diff --git a/.claude/skills/remote-training/SKILL.md b/.claude/skills/remote-training/SKILL.md
index fd29303af..5f6fe6d77 100644
--- a/.claude/skills/remote-training/SKILL.md
+++ b/.claude/skills/remote-training/SKILL.md
@@ -1,245 +1,317 @@
 ---
 name: remote-training
-description: Manages remote training infrastructure on Nebius VMs. Use for building/pushing Docker images, starting/stopping VM machines (train, train2, train3), running training jobs, dataset generation, and starting inference servers.
+description: Manages remote training infrastructure on Nebius Serverless (Jobs for convert/train, Endpoints for inference). Use for dataset conversion, training jobs, serving checkpoints, and end-to-end pipeline validation.
 ---
 
-# Remote Training Infrastructure
+# Remote Training Infrastructure (Nebius Serverless)
 
-This skill manages the Positronic training infrastructure on Nebius GPU VMs. It covers Docker image management, VM lifecycle, training jobs, dataset generation, and inference server deployment.
+This skill runs the Positronic convert → train → serve pipeline on
+[Nebius Serverless](https://docs.nebius.com/serverless): **Jobs** for batch work
+(dataset conversion, training) and **Endpoints** for HTTP inference servers.
+Compute is provisioned per job/endpoint and released automatically when it
+finishes, so there is no idle compute cost.
 
-## S3 Convention
+All operations go through the wrapper scripts in `workflows/nebius/`. Run them from
+the repo root. The full reference is `workflows/nebius/README.md`.
 
-```
-s3://interim/{dataset}/{vendor}/{codec}/          — converted LeRobot datasets
-s3://checkpoints/{dataset}/{vendor}/{codec_or_experiment}/  — training output
-s3://inference/{dataset}/{date_or_exp}/{vendor}/  — inference eval results
-```
+## Vendor tokens
 
-Every run writes `run_metadata_*.yaml` with the full CLI command — read it to reconstruct the pipeline.
+Every script takes a `<vendor>` positional that selects the container image and `uv`
+extra. Supported: `lerobot_0_3_3` (ACT), `lerobot` (SmolVLA), `openpi`, `gr00t`.
 
-### Current Artifacts
+| Vendor | Image | Train/serve hardware |
+|--------|-------|----------------------|
+| `lerobot_0_3_3` (ACT) | `positro/positronic:latest` | H100 |
+| `lerobot` (SmolVLA) | `positro/positronic:latest` | H100 |
+| `openpi` | `positro/openpi:latest` | H100 |
+| `gr00t` | `positro/gr00t:latest` | H100 |
 
-**sim_stack** (automated testing — `@positronic.cfg.ds.phail.sim_stack_cubes`):
+Conversion always runs on CPU (`cpu-e2`, `8vcpu-32gb`); train/serve on
+`gpu-h100-sxm` (`1gpu-16vcpu-200gb`). `openpi`/`gr00t` re-use the `lerobot_0_3_3`
+converter with their own codec namespace.
 
-| Vendor | Codec | Interim | Latest Checkpoint |
-|--------|-------|---------|-------------------|
-| groot | `ee_rot6d` | `s3://interim/sim_stack/groot/ee_rot6d/` | `s3://checkpoints/sim_stack/groot/ee_rot6d/230226/` |
-| lerobot (0.4.x) | `ee` | `s3://interim/sim_stack/lerobot_04/ee/` | `s3://checkpoints/sim_stack/lerobot_04/smolvla_150k/` |
-| lerobot (0.3.3) | `ee` | `s3://interim/sim_stack/lerobot/ee/` | `s3://checkpoints/sim_stack/lerobot/230226-ee/` |
-| openpi | `ee` | `s3://interim/sim_stack/openpi/ee/` | `s3://checkpoints/sim_stack/openpi/ee/pi05_positronic_lowmem/230226/` |
+## S3 Convention
 
-**phail_unified** (production — `@positronic.cfg.ds.phail.phail_unified`):
+```
+s3://interim/{dataset}/{vendor}/{codec}/                    — converted LeRobot datasets
+s3://checkpoints/{dataset}/{vendor}/{codec_or_experiment}/  — training output
+s3://inference/{dataset}/{date_or_exp}/{vendor}/            — inference eval results
+```
 
-| Vendor | Codec | Interim | Latest Checkpoint |
-|--------|-------|---------|-------------------|
-| smolvla (0.4.x) | `ee` | `s3://interim/phail_unified/smolvla/ee/` | `s3://checkpoints/phail_unified/smolvla/150315_ft_150k/` |
-| groot | `ee_rot6d` | `s3://interim/phail_unified/groot/ee_rot6d/` | — |
-| lerobot (0.4.x) / SmolVLA | `ee` | — | `s3://checkpoints/phail_unified/smolvla/170316_ee/` |
-| lerobot (0.3.3) | `ee` | `s3://interim/phail_unified/lerobot/ee/` | — |
-| openpi | `ee` | `s3://interim/phail_unified/openpi/ee/` | — |
+Every run writes a `run_metadata_*.yaml` into its output directory capturing the
+full CLI command plus a snapshot of the code state (`*.py`/`*.toml`). See
+[Analysing a past run](#analysing-a-past-run) to reconstruct what produced a
+given checkpoint.
 
-## Two LeRobot Versions
+### Where current paths live
 
-Positronic ships two LeRobot integrations because other vendors (GR00T, OpenPI) depend on the 0.3.3 dataset format:
+Concrete dataset/checkpoint S3 paths are intentionally **not** listed here — they
+rotate as new runs land and any list goes stale. The source of truth is the
+config in the codebase:
 
-| | LeRobot 0.4.x | LeRobot 0.3.3 |
-|---|---|---|
-| **Convert** | `lerobot-convert` | `lerobot-0_3_3-convert` |
-| **Train** | `lerobot-train` | `lerobot-0_3_3-train` |
-| **Serve** | `lerobot-server` | `lerobot-0_3_3-server` |
-| **Codecs** | `@positronic.vendors.lerobot.codecs.*` | `@positronic.vendors.lerobot_0_3_3.codecs.*` |
-| **GPU** | Desktop (consumer GPU) | Desktop (consumer GPU) |
+- **Datasets**: `positronic/cfg/ds/` (e.g. `@positronic.cfg.ds.phail.sim_stack_cubes`,
+  `@positronic.cfg.ds.phail.phail_unified`).
+- **Checkpoints**: the per-vendor server presets in
+  `positronic/vendors/<vendor>/server.py` — the named configs (e.g. `phail`,
+  `sim_stack`) set `checkpoints_dir=` to the path currently in use. Read those
+  for the live values rather than relying on a checkpoint path memorized
+  anywhere.
 
-`lerobot-0_3_3-convert` is also used for GR00T and OpenPI dataset conversion.
+To discover what physically exists, `aws s3 ls` under the convention above; to
+learn what produced a given checkpoint, read its `run_metadata_*.yaml` (see
+[Analysing a past run](#analysing-a-past-run)).
 
