v0.26.0:0 - disk-driven model menu (scan sparks; recipes; needs-setup)

The dashboard menu is now the set of models actually downloaded on the
Sparks, not a hard-coded catalog. models.yaml + overrides are reframed as
launch recipes matched to an on-disk model by repo; an on-disk model with
no recipe is flagged needs_setup and its launch settings are inferred from
its config.json for a one-time operator confirmation (discovery.py).

- delete now removes weights AND the menu card (delete_from_disk sweeps all
  hosts; the delete endpoint resolves keys via the live menu)
- new GET /api/models/suggest; /api/models returns the menu + a recipes list
  (download autocomplete); GET /api/models/disk-status removed
- dropped the two legacy Qwen recipes (235B FP8, 2.5 72B)
- tests: +test_discovery.py (cache parsing, infer_recipe, build_menu merge)

AGENTS.md

@@ -33,7 +33,7 @@ Subsystem guidance lives in `docs/guides/` and loads when matching files are tou
 
 - `image/app/` — FastAPI app (`server.py` entry, routers in sibling modules, `static/` dashboard UI).
 - `package/startos/` — StartOS manifest, interfaces, actions, version + release notes.
-- `docs/` — `AUDIO_API.md`, `EMBEDDINGS.md`, `REDACTION_GATEWAY.md` (consumer-facing API refs; update with API changes).
+- `docs/` — `AUDIO_API.md`, `EMBEDDINGS.md`, `REDACTION_GATEWAY.md`, `COORDINATION.md` (consumer-facing API refs; update with API changes).
 - `README.md` (overview), `HANDOFF.md` (fresh-user install guide), `runbook.md` (ops notes), `known-issues.md`, `ROADMAP.md` (longer-term backlog — items move into "Current state" below when picked up).
 
 ## Conventions
@@ -55,12 +55,16 @@ Subsystem guidance lives in `docs/guides/` and loads when matching files are tou
 
 ## Current state
 
-- **Working (v0.21.0:1, installed and serving):** swap dashboard; chat / transcribe / diarize(+chunk) / TTS proxies; embeddings + rerank + hybrid search (Qdrant); `/scrub` + `/rehydrate`; label-merge incl. dual-channel; per-Spark SSH-key copy + WireGuard `VPN <ip>` hardware-card badge. Spark 2 audio stack healthy. Security hardening (v0.19.0:0 — shellsafe SSH-injection guard, Qdrant path-injection, same-origin CSRF guard) shipped and stable; evidence in `EVALUATION.md`.
+- **Built, install pending: v0.26.0:0 — disk-driven model menu.** The dashboard now lists what's *actually downloaded* on the Sparks instead of a hard-coded catalog. `models.yaml` + overrides are reframed as **launch recipes** matched to an on-disk model by `repo` (no longer "the menu"); `image/app/discovery.py` does the merge: `build_menu` scans both Sparks (`disk.list_cached_models`, one `du` per host) ∪ recipes; an on-disk model with no recipe is `needs_setup` and `infer_recipe` reads its `config.json` to prefill a one-time setup form (operator confirms; saved to `/data` overrides). Delete now removes weights **and** the card (`delete_from_disk` sweeps all hosts; the delete endpoint resolves keys via the live menu so discovered models are deletable). New `GET /api/models/suggest`; `/api/models` returns the menu + a `recipes` list (download-box autocomplete); `GET /api/models/disk-status` removed (folded into `/api/models`). Dropped the two legacy Qwen recipes (235B FP8, 2.5 72B). Build/typecheck clean; **install (live-service restart) needs go/no-go.** Why a recipe layer survives a "menu = disk" redesign: a folder can't tell you parsers / solo-vs-cluster / MoE backend (Gemma MoE needs `marlin` on GB10) — disk drives *presence*, recipes drive *launch*.
-- **matrix-bridge bot tile (done, v0.21.0:1, verified live):** `bot`-kind service tile — status badge from docker-state only (no HTTP port), plus **Update** / Restart / Stop/Start / **View logs**. Code: `app/matrix_bridge.py` + `/api/matrix-bridge/{update,logs}` (update streams; 25-min cap; fail-loud). Driven directly as `modelo` on Spark 2 (**no `sudo -iu`** — spark2 has no passwordless sudo). User is a blank-default Configure-Sparks field (`matrix_bridge_user`); blank → tile hidden (portable). Host reuses `spark2_host` (`192.168.1.87` = the bot's box `spark-32d0`); container/dir/branch are env-overridable defaults. **Load-bearing ops dep:** Update's `git fetch` runs as `modelo`, which needs `modelo`'s `~/.ssh/config` pinning the Gitea deploy key with `IdentitiesOnly yes` — else the wrong key is offered and Gitea denies (publickey). Optional next, only if the bot dev asks: Docker `HEALTHCHECK` for running-but-disconnected detection (spec §Note).
+- **Live: v0.25.0:0** (installed 2026-06-18). The OpenClaw/Johnny-5 coexistence epic is fully shipped & live: configurable `VLLM_PORT` (v0.22, blank ⇒ 8888), local/fine-tuned models (v0.23), configurable topology (v0.24 — `VLLM_CONTAINER`, `DISABLED_SERVICES` hide-list, second-Spark `kind: vllm` monitor), coordination layer (v0.25 — swap reservation lock with `423`-enforced manual-swap pause + `?force=true` Release override, `swap_complete`/`swap_failed` webhook, read-only schedule registry; consumer API in `docs/COORDINATION.md`).
-- **Tests:** offline pytest harness in `image/tests/` — `cd image && .venv/bin/python -m pytest` (70 passing). Covers `build_launch_command` (incl. the shell-injection round-trip), the transcript↔diarizer label-merge, the `shellsafe` validators, and `matrix_bridge.build_update_command` (+ phase detection). Mock-heavy swap/proxy tests deliberately skipped (low ROI). Redaction + live-audio suites remain standalone scripts.
+- **Other live features:** swap dashboard; chat / transcribe / diarize(+chunk) / TTS proxies; embeddings + rerank + hybrid search (Qdrant); `/scrub` + `/rehydrate`; label-merge incl. dual-channel; per-Spark SSH-key copy + WireGuard `VPN <ip>` hardware badge. Security hardening (v0.19 — shellsafe SSH-injection guard, Qdrant path-injection, same-origin CSRF guard) stable (`EVALUATION.md`). Spark 2 audio/embeddings stack healthy.
+- **matrix-bridge bot tile (v0.21.0:1, live):** `bot`-kind tile (docker-state badge; Update/Restart/Stop-Start/View-logs) for the Matrix bot on Spark 2, driven as `modelo` (no `sudo -iu`; blank `matrix_bridge_user` ⇒ tile hidden; host reuses `spark2_host`). Code: `app/matrix_bridge.py` + `/api/matrix-bridge/{update,logs}`. **Load-bearing:** Update's `git fetch` runs as `modelo` and needs `modelo`'s `~/.ssh/config` pinning the Gitea deploy key with `IdentitiesOnly yes` (else publickey denial). Optional next only if the bot dev asks: Docker `HEALTHCHECK`.
+- **Tests:** offline pytest harness in `image/tests/` — `cd image && .venv/bin/python -m pytest` (137 passing). Covers `build_launch_command` (incl. the shell-injection round-trip + local-model bind-mount), the transcript↔diarizer label-merge, the `shellsafe` validators, `matrix_bridge.build_update_command` (+ phase detection), the configurable-topology layer (`test_topology.py`), the coordination layer (`test_coordination.py`: swap-lock lifecycle/expiry/token-auth, schedule-registry CRUD, webhook payload + HMAC signature — `now` is injected into the lock so expiry is tested without sleeping), and the disk-driven menu (`test_discovery.py`: cache-dirname↔repo parsing, the cache-listing parser incl. incomplete-download filtering, and `infer_recipe` family/mode mapping — Qwen3-MoE→flashinfer_cutlass, Gemma-MoE→marlin, vision caps, solo-vs-cluster by size/host-count). The `build_menu` merge + `/api/models/suggest` are exercised by hand against the live cluster (mock-heavy unit tests there would test the mocks). Redaction + live-audio suites remain standalone scripts.
 - **Signal Engine "flakiness":** diagnosed as *not* a server bug — transient 1–4s unresponsiveness while the single GPU is busy. Client-side remedy (in-flight cap 2 / ceiling 3 / retry-on-timeout+503) drafted and **forwarded to that dev (owner confirmed 2026-06-15)**. Awaiting whether they want the measured concurrency knee.
 - **Stance (decided, not built):** no public interface / no API-token auth — LAN + WireGuard/Tailscale split-tunnel only; the CSRF guard covers the browser-driven vector.
 - **Known limits:** `/health` blips while the GPU is busy (mitigated client-side); dual-channel can miss a quiet local word under loud remote bleed; connectivity log misses sub-5s outages between 5s polls; diarizer caps at 4 speakers; matrix-bridge badge won't visibly flip on a fast `docker restart` (status re-checked only after the command returns).
 - **Infra gotcha (safety):** passwordless sudo is NOT configured on spark2 — design unprivileged probes for any Spark feature (the badge uses `ip`, not `sudo wg show`). spark2 sits on the `starttunnel` WireGuard subnet (`10.59.211.6/24`, survives reboot). Owner declined SSH-key rotation after the 2026-06-12 history scrub (only the key *name* leaked) — don't re-flag.
 - **Hosting:** self-hosted Gitea — remote `gitea`, branch `master`, over SSH; push after committing. (Wart: commit `8d839e3` is mislabeled `v0.13.0:4` but contains through v0.18.0:0.)
-- **Next — committed 2026-06-17: OpenClaw/Johnny-5 coexistence epic (full plan + design stance in `ROADMAP.md` → "Cluster coordination").** Stance: Spark Control = control plane / GPU arbiter, **not** a job runner; business cron jobs live in separate services that *call* its swap API (swaps are already API-driven via `POST /api/swap`). Sequence: (1) **configurable `VLLM_PORT`** — DONE in tree, staged as **v0.22.0:0** (Configure-Sparks field, blank ⇒ 8888; + `_env_int` hardening in `config.py` so a blank/bad port no longer crashes startup, killing a P3 tech-debt item). **Not yet built/installed/committed — awaiting go/no-go.** (2) local-path/fine-tuned models (in ROADMAP under Dashboard). (3) configurable topology (service→Spark→port map + container names). (4) coordination layer (swap lock + swap webhook + schedule visibility) — only when our own automation lands. Still-open older threads: audio concurrency sweep (only if the Signal Engine dev wants the knee; needs a quiet window); optional matrix-bridge Docker `HEALTHCHECK` if the bot dev asks; Parakeet long-audio guard deferred (rationale in ROADMAP).
+- **Design stance (decided):** Spark Control = control plane / GPU arbiter, **not** a job runner; recurring business jobs live in separate services that *call* the swap API (`POST /api/swap`). Full epic history (v0.22→v0.25) is in git log + `ROADMAP.md` → "Cluster coordination".
+- **Usage note (2026-06-18):** owner's daily driver is the solo **Qwen3.6 35B**; the 235B `cluster` models are dormant. Keeping `launch-cluster.sh` (the `eugr/spark-vllm-docker` community standard, mirrors NVIDIA's `dgx-spark-playbooks` Ray+RoCE design) is still correct even single-node — it supplies the maintained, hardware-tuned vLLM images; raw docker would mean DIY image upkeep for no gain. Spark 2 stays the speech/embeddings box regardless.
+- **Next steps (all low-priority / externally gated; P2/P3 tech-debt backlog in `ROADMAP.md`):** (1) raw-`docker run` swap generalization — **DEFERRED** (rationale in ROADMAP; revisit only if an adopter wants Spark Control to *drive*, not just monitor, raw-docker swaps — cleanest fix is the adopter adopting `launch-cluster.sh`). (2) audio concurrency knee — only if the Signal Engine dev wants it (needs a quiet window). (3) matrix-bridge Docker `HEALTHCHECK` — only if the bot dev asks. (4) Parakeet long-audio guard — deferred (rationale in ROADMAP).

HANDOFF.md

@@ -92,7 +92,7 @@ Now that hosts are configured, Show Public Key will give you the paste-ready ins
 From the Spark Control service page, click the Web UI button. You should see:
 
 - A **top status bar** with the currently loaded LLM (or "no model loaded" if Spark 1's vLLM container is fresh).
-- An **LLM tab** with cards for each model in the bundled catalog. Models you've downloaded show "on disk" badges; others show "not downloaded".
+- An **LLM tab** whose cards are the models actually downloaded on your Sparks (the dashboard scans them on load). A model Spark Control doesn't yet know how to launch shows a "needs setup" card; the first switch reads its files, proposes settings, and asks you to confirm once. Use **+ Download a new model** to fetch one — it appears here when it finishes.
 - An **Audio / Speech tab** with health status and Install / Start / Stop / Restart buttons for Parakeet and Kokoro.
 
 If the dashboard loads and both Spark hardware cards show CPU/RAM/GPU stats, **you're in**.

README.md

@@ -112,14 +112,14 @@ Fields: `service` (required), `ok` (required), `source` (optional, free-form), `
 
 ## Status
 
-**v0.2.3 / s9pk version 0.13.0:4** — installed and verified on a Start9 server. Five bundled LLMs in the catalog (qwen3-vl, gemma4, qwen36, qwen3-235b-fp8, qwen2.5-72b), plus any custom models added through the UI.
+**s9pk version 0.26.0:0** — installed and verified on a Start9 server. The LLM menu is whatever's downloaded on the Sparks (scanned live, not hard-coded); bundled *launch recipes* (qwen3-vl, gemma4, gemma4-26b, qwen36) tell it how to launch known models, and anything else gets a "needs setup" card that infers + saves its settings on first use.
 
 ### What v0.2 added on top of v0.1
 
 - **Service discovery API** (`/api/endpoints`) for other LAN services
 - **Kokoro-82M TTS** replaces Magpie/Riva NIM as the default TTS backend (v0.14.0). Magpie's decoder had a ~30-50% truncation rate on multi-sentence inputs and ate 49 GB of GPU memory; Kokoro is 24/24 reliable at every input length tested, uses 1.3 GB GPU, and renders in ~1s. See HANDOFF.md and the release notes for the migration story.
 - **Always-on services panel** with Start/Stop/Restart for Parakeet + Kokoro, plus per-service host configuration in Configure Sparks (so they can live on Spark 1, Spark 2, or anywhere)
-- **Model download** from the dashboard — paste an HF repo, pick solo or cluster, watch percent progress with bytes/rate/ETA. After completion, an "Add to catalog" dialog appears pre-filled.
+- **Model download** from the dashboard — paste an HF repo (with autocomplete for known models), pick solo or cluster, watch percent progress with bytes/rate/ETA. After completion the model appears on the menu automatically; if it's unrecognized, a pre-filled "set up this model" dialog offers to configure it.
 - **spark-vllm-docker update check** — banner shows "N commits behind upstream"; Apply Update runs `git pull && ./build-and-copy.sh -c` over SSH with a streamed log
 - **Per-model Advanced settings** — knobs for max context, GPU memory %, and three optimization toggles (fastsafetensors, prefix caching, FP8 KV cache). Persisted to `/data/models-overrides.yaml` so they survive package updates. Bundled and custom models alike.
 - **Diarization with speaker fingerprints** via Sortformer + TitaNet, exposed at `/api/audio/diarize-chunk` for chunked workflows

ROADMAP.md

@@ -10,12 +10,25 @@ Driven by the one other Spark Control adopter (a colleague running OpenClaw + cr
 
 Sequenced:
 1. **Configurable `VLLM_PORT`** — DONE, v0.22.0:0. Field in Configure Sparks (blank ⇒ 8888); numeric-setting parsing hardened so a blank/bad value falls back instead of crashing startup. Was the immediate "vLLM unreachable" bug for an adopter on port 8000.
-2. **Local-path / fine-tuned model support** — see the dedicated item under "## Dashboard" below. Independently wanted; his merged `ten31-v2` (a directory, not an HF repo) is the motivating case.
+2. **Local-path / fine-tuned model support** — DONE, v0.23.0:0. Catalog/`ModelDef` gained `local_path` (exactly one of `repo`/`local_path`); swap bind-mounts the dir into the vLLM container at the same path via the launch script's `VLLM_SPARK_EXTRA_DOCKER_ARGS` hook (no `launch-cluster.sh` change); "+ Add local model" form + `local` badge; disk-delete refused for local models; `validate_local_path` boundary check. His merged `ten31-v2` was the motivating case.
-3. **Configurable topology** — make the service→Spark→port map and container names configurable so the package stops assuming our exact layout. Lets an adopter monitor vLLM on *both* Sparks, use a different container name, and stop the Parakeet probe from hitting a vLLM that shares its port — without forking. (Covers report P4 multi-Spark vLLM, P5 container name, and the Parakeet-port collision #6.)
+3. **Configurable topology** — DONE, v0.24.0:0. Three optional Configure-Sparks knobs: vLLM container name (`VLLM_CONTAINER`, blank ⇒ `vllm_node`; threaded through the swap log-tail + pre-flight validator via `quote_arg`); "services to hide" (`DISABLED_SERVICES`, comma list — hidden services show no tile and are skipped by status/deep-health/connectivity probes, killing the Parakeet-on-8000 collision); and a second-Spark vLLM monitor via a `kind: vllm` custom service in `services-overrides.yaml` (read-only tile probed through the shared `probe_vllm_endpoint`). `/api/endpoints` gained a `disabled` flag. Covers report P4/P5/#6. (Generalizing the *swap* mechanism to the adopter's raw `docker run` was deliberately left out — that's coordination, item 4; he swaps via his own crons and uses Spark Control to monitor.)
-4. **Coordination layer** — build when our own automation actually lands (zero value until something other than the dashboard swaps models):
+4. **Coordination layer** — DONE in tree, staged as **v0.25.0:0** (built/typechecked clean; install pending). All three primitives shipped; `image/app/coordination.py` + `docs/COORDINATION.md`. Brought forward 2026-06-17 on request rather than waiting for our own automation.
-   - **Swap lock** with holder + TTL (`POST` / `GET` / `DELETE /api/swap/lock`). An external scheduler acquires it before swapping; the dashboard then refuses manual swaps and shows who holds the GPU and until when. Enforced by the swap path, not advisory.
+   - **Swap lock** with holder + TTL (`POST` / `GET` / `DELETE /api/swap/lock`). Acquire returns a secret token; the swap endpoint refuses any real swap (`423`) that doesn't present it in `X-Swap-Lock-Token`, so the dashboard's manual swap is paused while a scheduler holds it (with a `?force=true` human override). In-memory + TTL-bounded → resets to unlocked on restart; re-acquire with the token extends. Enforced in `post_swap`, not advisory.
-   - **Swap-event webhook** (`swap_complete` / `swap_failed`) to a configurable URL, so downstream consumers update their provider config when the running model changes.
+   - **Swap-event webhook** (`swap_complete` / `swap_failed`) to a configurable URL (Configure-Sparks field), fired from `SwapManager._run` *outside* the swap lock; optional shared secret ⇒ `X-Spark-Signature` HMAC. Fire-and-forget (5 s, no retries); dry runs don't fire.
-   - **Schedule visibility** — read-only view the dashboard surfaces, *registered by* external schedulers (Spark Control does not own the schedule).
+   - **Schedule visibility** — `GET/POST/DELETE /api/schedule`; read-only "Scheduled jobs" dashboard panel, registered by external schedulers. Spark Control stores and displays, never executes.
+   - Tests: `image/tests/test_coordination.py` (22 cases — lock lifecycle/expiry/token, the single-read swap gate, schedule CRUD + id validation, webhook payload+signature). Known limit: lock + schedules are in-memory (a restart frees the lock and empties the registry until schedulers re-register) — persist to `/data` only if that bites.
+
+### Generalizing the swap mechanism to raw `docker run` — DEFERRED (decided 2026-06-18, research-backed; was item 4's last open thread)
+
+Our swap drives `~/spark-vllm-docker/launch-cluster.sh` over SSH on Spark 1 (`./launch-cluster.sh stop`, then `[VLLM_SPARK_EXTRA_DOCKER_ARGS=…] ./launch-cluster.sh [--solo ]-d exec vllm serve <model> <args>`, then `docker logs -f` until the ready marker). The OpenClaw adopter launches vLLM with a plain `docker run` instead, so the swap button can't drive his cluster — only monitor it. The portability fix would be a configurable "swap backend": keep `launch-cluster.sh` as the default and add a "bring your own command" mode (operator-authored stop/launch templates in `services-overrides.yaml` with quoted `{model}`/`{container}`/`{port}`/`{extra_args}` substitution; ready-detection unchanged; the vLLM-argparse pre-flight disabled for that backend).
+
+**Why deferred, not built:**
+- **Raw docker is not an upgrade for *us* — for half our catalog it's impossible.** `launch-cluster.sh` is the `eugr/spark-vllm-docker` community project (de-facto DGX Spark standard; mirrors NVIDIA's own `dgx-spark-playbooks` Ray+RDMA architecture). Its headline job is **multi-node** serving: our 235B `cluster` models (Qwen3-VL 235B, Qwen3 235B) exceed one Spark's 128 GB and *must* shard across both Sparks via Ray over the 200 Gbps ConnectX/RoCE link — plumbing (NCCL/MTU/per-node env) that a single-node `docker run` cannot do. So we keep the helper script; switching our own cluster to raw docker is off the table.
+- **The feature is therefore portability-only** (for differently-wired adopters), and the one known adopter doesn't need it — he swaps via his own crons and uses Spark Control to watch.
+- **Untestable on our hardware** — our cluster uses the helper script, so we can't validate a real raw-docker swap without risking the live vLLM.
+- The one real standing risk is eugr's single-maintainer status; fallback is community forks or migrating to NVIDIA's official `dgx-spark-playbooks` launcher (same design). No reason to switch now.
+
+**Revisit only if** an adopter explicitly wants Spark Control to *drive* (not just monitor) swaps on a raw-`docker run` cluster. At that point, get their actual working `docker run` command and build the command-template backend to it.
 
 ## Near term
 - parakeet-asr long-audio memory guard — **deferred 2026-06-15, low priority.** A duration cap on `/v1/audio/diarize`: Sortformer runs the whole file in one pass (`diarizer.py:128-135`) over Spark 2's *shared* 128 GB unified memory (also feeding Kokoro/embeddings/Qdrant), so one giant single file can thrash into swap. **Precautionary — no observed incident**, and the production consumer (Recap Relay) already chunks via `/diarize-chunk` (~5-min, already bounded), so the only exposed path is a consumer POSTing one huge file to the full `/diarize`. When picked up: add a configurable `MAX_DIARIZE_SECONDS` guard in `diarizer.py` right after `duration` is computed (~line 130) → raise → HTTP 413 in `main.py` (mirrors the existing `MAX_UPLOAD_MB` 413); ship via the Reapply-patches action (restarts the live parakeet-asr container → needs go/no-go). Leave transcription out of v1 (upstream/un-patched file; parakeet-TDT handles long audio better). Revisit only if a consumer starts sending long single files.
@@ -34,7 +47,6 @@ Sequenced:
 - Second audio worker / queueing layer; revisit which services share Spark 2.
 
