dsh-slot-health

A live health pane for local LLM endpoints in the DeepSeek Harness web UI. One rightbar tab, one second per sample, engines named by fingerprint — so a busy llama.cpp slot or a resident Ollama model reads as clearly as a dashboard instead of a raw dump.

Built and verified against live backends on the same box: llama.cpp on :8080 (Qwen3.8 27B) and Ollama on :11434 (including dual-GPU OLLAMA_SCHED_SPREAD with a 35B MoE).

See also: dsh-gpu-monitor-nvml — NVIDIA GPU util / VRAM / power pane for the same DSH web rightbar (same lab, same dual-face plugin shape).

Requirements

• DeepSeek Harness web profile (dsh web).
• A local OpenAI-compatible / native endpoint to watch:
  • llama.cpp llama-server (full support), or
  • Ollama (honest up-states + MODELS card), or
  • anything else reachable (fingerprinted as unknown / stub).

Configure the origin in cordis.patch.yml (default http://127.0.0.1:8080). Point it at :11434 when watching Ollama.

---

Screenshots

Light theme, matching the DSH default.

llama.cpp (busy, expanded)	Ollama (loaded, MODELS open)

llama.cpp (cards collapsed)	Ollama MODELS collapsed

Dock chip · llama	Dock chip · ollama

What you see

llama.cpp

Slot Health   llama.cpp  127.0.0.1:8080   · busy 2s · dec 59
  latency 2 ms · updated 0 s ago

SLOT 0 · busy
  ctx 8% · busy 2s · dec 59 · prompt 1%
  prompt   130 / 9,787 (1%)     ████░░░░░░
  decoded  59
  busy     2.0 s
  ttft     954 ms
  context  9,846 / 128,000 (8%) · 118,154 free

SERVER · 2m42s window
  58.3 / 4.2 tok/s · acc 38%
  prompt_tok/s · decode_tok/s · deferred · active · ctx peak
  acceptance + length lifetime figures from /metrics

• Busy from is_processing only — never from retained n_prompt_tokens (those stay high while idle).
• TTFT + busy-age latches computed host-side (the /slots payload has no timestamps).
• Wedged heuristic: prompt fully processed, zero decodes, busy ≥ 300 s (DSH streamIdleTimeoutMs) — called out in red, not silently “busy forever”.
• SERVER card from Prometheus /metrics (prompt/decode rates, deferred, active, context peak, lifetime acceptance/length). Figures use the sample envelope — bare lifetime .value is never mistaken for a percentage.

Ollama

Slot Health   ollama 0.22.1  127.0.0.1:11434   · loaded · model…
  latency <1 ms · updated 0 s ago

MODELS · 1 loaded / 12 in library
  batiai/qwen3.6-35b:iq3
    qwen35moe · 34.7B · 17.5 GB · VRAM 17.5 GB · keeps in 52 m …

• Honest states — not llama vocabulary:
  • up-no-model — reachable, /api/ps → {"models":[]} (up, nothing loaded; not an error).
  • up-loaded — ≥1 model resident. Resident ≠ busy — keep-alive evicts after the TTL; the dock chip says ollama · loaded, never busy.
• MODELS card: resident details + library count from /api/tags.

Fingerprint (never guessed)

Backend	Oracle shape
llama-cpp	/health → {"status":"ok"} and /slots → bare JSON array
ollama	/api/version → object with string version
vllm	/metrics text contains a vllm: series (stub — no dedicated sections yet)
unknown	reachable, none of the above

The engine word appears on the pane header and the dock chip (llama · busy, ollama · loaded).

Architecture

This is a dual-face plugin package: one npm package, two runtimes.

• Host face (lib/index.js, ESM) — registers the plugin, owns a 1 s sampling loop (SAMPLE_INTERVAL_MS = 1000), probes the configured origin, and serves a JSON snapshot at GET /api/dsh-slot-health on the same origin as the page.
• Client face (lib/client.js, CJS closure factory) — loaded by the web client via window.__ModuleLoader__.load({ id, factory }). Self-chaining 1 Hz fetch (no-store, abortable), renders the Slot Health rightbar tab and the footer dock chip.
• Glue — cordis.patch.yml inserts one Loader row; the browser half is discovered from the dsh.client declaration in package.json.