-## Machines
+## Docker Images
 
-| Context | GPU | Use Case |
-|---------|-----|----------|
-| `desktop` | RTX 3060 (12GB) | Dataset generation, lerobot training/inference, GR00T inference |
-| `notebook` | RTX 4060 Laptop (8GB) | Light tasks, testing, dataset generation |
-| `vm-train` | H100 (80GB) | GR00T/OpenPI training and inference |
-| `vm-train2` | H100 (80GB) | GR00T/OpenPI training and inference |
-| `vm-train3` | H100 (80GB) | GR00T/OpenPI training and inference |
+Serverless jobs/endpoints pull `positro/<vendor>:${NEBIUS_IMAGE_TAG:-latest}` from the
+registry — they do **not** mount local source, so any code change needs a rebuild + push
+before it takes effect remotely.
 
-Only GR00T and OpenPI training/inference require H100. Everything else runs on `desktop`.
+**Tag gotcha:** locally `make push-*` pushes `:<branch>` and `:<sha>` but **not** `:latest`
+(that only happens under `CI`). So after a code change, either:
+- `cd docker && CI=1 make push-<x>` — updates `:latest` (what the workflow pulls by default), or
+- `cd docker && make push-<x> IMAGE_TAG=<branch>` then run the workflow with
+  `NEBIUS_IMAGE_TAG=<branch>` — tests a branch build without clobbering `:latest`.
 
-### VM Management
+Plain `make push-<x>` with no `CI`/`IMAGE_TAG`/`NEBIUS_IMAGE_TAG` leaves serverless
+running the **old** `:latest` image — the change silently won't take effect.
 
-**IMPORTANT**: Always check if a VM is running a job before using it.
+| Image | Source | Used For |
+|-------|--------|----------|
+| `positro/positronic` | `positronic/docker/` | Conversion, lerobot / SmolVLA train+serve |
+| `positro/openpi` | `positronic/docker/` (depends on `positro/openpi-base`) | OpenPI train+serve, openpi stats |
+| `positro/gr00t` | `positronic/docker/` (depends on `positro/gr00t-base`) | GR00T train+serve |
 
 ```bash
-# Check connectivity and running containers
-ssh -o ConnectTimeout=5 vertix@vm-train 'echo connected' 2>&1
-docker --context vm-train ps 2>/dev/null
-
-# Start a stopped VM
-../internal/scripts/start.sh train   # or train2, train3
+cd docker
+make push-training  # positro/positronic
+make push-openpi    # positro/openpi (rebuild positro/openpi-base first if ../openpi changed)
+make push-groot     # positro/gr00t (rebuild positro/gr00t-base first if ../gr00t changed)
+make push           # all images
 ```
 
-## Docker Images
+Cross-repo base rebuilds: `cd ../openpi/docker && make push` (or `../gr00t/docker`),
+then `cd ../positronic/docker && make push-openpi`. See `docker/CONTEXTS.md`.
 
-| Image | Source | Used For |
-|-------|--------|----------|
-| `positro/positronic` | `positronic/docker/` | Dataset conversion, lerobot training/inference |
-| `positro/gr00t` | `positronic/docker/` (depends on `positro/gr00t-base`) | GR00T training and inference |
-| `positro/openpi` | `positronic/docker/` (depends on `positro/openpi-base`) | OpenPI training and inference |
-| `positro/dreamzero` | `positronic/docker/` (depends on `positro/dreamzero-base`) | DreamZero inference |
+## One-time setup
 
-Images are tagged by branch name. `make push` in `docker/` auto-derives the tag from the current git branch.
+The pipeline reads credentials from Nebius MysteryBox secrets and uses a shared
+filesystem for `uv`/HF/openpi caches. This is already provisioned for the
+Positronic-internal project. To (re)create it for a different project, follow
+"One-time setup" in `workflows/nebius/README.md` (four MysteryBox secrets +
+one `network_ssd` filesystem).
 
-```bash
-cd /home/vertix/dev/positronic/docker
-make push-training  # Just positro/positronic
-make push-groot     # positro/gr00t (rebuild base first if ../gr00t changed)
-make push-openpi    # positro/openpi (rebuild base first if ../openpi changed)
-make push           # All images
-```
+Defaults point at the Positronic-internal project; override via env when needed:
 
-For cross-repo base image rebuilds: `cd ../gr00t/docker && make push` then `cd ../positronic/docker && make push-groot`.
+| Variable | Default | Purpose |
+|---|---|---|
+| `NEBIUS_PARENT_ID` | `project-e00f38wexevrr52b8j` | Project to create jobs/endpoints in |
+| `NEBIUS_SUBNET_ID` | `vpcsubnet-e00pk1j1x6hjmr4m92` | VPC subnet |
+| `WANDB_SECRET` | `positronic-serverless-wandb-api-key` | MysteryBox name for WandB key. Set empty to disable wandb. |
+| `NEBIUS_CACHE_FS` | `computefilesystem-e00f6jyfr5wkawyrab` | Shared cache filesystem **ID** (mounted RW at `/cache`) |
 
 ## Pipeline
 
-All commands run from `docker/` directory. Use `CACHE_ROOT=/home/vertix` when targeting remote Docker contexts from Mac.
-
 ### 1. Convert Dataset
 
+`convert.sh` runs the right converter + codec for the vendor as a CPU Job. For
+`openpi` it **blocks until convert finishes, then chains a stats job** and prints
+the `--stats_path` to use for training.
+
 ```bash
-# GR00T / OpenPI / LeRobot 0.3.3 — use lerobot-0_3_3-convert
-CACHE_ROOT=/home/vertix docker --context desktop compose run --rm --pull always lerobot-0_3_3-convert convert \
+bash workflows/nebius/convert.sh lerobot_0_3_3 \
   --dataset.dataset=@positronic.cfg.ds.phail.sim_stack_cubes \
-  --dataset.codec=@positronic.vendors.gr00t.codecs.ee_rot6d \
-  --output_dir=s3://interim/sim_stack/groot/ee_rot6d/
+  --dataset.codec=@positronic.vendors.lerobot_0_3_3.codecs.ee \
+  --output_dir=s3://interim/sim_stack/lerobot/ee/
 
-# LeRobot 0.4.x (SmolVLA) — use lerobot-convert
-CACHE_ROOT=/home/vertix docker --context desktop compose run --rm --pull always lerobot-convert convert \
+bash workflows/nebius/convert.sh openpi \
   --dataset.dataset=@positronic.cfg.ds.phail.sim_stack_cubes \
-  --dataset.codec=@positronic.vendors.lerobot.codecs.ee \
-  --output_dir=s3://interim/sim_stack/lerobot_04/ee/
+  --dataset.codec=@positronic.vendors.openpi.codecs.ee \
+  --output_dir=s3://interim/sim_stack/openpi/ee/
+# → also submits openpi-stats-* ; note the printed --stats_path=<...>/stats/assets/
 ```
 
-**Default codecs**: groot `ee_rot6d`, lerobot `ee`, openpi `ee`.
+Default codecs: gr00t `ee_rot6d`, lerobot `ee`, openpi `ee`,
+lerobot_0_3_3 `ee`.
 