 ## Dashboard
-- Support local-path / fine-tuned models in the swap catalog. Today the catalog is static (`models.yaml` + custom overrides) and the "Add custom model" path (`POST /api/models`) only accepts an HF `org/name` repo (`shellsafe._HF_REPO_RE`), so a model that exists only as a directory on a Spark (the usual fine-tuning output) can't be registered or swapped. Needs: (a) a "local model" add form/field taking a Spark-side directory path, with its own safe validation instead of the `org/name` regex (path whitelist + `shlex.quote`, no traversal); (b) `models.build_launch_command` / `launch-cluster.sh` able to `vllm serve <path>`; (c) `disk.py` size-probe handling a path instead of deriving the HF cache dir from a repo id. Raised 2026-06-15 — a colleague's locally fine-tuned model doesn't appear because nothing scans the machine; the list is a curated catalog, not a discovery probe.
 - Per-model configurable vLLM flags editable from the UI (today: edit `models.yaml` and rebuild).
 - Spark host update actions (OS/driver) from the UI.
 - Open WebUI link-out integration; richer per-service detail views.

docs/COORDINATION.md

@@ -0,0 +1,157 @@
+# Cluster coordination through Spark Control (v0.25.0)
+
+Spark Control is the **GPU arbiter, not a job runner.** Your recurring pipelines
+(model-warming crons, "daily X" generators, batch jobs) live in your own
+services and *drive Spark Control's swap API*. This page documents the safety
+layer around that: a **swap reservation lock**, a **swap-event webhook**, and a
+**read-only schedule registry**.
+
+If only the dashboard ever swaps models, you don't need any of this — it's for
+when something automated also swaps.
+
+All endpoints are on the Spark Control host (same LAN/VPN URL as the LLM, audio,
+and embeddings proxies). There is no API-token auth by design (LAN + split-tunnel
+VPN only); a non-browser client passes the same-origin guard automatically.
+
+---
+
+## 1. Swap reservation lock
+
+A short, TTL-bounded reservation of the swap path. While a lock is held, **any
+real swap that doesn't present the holder's token is refused with `423 Locked`**
+— including the dashboard's manual swap. The holder *name* is descriptive; the
+returned **token** is the secret that authorises swaps and the release.
+
+The lock is in-memory: it resets to *unlocked* if Spark Control restarts (the
+safe-for-availability default), and the swap engine's own in-progress guard
+still prevents two swaps running at once.
+
+### `POST /api/swap/lock` — acquire (or extend)
+
+```json
+// request
+{ "holder": "openclaw-daily-vol", "ttl_seconds": 900, "note": "daily vol run" }
+
+// 200 response
+{
+  "held": true,
+  "holder": "openclaw-daily-vol",
+  "acquired_at": "2026-06-17T12:00:00+00:00",
+  "expires_at":  "2026-06-17T12:15:00+00:00",
+  "seconds_remaining": 900,
+  "note": "daily vol run",
+  "token": "a1b2c3…"        // SECRET — store it; needed to swap and to release
+}
+```
+
+- `ttl_seconds` is optional (default 900) and clamped to `[1, 86400]`.
+- **`409`** if a *different* holder already holds it (body includes the current
+  `lock` state). To **extend** your own lock, POST again with the same `holder`
+  **and** your `token` — the token is preserved and the window slides forward.
+
+### `GET /api/swap/lock` — status (no token)
+
+```json
+{ "held": true, "holder": "openclaw-daily-vol", "expires_at": "…", "seconds_remaining": 612, "note": "…" }
+// or
+{ "held": false }
+```
+
+### `DELETE /api/swap/lock` — release
+
+Send your token in the `X-Swap-Lock-Token` header (or `?token=`):
+
+```
+DELETE /api/swap/lock
+X-Swap-Lock-Token: a1b2c3…
+```
+
+- **`403`** if the token doesn't match. The dashboard's human override is
+  `DELETE /api/swap/lock?force=true` (no token).
+
+### Swapping while you hold the lock
+
+Pass the token on the swap call; the dashboard (no token) is then blocked:
+
+```
+POST /api/swap
+X-Swap-Lock-Token: a1b2c3…
+{ "model_key": "gemma-3-27b" }
+```
+
+Recommended scheduler flow: **acquire → swap (with token) → poll `/api/swap/{id}`
+→ release**. Always release in a `finally`; if you crash, the TTL frees it.
+
+> `POST /api/swap/{key}/validate` (pre-flight) and dry-run swaps are **not**
+> blocked by the lock — they don't touch the cluster.
+
+---
+
+## 2. Swap-event webhook
+
+Configure a URL in **Configure Sparks → "Swap webhook URL"**. After every real
+swap, Spark Control POSTs:
+
+```json
+{
+  "event": "swap_complete",   // or "swap_failed"
+  "job_id": "1a2b3c4d",
+  "model_key": "gemma-3-27b",
+  "state": "ready",           // or "failed"
+  "returncode": 0,
+  "started_at": "2026-06-17T12:00:00+00:00",
+  "finished_at": "2026-06-17T12:03:11+00:00",
+  "dry_run": false
+}
+```
+
+Headers: `X-Spark-Event: swap_complete`. If you set a **webhook secret**, the
+body is signed: `X-Spark-Signature: sha256=<hmac>` (HMAC-SHA256 of the raw body
+with the shared secret). Verify it like:
+
+```python
+import hmac, hashlib
+expected = "sha256=" + hmac.new(secret.encode(), raw_body, hashlib.sha256).hexdigest()
+assert hmac.compare_digest(expected, request.headers["X-Spark-Signature"])
+```
+
+Delivery is best-effort and fire-and-forget (5 s timeout, no retries) — a
+webhook failure never affects the swap itself. Dry runs don't fire.
+
+---
+
+## 3. Schedule registry (read-only display)
+
+So the dashboard can show *what's scheduled to touch the GPU and when*, your
+schedulers register their jobs here. **Spark Control only displays these — it
+never executes them.**
+
+### `POST /api/schedule` — register / update
+
+```json
+// request (pass a stable `id` to update in place on re-register)
+{ "id": "daily-vol", "name": "Daily Vol", "owner": "openclaw",
+  "cron": "0 6 * * *", "next_run": "2026-06-18T06:00:00Z",
+  "description": "Swaps to the big model, generates the vol report" }
+
+// response: the stored entry (generates an id if you omit one)
+```
+
+`name` is required; `id` (if given) must match `[A-Za-z0-9_.-]` (≤64 chars).
+
+### `GET /api/schedule` — list
+
+```json
+{ "schedules": [ { "id": "daily-vol", "name": "Daily Vol", "owner": "openclaw",
+  "cron": "0 6 * * *", "next_run": "…", "description": "…",
+  "registered_at": "…", "updated_at": "…" } ] }
+```
+
+### `DELETE /api/schedule/{id}` — deregister
+
+```json
+{ "deleted": true }
+```
+
+The registry is in-memory — re-register your schedules on your own startup so
+they survive a Spark Control restart.

docs/guides/fastapi-image.md

@@ -39,6 +39,7 @@ Two kinds, both run with the `image/.venv` interpreter (system python3 has no de
 ## Layout
 
 - `image/app/server.py` — FastAPI entry; routers live in sibling modules (`audio_proxy.py`, `llm_proxy.py`, `embeddings_proxy.py`, `redaction_gateway.py`, `swap.py`, `health.py`, `deep_health.py`, `connectivity.py`, …).
+- `image/app/discovery.py` — the disk-driven model menu. `/api/models` lists what's actually downloaded on the Sparks (via `disk.list_cached_models`); `models.yaml`/overrides are *launch recipes* matched by repo, not the menu. An on-disk model with no recipe is `needs_setup` → `infer_recipe` reads its `config.json` to prefill a setup form the operator confirms once.
 - `image/app/static/` — the dashboard UI.
-- `image/models.yaml` — vLLM model catalog bundled into the image.
+- `image/models.yaml` — bundled vLLM **launch recipes** (how to launch a known model), NOT the dashboard menu — the menu is the on-disk scan.
 - `image/spark_embed/` — Dockerfile + app for the embeddings container; built ON a Spark (ARM64, NGC PyTorch base — see the audio/cluster rule for NGC torch-pinning caveats).

docs/guides/startos-package.md

@@ -25,6 +25,22 @@ npm run prettier   # prettier --write startos (no semicolons, single quotes, tra
 - Version format is `X.Y.Z:N` (`:N` = revision). Bump in `package/startos/versions/v0_1_0.ts`; **replace** the release notes — never leave old notes behind under an extra key (any unknown key fails `tsc`).
 - New external-facing endpoints get noted in release notes for downstream app developers (Recap Relay, Ten31 Transcripts, CRM, Signal Engine consume these APIs).
 
+## Releasing to Gitea
+
+The s9pk is distributed via Gitea **Releases** (the binary is gitignored — never commit it). Adopters pull the latest asset with a read-only token. Per-version ritual:
+
+```bash
+# 1. bump version in startos/versions/v0_1_0.ts (+ replace release notes), then:
+cd package && make x86                       # build
+# 2. commit + push the source change
+git tag vX.Y.Z && git push gitea vX.Y.Z      # tag — plain vX.Y.Z, NO ':' (git refs forbid it)
+make install                                 # optional: sideload to your own server (restarts it — go/no-go)
+# 3. publish the s9pk as a release asset (needs a write-scoped token):
+GITEA_URL=https://<gitea-host> GITEA_TOKEN=<write-token> make release
+```
+
+`make release` → `scripts/gitea-release.sh`: creates/reuses the release for the tag and uploads (replacing) the s9pk asset; idempotent, fails loud on real HTTP errors. `GITEA_INSECURE=1` skips TLS verify for a self-signed LAN cert. Hand adopters a **read-only** token (repository: Read), ideally on a dedicated reader account; their agent then `GET`s `/api/v1/repos/<owner>/spark-control/releases/latest` and downloads the `.s9pk` asset. Note Gitea returns `browser_download_url` on its configured ROOT_URL (may be a `.local` name) — an off-LAN adopter pulls via whatever address actually reaches the Gitea.
+
 ## Layout
 
 - `package/startos/` — manifest, interfaces, actions (`configureSparks`, `showPublicKey`), `versions/v0_1_0.ts` (current version string + release notes).

image/app/config.py

@@ -1,13 +1,44 @@
 from __future__ import annotations
+import logging
 import os
 from dataclasses import dataclass
 from pathlib import Path
 
+from .shellsafe import validate_container
+
+log = logging.getLogger(__name__)
+
 
 def _env(name: str, default: str = "") -> str:
     return os.environ.get(name, default)
 
 
+def _env_container(name: str, default: str) -> str:
+    """Resolve a container-name env var, validating it at the config boundary.
+
+    The value flows into `docker logs`/`docker exec` over SSH, so it's quoted at
+    the sink — but per the repo's two-layer convention it's also whitelist-checked
+    here. A malformed optional value falls back to `default` rather than crashing
+    daemon startup (mirrors `_env_int` for VLLM_PORT)."""
+    val = os.environ.get(name, "") or default
+    try:
+        return validate_container(val)
+    except ValueError:
+        log.warning("ignoring invalid %s=%r; using %r", name, val, default)
+        return default
+
+
+def _env_set(name: str) -> frozenset[str]:
+    """Parse a comma-separated env var into a lowercased frozenset of keys.
+
+    Used by DISABLED_SERVICES so an adopter whose cluster doesn't run a given
+    support service can switch its tile + probes off entirely (rather than have
+    the probe hit whatever else listens on that port — e.g. a vLLM sharing
+    Parakeet's default 8000)."""
+    raw = os.environ.get(name, "")
+    return frozenset(part.strip().lower() for part in raw.split(",") if part.strip())
+
+
 def _env_int(name: str, default: int) -> int:
     """Parse an int env var, falling back to `default` when unset, blank, or
     malformed. The StartOS Configure panel passes optional numeric fields as an
@@ -63,6 +94,8 @@ class Settings:
     ssh_known_hosts: str
     models_yaml: str
     vllm_port: int
+    vllm_container: str
+    disabled_services: frozenset[str]
     parakeet_port: int
     kokoro_port: int
     embed_port: int
@@ -70,6 +103,8 @@ class Settings:
     bind_port: int
     open_webui_url: str
     ngc_api_key: str
+    swap_webhook_url: str
+    swap_webhook_secret: str
 
     @classmethod
     def from_env(cls) -> "Settings":
@@ -116,6 +151,15 @@ class Settings:
             ssh_known_hosts=_env("SSH_KNOWN_HOSTS"),
             models_yaml=_resolve_models_yaml(),
             vllm_port=_env_int("VLLM_PORT", 8888),
+            # Container name for the swappable vLLM on Spark 1. Defaults to the
+            # bundled launch-cluster.sh container; override if you named yours
+            # something else (the swap log-tail and pre-flight validator exec
+            # into it by name).
+            vllm_container=_env_container("VLLM_CONTAINER", "vllm_node"),
+            # Built-in support-service keys (parakeet, kokoro, embeddings,
+            # qdrant) the deployment doesn't run — hidden from the dashboard and
+            # never probed.
+            disabled_services=_env_set("DISABLED_SERVICES"),
             parakeet_port=_env_int("PARAKEET_PORT", 8000),
             kokoro_port=_env_int("KOKORO_PORT", 8880),
             embed_port=_env_int("EMBED_PORT", 8088),
@@ -123,6 +167,11 @@ class Settings:
             bind_port=_env_int("BIND_PORT", 9999),
             open_webui_url=_env("OPEN_WEBUI_URL", ""),
             ngc_api_key=_env("NGC_API_KEY", ""),
+            # Coordination layer: fire a swap-lifecycle webhook to this URL so
+            # downstream consumers re-point their model config on a swap. Blank
+            # ⇒ disabled. The optional secret HMAC-signs the body (X-Spark-Signature).
+            swap_webhook_url=_env("SWAP_WEBHOOK_URL", ""),
+            swap_webhook_secret=_env("SWAP_WEBHOOK_SECRET", ""),
         )
 
     @property

image/app/coordination.py

@@ -0,0 +1,342 @@
+"""Cluster-coordination layer: the GPU swap lock, swap-event webhook, and the
+read-only schedule registry.
+
+Spark Control is the **control plane / GPU arbiter, not a job runner.** Recurring
+business pipelines live in separate services that *call* the swap API. These
+three primitives add the *safety* layer around that:
+
+- **Swap lock** — a TTL-bounded reservation of the swap path. An external
+  scheduler acquires it before swapping; while held by someone else the
+  dashboard's manual swap is refused (enforced in the swap endpoint, not
+  advisory). Holder name is descriptive; the returned token is the secret that
+  authorises a swap or a release.
+- **Webhook** — fires `swap_complete` / `swap_failed` to a configurable URL so
+  downstream consumers re-point their provider config when the running model
+  changes. Optionally HMAC-signed.
+- **Schedule registry** — a read-only view the dashboard surfaces, *registered
+  by* external schedulers. Spark Control stores what it's told; it does not own
+  or execute any schedule.
+
+All state is in-memory (mirroring the swap/download/NIM job managers). On a
+restart the lock resets to *unlocked* — the available-by-default failure mode;
+the swap manager's own in-progress guard still prevents two swaps at once —
+and schedulers re-register their schedules.
+"""
+from __future__ import annotations
+import hashlib
+import hmac
+import json
+import logging
+import re
+import uuid
+from dataclasses import dataclass
+from datetime import datetime, timedelta, timezone
+from typing import Optional
+
+import httpx
+
+log = logging.getLogger(__name__)
+
+# A lock reserves the GPU for a window; clamp the TTL so a buggy client can
+# neither pin the cluster forever nor take a zero-length (useless) lock.
+LOCK_TTL_MIN = 1
+LOCK_TTL_MAX = 86_400      # 24h
+LOCK_TTL_DEFAULT = 900     # 15 min
+
+# Schedule ids are reflected to the dashboard and used as a URL path segment on
+# delete, so a caller-supplied id is whitelist-checked. Generated ids are hex.
+_SCHEDULE_ID_RE = re.compile(r"^[A-Za-z0-9_.-]{1,64}$")
+
+
+def valid_schedule_id(value: str) -> bool:
+    """Whitelist check for a caller-supplied schedule id (register and delete)."""
+    return bool(_SCHEDULE_ID_RE.match(value or ""))
+
+
+def _now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def _iso(dt: datetime) -> str:
+    return dt.isoformat()
+
+
+# ---------------------------------------------------------------- swap lock ----
+
+class LockHeld(Exception):
+    """The lock is held by a different holder. Carries the public lock state so
+    the endpoint can return holder + expiry in the 409 body."""
+
+    def __init__(self, state: dict) -> None:
+        self.state = state
+        super().__init__("swap lock is held by another holder")
+
+
+@dataclass
+class LockState:
+    holder: str
+    token: str
+    acquired_at: datetime
+    expires_at: datetime
+    note: str = ""
+
+    def public(self, now: datetime) -> dict:
+        """Token-free view safe to expose on GET / in error bodies."""
+        return {
+            "held": True,
+            "holder": self.holder,
+            "acquired_at": _iso(self.acquired_at),
+            "expires_at": _iso(self.expires_at),
+            "seconds_remaining": max(0, int((self.expires_at - now).total_seconds())),
+            "note": self.note,
+        }
+
+
+class SwapLockManager:
+    """In-memory, TTL-bounded reservation of the GPU swap path.
+
+    `now` is injectable on every method purely so the expiry logic is testable
+    without sleeping; production calls omit it and get wall-clock UTC.
+    """
+
+    def __init__(self) -> None:
+        self._lock: Optional[LockState] = None
+
+    def _active(self, now: Optional[datetime] = None) -> Optional[LockState]:
+        """The current lock if one is held and unexpired; lazily clears an
+        expired lock so it never lingers."""
+        now = now or _now()
+        if self._lock is not None and self._lock.expires_at <= now:
+            self._lock = None
+        return self._lock
+
+    def status(self, now: Optional[datetime] = None) -> dict:
+        now = now or _now()
+        active = self._active(now)
+        return active.public(now) if active else {"held": False}
+
+    def acquire(
+        self,
+        holder: str,
+        ttl_seconds: Optional[int] = None,
+        note: str = "",
+        token: Optional[str] = None,
+        *,
+        now: Optional[datetime] = None,
+    ) -> LockState:
+        """Acquire a free lock (new token), or extend one already held by
+        presenting its token. A request without the token is refused even if the
+        holder name matches — the name is descriptive, the token is the secret.
+        """
+        now = now or _now()
+        holder = (holder or "").strip()
+        if not holder:
+            raise ValueError("holder is required")
+        ttl = ttl_seconds if ttl_seconds is not None else LOCK_TTL_DEFAULT
+        try:
+            ttl = int(ttl)
+        except (TypeError, ValueError):
+            ttl = LOCK_TTL_DEFAULT
+        ttl = max(LOCK_TTL_MIN, min(LOCK_TTL_MAX, ttl))
+
+        active = self._active(now)
+        if active is not None:
+            # Held — only the token-holder may extend/re-acquire.
+            if not (token and hmac.compare_digest(active.token, token)):
+                raise LockHeld(active.public(now))
+            self._lock = LockState(
+                holder=holder or active.holder,
+                token=active.token,
+                acquired_at=active.acquired_at,
+                expires_at=now + timedelta(seconds=ttl),
+                note=note or active.note,
+            )
+            return self._lock
+
+        self._lock = LockState(
+            holder=holder,
+            token=uuid.uuid4().hex,
+            acquired_at=now,
+            expires_at=now + timedelta(seconds=ttl),
+            note=note,
+        )
+        return self._lock
+
+    def verify(self, token: Optional[str], now: Optional[datetime] = None) -> bool:
+        """True iff `token` matches the currently-active lock."""
+        active = self._active(now)
+        return bool(active and token and hmac.compare_digest(active.token, token))
+
+    def is_blocked_by(self, token: Optional[str], now: Optional[datetime] = None) -> Optional[dict]:
+        """Single-read swap gate. Returns the public lock state if an active
+        lock blocks a swap carrying this token, else None. Does exactly one
+        `_active()` read so the decision can't straddle a TTL expiry the way a
+        separate status()+verify() pair could (which, at the expiry tick, would
+        spuriously refuse a swap that should now be allowed)."""
+        now = now or _now()
+        active = self._active(now)
+        if active is None:
+            return None
+        if token and hmac.compare_digest(active.token, token):
+            return None
+        return active.public(now)
+
+    def release(
+        self,
+        token: Optional[str] = None,
+        *,
+        force: bool = False,
+        now: Optional[datetime] = None,
+    ) -> bool:
+        """Release the lock. Returns False if nothing was held. Requires the
+        matching token unless `force` (the human override from the dashboard)."""
+        active = self._active(now)
+        if active is None:
+            return False
+        if not force and not self.verify(token, now):
+            raise PermissionError("token does not hold the lock")
+        self._lock = None
+        return True
+
+
+# ----------------------------------------------------------------- webhook ----
+
+def build_webhook_payload(
+    *,
+    event: str,
+    job_id: str,
+    model_key: str,
+    state: str,
+    returncode: Optional[int],
+    started_at: Optional[str],
+    finished_at: Optional[str],
+    dry_run: bool,
+) -> dict:
+    return {
+        "event": event,            # swap_complete | swap_failed
+        "job_id": job_id,
+        "model_key": model_key,
+        "state": state,
+        "returncode": returncode,
+        "started_at": started_at,
+        "finished_at": finished_at,
+        "dry_run": dry_run,
+    }
+
+
+def sign_payload(secret: str, body: bytes) -> str:
+    """`X-Spark-Signature` value: sha256 HMAC of the exact JSON body the
+    consumer receives, so they can recompute and trust it."""
+    return "sha256=" + hmac.new(secret.encode(), body, hashlib.sha256).hexdigest()
+
+
+class WebhookNotifier:
+    """Fire-and-forget POST of swap-lifecycle events. A webhook failure is
+    logged and swallowed — it must never affect the swap outcome."""
+
+    def __init__(self, url: str, secret: str = "", timeout: float = 5.0) -> None:
+        self.url = (url or "").strip()
+        self.secret = secret or ""
+        self.timeout = timeout
+
+    @property
+    def enabled(self) -> bool:
+        return bool(self.url)
+
+    async def fire(self, event: str, payload: dict) -> None:
+        if not self.enabled:
+            return
+        body = json.dumps(payload).encode()
+        headers = {
+            "content-type": "application/json",
+            "user-agent": "spark-control-webhook",
+            "x-spark-event": event,
+        }
+        if self.secret:
+            headers["x-spark-signature"] = sign_payload(self.secret, body)
+        try:
+            async with httpx.AsyncClient(timeout=self.timeout) as client:
+                await client.post(self.url, content=body, headers=headers)
+        except Exception as e:  # noqa: BLE001 — best-effort, never propagate
+            log.warning("swap webhook to %s failed: %s", self.url, e)
+
+
+# -------------------------------------------------------- schedule registry ----
+
+@dataclass
+class ScheduleEntry:
+    id: str
+    name: str
+    owner: str = ""
+    cron: str = ""
+    next_run: str = ""
+    description: str = ""
+    registered_at: str = ""
+    updated_at: str = ""
+
+    def public(self) -> dict:
+        return {
+            "id": self.id,
+            "name": self.name,
+            "owner": self.owner,
+            "cron": self.cron,
+            "next_run": self.next_run,
+            "description": self.description,
+            "registered_at": self.registered_at,
+            "updated_at": self.updated_at,
+        }
+
+
+class ScheduleRegistry:
+    """What external schedulers tell us about their cron jobs. Read-only from the
+    dashboard's side; Spark Control never executes any of it."""
+
+    def __init__(self) -> None:
+        self._items: dict[str, ScheduleEntry] = {}
+
+    def list(self) -> list[dict]:
+        return [e.public() for e in self._items.values()]
+
+    def register(
+        self,
+        *,
+        name: str,
+        id: Optional[str] = None,
+        owner: str = "",
+        cron: str = "",
+        next_run: str = "",
+        description: str = "",
+    ) -> ScheduleEntry:
+        name = (name or "").strip()
+        if not name:
+            raise ValueError("name is required")
+        if id is not None:
+            id = id.strip()
+            if id and not valid_schedule_id(id):
+                raise ValueError("id must match [A-Za-z0-9_.-] (max 64 chars)")
+        ts = _iso(_now())
+        existing = self._items.get(id) if id else None
+        if existing is not None:
+            existing.name = name
+            existing.owner = owner.strip()
+            existing.cron = cron
+            existing.next_run = next_run
+            existing.description = description
+            existing.updated_at = ts
+            return existing
+        sid = id or uuid.uuid4().hex[:8]
+        entry = ScheduleEntry(
+            id=sid,
+            name=name,
+            owner=owner.strip(),
+            cron=cron,
+            next_run=next_run,
+            description=description,
+            registered_at=ts,
+            updated_at=ts,
+        )
+        self._items[sid] = entry
+        return entry
+
+    def delete(self, schedule_id: str) -> bool:
+        return self._items.pop(schedule_id, None) is not None

image/app/custom_services.py

@@ -10,6 +10,17 @@ Format:
         port: 8001
         health_path: /health
         image: nvcr.io/nim/nvidia/riva-multilingual:latest
+
+A `kind: vllm` entry monitors an additional vLLM on another Spark (read-only —
+the swap machinery only drives the primary Spark 1 vLLM). It gets a health tile
+probed via /v1/models plus container state and start/stop/restart:
+    custom:
+      - key: vllm-spark2
+        kind: vllm
+        host: <spark-2-ip>
+        user: <ssh-user>
+        container: vllm_node
+        port: 8000
 """
 from __future__ import annotations
 import os

