dsh-junie-subagent

@zumikua/dsh-junie-subagent

English | 中文

Delegates one task to the JetBrains Junie CLI over the Agent Client Protocol and returns Junie's final answer as a subagent result.

Requirements

• DSH >= 0.2.1-rc.2;

• the JetBrains Junie CLI installed and signed in. The provider looks for it in this order:

  1. the row command config;
  2. JUNIE_PATH;
  3. the managed install's ~/.local/share/junie/current pointer — <version>/junie/junie.exe on Windows, <version>/bin/junie elsewhere;
  4. junie.exe / junie on PATH, skipping a .bat shim.

  When none of them resolves, the run fails with every path it tried, and the settings page says so instead of showing an empty model list.

Install

Open the Plugins page in the DSH Web UI and install @zumikua/dsh-junie-subagent from the catalog; the plugin manager installs the package and selects its bundle in one step. The bundle patch inserts one provider and one dsh-tool-subagent instance named junie.

Restart DSH once after installing: the browser half (client.js) enters the page's module table at startup. Then open Settings → Junie and pick a model and thinking depth.

How one run works

spawn <junie.exe> --acp=true --skip-update-check
  → initialize
  → session/new { cwd, mcpServers: [] }
  → session/set_config_option (model / effort / brave_mode, when configured)
  → session/prompt { prompt: [{ type: "text", text }] }
  ← session/update … agent_message_chunk text …
  ← result { stopReason }
  → dispose (session/cancel, stdin EOF, terminate, wait for exit)

Each run owns its own process and session: the subagent seam has no continuation surface for out-of-process children, so a run accepts no follow-up message.

Confinement

The child is spawned through ctx.subprocess without a sandbox policy. It therefore runs as the signed-in user with that user's full file access — which is what lets Junie persist sessions under ~/.junie. Confinement is Junie's own permission model: brave_mode, the action allowlist, and session/request_permission, which this provider answers from permission.

Configuration

Key	Default	Meaning
providerName	junie	Name on ctx.subagents
command	auto-detected	Explicit launcher path; a missing file is reported instead of falling back
args	["--acp=true"]	Extra arguments; --acp=true and --skip-update-check are always ensured
model	mix.normal	Fallback model when the settings page has stored none
effort	high	Fallback effort when the settings page has stored none
braveMode	brave-auto	brave-auto / on / off
permission	reject	Auto-answer session/request_permission with reject (default) or the first allow option
timeoutMs	900000	Whole-run bound
graceMs	3000	Subprocess teardown grace

Choosing a model

The bundle registers exactly one provider (junie) and one model-facing tool (junie). The model and the thinking depth are chosen on the settings page and stored by the Host; the model / effort keys above are only the fallback used before anything is stored.

Per-call model selection through the shared provider / model / reasoning_effort tool fields is not available: dsh-tool-subagent validates those against the Host's LLM catalog (llm.resolveCallConfig), and Junie model ids are not LLM routes — which is also why the official Codex, Claude Code, and ACP providers advertise no agentOptions capability.

Settings page and storage

1. The page is the Junie section in the settings panel; it picks the model, the thinking depth, and Junie's brave mode (auto / on / off).
2. Values are read and written through a same-origin route the Host registers — GET/PUT /junie-subagent/config — and stored in ~/.dsh/junie-settings.json.
3. The model list comes from the configOptions of Junie's own ACP session/new (rendered by group, cached for ten minutes, costs no credits).
4. Thinking depths are model-specific. Saving validates the pair against a throwaway Junie session; an unsupported combination answers 400 with the values that model accepts, and the option list in the page is left untouched.
5. A change applies from the next delegation. The provider never prompts: it rejects every permission request, so brave-auto allows safe work and rejects the rest, off rejects everything that would need approval, and on runs everything.