 ### 2. Train
 
-Each vendor has a training script at `positronic/vendors/{vendor}/train.py` with usage examples in its docstring. Read the script's docstring for available subcommands and parameters.
+`train.sh` runs `python -m positronic.vendors.<vendor>.train` as an H100 Job. The
+dataset bucket is mounted read-only via Mountpoint-S3 at `/mnt/input` for
+`lerobot_0_3_3`; other vendors stream via `pos3` from the `s3://` path directly.
+`--output_dir` / `--output_path` stays an `s3://` URL (handled by `pos3`).
 
-General pattern:
+Read each vendor's `positronic/vendors/<vendor>/train.py` docstring for its exact
+flags. `--resume=true` resumes an interrupted run.
 
 ```bash
-# From docker/ directory
-[CACHE_ROOT=/home/vertix] docker [--context <machine>] compose run --rm --pull always <service> \
-  [subcommand] --input_path=<interim_path> --exp_name=<name> --output_dir=<checkpoint_path> ...
+# ACT / SmolVLA
+bash workflows/nebius/train.sh lerobot_0_3_3 \
+  --input_path=s3://interim/sim_stack/lerobot/ee/ \
+  --exp_name=act_sim_stack_v1 \
+  --output_dir=s3://checkpoints/sim_stack/lerobot/ \
+  --num_train_steps=50000 --save_freq=10000
+
+# OpenPI — needs --stats_path from the convert step's chained stats job
+bash workflows/nebius/train.sh openpi \
+  --input_path=s3://interim/sim_stack/openpi/ee/ \
+  --stats_path=s3://interim/sim_stack/openpi/stats/assets/ \
+  --output_path=s3://checkpoints/sim_stack/openpi/ \
+  --exp_name=pi_sim_stack_v1 \
+  --num_train_steps=30000
+# openpi checkpoint lands at <output_path>/pi05_positronic_lowmem/<exp_name>/
 ```
 
-| Vendor | Docker service | Machine | Script (read docstring for usage) |
-|--------|---------------|---------|-----------------------------------|
-| LeRobot 0.4.x (SmolVLA) | `lerobot-train` | desktop | `positronic/vendors/lerobot/train.py` |
-| LeRobot 0.3.3 (ACT) | `lerobot-0_3_3-train` | desktop | `positronic/vendors/lerobot_0_3_3/train.py` |
-| GR00T | `groot-train` | H100 | `positronic/vendors/gr00t/train.py` |
-| OpenPI | `openpi-train` (needs `openpi-stats` first) | H100 | `positronic/vendors/openpi/train.py` |
+The first job after a dependency change pays the full `uv`/HF cold-download
+(~10 min); later jobs reuse `/cache` and start faster.
+
+### 3. Serve a Checkpoint
+
+`serve.sh <vendor> <unique-endpoint-name> [server args...]` creates a public
+Endpoint on H100 port 8000, blocks until a public IP is allocated, and prints a
+banner containing `Endpoint URL:  http://<IP>` (the endpoint IP), the endpoint
+ID/name, and the teardown command. The container then takes ~10–15 min more
+to `uv sync` and load the model.
+
+```bash
+# Named preset — checkpoint path comes from the vendor's server.py config
+# (e.g. `demo`, `sim_stack`, `phail`). These are the source of truth; prefer them.
+bash workflows/nebius/serve.sh lerobot_0_3_3 my-act-demo demo
+
+# Explicit checkpoint dir (subcommand differs per vendor). Get <ckpt-dir> from the
+# vendor's server.py preset or `aws s3 ls` under the S3 convention — not memorized.
+bash workflows/nebius/serve.sh lerobot_0_3_3 act-server serve \
+  --checkpoints_dir=<ckpt-dir>
 
-**Resume any training**: add `--resume=true` to the same command.
+bash workflows/nebius/serve.sh openpi pi-server serve \
+  --checkpoints_dir=<ckpt-dir>
 
-### 3. Start Inference Server
+bash workflows/nebius/serve.sh gr00t groot-server ee_rot6d \
+  --checkpoints_dir=<ckpt-dir>
+```
 
-All servers use subcommands: `serve` for custom checkpoints, or named presets like `phail`, `sim_stack`.
+Sanity-check once warm:
 
 ```bash
-# LeRobot 0.4.x SmolVLA — preset (desktop)
-CACHE_ROOT=/home/vertix docker --context desktop compose run --rm --pull always --service-ports lerobot-server \
-  phail
-
-# LeRobot 0.4.x SmolVLA — custom checkpoint (desktop)
-CACHE_ROOT=/home/vertix docker --context desktop compose run --rm --pull always --service-ports lerobot-server \
-  --checkpoints_dir=s3://checkpoints/sim_stack/lerobot_04/smolvla_150k/
-
-# LeRobot 0.3.3 ACT (desktop) — requires subcommand
-CACHE_ROOT=/home/vertix docker --context desktop compose run --rm --pull always --service-ports lerobot-0_3_3-server \
-  serve \
-  --checkpoints_dir=s3://checkpoints/sim_stack/lerobot/230226-ee/
-
-# GR00T (desktop or H100) — codec subcommand required
-CACHE_ROOT=/home/vertix docker --context desktop compose run --rm --pull always --service-ports groot-server \
-  ee_rot6d \
-  --checkpoints_dir=s3://checkpoints/sim_stack/groot/ee_rot6d/230226/
-
-# OpenPI (H100) — requires subcommand
-docker --context vm-train compose run --rm --pull always --service-ports openpi-server \
-  serve \
-  --checkpoints_dir=s3://checkpoints/sim_stack/openpi/ee/pi05_positronic_lowmem/230226/
+curl http://<endpoint-ip>:8000/api/v1/models   # → {"models": ["<step>"]}
 ```
 
-All servers expose WebSocket API on port 8000. Available presets per server:
+Tear down (releases compute + public IP):
 
-| Server | Presets |
-|--------|---------|
-| `lerobot-server` | `serve`, `phail` |
-| `lerobot-0_3_3-server` | `serve`, `phail`, `sim_stack` |
-| `groot-server` | `serve`, `ee`, `ee_rot6d`, `phail`, `sim_stack`, ... |
-| `openpi-server` | `serve`, `phail`, `sim_stack` |
+```bash
+bash workflows/nebius/stop.sh my-act-demo
+```
+
+To pause without releasing the static IP: `nebius ai endpoint stop <id>`
+(`start` resumes).
 
