dsh-opencode-provider

Use models available through a local OpenCode server inside DeepSeek Harness without a DeepSeek API key.

Architecture

DeepSeek Harness (DSH)
  └─ LLM Runtime
       └─ opencode-local provider route
            └─ OpenCodeAdapter (dsh-opencode-provider)
                 ├─ POST /api/session              → create session
                 ├─ POST /api/session/{id}/prompt   → serialized conversation
                 ├─ GET  /api/session/{id}/event    → SSE stream
                 └─ POST /api/session/{id}/interrupt → best-effort cancel
                          ↕ HTTP + SSE
                   OpenCode Server (127.0.0.1:4096)
                          ↕
                   Model (mimo-v2.5-free)

The adapter serializes the full DSH conversation (system prompt + messages) into a single prompt string, sends it to the OpenCode server, and translates the SSE event stream into DSH StreamChunk values. OpenCode runs its own internal agent/tool loop, so the adapter exposes text output only — structured DSH tool calling is not available.

Verified Features

• No API key required — connects to a local OpenCode server over plain HTTP
• Streaming — text deltas are delivered as they arrive
• Cancellation — abort signal tears down the SSE stream and interrupts the session
• Timeouts — configurable SSE establishment (10 s), prompt admission (10 s), and terminal completion (120 s) limits
• Retry policy — provider-level retry config forwarded to DSH runtime
• Live SSE event formats — handles both legacy (session.text.delta, session.execution.succeeded) and current (session.next.text.delta, session.next.step.ended) event envelopes
• Graceful fallback — unknown SSE events are skipped; non-JSON lines are ignored; data.wrapper envelopes are unwrapped defensively

Requirements

• Node.js ≥ 22
• OpenCode server running locally (default: http://127.0.0.1:4096)
• DeepSeek Harness ≥ 0.1.0-rc.8

Installation

Via DSH plugin command

dsh plugin --profile web add dsh-opencode-provider

From source checkout

git clone https://github.com/goku54477/dsh-opencode-provider.git
cd dsh-opencode-provider
pnpm dsh

Quick Start

1. Start the OpenCode server:

opencode serve --hostname 127.0.0.1 --port 4096

2. The adapter auto-configures with the default base URL (http://127.0.0.1:4096). No additional setup is needed.

3. In DeepSeek Harness, select the model:

opencode-local/mimo-v2.5-free

Configuration

All fields are optional. Defaults work out of the box with a local OpenCode server.

# Inside your DSH cordis.yml or plugin config
llm-opencode:
  # Endpoint of the OpenCode server
  baseURL: "http://127.0.0.1:4096"

  # Context window reported to the harness (default: 128000)
  defaultContextWindow: 128000

  # Max output tokens per request (default: 16384)
  maxTokens: 16384

  # Advisory model list for discovery consumers
  models:
    - id: mimo-v2.5-free
      name: mimo-v2.5-free
      description: "OpenCode local agent (text output only)"
      contextWindow: 128000

  # Retry policy
  retryPolicy:
    mode: always
    backoff:
      initialDelayMs: 100
      maxDelayMs: 5000
      jitterRatio: 0.2

How It Works

1. Session creation — POST /api/session with { model: { id, providerID: "opencode" } }
2. Concurrent SSE + prompt — the adapter fires the SSE subscription (GET /api/session/{id}/event) and prompt delivery (POST /api/session/{id}/prompt) concurrently. This avoids a deadlock where the OpenCode server withholds SSE headers until the session has activity.
3. Conversation serialization — the full DSH conversation is formatted as [System]\n...\n\n[User]\n...\n\n[Assistant]\n... and embedded in the prompt.
4. Event translation — SSE events (session.text.delta, session.next.step.ended, session.execution.succeeded, etc.) are mapped to DSH StreamChunk types (block-start, text-delta, block-end, finish).
5. Cleanup — on completion, error, or abort, the adapter sends a best-effort interrupt to the session.

Development

# Build
npm run build

# Type-check
npm run typecheck

# Run tests
npm run test

# Build + test
npm run check

# Clean output
npm run clean

Security Warning

This plugin communicates with a local HTTP server (no TLS by default). Do not expose the OpenCode server to untrusted networks. The default binding (127.0.0.1) restricts access to the local machine.

Platform Compatibility

v0.1.0 has been tested on the following exact environment:

Other platforms, Node versions, harness releases, OpenCode versions, and models have not been verified. The adapter may work elsewhere, but no guarantees are made.

Limitations

• Text output only — OpenCode runs its own agent/tool loop. Structured tool calling exposed by DSH is not available through this adapter.
• No multimodal support — the adapter reports inputModalities: ["text"] only. Image/audio/video inputs are not forwarded.
• Single model advertised — the default catalog lists mimo-v2.5-free. Adding models requires manual config; the adapter does not dynamically discover available models from the server.
• Prompt size — the entire conversation is serialized into a single prompt string. Very long conversations may exceed the model's context window.
• No persistent sessions — each stream() call creates a fresh OpenCode session. Conversation state is not preserved across requests.
• Local-only by default — the adapter expects the OpenCode server on 127.0.0.1. Remote servers require manual baseURL configuration and appropriate network access.

License

MIT