Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 5 additions & 2 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,9 @@ All notable changes to `@developerz.ai/ts23d`. Format: [Keep a Changelog](https:

## [Unreleased]

Several agents can share one machine through a resource queue, CI is balanced by measured cost, and the check that committed renders are fresh now works.
## [0.14.0] - 2026-11-19

Several agents can share one machine through a resource queue, the npm package shrinks from 40 MB to 4 MB, CI runs in under five minutes, and the check that committed renders are fresh now works.

### Added
- Machine-wide resource queue (ADR 017): agents sharing a machine take turns at `cpu` tokens, `blender` and `gpu` through ticket files under `$TS23D_HOME/queue/`, first come, first served, no daemon. A dead or killed command frees its place; a queue that cannot be written warns once and the command runs unqueued.
Expand Down Expand Up @@ -489,7 +491,8 @@ The first published version: the asset DSL with primitives, organic sculpting, t
- Animations: `toGlb()` exports none unless `{ animations: true }` (glTF has no autoplay flag; engines drive named pivots from code). `t3d build` bakes the `run` clip by default so viewers show play controls; `--no-animations` skips it.
- CLI output is one folder per asset: `out/<asset>/{lod0,lod1,lod2}.glb`, `sheet.png`, `manifest.json`, plus `out/index.json`. `t3d build` renders the sheet too (`--no-shot` to skip); `t3d shot` writes sheet + manifest only.

[Unreleased]: https://github.com/developerz-ai/ts23d/compare/v0.13.0...HEAD
[Unreleased]: https://github.com/developerz-ai/ts23d/compare/v0.14.0...HEAD
[0.14.0]: https://github.com/developerz-ai/ts23d/compare/v0.13.0...v0.14.0
[0.13.0]: https://github.com/developerz-ai/ts23d/compare/v0.12.0...v0.13.0
[0.12.0]: https://github.com/developerz-ai/ts23d/compare/v0.11.0...v0.12.0
[0.11.0]: https://github.com/developerz-ai/ts23d/compare/v0.10.0...v0.11.0
Expand Down
2 changes: 1 addition & 1 deletion docs/changelog/0.13.0.md
Original file line number Diff line number Diff line change
Expand Up @@ -108,4 +108,4 @@ A correctness sweep across the library, CLI, gates, gallery and recipes: over a

---

Older: [0.12.0](0.12.0.md)
Older: [0.12.0](0.12.0.md) · Newer: [0.14.0](0.14.0.md)
35 changes: 35 additions & 0 deletions docs/changelog/0.14.0.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,35 @@
<!-- Generated from CHANGELOG.md by `bun run changelog`. Edit CHANGELOG.md, never this file. -->

# ts23d 0.14.0

Released 2026-11-19 · [npm](https://www.npmjs.com/package/@developerz.ai/ts23d/v/0.14.0) · [diff v0.13.0...v0.14.0](https://github.com/developerz-ai/ts23d/compare/v0.13.0...v0.14.0) · [all releases](README.md) · [CHANGELOG](../../CHANGELOG.md)

Several agents can share one machine through a resource queue, the npm package shrinks from 40 MB to 4 MB, CI runs in under five minutes, and the check that committed renders are fresh now works.

### Added
- Machine-wide resource queue (ADR 017): agents sharing a machine take turns at `cpu` tokens, `blender` and `gpu` through ticket files under `$TS23D_HOME/queue/`, first come, first served, no daemon. A dead or killed command frees its place; a queue that cannot be written warns once and the command runs unqueued.
- `t3d queue status` (`--json`), `t3d queue clear --stale | --pid <n> | --all`, `t3d queue run [--cpu <n|all>] [--blender] [--gpu] -- <command…>`, `t3d queue hold -- <command…>`.
- `--no-wait` and `--max-wait <s>` on every command; `T3D_QUEUE=wait|fail|off`, `T3D_QUEUE_MAX_WAIT`; machine settings in `$TS23D_HOME/queue.json`. Off when `CI` is set.
- Error codes `T3D_QUEUE_BUSY` and `T3D_QUEUE_TIMEOUT`, each naming the holder and the fix.
- `t3d doctor` has a `queue` group; `onToolChild` in `/neural` reports each tool as it starts.
- `bun run scripts/recipe-costs.ts [--tests] [--plan]` re-measures the per-recipe and per-test costs CI shards by; `bun run recipes <name…>` checks only the named folders.

