Configuring opencode2dsh
DeepSeek Harness reads the plugin configuration from the profile’s cordis.patch.yml. The default mode is the native adapter.
| Key |
Default |
Function |
mode |
adapter |
Selects native streaming or the legacy sidecar mode. |
providerId |
opencode2dsh |
Provider name shown in DeepSeek Harness. |
refreshSeconds |
300 |
Live catalogue refresh interval. |
allowedModels |
[] |
Adapter-only exact model whitelist. An empty list exposes eligible free-lane models. |
firstEventMs |
30000 |
Adapter first-event watchdog, in milliseconds. |
bodyIdleMs |
120000 |
Mid-response body-idle watchdog. |
responsesBodyIdleMs |
300000 |
Body-idle floor for muse-spark-* Responses models. |
The watchdog values accept integers from 1 to 600000; changing them requires restarting DeepSeek Harness. Pricing metadata refreshes every 24 hours. sidecar mode is not bundled: it requires an agent built from legacy/agent, plus agentPath and optional agentArgs. Its restart controls are restartDelayMs, restartMaxDelayMs, and maxConsecutiveCrashes.
How opencode2dsh behaves
DeepSeek Harness receives streamed blocks for text, usage, and completion from the registered LlmAdapter. Chat models use pi-ai OpenAI completions; muse-spark-* uses OpenAI Responses. Session and project correlation values are derived per conversation. The catalogue uses the live upstream list intersected with free-by-metadata eligibility, then falls back to an offline cache and verified static list. Fast startup retries and periodic refresh support recovery, while classified upstream failures remain visible to DeepSeek Harness and retry ownership stays there.
Commands and verification
After adding the package, restart dsh web, open the model picker, and select a model in the opencode2dsh group. DeepSeek Harness requires a web profile for this integration.
Requirements and limits
The plugin requires DeepSeek Harness 0.1.7 or newer and Node.js 20 or newer. The runtime needs outbound HTTPS access to opencode.ai and models.dev. Image attachments are accepted only for models whose models.dev metadata declares image input, with a verified vision-family fallback when metadata is silent. The optional delegation.enabled path is disabled by default and, when enabled, seeds only the opencode credential into an isolated XDG root and runs a managed local opencode serve after the anonymous lane and sibling retry hit the credential gate.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub