Every excerpt below is copied from the repository at commit c7a2f8fed166b8b63e679d8655af4886602319ac. Nothing was run against RunPod to produce it. Lines longer than 1,600 characters are cut and marked.
1. The development environment of infra/runpod/endpoints/desired-state.json#
Command: git show c7a2f8fed166b8b63e679d8655af4886602319ac:infra/runpod/endpoints/desired-state.json, then the environments.development object, re-indented.
json
{
"endpoints": [
{
"key": "comfyui-chroma",
"managed": true,
"endpointIdEnv": "RUNPOD_COMFYUI_ENDPOINT_ID",
"canaryEndpointIdEnv": "RUNPOD_COMFYUI_CANARY_ENDPOINT_ID",
"canaryTrafficPercentage": 0,
"name": "isis-comfyui-chroma",
"gpuIds": "NVIDIA GeForce RTX 4090,NVIDIA GeForce RTX 5090,NVIDIA RTX A5000,NVIDIA L4,NVIDIA GeForce RTX 3090",
"templateName": "isis-comfyui-chroma-template",
"templateId": "vfl3iu1dr6",
"workersMin": 0,
"workersMax": 1,
"idleTimeout": 10,
"executionTimeout": 300,
"flashBoot": true,
"dataCenterId": "EU-RO-1",
"networkVolumeId": "5bst2pgw7y",
"containerDiskInGb": 30,
"image": null,
"imageRepository": "ghcr.io/greychimp/oshun/runpod-comfyui-slim",
"registryAuthName": "oshun-ghcr-pull",
"workerKind": "worker-comfyui",
"env": {
"COMFY_POLLING_MAX_RETRIES": "1500",
"COMFY_POLLING_INTERVAL_MS": "250",
"COMFY_LOG_LEVEL": "INFO",
"BUCKET_NAME": "app-dev-server",
"BUCKET_KEY_PREFIX": "isis/outputs",
"BUCKET_REGION": "hel1"
},
"notes": "Cost-bounded dev endpoint (ADR-0005). Image is built by RunPod from the GitHub repository (console: Serverless → New Endpoint → Import Git Repository; GitHub Actions is billing-blocked on this account as of 2026-09-11). `image` stays null for RunPod-built images; the live endpoint id is recorded after creation. `registryAuthId` is kept for the GHCR fallback path.",
"registryAuthId": "cmtx2h80b001kiob5v4p8dtwc",
"imageSource": {
"kind": "runpod-github",
"repository": "GreyChimp/oshun",
"branch": "runpod-comfyui-slim",
"dockerfilePath": "docker/runpod/comfyui-slim/Dockerfile",
"rebuildTrigger": "github-release",
"branchNote": "Quiet branch: every push to the source branch restarts the RunPod build and cancels the in-flight one, and main receives pushes from several agents. Fast-forward this branch deliberately when the worker image must change."
},
"endpointId": "6jye7vngg4i685",
"allowedCudaVersions": [
"12.8",
"12.9",
"13.0",
"13.1",
"13.2",
"13.3"
],
"gpuIdsNote": "RunPod GPU pools (saveEndpoint gpuIds): ADA_24 = 4090; AMPERE_24 = L4, A5000, 3090; BLACKWELL_32 = 5090, RTX PRO 4500 Blackwell. Widened 2026-09-11 because EU-RO-1 (the volume's datacenter) had no free 4090/A5000/3090 and workers sat throttled.",
"envFromEnvironment": [
"BUCKET_ENDPOINT_URL",
"BUCKET_ACCESS_KEY_ID",
"BUCKET_SECRET_ACCESS_KEY"
],
"secretRefs": {
"BUCKET_ACCESS_KEY_ID": "{{ RUNPOD_SECRET_oshun-hetzner-s3-access-key }}",
"BUCKET_SECRET_ACCESS_KEY": "{{ RUNPOD_SECRET_oshun-hetzner-s3-secret-key }}"
},
"workerOutputNote": "C.12.06 / V.12.05 (user decision 2026-09-13): outputs go to Hetzner Object Storage, bucket app-dev-server, prefix isis/outputs, through the image's oshun_uploader .pth hook. Console env set 2026-09-13 15:4xZ with the two keys as RunPod secret references (secretRefs); first live s3_url output 15:51Z. BUCKET_ENDPOINT_URL is https://hel1.your-objectstorage.com; the two keys are HETZNER_APP_DEV_S3_ACCESS_KEY / _SECRET_KEY in env-master.env. A git-integration template's env is console-only; runbook §5b."
