dsh-plugin-opencode

Unofficial bridge between the OpenCode CLI and DeepSeek Harness: an opencode delegation tool plus a /opencode command running opencode run (headless), and an opencode-agent LLM provider route with automatic model discovery via opencode models — all over the ctx.subprocess seam. Not affiliated with the OpenCode project.

What the agent gets

• opencode tool — spawns opencode run --format json "<prompt>" as a managed child process. Per-call cwd, model (provider/model), agent, session, continueLast, auto, and timeoutMs. Returns { ok, exitCode, stdout, stderr, timedOut, truncated, sessionId, inputTokens, outputTokens }. The ndjson event stream is parsed into the answer text; sessionId can be passed back as session to continue a run.
• /opencode <task> — slash command shortcut (when the profile composes the commands service).

Cancelling the tool call aborts the child via the subprocess seam's terminate escalation (SIGTERM → grace → SIGKILL on the process tree); output capture is bounded and keeps the tail on overflow.

The provider route

@quonaro/dsh-plugin-opencode/provider registers opencode-agent with ctx.llm.registerAdapter, so OpenCode appears next to ordinary model providers (Settings → Models, /model, agent presets). Each generation spawns opencode run --format json with the flattened transcript and streams its text back; token usage is real, summed from step_finish events.

The route is opencode-agent, not opencode: the pi-ai catalog already owns opencode (OpenCode's own HTTP API), and the configurable-provider directory rejects duplicates.

The adapter lists opencode models for the agent's real model catalog and caches the result; disable via autoDiscoverModels to use the static models table.

OpenCode is an agent, not a model — each call is a full autonomous session. Use it to hand whole tasks to OpenCode; latency is seconds to minutes.

Prerequisites

• opencode CLI on PATH (or set opencodePath to an absolute path)
• opencode auth login completed (credentials live in ~/.local/share/opencode, reached via HOME/XDG_* in forwardEnv)

Install

# from a local checkout
dsh plugin --profile <your-profile> add /path/to/dsh-plugin-opencode

The package declares dsh.bundle, so dsh plugin add applies cordis.patch.yml to the profile automatically. On pnpm ≥ 10 the first git install refuses the prepare build — copy the package key pnpm prints into the profile's pnpm-workspace.yaml allowBuilds: and re-run.

Configuration

Edit the dsh-plugin-opencode row in the profile's cordis.patch.yml, or the derived settings namespace in the GUI. All fields are volatile — changes apply to the next call without a restart.

Key	Default	Notes
opencodePath	opencode	Executable name or absolute path
auto	true	--auto auto-approves permissions not explicitly denied — headless runs can't answer prompts; marked dangerous upstream, tighten via permission in opencode.jsonc
model	''	-m provider/model; empty = opencode config default
agent	''	--agent sub-agent (e.g. build, plan)
attach	''	--attach <url> drives a running opencode serve instead of a local instance
pure	false	--pure disables external opencode plugins
timeoutMs	600000	Cooperative timeout per delegation call
maxOutputBytes	1048576	In-memory cap per captured stream; overflow keeps the tail
forwardEnv	PATH/HOME/USER/XDG**/OPENCODE**	Env vars forwarded to the child
extraArgs	[]	Extra CLI flags appended before the prompt

Provider row (dsh-plugin-opencode-provider) adds cwd, stderrMaxBytes, models (static route→-m table), brief, localSessionTitles, autoDiscoverModels, discoveryTimeoutMs, discoveryCacheMs.

Notes

• --session/continueLast/fork/agent are per-call tool parameters; session/continueLast/fork live on the tool, not the provider route (a model route has no session handle).
• The tool result's sessionId (ses_…) enables multi-turn delegation across calls.