dispatch

Forward the user request to the cursor:cursor-dispatch subagent. The subagent invokes cursor-companion.mjs dispatch once via the --raw-args-stdin channel (codex F2 — $ARGUMENTS may contain shell metacharacters; never interpolate it into a bash pipeline) and returns stdout verbatim.

Sandbox model (v0.3+): by default, cursor edits the caller's cwd directly (--in-place). Commits land on the current branch — no cherry-pick needed. Pass --isolated to opt into the legacy .cursor/worktrees/<jobId> sandbox with auto-commit to a <jobId> branch. --background defaults to --isolated (a background agent racing your edits is dangerous); pass --in-place explicitly to override.

Raw user request: $ARGUMENTS

Routing rules:

• If the request contains --wait, run the subagent in foreground.
• If the request contains --background, run the subagent in background (Agent({..., run_in_background: true})).
• Otherwise, decide based on prompt size:
  • For tiny / clearly bounded tasks, recommend --wait.
  • For multi-step / open-ended / "this could take a while" tasks, recommend --background.
  • When in doubt, recommend --background.
  • Use AskUserQuestion exactly once with two options:
    • Wait for results
    • Run in background
  • Put the recommended option first and suffix its label with (Recommended).

Mode routing (cursor-agent supports plan, ask, agent):

• If the request already contains --mode <plan|ask|agent> or --plan-only, do not ask — honor it.
• Otherwise infer from intent and ask the user once via AskUserQuestion:
  • Pure question / explanation / read-only inspection ("what does X do", "why is Y", "解释一下", "看看") → recommend Ask only (read-only Q&A) first.
  • Wants a written plan or design before any edit ("draft a plan", "先出个方案", "怎么改") → recommend Plan only (no edits) first.
  • Wants the agent to actually do the work ("fix", "implement", "改一下", "动手") → recommend Agent (default, may edit) first.
  • When in doubt, recommend Agent (default, may edit) first.
• Three options total; never offer more.
• On Ask only, append --mode ask to the forwarded text.
• On Plan only, append --mode plan.
• On Agent, append nothing (default).

Model routing:

• If the request already contains --model <m>, leave it untouched.
• If the request contains --resume <jobId>, append nothing — the existing cursor thread already has a bound model.
• Otherwise silently append --model auto. auto is cursor's per-request router (visible in cursor-agent --list-models), so the slash command never enumerates specific model IDs and never goes stale as cursor's lineup changes. Users wanting a specific model can override with --model <id>; run cursor-agent --list-models to see currently available IDs.

Resume routing:

• If the request contains --resume or --fresh, do not ask.
• Otherwise, run:

  node "${CLAUDE_PLUGIN_ROOT}/scripts/cursor-companion.mjs" dispatch-resume-candidate --json
  

  • If available: true, ask the user (AskUserQuestion) whether to continue the previous cursor thread or start a new one.
  • Choices:
    • Continue current cursor thread
    • Start a new dispatch
  • If the user's request reads like a follow-up ("再", "继续", "调整一下", "基础上", "another pass"), put Continue current cursor thread (Recommended) first.
  • Otherwise put Start a new dispatch (Recommended) first.
  • On Continue, append --resume <returned-jobId> to the forwarded text.
  • On Start a new dispatch, append --fresh.

Forwarding (codex F2 — the full forwarded text goes on stdin via a heredoc; sh never re-evaluates it):

Pass the prompt + flags to the subagent through stdin. The subagent runs:

node "${CLAUDE_PLUGIN_ROOT}/scripts/cursor-companion.mjs" dispatch --raw-args-stdin <<'__CURSOR_RAW_ARGS_END_a8f3c91d4b__'
<the full forwarded request, including any --resume/--fresh/--wait/--background/--isolated/--in-place/--mode you appended>
__CURSOR_RAW_ARGS_END_a8f3c91d4b__