep the convention consistent if one comes up alongside a batch): <pkg>: init at <version> — e.g. defuddle: init at 0.9.0, esa: init at v0.2.0. For packages pinned to a commit instead of a tag, use nixpkgs' unstable- date format: hunk: init at unstable-2026-05-08.
• Never combine multiple packages' descriptions in one message (no probe/prr/toffee: ...) — if you catch yourself writing a message with a slash or "and", that's a signal the commit needs splitting per the one-package-per-commit rule.
• Don't add prefixes like fix:/chore:/feat: — this repo doesn't use conventional commits, just <package-name>: <description>.

Known Issues

• darwin.apple_sdk_11_0 has been removed when building via nix build (flake-pinned nixpkgs) for packages still using darwin.apple_sdk.frameworks.* in buildInputs. Fix: replace the darwin callPackage arg with apple-sdk, and swap darwin.apple_sdk.frameworks.X for just apple-sdk in the darwin lib.optionals block. Works on both the old <nixpkgs> channel and the flake's pinned nixpkgs. Any cargoHash/vendorHash will need recomputation after this since it changes the dependency closure.
• Local <nixpkgs> channel has an older rustc than the flake's pinned nixpkgs (observed: channel rustc 1.82 vs flake nixpkgs-unstable rustc 1.96+). If a Rust package's MSRV exceeds the channel's rustc, nix-build will fail with rustc X.Y.Z is not supported by ... — use nix build .#<pkg> instead, which resolves against the flake's own nixpkgs input.
• hunk is provided via a separate flake input (see flake.nix), not a pkgs/hunk directory — skip it in bulk nix-update runs, it'll error with a sanitizePositions/package.nix is not in ... eval error which is expected and not fixable from this repo.