dsh-malko-prefs

dsh-malko-prefs

Vibe coded. This plugin was written almost entirely by an AI assistant (OpenCode, deepseek-v4.1-flash) from natural-language prompts, with light human review and testing. It is not a carefully hand-engineered product: read the code before trusting it, expect rough edges, and check what it changes in your dsh profile — in particular it overrides your agent presets with a retargeted copy of the installed ones (see Install). Use at your own risk.

Personal preferences for DeepSeek Harness (dsh): a tunable companion to the official compaction engine, a llama.cpp model enrichment for the Models page, and tab notifications, packaged as a single dsh bundle.

• Compaction — the official @deepseek-ai/dsh-compaction-basic engine is the reference behaviour; this plugin wraps it with a settings page so the threshold, the retention, the summarization model/reasoning and an optional end-of-turn compaction become user-configurable. It also adds a /force-compact command.
• Model enrichment — reads the richer model metadata a llama.cpp server exposes (meta.n_ctx, architecture.input_modalities, aliases) and fills the corresponding llm-pi-ai model entries (context window, text/image input).
• Notifications — the browser tab icon turns amber while a question/approval waits for you, blue while a session works, and green when a session finishes unattended; the browser raises a notification on either event with an optional sound. Ported from dsh-notice-center (MIT).

Requirements

• dsh ≥ 0.1.7-rc.1 — built on the current settings domain (Config fields marked .volatile() on the Host, ctx.configForms on the Client) and on BasicCompactionEngine. Tested on 0.1.7-rc.2 and 0.2.0-rc.2.
• A working llm-pi-ai provider (e.g. your llama.cpp server) for the model enrichment.

Install

From the registry (once published):

dsh plugin --profile web add dsh-malko-prefs
# then restart dsh web

Or from a local checkout / tarball:

dsh plugin --profile web add /path/to/dsh-malko-prefs
dsh plugin --profile web add ./dsh-malko-prefs-0.1.0.tgz

The bundle layer does two things (see cordis.patch.yml):

1. inserts the Host half: the settings service malkoPrefs and the malkoModels probe Service, whose strict Typert wire definition is contributed through the package ./typert export that @deepseek-ai/dsh-typert-loader registers automatically on mount;
2. overrides every shipped agent preset (preset-cordis, preset-minimal, preset-ptc, preset-standard) with the preset the installed dsh actually ships, retargeted at dsh-malko-prefs/compaction.

A patch replaces a preset row's config wholesale, and the compaction backend lives nested inside the preset, so the whole preset must be supplied. Instead of freezing a copy, cordis.patch.yml asks dsh-malko-prefs/preset-config for the body at load time:

- id: preset-standard
  name: 'dsh-malko-prefs/preset'
  config: !!js process.getBuiltinModule('node:module').createRequire(baseUrl)('dsh-malko-prefs/preset-config')('standard', baseUrl, ctx)

dsh-malko-prefs/preset is a small wrapper around the upstream @deepseek-ai/dsh-agent-preset: the upstream row is an EntryGroup, so the Loader deliberately skips interpolation of its config; the wrapper is an ordinary plugin, so the computed config is evaluated and then registered with the same agentPresets.register() call.

src/preset-config.cjs locates @deepseek-ai/dsh-web-app/presets/<id>.patch.yml in the running dsh, parses it with the Loader's YAML dialect (so !!js nodes survive), evaluates those nodes against the Loader context (they drive disabled rows and customSkillDirs), and swaps @deepseek-ai/dsh-compaction-basic for dsh-malko-prefs/compaction. The plugin therefore follows the installed dsh version instead of requiring a regeneration after every upgrade.

Rollback: dsh plugin --profile web remove dsh-malko-prefs (then restart).

Compatibility

One release supports every dsh that ships @deepseek-ai/dsh-web-app/presets/{cordis,minimal,ptc,standard}.patch.yml (currently 0.1.7-rc.1 and later). Because the preset bodies are read from the running dsh, a dsh upgrade changes them without a plugin release; engines.dsh declares the lower bound. There is no separate “old dsh” package line.

Settings — “Malko's prefs”

Adds a section to Settings, split into three sub-tabs: Context compaction, Notifications and llama.cpp models (the tab bar stays pinned while the panel scrolls). All fields are applied live (no restart).

Context compaction

Compaction threshold

Effective trigger: min(window × ratio, window − output − headroom); when an absolute threshold and a context window are set, it becomes exactly that threshold. Default: ratio 0.8 + headroom 32768 → compacts around 160k on a 200k window.

Retention

Keep last (tokens) (absolute, e.g. 32k) or Keep ratio — the verbatim recent-context budget.

Behaviour

• Automatic compaction — the official between-step pressure compaction and context-overflow recovery.
• Compact at end of turn — runs one more compaction when the agent goes idle.

Summarization

• Model — Session model (default) or Custom model.
• Provider / Model — shown when Custom model; populated from the llm-pi-ai catalog.
• Reasoning — Default, Off, or any level the selected model declares (session mode exposes only Default / Off). The instruction is appended after the replayed prefix, so the provider's prompt cache stays warm.