image/app/deep_health.py

@@ -377,6 +377,10 @@ class DeepHealth:
     async def run_all(self) -> dict[str, ProbeResult]:
         results = {}
         for name in self.PROBES:
+            # Don't deep-probe a service the deployment switched off — its port
+            # may be answered by something else (e.g. a vLLM on Parakeet's 8000).
+            if name in self.settings.disabled_services:
+                continue
             results[name] = await self.run_one(name)
         return results
 

image/app/discovery.py

@@ -0,0 +1,209 @@
+"""Disk-driven model menu + launch-recipe inference.
+
+The dashboard's model list is whatever is actually downloaded on the Sparks
+(see `disk.list_cached_models`), NOT a hard-coded catalog. The bundled/overridden
+catalog entries are *launch recipes*: matched to an on-disk model by repo, they
+say HOW to launch it. A completed model on disk with no matching recipe shows up
+as `needs_setup` — the first switch reads its `config.json`, proposes a recipe
+(`infer_recipe`) the operator confirms once, and that confirmed recipe is saved
+to /data so it's a normal card from then on.
+
+Why a recipe layer at all, if the menu is the disk? Because a folder on disk
+doesn't say how to launch it: the per-family parsers (`--reasoning-parser`,
+`--tool-call-parser`), the MoE backend (some Gemma MoE checkpoints need
+`marlin` on GB10), and solo-vs-cluster topology can't be read off a directory.
+We infer a best guess from the model's own config + size, but the operator
+confirms it — a wrong guess is cheap, a wrong launch is not.
+"""
+from __future__ import annotations
+import asyncio
+import re
+
+from .config import Settings
+from .disk import list_cached_models, probe_disk
+from .overrides import extract_knobs_from_args
+
+
+# A model whose weights exceed this can't fit one Spark's 128 GB beside a KV
+# cache, so it must shard across both via Ray. A heuristic prefill only — the
+# operator confirms mode in the setup form, so the exact cutoff isn't critical.
+SINGLE_SPARK_BYTES = 115 * 1000 ** 3
+
+# Generic knob defaults applied to every inferred recipe (the operator can tweak
+# these in the setup form). Family-specific flags (parsers, MoE backend) are
+# layered on separately by `_detect_family`.
+_COMMON_KNOBS = {
+    "max_model_len": 32768,
+    "gpu_memory_utilization": 0.85,
+    "fastsafetensors": True,
+    "prefix_caching": True,
+    "kv_cache_dtype": "fp8",
+}
+
+
+def repo_to_key(repo: str) -> str:
+    """Stable, URL-safe menu key for a discovered model with no recipe key yet.
+
+    'RedHatAI/Qwen3.6-35B-A3B-NVFP4' -> 'redhatai-qwen3-6-35b-a3b-nvfp4'. The same
+    slug is used by the menu, the setup form, and `_identify_current_model`, so a
+    loaded-but-unconfigured model still highlights as active."""
+    return re.sub(r"[^a-z0-9_-]+", "-", repo.lower()).strip("-")
+
+
+def _detect_family(config: dict) -> tuple[str, list[str], list[str]]:
+    """Return (family_label, vllm_flags, capabilities) inferred from config.json.
+
+    Only family-specific, non-knob flags (parsers, MoE backend) go in vllm_flags;
+    generic knob defaults are handled by the caller. Best-effort and operator-
+    confirmed, so a wrong guess is cheap."""
+    arch = " ".join(config.get("architectures") or [])
+    mtype = str(config.get("model_type") or "")
+    s = (arch + " " + mtype).lower()
+    is_moe = (
+        "moe" in s
+        or any(config.get(k) for k in ("num_experts", "n_routed_experts", "num_local_experts"))
+    )
+    is_vision = (
+        "conditionalgeneration" in s
+        or "vision" in s
+        or "vlforcausallm" in s
+        or "vision_config" in config
+        or "image_token_index" in config
+    )
+    flags: list[str] = []
+    caps: list[str] = []
+    label = "Generic"
+    if mtype.startswith("qwen3") or "qwen3" in s:
+        label = "Qwen3 (MoE)" if is_moe else "Qwen3"
+        flags.append("--reasoning-parser=qwen3")
+        caps.append("reasoning")
+        if is_moe:
+            flags.append("--moe_backend=flashinfer_cutlass")
+    elif "gemma" in s:
+        label = "Gemma (MoE)" if is_moe else "Gemma"
+        flags += ["--reasoning-parser=gemma4", "--tool-call-parser=gemma4", "--enable-auto-tool-choice"]
+        caps += ["reasoning", "tools"]
+        if is_moe:
+            # The fast flashinfer/CUTLASS FP4 path errors on GB10 for Gemma MoE;
+            # marlin is the working fallback (see the Gemma 26B trial notes).
+            flags.append("--moe_backend=marlin")
+    if is_vision and "vision" not in caps:
+        caps.append("vision")
+    return label, flags, caps
+
+
+def _infer_mode(total_bytes: int, on_host_count: int) -> str:
+    """Solo unless the weights are present on both Sparks or too big for one."""
+    if on_host_count >= 2 or total_bytes > SINGLE_SPARK_BYTES:
+        return "cluster"
+    return "solo"
+
+
+def infer_recipe(repo: str, config: dict, total_bytes: int, on_host_count: int) -> dict:
+    """Propose a launch recipe for a discovered model — prefills the setup form."""
+    label, flags, caps = _detect_family(config or {})
+    mode = _infer_mode(total_bytes, on_host_count)
+    vllm_args = list(flags)
+    vllm_args.append("--max-num-batched-tokens=16384")
+    knobs = dict(_COMMON_KNOBS)
+    if mode == "cluster":
+        # Large models shard across both Sparks via Ray; leave more headroom.
+        vllm_args += ["-tp=2", "--distributed-executor-backend=ray"]
+        knobs["gpu_memory_utilization"] = 0.7
+    return {
+        "key": repo_to_key(repo),
+        "repo": repo,
+        "display_name": repo.split("/")[-1],
+        "mode": mode,
+        "capabilities": caps,
+        "vllm_args": vllm_args,
+        "knobs": knobs,
+        "family": label,
+    }
+
+
+def _menu_entry_from_recipe(m, *, on_disk: bool, total_bytes: int, per_host: list[dict]) -> dict:
+    d = m.model_dump()
+    d["effective_knobs"] = {**extract_knobs_from_args(m.vllm_args), **(m.knobs or {})}
+    d["needs_setup"] = False
+    d["on_disk"] = on_disk
+    d["total_bytes"] = total_bytes
+    d["per_host"] = per_host
+    return d
+
+
+async def build_menu(settings: Settings, catalog) -> dict[str, dict]:
+    """The disk-driven model menu: every completed model on the Sparks, annotated
+    with its launch recipe (matched by repo) or flagged `needs_setup` if none.
+
+    Two SSH scans total (one per Spark), run in parallel — much cheaper than the
+    old per-recipe disk probe. A host that errors is skipped, not fatal."""
+    hosts = [(settings.spark1_host, settings.spark1_user)]
+    if settings.spark2_host:
+        hosts.append((settings.spark2_host, settings.spark2_user))
+    scans = await asyncio.gather(
+        *(list_cached_models(h, u, settings) for h, u in hosts),
+        return_exceptions=True,
+    )
+    by_repo: dict[str, dict] = {}
+    for (h, _u), res in zip(hosts, scans):
+        if isinstance(res, Exception):
+            continue
+        for repo, size, complete in res:
+            e = by_repo.setdefault(repo, {"total_bytes": 0, "per_host": [], "complete": False})
+            e["total_bytes"] += size
+            e["per_host"].append({"host": h, "size_bytes": size})
+            e["complete"] = e["complete"] or complete
+
+    recipe_by_repo = {m.repo: (k, m) for k, m in catalog.models.items() if m.repo}
+
+    menu: dict[str, dict] = {}
+    for repo, info in by_repo.items():
+        # Skip half-fetched / corrupt caches (no finished snapshot) — they'd show
+        # as broken cards. In-flight downloads surface in the download panel.
+        if not info["complete"]:
+            continue
+        if repo in recipe_by_repo:
+            key, m = recipe_by_repo[repo]
+            menu[key] = _menu_entry_from_recipe(
+                m, on_disk=True, total_bytes=info["total_bytes"], per_host=info["per_host"]
+            )
+        else:
+            key = repo_to_key(repo)
+            menu[key] = {
+                "display_name": repo.split("/")[-1],
+                "repo": repo,
+                "local_path": None,
+                "size_gb": round(info["total_bytes"] / 1e9, 1),
+                "mode": _infer_mode(info["total_bytes"], len(info["per_host"])),
+                "capabilities": [],
+                "expected_ready_seconds": 300,
+                "vllm_args": [],
+                "description": None,
+                "knobs": None,
+                "custom": False,
+                "needs_setup": True,
+                "effective_knobs": {},
+                "on_disk": True,
+                "total_bytes": info["total_bytes"],
+                "per_host": info["per_host"],
+            }
+
+    # Local/fine-tuned recipes live as a directory, not an HF cache entry — probe
+    # each by path and include it if present. Their keys are unique catalog keys
+    # (and local models carry repo="" per ModelDef), so they never collide with a
+    # discovered repo's slug or an HF recipe key above.
+    for key, m in catalog.models.items():
+        if not m.local_path:
+            continue
+        st = await probe_disk(m.repo, m.mode, settings, local_path=m.local_path)
+        if not st.on_disk:
+            continue
+        menu[key] = _menu_entry_from_recipe(
+            m,
+            on_disk=True,
+            total_bytes=st.total_bytes,
+            per_host=[{"host": r.host, "size_bytes": r.size_bytes} for r in st.per_host if r.on_disk],
+        )
+
+    return menu

image/app/disk.py

@@ -10,11 +10,13 @@ model or one tied to an in-flight swap/download.
 """
 from __future__ import annotations
 import asyncio
+import json
 import re
 from dataclasses import dataclass
 from typing import Optional
 
 from .config import Settings
+from .shellsafe import quote_arg
 from .ssh import ssh_run
 
 
@@ -35,6 +37,87 @@ def repo_to_cache_dirname(repo: str) -> str:
     return dn
 
 
+def cache_dirname_to_repo(dirname: str) -> Optional[str]:
+    """Inverse of `repo_to_cache_dirname`: 'models--org--name' -> 'org/name'.
+
+    A repo has exactly one '/', so the org is the first '--'-segment and the name
+    is everything after (names may themselves contain single dashes). Returns
+    None for anything that isn't a model cache dir."""
+    if not dirname.startswith("models--"):
+        return None
+    parts = dirname[len("models--"):].split("--")
+    if len(parts) < 2 or not parts[0] or not parts[1]:
+        return None
+    return f"{parts[0]}/{'--'.join(parts[1:])}"
+
+
+def parse_cache_listing(out: str) -> list[tuple[str, int, bool]]:
+    """Parse the 'size|complete|dirname' lines from `list_cached_models`'s scan.
+
+    Returns [(repo, size_bytes, complete), ...], skipping non-model lines. Pure
+    function so the parsing is unit-testable without SSH."""
+    items: list[tuple[str, int, bool]] = []
+    for line in out.splitlines():
+        line = line.strip()
+        if line.count("|") < 2:
+            continue
+        size_s, complete_s, dirname = line.split("|", 2)
+        repo = cache_dirname_to_repo(dirname.strip())
+        if not repo:
+            continue
+        try:
+            size = int(size_s)
+        except ValueError:
+            size = 0
+        items.append((repo, size, complete_s.strip() == "1"))
+    return items
+
+
+async def list_cached_models(host: str, user: str, settings: Settings) -> list[tuple[str, int, bool]]:
+    """Enumerate every Hugging Face model cached on a host: (repo, size_bytes, complete).
+
+    'complete' = the cache has at least one snapshot carrying a config.json (a
+    finished download, not a half-fetched/corrupt dir). One SSH round-trip; the
+    glob's no-match case is handled by the `[ -d ]` guard."""
+    if not host or not user:
+        return []
+    cmd = (
+        'HUB="$HOME/.cache/huggingface/hub"; '
+        'for d in "$HUB"/models--*; do '
+        '[ -d "$d" ] || continue; '
+        'n=$(basename "$d"); '
+        'sz=$(du -sb "$d" 2>/dev/null | cut -f1); sz=${sz:-0}; '
+        'if ls "$d"/snapshots/*/config.json >/dev/null 2>&1; then c=1; else c=0; fi; '
+        'echo "${sz}|${c}|${n}"; '
+        'done'
+    )
+    rc, out, err = await ssh_run(host, user, cmd, settings, timeout=30.0)
+    if rc != 0:
+        return []
+    return parse_cache_listing(out)
+
+
+async def read_model_config(host: str, user: str, repo: str, settings: Settings) -> Optional[dict]:
+    """Read a cached model's config.json (first snapshot) for launch inference.
+
+    Returns the parsed dict, or None if absent/unreadable. The dirname is
+    whitelisted (repo_to_cache_dirname) so it's safe to embed unquoted."""
+    if not host or not user:
+        return None
+    dn = repo_to_cache_dirname(repo)
+    cmd = (
+        f'D=$(ls -d "$HOME/.cache/huggingface/hub/{dn}/snapshots/"*/ 2>/dev/null | head -1); '
+        f'[ -n "$D" ] && cat "${{D}}config.json" 2>/dev/null'
+    )
+    rc, out, err = await ssh_run(host, user, cmd, settings, timeout=20.0)
+    if rc != 0 or not out.strip():
+        return None
+    try:
+        return json.loads(out)
+    except (ValueError, TypeError):
+        return None
+
+
 @dataclass
 class HostDiskResult:
     host: str
@@ -76,16 +159,52 @@ async def probe_host(host: str, user: str, repo: str, settings: Settings) -> Hos
     return HostDiskResult(host=host, on_disk=True, size_bytes=size)
 
 
-async def probe_disk(repo: str, mode: str, settings: Settings) -> DiskStatus:
+async def probe_local_host(host: str, user: str, path: str, settings: Settings) -> HostDiskResult:
-    """Probe one model across the relevant Sparks based on its mode (solo|cluster)."""
+    """Return whether a local model directory exists on this host and its size.
+
+    For locally fine-tuned models (a Spark directory, not an HF cache entry). The
+    path is whitelisted at the API boundary (shellsafe.validate_local_path); we
+    shlex-quote it here in depth.
+    """
+    if not host or not user:
+        return HostDiskResult(host=host or "?", on_disk=False, error="host not configured")
+    qp = quote_arg(path)
+    cmd = f"if [ -d {qp} ]; then du -sb {qp} 2>/dev/null | cut -f1; else echo MISSING; fi"
+    rc, out, err = await ssh_run(host, user, cmd, settings, timeout=20.0)
+    if rc != 0:
+        return HostDiskResult(host=host, on_disk=False, error=(err or out).strip() or f"rc={rc}")
+    raw = out.strip()
+    if raw == "MISSING" or raw == "":
+        return HostDiskResult(host=host, on_disk=False)
+    try:
+        size = int(raw.splitlines()[-1])
+    except ValueError:
+        return HostDiskResult(host=host, on_disk=False, error=f"unparsable du output: {raw!r}")
+    return HostDiskResult(host=host, on_disk=True, size_bytes=size)
+
+
+async def probe_disk(
+    repo: str, mode: str, settings: Settings, *, local_path: str | None = None
+) -> DiskStatus:
+    """Probe one model across the relevant Sparks based on its mode (solo|cluster).
+
+    A local model (local_path set) is probed by directory; otherwise by HF cache.
+    """
     hosts: list[tuple[str, str]] = [(settings.spark1_host, settings.spark1_user)]
     if mode == "cluster" and settings.spark2_host:
         hosts.append((settings.spark2_host, settings.spark2_user))
 
-    results = await asyncio.gather(*(probe_host(h, u, repo, settings) for h, u in hosts))
+    if local_path:
+        results = await asyncio.gather(
+            *(probe_local_host(h, u, local_path, settings) for h, u in hosts)
+        )
+        key = local_path
+    else:
+        results = await asyncio.gather(*(probe_host(h, u, repo, settings) for h, u in hosts))
+        key = repo
     on_disk = any(r.on_disk for r in results)
     total = sum(r.size_bytes for r in results)
-    return DiskStatus(repo=repo, on_disk=on_disk, total_bytes=total, per_host=list(results))
+    return DiskStatus(repo=key, on_disk=on_disk, total_bytes=total, per_host=list(results))
 
 
 async def delete_host(host: str, user: str, repo: str, settings: Settings) -> HostDiskResult:
@@ -122,10 +241,14 @@ async def delete_host(host: str, user: str, repo: str, settings: Settings) -> Ho
     return HostDiskResult(host=host, on_disk=False, size_bytes=freed)
 
 
-async def delete_from_disk(repo: str, mode: str, settings: Settings) -> DiskStatus:
+async def delete_from_disk(repo: str, settings: Settings) -> DiskStatus:
-    """rm -rf the model's cache dir on the relevant Sparks. Idempotent."""
+    """rm -rf the model's cache dir on ALL configured Sparks. Idempotent.
+
+    We sweep both Sparks regardless of the model's declared mode: a 'remove from
+    disk & menu' must leave nothing behind, and rm of an absent dir reports 0
+    bytes freed (FREED 0), so an extra host is harmless."""
     hosts: list[tuple[str, str]] = [(settings.spark1_host, settings.spark1_user)]
-    if mode == "cluster" and settings.spark2_host:
+    if settings.spark2_host:
         hosts.append((settings.spark2_host, settings.spark2_user))
 
     results = await asyncio.gather(*(delete_host(h, u, repo, settings) for h, u in hosts))

image/app/health.py

@@ -6,17 +6,28 @@ from .config import Settings
 _TIMEOUT = 3.0
 
 
-async def check_vllm(settings: Settings) -> dict:
+def _disabled(settings: Settings, key: str) -> dict | None:
-    base_url = (
+    """A clean 'disabled' verdict if `key` is in DISABLED_SERVICES, else None.
-        f"http://{settings.spark1_host}:{settings.vllm_port}/v1"
+
-        if settings.spark1_host
+    Lets an adopter who doesn't run a given support service switch its probe off
-        else None
+    entirely — so the probe never hits whatever else listens on that port, and
-    )
+    the connectivity log doesn't record it as perpetually down."""
-    if not settings.spark1_host:
+    if key in settings.disabled_services:
-        return {"ok": False, "error": "spark1 not configured", "base_url": base_url}
+        return {"ok": False, "disabled": True, "error": "disabled", "base_url": None}
+    return None
+
+
+async def probe_vllm_endpoint(host: str, port: int) -> dict:
+    """Probe any OpenAI-compatible vLLM at host:port via /v1/models.
+
+    Shared by the primary (Spark 1) health check and any extra vLLM registered
+    as a custom service (kind: vllm) to monitor a second Spark."""
+    base_url = f"http://{host}:{port}/v1" if host else None
+    if not host:
+        return {"ok": False, "error": "vllm host not configured", "base_url": base_url}
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as c:
-            r = await c.get(f"http://{settings.spark1_host}:{settings.vllm_port}/v1/models")
+            r = await c.get(f"http://{host}:{port}/v1/models")
             r.raise_for_status()
             ids = [m["id"] for m in r.json().get("data", [])]
             return {
@@ -29,7 +40,15 @@ async def check_vllm(settings: Settings) -> dict:
         return {"ok": False, "error": str(e), "base_url": base_url}
 
 
+async def check_vllm(settings: Settings) -> dict:
+    if not settings.spark1_host:
+        return {"ok": False, "error": "spark1 not configured", "base_url": None}
+    return await probe_vllm_endpoint(settings.spark1_host, settings.vllm_port)
+
+
 async def check_parakeet(settings: Settings) -> dict:
+    if d := _disabled(settings, "parakeet"):
+        return d
     base_url = (
         f"http://{settings.parakeet_host}:{settings.parakeet_port}"
         if settings.parakeet_host
@@ -47,6 +66,8 @@ async def check_parakeet(settings: Settings) -> dict:
 
 
 async def check_kokoro(settings: Settings) -> dict:
+    if d := _disabled(settings, "kokoro"):
+        return d
     base_url = (
         f"http://{settings.kokoro_host}:{settings.kokoro_port}"
         if settings.kokoro_host
@@ -68,6 +89,8 @@ async def check_kokoro(settings: Settings) -> dict:
 
 
 async def check_embeddings(settings: Settings) -> dict:
+    if d := _disabled(settings, "embeddings"):
+        return d
     base_url = (
         f"http://{settings.embed_host}:{settings.embed_port}"
         if settings.embed_host
@@ -89,6 +112,8 @@ async def check_embeddings(settings: Settings) -> dict:
 
 
 async def check_qdrant(settings: Settings) -> dict:
+    if d := _disabled(settings, "qdrant"):
+        return d
     base_url = (
         f"http://{settings.qdrant_host}:{settings.qdrant_port}"
         if settings.qdrant_host

image/app/models.py

@@ -1,15 +1,33 @@
 from __future__ import annotations
+import logging
 from typing import Literal, Optional
 import yaml
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, model_validator
 
 from .overrides import apply_knobs_to_args, load_overrides
-from .shellsafe import quote_arg, quote_args
+from .shellsafe import quote_arg, quote_args, validate_local_path
+
+log = logging.getLogger(__name__)
+
+
+def _chat_template_path(vllm_args: list[str]) -> str | None:
+    """Extract the path from a `--chat-template=<path>` arg, if present."""
+    for a in vllm_args:
+        if a.startswith("--chat-template="):
+            return a.split("=", 1)[1]
+    return None
+
+
+def _is_within(path: str, base: str) -> bool:
+    """True if `path` is `base` itself or lives inside it (lexical check)."""
+    base = base.rstrip("/")
+    return path == base or path.startswith(base + "/")
 
 
 class ModelDef(BaseModel):
     display_name: str
-    repo: str
+    repo: str = ""                   # HF 'org/name'; empty for a local model
+    local_path: str | None = None    # absolute dir on the Spark; set => local model
     size_gb: float
     mode: Literal["solo", "cluster"]
     capabilities: list[str] = Field(default_factory=list)
@@ -19,6 +37,38 @@ class ModelDef(BaseModel):
     knobs: dict | None = None       # user-customized; merged at launch time
     custom: bool = False             # True if this came from /data overrides
 
+    @model_validator(mode="after")
+    def _validate_source(self) -> "ModelDef":
+        if bool(self.repo) == bool(self.local_path):
+            raise ValueError(
+                f"model {self.display_name!r} must set exactly one of 'repo' (HF) "
+                f"or 'local_path' (Spark directory)"
+            )
+        if self.local_path:
+            # Single place that enforces the path whitelist, so YAML/override
+            # entries get the same boundary check as the API. The quote_arg sink
+            # is still defense-in-depth.
+            validate_local_path(self.local_path)
+            # Only local_path is bind-mounted into the vLLM container, so any
+            # --chat-template path must live inside it or vLLM can't find it.
+            tmpl = _chat_template_path(self.vllm_args)
+            if tmpl is not None and not _is_within(tmpl, self.local_path):
+                raise ValueError(
+                    f"--chat-template path {tmpl!r} must be inside the model "
+                    f"directory {self.local_path!r} (only that directory is mounted "
+                    f"into the container)"
+                )
+        return self
+
+    @property
+    def is_local(self) -> bool:
+        return bool(self.local_path)
+
+    @property
+    def source(self) -> str:
+        """What `vllm serve` is pointed at: the local dir if set, else the HF repo."""
+        return self.local_path if self.local_path else self.repo
+
 
 class Defaults(BaseModel):
     port: int = 8888
@@ -47,7 +97,8 @@ def _merge_overrides(catalog: Catalog) -> Catalog:
             continue
         defaults_dump = {
             "display_name": entry.get("display_name", key),
-            "repo": entry["repo"],
+            "repo": entry.get("repo", ""),
+            "local_path": entry.get("local_path"),
             "size_gb": float(entry.get("size_gb", 0)),
             "mode": entry.get("mode", "solo"),
             "capabilities": entry.get("capabilities") or [],
@@ -57,7 +108,12 @@ def _merge_overrides(catalog: Catalog) -> Catalog:
             "knobs": entry.get("knobs"),
             "custom": True,
         }
-        new_models[key] = ModelDef.model_validate(defaults_dump)
+        # A single malformed override entry (bad path, missing source, etc.) must
+        # not take down the whole catalog — skip it and keep the rest loadable.
+        try:
+            new_models[key] = ModelDef.model_validate(defaults_dump)
+        except Exception as e:
+            log.warning("skipping invalid custom model %r: %s", key, e)
 
     return Catalog(defaults=catalog.defaults, models=new_models)
 
