dsh-raincode

Make raincode the model layer of DeepSeek Harness (dsh): a cheap model pool with StablePrefix prompt caching, retries, and failover, driven by a local OpenAI-compatible gateway. Inside dsh you can pick any model from your raincode pool and browse the raincode skill library with /skills.

The plugin is the "mouth" — it translates dsh's GenerateOptions to the OpenAI wire format and streams chunks back — while the raincode gateway (Rust) is the "brain" that owns profile selection, caching, and retries. The plugin holds no model logic.

Prerequisites

1. Install raincode (>= 0.1.0 with gateway /v1/models and /skills):

   cd raincode
   cargo install --path crates/rc-cli
   # or use ./install.sh
   

2. Configure profiles (your model pool):

   raincode setup
   # writes ~/.raincode/profiles.toml
   

3. Start the gateway (keep it running):

   raincode proxy --port 8787
   

   Verify: curl http://127.0.0.1:8787/health → {"ok":true,"service":"raincode-gateway"}.

Install in dsh

dsh plugin --profile web add dsh-plugins-raincode

Then add the row to the profile's patch layer (~/.dsh/profiles/web/cordis.patch.yml, or ~/.dsh/cordis.patch.yml for every profile):

- insert:
    - id: raincode
      name: dsh-plugins-raincode
      config:
        baseURL: http://127.0.0.1:8787
        apiKeyEnv: RAINCODE_API_KEY

dsh --profile web --dump-config should show the raincode row; restart dsh web to load it.

Configuration

Field	Default	Meaning
baseURL	http://127.0.0.1:8787	raincode gateway address
apiKeyEnv	RAINCODE_API_KEY	dsh credential seam only; a loopback gateway does not verify keys (missing credentials don't block it)
models	[]	optional static catalog override; empty fetches the pool from the gateway's /v1/models
streamIdleTimeoutMs	300000	stream idle timeout (dsh-timeout helper)
retryPolicy	dsh default	retry policy via RetryPolicySchema

What you get

• Model selection — dsh's model picker shows the raincode pool (fetched from the gateway's /v1/models); the chosen model maps exactly to a gateway profile.
• /skills browsing — run /skills to list the raincode skill library, /skills <name> to read a skill's body. Data comes from the gateway's /skills endpoint (~/.raincode/skills/).
• Caching & retries — prompt caching (sessionId → prompt_cache_key) and retry/failover are handled by the gateway per profile; nothing to configure in dsh.
• Parallel tool calls — tool calls stay separate blocks keyed by call index; fragmented arguments accumulate per call.

FAQ

• Gateway not running? You'll see TRANSPORT-style errors from the model call. Start it with raincode proxy --port 8787.
• Why is the API key a formality? A loopback gateway trusts the machine and resolves real upstream keys from ~/.raincode/profiles.toml (inline, key file, or env). The plugin only demands a credential when baseURL points at a remote gateway.
• How do I add a model? raincode setup (or edit ~/.raincode/profiles.toml and restart the gateway).

License

MIT

中文说明见 README.zh.md.