Advanced

Summary output cap, Extra compaction attempts, Overflow recovery attempts.

llama.cpp models

For every llm-pi-ai provider that declares a baseURL, Fetch from server reads GET {baseURL}/models. It then shows a review of every discovered model and what importing would change: context window (from meta.n_ctx/n_ctx_train), output cap, and input modalities (from architecture.input_modalities); models that do not exist yet are marked new model. Tick the models you want (or use Select all) and click Import — only the selected ones are written (existing entries are updated, new ones appended). The request authenticates as the provider does: the endpoint's apiKeyEnv credential is resolved (dsh credentials, then the host environment) and sent as Authorization: Bearer …, so an api-key-protected server works.

Notifications

Two independent features, both driven by the official client signals (sessions, uiSession.sessionStatus); nothing is persisted beyond the config.

Tab status light — recolours the browser tab icon:

Priority is amber > working > green > idle, and sub-agent sessions are ignored. All icon links are repainted together (DSH ships a dark and a light favicon selected by prefers-color-scheme).

System notifications — settings are grouped by notification type. One global master switch (Enable notifications, opt-in; the browser asks for permission once), one global Notify in the foreground (off = only fire when the tab is not visible and focused) and one global Volume. Then one group per type — When a session finishes and While waiting for you (question / approval / plan review) — each with its own Notify enable, Keep on screen (stay until dismissed) and sound.

Sound (per type) — each group's selector defaults to No sound (both defaults are silent on a fresh install). The list is: No sound, two built-in synthesized chimes (Chime Up / Chime Down, no asset needed) and 45 bundled opencode sounds (see assets/audio/README.md); picking one previews it and the Host serves the mp3s at /malko-prefs-sounds/<id>.mp3. If that route is unavailable the player falls back to a chime — but browsers may block audio until you have interacted with the page at least once.

The sound is decoupled from the banner: once a sound is chosen for a type it plays on that event (completion / waiting) even when notifications — or that type — are disabled; with No sound nothing plays.

System notifications need a secure context — http://127.0.0.1:PORT or localhost. Opened over a LAN IP the Notification API is unavailable (a browser rule, not a plugin one); the tab status light still works.

Commands

• /compact — the official dsh command (idle manual compaction).
• /force-compact — compact now: immediately when the agent is idle, otherwise queued and consumed at the next model step (bypasses the threshold).

How it works

host plugin  (lib/index.mjs)
  ├─ Config (volatile)  ──▶ settings page "malko-prefs"
  ├─ service malkoPrefs ──▶ live prefs + force queue
  ├─ service malkoModels ──▶ llama.cpp probe (wire def in lib/typert.host.mjs)
  ├─ /malko-prefs-sounds/<id>.mp3 ──▶ bundled sound library
  └─ /force-compact command

client plugin (lib/client.js)
  ├─ mounts its own malkoModels Remote via ctx.remote.$mount()
  ├─ tab status light + notifications (src/notify.ts, optional uiSession)
  └─ settings page: 3 sub-tabs (compaction / models / notifications)

preset plugin (lib/compaction.mjs)
  └─ MalkoCompactionEngine extends BasicCompactionEngine
       ├─ rebuilds its policy from malkoPrefs before each operation
       ├─ summarize() → model + reasoning control (cache-safe order)
       └─ agent/status idle → optional turn-end compaction

The engine reuses the official engine for everything else (durable transactions, pruning, overflow recovery), so behaviour matches compaction-basic unless a preference changes it.

Note: service classes avoid JS private members (#x) — cordis binds service methods to a proxy, and private members would throw Receiver must be an instance of class ....

Development

node build.mjs            # esbuild → lib/{index,compaction,typert.host}.mjs + lib/client.js
node scripts/check.mjs    # static conformance checks

Publishing

The package is publication-ready:

• private is removed; LICENSE (MIT) and author are set; files ships lib/, cordis.patch.yml, scripts/, assets/ and LICENSE.
• prepublishOnly runs the build and scripts/check.mjs.
• Conformant bundle: dsh.bundle.patch, exports["./client"] and exports["./package.json"], the client bundle id equals the package name, and the client requires only baseline platform modules (react, @deepseek-ai/dsh-client-ui-primitives).
• The name is free on npmjs.com; release with npm publish (unscoped/public).

Caveat: engines.dsh declares the supported dsh range. The preset bodies are read from the installed dsh at load time, so the plugin follows preset content changes automatically; only a new preset id requires adding a row to cordis.patch.yml and src/preset-config.cjs.

Limits

• The preset retarget reads @deepseek-ai/dsh-web-app/presets/*.patch.yml through the dsh install and uses js-yaml; a future dsh that moves or reformats those files will fail loudly at preset mount instead of silently mounting a stale copy.
• The /force-compact busy path uses the engine's context-overflow entry, which compacts the maximal safe head (retention 0).
• The model enrichment is host-side (the browser cannot reach a local llama.cpp server cross-origin).
• Browser notifications require a secure context (127.0.0.1/localhost); the tab status light's green/amber state is in-memory and resets on reload.
• Notification sound is best-effort: browsers may block audio until the page has had a user gesture, and if the /malko-prefs-sounds route is unavailable the player falls back to a synthesized chime.