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 restates your agent presets (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. restates every shipped agent preset (preset-cordis, preset-ptc, preset-standard) with the official compaction-basic row retargeted at dsh-malko-prefs/compaction.

This second point is why the file is generated, not hand-written: a patch replaces a row's config wholesale and the compaction backend lives nested inside the preset, so swapping it means copying the whole preset. Regenerate it after a dsh upgrade that touches the presets:

node scripts/gen-preset-override.mjs                 # auto-detects dsh on PATH
node scripts/gen-preset-override.mjs <presets-dir>   # or pass it explicitly
DSH_PRESETS_DIR=<presets-dir> node scripts/gen-preset-override.mjs

The generated cordis.patch.yml is tied to the preset composition of the dsh version it was generated against — the plugin targets whichever dsh ships those presets (declared through its optional peerDependencies).

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

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: cordis.patch.yml is generated from the shipped presets, so regenerate it against the dsh version you target and keep the peerDependencies ranges in sync before publishing.

Limits

• The generated cordis.patch.yml copies the shipped presets; re-run gen-preset-override.mjs after a dsh upgrade.
• 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.