@@ -78,7 +134,21 @@ def build_launch_command(key: str, model: ModelDef, defaults: Defaults) -> str:
     solo = "--solo " if model.mode == "solo" else ""
     base_args = apply_knobs_to_args(list(model.vllm_args), model.knobs)
     args = [f"--port={defaults.port}", f"--host={defaults.host}", *base_args]
-    # repo + args are user-controlled (custom models, knobs); shlex.quote each so
+    # source + args are user-controlled (custom models, knobs); shlex.quote each
-    # they cannot break out of the SSH shell command. shlex.split (used by the
+    # so they cannot break out of the SSH shell command. shlex.split (used by the
     # vLLM pre-flight validator) cleanly reverses this quoting.
-    return f"./launch-cluster.sh {solo}-d exec vllm serve {quote_arg(model.repo)} {quote_args(args)}"
+    prefix = ""
+    if model.local_path:
+        # A local model's directory isn't in the HF cache the launch script
+        # already mounts, so bind-mount it at the SAME path inside the vllm
+        # container via the script's VLLM_SPARK_EXTRA_DOCKER_ARGS hook. Same
+        # path inside and out means `vllm serve <dir>` and any
+        # `--chat-template=<dir>/...` arg both resolve. No launch-cluster.sh
+        # change needed. (The env assignment sits before the script, so the
+        # validator's `serve`-keyed shlex round-trip is unaffected.)
+        mount = quote_arg(f"-v {model.local_path}:{model.local_path}")
+        prefix = f"VLLM_SPARK_EXTRA_DOCKER_ARGS={mount} "
+    return (
+        f"{prefix}./launch-cluster.sh {solo}-d exec vllm serve "
+        f"{quote_arg(model.source)} {quote_args(args)}"
+    )

image/app/overrides.py

@@ -14,7 +14,7 @@ Shape:
     custom:
       - key: my-new-model
         display_name: My New Model (from download)
-        repo: my-org/my-model
+        repo: my-org/my-model        # an HF repo; OR set local_path instead (exactly one)
         size_gb: 20
         mode: solo
         description: null
@@ -25,6 +25,12 @@ Shape:
           fastsafetensors: true
           prefix_caching: true
           kv_cache_dtype: fp8
+      - key: my-finetune                       # a local/fine-tuned model (a directory on the Spark)
+        display_name: My Fine-tune
+        local_path: /home/you/models/my-finetune
+        size_gb: 59
+        mode: solo
+        vllm_args: [--chat-template=/home/you/models/my-finetune/chat_template.jinja]
 """
 from __future__ import annotations
 import os

image/app/server.py

@@ -6,25 +6,27 @@ from pathlib import Path
 from fastapi import FastAPI, HTTPException, Query, Request
 from fastapi.responses import FileResponse, JSONResponse, StreamingResponse
 from fastapi.staticfiles import StaticFiles
-from pydantic import BaseModel
+from pydantic import BaseModel, ValidationError
 from typing import Literal
 
 from .config import Settings
 from .connectivity import get_mac, record_report, record_state, summary as connectivity_summary
+from .coordination import LockHeld, ScheduleRegistry, SwapLockManager, WebhookNotifier, valid_schedule_id
 from .custom_services import add_custom_service, delete_custom_service
 from .audio_proxy import build_router as build_audio_router
 from .deep_health import DeepHealth
-from .disk import delete_from_disk, probe_disk
+from .discovery import build_menu, infer_recipe, repo_to_key
+from .disk import delete_from_disk, probe_host, read_model_config
 from .download import DownloadManager
 from .llm_proxy import build_router as build_llm_router
 from .embeddings_proxy import build_router as build_embeddings_router
 from .redaction_gateway import build_router as build_redaction_router, MapStore
 from .hardware import HardwareProbe
-from .health import check_kokoro, check_parakeet, check_vllm, check_embeddings, check_qdrant
+from .health import check_kokoro, check_parakeet, check_vllm, check_embeddings, check_qdrant, probe_vllm_endpoint
 from .matrix_bridge import MatrixBridgeManager
-from .models import load_catalog
+from .models import ModelDef, load_catalog
 from .nim import SUGGESTED_NIMS, CATALOG_URL, NimManager
-from .overrides import add_custom, delete_custom, extract_knobs_from_args, load_overrides, set_knobs
+from .overrides import add_custom, delete_custom, load_overrides, set_knobs
 from .services import docker_state, run_action, services_from_settings
 from .shellsafe import validate_container, validate_image, validate_repo
 from .speech_models import SpeechModelsManager
@@ -37,7 +39,12 @@ from .wol import send_local_broadcast, send_via_peer
 
 settings = Settings.from_env()
 catalog = load_catalog(settings.models_yaml)
-swap_manager = SwapManager(settings, catalog)
+# Coordination layer (GPU arbiter): swap-lifecycle webhook, the swap reservation
+# lock, and the read-only schedule registry. See coordination.py.
+swap_webhook = WebhookNotifier(settings.swap_webhook_url, settings.swap_webhook_secret)
+swap_lock = SwapLockManager()
+schedule_registry = ScheduleRegistry()
+swap_manager = SwapManager(settings, catalog, notifier=swap_webhook)
 download_manager = DownloadManager(settings)
 update_manager = UpdateManager(settings)
 hardware_probe = HardwareProbe(settings)
@@ -67,6 +74,10 @@ _CSRF_EXEMPT_PREFIXES = (
     "/api/audio/",        # diarize-chunk / label-merge / transcribe-with-speakers
     "/api/health-event",  # health reports posted by consumer apps
 )
+# Note: the coordination endpoints (/api/swap/lock, /api/schedule) are
+# intentionally NOT exempt. External schedulers are non-browser clients (no
+# Origin header) so they pass the guard already — same as /api/swap — while a
+# malicious page can't drive them from the operator's browser. Don't add them.
 
 
 @app.middleware("http")
@@ -151,20 +162,65 @@ def _reload_catalog() -> None:
     swap_manager.reload_catalog(catalog)
 
 
+def _recipe_summaries() -> list[dict]:
+    """Known launch recipes (bundled + saved), for the download panel's autocomplete.
+
+    These are NOT the menu — the menu is what's on disk. This is just the set of
+    repos Spark Control already knows how to launch, so the download box can
+    suggest them by name without putting phantom cards on the dashboard."""
+    out = []
+    for m in catalog.models.values():
+        if m.repo:
+            out.append({"repo": m.repo, "display_name": m.display_name, "mode": m.mode})
+    return out
+
+
 @app.get("/api/models")
 async def get_models() -> dict:
-    out_models: dict[str, dict] = {}
+    """The model menu = what's actually downloaded on the Sparks (one scan per
-    for key, m in catalog.models.items():
+    Spark), each annotated with its launch recipe or flagged `needs_setup`.
-        d = m.model_dump()
+
-        # Always include effective knobs for the UI (defaults from base args + any overrides)
+    Does SSH, so it's the slower of the model endpoints; the front-end calls it on
-        d["effective_knobs"] = {**extract_knobs_from_args(m.vllm_args), **(m.knobs or {})}
+    load, after a swap/download/delete, and on a slow timer — not every poll."""
-        out_models[key] = d
+    if not settings.configured:
+        return {"configured": False, "defaults": catalog.defaults.model_dump(), "models": {}, "recipes": []}
+    menu = await build_menu(settings, catalog)
     return {
+        "configured": True,
         "defaults": catalog.defaults.model_dump(),
-        "models": out_models,
+        "models": menu,
+        "recipes": _recipe_summaries(),
     }
 
 
+@app.get("/api/models/suggest")
+async def suggest_model(repo: str = Query(...)) -> dict:
+    """Read a downloaded model's config.json + size and propose a launch recipe.
+
+    Prefills the 'set up this model' form for an on-disk model that has no recipe
+    yet. The operator confirms/edits, then POSTs it to /api/models to save."""
+    if not settings.configured:
+        raise HTTPException(503, "spark1 not configured")
+    try:
+        validate_repo(repo)
+    except ValueError as e:
+        raise HTTPException(400, str(e))
+    hosts = [(settings.spark1_host, settings.spark1_user)]
+    if settings.spark2_host:
+        hosts.append((settings.spark2_host, settings.spark2_user))
+    # Config from whichever Spark has it; size summed across the Sparks that do.
+    sizes = await asyncio.gather(*(probe_host(h, u, repo, settings) for h, u in hosts))
+    total = sum(r.size_bytes for r in sizes if r.on_disk)
+    on_hosts = sum(1 for r in sizes if r.on_disk)
+    config = None
+    for (h, u), r in zip(hosts, sizes):
+        if r.on_disk:
+            config = await read_model_config(h, u, repo, settings)
+            if config is not None:
+                break
+    return infer_recipe(repo, config or {}, total, on_hosts)
+
+
 class KnobsBody(BaseModel):
     knobs: dict
 
@@ -183,7 +239,8 @@ async def put_model_knobs(key: str, body: KnobsBody) -> dict:
 class CustomModelBody(BaseModel):
     key: str
     display_name: str
-    repo: str
+    repo: str = ""
+    local_path: str | None = None
     size_gb: float = 0
     mode: Literal["solo", "cluster"] = "solo"
     description: str | None = None
@@ -196,8 +253,17 @@ class CustomModelBody(BaseModel):
 async def post_model(body: CustomModelBody) -> dict:
     if not body.key or not body.key.replace("-", "").replace("_", "").isalnum():
         raise HTTPException(400, "key must be alphanumeric/-/_ only")
+    # Validate the full entry BEFORE persisting (exactly-one source, local-path
+    # whitelist, chat-template location). Doing it via ModelDef means the API and
+    # the YAML-override path share one set of rules, and a bad entry can't be
+    # written to /data and then break catalog load.
     try:
-        validate_repo(body.repo)
+        ModelDef.model_validate(body.model_dump())
+        if body.repo:
+            validate_repo(body.repo)  # HF charset (the model only validates local paths)
+    except ValidationError as e:
+        msg = e.errors()[0]["msg"] if e.errors() else str(e)
+        raise HTTPException(400, msg.removeprefix("Value error, "))
     except ValueError as e:
         raise HTTPException(400, str(e))
     if body.key in catalog.models and not catalog.models[body.key].custom:
@@ -218,57 +284,43 @@ async def del_model(key: str) -> dict:
     return {"ok": True, "key": key}
 
 
-@app.get("/api/models/disk-status")
-async def get_models_disk_status() -> dict:
-    """Probe each catalog model's HF cache on the appropriate Spark(s) in parallel.
-
-    Result is keyed by model key: {on_disk, total_bytes, per_host:[{host,on_disk,size_bytes,error?}]}.
-    Designed to be called once on dashboard load; takes ~1–3s depending on Spark count.
-    """
-    if not settings.configured:
-        return {"configured": False, "models": {}}
-    keys = list(catalog.models.keys())
-    statuses = await asyncio.gather(*(
-        probe_disk(catalog.models[k].repo, catalog.models[k].mode, settings) for k in keys
-    ), return_exceptions=True)
-    out: dict[str, dict] = {}
-    for k, s in zip(keys, statuses):
-        if isinstance(s, Exception):
-            out[k] = {"on_disk": False, "total_bytes": 0, "per_host": [], "error": str(s)}
-            continue
-        out[k] = {
-            "on_disk": s.on_disk,
-            "total_bytes": s.total_bytes,
-            "per_host": [
-                {"host": r.host, "on_disk": r.on_disk, "size_bytes": r.size_bytes, **({"error": r.error} if r.error else {})}
-                for r in s.per_host
-            ],
-        }
-    return {"configured": True, "models": out}
-
-
 @app.delete("/api/models/{key}/disk")
 async def del_model_disk(key: str) -> dict:
-    """Delete a model's weights from the Spark filesystem(s). The catalog entry stays.
+    """Remove a model's weights from the Sparks — and thus from the menu, since the
+    menu IS the disk. Resolves the key against the live menu, so a discovered
+    model (no saved recipe) is deletable too.
 
     Safety rails:
+      - Refuses a local/fine-tuned directory (hand-placed, not re-downloadable).
       - Refuses if the model is currently loaded on vLLM.
-      - Refuses if a swap or download is in flight.
+      - Refuses if a swap or this model's own download is in flight.
-      - Idempotent: if the cache dir is already gone on a host, that host reports 0 bytes freed.
+      - Idempotent across both Sparks: an already-absent cache dir frees 0 bytes.
     """
-    if key not in catalog.models:
+    if not settings.configured:
+        raise HTTPException(503, "spark1 not configured")
+    menu = await build_menu(settings, catalog)
+    entry = menu.get(key)
+    if entry is None:
         raise HTTPException(404, f"unknown model: {key}")
-    m = catalog.models[key]
+
+    # Never rm a local fine-tune directory from the dashboard — it's irreplaceable
+    # training output the user placed by hand, not a re-downloadable HF cache.
+    if entry.get("local_path"):
+        raise HTTPException(
+            400,
+            "this is a local model; its directory must be managed on the Spark, not deleted from here",
+        )
+    repo = entry["repo"]
 
     # Refuse if currently loaded
     try:
         vllm = await check_vllm(settings)
     except Exception:
         vllm = {}
-    if vllm.get("ok") and vllm.get("current_model") == m.repo:
+    if vllm.get("ok") and vllm.get("current_model") == repo:
         raise HTTPException(
             409,
-            f"'{m.display_name}' is the currently loaded model. Switch to a different model first, then try again."
+            f"'{entry['display_name']}' is the currently loaded model. Switch to a different model first, then try again."
         )
 
     # Refuse if a swap is in flight
@@ -278,10 +330,10 @@ async def del_model_disk(key: str) -> dict:
     # Refuse if a download is in flight for this same repo (a different model's download is fine)
     if download_manager.current_job_id:
         job = download_manager.get(download_manager.current_job_id)
-        if job and job.repo == m.repo:
+        if job and job.repo == repo:
             raise HTTPException(409, "this model is currently downloading; cancel or wait for it to finish")
 
-    status = await delete_from_disk(m.repo, m.mode, settings)
+    status = await delete_from_disk(repo, settings)
     # Audit log
     record_report(
         f"disk:{key}",
@@ -292,7 +344,7 @@ async def del_model_disk(key: str) -> dict:
     return {
         "ok": True,
         "key": key,
-        "repo": m.repo,
+        "repo": repo,
         "bytes_freed": status.total_bytes,
         "per_host": [
             {"host": r.host, "size_bytes": r.size_bytes, **({"error": r.error} if r.error else {})}
@@ -476,6 +528,10 @@ async def get_services() -> dict:
             http = await check_embeddings(settings)
         elif name == "qdrant":
             http = await check_qdrant(settings)
+        elif svc.kind == "vllm":
+            # An extra vLLM monitored on another Spark (registered as a custom
+            # service). Probe its own host/port, not the primary Spark 1 one.
+            http = await probe_vllm_endpoint(svc.host, svc.port)
         elif svc.kind == "bot":
             # No HTTP health endpoint (host networking, no port) — judged purely
             # by docker state. http_ready stays None so the badge isn't pinned
@@ -497,7 +553,7 @@ async def get_services() -> dict:
             # Prefer the check fn's own top-level model key (embeddings reports
             # it there); fall back to a model field inside detail for services
             # whose /health embeds it (parakeet).
-            "model": http.get("model") or ((http.get("detail") or {}).get("model") if isinstance(http.get("detail"), dict) else None),
+            "model": http.get("model") or http.get("current_model") or ((http.get("detail") or {}).get("model") if isinstance(http.get("detail"), dict) else None),
             "docker_state": docker.get("state"),
             "restart_count": docker.get("restart_count"),
             "started_at": docker.get("started_at"),
@@ -775,17 +831,20 @@ async def get_endpoints() -> dict:
             "base_url": vllm.get("base_url"),
             "model": vllm.get("current_model"),
             "openai_compat": True,
+            "disabled": bool(vllm.get("disabled")),
         },
         "parakeet": {
             "ready": bool(parakeet.get("ok")),
             "base_url": parakeet.get("base_url"),
             "kind": "stt",
             "model": (parakeet.get("detail") or {}).get("model") if isinstance(parakeet.get("detail"), dict) else None,
+            "disabled": bool(parakeet.get("disabled")),
         },
         "kokoro": {
             "ready": bool(kokoro.get("ok")),
             "base_url": kokoro.get("base_url"),
             "kind": "tts",
+            "disabled": bool(kokoro.get("disabled")),
         },
         "embeddings": {
             "ready": bool(embeddings.get("ok")),
@@ -794,12 +853,14 @@ async def get_endpoints() -> dict:
             "model": embeddings.get("model"),
             # The proxied OpenAI-compatible endpoints live on Spark Control itself.
             "openai_endpoints": ["/v1/embeddings", "/v1/rerank", "/api/search"],
+            "disabled": bool(embeddings.get("disabled")),
         },
         "qdrant": {
             "ready": bool(qdrant.get("ok")),
             "base_url": qdrant.get("base_url"),
             "kind": "vectordb",
             "collection": settings.qdrant_collection or None,
+            "disabled": bool(qdrant.get("disabled")),
         },
     }
 
@@ -813,12 +874,15 @@ async def get_status() -> dict:
         check_embeddings(settings),
         check_qdrant(settings),
     )
-    # Feed health into the connectivity log (deduped — only logs on transition)
+    # Feed health into the connectivity log (deduped — only logs on transition).
-    record_state("vllm", bool(vllm.get("ok")))
+    # Skip services switched off via DISABLED_SERVICES — they'd otherwise log as
-    record_state("parakeet", bool(parakeet.get("ok")))
+    # perpetually down.
-    record_state("kokoro", bool(kokoro.get("ok")))
+    for _name, _r in (
-    record_state("embeddings", bool(embeddings.get("ok")))
+        ("vllm", vllm), ("parakeet", parakeet), ("kokoro", kokoro),
-    record_state("qdrant", bool(qdrant.get("ok")))
+        ("embeddings", embeddings), ("qdrant", qdrant),
+    ):
+        if not _r.get("disabled"):
+            record_state(_name, bool(_r.get("ok")))
     current_key = _identify_current_model(vllm.get("current_model"))
     return {
         "configured": settings.configured,
@@ -835,10 +899,13 @@ async def get_status() -> dict:
 def _identify_current_model(repo: str | None) -> str | None:
     if not repo:
         return None
+    # A recipe-backed model keys by its recipe key; a discovered model (loaded but
+    # not yet set up) keys by the same slug build_menu uses, so it still
+    # highlights as the active card.
     for key, m in catalog.models.items():
         if m.repo == repo:
             return key
-    return None
+    return repo_to_key(repo)
 
 
 class SwapRequest(BaseModel):
@@ -856,9 +923,21 @@ async def validate_swap(key: str) -> dict:
 
 
 @app.post("/api/swap")
-async def post_swap(req: SwapRequest) -> dict:
+async def post_swap(req: SwapRequest, request: Request) -> dict:
     if not settings.configured and not req.dry_run:
         raise HTTPException(503, "spark1 not configured")
+    # Enforce the swap reservation lock (the GPU arbiter). A held lock blocks any
+    # real swap that doesn't present the holder's token in X-Swap-Lock-Token — so
+    # an external scheduler that holds the lock can swap, but the dashboard (no
+    # token) is refused while someone else holds it. Dry runs don't touch the
+    # cluster, so they're exempt.
+    if not req.dry_run:
+        blocked = swap_lock.is_blocked_by(request.headers.get("x-swap-lock-token"))
+        if blocked is not None:
+            raise HTTPException(status_code=423, detail={
+                "error": "the GPU swap path is reserved by another holder",
+                "lock": blocked,
+            })
     try:
         job = await swap_manager.trigger(req.model_key, dry_run=req.dry_run)
     except KeyError:
@@ -913,6 +992,89 @@ async def stream_swap(job_id: str):
     return StreamingResponse(gen(), media_type="text/event-stream")
 
 
+# ---- Coordination layer: swap lock + schedule registry ----
+# Endpoints are control-surface, not browser-exempt: an external scheduler is a
+# non-browser client (no Origin header) so it passes the CSRF guard already, the
+# same way it calls /api/swap today; the dashboard is same-origin.
+
+class LockAcquireRequest(BaseModel):
+    holder: str
+    ttl_seconds: int | None = None
+    note: str = ""
+    token: str | None = None   # present only to extend an existing hold
+
+
+@app.post("/api/swap/lock")
+async def acquire_swap_lock(req: LockAcquireRequest) -> dict:
+    """Reserve the GPU swap path. Returns a secret token used to swap (header
+    X-Swap-Lock-Token) and to release. 409 if held by another holder."""
+    try:
+        lock = swap_lock.acquire(req.holder, req.ttl_seconds, req.note, token=req.token)
+    except ValueError as e:
+        raise HTTPException(422, str(e))
+    except LockHeld as e:
+        raise HTTPException(status_code=409, detail={
+            "error": "swap lock is held by another holder",
+            "lock": e.state,
+        })
+    return {**swap_lock.status(), "token": lock.token}
+
+
+@app.get("/api/swap/lock")
+async def get_swap_lock() -> dict:
+    """Public, token-free view of the reservation: held? who? until when?"""
+    return swap_lock.status()
+
+
+@app.delete("/api/swap/lock")
+async def release_swap_lock(request: Request, force: bool = Query(False)) -> dict:
+    """Release the reservation. Needs the matching X-Swap-Lock-Token unless
+    ?force=true (the human override from the dashboard)."""
+    token = request.headers.get("x-swap-lock-token") or request.query_params.get("token")
+    try:
+        released = swap_lock.release(token, force=force)
+    except PermissionError as e:
+        raise HTTPException(403, str(e))
+    return {"released": released, **swap_lock.status()}
+
+
+class ScheduleRequest(BaseModel):
+    name: str
+    id: str | None = None
+    owner: str = ""
+    cron: str = ""
+    next_run: str = ""
+    description: str = ""
+
+
+@app.get("/api/schedule")
+async def list_schedules() -> dict:
+    return {"schedules": schedule_registry.list()}
+
+
+@app.post("/api/schedule")
+async def register_schedule(req: ScheduleRequest) -> dict:
+    """Register (or update, by id) a schedule an external scheduler owns. Spark
+    Control only stores it for the dashboard — it never executes it."""
+    try:
+        entry = schedule_registry.register(
+            name=req.name, id=req.id, owner=req.owner,
+            cron=req.cron, next_run=req.next_run, description=req.description,
+        )
+    except ValueError as e:
+        raise HTTPException(422, str(e))
+    return entry.public()
+
+
+@app.delete("/api/schedule/{schedule_id}")
+async def delete_schedule(schedule_id: str) -> dict:
+    # Whitelist the path segment at the boundary (repo convention), even though
+    # it's only ever a dict key — keeps it from being reflected or logged raw.
+    if not valid_schedule_id(schedule_id):
+        raise HTTPException(422, "invalid schedule id")
+    return {"deleted": schedule_registry.delete(schedule_id)}
+
+
 class DownloadRequest(BaseModel):
     repo: str
     mode: Literal["spark1", "spark2", "cluster"] = "spark1"

image/app/services.py

@@ -5,6 +5,7 @@ machinery. We just run `docker start|stop|restart <container>` via SSH on the
 appropriate host.
 """
 from __future__ import annotations
+import logging
 import time
 from dataclasses import dataclass
 from typing import Literal, Optional
@@ -13,6 +14,8 @@ from .config import Settings
 from .shellsafe import quote_arg
 from .ssh import ssh_run
 
+log = logging.getLogger(__name__)
+
 
 # Cache the "unreachable" verdict per (host, user) for a short period so that a
 # repeated docker_state call doesn't re-pay the 6 s SSH connect timeout each time.
@@ -103,7 +106,13 @@ def services_from_settings(s: Settings) -> dict[str, ServiceDef]:
     }
     for entry in load_custom_services():
         key = entry.get("key")
-        if not key or key in out:
+        if not key:
+            continue
+        if key in out:
+            # A custom entry can't shadow a built-in (parakeet/kokoro/…); warn so
+            # an adopter who picked a colliding key for, say, a second vLLM sees
+            # why no tile appeared instead of a silent no-op.
+            log.warning("custom service %r collides with a built-in name; ignoring", key)
             continue
         out[key] = ServiceDef(
             name=key,
@@ -113,7 +122,9 @@ def services_from_settings(s: Settings) -> dict[str, ServiceDef]:
             container=entry.get("container", key),
             port=int(entry.get("port", 0)),
         )
-    return out
+    # Drop services the deployment has switched off (DISABLED_SERVICES) so they
+    # show no tile and are never probed/auto-restarted.
+    return {k: v for k, v in out.items() if k not in s.disabled_services}
 
 
 async def docker_state(settings: Settings, svc: ServiceDef) -> dict:

image/app/shellsafe.py

@@ -28,6 +28,12 @@ _IMAGE_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:/@-]*$")
 # Docker container / volume name (Docker's own rule).
 _CONTAINER_RE = re.compile(r"^[A-Za-z0-9][A-Za-z0-9_.-]*$")
 