 ### 4. Run Inference Client
 
+`serve.sh` prints the endpoint IP in its banner. To read it again later (by
+endpoint name):
+
 ```bash
-# With GUI
-uv run positronic-inference sim \
-  --policy=.remote --policy.host=desktop --policy.port=8000 \
-  --driver.show_gui
+nebius ai endpoint list --parent-id "$NEBIUS_PARENT_ID" --format json \
+  | jq -r --arg n "<endpoint-name>" \
+    '.items[] | select(.metadata.name==$n) | .status.public_endpoints[0]'
+# → the public IP; the endpoint serves on port 8000
+```
+
+Point the `positronic-inference` CLI at the endpoint IP:
 
-# Headless
-MUJOCO_GL=egl uv run positronic-inference sim \
-  --policy=.remote --policy.host=desktop --policy.port=8000 \
-  --driver.show_gui=False --driver.simulation_time=10
+```bash
+uv run positronic-inference sim \
+  --policy=.remote --policy.host=<endpoint-ip> --policy.port=8000 \
+  --output_dir=s3://inference/sim_stack_validation/<run_name>/<vendor>/
 ```
 
-## Sim Eval End-to-End
+View results locally (top-level dir compares multiple runs):
 
 ```bash
-# 1. Start server in background (-d flag)
-# Check the server script's cfn.cli() to determine if a subcommand is needed (see table above)
-CACHE_ROOT=/home/vertix docker --context <machine> compose run -d --rm --pull always --service-ports <server-service> \
-  [subcommand] --checkpoints_dir=<checkpoint_path>
-
-# Wait for ready
-docker --context <machine> logs --tail 5 <container_id>
-# Look for: "Uvicorn running on http://0.0.0.0:8000"
-
-# 2. Run sim episodes
-CACHE_ROOT=/home/vertix docker --context <machine> compose run --rm --pull always positronic-inference \
-  sim --policy=.remote --policy.host=<server_machine> --policy.port=8000 \
-  --driver.num_iterations=50 --driver.simulation_time=30 \
-  --output_dir=s3://inference/sim_stack_validation/<run_name>/<model_type>
-
-# 3. View results (locally) — pass top-level dir to compare multiple runs
 uv run python -m positronic.cfg.eval sim \
   --dataset.base.path=s3://inference/sim_stack_validation/<run_name> --reset_cache --https
-# Opens http://localhost:5001
+# http://localhost:5001
+```
+
+## End-to-End Validation
 
-# 4. Clean up
-docker --context <machine> stop <container_id>
+`e2e.sh` runs the whole pipeline for one vendor (convert → train 200 steps → serve
+→ `/api/v1/models` smoke → teardown), polling Nebius and printing a per-stage
+status line. ~$2–5 per run. Use it to verify a vendor still works after an
+image/dependency/script change.
+
+```bash
+bash workflows/nebius/e2e.sh openpi
+
+# All four — seed the cache with one first, then fan out warm
+bash workflows/nebius/e2e.sh lerobot_0_3_3
+for v in lerobot openpi gr00t; do bash workflows/nebius/e2e.sh "$v" & done; wait
+```
+
+Override via env: `E2E_S3_BASE`, `E2E_EXP_NAME`, `E2E_LOG_ROOT`, `E2E_DATASET`.
+
+## Analysing a past run
+
+Each convert/train/serve run writes `run_metadata_*.yaml` into its S3 output
+directory. It records the exact CLI command that produced the artifact and a
+snapshot of the relevant source files (`*.py`/`*.toml`), so a checkpoint or
+dataset can be traced back to the code and arguments that made it — without
+guessing.
+
+```bash
+# List the metadata files under a checkpoint/dataset output dir
+aws s3 ls s3://checkpoints/sim_stack/openpi/ee/pi05_positronic_lowmem/<exp>/ \
+  --recursive | grep run_metadata_
+
+# Read one (full command + code snapshot)
+aws s3 cp s3://checkpoints/.../run_metadata_YYYYMMDD_HHMMSS.yaml - | less
 ```
 
-**Naming**: `s3://inference/sim_stack_validation/<DDMMYY[-suffix]>/<model_type>/` where model_type is `lerobot`, `groot`, `openpi`, or `dreamzero`.
+To reproduce a run, copy the command recorded in `run_metadata_*.yaml` and
+resubmit it via the matching `workflows/nebius/*.sh` wrapper (set
+`NEBIUS_IMAGE_TAG` if the run used a non-`latest` image).
+
+For inference runs, each episode also has a `static.json` alongside the recorded
+data; the eval viewer in [4. Run Inference Client](#4-run-inference-client)
+(`positronic.cfg.eval sim --dataset.base.path=…`) renders these for inspection
+and side-by-side comparison of multiple runs.
 
-## Monitoring Background Jobs
+## Monitoring Jobs & Endpoints
 
 ```bash
-grep -o '[0-9]*%' /tmp/claude/-home-vertix-dev-positronic/tasks/<task_id>.output | tail -1
-tail -50 /tmp/claude/-home-vertix-dev-positronic/tasks/<task_id>.output
-grep -i "error\|complete\|finished" /tmp/claude/-home-vertix-dev-positronic/tasks/<task_id>.output
+# Jobs
+nebius ai job get  <aijob-id>                 # state: PROVISIONING/STARTING/RUNNING/COMPLETED/FAILED
+nebius ai job logs <aijob-id> --follow
+nebius ai job list --parent-id "$NEBIUS_PARENT_ID" --format json | jq '.items[].metadata.name'
+
+# Endpoints
+nebius ai endpoint get  <endpoint-id>
+nebius ai endpoint logs <endpoint-id> --follow   # wait for "INFO Started server process"
+nebius ai endpoint list --parent-id "$NEBIUS_PARENT_ID" --format json
 ```
 
+The `create` call streams the job ID and ready-to-paste follow-up commands.
+
 ## Common Issues
 
-- **CUDA OOM**: Each GR00T server uses ~6GB. On 12GB GPUs (desktop), only one server at a time.
-- **Port conflict**: `docker ps -a | grep -E "groot-server|openpi-server"` then `docker stop <id>`.
-- **VM unreachable**: `../internal/scripts/start.sh train2` then verify SSH.
-- **Headless rendering**: Use `MUJOCO_GL=egl` env var.
+- **Job stuck in PROVISIONING/STARTING**: normal — image pull + `uv` resolve.
+  First run on a cold cache is ~10 min; check `nebius ai job logs <id> --follow`.
+- **OpenPI train can't find stats**: pass `--stats_path=<output_dir-sibling>/stats/assets/`
+  exactly as printed by `convert.sh openpi`. Stats must be a *sibling* of the
+  dataset dir (pos3 forbids upload-inside-download).
+- **gr00t with read-only input mount fails**: only `lerobot_0_3_3` uses the RO
+  Mountpoint-S3 mount; gr00t writes back into the dataset dir, so it streams via
+  `pos3` from the `s3://` path instead (handled automatically by `train.sh`).
+- **Endpoint name collision**: names must be unique in the project. Pick a fresh
+  name or `stop.sh` the old one.
+- **Cold cache + parallel fan-out**: 4 jobs racing to populate `/cache` thrash;
+  seed with one vendor, then fan out the rest warm (see `e2e.sh` header).
+- **Wipe the shared cache**: throwaway `busybox` job mounting the FS — see
+  "Cleaning the shared cache" in `workflows/nebius/README.md`.
 
