Provider Quick Config — a dsh dynamic plugin

中文版 | English

Plugin source lives in this repo: plugin/ holds the dynamic-plugin form (host.js = Host half, client.js = Client half, manifest.json = restorable definition); dsh-provider-quick-config/ is the formally installable npm package (Host via dsh.bundle, Web UI via dsh.client) — zero changes to deepseek-harness source (the repo's tracked files show no modifications; it is byte-identical to the cloud origin/master).

Screenshot

Two ways to use this plugin

Formal install (already done on this machine)

The package was built, packed (dsh-provider-quick-config-0.5.4.tgz) and installed into the web profile:

cd /path/to/deepseek-harness
# pack the package (in dsh-provider-quick-config/): npm pack
node --import tsx/esm apps/cli/src/bin.ts plugin --profile web add \
  file:/path/to/dsh-provider-quick-config/dsh-provider-quick-config-0.5.4.tgz
cd ~/.dsh/profiles/web && pnpm install   # only needed when the tarball path changed

dsh plugin add runs pnpm add in the profile and appends the package to dsh.profile.bundles when it declares dsh.bundle.patch. Verify:

cat ~/.dsh/profiles/web/package.json
# dsh.profile.bundles should include "dsh-provider-quick-config"

Then restart dsh web (e.g. deepseek.sh restart) — the Host half mounts from the profile bundle, and dsh-client-modules serves the Web half at /plugins/dsh-provider-quick-config/client.js. After restart: the + button appears next to Send (no per-session approval needed anymore).

Update after code changes: bump the version, re-pack, dsh plugin --profile web add file:…<new>.tgz (or update the dependency spec + pnpm install), restart. Uninstall: dsh plugin --profile web remove dsh-provider-quick-config, restart.

A + button next to the Send button in the DeepSeek Harness Web GUI. Click it to configure model providers without touching config files by hand:

• List every configured provider route (llm-pi-ai.providers.*) with credential status and the models each one serves.
• One-click vendor presets: 智谱 GLM, MiniMax, OpenAI GPT, Anthropic Claude, Local models (Ollama) — endpoint / protocol / thinkingFormat / model list are pre-filled.
• Model picker: per-vendor quick-model chips (click to select/deselect) plus free-form rows (id / ctx / max) — e.g. switch MiniMax to M3 by clicking the MiniMax-M3 chip, no JSON editing.
• Auto-validate model names on save (ping): before saving, the plugin queries the endpoint's GET /models and compares (case/dash-insensitive). A model that does not exist (e.g. typing glm4.6v instead of GLM-4.6) fails with a red error and blocks the save. Endpoints that do not support listing are skipped and saved normally.
• Auto-load local models: picking the Ollama preset immediately fetches GET /models and fills the form (a "Fetch models" button exists on every custom route too).
• Auto-sync local models (syncModels): routes marked syncModels: true (Ollama preset on by default, toggle in the form, badge in the list) are compared against their endpoint's GET /models every 60s in the background; when the endpoint list changes (e.g. you ollama pull a new model) the plugin writes the new list back to settings.yaml automatically and the dropdown follows — no manual editing. Endpoint order wins, capacities of already-configured models are preserved. Verified end-to-end.
• Multiple API keys per vendor: add the same vendor again and the route key auto-numbers (glm → glm2 → glm3…), display name gets · 号N.
• Media showcase (🎞): a persistent right-side panel that scans the current session's history and shows every media file the conversation mentioned — images (PNG/JPEG/WebP/GIF/AVIF/BMP/SVG/ICO), videos, recordings. SVG renders natively in <img>. Files are resolved from the session's working directory, with a filename search fallback (a file whose text reference is missing a path prefix, e.g. tiile-20260815/x.mp4 actually under outputs/tiile-20260815/, is still found). Items are grouped by conversation turn ("第 1 轮", "第 2 轮"…). Media data never enters the model context — display only. Each item can insert its path into the composer (inputActions.setDraft), and raster images can additionally be inserted as an attachment. The panel also auto-scans <cwd>/.uploads/ (SVG/通用上传) and <cwd>/.screenshots/ (截图库) so files written there appear automatically, grouped into separate "upload library" and "screenshot library" turns.
• SVG upload + insert path: the upload button accepts .svg (PNG/JPEG/WebP/GIF still go through as image attachments). SVG files are saved to <session cwd>/.uploads/<name> via the new save-media endpoint, and the path is appended to the composer as ![name](<abs path>) — visible inline once sent, and the SVG appears in the media showcase (it can also be dragged or pasted from the file system into <cwd>/.uploads/ to land in the showcase).
• Screenshot 📷: captures the screen via getDisplayMedia, saves the PNG into <session cwd>/.screenshots/ (the media-showcase library folder, gitignored) and writes the returned path into the composer — you decide whether to send it. The saved screenshot automatically appears in the showcase.
• Custom OpenAI-compatible providers (any gateway / self-hosted server) with hand-written api, baseURL, thinkingFormat, and model list.
• Edit / delete existing routes; set or update API keys.

