RFD 1057 details: the open-work list
Delete an entry when it closes. This file shrinks, and it never grows a history section.
Verified, and running
| What | Evidence |
|---|---|
| The planner composes three documents | pipeline_test.exs, all 105 pass on this box now that taskweft_nif is rebuilt for x86_64 |
| The model image serves HTTP | RFD 1040, run in Docker on this machine |
| CockroachDB provisions and runs | RFD 1020, 92 tests with the node up |
| taskweft composition | PRs 207, 208, 209, merged upstream |
| Both Quadlets run end to end | RFD 1058, weftspun.service and weftspun-crdb.service both active (running), /api/v1/health and /api/v1/models answered over weftspun.network |
taskweft_nif runs on x86_64 |
make clean && mix deps.compile taskweft_nif --force from deps/taskweft_nif/. mix test then runs 105/105 with no arch-mismatch failure |
Written, and never run
The dev container. RFD 1056. The image does not build yet. The Debian attempt failed at mix local.hex, and the Fedora rewrite answers that by reading the error. Build it before trusting it.
The Pixal3D worker stage. RFD 1040. Only the contract stage ran. The worker stage pulls 24.045 GB and needs an NVIDIA device, thus it needs a rented card.
The local worker. RFD 1055 Phase 2. No worker service runs on this box yet, only a manual contract-stage test (RFD 1040). No instance is rented either, and RFD 1062’s Gall’s law says none should be, until this box is not enough.
The Fly.io / 4090 split. RFD 1062. No fly.toml, no worker-side job-receiving adapter, no Tailscale join between a Fly machine and this box, no CockroachDB migration off this box. RFD 1062 names the target. None of it runs yet. The adapter’s asset-transport half is settled, idtx_transport/aria-storage, per RFD 1062’s own DETAILS.md. Open work narrows to the job-control envelope, and whether an aria-storage instance runs anywhere reachable yet.
Measured, and not built
Fourteen model folders. RFD 1036. Each still carries a cog.yaml and a predict.py, and RFD 1036 no longer selects Cog. Convert one when its model is next worked on, and not in a sweep. The folder names are already converted, and the files inside are not.
_to_usd in the worker. RFD 1053. The layer records the GLB as an asset attribute, because usd-core alone reads no glTF. A glTF file format plugin would let it be a reference arc.
The idtx_core NIF adapter. RFD 1061. flow/adapters/ in thirdparty/fabric-flow-adapters/ holds three hosts — Godot, Unity, CLI — and no Elixir one. Needs a fourth adapter, a weftspun_studio route, and the browser call site swapped over. RFD 1061’s stopgap (prepareGlbForApiUpload in glbCompress.js) stays until this lands.
Unknown, and blocking a number
The Q4_K_M quality cost. RFD 1043. Quantization is a price choice now, and no measurement compares the two formats.
The browser client’s own test suite. RFD 1059. Failures span skintokensLoadOrientation.test.js, taskAdvancedOptionsDecimation.test.js (multiple, past the three already skipped), viewportAnimationTarget.test.js, and src/library/taskModelUrl.test.js, unrelated to each other and to any one commit. .github/workflows/main.yml is deleted until this is fixed, not patched test by test under a deadline.
Decided, and waiting on order
The Replicate passthrough. RFD 1055 Phase 1 step 3 removes it. RFD 1055 also records why it stays until this box’s own worker answers.
The abandoned world cluster’s stale references. RFD 1064 pivots the roadmap to character concepts. RFD 1049, RFD 1050, RFD 1051, and RFD 1052 record the abandonment, but RFD 1016, RFD 1017, RFD 1026, RFD 1027, RFD 1037, and RFD 1038 still cite all four as active work. lib/weftspun_studio/inventory.ex still marks all four :active. Update each reference, or mark the four :vetoed, when the catalog is next touched.