 ### Nebius Auth (Headless)
 
 1. `nebius --no-browser --auth-timeout 5m iam whoami 2>&1` — extract auth URL
 2. User clicks URL, browser redirects to `http://127.0.0.1:PORT/?code=XXX&state=YYY`
 3. `curl -s "http://127.0.0.1:PORT/?code=XXX&state=YYY"` on the machine running nebius
-4. Auth completes, VM scripts work
+4. Auth completes, scripts work
diff --git a/positronic/offboard/vendor_server.py b/positronic/offboard/vendor_server.py
index a71c40732..1614c9a61 100644
--- a/positronic/offboard/vendor_server.py
+++ b/positronic/offboard/vendor_server.py
@@ -2,6 +2,7 @@
 
 import asyncio
 import logging
+import time
 from abc import ABC, abstractmethod
 from collections.abc import Callable
 from typing import Any
@@ -113,13 +114,24 @@ class VendorServer(ABC):
         resolve_model(None) → create_policy → reset → warmup
     """
 
-    def __init__(self, codec: Codec | None, host: str = '0.0.0.0', port: int = 8000, recording_dir: str | None = None):
+    def __init__(
+        self,
+        codec: Codec | None,
+        host: str = '0.0.0.0',
+        port: int = 8000,
+        recording_dir: str | None = None,
+        idle_timeout_min: float | None = None,
+    ):
         self.codec = codec
         self.host = host
         self.port = port
         if recording_dir:
             self.codec = RecordingCodec(self.codec, pos3.sync(recording_dir))
 
+        self.idle_timeout_min = idle_timeout_min
+        self._active_sessions = 0
+        self._last_activity = time.monotonic()
+
         self.metadata: dict[str, Any] = {}
 
         self.app = FastAPI()
@@ -168,6 +180,8 @@ async def websocket_endpoint(self, websocket: WebSocket, model_id: str | None =
         await websocket.accept()
         logger.info(f'Connected to {websocket.client} requesting {model_id or "default"}')
 
+        self._active_sessions += 1
+        self._last_activity = time.monotonic()
         model_handle = None
         try:
             model_handle, extra_meta = await self.resolve_model(model_id, websocket)
@@ -180,6 +194,7 @@ async def websocket_endpoint(self, websocket: WebSocket, model_id: str | None =
             try:
                 while True:
                     message = await websocket.receive_bytes()
+                    self._last_activity = time.monotonic()
                     try:
                         raw_obs = deserialise(message)
                         actions = policy.select_action(raw_obs)
@@ -198,6 +213,8 @@ async def websocket_endpoint(self, websocket: WebSocket, model_id: str | None =
             except Exception:
                 logger.debug('Failed to send error to client', exc_info=True)
         finally:
+            self._active_sessions = max(0, self._active_sessions - 1)
+            self._last_activity = time.monotonic()
             if model_handle is not None:
                 await self.release_policy(model_handle)
 
@@ -207,11 +224,33 @@ async def _startup(self):
         policy.reset()
         await self.warmup(policy)
 
+    async def _idle_watchdog(self, server: uvicorn.Server):
+        timeout_s = self.idle_timeout_min * 60
+        poll = min(timeout_s, 30)
+        while not server.should_exit:
+            await asyncio.sleep(poll)
+            if self._active_sessions > 0:
+                continue
+            idle = time.monotonic() - self._last_activity
+            if idle >= timeout_s:
+                logger.warning(f'No activity for {idle:.0f}s (idle timeout {timeout_s:.0f}s); shutting down server')
+                server.should_exit = True
+                return
+
     def serve(self):
         async def _run():
             await self._startup()
             config = uvicorn.Config(self.app, host=self.host, port=self.port, log_level='info')
-            await uvicorn.Server(config).serve()
+            server = uvicorn.Server(config)
+            self._last_activity = time.monotonic()
+            watchdog = None
+            if self.idle_timeout_min and self.idle_timeout_min > 0:
+                watchdog = asyncio.create_task(self._idle_watchdog(server))
+            try:
+                await server.serve()
+            finally:
+                if watchdog is not None:
+                    watchdog.cancel()
 
         try:
             asyncio.run(_run())
diff --git a/positronic/vendors/dreamzero/server.py b/positronic/vendors/dreamzero/server.py
index eee199810..ac09ce065 100644
--- a/positronic/vendors/dreamzero/server.py
+++ b/positronic/vendors/dreamzero/server.py
@@ -202,8 +202,11 @@ def __init__(
         roboarena_port: int = 1234,
         enable_dit_cache: bool = True,
         recording_dir: str | None = None,
+        idle_timeout_min: float | None = None,
     ):
-        super().__init__(codec=codec, host=host, port=port, recording_dir=recording_dir)
+        super().__init__(
+            codec=codec, host=host, port=port, recording_dir=recording_dir, idle_timeout_min=idle_timeout_min
+        )
         self.model_path = model_path
         self.dreamzero_venv = Path(dreamzero_venv)
         self.backbone = backbone
@@ -272,6 +275,7 @@ def shutdown_model(self):
     port=8000,
     enable_dit_cache=True,
     recording_dir=None,
+    idle_timeout_min=None,
 )
 def server(
     codec: Codec | None,
@@ -282,6 +286,7 @@ def server(
     port: int,
     enable_dit_cache: bool,
     recording_dir: str | None,
+    idle_timeout_min: float | None,
 ):
     """Starts the DreamZero inference server."""
     with pos3.mirror():
@@ -294,6 +299,7 @@ def server(
             port=port,
             enable_dit_cache=enable_dit_cache,
             recording_dir=recording_dir,
+            idle_timeout_min=idle_timeout_min,
         ).serve()
 
 
diff --git a/positronic/vendors/gr00t/server.py b/positronic/vendors/gr00t/server.py
index 123856cd3..e844af450 100644
--- a/positronic/vendors/gr00t/server.py
+++ b/positronic/vendors/gr00t/server.py
@@ -278,8 +278,11 @@ def __init__(
         zmq_port: int = 5555,
         recording_dir: str | None = None,
         ready_timeout: float = 120.0,
+        idle_timeout_min: float | None = None,
     ):
-        super().__init__(codec=codec, host=host, port=port, recording_dir=recording_dir)
+        super().__init__(
+            codec=codec, host=host, port=port, recording_dir=recording_dir, idle_timeout_min=idle_timeout_min
+        )
         self.checkpoints_dir = checkpoints_dir.rstrip('/')
         self.checkpoint = checkpoint
         self.modality_config = modality_config
@@ -376,6 +379,7 @@ def shutdown_model(self):
     modality_config='ee',
     recording_dir=None,
     ready_timeout=120.0,
+    idle_timeout_min=None,
 )
 def server(
     codec: Codec,
@@ -386,6 +390,7 @@ def server(
     modality_config: str,
     recording_dir: str | None,
     ready_timeout: float,
+    idle_timeout_min: float | None,
 ):
     """Starts the GR00T inference server with encoding/decoding."""
 