How it works (no harness code changes — pure config files, hot-reload)

The plugin does not invent its own config store; it drives the two seams the harness already ships:

Only credential references (environment-variable names) are stored by the plugin; key values go straight into .credentials.yaml and never come back to the GUI.

Route writes go through settings.mutate, which validates against the llm-pi-ai namespace schema + assertServiceable: a wrong protocol, empty baseURL, or invalid model is rejected at write time — a broken route can never be stored.

Install / run

1. Approve the Run card in this session (single check is enough; double check auto-runs future versions).
2. Refresh the page — the + appears in the composer's tool row, next to the model selector / send button.
3. Click + → Add Provider → pick a preset or custom → fill the form → Save.

The official Settings → Models page still works alongside; this plugin is the quick entry point beside the input box.

Restoring the plugin after a process restart

Dynamic plugins live only in process memory: restarting dsh web removes pprov-1. Restore from this repo (the source never goes away):

1. code.host ← full content of plugin/host.js
2. code.client ← full content of plugin/client.js
3. Re-define with cordis_define (kind: "new", idPrefix: "pprov"; copy name / purpose from plugin/manifest.json), then cordis_run.

Your configuration (~/.dsh/settings.yaml + .credentials.yaml) is untouched, so routes reappear immediately.

Form fields

Pitfalls (learned the hard way)

• Ollama still wants a key: the OpenAI-compatible implementation always sends an Authorization header, so give it any non-empty placeholder (e.g. local).
• Catalog routes don't need a model list: key + credential only; endpoint/protocol/models are inherited from the pi-ai catalog. Hand-written routes declare models explicitly.
• .credentials.yaml is strict: a flat name: value map; the root must be a mapping, values must be non-empty strings, and duplicate keys fail the whole file at startup.
• *⚠️ Do NOT use YAML anchors (&anchor / `anchor) in settings.yaml**: the settings service saves with node-preserving merges; once a save replaces the anchor's owner (e.g.glm'smodels: &glm_models), remaining*glm_modelsaliases dangle and the whole document fails to serialize withUnresolved alias (the anchor must be set before the alias)` — both this plugin and the official Settings → Models page hit it. Write shared model lists explicitly (or ask the plugin/me to flatten the file first).
• Environment variables are a startup snapshot: export after launch does nothing until restart; use this plugin's key field (writes .credentials.yaml) instead.
• apiKeyEnv set but unresolvable → MISSING_CREDENTIAL; it will not fall back to some other key that happens to be in the environment.
• Safari screen-capture permission (📷 button does nothing / picker never opens): Safari's screen-share permission lives inside Safari itself, not System Settings. Fix: Safari → menu bar Safari → Settings… (⌘,) → Websites tab → Screen Sharing (or "Screen Recording") in the left list → find your host (127.0.0.1:<port> or the domain you use) → change it from Deny to Ask (or Allow) → completely quit Safari (⌘Q) and reopen (a plain refresh is not enough — WebKit bug 253024 keeps the failure sticky across reloads). Also confirm System Settings → Privacy & Security → Screen Recording has Safari checked. If you once clicked "Don't Allow" and it never asks again: that decision is remembered per-site — go back to Safari → Settings → Websites → Screen Sharing, find the site, and delete the row entirely (remove, don't just switch the state), or set it to Ask; then quit Safari (⌘Q) and reopen — the picker prompt comes back. If it still fails, use the system screenshot instead: ⌘⇧4 then paste (⌘V) into the input box (the composer natively accepts pasted images).
• Sessions that already sent requests keep the model recorded in their log; if the default model points to a deleted provider, the picker asks you to choose again.

Technical notes (for plugin maintainers)

• Host half (harness.handle RPC, all through ctx.settings / ctx.credentials / ctx.llm):
  • providers.list → configured routes + credential state + pi-ai directory + protocol/dialect enums
  • providers.save / providers.remove → ops are built Client-side and arrive over the wire; the Host only forwards them to settings.mutate('llm-pi-ai', ops, revision)
  • providers.discover → ctx.llm.discoverModels('llm-pi-ai', { baseURL, api, apiKey })
  • providers.ping → same discovery, then validates each model id (normalized) against the listing
  • credentials.set / credentials.unset
• ⚠️ vm-sandbox realm trap: dynamic Host code runs in a node:vm realm, so any object literal is not a host-realm plain object — passing one to a service that checks isPlainObject (like settings.mutate / settings.update) throws ops must be {op:'set'|'unset', path}. Any structured data for such services must be constructed Client-side and passed through the host.call JSON wire (wire-decoded values are host-realm); the Host only forwards. Return values are fine — the guard re-materializes them.
• Client half: the + button registers in conversation.input.right (tool row before the send button); the panel registers in conversation.input.overlay (the composer's floating anchor); all colors use --dsw-* theme variables.
• Writes carry expectedRevision; a concurrent edit raises SETTINGS_CONFLICT and the client re-reads and retries once.

Repository layout

DeepSeek-Harness-provider/
├── README.md          # this file (English)
├── README.zh.md       # 中文版
├── plugin/
│   ├── host.js        # Host half source (code.host body)
│   ├── client.js      # Client half source (code.client body)
│   └── manifest.json  # restorable definition (name/purpose/code)
└── .gitignore