+# Absolute filesystem path to a local model directory on a Spark. Conservative
+# charset (letters, digits, and safe path punctuation) with a required leading
+# '/', so it carries no shell metacharacters and no whitespace. Traversal ('.'
+# and '..' segments) is rejected separately in validate_local_path.
+_LOCAL_PATH_RE = re.compile(r"^/[A-Za-z0-9._+/-]+$")
+
 
 def validate_repo(repo: str) -> str:
     """Return `repo` if it is a well-formed 'org/name'; else raise ValueError."""
@@ -50,6 +56,25 @@ def validate_container(name: str) -> str:
     return name
 
 
+def validate_local_path(path: str) -> str:
+    """Return `path` if it is a safe absolute model directory path; else ValueError.
+
+    For locally fine-tuned models served by directory (not an HF repo). Requires
+    an absolute path, a metacharacter-free charset, and no '.'/'..' segments so a
+    caller cannot traverse out of an intended models directory. The `quote_arg`
+    sink still quotes it in depth — this is the boundary check.
+    """
+    p = path or ""
+    if len(p) > 512 or not _LOCAL_PATH_RE.fullmatch(p):
+        raise ValueError(
+            f"invalid local model path (expected an absolute path, no spaces or "
+            f"shell metacharacters): {path!r}"
+        )
+    if any(seg in (".", "..") for seg in p.split("/")):
+        raise ValueError(f"local model path must not contain '.' or '..' segments: {path!r}")
+    return p
+
+
 def quote_arg(value: object) -> str:
     """shlex.quote a single token for safe embedding in a shell command string."""
     return shlex.quote(str(value))

image/app/static/app.js

@@ -19,13 +19,21 @@ const state = {
   configured: true,
   timer_handle: null,
   deep_health: {},
-  disk_status: {},         // keyed by model key: { on_disk, total_bytes, per_host }
+  models_loaded: false,    // true once the first disk scan (/api/models) returns
-  disk_status_loaded: false,
+  recipes: [],             // known launch recipes (for the download autocomplete)
+  lock: { held: false },   // GPU swap reservation (coordination layer)
+  schedules: [],           // schedules external automation has registered
 };
 
 const el = (sel) => document.querySelector(sel);
 const $$ = (sel) => document.querySelectorAll(sel);
 
+// ISO timestamp -> local clock string (e.g. "2:45:10 PM"); '' if unparseable.
+function fmtClock(iso) {
+  const t = Date.parse(iso);
+  return isNaN(t) ? '' : new Date(t).toLocaleTimeString();
+}
+
 function escapeHtml(s) {
   if (s == null) return '';
   return String(s)
@@ -51,69 +59,86 @@ function renderCards() {
   const root = el('#cards');
   root.innerHTML = '';
   const isSwapping = !!state.swap_job_id;
-  for (const key of Object.keys(state.models)) {
+  // GPU reserved by external automation — manual swaps are refused server-side
+  // (423); reflect that in the buttons so the click never bounces.
+  const locked = !!(state.lock && state.lock.held);
+  const lockTip = locked
+    ? `Reserved by ${state.lock.holder || 'automation'}${state.lock.expires_at ? ' until ' + fmtClock(state.lock.expires_at) : ''}`
+    : '';
+  const keys = Object.keys(state.models);
+  if (keys.length === 0) {
+    // The menu is the disk: nothing downloaded (or the scan hasn't returned yet).
+    root.innerHTML = state.models_loaded
+      ? `<div class="empty-menu muted">No models downloaded on the Sparks yet. Use <strong>+ Download a new model</strong> above to fetch one — it'll appear here when it's done.</div>`
+      : `<div class="empty-menu muted">Scanning the Sparks for downloaded models…</div>`;
+    return;
+  }
+  for (const key of keys) {
     const m = state.models[key];
     const isActive = key === state.current_model_key;
     const card = document.createElement('div');
-    card.className = 'card' + (isActive ? ' active' : '');
+    card.className = 'card' + (isActive ? ' active' : '') + (m.needs_setup ? ' needs-setup' : '');
     const desc = m.description
       ? `<div class="desc">${escapeHtml(m.description)}</div>`
       : '';
     const customPill = m.custom ? `<span class="tag custom-pill">custom</span>` : '';
-    // Disk-presence pill + trash button. Until /api/models/disk-status comes back,
+    const localPill = m.local_path ? `<span class="tag local-pill" title="Served from a directory on the Spark, not Hugging Face">local</span>` : '';
-    // we don't know — render a neutral placeholder.
+    // Every card on the menu is on disk by definition — show its real size.
-    const disk = state.disk_status[key];
+    const gb = (m.total_bytes || 0) / 1e9;
-    let diskPill = '';
+    const diskPill = gb > 0
-    if (state.disk_status_loaded) {
+      ? `<span class="tag on-disk" title="Weights present on the Spark(s)">on disk · ${gb.toFixed(1)} GB</span>`
-      if (disk && disk.on_disk) {
+      : '';
-        const gb = (disk.total_bytes / 1e9);
+    const setupPill = m.needs_setup
-        diskPill = `<span class="tag on-disk" title="Weights present on disk">on disk · ${gb.toFixed(1)} GB</span>`;
+      ? `<span class="tag setup-pill" title="On disk, but Spark Control hasn't been told how to launch it">needs setup</span>`
-      } else {
+      : '';
-        diskPill = `<span class="tag not-on-disk" title="Weights not downloaded">not downloaded</span>`;
+    // Trash = remove weights from disk AND from the menu. Disabled if active / mid-swap.
-      }
+    // Never offered for local models: their directory is hand-placed training output,
-    }
+    // not a re-downloadable HF cache (the server refuses the delete too).
-    // Trash button — hidden if not on disk; disabled (with tooltip) if currently loaded.
     let trashBtn = '';
-    if (state.disk_status_loaded && disk && disk.on_disk) {
+    if (!m.local_path) {
       const disabled = isActive || isSwapping;
       const tip = isActive
         ? 'Currently loaded — switch to another model first'
         : isSwapping
         ? 'A swap is in progress'
-        : 'Delete weights from disk';
+        : 'Remove weights from disk & menu';
-      trashBtn = `<button class="icon-btn danger" data-disk-del-key="${key}" title="${escapeHtml(tip)}" aria-label="Delete from disk" ${disabled ? 'disabled' : ''}>${trashIcon}</button>`;
+      trashBtn = `<button class="icon-btn danger" data-disk-del-key="${key}" title="${escapeHtml(tip)}" aria-label="Remove from disk and menu" ${disabled ? 'disabled' : ''}>${trashIcon}</button>`;
     }
-    // Primary card action: "Switch to this" (green) when on disk; "Download" (blue) when not.
+    // Primary action: "Current" / "Switch to this", or "Set up & switch" for a
-    // Before disk-status loads we render the swap button as a sensible default.
+    // model on disk that has no launch recipe yet.
-    const isOnDisk = !state.disk_status_loaded || (disk && disk.on_disk);
+    const swapBlocked = isSwapping || locked;
-    const dlInFlight = !!(typeof dlState !== 'undefined' && dlState && dlState.job_id);
+    const lockTipAttr = locked ? ` title="${escapeHtml(lockTip)}"` : '';
     let primaryBtn = '';
     if (isActive) {
       primaryBtn = `<button class="btn" disabled>Current</button>`;
-    } else if (isOnDisk) {
+    } else if (m.needs_setup) {
-      primaryBtn = `<button class="btn primary" data-swap-key="${key}" ${isSwapping ? 'disabled' : ''}>Switch to this</button>`;
+      primaryBtn = `<button class="btn primary" data-setup-key="${key}"${lockTipAttr} ${swapBlocked ? 'disabled' : ''}>Set up &amp; switch</button>`;
     } else {
-      const tip = dlInFlight ? 'A download is already in progress' : 'Download weights to the Spark(s)';
+      primaryBtn = `<button class="btn primary" data-swap-key="${key}"${lockTipAttr} ${swapBlocked ? 'disabled' : ''}>Switch to this</button>`;
-      primaryBtn = `<button class="btn info" data-download-key="${key}" title="${escapeHtml(tip)}" ${dlInFlight ? 'disabled' : ''}>Download</button>`;
     }
+    // The Test/Advanced controls need a saved recipe; hide them until setup is done.
+    const recipeActions = m.needs_setup ? '' : `
+        <button class="btn test-btn" data-test-key="${key}" title="Pre-flight check the launch command without starting the engine">Test</button>
+        <button class="btn adv-btn" data-adv-key="${key}" title="Advanced settings">Advanced</button>`;
     card.innerHTML = `
       <div class="name">${escapeHtml(m.display_name)}</div>
       <div class="meta">
         <span class="tag mode-${m.mode}">${m.mode}</span>
-        <span class="tag">${m.size_gb} GB</span>
-        ${customPill}
         ${diskPill}
+        ${setupPill}
+        ${customPill}
+        ${localPill}
         ${(m.capabilities || []).map(c => `<span class="tag cap">${escapeHtml(c)}</span>`).join('')}
       </div>
       ${desc}
       <div class="muted small repo">
-        <a href="https://huggingface.co/${encodeURIComponent(m.repo)}" target="_blank" rel="noopener" title="View on Hugging Face">${escapeHtml(m.repo)} <span class="hf-icon">↗</span></a>
+        ${m.local_path
+          ? `<span class="local-path" title="Local model directory on the Spark">${escapeHtml(m.local_path)}</span>`
+          : `<a href="https://huggingface.co/${encodeURIComponent(m.repo)}" target="_blank" rel="noopener" title="View on Hugging Face">${escapeHtml(m.repo)} <span class="hf-icon">↗</span></a>`}
       </div>
       <div class="spacer"></div>
       <div class="card-actions">
-        ${primaryBtn}
+        ${primaryBtn}${recipeActions}
-        <button class="btn test-btn" data-test-key="${key}" title="Pre-flight check the launch command without starting the engine">Test</button>
-        <button class="btn adv-btn" data-adv-key="${key}" title="Advanced settings">Advanced</button>
         ${trashBtn}
       </div>
       <div class="test-result hidden" data-test-result-for="${key}"></div>
@@ -123,8 +148,8 @@ function renderCards() {
   for (const btn of root.querySelectorAll('[data-swap-key]')) {
     btn.addEventListener('click', () => triggerSwap(btn.dataset.swapKey));
   }
-  for (const btn of root.querySelectorAll('[data-download-key]')) {
+  for (const btn of root.querySelectorAll('[data-setup-key]')) {
-    btn.addEventListener('click', () => triggerDownloadForKey(btn.dataset.downloadKey));
+    btn.addEventListener('click', () => openSetupForKey(btn.dataset.setupKey));
   }
   for (const btn of root.querySelectorAll('[data-adv-key]')) {
     btn.addEventListener('click', () => openAdvanced(btn.dataset.advKey));
@@ -923,6 +948,10 @@ function renderHealth(status) {
   function setDot(id, ok, payload) {
     const item = el(id);
     if (!item) return;
+    // A service switched off via DISABLED_SERVICES isn't part of this
+    // deployment — hide its indicator entirely rather than show it as down.
+    if (payload && payload.disabled) { item.classList.add('hidden'); return; }
+    item.classList.remove('hidden');
     const dot = item.querySelector('.dot');
     dot.classList.remove('ok', 'bad', 'warn');
     if (ok === true) dot.classList.add('ok');
@@ -1141,24 +1170,44 @@ async function pollStatus() {
   }
 }
 
+let menuLoadInFlight = false;
+
 async function loadModels() {
-  const data = await fetchJSON('/api/models');
+  // The menu is whatever's downloaded on the Sparks — /api/models does the scan
-  state.defaults = data.defaults || {};
+  // (SSH), so this is the slower model call. Best-effort: a transient failure
-  state.models = data.models || {};
+  // leaves the previous menu in place rather than blanking the dashboard.
+  // Guard against overlap: init() fires this un-awaited and pollStatus()'s
+  // empty-menu fallback may call it again before the scan returns.
+  if (menuLoadInFlight) return;
+  menuLoadInFlight = true;
+  try {
+    const data = await fetchJSON('/api/models');
+    state.defaults = data.defaults || {};
+    state.models = data.models || {};
+    state.recipes = data.recipes || [];
+    state.models_loaded = true;
+    populateDownloadSuggestions();
+    renderCards();
+  } catch (e) {
+    console.warn('model menu load failed:', e.message);
+  } finally {
+    menuLoadInFlight = false;
+  }
 }
 
-async function loadDiskStatus() {
+// Populate the download box's autocomplete with known recipes not currently on
-  // Probes each catalog model's HF cache over SSH; takes a beat. Best-effort.
+// disk — so common/bundled models stay discoverable without phantom menu cards.
-  try {
+function populateDownloadSuggestions() {
-    const r = await fetchJSON('/api/models/disk-status');
+  const dl = el('#dl-suggestions');
-    if (r && r.models) {
+  if (!dl) return;
-      state.disk_status = r.models;
+  const onDiskRepos = new Set(Object.values(state.models).map(m => m.repo).filter(Boolean));
-      state.disk_status_loaded = true;
+  dl.innerHTML = '';
-      renderCards();
+  for (const r of state.recipes || []) {
-    }
+    if (onDiskRepos.has(r.repo)) continue;
-  } catch (e) {
+    const opt = document.createElement('option');
-    // Silent — pills just won't render. Don't block dashboard.
+    opt.value = r.repo;
-    console.warn('disk-status probe failed:', e.message);
+    opt.label = `${r.display_name} (${r.mode})`;
+    dl.appendChild(opt);
   }
 }
 
@@ -1172,14 +1221,12 @@ function fmtBytesShort(n) {
 
 function openDiskDeleteDialog(key) {
   const m = state.models[key];
-  const disk = state.disk_status[key];
+  if (!m || !m.on_disk) return;
-  if (!m || !disk || !disk.on_disk) return;
   const dlg = el('#disk-delete-dialog');
-  el('#dd-summary').innerHTML = `Free <strong>${fmtBytesShort(disk.total_bytes)}</strong> by removing <strong>${escapeHtml(m.display_name)}</strong> (<code>${escapeHtml(m.repo)}</code>) from disk.`;
+  el('#dd-summary').innerHTML = `Free <strong>${fmtBytesShort(m.total_bytes)}</strong> by removing <strong>${escapeHtml(m.display_name)}</strong> (<code>${escapeHtml(m.repo)}</code>) from the Sparks. This also takes it off the menu.`;
   const hostsEl = el('#dd-hosts');
   hostsEl.innerHTML = '';
-  for (const h of (disk.per_host || [])) {
+  for (const h of (m.per_host || [])) {
-    if (!h.on_disk) continue;
     const li = document.createElement('li');
     li.innerHTML = `<code>${escapeHtml(h.host)}</code> — ${fmtBytesShort(h.size_bytes)}`;
     hostsEl.appendChild(li);
@@ -1198,20 +1245,19 @@ function openDiskDeleteDialog(key) {
     try {
       const r = await fetchJSON(`/api/models/${encodeURIComponent(key)}/disk`, { method: 'DELETE' });
       dlg.close();
-      // Optimistically clear local disk state for this key, then refresh.
+      // Optimistically drop the card, then re-scan the menu (it's gone from disk).
-      delete state.disk_status[key];
+      delete state.models[key];
       renderCards();
-      // Eagerly re-probe so size is accurate (and shows "not downloaded" pill).
+      await loadModels();
-      loadDiskStatus();
       const freed = r && typeof r.bytes_freed === 'number' ? fmtBytesShort(r.bytes_freed) : '';
-      console.log(`Deleted ${m.display_name} from disk${freed ? ` — freed ${freed}` : ''}.`);
+      console.log(`Removed ${m.display_name} from disk${freed ? ` — freed ${freed}` : ''}.`);
     } catch (e) {
       errEl.textContent = e.message || 'Delete failed';
       errEl.classList.remove('hidden');
     } finally {
       confirm.disabled = false;
       cancel.disabled = false;
-      confirm.textContent = 'Delete from disk';
+      confirm.textContent = 'Remove from disk & menu';
     }
   };
   cancel.onclick = onCancel;
@@ -1221,6 +1267,11 @@ function openDiskDeleteDialog(key) {
 
 async function triggerSwap(modelKey) {
   if (state.swap_job_id) return;
+  if (state.lock && state.lock.held) {
+    const until = state.lock.expires_at ? ' until ' + fmtClock(state.lock.expires_at) : '';
+    alert(`The GPU swap path is reserved by ${state.lock.holder || 'automation'}${until}. Use "Release" on the reservation banner to override.`);
+    return;
+  }
   try {
     const r = await fetchJSON('/api/swap', {
       method: 'POST',
@@ -1229,42 +1280,84 @@ async function triggerSwap(modelKey) {
     });
     attachToSwap(r.job_id, /*needsBackfill=*/false);
   } catch (e) {
-    alert('Failed to start swap: ' + e.message);
+    // 423 Locked: a reservation was acquired between our last poll and this click.
+    if (e.message && e.message.startsWith('423')) {
+      alert('The GPU swap path was just reserved by automation. Refreshing…');
+      pollCoordination();
+    } else {
+      alert('Failed to start swap: ' + e.message);
+    }
   }
 }
 
-async function triggerDownloadForKey(modelKey) {
+// ---- coordination layer: swap lock + schedule registry ----
-  const m = state.models[modelKey];
+
-  if (!m) return;
+async function pollCoordination() {
-  if (dlState.job_id) {
-    alert('A download is already in progress; wait for it to finish.');
-    return;
-  }
-  // Pick the download target from the model's mode:
-  //   solo    -> spark1 only
-  //   cluster -> both Sparks (fetch on Spark 1, rsync to Spark 2 in parallel)
-  const dlMode = m.mode === 'cluster' ? 'cluster' : 'spark1';
-  const sizeNote = m.size_gb ? ` (~${m.size_gb} GB)` : '';
-  const target = m.mode === 'cluster' ? 'both Sparks' : 'Spark 1';
-  if (!confirm(`Download "${m.display_name}"${sizeNote} to ${target}? Large models can take a while; you can watch progress in the download panel.`)) {
-    return;
-  }
-  dlState.last_repo = m.repo;
-  dlState.last_mode = dlMode;
   try {
-    const r = await fetchJSON('/api/download', {
+    state.lock = await fetchJSON('/api/swap/lock');
-      method: 'POST',
+  } catch { state.lock = { held: false }; }
-      headers: { 'content-type': 'application/json' },
+  try {
-      body: JSON.stringify({ repo: m.repo, mode: dlMode }),
+    const r = await fetchJSON('/api/schedule');
-    });
+    state.schedules = r.schedules || [];
-    // Open the download panel + attach to progress stream
+  } catch { state.schedules = []; }
-    openDownloadForm();
+  renderLockBanner();
-    attachToDownload(r.job_id);
+  renderSchedules();
-  } catch (e) {
+  renderCards();  // reflect lock state on the swap buttons
-    alert('Failed to start download: ' + e.message);
+}
+
+function renderLockBanner() {
+  const banner = el('#lock-banner');
+  if (!banner) return;
+  const lock = state.lock;
+  if (lock && lock.held) {
+    const until = lock.expires_at ? ` until ${fmtClock(lock.expires_at)}` : '';
+    const note = lock.note ? ` — ${escapeHtml(lock.note)}` : '';
+    el('#lock-text').innerHTML =
+      `GPU swap path reserved by <strong>${escapeHtml(lock.holder || 'automation')}</strong>${until}${note}. Manual swaps are paused.`;
+    banner.classList.remove('hidden');
+  } else {
+    banner.classList.add('hidden');
   }
 }
 
+function renderSchedules() {
+  const panel = el('#schedule-panel');
+  const list = el('#schedule-list');
+  if (!panel || !list) return;
+  const items = state.schedules || [];
+  if (!items.length) {
+    panel.classList.add('hidden');
+    list.innerHTML = '';
+    return;
+  }
+  list.innerHTML = items.map((s) => {
+    const meta = [
+      s.cron ? `<code>${escapeHtml(s.cron)}</code>` : '',
+      s.next_run ? `next: ${escapeHtml(s.next_run)}` : '',
+      s.owner ? `by ${escapeHtml(s.owner)}` : '',
+    ].filter(Boolean).join(' · ');
+    const desc = s.description ? `<div class="desc">${escapeHtml(s.description)}</div>` : '';
+    return `<div class="schedule-item">
+      <div class="name">${escapeHtml(s.name)}</div>
+      <div class="muted small">${meta}</div>
+      ${desc}
+    </div>`;
+  }).join('');
+  panel.classList.remove('hidden');
+}
+
+async function releaseLock() {
+  const lock = state.lock || {};
+  const who = lock.holder || 'automation';
+  if (!confirm(`Force-release the GPU reservation held by ${who}? Any job relying on it may then collide with a manual swap.`)) return;
+  try {
+    await fetchJSON('/api/swap/lock?force=true', { method: 'DELETE' });
+  } catch (e) {
+    alert('Failed to release: ' + e.message);
+  }
+  pollCoordination();
+}
+
 async function attachToSwap(jobId, needsBackfill) {
   if (state.swap_eventsource) {
     state.swap_eventsource.close();
@@ -1495,12 +1588,14 @@ function handleDownloadDone(d) {
     el('#dl-title').textContent = 'Done';
     el('#dl-phase').textContent = 'Done ✓';
     el('#dl-progress-fill').style.width = '100%';
-    // Offer to add to catalog
+    // The new model now appears on the menu (the menu is the disk). If it matched
+    // a known recipe it's ready to switch to; if not, offer to set it up.
     const repo = dlState.last_repo;
-    const mode = dlState.last_mode;
+    loadModels().then(() => {
-    if (repo) {
+      if (!repo) return;
-      setTimeout(() => openCatalogDialog(repo, mode), 600);
+      const entry = Object.values(state.models).find(m => m.repo === repo);
-    }
+      if (entry && entry.needs_setup) setTimeout(() => openSetupDialog(repo, { thenSwap: false }), 600);
+    });
   }
   dlState.job_id = null;
 }
@@ -1613,21 +1708,67 @@ function openAdvanced(key) {
   dlg.showModal();
 }
 