@@ -399,6 +404,7 @@ def server(
             port=port,
             recording_dir=recording_dir,
             ready_timeout=ready_timeout,
+            idle_timeout_min=idle_timeout_min,
         ).serve()
 
 
diff --git a/positronic/vendors/lerobot/server.py b/positronic/vendors/lerobot/server.py
index 857ad1fe9..60afe1a61 100644
--- a/positronic/vendors/lerobot/server.py
+++ b/positronic/vendors/lerobot/server.py
@@ -33,8 +33,11 @@ def __init__(
         port: int = 8000,
         device: str | None = None,
         recording_dir: str | None = None,
+        idle_timeout_min: float | None = None,
     ):
-        super().__init__(codec=codec, host=host, port=port, recording_dir=recording_dir)
+        super().__init__(
+            codec=codec, host=host, port=port, recording_dir=recording_dir, idle_timeout_min=idle_timeout_min
+        )
         self.checkpoints_dir = str(checkpoints_dir).rstrip('/') + '/checkpoints'
         self.checkpoint = checkpoint
         self.device = device or _detect_device()
@@ -73,10 +76,28 @@ async def release_policy(self, model_handle):
         await self.policy_manager.release_session()
 
 
-@cfn.config(codec=lerobot_codecs.ee, checkpoint=None, port=8000, host='0.0.0.0', recording_dir=None)
-def main(checkpoints_dir: str, checkpoint: str | None, codec, port: int, host: str, recording_dir: str | None):
+@cfn.config(
+    codec=lerobot_codecs.ee, checkpoint=None, port=8000, host='0.0.0.0', recording_dir=None, idle_timeout_min=None
+)
+def main(
+    checkpoints_dir: str,
+    checkpoint: str | None,
+    codec,
+    port: int,
+    host: str,
+    recording_dir: str | None,
+    idle_timeout_min: float | None,
+):
     checkpoints_dir = str(pos3.download(checkpoints_dir))
-    InferenceServer(codec, checkpoints_dir, checkpoint, host=host, port=port, recording_dir=recording_dir).serve()
+    InferenceServer(
+        codec,
+        checkpoints_dir,
+        checkpoint,
+        host=host,
+        port=port,
+        recording_dir=recording_dir,
+        idle_timeout_min=idle_timeout_min,
+    ).serve()
 
 
 phail = main.override(
diff --git a/positronic/vendors/lerobot_0_3_3/server.py b/positronic/vendors/lerobot_0_3_3/server.py
index df6450b16..2782d9361 100644
--- a/positronic/vendors/lerobot_0_3_3/server.py
+++ b/positronic/vendors/lerobot_0_3_3/server.py
@@ -43,8 +43,11 @@ def __init__(
         metadata: dict[str, Any] | None = None,
         device: str | None = None,
         recording_dir: str | None = None,
+        idle_timeout_min: float | None = None,
     ):
-        super().__init__(codec=codec, host=host, port=port, recording_dir=recording_dir)
+        super().__init__(
+            codec=codec, host=host, port=port, recording_dir=recording_dir, idle_timeout_min=idle_timeout_min
+        )
         self.policy_factory = policy_factory
         self.checkpoints_dir = str(checkpoints_dir).rstrip('/') + '/checkpoints'
         self.checkpoint = checkpoint
@@ -96,7 +99,15 @@ def act(checkpoint_path: str) -> PreTrainedPolicy:
     return policy
 
 
-@cfn.config(policy_factory=act, codec=lerobot_codecs.ee, checkpoint=None, port=8000, host='0.0.0.0', recording_dir=None)
+@cfn.config(
+    policy_factory=act,
+    codec=lerobot_codecs.ee,
+    checkpoint=None,
+    port=8000,
+    host='0.0.0.0',
+    recording_dir=None,
+    idle_timeout_min=None,
+)
 def main(
     policy_factory: Callable[[str], PreTrainedPolicy],
     checkpoints_dir: str,
@@ -105,10 +116,18 @@ def main(
     port: int,
     host: str,
     recording_dir: str | None,
+    idle_timeout_min: float | None,
 ):
     checkpoints_dir = str(pos3.download(checkpoints_dir))
     InferenceServer(
-        policy_factory, codec, checkpoints_dir, checkpoint, host=host, port=port, recording_dir=recording_dir
+        policy_factory,
+        codec,
+        checkpoints_dir,
+        checkpoint,
+        host=host,
+        port=port,
+        recording_dir=recording_dir,
+        idle_timeout_min=idle_timeout_min,
     ).serve()
 
 
@@ -123,12 +142,16 @@ def main(
 _DEMO_CHECKPOINT = 's3://positronic-public/checkpoints/sim_stack_cubes/act/'
 
 
-@cfn.config(policy_factory=act, codec=lerobot_codecs.ee, checkpoint=None, port=8000, host='0.0.0.0')
-def demo(policy_factory, checkpoint, codec, port, host):
+@cfn.config(
+    policy_factory=act, codec=lerobot_codecs.ee, checkpoint=None, port=8000, host='0.0.0.0', idle_timeout_min=None
+)
+def demo(policy_factory, checkpoint, codec, port, host, idle_timeout_min):
     from positronic.cfg.ds import PUBLIC
 
     checkpoints_dir = str(pos3.download(_DEMO_CHECKPOINT, profile=PUBLIC))
-    InferenceServer(policy_factory, codec, checkpoints_dir, checkpoint, host=host, port=port).serve()
+    InferenceServer(
+        policy_factory, codec, checkpoints_dir, checkpoint, host=host, port=port, idle_timeout_min=idle_timeout_min
+    ).serve()
 
 
 if __name__ == '__main__':
diff --git a/positronic/vendors/openpi/server.py b/positronic/vendors/openpi/server.py
index d5f4c5e22..9b0ea26f4 100644
--- a/positronic/vendors/openpi/server.py
+++ b/positronic/vendors/openpi/server.py
@@ -190,8 +190,11 @@ def __init__(
         openpi_ws_port: int = 8001,
         metadata: dict[str, Any] | None = None,
         recording_dir: str | None = None,
+        idle_timeout_min: float | None = None,
     ):
-        super().__init__(codec=codec, host=host, port=port, recording_dir=recording_dir)
+        super().__init__(
+            codec=codec, host=host, port=port, recording_dir=recording_dir, idle_timeout_min=idle_timeout_min
+        )
         self.checkpoints_dir = str(checkpoints_dir).rstrip('/')
         self.config_name = config_name
         self.checkpoint = checkpoint
@@ -280,6 +283,7 @@ def shutdown_model(self):
     port=8000,
     openpi_ws_port=8001,
     recording_dir=None,
+    idle_timeout_min=None,
 )
 def server(
     codec,
@@ -290,6 +294,7 @@ def server(
     port: int,
     openpi_ws_port: int,
     recording_dir: str | None,
+    idle_timeout_min: float | None,
 ):
     """OpenPI inference server.
 
@@ -316,6 +321,7 @@ def server(
         port=port,
         openpi_ws_port=openpi_ws_port,
         recording_dir=recording_dir,
+        idle_timeout_min=idle_timeout_min,
     ).serve()
 
 
diff --git a/workflows/nebius/README.md b/workflows/nebius/README.md
index 2cb37a45a..1e0ea156c 100644
--- a/workflows/nebius/README.md
+++ b/workflows/nebius/README.md
@@ -70,7 +70,9 @@ nebius mysterybox secret create \
 The names matter — `convert.sh`, `train.sh`, and `serve.sh` reference the secrets by name. If a
 secret with one of these names already exists, the create call fails; skip it.
 
