dsh-sidecar

Fire focused subagents in the background with unique names and separate spaces, then review each result — task, answer and evidence trail, plus the full transcript on demand — on your own schedule. Separate plugin: no shared state with dsh-autocontinue (continuation/restarts) or the dsh-update updater.

Commands and tools

/sidecar <task>                              # auto name (sidecar-1, -2, ...), parent cwd
/sidecar --name alpha <task>                 # unique name, pinned as the child session title
/sidecar --cwd /tmp/alpha-space <task>       # separate working directory
/sidecar --worktree --name wt-fix <task>     # isolated git worktree + branch sidecar/wt-fix
/sidecar list                                # running sidecars: name, id, age, cwd
/sidecar result <id>                         # read and consume a stored result now
/sidecar log <id>                            # full child transcript on demand
/sidecar cancel <id>                         # interrupt one

Multiple sidecars run concurrently up to the subagents service cap (maxActiveSubagents, default 8). Each gets a unique name (requested names are suffixed -2, -3, ... if taken), a distinct child session id, and its title pinned to the name so the GUI list is unambiguous. They share the parent cwd by default; use --cwd for a separate workspace or --worktree for a real filesystem/branch isolation. Worktrees start from HEAD — uncommitted parent changes are intentionally not visible there.

Model-facing tools (used when the parent chooses to review):

• sidecar_result — { id } returns the stored review body and removes it from the queue.
• sidecar_list — lists stored results waiting for review.

Delivery: notice, not a queued user message

The child is spawned through ctx.subagents.startActivation() with delivery: 'caller'; its prompt forbids send_message and further delegation.

When it finishes, the plugin:

1. stores the full result at ~/.dsh/sidecar/results/<parent>/<child>.json;
2. calls agent.inject(notice) — a next-step notice with no wake. This is not followup: it does not create a queued user message and does not start a turn. A running parent sees it at its next step; an idle parent sees it when it next runs. The notice says only that the result is stored and that sidecar_result will return it;
3. never injects the full result into the transcript. The parent reviews by calling sidecar_result (or the user runs /sidecar result <id>), which returns the result as a tool result and consumes the file.

The stored review is deliberately compact:

• the complete task the sidecar was given (so the parent knows what the result is for),
• the final answer,
• a compact evidence trail: every tool call and its result, in order (capped at 8 KB), and
• the child session id with a pointer to sidecar_log.

The child's complete session log — including workspace instructions, reasoning and every step — stays on disk. sidecar_log (or /sidecar log <id>) returns it on demand. Pushing the full raw log into the parent by default would bury the useful task/answer under pages of workspace boilerplate, so it is opt-in.

If the parent closes before ever seeing the notice, the file persists; the next time that parent agent is created (session reopened or resumed) the notice is injected again. sidecar_list is the discovery path if the notice was already consumed or missed.

Install

dsh plugin --profile web add github:kenjoe41/dsh-sidecar
# or pin a release
dsh plugin --profile web add github:kenjoe41/dsh-sidecar#v0.4.0

For a local checkout, link it instead; a linked install needs node_modules/@deepseek-ai/{dsh-llm,dsh-tools} shims so the profile's peers resolve. Reload or restart the profile after installing.