dsh-opencode-models

English | 中文

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

What it does

DeepSeek Harness serves OpenCode Zen through two llm-pi-ai provider routes configured in ~/.dsh/settings.yaml:

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.

Agent tools

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.

"OpenCode Models" settings section

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.

Install

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.

Configuration

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.

Assumed capacities

The zen listing discloses nothing but ids, so any entry adopted by id needs capacities from somewhere. This plugin's policy:

• Explicit models entries win — pass real figures when you know them.
• Otherwise assumeDefaults: true (tools) or accepting the panel's checkbox flow fills contextWindow: 128000, maxTokens: 32000, input: ["text"].
• Every assumed value is reported back by name; correct them via 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.

Development

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).

Verification

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

License

MIT

Muse 接入方案

OpenCode Go 的 muse-spark-1.2-contributor（地区受限模型）接入指南（独立路由 + 代理 + 思考档位），见 MUSE.md。