-Also create one shared filesystem for the dependency caches. Every job and endpoint mounts it
+## Shared cache filesystem
+
+Create one shared filesystem for the dependency caches. Every job and endpoint mounts it
 read-write at `/cache`; `uv`, HuggingFace, and openpi asset downloads land there and persist
 across cold starts, so only the first run after a dependency change pays the download cost:
 
@@ -91,38 +93,8 @@ The filesystem is RWX — many jobs/endpoints attach it concurrently. pos3's own
 (`~/.cache/positronic/s3/`) is deliberately *not* redirected here; it stays on each container's
 local disk and re-fetches from S3 by design.
 
-### Cleaning the shared cache
-
-There is no file browser for the filesystem — to inspect or wipe it you mount it in a
-throwaway job. Make sure no jobs/endpoints are using the cache first (a wipe while a
-warm job reads it will break that job).
-
-Inspect usage:
-
-```bash
-nebius ai job create --parent-id "$PARENT_ID" --subnet-id "$SUBNET_ID" \
-  --name cache-du --image busybox:latest \
-  --container-command du --args '-sh /cache /cache/uv /cache/hf /cache/openpi' \
-  --platform cpu-e2 --preset 4vcpu-16gb --timeout 1h \
-  --volume "$NEBIUS_CACHE_FS:/cache:rw"
-# then: nebius ai job logs <aijob-id>
-```
-
-Wipe everything (full reset — the next run repays the cold download):
-
-```bash
-nebius ai job create --parent-id "$PARENT_ID" --subnet-id "$SUBNET_ID" \
-  --name cache-wipe --image busybox:latest \
-  --container-command find --args '/cache -mindepth 1 -delete' \
-  --platform cpu-e2 --preset 4vcpu-16gb --timeout 1h \
-  --volume "$NEBIUS_CACHE_FS:/cache:rw"
-```
-
-To clear only one tool's cache, target its subdir, e.g. `--args '/cache/uv -mindepth 1
--delete'`. Two gotchas: `--volume` needs the filesystem **ID** (not name), and Nebius
-space-splits `--args`, so use a no-shell command (`find`/`du`) — a quoted `sh -c "..."`
-gets torn apart. Deleting and recreating the filesystem also works but loses the warm
-cache for every workflow.
+To inspect or wipe this filesystem later, see
+[Appendix: Cleaning the shared cache](#appendix-cleaning-the-shared-cache).
 
 ## Convert a Positronic dataset
 
@@ -311,6 +283,7 @@ override them** with their own project + subnet IDs:
 | `NEBIUS_SUBNET_ID` | `vpcsubnet-e00pk1j1x6hjmr4m92` | VPC subnet for the compute instance |
 | `WANDB_SECRET` | `positronic-serverless-wandb-api-key` | MysteryBox secret name for the WandB key. Set empty (`WANDB_SECRET=`) to skip wandb entirely. |
 | `NEBIUS_CACHE_FS` | `computefilesystem-e00f6jyfr5wkawyrab` | Shared filesystem **ID** (not name — `--volume` rejects names) mounted RW at `/cache` for the `uv`/HF/openpi caches (`UV_CACHE_DIR`, `HF_HOME`, `OPENPI_DATA_HOME`). Not used by pos3. The default is Positronic-internal; external users must override with their own filesystem ID. |
+| `NEBIUS_IMAGE_TAG` | `latest` | Docker image tag the job/endpoint pulls (`positro/<image>:<tag>`). `cd docker && make push-* IMAGE_TAG=<branch>` pushes that tag unconditionally; set `NEBIUS_IMAGE_TAG=<branch>` to run a branch build remotely without clobbering `:latest`. `make push-*` only updates `:latest` when run with `CI` set. Note `convert.sh openpi` chains a stats job on the `positro/openpi` image, so with `NEBIUS_IMAGE_TAG=<branch>` you must also have pushed `positro/openpi:<branch>` (not just `positro/positronic:<branch>`); otherwise leave `NEBIUS_IMAGE_TAG` unset so stats uses `:latest`. |
 
 Other operational settings (platform/preset, MysteryBox secret names, S3 endpoint URL, region)
 are hardcoded — change them by editing the script directly. The vendor positional arg selects
@@ -322,3 +295,36 @@ the container image and `uv` extras:
 | `lerobot` (SmolVLA) | `positro/positronic` | `--extra lerobot` |
 | `openpi` | `positro/openpi` | `--extra openpi` (serve); none for train/stats |
 | `gr00t` | `positro/gr00t` | _(none — `/gr00t` is co-installed)_ |
+
+## Appendix: Cleaning the shared cache
+
+There is no file browser for the [shared cache filesystem](#shared-cache-filesystem) — to
+inspect or wipe it you mount it in a throwaway job. Make sure no jobs/endpoints are using the
+cache first (a wipe while a warm job reads it will break that job).
+
+Inspect usage:
+
+```bash
+nebius ai job create --parent-id "$PARENT_ID" --subnet-id "$SUBNET_ID" \
+  --name cache-du --image busybox:latest \
+  --container-command du --args '-sh /cache /cache/uv /cache/hf /cache/openpi' \
+  --platform cpu-e2 --preset 4vcpu-16gb --timeout 1h \
+  --volume "$NEBIUS_CACHE_FS:/cache:rw"
+# then: nebius ai job logs <aijob-id>
+```
+
+Wipe everything (full reset — the next run repays the cold download):
+
+```bash
+nebius ai job create --parent-id "$PARENT_ID" --subnet-id "$SUBNET_ID" \
+  --name cache-wipe --image busybox:latest \
+  --container-command find --args '/cache -mindepth 1 -delete' \
+  --platform cpu-e2 --preset 4vcpu-16gb --timeout 1h \
+  --volume "$NEBIUS_CACHE_FS:/cache:rw"
+```
+
+To clear only one tool's cache, target its subdir, e.g. `--args '/cache/uv -mindepth 1
+-delete'`. Two gotchas: `--volume` needs the filesystem **ID** (not name), and Nebius
+space-splits `--args`, so use a no-shell command (`find`/`du`) — a quoted `sh -c "..."`
+gets torn apart. Deleting and recreating the filesystem also works but loses the warm
+cache for every workflow.
diff --git a/workflows/nebius/convert.sh b/workflows/nebius/convert.sh
index f9a507ab6..1187e3781 100644
--- a/workflows/nebius/convert.sh
+++ b/workflows/nebius/convert.sh
@@ -18,6 +18,10 @@ SUBNET_ID="${NEBIUS_SUBNET_ID:-vpcsubnet-e00pk1j1x6hjmr4m92}"
 # Shared filesystem (RWX) holding the uv / HF / openpi caches across cold starts.
 # pos3's cache stays on local disk (~/.cache/positronic/s3) — never redirected here.
 CACHE_FS="${NEBIUS_CACHE_FS:-computefilesystem-e00f6jyfr5wkawyrab}"
+# Docker image tag pulled by the job. `make push-*` only updates `:latest` under
+# CI; locally it pushes `:<branch>`/`:<sha>`. To convert with a branch build:
+# `make push-training IMAGE_TAG=<branch>` then run with `NEBIUS_IMAGE_TAG=<branch>`.
+IMAGE_TAG="${NEBIUS_IMAGE_TAG:-latest}"
 
