dsh-opencode-models
在 settings.yaml 中管理 OpenCode Zen 免费层级和 Go 层级模型：从 opencode.ai 获取实时模型列表，查看每个路由的已配置与实时状态差异，并通过四个代理工具及 OpenCode Models 设置部分进行添加/移除。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:wyouwd1/dsh-opencode-models#3783acac043f158ca0f23470aea7f29ee3f9ec99Opencode Models 以 dsh-opencode-models 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Manage OpenCode Zen free-tier and Go-tier models in DeepSeek Harness: fetch the live model listings from opencode.ai on demand, compare them with what your routes actually serve, and add or remove entries — through four agent tools and an "OpenCode Models" settings page. The free tier is an explicit official allowlist from the OpenCode Zen docs pricing table (big-pickle, ox-alpha-free, mimo-v2.5-free, hy3-free, nemotron-3-ultra-free, nemotron-3.5-lightning-free, muse-spark-1.2-contributor-free): paid ids the free endpoint also advertises never count, drift or get offered, and suffix-only ids outside that table are not adoptable either.
dsh plugin --profile web add github:wyouwd1/dsh-opencode-models
DeepSeek Harness serves OpenCode Zen through two llm-pi-ai provider routes configured in ~/.dsh/settings.yaml:
| Route | Tier | Endpoint | Model ids |
|---|---|---|---|
opencode | Free (PI_AI_API_KEY) | https://opencode.ai/zen/v1 | usually -free-suffixed |
opencode-go | Go subscription | https://opencode.ai/zen/go/v1 | unsuffixed |
The listing endpoints disclose ids only — no context windows, output caps, modalities, or reasoning levels — and the lists change over time (limited-time models get delisted). This plugin keeps the two layers honest against each other. The same model often has different ids per tier (x-preview-f-free vs x-preview-f); it never mixes them.
| Tool | What it does |
|---|---|
oc_model_status | Fetches both listings live and reports per tier: configured count, live count, online-but-not-configured ids, delisted-but-still-configured ids. Read-only. |
oc_model_add | Adds entries to one route. Pass ids copied verbatim from a status report or full models entries. Ids whose capacities the listing hides require assumeDefaults: true (contextWindow 128000, maxTokens 32000, input text) — every assumption is reported so you can correct it later. An id present only in the other tier's listing is refused with the two-tier id rule explained. |
oc_model_remove | Removes configured ids from one route; reports not-found ids and skips the write when nothing matched. |
oc_model_sync | Previews drift for both tiers by default. With apply: true it adds every online-not-configured id (assumed capacities flagged); with pruneStale: true it also removes delisted ids. |
Model-list edits take effect on the next request — no restart. This plugin never adds or removes whole routes; declaring a new route (apiKeyEnv/baseURL) stays on the Models page and still needs a restart.
A section pinned to the TOP of the web settings sidebar shows exactly two cards in order — OpenCode Zen free tier, then Go tier — nothing else. Every configured row has a checkbox and a bulk delete-selected bar trims any mix of models across the two lists in one confirmation, removing them from the conversation model list on the next request: configured entries with a "delisted" flag on stale ones, checkboxes over online-not-configured ids, per-row remove, and a two-click sync (preview → confirm). It reads and writes exclusively through the existing configuration-page contracts — settings.describe / settings.update with expectedRevision, and llm.discoverModels for the live listings — and refreshes only while open when pushed invalidations arrive.
Writes are guarded by the settings revision: if the Models page or another session writes first, this plugin re-reads once instead of overwriting it, and schema-invalid candidates are refused before anything persists.
From a checkout:
dsh plugin --profile web add /path/to/dsh-opencode-models
dsh --profile web
From GitHub (sources install without any build step — the shipped lib/ artifacts are plain ESM):
dsh plugin --profile web add github:wyouwd1/dsh-opencode-models
Requirements: DeepSeek Harness ≥ 0.1.1-rc.1 (the llm-pi-ai adapter family, settings seam, and web app), Node ≥ 22.
No config keys. The two tier definitions (route keys, base URLs) follow the official OpenCode Zen endpoint layout and the shared credential comes from each route's existing apiKeyEnv (PI_AI_API_KEY). Missing credentials surface the endpoint's 401 message verbatim in tool output and in the panel.
The zen listing discloses nothing but ids, so any entry adopted by id needs capacities from somewhere. This plugin's policy:
models entries win — pass real figures when you know them.assumeDefaults: true (tools) or accepting the panel's checkbox flow fills contextWindow: 128000, maxTokens: 32000, input: ["text"].oc_model_remove + a full re-add whenever you learn better numbers.Assuming wrong capacities degrades context filing and output caps; it never blocks requests.
npm test # node:test suites over lib/shared.js, lib/writer.js, lib/tools.js
The host face (lib/index.js) imports no @deepseek-ai/* package: tools register as plain definitions and all services resolve through the cordis context at call time. The browser face (lib/client.js) ships as the closure-factory artifact the module loader expects and requires only platform-seeded modules (react, dsh-client-ui-primitives).
Reproduce the install smoke test against an isolated DSH home (no effect on your real ~/.dsh):
DSH_HOME=/tmp/ocmm-home dsh plugin --profile web add /path/to/dsh-opencode-models
DSH_HOME=/tmp/ocmm-home dsh --profile web --dump-config | grep opencode-model # shows the "# == dsh-opencode-models" layer
DSH_HOME=/tmp/ocmm-home dsh --profile web --port 3101 --no-open # then open http://127.0.0.1:3101
curl -f http://127.0.0.1:3101/plugins/dsh-opencode-models/client.js # the served browser half