dsh-ofm-model-manager
English | 中文
Per-model enable/disable and renaming for every provider in DeepSeek Harness.
Adds a panel at the bottom of Settings → Models that toggles each model of each provider and gives it a custom display name.
- Disable a model → it disappears from the composer's model picker and the
/model dialog. A provider whose models are all disabled disappears from the picker as a group.
- Give it a custom name → the picker shows your name instead of the built-in one. The model id is untouched.
- Preferences live in the profile and apply immediately — no page reload, no DSH restart.
Model manager panel at the bottom of Settings → Models
The panel at the bottom of Settings → Models: search box, status chips and provider filter on top, per-provider groups with per-model switches and custom names below.
A row with a custom name enabled
A row with "custom name" enabled — switch, checkbox, name field and "restore default".
Install
One line from GitHub:
sh dsh plugin --profile desktop add github:DeepseekDays/dsh-ofm-model-manager
(use your own profile name instead of desktop if different). Then fully quit DSH, tray included, and start it again — new bundles are read at startup.
From a checkout, node install.mjs does the same with a directory junction and without running pnpm (running pnpm while DSH is live rebuilds node_modules and can interrupt a running session):
powershell node install.mjs # default: desktop profile node install.mjs --profile web # another profile node install.mjs --dry # show what would change node install.mjs --uninstall # remove junction + both manifest entries, keep the folder
It backs up <profile>/package.json before touching it.
What the panel gives you
| Control | Scope |
|---|
| Search box | exact match on name, custom name, model id, description, provider name |
| Status chips | All / Enabled / Region-restricted / Disabled, each with a count |
| Provider dropdown | one route at a time, with its model count |
The three are ANDed. Rows are grouped by provider, and each group header shows name · enabled/total plus a group-level switch ("enable all" when everything is off, otherwise "disable all"). Groups collapse; collapsing is render-only and does not change filtering or counts.
Batch buttons act on the current view (the filtered result) and say so in small print underneath — there is deliberately no silent "select the whole catalog".
Each row is effective name + status badge / model id / description, followed by switch → "custom name" checkbox → name field (disabled until the checkbox is on, showing the built-in name). A custom tag marks rows that use a custom name; a wildcard tag marks names that apply to every provider (see below).
Long lists render 40 rows at a time with a "show more (N left)" button, and the budget resets whenever the filter changes — no virtual-scrolling library, this is a hand-written ModuleLoader bundle.
Empty and error states are explicit: nothing listed at all, everything disabled (with a one-click "enable all"), no search results (echoing the keyword), plus a read failure state with a retry and a write failure state that rolls back.
Requirements and compatibility
- A DSH that provides the
settings.models.footer slot and @deepseek-ai/dsh-client-ui-primitives@ (tested against ``deepseek-ai/dsh` 0.2.0-rc.2).
- Declares
dsh.bundle in package.json, so it installs with plain dsh plugin add.
- No kernel files are modified. The plugin wraps
ctx.llm.listModels() at runtime and restores it on unload (reference-counted).
- Coexists with
dsh-provider-toggle (provider-level switching). Both wrap the same own property on the llm instance; providers-level unloading deletes it, so this plugin re-installs its wrapper on the next config read whenever the own property has disappeared entirely — and deliberately does not stack a second wrapper when someone else merely wrapped it again.
- Unaffected by plugins that inject models deeper down (e.g. pi-ai
getModels): filtering happens at the listModels exit, so injected models are managed and displayed like any other.
- Runtime dependencies: none.
Preference keys
provider|modelId only that model on that route modelId no separator = every provider (wildcard)
Why this shape: model ids repeat across providers (deepseek-v4.1-flash exists under several), so a bare-id key silently disabled all of them — v2 keys carry the provider. Bare ids from v1 config read as wildcards, so v1 configuration keeps working with zero migration. One field with one element type keeps the schema trivial, and the pipe is a plain character that YAML does not need to quote, so keys stay readable inside cordis.patch.yml.
Exact key beats wildcard key, so "rename for everyone" and "rename this one specifically" can coexist. The panel only ever writes exact keys; wildcards come from v1 config or a hand-edited file, and are shown as a wildcard tag in the UI.
