dsh-lineup-manager

Pick which model lineups appear in DeepSeek Harness's model picker — grouped by provider, and refreshed live from each platform.

English · 简体中文 · Español · Português · हिन्दी

Compatibility

The problem

DSH groups its model dropdown by provider, but how many entries show up depends on your config:

• a provider that declares a models: list in cordis.patch.yml → only those lineups appear
• a provider with no models: list → the picker shows the adapter's entire catalog, which can be hundreds of entries

So keeping only the free lineups means hand-editing YAML. This plugin turns that into checkboxes in Settings.

Platforms change their lineup every few days — new models appear, free tiers vanish. Refresh pulls the live list and marks the difference, so you can see what moved without leaving the panel.

Features

• Per-provider checkboxes for every declared lineup
• Bulk per provider: All / None / Free only
• Refresh — pulls the platform's current lineup list and flags new and gone entries. Available per provider and as Refresh all
• Live-added models become selectable immediately, before they exist in your config
• Writes are surgical and reversible: only the target provider's models: block is replaced, with a timestamped backup, a structural check, and a read-back verification that rolls back on mismatch

Install

From a tarball (recommended)

pnpm pack                       # produces dsh-external-dsh-lineup-manager-.tgz
dsh plugin add ./dsh-external-dsh-lineup-manager-0.2.0.tgz

Restart DSH afterwards.

From a git checkout

dsh plugin add github:/dsh-lineup-manager

The prepare script builds dist/ on install. pnpm >=10 refuses to run git dependencies' build scripts by default; if the first add fails, add this to the profile's pnpm-workspace.yaml and retry:

allowBuilds:
"@dsh-external/dsh-lineup-manager": true

Only allow this for sources you trust, and pin a commit: github:/dsh-lineup-manager#.

Manual

pnpm install
pnpm build                     # emits dist/ and verifies it

Then copy the package into $DSH_HOME/profiles//node_modules/ and restart DSH.

Usage

1. Open Settings → the "Lineups" button at the top
2. The panel lists every provider in your cordis.patch.yml with a checkbox per lineup
3. Per provider, use All / None / Free only, then Save that group
4. Restart DSH for the model picker to reflect the change

Unsaved changes are flagged; closing the panel asks for confirmation.

Refresh

Press Refresh on a provider (or Refresh all in the footer) to pull that platform's current list:

• a green new badge = the platform offers it, you have not checked it
• a struck-through grey gone badge = you checked it, the platform no longer offers it

Refresh never writes. It only labels. Nothing reaches the config until you check boxes and press Save, which still goes through the backup / structural check / read-back-rollback path.

Strongly recommended: declare a baseURL for every provider

openrouter:
apiKeyEnv: OPENROUTER_API_KEY
api: openai-completions
baseURL: https://openrouter.ai/api/v1

Without a baseURL, the plugin must infer the endpoint by reading the installed pi-ai catalog out of app.asar. That path depends on the host environment and fails poorly when it does — as an unwritten response the client can only report as HTTP 400 (empty body). Declaring baseURL goes straight to the network and never touches that code.

Verified endpoints:

A baseURL change needs a DSH restart.

Safety

Editing a config file is risky — a corrupted cordis.patch.yml can stop DSH from starting — so writes are guarded four ways:

Also: unknown providers are refused with 409; duplicate or whitespace-containing ids are rejected; request bodies are capped at 256 KB.

Backups accumulate. Clean up cordis.patch.yml.bak-lineup-* once you are happy with the config.

How refresh works (and the traps)

It must bypass pi-ai's built-in catalog

discoverModels returns the installed catalog — frozen at your DSH version — the moment it receives a provider argument. That is silent staleness: the feature looks like it works while showing months-old data. Refresh therefore sends baseURL and never provider. Guarded by test-refresh.mjs (R2).

Namespace

dsh-llm-pi-ai registers model discovery under its own loader entry id — llm-pi-ai as declared in the profile. This plugin runs under a different entry, so deriving the namespace from ctx.fiber.entry.options.id yields dsh-lineup-manager and the kernel answers NO_DISCOVERY. The namespace is a constant. Guarded by test-refresh.mjs (R9).

asar reads

The catalog lives inside app.asar (~174 MB). Reading it whole costs ~125 ms and hundreds of MB of RSS, multiplied across a refresh-all. The reader seeks: header plus the one entry's byte range. Measured 4 concurrent resolutions drop from ~905 MB to ~69 MB. Guarded by test-asar.mjs.

Keys that expired

A platform answering 401 api_key_expired makes that provider show a failure badge. Other providers are unaffected and your config is untouched — refresh reporting this is the point, not a bug.

Layout

src/                  runtime sources
index.js            host half: HTTP API over cordis.patch.yml
client.js           browser half: the Settings panel
cordis.patch.yml    bundle patch
scripts/build.mjs     emits dist/, verifies the copy
tests/                plain-Node test suites
dist/                 build output (not tracked)

Development

pnpm build        # emit + verify dist/
pnpm test         # 66 checks across four suites
pnpm test:live    # hits real provider endpoints; needs real credentials

The four suites run entirely in temporary directories and never touch a real profile. test-live.mjs does hit the network and reads your real credentials; it also asserts the config is unchanged afterwards.

License

MIT