agent

Invoke the cursor:cursor-agent subagent via the Agent tool (subagent_type: "cursor:cursor-agent"), forwarding the natural-language prompt as described below (raw task text with routing flags stripped, or the combined context-plus-task string when --context was used). The final user-visible response must be Cursor's output verbatim.

Raw user request: $ARGUMENTS

Context flag:

• If the request includes --context, handle it only in this command. Do not forward --context to the subagent or to task.
• Before anything else below, if --context was present: build a short (~5 lines, one per field) context block from this Claude session’s recent conversation and prepend it to the user’s task text. Use that combined string as the natural-language prompt passed to the subagent. Remove --context from both routing flags and task text.
• Fixed template:

## Context (from Claude session)
- Goal: <one line>
- Recent decisions: <one line>
- Files touched: <comma-separated paths, or "n/a">
- Current state: <one line>
- Open question / blocker: <one line, or "none">

## Task
<original user task text>

• Block rules: facts already stated in the conversation only; no invention. No literal transcript, no code snippets, no secrets, keys, or tokens. If a field does not apply, use n/a or none. If nothing concrete exists to summarize (e.g. first message in session), omit the context block and tell the user briefly before continuing.
• If the user combines --context with --continue or --resume, still prepend the block as above.
• --context is exclusive to this slash command and is never a cursor-agent or companion flag.

Execution mode:

• If the request includes --wait, run the cursor:cursor-agent subagent in the foreground.
• Otherwise (no flag or explicit --background), run the cursor:cursor-agent subagent in the background. Background is the default.
• --background and --wait are execution flags for Claude Code. Do not forward them to task, and do not treat them as part of the natural-language task text.
• --model, --mode, and --read-only are runtime flags. Preserve them for the forwarded task call, but do not treat them as part of the natural-language task text.
• --context is handled only here (see Context flag); never forward it to task.
• If the request includes --continue, --resume <id>, or --fresh, do not ask whether to continue. The user already chose.
• Otherwise, before starting Cursor, check for a resumable thread from this Claude session by running:

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

• If that helper reports available: true, use AskUserQuestion exactly once to ask whether to continue the current Cursor thread or start a new one.
• The two choices must be:
  • Continue current Cursor thread
  • Start a new Cursor thread
• If the user is clearly giving a follow-up instruction such as "continue", "keep going", "resume", "apply the top fix", or "dig deeper", put Continue current Cursor thread (Recommended) first.
• Otherwise put Start a new Cursor thread (Recommended) first.
• If the user chooses continue, add --continue before routing to the subagent.
• If the user chooses a new thread, add --fresh before routing to the subagent.
• If the helper reports available: false, do not ask. Route normally.

Operating rules:

• The subagent is a thin forwarder only. It should use one Bash call to invoke node "${CLAUDE_PLUGIN_ROOT}/scripts/cursor-companion.mjs" task ... and return that command's stdout as-is.
• Return the Cursor companion stdout verbatim to the user.
• Do not paraphrase, summarize, rewrite, or add commentary before or after Cursor’s stdout (except building the optional session context block when --context was requested).
• Default to write-capable runs. --read-only suppresses --force.
• If the helper reports that Cursor is missing or unauthenticated, stop and tell the user to run /cursor:setup.
• If the user did not supply a request, ask what Cursor should investigate or fix.