-function openCatalogDialog(repo, mode) {
+// Context carried from openSetupDialog -> the submit handler: the inferred
+// launch flags (parsers/MoE backend) and whether to swap right after saving.
+let setupCtx = { key: '', repo: '', vllm_args: [], thenSwap: false };
+
+// "Set up & switch" on a needs-setup card.
+async function openSetupForKey(key) {
+  const m = state.models[key];
+  if (!m) return;
+  if (state.lock && state.lock.held) {
+    const until = state.lock.expires_at ? ' until ' + fmtClock(state.lock.expires_at) : '';
+    alert(`The GPU swap path is reserved by ${state.lock.holder || 'automation'}${until}. Use "Release" on the reservation banner to override.`);
+    return;
+  }
+  await openSetupDialog(m.repo, { thenSwap: true });
+}
+
+// Open the "set up this model" dialog, prefilled from inference (config.json +
+// size). The operator confirms once; on save the recipe persists and (if
+// thenSwap) we switch to it.
+async function openSetupDialog(repo, opts = {}) {
   const dlg = el('#catalog-dialog');
-  const key = repo.split('/').pop().toLowerCase().replace(/[^a-z0-9_-]/g, '-');
+  let sug = null;
-  el('#cd-key').value = key;
+  try {
-  el('#cd-name').value = repo.split('/').pop();
+    sug = await fetchJSON(`/api/models/suggest?repo=${encodeURIComponent(repo)}`);
+  } catch (e) {
+    console.warn('recipe suggestion failed:', e.message);
+  }
+  const fallbackKey = repo.toLowerCase().replace(/[^a-z0-9_-]+/g, '-').replace(/^-+|-+$/g, '');
+  setupCtx = {
+    key: (sug && sug.key) || fallbackKey,
+    repo,
+    vllm_args: (sug && sug.vllm_args) || [],
+    thenSwap: !!opts.thenSwap,
+  };
+  el('#cd-key').value = setupCtx.key;
+  el('#cd-name').value = (sug && sug.display_name) || repo.split('/').pop();
   el('#cd-repo').value = repo;
   el('#cd-size').value = '';
-  el('#cd-mode').value = mode || 'solo';
+  el('#cd-mode').value = (sug && sug.mode) || 'solo';
   el('#cd-desc').value = '';
-  el('#cd-mml').value = 32768;
+  const knobs = (sug && sug.knobs) || {};
-  el('#cd-gmu').value = 0.85;
+  el('#cd-mml').value = knobs.max_model_len || 32768;
-  el('#cd-gmu-out').value = '0.85';
+  el('#cd-gmu').value = knobs.gpu_memory_utilization || 0.85;
-  el('#cd-fst').checked = true;
+  el('#cd-gmu-out').value = parseFloat(el('#cd-gmu').value).toFixed(2);
-  el('#cd-pcache').checked = true;
+  el('#cd-fst').checked = knobs.fastsafetensors !== false;
-  el('#cd-fp8').checked = true;
+  el('#cd-pcache').checked = knobs.prefix_caching !== false;
+  el('#cd-fp8').checked = (knobs.kv_cache_dtype || 'fp8') === 'fp8';
+
+  const det = el('#cd-detected');
+  if (det) {
+    if (sug) {
+      const caps = (sug.capabilities || []).join(', ');
+      const flags = setupCtx.vllm_args.length ? `: <code>${escapeHtml(setupCtx.vllm_args.join(' '))}</code>` : '';
+      det.innerHTML = `Detected <strong>${escapeHtml(sug.family || 'Generic')}</strong>${caps ? ` · ${escapeHtml(caps)}` : ''}. Launch flags set automatically${flags}.`;
+    } else {
+      det.textContent = "Couldn't auto-detect this model's settings — pick mode and knobs manually.";
+    }
+    det.classList.remove('hidden');
+  }
+  const submit = el('#cd-submit');
+  if (submit) submit.textContent = setupCtx.thenSwap ? 'Save & switch' : 'Save settings';
   dlg.showModal();
 }
 
@@ -1637,13 +1778,15 @@ function setupCatalogDialog() {
   el('#catalog-form').addEventListener('submit', async (e) => {
     e.preventDefault();
     const body = {
-      key: el('#cd-key').value.trim(),
+      key: el('#cd-key').value.trim() || setupCtx.key,
       display_name: el('#cd-name').value.trim(),
       repo: el('#cd-repo').value.trim(),
       size_gb: parseFloat(el('#cd-size').value) || 0,
       mode: el('#cd-mode').value,
       description: el('#cd-desc').value.trim() || null,
-      vllm_args: [],
+      // The inferred family flags (parsers / MoE backend); knob-controlled flags
+      // are layered on by the server from `knobs`, so no duplication.
+      vllm_args: setupCtx.vllm_args || [],
       knobs: {
         max_model_len: parseInt(el('#cd-mml').value, 10) || 32768,
         gpu_memory_utilization: parseFloat(el('#cd-gmu').value),
@@ -1661,8 +1804,9 @@ function setupCatalogDialog() {
       el('#catalog-dialog').close();
       closeDownloadPanel();
       await loadModels();
+      if (setupCtx.thenSwap) triggerSwap(body.key);
       pollStatus();
-    } catch (e) { alert('Add to catalog failed: ' + e.message); }
+    } catch (e) { alert('Saving the model setup failed: ' + e.message); }
   });
 }
 
@@ -1671,6 +1815,60 @@ function setupAdvancedDialog() {
   el('#adv-gmu').addEventListener('input', (e) => { el('#adv-gmu-out').value = parseFloat(e.target.value).toFixed(2); });
 }
 
+function openLocalModelDialog() {
+  const dlg = el('#local-model-dialog');
+  el('#lm-key').value = '';
+  el('#lm-name').value = '';
+  el('#lm-path').value = '';
+  el('#lm-chat').value = '';
+  el('#lm-size').value = '';
+  el('#lm-mode').value = 'solo';
+  el('#lm-desc').value = '';
+  el('#lm-mml').value = 32768;
+  el('#lm-gmu').value = 0.85;
+  el('#lm-gmu-out').value = '0.85';
+  el('#lm-fst').checked = true;
+  el('#lm-pcache').checked = true;
+  el('#lm-fp8').checked = true;
+  dlg.showModal();
+}
+
+function setupLocalModelDialog() {
+  el('#lm-cancel').addEventListener('click', () => el('#local-model-dialog').close());
+  el('#lm-gmu').addEventListener('input', (e) => { el('#lm-gmu-out').value = parseFloat(e.target.value).toFixed(2); });
+  el('#local-model-form').addEventListener('submit', async (e) => {
+    e.preventDefault();
+    const chat = el('#lm-chat').value.trim();
+    const body = {
+      key: el('#lm-key').value.trim(),
+      display_name: el('#lm-name').value.trim(),
+      local_path: el('#lm-path').value.trim(),
+      size_gb: parseFloat(el('#lm-size').value) || 0,
+      mode: el('#lm-mode').value,
+      description: el('#lm-desc').value.trim() || null,
+      // A fine-tune's chat template (if any) rides along as a launch flag.
+      vllm_args: chat ? [`--chat-template=${chat}`] : [],
+      knobs: {
+        max_model_len: parseInt(el('#lm-mml').value, 10) || 32768,
+        gpu_memory_utilization: parseFloat(el('#lm-gmu').value),
+        fastsafetensors: el('#lm-fst').checked,
+        prefix_caching: el('#lm-pcache').checked,
+        kv_cache_dtype: el('#lm-fp8').checked ? 'fp8' : 'auto',
+      },
+    };
+    try {
+      await fetchJSON('/api/models', {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify(body),
+      });
+      el('#local-model-dialog').close();
+      await loadModels();
+      pollStatus();
+    } catch (e) { alert('Add local model failed: ' + e.message); }
+  });
+}
+
 // ===================== NIM installer =====================
 
 const nimState = {
@@ -2034,8 +2232,11 @@ async function init() {
     if (kbtn) { copySparkSshKey(kbtn.dataset.sshKey, kbtn); return; }
   });
   el('#sshkey-close').addEventListener('click', () => el('#sshkey-dialog').close());
+  el('#open-local').addEventListener('click', openLocalModelDialog);
+  el('#lock-release').addEventListener('click', releaseLock);
   setupCatalogDialog();
   setupAdvancedDialog();
+  setupLocalModelDialog();
   // Open WebUI link from /api/config
   try {
     state.config = await fetchJSON('/api/config');
@@ -2047,19 +2248,22 @@ async function init() {
   } catch {}
   setupDashboardTabs();
   setupEndpointCollapse();
-  await loadModels();
+  // Fire the (SSH-backed) menu scan without awaiting — it self-renders a
+  // "Scanning…" state and fills in when it returns, so a slow/unreachable
+  // cluster never blocks first paint. pollStatus() below paints the rest.
+  loadModels();
   await pollStatus();
   await renderServices();
+  pollCoordination();
   pollHardware();
   pollUpdates();
-  // Disk-status probe runs after first paint — slow over SSH and not blocking.
-  loadDiskStatus();
   // Speech-model patches panel — slow over SSH, runs after first paint.
   renderSpeechModels();
   setInterval(pollStatus, 5000);
+  setInterval(pollCoordination, 5000); // swap lock + schedule registry
   setInterval(pollHardware, 8000);    // every 8s
   setInterval(pollUpdates, 300000);  // every 5 min
-  setInterval(loadDiskStatus, 60000); // every 60s — disk state changes rarely
+  setInterval(loadModels, 60000); // every 60s — re-scan the Sparks for added/removed models
   setInterval(renderSpeechModels, 120000); // every 2 min — patches change rarely
 }
 

image/app/static/index.html

@@ -96,6 +96,13 @@
       </details>
     </section>
 
+    <section id="lock-banner" class="banner lock-banner hidden">
+      <span class="lock-icon" aria-hidden="true">🔒</span>
+      <span id="lock-text">GPU swap path reserved</span>
+      <span class="spacer"></span>
+      <button id="lock-release" class="btn small-btn">Release</button>
+    </section>
+
     <nav id="dashboard-tabs" class="dashboard-tabs hidden" role="tablist">
       <button type="button" class="dashboard-tab" data-tab="llm" role="tab" aria-selected="true">LLM</button>
       <button type="button" class="dashboard-tab" data-tab="audio" role="tab" aria-selected="false">Audio / Speech</button>
@@ -229,13 +236,15 @@
       <div class="section-header">
         <h2 class="section-title">LLM swap</h2>
         <button id="open-download" class="btn small-btn">+ Download a new model</button>
+        <button id="open-local" class="btn small-btn">+ Add local model</button>
       </div>
 
       <dialog id="catalog-dialog" class="modal">
         <form method="dialog" class="modal-form" id="catalog-form">
-          <h3>Add downloaded model to catalog</h3>
+          <h3>Set up this model</h3>
-          <p class="muted small">It will appear as a new card you can swap to. Knob values become its default launch flags — you can tweak later via the model's "Advanced" panel.</p>
+          <p class="muted small">This model is downloaded, but Spark Control needs to know how to launch it. We've guessed from the model's own files — confirm or adjust, and it's saved so you're never asked again.</p>
-          <label class="modal-row"><span>Key (URL-safe id)</span><input type="text" id="cd-key" required pattern="[a-zA-Z0-9_-]+"></label>
+          <p id="cd-detected" class="muted small cd-detected hidden"></p>
+          <label class="modal-row"><span>Key (URL-safe id)</span><input type="text" id="cd-key" required pattern="[a-zA-Z0-9_-]+" readonly></label>
           <label class="modal-row"><span>Display name</span><input type="text" id="cd-name" required></label>
           <label class="modal-row"><span>Repo (read-only)</span><input type="text" id="cd-repo" readonly></label>
           <label class="modal-row"><span>Size (GB)</span><input type="number" id="cd-size" step="0.1" min="0"></label>
@@ -256,21 +265,52 @@
           </fieldset>
           <div class="modal-actions">
             <button type="button" id="cd-cancel" class="btn">Cancel</button>
-            <button type="submit" class="btn primary">Add to catalog</button>
+            <button type="submit" id="cd-submit" class="btn primary">Save settings</button>
+          </div>
+        </form>
+      </dialog>
+
+      <dialog id="local-model-dialog" class="modal">
+        <form method="dialog" class="modal-form" id="local-model-form">
+          <h3>Add a local / fine-tuned model</h3>
+          <p class="muted small">For a model that lives as a directory on a Spark (e.g. a fine-tune), not a Hugging Face repo. The directory is bind-mounted into the vLLM container at the same path when you swap to it. It must already exist on the Spark.</p>
+          <label class="modal-row"><span>Key (URL-safe id)</span><input type="text" id="lm-key" required pattern="[a-zA-Z0-9_-]+"></label>
+          <label class="modal-row"><span>Display name</span><input type="text" id="lm-name" required></label>
+          <label class="modal-row"><span>Model directory (absolute path on the Spark)</span><input type="text" id="lm-path" required placeholder="e.g. /home/you/models/my-finetune"></label>
+          <label class="modal-row"><span>Chat template path (optional)</span><input type="text" id="lm-chat" placeholder="e.g. /home/you/models/my-finetune/chat_template.jinja"></label>
+          <label class="modal-row"><span>Size (GB)</span><input type="number" id="lm-size" step="0.1" min="0"></label>
+          <label class="modal-row"><span>Mode</span>
+            <select id="lm-mode">
+              <option value="solo">solo (Spark 1 only)</option>
+              <option value="cluster">cluster (both Sparks via Ray)</option>
+            </select>
+          </label>
+          <label class="modal-row"><span>Description (optional)</span><textarea id="lm-desc" rows="3"></textarea></label>
+          <fieldset class="modal-fieldset">
+            <legend>Default launch knobs</legend>
+            <label class="modal-row"><span>Max context (tokens)</span><input type="number" id="lm-mml" step="1024" min="1024" value="32768"></label>
+            <label class="modal-row"><span>GPU memory %</span><input type="range" id="lm-gmu" min="0.5" max="0.95" step="0.01" value="0.85"> <output id="lm-gmu-out">0.85</output></label>
+            <label class="modal-row inline"><input type="checkbox" id="lm-fst" checked> Fast safetensors loading</label>
+            <label class="modal-row inline"><input type="checkbox" id="lm-pcache" checked> Prefix caching</label>
+            <label class="modal-row inline"><input type="checkbox" id="lm-fp8" checked> FP8 KV cache</label>
+          </fieldset>
+          <div class="modal-actions">
+            <button type="button" id="lm-cancel" class="btn">Cancel</button>
+            <button type="submit" class="btn primary">Add local model</button>
           </div>
         </form>
       </dialog>
 
       <dialog id="disk-delete-dialog" class="modal">
         <form method="dialog" class="modal-form">
-          <h3>Delete model weights from disk?</h3>
+          <h3>Remove this model from the Sparks?</h3>
           <p id="dd-summary" class="muted small"></p>
           <ul class="muted small dd-hosts" id="dd-hosts"></ul>
-          <p class="muted small">This is reversible — you can re-download from the catalog at any time. The catalog entry stays intact.</p>
+          <p class="muted small">This deletes the weights and removes the card from the menu. You can always download it again later (re-downloading restores its saved settings).</p>
           <p id="dd-error" class="muted small dd-error hidden"></p>
           <div class="modal-actions">
             <button type="button" id="dd-cancel" class="btn">Cancel</button>
-            <button type="button" id="dd-confirm" class="btn danger">Delete from disk</button>
+            <button type="button" id="dd-confirm" class="btn danger">Remove from disk &amp; menu</button>
           </div>
         </form>
       </dialog>
@@ -315,11 +355,12 @@
         <div class="download-form" id="download-form">
           <label class="dl-row">
             <span class="dl-label">HuggingFace repo</span>
-            <input type="text" id="dl-repo" placeholder="e.g. RedHatAI/Qwen3.6-35B-A3B-NVFP4" autocomplete="off">
+            <input type="text" id="dl-repo" placeholder="e.g. RedHatAI/Qwen3.6-35B-A3B-NVFP4" autocomplete="off" list="dl-suggestions">
+            <datalist id="dl-suggestions"></datalist>
             <a id="dl-hf-link" class="dl-hf-link hidden" href="#" target="_blank" rel="noopener" title="Open on Hugging Face">↗</a>
           </label>
           <div class="dl-help muted small">
-            <a href="https://huggingface.co/models?other=vllm" target="_blank" rel="noopener">Browse vLLM-compatible models</a>
+            Type any repo, or pick a known one from the list. <a href="https://huggingface.co/models?other=vllm" target="_blank" rel="noopener">Browse vLLM-compatible models</a>
             · NVFP4-quantized models (e.g. <code>RedHatAI/...</code>) are best for Blackwell hardware
           </div>
           <div class="dl-row">
@@ -362,6 +403,14 @@
       <section id="cards" class="cards"></section>
     </section>
 
+    <section id="schedule-panel" class="schedule-panel hidden">
+      <div class="section-header">
+        <h2 class="section-title">Scheduled jobs</h2>
+      </div>
+      <p class="muted small">Registered by your own automation. Spark Control only displays these — it doesn't run them.</p>
+      <div id="schedule-list" class="schedule-list"></div>
+    </section>
+
     <section id="update-banner" class="update-banner hidden">
       <div class="ub-context muted small">
         Updates to <strong><a href="https://github.com/eugr/spark-vllm-docker" target="_blank" rel="noopener">eugr/spark-vllm-docker</a></strong>

image/app/static/style.css

@@ -74,6 +74,42 @@ main {
 }
 .banner em { font-style: normal; background: rgba(245, 158, 11, 0.15); padding: 2px 6px; border-radius: 4px; }
 
+/* GPU swap reservation (coordination layer) — informational, not a warning. */
+.lock-banner {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  border-color: var(--info);
+  color: var(--info);
+}
+.lock-banner .lock-icon { font-size: 16px; }
+.lock-banner strong { color: var(--text); }
+.lock-banner .spacer { flex: 1; }
+
+/* Scheduled-jobs panel — read-only view of what external automation registered. */
+.schedule-panel { margin-top: 8px; }
+.schedule-list {
+  display: grid;
+  grid-template-columns: repeat(auto-fill, minmax(240px, 1fr));
+  gap: 12px;
+  margin-top: 8px;
+}
+.schedule-item {
+  background: var(--surface);
+  border: 1px solid var(--border);
+  border-radius: var(--radius);
+  padding: 12px 14px;
+}
+.schedule-item .name { font-weight: 600; margin-bottom: 4px; }
+.schedule-item code {
+  background: var(--surface-2);
+  border: 1px solid var(--border);
+  border-radius: 4px;
+  padding: 1px 5px;
+  font-size: 12px;
+}
+.schedule-item .desc { margin-top: 6px; color: var(--muted); font-size: 13px; }
+
 /* ===== Endpoint panel ===== */
 
 .endpoint-panel {
@@ -694,6 +730,7 @@ main {
 .card .repo a { color: inherit; text-decoration: none; }
 .card .repo a:hover { color: var(--info); text-decoration: underline; }
 .card .repo .hf-icon { font-size: 13px; opacity: 0.7; }
+.card .repo .local-path { font-family: var(--mono, ui-monospace, monospace); opacity: 0.85; }
 .tag {
   background: var(--surface-2);
   border: 1px solid var(--border);
@@ -738,8 +775,15 @@ main {
 .card .adv-btn,
 .card .test-btn { padding: 8px 12px; font-size: 12px; }
 .card .custom-pill { color: var(--info); border-color: rgba(96, 165, 250, 0.4); }
+.card .local-pill { color: var(--warn); border-color: rgba(245, 158, 11, 0.4); }
 .tag.on-disk { color: var(--accent); border-color: rgba(74, 222, 128, 0.4); }
 .tag.not-on-disk { color: var(--muted); border-color: var(--border); opacity: 0.7; }
+.tag.setup-pill { color: var(--warn); border-color: rgba(245, 158, 11, 0.4); }
+.card.needs-setup { border-style: dashed; }
+.card-actions .btn[data-setup-key] { flex: 1; }
+.empty-menu { grid-column: 1 / -1; padding: 28px 16px; text-align: center; border: 1px dashed var(--border); border-radius: 10px; }
+.cd-detected { padding: 8px 10px; border: 1px solid var(--border); border-radius: 8px; background: rgba(255,255,255,0.02); }
+.cd-detected code { word-break: break-all; }
 .card-actions .icon-btn.danger { color: var(--error); border-color: rgba(239, 68, 68, 0.3); margin-left: auto; }
 .card-actions .icon-btn.danger:hover:not(:disabled) { background: rgba(239, 68, 68, 0.08); border-color: var(--error); color: var(--error); }
 .card-actions .icon-btn.danger:disabled { opacity: 0.35; cursor: not-allowed; }

image/app/swap.py

@@ -6,7 +6,9 @@ from datetime import datetime, timezone
 from typing import Optional
 
 from .config import Settings
+from .coordination import WebhookNotifier, build_webhook_payload
 from .models import Catalog, build_launch_command
+from .shellsafe import quote_arg
 from .ssh import ssh_run, ssh_stream, StreamHandle
 
 
@@ -32,9 +34,15 @@ class SwapJob:
 
 
 class SwapManager:
-    def __init__(self, settings: Settings, catalog: Catalog) -> None:
+    def __init__(
+        self,
+        settings: Settings,
+        catalog: Catalog,
+        notifier: Optional[WebhookNotifier] = None,
+    ) -> None:
         self.settings = settings
         self.catalog = catalog
+        self.notifier = notifier
         self.lock = asyncio.Lock()
         self.jobs: dict[str, SwapJob] = {}
         self.current_job_id: Optional[str] = None
@@ -77,6 +85,21 @@ class SwapManager:
                 job.finished_at = datetime.now(timezone.utc).isoformat()
                 if self.current_job_id == job.id:
                     self.current_job_id = None
+        # Outside the swap lock (so a webhook POST can't stall a queued swap) and
+        # only for real swaps — a dry run never changes the running model. A
+        # webhook failure is logged inside fire(), never raised.
+        if self.notifier is not None and self.notifier.enabled and not job.dry_run:
+            event = "swap_complete" if job.state == "ready" else "swap_failed"
+            await self.notifier.fire(event, build_webhook_payload(
+                event=event,
+                job_id=job.id,
+                model_key=job.model_key,
+                state=job.state,
+                returncode=job.returncode,
+                started_at=job.started_at,
+                finished_at=job.finished_at,
+                dry_run=job.dry_run,
+            ))
 
     async def _do(self, job: SwapJob) -> None:
         model = self.catalog.models[job.model_key]
@@ -112,7 +135,7 @@ class SwapManager:
 
         # Step 3: tail logs until the ready marker (or timeout)
         job.state = "tailing"
-        tail_cmd = "docker logs -f --tail 50 vllm_node"
+        tail_cmd = f"docker logs -f --tail 50 {quote_arg(s.vllm_container)}"
         job.append(f"$ {tail_cmd}")
         timeout = max(model.expected_ready_seconds * 2, 600)
         handle = StreamHandle()

image/app/validate.py

@@ -22,6 +22,7 @@ from typing import Any
 
 from .config import Settings
 from .models import Catalog, build_launch_command
+from .shellsafe import quote_arg
 from .ssh import ssh_run
 
 
@@ -114,7 +115,7 @@ async def validate_launch(key: str, catalog: Catalog, settings: Settings) -> dic
     # Pipe the JSON args list to a here-doc Python invocation. The validator
     # reads from stdin to avoid shell-escaping the args themselves.
     cmd = (
-        f"echo '{payload}' | docker exec -i vllm_node python3 -c "
+        f"echo '{payload}' | docker exec -i {quote_arg(settings.vllm_container)} python3 -c "
         + shlex.quote(_VALIDATOR_SCRIPT)
     )
 

image/models.yaml

@@ -1,9 +1,14 @@
-# spark-control model catalog
+# spark-control launch recipes
 #
-# Edit this file (or override at runtime via the StartOS "Edit Model Catalog"
+# These are NOT the dashboard menu. The menu is whatever is actually downloaded
-# action) to add or change available models.
+# on the Sparks — Spark Control scans the Hugging Face cache on each load and
+# shows what it finds. These entries are launch *recipes*: matched to an on-disk
+# model by `repo`, they say HOW to launch it. A downloaded model with no recipe
+# here shows up as "needs setup", and the dashboard infers + saves one on first
+# use (from the model's own config.json). Add a recipe to make a known model
+# launch correctly the moment it's downloaded, with no setup prompt.
 #
-# Each model entry produces this command on Spark 1:
+# Each recipe produces this command on Spark 1:
 #   cd ~/spark-vllm-docker
 #   ./launch-cluster.sh [--solo] -d exec vllm serve <repo> \
 #     --port=<defaults.port> --host=<defaults.host> <vllm_args...>
@@ -54,6 +59,34 @@ models:
       - --enable-prefix-caching
       - --kv-cache-dtype=fp8
 