},
{
"key": "comfyui-video",
"managed": true,
"endpointIdEnv": "RUNPOD_COMFYUI_VIDEO_ENDPOINT_ID",
"name": "isis-comfyui-video",
"gpuIds": "NVIDIA RTX PRO 6000 Blackwell Server Edition,NVIDIA RTX PRO 6000 Blackwell Workstation Edition,NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition,NVIDIA H100 PCIe,NVIDIA H100 80GB HBM3,NVIDIA H100 NVL,NVIDIA H200,NVIDIA A100 80GB PCIe,NVIDIA A100-SXM4-80GB",
"gpuSelectionNote": "Verified live 2026-09-12. `gpuPools` returns AMPERE_16, AMPERE_24, ADA_24, BLACKWELL_32, ADA_32_PRO, AMPERE_48, ADA_48_PRO, AMPERE_80, ADA_80_PRO, BLACKWELL_96, HOPPER_141, BLACKWELL_180, BLACKWELL_288 — there is NO 80 GB Hopper pool, so the chroma endpoint's pool-id form (ADA_24,...) cannot express \"H100 80 GB\". These are the exact `gpuTypes` ids: \"NVIDIA H100 80GB HBM3\" is SXM and \"NVIDIA H100 PCIe\" is PCIe, both 80 GB (\"NVIDIA H100 NVL\" is 94 GB and is excluded as a different price class). V.03.04 must read the console back after creation and record what it actually shows — the console is the authority on placement (rule 9), and REST GET echoes a stale list.",
"gpuFallbackNote": "If neither H100 type places a worker in EU-RO-1, the documented options are HOPPER_141 (H200 141 GB, a higher price class) or moving the video endpoint to a datacenter with H100 stock — which would need a second network volume, since volumes cannot move. Do NOT fall back to a smaller GPU: a 28.6 GB fp16 Wan expert does not fit 48 GB beside its encoders, so the job would OOM after the load was billed (V.03.05).",
"templateName": "isis-comfyui-video-template",
"templateId": "qcxiqduk01",
"workersMin": 0,
"workersMax": 1,
"idleTimeout": 10,
"executionTimeout": 900,
"flashBoot": true,
"dataCenterId": "EU-RO-1",
"networkVolumeId": "5bst2pgw7y",
"containerDiskInGb": 30,
"image": null,
"imageRepository": "ghcr.io/greychimp/oshun/runpod-comfyui-slim",
"registryAuthName": "oshun-ghcr-pull",
"registryAuthId": "cmtx2h80b001kiob5v4p8dtwc",
"workerKind": "worker-comfyui",
"env": {
"COMFY_POLLING_MAX_RETRIES": "3600",
"COMFY_POLLING_INTERVAL_MS": "250",
"COMFY_LOG_LEVEL": "INFO",
"BUCKET_NAME": "app-dev-server",
"BUCKET_KEY_PREFIX": "isis/outputs",
"BUCKET_REGION": "hel1"
},
"envFromEnvironment": [
"BUCKET_ENDPOINT_URL",
"BUCKET_ACCESS_KEY_ID",
"BUCKET_SECRET_ACCESS_KEY"
],
"notes": "Second endpoint for the `motion` catalog (decision D-V1): 80 GB so fp16 14B Wan experts, Animate, VACE and LTX-2.3 22B load without CPU offload, and a 900 s execution cap because a clip is minutes, not seconds. The image endpoint keeps its 24 GB / 300 s bounds untouched. Same worker image and the same network volume as comfyui-chroma (one image, D-V2). The three BUCKET_* names are listed by NAME only and injected by runpod-provision.mjs from the environment — never written into this file — so the worker can PUT video output straight to the volume's S3 endpoint instead of base64 (D-V4). NOT CREATED YET: V.03.04 creates it and records the live id.",
"imageSource": {
"kind": "runpod-github",
"repository": "GreyChimp/oshun",
"branch": "runpod-comfyui-slim",
"dockerfilePath": "docker/runpod/comfyui-slim/Dockerfile",
"rebuildTrigger": "github-release",
"branchNote": "Quiet branch: every push to the source branch restarts the RunPod build and cancels the in-flight one, and main receives pushes from several agents. Fast-forward this branch deliberately when the worker image must change."
},
"endpointId": "2lezocymx2rgot",
"secretRefs": {
"BUCKET_ACCESS_KEY_ID": "{{ RUNPOD_SECRET_oshun-hetzner-s3-access-key }}",
"BUCKET_SECRET_ACCESS_KEY": "{{ RUNPOD_SECRET_oshun-hetzner-s3-secret-key }}"
},
"workerOutputNote": "C.12.06 / V.12.05 (user decision 2026-09-13): outputs go to Hetzner Object Storage, bucket app-dev-server, prefix isis/outputs, through the image's oshun_uploader .pth hook. Console env set 2026-09-13 15:4xZ with the two keys as RunPod secret references (secretRefs); first live s3_url output 15:51Z. BUCKET_ENDPOINT_URL is https://hel1.your-objectstorage.com; the two keys are HETZNER_APP_DEV_S3_ACCESS_KEY / _SECRET_KEY in env-master.env. A git-integration template's env is console-only; runbook §5b."
}
]
}
2. docker/runpod/comfyui/RETIRED.md, whole#
Command: git show c7a2f8fed166b8b63e679d8655af4886602319ac:docker/runpod/comfyui/RETIRED.md
text
# This image lineage is retired — do not revive it without a decision
**Status:** retired, never published. Recorded 2026-09-12 (tracker C.12.05).
`oshunai/runpod-comfyui` **does not exist on Docker Hub** (404, verified
2026-09-11) and never will: no Docker Hub account is being created (user
decision, 2026-09-11). Every image this directory builds is tagged
`oshunai/runpod-comfyui:<flavor>`, so nothing built here has ever been pushed
anywhere, and the workflow that pushes it cannot succeed.