 if [ $# -lt 1 ]; then
   cat >&2 <<'EOF'
@@ -83,7 +87,7 @@ CREATE_OUT=$(nebius ai job create \
   --parent-id "$PARENT_ID" \
   --subnet-id "$SUBNET_ID" \
   --name "$JOB_NAME" \
-  --image positro/positronic:latest \
+  --image "positro/positronic:${IMAGE_TAG}" \
   --container-command uv \
   --args "$CONVERT_ARGS" \
   --platform cpu-e2 \
@@ -149,7 +153,7 @@ nebius ai job create \
   --parent-id "$PARENT_ID" \
   --subnet-id "$SUBNET_ID" \
   --name "$STATS_JOB_NAME" \
-  --image positro/openpi:latest \
+  --image "positro/openpi:${IMAGE_TAG}" \
   --container-command uv \
   --args "$STATS_ARGS" \
   --platform cpu-e2 \
diff --git a/workflows/nebius/serve.sh b/workflows/nebius/serve.sh
index fda3ba318..045f5d1c2 100644
--- a/workflows/nebius/serve.sh
+++ b/workflows/nebius/serve.sh
@@ -16,6 +16,10 @@ SUBNET_ID="${NEBIUS_SUBNET_ID:-vpcsubnet-e00pk1j1x6hjmr4m92}"
 # Shared filesystem (RWX) holding the uv / HF / openpi caches across cold starts.
 # pos3's cache stays on local disk (~/.cache/positronic/s3) — never redirected here.
 CACHE_FS="${NEBIUS_CACHE_FS:-computefilesystem-e00f6jyfr5wkawyrab}"
+# Docker image tag pulled by the endpoint. `make push-*` only updates `:latest`
+# under CI; locally it pushes `:<branch>`/`:<sha>`. To serve a branch build:
+# `make push-<x> IMAGE_TAG=<branch>` then run with `NEBIUS_IMAGE_TAG=<branch>`.
+IMAGE_TAG="${NEBIUS_IMAGE_TAG:-latest}"
 
 if [ $# -lt 2 ]; then
   cat >&2 <<'EOF'
@@ -55,17 +59,25 @@ NAME="$2"
 shift 2
 
 case "$VENDOR" in
-  lerobot_0_3_3) IMAGE="positro/positronic:latest"; EXTRA="--extra lerobot_0_3_3 " ;;
-  lerobot)       IMAGE="positro/positronic:latest"; EXTRA="--extra lerobot " ;;
+  lerobot_0_3_3) IMAGE="positro/positronic:${IMAGE_TAG}"; EXTRA="--extra lerobot_0_3_3 " ;;
+  lerobot)       IMAGE="positro/positronic:${IMAGE_TAG}"; EXTRA="--extra lerobot " ;;
   # openpi.server imports `openpi_client` at module top → needs --extra openpi
-  openpi)        IMAGE="positro/openpi:latest";     EXTRA="--extra openpi " ;;
-  gr00t)         IMAGE="positro/gr00t:latest";      EXTRA="" ;;
+  openpi)        IMAGE="positro/openpi:${IMAGE_TAG}";     EXTRA="--extra openpi " ;;
+  gr00t)         IMAGE="positro/gr00t:${IMAGE_TAG}";      EXTRA="" ;;
   *)
     echo "Unknown vendor: '$VENDOR'. Supported: lerobot_0_3_3 | lerobot | openpi | gr00t" >&2
     exit 1
     ;;
 esac
 
+# Serverless endpoints have no native idle/scale-to-zero, so opt the server into
+# self-shutdown (the base default is no timeout). Override the window with
+# NEBIUS_IDLE_TIMEOUT_MIN; skip injection if the caller already passed one.
+case " $* " in
+  *" --idle_timeout_min="*|*" --idle_timeout_min "*) ;;
+  *) set -- "$@" "--idle_timeout_min=${NEBIUS_IDLE_TIMEOUT_MIN:-20}" ;;
+esac
+
 SERVER_ARGS="run --python 3.11 ${EXTRA}python -m positronic.vendors.${VENDOR}.server $*"
 
 echo "Creating $VENDOR endpoint '$NAME'..."
diff --git a/workflows/nebius/train.sh b/workflows/nebius/train.sh
index e8400ed9b..eb081dea8 100644
--- a/workflows/nebius/train.sh
+++ b/workflows/nebius/train.sh
@@ -20,6 +20,10 @@ WANDB_SECRET="${WANDB_SECRET-positronic-serverless-wandb-api-key}"
 # Shared filesystem (RWX) holding the uv / HF / openpi caches across cold starts.
 # pos3's cache stays on local disk (~/.cache/positronic/s3) — never redirected here.
 CACHE_FS="${NEBIUS_CACHE_FS:-computefilesystem-e00f6jyfr5wkawyrab}"
+# Docker image tag pulled by the job. `make push-*` only updates `:latest` under
+# CI; locally it pushes `:<branch>`/`:<sha>`. To test a branch build remotely:
+# `make push-<x> IMAGE_TAG=<branch>` then run with `NEBIUS_IMAGE_TAG=<branch>`.
+IMAGE_TAG="${NEBIUS_IMAGE_TAG:-latest}"
 
 if [ $# -lt 1 ]; then
   cat >&2 <<'EOF'
@@ -42,10 +46,10 @@ VENDOR="$1"
 shift
 
 case "$VENDOR" in
-  lerobot_0_3_3) IMAGE="positro/positronic:latest"; EXTRA="--extra lerobot_0_3_3 " ;;
-  lerobot)       IMAGE="positro/positronic:latest"; EXTRA="--extra lerobot " ;;
-  openpi)        IMAGE="positro/openpi:latest";     EXTRA="" ;;
-  gr00t)         IMAGE="positro/gr00t:latest";      EXTRA="" ;;
+  lerobot_0_3_3) IMAGE="positro/positronic:${IMAGE_TAG}"; EXTRA="--extra lerobot_0_3_3 " ;;
+  lerobot)       IMAGE="positro/positronic:${IMAGE_TAG}"; EXTRA="--extra lerobot " ;;
+  openpi)        IMAGE="positro/openpi:${IMAGE_TAG}";     EXTRA="" ;;
+  gr00t)         IMAGE="positro/gr00t:${IMAGE_TAG}";      EXTRA="" ;;
   *)
     echo "Unknown vendor: '$VENDOR'. Supported: lerobot_0_3_3 | lerobot | openpi | gr00t" >&2
     exit 1