+  gemma4-26b:
+    display_name: "Gemma 4 26B-A4B (vision, light)"
+    description: >-
+      Lighter, faster sibling of the Gemma 4 31B above: a Mixture-of-Experts
+      model with 26B total parameters but only ~4B active per token, so it
+      generates quickly. Takes images as well as text (good for tasks like
+      reading a business card into structured text). Reasoning is a bit
+      shallower than the dense 31B. Runs solo on one Spark.
+    repo: nvidia/Gemma-4-26B-A4B-NVFP4
+    size_gb: 17
+    mode: solo
+    capabilities: [vision, reasoning, tools]
+    expected_ready_seconds: 240
+    vllm_args:
+      - --gpu-memory-utilization=0.8
+      - --max-model-len=32768
+      - --max-num-batched-tokens=16384
+      - --reasoning-parser=gemma4
+      - --tool-call-parser=gemma4
+      - --enable-auto-tool-choice
+      # MoE backend: research found this model's expert layers fall back to
+      # 'marlin' on GB10 (the fast flashinfer_cutlass path errors on sm_121).
+      # If a swap fails to start, this flag is the first thing to flip.
+      - --moe_backend=marlin
+      - --load-format=fastsafetensors
+      - --enable-prefix-caching
+      - --kv-cache-dtype=fp8
+
   qwen36:
     display_name: "Qwen3.6 35B-A3B (daily driver)"
     description: >-
@@ -74,36 +107,3 @@ models:
       - --load-format=fastsafetensors
       - --enable-prefix-caching
       - --kv-cache-dtype=fp8
-
-  qwen3-235b-fp8:
-    display_name: "Qwen3 235B-A22B FP8 (legacy)"
-    description: >-
-      Earlier generation of the Qwen 235B family in native FP8 precision.
-      Runs across both Sparks. Mostly superseded by Qwen3-VL above; keep
-      around for text-only baseline comparisons.
-    repo: Qwen/Qwen3-235B-A22B-FP8
-    size_gb: 220
-    mode: cluster
-    capabilities: []
-    expected_ready_seconds: 360
-    vllm_args:
-      - --gpu-memory-utilization=0.7
-      - -tp=2
-      - --distributed-executor-backend=ray
-      - --max-model-len=32768
-
-  qwen25-72b:
-    display_name: "Qwen2.5 72B (legacy)"
-    description: >-
-      Last-generation 72B dense model. Cluster mode required due to size.
-      Kept for compatibility and baseline comparison against newer Qwens.
-    repo: Qwen/Qwen2.5-72B-Instruct
-    size_gb: 145
-    mode: cluster
-    capabilities: []
-    expected_ready_seconds: 360
-    vllm_args:
-      - --gpu-memory-utilization=0.7
-      - -tp=2
-      - --distributed-executor-backend=ray
-      - --max-model-len=32768

image/tests/test_coordination.py

@@ -0,0 +1,201 @@
+"""Coordination layer: swap lock lifecycle/expiry, schedule registry CRUD, and
+the webhook payload+signature. All offline — the lock takes an injectable `now`
+so expiry is tested without sleeping, and the webhook is exercised only on the
+disabled (no-network) path plus its pure payload/signature helpers.
+"""
+import asyncio
+from datetime import datetime, timedelta, timezone
+
+import pytest
+
+from app.coordination import (
+    LOCK_TTL_MAX,
+    LOCK_TTL_MIN,
+    LockHeld,
+    ScheduleRegistry,
+    SwapLockManager,
+    WebhookNotifier,
+    build_webhook_payload,
+    sign_payload,
+    valid_schedule_id,
+)
+
+T0 = datetime(2026, 6, 17, 12, 0, 0, tzinfo=timezone.utc)
+
+
+# ----------------------------------------------------------------- swap lock ----
+
+def test_acquire_free_lock_returns_token_and_status_held():
+    mgr = SwapLockManager()
+    lock = mgr.acquire("openclaw", ttl_seconds=60, note="daily vol", now=T0)
+    assert lock.token
+    st = mgr.status(now=T0)
+    assert st["held"] is True
+    assert st["holder"] == "openclaw"
+    assert st["note"] == "daily vol"
+    assert st["seconds_remaining"] == 60
+    assert "token" not in st  # public view never leaks the token
+
+
+def test_acquire_requires_holder():
+    with pytest.raises(ValueError):
+        SwapLockManager().acquire("  ", now=T0)
+
+
+def test_acquire_held_by_other_raises_lockheld_with_state():
+    mgr = SwapLockManager()
+    mgr.acquire("openclaw", ttl_seconds=60, now=T0)
+    with pytest.raises(LockHeld) as ei:
+        mgr.acquire("johnny5", ttl_seconds=60, now=T0)
+    assert ei.value.state["holder"] == "openclaw"
+
+
+def test_reacquire_with_token_extends_and_keeps_token():
+    mgr = SwapLockManager()
+    first = mgr.acquire("openclaw", ttl_seconds=60, now=T0)
+    later = T0 + timedelta(seconds=30)
+    second = mgr.acquire("openclaw", ttl_seconds=60, token=first.token, now=later)
+    assert second.token == first.token
+    # window extended from the later moment, not the original
+    assert mgr.status(now=later)["seconds_remaining"] == 60
+    assert second.acquired_at == first.acquired_at  # acquired_at preserved
+
+
+def test_reacquire_without_token_is_refused_even_for_same_holder_name():
+    # Holder name is descriptive, not a secret — matching it must not grant access.
+    mgr = SwapLockManager()
+    mgr.acquire("openclaw", ttl_seconds=60, now=T0)
+    with pytest.raises(LockHeld):
+        mgr.acquire("openclaw", ttl_seconds=60, now=T0)
+
+
+def test_ttl_is_clamped():
+    mgr = SwapLockManager()
+    mgr.acquire("a", ttl_seconds=0, now=T0)
+    assert mgr.status(now=T0)["seconds_remaining"] == LOCK_TTL_MIN
+    mgr2 = SwapLockManager()
+    mgr2.acquire("b", ttl_seconds=10**9, now=T0)
+    assert mgr2.status(now=T0)["seconds_remaining"] == LOCK_TTL_MAX
+
+
+def test_lock_expires_and_clears_lazily():
+    mgr = SwapLockManager()
+    tok = mgr.acquire("openclaw", ttl_seconds=10, now=T0).token
+    after = T0 + timedelta(seconds=11)
+    assert mgr.status(now=after) == {"held": False}
+    assert mgr.verify(tok, now=after) is False
+    # an expired lock is free to re-take by anyone
+    mgr.acquire("johnny5", ttl_seconds=10, now=after)
+    assert mgr.status(now=after)["holder"] == "johnny5"
+
+
+def test_verify_matches_only_active_token():
+    mgr = SwapLockManager()
+    tok = mgr.acquire("openclaw", ttl_seconds=60, now=T0).token
+    assert mgr.verify(tok, now=T0) is True
+    assert mgr.verify("nope", now=T0) is False
+    assert mgr.verify(None, now=T0) is False
+
+
+def test_release_requires_token_then_frees():
+    mgr = SwapLockManager()
+    tok = mgr.acquire("openclaw", ttl_seconds=60, now=T0).token
+    with pytest.raises(PermissionError):
+        mgr.release("wrong", now=T0)
+    assert mgr.release(tok, now=T0) is True
+    assert mgr.status(now=T0) == {"held": False}
+
+
+def test_force_release_skips_token_and_release_of_free_lock_is_false():
+    mgr = SwapLockManager()
+    mgr.acquire("openclaw", ttl_seconds=60, now=T0)
+    assert mgr.release(force=True, now=T0) is True
+    assert mgr.release(force=True, now=T0) is False  # nothing held now
+
+
+def test_is_blocked_by_is_the_swap_gate():
+    # Mirrors the single-read decision the /api/swap endpoint makes.
+    mgr = SwapLockManager()
+    assert mgr.is_blocked_by(None, now=T0) is None          # free lock blocks nobody
+    tok = mgr.acquire("openclaw", ttl_seconds=10, now=T0).token
+    blocked = mgr.is_blocked_by(None, now=T0)               # no token -> blocked
+    assert blocked is not None and blocked["holder"] == "openclaw"
+    assert mgr.is_blocked_by("wrong", now=T0) is not None   # wrong token -> blocked
+    assert mgr.is_blocked_by(tok, now=T0) is None           # holder's token -> allowed
+    # At/after expiry the gate is open even without a token (the bug a separate
+    # status()+verify() pair would get wrong).
+    assert mgr.is_blocked_by(None, now=T0 + timedelta(seconds=11)) is None
+
+
+# ------------------------------------------------------------------- webhook ----
+
+def test_build_webhook_payload_shape():
+    p = build_webhook_payload(
+        event="swap_complete", job_id="abc123", model_key="gemma",
+        state="ready", returncode=0, started_at="t0", finished_at="t1",
+        dry_run=False,
+    )
+    assert p == {
+        "event": "swap_complete", "job_id": "abc123", "model_key": "gemma",
+        "state": "ready", "returncode": 0, "started_at": "t0",
+        "finished_at": "t1", "dry_run": False,
+    }
+
+
+def test_sign_payload_is_deterministic_and_prefixed():
+    body = b'{"event":"swap_complete"}'
+    sig = sign_payload("s3cr3t", body)
+    assert sig.startswith("sha256=")
+    assert sig == sign_payload("s3cr3t", body)
+    assert sig != sign_payload("other", body)
+
+
+def test_disabled_webhook_fire_is_noop():
+    n = WebhookNotifier("", "")
+    assert n.enabled is False
+    # Must not attempt any network call or raise when no URL is configured.
+    assert asyncio.run(n.fire("swap_complete", {"x": 1})) is None
+
+
+# --------------------------------------------------------- schedule registry ----
+
+def test_register_and_list_schedule():
+    reg = ScheduleRegistry()
+    e = reg.register(name="Daily Vol", owner="openclaw", cron="0 6 * * *")
+    assert e.id and e.registered_at and e.updated_at
+    listed = reg.list()
+    assert len(listed) == 1 and listed[0]["name"] == "Daily Vol"
+
+
+def test_register_with_id_updates_in_place():
+    reg = ScheduleRegistry()
+    reg.register(name="Daily Vol", id="dv", owner="openclaw", cron="0 6 * * *")
+    reg.register(name="Daily Vol v2", id="dv", owner="openclaw", cron="0 7 * * *")
+    listed = reg.list()
+    assert len(listed) == 1
+    assert listed[0]["name"] == "Daily Vol v2" and listed[0]["cron"] == "0 7 * * *"
+
+
+def test_register_requires_name_and_validates_id():
+    reg = ScheduleRegistry()
+    with pytest.raises(ValueError):
+        reg.register(name="  ")
+    with pytest.raises(ValueError):
+        reg.register(name="ok", id="bad id; rm -rf")
+
+
+def test_delete_schedule():
+    reg = ScheduleRegistry()
+    reg.register(name="Daily Vol", id="dv")
+    assert reg.delete("dv") is True
+    assert reg.delete("dv") is False
+    assert reg.list() == []
+
+
+def test_valid_schedule_id():
+    assert valid_schedule_id("daily-vol")
+    assert valid_schedule_id("a.b_c-1")
+    assert not valid_schedule_id("")
+    assert not valid_schedule_id("../etc")
+    assert not valid_schedule_id("has space")
+    assert not valid_schedule_id("x" * 65)

image/tests/test_discovery.py

@@ -0,0 +1,190 @@
+"""Disk-driven menu helpers: cache-dir parsing + launch-recipe inference.
+
+All offline — pure functions over a fake cache listing and fake config.json
+dicts. The SSH scan, the menu merge, and the suggest endpoint that wire these
+together are exercised by hand against the live cluster (mock-heavy unit tests of
+those would test the mocks).
+"""
+import asyncio
+
+from app import discovery
+from app.config import Settings
+from app.disk import DiskStatus, cache_dirname_to_repo, parse_cache_listing
+from app.discovery import repo_to_key, infer_recipe, _detect_family
+from app.models import load_catalog
+
+
+# ---- cache dirname <-> repo ----
+
+def test_cache_dirname_to_repo_roundtrip():
+    assert cache_dirname_to_repo("models--RedHatAI--Qwen3.6-35B-A3B-NVFP4") == "RedHatAI/Qwen3.6-35B-A3B-NVFP4"
+
+
+def test_cache_dirname_name_with_double_dash():
+    # The org is the first segment; everything after is the name (single '/').
+    assert cache_dirname_to_repo("models--org--weird--name") == "org/weird--name"
+
+
+def test_cache_dirname_rejects_non_model_dirs():
+    assert cache_dirname_to_repo("datasets--foo--bar") is None
+    assert cache_dirname_to_repo("models--onlyorg") is None
+    assert cache_dirname_to_repo("random") is None
+
+
+# ---- parse_cache_listing ----
+
+def test_parse_cache_listing_complete_and_incomplete():
+    out = (
+        "20000000000|1|models--RedHatAI--Qwen3.6-35B-A3B-NVFP4\n"
+        "5000000000|0|models--some--half-downloaded\n"
+        "\n"
+        "garbage line with no pipes\n"
+        "123|1|not-a-model-dir\n"
+    )
+    items = parse_cache_listing(out)
+    assert items == [
+        ("RedHatAI/Qwen3.6-35B-A3B-NVFP4", 20000000000, True),
+        ("some/half-downloaded", 5000000000, False),
+    ]
+
+
+def test_parse_cache_listing_bad_size_defaults_zero():
+    items = parse_cache_listing("notanumber|1|models--a--b")
+    assert items == [("a/b", 0, True)]
+
+
+# ---- repo_to_key ----
+
+def test_repo_to_key_is_url_safe_and_stable():
+    assert repo_to_key("RedHatAI/Qwen3.6-35B-A3B-NVFP4") == "redhatai-qwen3-6-35b-a3b-nvfp4"
+    # Idempotent enough to be a stable id across calls.
+    assert repo_to_key("nvidia/Gemma-4-26B-A4B-NVFP4") == "nvidia-gemma-4-26b-a4b-nvfp4"
+
+
+# ---- family detection ----
+
+def test_detect_qwen3_moe():
+    cfg = {"architectures": ["Qwen3MoeForCausalLM"], "model_type": "qwen3_moe", "num_experts": 128}
+    label, flags, caps = _detect_family(cfg)
+    assert "--reasoning-parser=qwen3" in flags
+    assert "--moe_backend=flashinfer_cutlass" in flags
+    assert "reasoning" in caps
+    assert "MoE" in label
+
+
+def test_detect_gemma_moe_uses_marlin():
+    cfg = {"architectures": ["Gemma4MoeForConditionalGeneration"], "model_type": "gemma4_moe", "num_local_experts": 8}
+    label, flags, caps = _detect_family(cfg)
+    assert "--reasoning-parser=gemma4" in flags
+    assert "--tool-call-parser=gemma4" in flags
+    assert "--moe_backend=marlin" in flags        # NOT flashinfer_cutlass — GB10 footgun
+    assert "vision" in caps                         # ConditionalGeneration => multimodal
+    assert "tools" in caps
+
+
+def test_detect_generic_has_no_family_flags():
+    label, flags, caps = _detect_family({"architectures": ["LlamaForCausalLM"], "model_type": "llama"})
+    assert flags == []
+    assert label == "Generic"
+
+
+def test_detect_vision_from_config_keys():
+    _, _, caps = _detect_family({"model_type": "qwen3", "vision_config": {"x": 1}})
+    assert "vision" in caps
+
+
+# ---- infer_recipe (the prefill the setup form receives) ----
+
+def test_infer_recipe_solo_small_model():
+    cfg = {"architectures": ["Qwen3ForCausalLM"], "model_type": "qwen3"}
+    rec = infer_recipe("RedHatAI/Qwen3.6-35B-A3B-NVFP4", cfg, total_bytes=20_000_000_000, on_host_count=1)
+    assert rec["mode"] == "solo"
+    assert rec["key"] == "redhatai-qwen3-6-35b-a3b-nvfp4"
+    assert rec["repo"] == "RedHatAI/Qwen3.6-35B-A3B-NVFP4"
+    assert "--reasoning-parser=qwen3" in rec["vllm_args"]
+    assert "-tp=2" not in rec["vllm_args"]
+    assert rec["knobs"]["kv_cache_dtype"] == "fp8"
+
+
+def test_infer_recipe_cluster_when_on_both_hosts():
+    rec = infer_recipe("org/big", {}, total_bytes=10_000_000_000, on_host_count=2)
+    assert rec["mode"] == "cluster"
+    assert "-tp=2" in rec["vllm_args"]
+    assert "--distributed-executor-backend=ray" in rec["vllm_args"]
+    assert rec["knobs"]["gpu_memory_utilization"] == 0.7
+
+
+def test_infer_recipe_cluster_when_too_big_for_one_spark():
+    rec = infer_recipe("org/huge", {}, total_bytes=200_000_000_000, on_host_count=1)
+    assert rec["mode"] == "cluster"
+
+
+# ---- build_menu merge (disk scan ∪ recipes) ----
+
+def _both_spark_settings(monkeypatch) -> Settings:
+    for k in ("SPARK1_HOST", "SPARK1_USER", "SPARK2_HOST", "SPARK2_USER"):
+        monkeypatch.delenv(k, raising=False)
+    monkeypatch.setenv("SPARK1_HOST", "1.1.1.1")
+    monkeypatch.setenv("SPARK1_USER", "u")
+    monkeypatch.setenv("SPARK2_HOST", "2.2.2.2")
+    monkeypatch.setenv("SPARK2_USER", "u")
+    return Settings.from_env()
+
+
+def test_build_menu_merges_recipe_discovered_and_hides_incomplete(monkeypatch):
+    cat = load_catalog("models.yaml")  # bundled recipes incl. qwen36 + gemma4
+    settings = _both_spark_settings(monkeypatch)
+
+    async def fake_list(host, user, s):
+        if host == "1.1.1.1":
+            return [
+                ("RedHatAI/Qwen3.6-35B-A3B-NVFP4", 20_000_000_000, True),  # recipe match
+                ("someorg/mystery-7B", 7_000_000_000, True),               # needs setup
+                ("broken/half", 1_000_000_000, False),                     # incomplete -> hidden
+            ]
+        return []  # spark2 empty
+
+    async def fake_probe(repo, mode, s, *, local_path=None):
+        return DiskStatus(repo=local_path or repo, on_disk=False, total_bytes=0, per_host=[])
+
+    monkeypatch.setattr(discovery, "list_cached_models", fake_list)
+    monkeypatch.setattr(discovery, "probe_disk", fake_probe)
+
+    menu = asyncio.run(discovery.build_menu(settings, cat))
+
+    # Recipe-matched: keyed by recipe key, ready (not needs_setup), real size.
+    assert "qwen36" in menu
+    assert menu["qwen36"]["needs_setup"] is False
+    assert menu["qwen36"]["total_bytes"] == 20_000_000_000
+
+    # Discovered-without-recipe: slug key, needs_setup.
+    slug = repo_to_key("someorg/mystery-7B")
+    assert menu[slug]["needs_setup"] is True
+
+    # Incomplete download is filtered out entirely.
+    assert all("half" not in k for k in menu)
+
+    # A recipe with nothing on disk (e.g. gemma4) must NOT appear — the menu is the disk.
+    assert "gemma4" not in menu
+
+
+def test_build_menu_sums_cluster_model_across_both_sparks(monkeypatch):
+    cat = load_catalog("models.yaml")
+    settings = _both_spark_settings(monkeypatch)
+
+    async def fake_list(host, user, s):
+        # Same repo present on BOTH Sparks — one card, sizes summed (not two cards).
+        return [("org/sharded-235B", 70_000_000_000, True)]
+
+    async def fake_probe(repo, mode, s, *, local_path=None):
+        return DiskStatus(repo=repo, on_disk=False, total_bytes=0, per_host=[])
+
+    monkeypatch.setattr(discovery, "list_cached_models", fake_list)
+    monkeypatch.setattr(discovery, "probe_disk", fake_probe)
+
+    menu = asyncio.run(discovery.build_menu(settings, cat))
+    key = repo_to_key("org/sharded-235B")
+    assert list(menu) == [key]                       # exactly one card
+    assert menu[key]["total_bytes"] == 140_000_000_000  # summed across both hosts
+    assert len(menu[key]["per_host"]) == 2
+    assert menu[key]["mode"] == "cluster"            # present on 2 hosts -> cluster

image/tests/test_launch_command.py

@@ -7,6 +7,9 @@ the command back into the exact token list. The vLLM pre-flight validator
 """
 import shlex
 
+import pytest
+from pydantic import ValidationError
+
 from app.models import Defaults, ModelDef, build_launch_command
 
 DEFAULTS = Defaults(port=8888, host="0.0.0.0")
@@ -65,3 +68,81 @@ def test_injection_via_vllm_arg_stays_literal():
     payload = "--foo=$(touch /tmp/pwned)"
     cmd = build_launch_command("k", _model(vllm_args=[payload]), DEFAULTS)
     assert payload in shlex.split(cmd)  # preserved as one inert token
+
+
+# ---- local / fine-tuned models (served by directory, not HF repo) ----
+
+def test_local_model_bind_mounts_dir_and_serves_the_path():
+    m = _model(repo="", local_path="/home/u/models/ft-v2", vllm_args=["--max-model-len=2048"])
+    cmd = build_launch_command("k", m, DEFAULTS)
+    tokens = shlex.split(cmd)
+    # The launch script's hook bind-mounts the host dir at the SAME container path.
+    assert tokens[0] == (
+        "VLLM_SPARK_EXTRA_DOCKER_ARGS=-v /home/u/models/ft-v2:/home/u/models/ft-v2"
+    )
+    # vLLM is pointed at the directory, not an HF repo id.
+    i = tokens.index("serve")
+    assert tokens[i + 1] == "/home/u/models/ft-v2"
+    assert "--max-model-len=2048" in tokens
+
+
+def test_local_model_chat_template_arg_survives_round_trip():
+    m = _model(
+        repo="",
+        local_path="/m/ft",
+        vllm_args=["--chat-template=/m/ft/chat_template.jinja"],
+    )
+    cmd = build_launch_command("k", m, DEFAULTS)
+    assert "--chat-template=/m/ft/chat_template.jinja" in shlex.split(cmd)
+
+
+def test_local_path_with_metacharacters_is_quoted_not_executed():
+    # The validator rejects a hostile path at the boundary; bypass it with
+    # model_construct to prove the quote_arg sink is safe in depth even if a bad
+    # value somehow reaches build_launch_command.
+    evil = "/m/ft; rm -rf ~"
+    m = ModelDef.model_construct(
+        display_name="X", repo="", local_path=evil, size_gb=1.0, mode="solo",
+        vllm_args=[], knobs=None, custom=False, capabilities=[],
+        expected_ready_seconds=300, description=None,
+    )
+    cmd = build_launch_command("k", m, DEFAULTS)
+    tokens = shlex.split(cmd)
+    i = tokens.index("serve")
+    assert tokens[i + 1] == evil  # recovered as one literal token, not executed
+    assert tokens[0] == f"VLLM_SPARK_EXTRA_DOCKER_ARGS=-v {evil}:{evil}"
+
+
+def test_model_requires_exactly_one_source():
+    with pytest.raises(ValidationError):
+        ModelDef(display_name="x", size_gb=1, mode="solo")  # neither repo nor local_path
+    with pytest.raises(ValidationError):
+        ModelDef(display_name="x", repo="o/n", local_path="/p", size_gb=1, mode="solo")  # both
+
+
+def test_local_model_rejects_chat_template_outside_dir():
+    # Only local_path is mounted into the container, so a chat-template elsewhere
+    # would silently 404 inside vLLM — reject it up front.
+    with pytest.raises(ValidationError):
+        ModelDef(
+            display_name="x", repo="", local_path="/m/ft", size_gb=1, mode="solo",
+            vllm_args=["--chat-template=/other/dir/t.jinja"],
+        )
+
+
+def test_invalid_local_path_rejected_by_model():
+    with pytest.raises(ValidationError):
+        ModelDef(display_name="x", repo="", local_path="/m/../etc", size_gb=1, mode="solo")
+
+
+def test_merge_overrides_loads_local_and_skips_invalid(monkeypatch):
+    # YAML/override-added local models get the same validation as the API; a single
+    # bad entry is skipped (logged) rather than breaking the whole catalog load.
+    from app import models as M
+    monkeypatch.setattr(M, "load_overrides", lambda: {"knobs": {}, "custom": [
+        {"key": "good", "display_name": "G", "local_path": "/home/u/m", "size_gb": 1, "mode": "solo"},
+        {"key": "bad", "display_name": "B", "local_path": "/home/u/../etc", "size_gb": 1, "mode": "solo"},
+    ]})
+    cat = M._merge_overrides(M.Catalog(models={}))
+    assert cat.models["good"].is_local and cat.models["good"].source == "/home/u/m"
+    assert "bad" not in cat.models  # traversal path skipped, not catalog-fatal

image/tests/test_shellsafe.py

@@ -6,7 +6,12 @@ use `validate_x(v)` inline.
 """
 import pytest
 
-from app.shellsafe import validate_container, validate_image, validate_repo
+from app.shellsafe import (
+    validate_container,
+    validate_image,
+    validate_local_path,
+    validate_repo,
+)
 
