dsh-llm-opencode-go

English | 简体中文

Introduction

The usage panel and the quota capsule, as they ship on the web client.

A DeepSeek Harness (DSH) plugin that serves OpenCode Go (https://opencode.ai/zen/go/v1) as a native DSH model provider.

The plugin implements OpenCode Go's three wire formats itself: it translates harness messages, tools, and reasoning into each protocol's request, and translates the streamed response back into DSH StreamChunk values.

- insert:
    - id: opencode-go
      name: dsh-llm-opencode-go
      config:
        apiKeyEnv: OPENCODE_GO_API_KEY

That is the whole required configuration. apiKeyEnv names a credential reference, not the key itself: the key value never enters the config file, and the plugin resolves the reference through the harness credential seam on every request, so a key written by the web Models page reaches the next call with no restart.

Features

Provider route	opencode-go (configurable)
Protocols	OpenAI Responses, OpenAI-compatible Chat Completions, Anthropic Messages
Model discovery	GET /models, with a memory + on-disk cache
Session affinity	x-opencode-session, one stable id per conversation
Tool calling	Full round trip for every dispatched call: definitions, fragmented arguments, tool results, follow-up turn
Reasoning	Per-protocol effort mapping, including Anthropic thinking budgets
Streaming	SSE → StreamChunk, with usage, finish reason, and an idle watchdog
Failures	Stable provider-neutral codes (AUTH, RATE_LIMIT, QUOTA, CONTEXT_WINDOW_EXCEEDED, …)
Configuration UI	One page in the web client, with a row in Settings' own navigation and a Configure control on the plugin's row: API key, credential reference, model visibility, model variants, fetching the model list, usage
Fast entries	A quota capsule at the sidebar's foot (one ring per window — 5H / Wk / Mo — with what is left of each), a usage panel of its own (from that row, or the keyboard), and a usage page under Settings
Runtime	Node.js ≥ 20; no runtime dependencies

Install and use

Install

dsh plugin add dsh-llm-opencode-go

Then fully restart the profile: bundle layers are read at startup. The startup log should carry:

dsh-llm-opencode-go: provider "opencode-go" ready at https://opencode.ai/zen/go/v1 (credential OPENCODE_GO_API_KEY, models discover)

Give it a key

Either export OPENCODE_GO_API_KEY in the environment that launches DSH, or store it through the credentials seam — the web Models page writes it. The plugin resolves the reference per request, so both work and neither requires a restart.

See the usage

Usage has three entries, all reading the same facts:

Entry	Where
The quota capsule at the sidebar's foot	Beside Settings. One row, three groups — ◉ 5H 92% │ ◉ 周 58% │ ◉ 月 9% — visible without a click; that row opens the usage panel, and the gear beside it opens the plugin's configuration page.
The usage panel	From that row, or the keyboard (desktop Ctrl/Cmd+U): the subscription's three windows first, this route's counters under them, and a header that stays on screen as they scroll — the way to the plugin's configuration page, and an × back to the conversation.
Settings → Built-in plugins	A tab carrying the same content as the panel.

All three need the web client; a deployment without one has none of them and is otherwise unchanged. See Configuration.

Configuration

The config block above is all that is required; every other key is optional. The full field table, catalog shaping, and image input live in Configuration.

Point it at another credential reference, if you would rather the plugin read a differently-named variable:

config:
  apiKeyEnv: OC_KEY

The value is never read from the config file itself, only the variable name.

Or configure it from the web client. Settings carries an OpenCode Go settings row of its own, beside the pages the shell ships, and the same page is behind the sidebar's Plugins page: the dsh-llm-opencode-go bundle's opencode-go row has a Configure control. Either door opens the plugin's own page: the API key, the credential reference it is stored under, one switch per catalog model for what the picker offers, the model variants, a control that re-reads the service's model list, and what this route has spent. The key is written to the credential store, never to cordis.patch.yml, and the page is mounted only by a deployment that serves the web client. See Configuration.

Docs

Details live under docs/; start with what you need:

• Models — per-model protocol table, training-data gate, capacity snapshot, text/image split
• Configuration — full field table, shaping the catalog, image input, the graphical and fast entries
• Wire protocol — endpoints and auth, x-opencode-session, reasoning and tool-calling translations
• Reliability — stable failure codes and limits
• Development — project status, verification suites, re-measure probes, architecture

License

MIT