Repository navigation
Releases: rig-c/core
Release list
v2.64.0
2.64.0 (2026-10-19)
Features
- review: rigc review is the launcher — a bare target opens, never installs, reads review's Bun floor, refuses by name (#1492) (0a91020), closes #1486
Bug Fixes
- bin: read the CLI's exit and signal before spawnSync's error, so a refusal under Bun is not "could not launch bun" (#1496) (e302d72), closes #1488
- guide: write the guide whole through a loop that survives short writes, so Bun 1.2.0 stops corrupting a byte (#1495) (a3703dc), closes #1487
- selftest: one memory base row per platform and Bun line, so the declared minimum Bun is held to its own ceiling (#1497) (d5e6347), closes #1490
v2.63.0
v2.62.0
v2.61.1
v2.61.0
v2.60.0
Every file build writes is byte-identical to v2.59.0's
Three additions, none of which changes what build emits:
rigc review <dir> [...]hands a build to a person. It runs therig-reviewbin of@rig-c/reviewwith the arguments untouched and returns its exit code. When the package is not installed, it refuses by name:npm i -D @rig-c/review.@rig-c/reviewis not a dependency ofrig-c.rigc render --seam <a>:<b>[,…]is a reading, never a gate: it does not move the exit status. It writesseam.json(rigc-seam/1) beside each rendered set, reporting per frame where two slots that meet open a gap or show a ghost. Each entry gives frame index, time, px and a Spine-world box, plus spans and the worst frame. The contract is AUTHORING §8.8.- The
stopshalf-turn refusal (CE_STOP_LOCAL_BOUND_RANGE) now names the link, its root, the touch's world position, rest, the end, and the remedy. The code and the old text, now a prefix, are unchanged.
Features
- cli: rigc review delegates to the installed @rig-c/review bin (#1462) (70ecefa)
- render: render --seam — per-frame seam gap and ghost between two slots, a reading that never moves the exit status (#1464) (24b4714)
Bug Fixes
v2.59.0
v2.58.1
Every file build writes is byte-identical to v2.58.0's
Read a guide section before expanding it (#1447, #1448). rigc guide §5.1 now prints the compile-error introduction and a group index within 30,000 printed UTF-8 bytes. A parent section shows its own prose and child addresses with their full-span sizes; --all explicitly includes the children's bodies. Leaf sections still print their full text either way, and a CE_ code still resolves to its one documented row. --all with a rule name is a usage refusal, exit 2.
The release's dependant example is runnable. The documented comparison now uses the public scarf example and checks its config, animation file, painting, both layer directories and both layers.json files by name before packing or installation. Repository verification also makes GT05 report each unmet population bound in the refusal-recipe control's failure detail.
The byte comparison measured all 19 public-example and gallery recipes on Bun 1.4.2: 57 files, including skeleton.model.json, with every build command exiting 0. The tracked tools/emit_hashes.base.json is unchanged and its fresh check is CURRENT. The current @rig-c/parts 1.24.0 checkout type-checks against this release, and its public scarf build writes the same 106 files as on 2.58.0; that count comes from the two actual builds.
Editor round trip: target 0 this cut; no new round trip was run. The emitted files are byte-identical, and HQ authorized reuse of the previous format round trip for this follow-up.
Bug Fixes
v2.58.0
Every file build writes is byte-identical to v2.57.0's
Compile errors point to one guide row (#1399, #1445). A source compile refusal now carries a stable CE_ code separately from its original message, and its CLI line ends with (see rigc guide <CODE>). Run that command to read exactly the documented row for the refusal, with the shipped file, section, line and heading named. rigc guide §5.1.<k> reads a whole numbered group when more context is needed.
Library callers can read CompileError.code without parsing the message. Existing uncoded CompileError construction remains supported and points to §5.1.
The byte comparison measured all 19 public-example and gallery recipes on the same runtime: 57 files, including skeleton.model.json, with every build command exiting 0. The tracked tools/emit_hashes.base.json is unchanged and its fresh check is CURRENT. A copy of @rig-c/parts 1.24.0 type-checks against this release, and its public scarf example builds to the same 106 files as on 2.57.0.
Editor round trip: target 0 this cut; no new round trip was run. The emitted files are byte-identical, and HQ accepted reuse of the previous format round trip.
Features
v2.57.0
Every file build writes is byte-identical to v2.56.0's.
rigc help <command> (#1440, #1443) prints exactly what rigc <command> --help prints and exits the same way, on both entries. It is the spelling an agent reaches for first. rigc help on its own prints the top-level usage, and rigc help <unknown> is refused by naming the nearest command.
rigc split (#1414, #1438; follow-ups #1442, #1444) writes a rig spec and a motion spec with an authored mesh cut into its connected pieces. It re-slices every deform key and renumbers UVs, weights and triangles, and it never edits the author's files. The compile refusal for a mesh in pieces points to the command, with the names shell-quoted so the command copies and runs. A stated edge between two orphan vertices is dropped by name. The invariants.meshSlots line states how far the budget has to be raised (AUTHORING §3.4.11).
Editor round trip: no target this cut. Nothing this release changes reaches the emitted skeleton or atlas, as the byte-identity line above confirms; the 2.56.0 trip covers the format as it stands.
Features
- cli: rigc help <command> prints what <command> --help prints, the spelling an agent reaches for first (#1440) (#1443) (5ada1d2)
- cli: rigc split writes a rig spec and motion spec with an authored mesh cut into its connected pieces, deform keys re-sliced, so a mesh the compiler refuses as several pieces does not have to be split by hand (#1414) (#1438) (cef502f)