Client constraints, honored: only frozen PLATFORM_MODULES may be required at runtime (react, cordis, client store, ui slots/primitives/dockkit); everything else is inlined by esbuild. Presentation is inline styles only — no CSS files in the bundle. The plugin is read-only toward the origin (no load/unload/kill).

dsh-slot-health/
├── package.json          # dual-face exports: "." (host) and "./client" (browser)
├── build.mjs             # esbuild, two configs (host ESM + client CJS closure factory)
├── cordis.patch.yml      # Loader row + origin
├── media/                # README screenshots
├── src/
│   ├── host/
│   │   ├── index.ts      # plugin registration + sampling loop
│   │   ├── route.ts      # GET /api/dsh-slot-health
│   │   ├── collect.ts    # probes + fingerprint dispatch
│   │   ├── fingerprint.ts
│   │   ├── latch.ts      # busy-age / TTFT latches
│   │   ├── metrics.ts    # /metrics parse → SERVER card
│   │   └── ollama.ts     # /api/ps + /api/tags
│   ├── client/
│   │   ├── index.tsx     # slot + dock injection
│   │   ├── SlotBody.tsx  # pane: SLOT / SERVER / MODELS cards
│   │   ├── SlotTitle.tsx # tab chip
│   │   └── SlotDockChip.tsx
│   └── shared/
│       └── types.ts      # HealthSnapshot / SlotSample / Backend
└── lib/                  # prebuilt output (committed for install-without-toolchain)

Install

Requires DeepSeek Harness with a web profile and a local endpoint to watch.

From GitHub (users)

dsh plugin --profile web add github:janpauldahlke/dsh-slot-health
# restart dsh web (or rely on live patch reload), then hard-refresh the browser

lib/ is committed, so install does not require a local TypeScript/esbuild toolchain.

Point origin in the installed cordis.patch.yml (or your profile override) at the endpoint you want:

config:
  origin: http://127.0.0.1:8080   # llama.cpp
  # origin: http://127.0.0.1:11434  # Ollama

From a git checkout (developers)

git clone https://github.com/janpauldahlke/dsh-slot-health.git
cd dsh-slot-health
npm install          # prepare builds lib/ if toolchain present
node build.mjs       # optional: force rebuild → lib/index.js + lib/client.js

# wire into your web profile (absolute path; link: dep + bundle entry)
dsh plugin --profile web add "$PWD"

That updates ~/.dsh/profiles/web/package.json roughly to:

"dependencies": {
  "dsh-slot-health": "link:/abs/path/to/dsh-slot-health"
},
"dsh": {
  "profile": {
    "bundles": [
      "@deepseek-ai/dsh-base",
      "@deepseek-ai/dsh-web-app",
      "dsh-slot-health"
    ]
  }
}

Restart dsh web, hard-refresh. The rightbar Slot Health tab and footer dock chip should appear.

Dev loop

# edit src/ → rebuild (profile already link:s this tree)
node build.mjs
# host half often hot-reloads with patchReload: live; client half: hard-refresh

If the profile cannot resolve the package name:

ln -sfn "$PWD" "$HOME/.dsh/profiles/web/node_modules/dsh-slot-health"

Do not link only into a harness monorepo node_modules — Cordis resolves from the profile.

Verify

dsh --profile web --dump-config | grep dsh-slot-health
curl -s http://127.0.0.1:3080/api/dsh-slot-health | head   # adjust port
# expect JSON: state, backend, slots[] and/or ollama, optional metrics

Remove:

dsh plugin --profile web remove dsh-slot-health

How this was built

Closed-loop iteration on live hardware: research the DSH out-of-tree plugin contract, scaffold a dual-face package, build a read-only sampler against llama.cpp /health /slots /metrics, add Ollama fingerprint + MODELS, polish cards/tooltips/dock chip, and re-check every pass on a dedicated acceptance port — without touching sacred ports (:3080 main dsh, :8080 llama-server, :11434 ollama). Built as a companion to dsh-gpu-monitor-nvml on the same lab.

Authors

Built as a human ↔ local-model pair, not “AI did it” and not “human only reviewed.”

• Jan (hagbardCeline) — software engineer. Set the goal and honesty rules (fingerprint by shape, Ollama ≠ llama vocabulary, never lie about busy), drove the accept/reject loop against real endpoints, killed bad paths, and owned the final “ship this” call.
• Qwen3.8 27B Q6 HauHau (coauthor) — wrote the bulk of the implementation under that loop: host sampler, client UI, build glue. Served by llama.cpp across dual NVIDIA GPUs (RTX 4080 SUPER + RTX A4000). Ollama smoke used batiai/qwen3.6-35b:iq3 with OLLAMA_SCHED_SPREAD.
• Hardware — Ryzen 7 7800X3D, 61 GiB RAM, dual 16 GiB NVIDIA GPUs.

The interesting part is the recursion: an agent helping build a health pane for the endpoint it is running on, with a human in the loop the whole way.

---

License: MIT · Contributing