### Changed
- `build`, `shot`, `hero`, `review`, `critique`, `blueprint`, `icon`, `gaps`, `tune`, `finish`, `concept` and `neural run` take one CPU token; `hero --quality path` takes 8 and traces on the threads granted (same bytes). `blender review` and `blender session start|run|finish` take `blender`; Blender adapter jobs and jobs sent to a loopback model URL take `blender` / `gpu` while they run. A waiting command prints one `t3d queue: waiting …` line on stderr.
- `bun run verify` holds one place in the queue for the whole gate; tests run with the queue off.
- CI: `gate` and `shots` are one 16-shard `gate` job (26 jobs → 18) balanced by measured cost (`scripts/lib/recipe-costs.json`, `scripts/lib/test-timings.json`); `bun test` is split across the shards; docs-only changes skip the recipe gate and the renders. `verify` stays the single required check.
- `bun run shots` renders each folder's sheet, hero, blueprint and icon in one process and from one compile per asset and LOD (`human_head` renders 242 s → 88 s locally), byte-identical to the four separate commands. The shared compile is off everywhere else, so `t3d mcp` never serves an old compile.
- `bear_hybrid`, `wolf_hybrid` and `dragon_hybrid` are model-first with no CPU build: `t3d build` with a model output missing stops with one `T3D_NEURAL_MISSING` instead of building a code creature that fails QA. `modelChain` takes `required: true` and `ctx.neural` asks take `{ required: true }`; `ranger` and `human_head` keep their CPU fallback.
- CI: 18 gate shards with costs calibrated from a real run: `bun run scripts/recipe-costs.ts --from-run <run id>` reads a green `ci` run's logs and rewrites the cost files in GitHub seconds; `--plan` marks the shard count to pick. Measured 5.5 min and 67 runner-minutes before rebalancing (jobs 155–299 s); measured 4.6 min push → green after (slowest job 269 s), under the 5-minute target for the first time since it was set.

### Fixed
- The npm package is 4.2 MB instead of 40.6 MB: model recipes' generated outputs (`recipes/*/neural/`) stay in git. 0.13.0 shipped them without the concept images they need, so they never built from the package. The `pack` gate step now fails when the tarball goes over 5.3 MB packed or 16 MB unpacked, holds a file over 1 MB, or holds anything under `recipes/*/neural/`.
- `T3D_NEURAL_MISSING` for a recipe inside the installed package says the package ships no model outputs and to clone the repo or copy the recipe folder; a `--require-neural` miss no longer says "using the CPU path".
- Model creatures: a rigger's terminal stub bone on a leg (UniRig, 6 mm past the toe) no longer shifts hip, stifle, hock and paw one joint down and gets the rig refused (`T3D_MODEL_SKELETON`).
- Critiques: every committed asset has a well-formed critique bound to its current renders, re-judged from the images, with a fix note under every score below 10. Malformed critiques had been hiding their scores from the gate.
- CI's "committed renders match a fresh render" check never failed: its pathspec `recipes/*/out` matches no file under `out/`. It now uses `:(glob)recipes/*/out/**`, and a test proves the pathspec lists the committed outputs (D286).
- `rts_faction`: committed renders refreshed; they still showed the tank's old belt weathering, which the broken check let through.

---

Older: [0.13.0](0.13.0.md)
1 change: 1 addition & 0 deletions docs/changelog/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@ Every published version of `@developerz.ai/ts23d`, newest first: one page each w

| Version | Date | Summary |
|---|---|---|
| [0.14.0](0.14.0.md) | 2026-11-19 | Several agents can share one machine through a resource queue, the npm package shrinks from 40 MB to 4 MB, CI runs in under five minutes, and the check that committed renders are fresh now works. |
| [0.13.0](0.13.0.md) | 2026-11-19 | A correctness sweep across the library, CLI, gates, gallery and recipes: over a hundred and fifty defects fixed, stricter input validation and QA gates, track belts that move in any glTF reader, and PNG bytes that no longer depend on the runtime. |
| [0.12.0](0.12.0.md) | 2026-11-18 | Swimming and many-limbed creatures with motion from published biomechanics, one open Blender session for chained work, and brushes for existing meshes. |
| [0.11.0](0.11.0.md) | 2026-11-17 | Two more model-first creatures, a brown bear and a winged drake, and the library options and QA rules they needed. |
Expand Down
2 changes: 1 addition & 1 deletion package.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "@developerz.ai/ts23d",
"version": "0.13.0",
"version": "0.14.0",
"description": "TypeScript to 3D (ts23d): free, deterministic 3D asset generation in TypeScript — a DSL your coding agent writes, compiled to game-ready glTF. No GPU or credits needed; optional adapters add local GPU models, Blender and hosted services.",
"type": "module",
"license": "MIT",
Expand Down
2 changes: 1 addition & 1 deletion src/export/version.ts
Original file line number Diff line number Diff line change
Expand Up @@ -2,5 +2,5 @@
// constant because the library may not import package.json (rootDir is src); version.test.ts
// fails the gate when the two disagree, and `bun run release` stamps both.

export const VERSION = '0.13.0';
export const VERSION = '0.14.0';
export const GENERATOR = `@developerz.ai/ts23d ${VERSION}`;