The design is also the one the Chroma/RunPod work deliberately moved away from:
each flavor **bakes** its own model set into the image (the `flagship-image`
inventory alone lists 47 baked artifacts). The replacement keeps models on a
RunPod network volume and loads them on demand, from a single slim image built
by RunPod's own git integration — `docker/runpod/comfyui-slim/`, which is what
the live endpoints actually run.
## Why it is still here
C.12.05 says to retire this "once nothing references them". Things still do.
This is the caller scan, run 2026-09-12, so the next person does not have to
redo it:
| Reference | What it does |
| ------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------- |
| `scripts/isis/generate_runpod_endpoint_image_manifests.mjs` | **Syncs INTO this directory** (`workflow_catalog`, `isis_nodes`) and is a CI gate over node/model consistency |
| `scripts/isis/validate_runpod_endpoint_image_manifests.mjs` | Validates the manifest the generator emits |
| `infra/runpod/endpoints/image-build-manifest.json` | **13 of its 17 endpoints** build from this context |
| `infra/runpod/endpoints/image-build-config.json` | The generator's input; the flavors are declared there |
| `.github/workflows/deploy-runpod.yml` | Path filter (line ~97) and a buildx job with this context (lines ~305–306) |
| `docker/runpod/comfyui/tests/test_catalog_renderer.py` | Unit tests for `catalog_renderer.py` |
| `DEPENDENCIES.md`, `V1/features/external-models-and-execution.md`, `ISIS_TODOS_2.md` | Prose references |
Roughly forty `scripts/isis/generate-v2-*`/`validate-v2-*` scripts read
`image-build-manifest.json` itself. They do not name this directory, but they do
consume what the generator produces from it, so the manifest cannot simply lose
13 endpoints without checking them.
## The order a removal has to happen in
Deleting this directory first would break a CI gate and a deploy workflow, which
is why C.12.05 is still open rather than done:
1. Decide what happens to the 13 flavor endpoints in
`infra/runpod/endpoints/image-build-config.json` — removed, or marked retired
the way `comfyui-vace` / `comfyui-ltx2` / `comfyui-flagship-video` already
are in `infra/runpod/endpoints/inventory/`
(`status: "retired-never-published"`).
2. Regenerate `image-build-manifest.json` and check every `generate-v2-*` /
`validate-v2-*` consumer still passes against the smaller manifest.
3. Remove the comfyui build job and path filter from
`.github/workflows/deploy-runpod.yml`.
4. Move `catalog_renderer.py` + its test, or drop them — the TypeScript port
(`libs/isis/workflows/src/templates/catalog-renderer.ts`) is what the live
path uses; this copy is the original the port was made from.
5. Then, and only then, delete this directory.
Until step 5, treat everything here as frozen: it is not the build path, it is
not deployed, and changes to it do not reach any running endpoint.
3. Tracker lines that open each cited item#
File ISIS_CHROMA_RUNPOD_MVP_TODOS_2026-09-11.md. For each item, the lines where its id appears in bold as the item's own heading (**<id>**).
C.12.12: 1 heading line(s)#
text
2086 - [x] **C.12.12** ~~Slim worker image fails at container create on RunPod.~~
V.03.04: 1 heading line(s)#
text
3065 - [x] **V.03.04** Create the endpoint (`endpoint create --yes`), drift check
V.01.09: 1 heading line(s)#
text
2874 - [x] **V.01.09** ComfyUI's own validator, in CI, for $0: after the smoke test,
V.01.07: 1 heading line(s)#
text
2831 - [x] **V.01.07** Point BOTH endpoints (`comfyui-chroma`, `comfyui-video`) at
A.05.01: 1 heading line(s)#
text
5902 - [x] **A.05.01** One worker-image build after the new goldens exist: the
C.12.18: 1 heading line(s)#
text
2103 - [x] **C.12.18** _(found by C.11.08; fixed and proved 2026-09-12)_
C.11.07: 1 heading line(s)#
text
1699 - [x] **C.11.07** `matting-remove-background` × 4 models on the C.11.02 output;
C.11.01: 1 heading line(s)#
text
1665 - [x] **C.11.01** `chroma-txt2img` with `chroma1-hd`, square*1024, default steps
V.01.04: 1 heading line(s)#
text
2751 - [x] **V.01.04** README: add the video contract — video outputs arrive in
L.06.03: 1 heading line(s)#
text
6313 - [x] **L.06.03** Point both endpoints at the new volume. Try the console's
V.12.05: 1 heading line(s)#
text
5016 - [ ] **V.12.05** Webhook completion + direct S3 output are already the
A.01.09: 1 heading line(s)#
text
5428 - [x] **A.01.09** SAM 3.1 on core nodes: row
C.12.05: 1 heading line(s)#
text
1952 - [ ] **C.12.05** Retire `docker/runpod/comfyui/` flavor images formally (delete