Cursor CLI

Cursor CLI

Announcement: Begin with: "I'm using the cursor-cli skill for cross-AI validation with Cursor."

Use this wrapper for a one-shot Cursor review or approach validation. Read the shared lifecycle in ../codex-cli/references/cross-ai-run.md first.

Provider adapter

1. Check command -v agent, agent --version, and agent --help. Verify the binary is Cursor Agent and supports --print, --mode ask, and --output-format json. cursor-agent may be an installation alias; do not substitute the Cursor editor executable.

2. When the caller requests a particular model or an independent model family, inspect agent models and select an exact listed slug. Otherwise omit the model override. Cursor supports multiple providers; auto and an unreported default do not establish model-level independence. Record the requested slug separately from any model identity actually reported by the provider.

3. Resolve the loaded skill directory and run the adapter from the target repository root with Python 3.10+:

   CURSOR_SKILL_DIR="/absolute/path/to/loaded/cursor-cli"
   python3 "$CURSOR_SKILL_DIR/scripts/review.py" \
     --prompt-file /tmp/cursor-prompt.md \
     --file path/to/file.py \
     --timeout 120 \
     --output-dir /tmp/cursor-review-unique
   

   Set CURSOR_SKILL_DIR to the directory containing this loaded skill, not a path relative to the reviewed repository. --file is repeatable and pre-reads verified UTF-8 files into the prompt. The combined input limit is 96 KiB; narrow larger reviews explicitly. Add --model '<exact listed slug>' only when selected. Each invocation requires a fresh output directory.

4. Inspect receipt.json, stdout.json, stderr.log, version.stderr.log, and response.txt on success. A completed JSON result requires exit 0, type: result, subtype: success, is_error: false, and a non-empty string result. An error envelope, timeout, malformed output, trust rejection, or empty answer is incomplete. Also read the answer for blocked reads, omitted scope, or other visible limitations: a valid envelope alone does not prove the review was completed. Return the shared receipt and evidence gaps.

The adapter always uses agent -p --mode ask --output-format json. Print mode alone can use write and shell tools; keep the explicit read-only mode. Workspace trust is a separate prerequisite: the adapter preserves a trust rejection and stops. It does not pass --trust, force/yolo, automatic MCP approval, or sandbox bypass flags, retry, authenticate, update the CLI, or modify settings. Resolve missing capabilities or trust through a separate user-authorized action.

Prompt requirements

Include the repository root, task/requirements path when task-attached, changed files or diff scope, and concrete questions. Cursor starts without this conversation's context. Resolve task evidence through ../setup/references/task-context.md; standalone calls do not create a task.

When not to use

• The caller did not explicitly choose Cursor
• The request needs interactive refinement
• The primary review workflow already answers the request

References

• ../codex-cli/references/cross-ai-run.md — shared lifecycle
• reference.md — native output contract, trust boundary, and official sources
• templates.md — prompt shapes; resolve placeholders before execution