 # Shell metacharacters that must never survive any validator — these are the
 # actual injection vectors. (Path traversal like "../" is NOT in scope here:
@@ -96,3 +101,27 @@ def test_container_valid_passes_through_unchanged(name):
 def test_container_rejects_malformed_and_hostile(name):
     with pytest.raises(ValueError):
         validate_container(name)
+
+
+# ---- validate_local_path: absolute model dir, no traversal/metacharacters ----
+
+@pytest.mark.parametrize("path", [
+    "/home/modelo/models/gemma-4-31B-ten31-v2",
+    "/data/models/ft.v2_1",
+    "/srv/m/a-b/c",
+])
+def test_local_path_valid_passes_through_unchanged(path):
+    assert validate_local_path(path) == path
+
+
+@pytest.mark.parametrize("path", [
+    "",
+    "relative/path",            # must be absolute
+    "~/models/x",               # no ~ expansion
+    "/models/../etc/shadow",    # '..' traversal
+    "/models/./x",              # '.' segment
+    "/a" * 300,                 # over the 512 cap (600 chars)
+] + [f"/models/x{h}" for h in HOSTILE])
+def test_local_path_rejects_relative_traversal_and_hostile(path):
+    with pytest.raises(ValueError):
+        validate_local_path(path)

image/tests/test_topology.py

@@ -0,0 +1,120 @@
+"""Configurable topology: DISABLED_SERVICES, vLLM container override, and the
+extra-vLLM probe. All offline — the disabled checks short-circuit before any
+network call, and the probes are exercised only on the not-configured path.
+"""
+import asyncio
+
+from app.config import Settings
+from app.health import (
+    check_embeddings,
+    check_kokoro,
+    check_parakeet,
+    check_qdrant,
+    check_vllm,
+    probe_vllm_endpoint,
+)
+from app.services import services_from_settings
+
+
+def _settings(monkeypatch, **env) -> Settings:
+    # Pin the topology env vars under test; default the rest to blank so a stray
+    # value in the real environment can't leak into the assertion.
+    keys = [
+        "SPARK1_HOST", "SPARK1_USER", "SPARK2_HOST", "SPARK2_USER",
+        "DISABLED_SERVICES", "VLLM_CONTAINER",
+    ]
+    for k in keys:
+        monkeypatch.delenv(k, raising=False)
+    for k, v in env.items():
+        monkeypatch.setenv(k, v)
+    return Settings.from_env()
+
+
+# ---- DISABLED_SERVICES parsing ----
+
+def test_disabled_services_parsed_lowercased_and_trimmed(monkeypatch):
+    s = _settings(monkeypatch, DISABLED_SERVICES="parakeet, Kokoro ,,")
+    assert s.disabled_services == frozenset({"parakeet", "kokoro"})
+
+
+def test_disabled_services_blank_is_empty(monkeypatch):
+    assert _settings(monkeypatch).disabled_services == frozenset()
+
+
+# ---- vLLM container override ----
+
+def test_vllm_container_defaults_to_vllm_node(monkeypatch):
+    assert _settings(monkeypatch).vllm_container == "vllm_node"
+
+
+def test_vllm_container_override(monkeypatch):
+    assert _settings(monkeypatch, VLLM_CONTAINER="vllm-gemma4").vllm_container == "vllm-gemma4"
+
+
+def test_vllm_container_invalid_falls_back(monkeypatch):
+    # A malformed value (space / shell metachar) is rejected at the boundary and
+    # falls back to the default rather than crashing startup or reaching a sink.
+    assert _settings(monkeypatch, VLLM_CONTAINER="bad name; rm -rf").vllm_container == "vllm_node"
+
+
+# ---- services map honors the disable list ----
+
+def test_services_from_settings_drops_disabled(monkeypatch):
+    s = _settings(
+        monkeypatch,
+        SPARK1_HOST="10.0.0.1", SPARK1_USER="u",
+        SPARK2_HOST="10.0.0.2", SPARK2_USER="u",
+        DISABLED_SERVICES="parakeet,qdrant",
+    )
+    svcs = services_from_settings(s)
+    assert "parakeet" not in svcs and "qdrant" not in svcs
+    assert "kokoro" in svcs and "embeddings" in svcs
+
+
+def test_custom_vllm_service_registered(monkeypatch):
+    from app import custom_services
+    monkeypatch.setattr(custom_services, "load_custom_services", lambda: [
+        {"key": "vllm-spark2", "kind": "vllm", "host": "10.0.0.2",
+         "user": "u", "container": "vllm_node", "port": 8000},
+    ])
+    s = _settings(monkeypatch, SPARK1_HOST="10.0.0.1", SPARK1_USER="u",
+                  SPARK2_HOST="10.0.0.2", SPARK2_USER="u")
+    svc = services_from_settings(s)["vllm-spark2"]
+    assert svc.kind == "vllm" and svc.port == 8000 and svc.container == "vllm_node"
+
+
+def test_custom_service_colliding_with_builtin_is_ignored(monkeypatch):
+    # A custom entry can't shadow a built-in key — the built-in wins.
+    from app import custom_services
+    monkeypatch.setattr(custom_services, "load_custom_services", lambda: [
+        {"key": "parakeet", "kind": "vllm", "host": "10.0.0.9", "user": "u", "port": 8000},
+    ])
+    s = _settings(monkeypatch, SPARK1_HOST="10.0.0.1", SPARK1_USER="u",
+                  SPARK2_HOST="10.0.0.2", SPARK2_USER="u")
+    assert services_from_settings(s)["parakeet"].kind == "stt"
+
+
+# ---- disabled health checks short-circuit (no network) ----
+
+def test_disabled_check_returns_disabled_verdict(monkeypatch):
+    s = _settings(
+        monkeypatch,
+        SPARK2_HOST="10.0.0.2", SPARK2_USER="u",  # host set, but disable wins
+        DISABLED_SERVICES="parakeet,kokoro,embeddings,qdrant",
+    )
+    for check in (check_parakeet, check_kokoro, check_embeddings, check_qdrant):
+        r = asyncio.run(check(s))
+        assert r == {"ok": False, "disabled": True, "error": "disabled", "base_url": None}
+
+
+# ---- vLLM probe: not-configured path is pure ----
+
+def test_probe_vllm_endpoint_unconfigured(monkeypatch):
+    r = asyncio.run(probe_vllm_endpoint("", 8000))
+    assert r["ok"] is False and "not configured" in r["error"]
+
+
+def test_check_vllm_unconfigured_without_spark1(monkeypatch):
+    s = _settings(monkeypatch)  # no SPARK1_HOST
+    r = asyncio.run(check_vllm(s))
+    assert r["ok"] is False and "spark1 not configured" in r["error"]

package/startos/actions/configureSparks.ts

@@ -49,6 +49,24 @@ const inputSpec = InputSpec.of({
     placeholder: 'leave blank for 8888',
     masked: false,
   }),
+  vllm_container: Value.text({
+    name: 'vLLM container name (optional)',
+    description:
+      'Docker container name for the swappable vLLM on Spark 1. Defaults to "vllm_node" (what the bundled launch-cluster.sh creates). Change this only if you run your vLLM under a different container name — the model-swap log view and the pre-flight validator exec into it by name.',
+    required: false,
+    default: null,
+    placeholder: 'leave blank for vllm_node',
+    masked: false,
+  }),
+  disabled_services: Value.text({
+    name: 'Services to hide (optional)',
+    description:
+      "Comma-separated list of built-in services your cluster doesn't run, so Spark Control hides their tiles and stops probing them. Valid names: parakeet, kokoro, embeddings, qdrant. Example: if you only run vLLM, set this to 'parakeet,kokoro,embeddings,qdrant'. Leave blank to monitor all of them. (Useful when, say, your vLLM shares port 8000 with Parakeet's default — hide Parakeet so its probe doesn't hit vLLM.)",
+    required: false,
+    default: null,
+    placeholder: 'e.g. parakeet,kokoro',
+    masked: false,
+  }),
   parakeet_host: Value.text({
     name: 'Parakeet host (optional)',
     description:
@@ -155,6 +173,24 @@ const inputSpec = InputSpec.of({
     placeholder: 'starts with "nvapi-..."',
     masked: true,
   }),
+  swap_webhook_url: Value.text({
+    name: 'Swap webhook URL (optional)',
+    description:
+      'If you run automation that needs to know when the loaded model changes, paste a URL here. Spark Control POSTs a small JSON event (swap_complete / swap_failed) to it after every model swap, so the consumer can re-point its config to the new model. Leave blank to disable. Only needed if something other than this dashboard cares about swaps.',
+    required: false,
+    default: null,
+    placeholder: 'e.g. https://my-service.local/spark-swap',
+    masked: false,
+  }),
+  swap_webhook_secret: Value.text({
+    name: 'Swap webhook secret (optional)',
+    description:
+      'Optional shared secret. If set, each webhook is signed with an "X-Spark-Signature: sha256=…" header (HMAC of the body) so the receiver can verify it really came from Spark Control. Leave blank to send the webhook unsigned.',
+    required: false,
+    default: null,
+    placeholder: 'a random string the receiver also knows',
+    masked: true,
+  }),
 })
 
 export const configureSparks = sdk.Action.withInput(

package/startos/fileModels/sparkConfig.yaml.ts

@@ -9,6 +9,11 @@ export const sparkConfigSchema = z.object({
   spark2_user: z.string().catch(''),
   // Optional vLLM port override (Spark 1). Blank => 8888 (launch-cluster.sh default).
   vllm_port: z.string().catch(''),
+  // Optional vLLM container-name override (Spark 1). Blank => "vllm_node".
+  vllm_container: z.string().catch(''),
+  // Optional comma-separated list of built-in services to switch off
+  // (parakeet, kokoro, embeddings, qdrant). Blank => all enabled.
+  disabled_services: z.string().catch(''),
   // Optional per-service overrides. Blank => use spark2_host / spark2_user.
   parakeet_host: z.string().catch(''),
   parakeet_user: z.string().catch(''),
@@ -30,6 +35,11 @@ export const sparkConfigSchema = z.object({
   open_webui_url: z.string().catch(''),
   // Optional NGC API key for pulling NIM containers from nvcr.io/nim/...
   ngc_api_key: z.string().catch(''),
+  // Optional coordination webhook: POSTed on swap_complete/swap_failed so
+  // downstream consumers re-point their model config. Blank => disabled.
+  swap_webhook_url: z.string().catch(''),
+  // Optional shared secret; if set, the webhook body is HMAC-signed.
+  swap_webhook_secret: z.string().catch(''),
 })
 
 export type SparkConfig = z.infer<typeof sparkConfigSchema>

package/startos/main.ts

@@ -14,6 +14,8 @@ export const main = sdk.setupMain(async ({ effects }) => {
     spark2_host: '',
     spark2_user: '',
     vllm_port: '',
+    vllm_container: '',
+    disabled_services: '',
     parakeet_host: '',
     parakeet_user: '',
     parakeet_container: '',
@@ -30,6 +32,8 @@ export const main = sdk.setupMain(async ({ effects }) => {
     matrix_bridge_user: '',
     open_webui_url: '',
     ngc_api_key: '',
+    swap_webhook_url: '',
+    swap_webhook_secret: '',
   }
 
   return sdk.Daemons.of(effects).addDaemon('primary', {
@@ -52,6 +56,8 @@ export const main = sdk.setupMain(async ({ effects }) => {
         SPARK2_HOST: cfg.spark2_host,
         SPARK2_USER: cfg.spark2_user,
         VLLM_PORT: cfg.vllm_port,
+        VLLM_CONTAINER: cfg.vllm_container,
+        DISABLED_SERVICES: cfg.disabled_services,
         PARAKEET_HOST: cfg.parakeet_host,
         PARAKEET_USER: cfg.parakeet_user,
         PARAKEET_CONTAINER: cfg.parakeet_container,
@@ -71,6 +77,8 @@ export const main = sdk.setupMain(async ({ effects }) => {
         CONNECTIVITY_LOG: '/data/connectivity.json',
         OPEN_WEBUI_URL: cfg.open_webui_url,
         NGC_API_KEY: cfg.ngc_api_key,
+        SWAP_WEBHOOK_URL: cfg.swap_webhook_url,
+        SWAP_WEBHOOK_SECRET: cfg.swap_webhook_secret,
         BIND_PORT: String(uiPort),
       },
     },

package/startos/versions/v0_1_0.ts

@@ -1,10 +1,10 @@
 import { VersionInfo, IMPOSSIBLE } from '@start9labs/start-sdk'
 
 export const v0_1_0 = VersionInfo.of({
-  version: '0.22.0:0',
+  version: '0.26.0:0',
   releaseNotes: {
     en_US:
-      "v0.22.0:0 — configurable vLLM port. The port Spark Control uses to reach vLLM on Spark 1 (the health check and the chat proxy) is now a field in the Configure Sparks action, so you can point it at a vLLM that listens on a non-default port without rebuilding the package. Leave it blank to keep the previous default of 8888 — what the bundled launch-cluster.sh wrapper uses; set it to 8000 (vLLM's own default) or any other port if your vLLM listens elsewhere. Also hardened numeric-setting parsing so a blank or malformed port value falls back to its default instead of crashing daemon startup.",
+      "v0.26.0:0 — the model menu is now what's actually on your Sparks. The dashboard scans both Sparks for downloaded models and shows exactly those — no more hard-coded list. (1) Delete means delete: removing a model frees its weights AND takes the card off the menu (re-download later to bring it back, with its saved settings). (2) Download a new model and it appears on the menu by itself when it finishes. (3) Models Spark Control doesn't recognize show a \"needs setup\" card — the first time you switch to one, it reads the model's own files, guesses how to launch it (which family, solo vs both Sparks, the right vLLM flags), and asks you to confirm once; after that it's a normal card. (4) The download box now autocompletes known-good models. (5) Each install shows its own Sparks' models, so a shared copy no longer displays someone else's list. Removed the two legacy Qwen entries (235B FP8, 2.5 72B) — they'll still appear if you actually have them downloaded. No consumer-API changes; the /v1 proxy and swap API are unchanged.",
   },
   migrations: {
     up: async ({ effects }) => {},

runbook.md

@@ -52,13 +52,43 @@ The **Update** button runs `git fetch && git reset --hard origin/<branch> && doc
 
 3. Spark Control's own package key must be authorized for that SSH user (Show Public Key → add to their `authorized_keys`) unless it's the same user Spark Control already uses for that Spark.
 
+## Configurable topology (v0.24.0+)
+
+For a cluster wired differently from the reference layout, three optional knobs in **Configure Sparks** (no fork needed):
+
+- **vLLM container name** — defaults to `vllm_node`. Set it if your swappable vLLM on Spark 1 runs under a different container name; the swap log-tail and the pre-flight validator `docker exec` into it by name.
+- **Services to hide** — comma-separated `parakeet,kokoro,embeddings,qdrant`. Hidden services show no tile and are never probed (status, deep-health, or connectivity log). Use this when a service you don't run would otherwise be probed at a port something else answers — e.g. a vLLM on port 8000 colliding with Parakeet's default.
+- **Monitor a second vLLM** — the swap machinery only drives the Spark 1 vLLM, but you can *monitor* a vLLM on another Spark by adding a custom service of `kind: vllm` to `/data/services-overrides.yaml`:
+
+  ```yaml
+  custom:
+    - key: vllm-spark2
+      kind: vllm
+      host: <spark-2-ip>
+      user: <ssh-user>
+      container: vllm_node
+      port: 8000
+  ```
+
+  It gets a read-only tile: loaded model (via `/v1/models`), container state, and start/stop/restart. (Spark Control's SSH key must be authorized for that user — Show Public Key.)
+
 ## Adding a new model
 
-1. Add an entry to `image/models.yaml`. Required fields: `display_name`, `repo`, `size_gb`, `mode` (`solo` or `cluster`), `vllm_args`. Optional but recommended: `description` (one paragraph — what the model is, what it's good for, how it differs from others; renders below the meta tags in each card), `capabilities` (tags like `[vision, reasoning, tools]`), `expected_ready_seconds`.
+The menu is whatever's downloaded on the Sparks, so the normal path is just:
-2. Confirm the weights are on the Spark: `ssh <spark-user>@<spark-1-host> 'ls ~/.cache/huggingface/hub/'`. If not, download with `./hf-download.sh <repo>` on Spark 1.
+**download it, then set it up once.**
-3. Rebuild + redeploy the package: `cd package && make x86 && make install`.
 
-If `description` is omitted, the card simply hides that section — no need to populate it for every model. Keep descriptions generic (not user-specific) so the catalog stays portable.
+1. **Download** from the dashboard (**+ Download a new model**, paste the HF repo) or on Spark 1 with `./hf-download.sh <repo>`. When it finishes it appears on the menu by itself.
+2. **Set it up.** If Spark Control already has a recipe for it (see below), it's ready to switch to. Otherwise it shows a **"needs setup"** card: the first switch reads the model's `config.json`, proposes how to launch it (family/parsers, solo vs cluster, vLLM flags), and you confirm once. The confirmed recipe persists to `/data/models-overrides.yaml` (survives package updates).
+
+### Bundling a launch recipe (optional — skips the setup prompt)
+
+To make a known model launch correctly the instant it's downloaded, add a *recipe* to `image/models.yaml`. These are **not** the menu — they're matched to an on-disk model by `repo`. Required: `display_name`, `repo`, `size_gb`, `mode` (`solo`/`cluster`), `vllm_args`. Optional: `description`, `capabilities` (e.g. `[vision, reasoning, tools]`), `expected_ready_seconds`. Then rebuild + redeploy: `cd package && make x86 && make install`. Keep descriptions generic (not user-specific) so the recipes stay portable.
+
+### Local / fine-tuned models (v0.23.0+)
+
+A model that lives as a directory on a Spark (e.g. a LoRA-merged fine-tune) instead of an HF repo: use the **"+ Add local model"** button under LLM swap (or a `custom:` entry with `local_path` instead of `repo` in the override YAML). The directory must already exist on the Spark; only its parent dir is mounted, so a `--chat-template` must live **inside** `local_path`.
+
+**Load-bearing contract:** on swap, spark-control prefixes the launch with `VLLM_SPARK_EXTRA_DOCKER_ARGS="-v <path>:<path>"` so `launch-cluster.sh` bind-mounts the dir into the vLLM container at the same path. This relies on the upstream `eugr/spark-vllm-docker` `launch-cluster.sh` expanding `$VLLM_SPARK_EXTRA_DOCKER_ARGS` **unquoted** into its `docker run` (verified against the on-Spark script 2026-06-17: line ~11 appends it to `DOCKER_ARGS`, used unquoted in `docker run`). If a future upstream version quotes that variable, local-model mounts would silently fail — re-check this before pulling launch-cluster.sh updates.
 
 ## Manual swap fallback
 
@@ -75,6 +105,17 @@ cd ~/spark-vllm-docker
 docker logs -f vllm_node      # wait for "Application startup complete."
 ```
 
+## Sideload (`make install`) can't reach the server
+
+Symptom: `make install` fails with `package.sideload: error sending request for url (https://immense-voyage.local/rpc/v1)`. Cause seen 2026-06-17: `immense-voyage.local` stopped resolving via mDNS from the Mac (`curl https://immense-voyage.local/...` → exit 6, "couldn't resolve host"), even though the server is up — `curl -sk https://<server-ip>/rpc/v1` returns 200.
+
+- **Don't** work around it with `start-cli -H https://<server-ip> package install`: TLS connects but it returns `UNAUTHORIZED`, because start-cli's stored credential is bound to the registered `.local` host, not the IP.
+- **Fix:** make the name resolve again, then re-run `make install`:
+  - `sudo dscacheutil -flushcache && sudo killall -HUP mDNSResponder` (flush mDNS), or
+  - `echo "<server-ip> immense-voyage.local" | sudo tee -a /etc/hosts` (deterministic; remove later).
+
+Note this only blocks installing to *your own* Start9 — building and publishing the s9pk to Gitea Releases is unaffected (adopters still pull the latest).
+
 ## Diagnostics
 
 ```bash

scripts/gitea-release.sh

@@ -8,38 +8,58 @@
 # The git tag (vX.Y.Z, derived from the version) must already exist and be pushed
 # (`git tag v0.22.0 && git push gitea v0.22.0`). Re-running is idempotent: it
 # reuses an existing release for the tag and replaces a same-named asset.
+# Set GITEA_INSECURE=1 to skip TLS verification (self-signed cert on a LAN box).
 set -euo pipefail
 
 VERSION="${1:-}"; S9PK="${2:-}"
 [ -n "$VERSION" ] && [ -n "$S9PK" ] || {
   echo "usage: GITEA_URL=.. GITEA_TOKEN=.. $0 <version e.g. 0.22.0:0> <s9pk path>" >&2; exit 2; }
 : "${GITEA_URL:?set GITEA_URL to your Gitea base URL, e.g. https://gitea.lan:3000}"
-: "${GITEA_TOKEN:?set GITEA_TOKEN to a token with repository write access}"
+: "${GITEA_TOKEN:?set GITEA_TOKEN to a token with repository read+write access}"
 [ -f "$S9PK" ] || { echo "s9pk not found: $S9PK" >&2; exit 1; }
 
 TAG="v${VERSION%%:*}"                      # 0.22.0:0 -> v0.22.0
 ASSET="$(basename "$S9PK")"
 SLUG="$(git remote get-url gitea | sed -E 's#.*[:/]([^/:]+/[^/]+)\.git$#\1#')"  # grant/spark-control
 API="${GITEA_URL%/}/api/v1/repos/${SLUG}"
-AUTH=(-H "Authorization: token ${GITEA_TOKEN}")
+CURL=(curl -sS)                            # no -f: we inspect HTTP codes ourselves
+[ "${GITEA_INSECURE:-}" = "1" ] && CURL+=(-k)
 
 echo "repo ${SLUG} | tag ${TAG} | asset ${ASSET} | ${GITEA_URL}"
 
+# api METHOD URL [extra curl args...] -> sets globals HTTP_CODE and BODY
+api() {
+  local method="$1" url="$2"; shift 2
+  local out
+  out="$("${CURL[@]}" -X "$method" -H "Authorization: token ${GITEA_TOKEN}" "$@" \
+        -w $'\n%{http_code}' "$url")"
+  HTTP_CODE="${out##*$'\n'}"
+  BODY="${out%$'\n'*}"
+}
+
 # Reuse an existing release for this tag, otherwise create one.
-id="$(curl -fsS "${AUTH[@]}" "$API/releases/tags/$TAG" 2>/dev/null | jq -r '.id // empty')"
+api GET "$API/releases/tags/$TAG"
-if [ -z "$id" ]; then
+if [ "$HTTP_CODE" = 200 ]; then
-  id="$(curl -fsS -X POST "${AUTH[@]}" -H 'Content-Type: application/json' \
+  id="$(printf '%s' "$BODY" | jq -r '.id')"
+elif [ "$HTTP_CODE" = 404 ]; then
+  api POST "$API/releases" -H 'Content-Type: application/json' \
     --data "$(jq -n --arg t "$TAG" --arg n "$VERSION" \
-      '{tag_name:$t, name:$n, body:("Spark Control "+$n+". See AGENTS.md / release notes.")}')" \
+      '{tag_name:$t, name:$n, body:("Spark Control "+$n+". See AGENTS.md / release notes.")}')"
-    "$API/releases" | jq -r '.id')"
+  [ "$HTTP_CODE" = 201 ] || { echo "create release failed (HTTP $HTTP_CODE): $BODY" >&2; exit 1; }
+  id="$(printf '%s' "$BODY" | jq -r '.id')"
+else
+  echo "release lookup failed (HTTP $HTTP_CODE) — check GITEA_URL and the token's scope: $BODY" >&2
+  exit 1
 fi
-[ -n "$id" ] && [ "$id" != null ] || { echo "could not obtain release id (check URL/token/tag)" >&2; exit 1; }
+[ -n "$id" ] && [ "$id" != null ] || { echo "could not parse release id: $BODY" >&2; exit 1; }
 
 # Replace a same-named asset so re-runs don't 409.
-old="$(curl -fsS "${AUTH[@]}" "$API/releases/$id/assets" | jq -r --arg n "$ASSET" '.[] | select(.name==$n) | .id')"
+api GET "$API/releases/$id/assets"
-[ -n "$old" ] && curl -fsS -X DELETE "${AUTH[@]}" "$API/releases/$id/assets/$old" >/dev/null || true
+old="$(printf '%s' "$BODY" | jq -r --arg n "$ASSET" '.[]? | select(.name==$n) | .id')"
+[ -n "$old" ] && { api DELETE "$API/releases/$id/assets/$old"; }
 
-curl -fsS -X POST "${AUTH[@]}" -F "attachment=@${S9PK};type=application/octet-stream" \
+api POST "$API/releases/$id/assets?name=$ASSET" \
-  "$API/releases/$id/assets?name=$ASSET" >/dev/null
+  -F "attachment=@${S9PK};type=application/octet-stream"
+[ "$HTTP_CODE" = 201 ] || { echo "asset upload failed (HTTP $HTTP_CODE): $BODY" >&2; exit 1; }
 
 echo "published: ${GITEA_URL%/}/${SLUG}/releases/tag/${TAG}"