Development
sh node test/all.mjs
Four offline suites — no DSH, no network, no browser:
| File | Covers |
|---|
test/manager.test.mjs | pure logic: key format, wildcard/exact precedence, route claim, catalog merge, three states, search, TTL cache |
test/host.test.mjs | apply() against a fake ctx: does a disabled model really leave the catalog, is the panel's read-only endpoint still served, broadcast, trust gate, shared patch slot with provider-toggle |
test/client.test.mjs | loads client.js through the ModuleLoader protocol and really renders the panel with a minimal React stand-in: grouping, counts, collapsing, empty states, batch scope, partial rendering, write shape and rollback |
test/manifest.test.mjs | manifest consistency: package name, bundle, slot, field names, separator, self-containment |
The suites re-implement the expected behaviour (they replicate DSH's groups.filter(group => group.models.length > 0) rather than calling the code under test) so they cannot pass by agreeing with themselves.
Troubleshooting
| Symptom | Cause / fix |
|---|
| Panel does not appear | the bundle was not loaded — check dsh.profile.bundles in the profile package.json, then fully quit DSH (tray included) and start again |
| "cannot read this plugin's settings namespace" | same as above; also check that the entry id in cordis.patch.yml matches the injected namespace |
| The switch seems to do nothing | the wrapper was removed when another plugin unloaded. It is re-installed on the next config read; restart DSH if it persists |
| No name field | the host half is still v1 (no names field in the namespace) — fully restart DSH |
| Picker did not change after renaming | wait for the 500 ms write debounce, then look for a failed settings.mutate in the console |
| Panel flickers while typing | fixed (see limit 8). A browser-half hot reload is enough; if it persists, look for this plugin's listener failed or settings/conflict |
| Start from scratch | delete the config of the id: dsh-ofm-model-manager entry in the profile's cordis.patch.yml |
Known limits
- Disabling affects the catalog, not routing. A session already using the model keeps working; new sessions simply cannot pick it. That is the kernel's behaviour and this plugin does not change it.
- Renaming only changes the display name, never the model id — config, session logs and credential references keep the original id.
- Provider-level switching is not taken over: that is
dsh-provider-toggle's job. Both can be installed and active at once.
- Settings are written as one whole volatile form, so preferences are last-writer-wins; two tabs editing at once get a
settings/conflict (the panel reports it and rolls back rather than dropping data silently).
- Wildcard keys are a global broadcast — tagged in the UI so "I changed one model, why did everything change" cannot happen unnoticed.
- Partial rendering instead of virtual scrolling: more than 40 rows need "show more"; filtering is the intended tool for long catalogs.
- The read-only endpoint lives under
/api/ofm-model-manager (longer prefix than the kernel's /api, and the web server prefers the longest prefix, so it runs ahead of connection's own admission check) — hence its own equivalent trust gate (loopback host, cross-site fetch refused, same-origin Origin/Referer), preferring the connection service's admission check when the composition provides it.
- Renaming does not lock or repaint the panel: name writes use a separate path that never sets
busy (otherwise every row's checkbox would grey out and back while you type), and only fields the UI actually renders trigger a repaint (RENDER_KEYS in client.js). Accounting-only fields are excluded — add new fields to RENDER_KEYS or their updates will silently not repaint (there is a case pinning this).
- On hot-reload the pending debounced write is flushed before the old instance hands over (
clearTimeout then an immediate flushNames()), so what you typed in the last 500 ms survives an HMR swap instead of being written from stale state.
The Chinese README (README.zh.md) documents the implementation in more depth: how the catalog is filtered, why disabling the last model of a route removes the whole group, the read-only endpoint's trust gate, the optimistic-update and rollback path, and the transition state between a hot-reloaded browser half and a not-yet-restarted host half.
Files
index.js host half: catalog filtering + read-only endpoint + state client.js browser half: the Settings panel src/manager.js pure logic: key format, catalog merge, cache, search src/trust.js trust gate for the read-only endpoint cordis.patch.yml bundle declaration install.mjs install / uninstall / dry run test/ offline suites assets/ screenshots used by this README
License
MIT — see LICENSE.
