Cursor Council
ReviewAudited by ClawScan on May 10, 2026.
Overview
Cursor Council is purpose-aligned, but its default workflow launches multiple logged-in coding agents in detached tmux sessions with force/approval-bypass patterns, so it needs careful review before use.
Install only if you are comfortable supervising several coding agents at once. Use separate branches or worktrees, remove `--force` unless truly needed, review each sub-agent's changes before merging, avoid putting secrets in prompts, and shut down tmux sessions when the work is complete.
Findings (5)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Multiple agents could edit project files or continue high-impact coding actions before the user has reviewed each change.
The default orchestration examples run agent commands with `--force` and explicitly instruct accepting approval prompts, reducing per-action human review across multiple coding agents.
tmux send-keys -t cursor-frontend "agent -p '...' --force" Enter ... 看到 "waiting for approval" 就 `tmux send-keys -t $session y`
Avoid `--force` by default, require explicit user confirmation for each sub-agent task, use separate branches/worktrees, and review diffs before merging.
A poor task split could cause conflicting edits, broken builds, or hard-to-review changes across the same repository.
The skill acknowledges that parallel agents can conflict and relies on the user or supervising agent to split tasks safely; the boundary is recommended but not enforced.
每个 Cursor 负责不同的文件/模块，**绝对不能有文件冲突** ... 最好每个 Cursor 在不同分支上工作
Use isolated branches or worktrees, give each agent explicit file/module limits, and run tests and manual review before combining results.
Agents may keep running, consuming resources or account quota, after the user has stopped watching the terminal.
Detached tmux sessions are central to the skill's design, but they create background agent sessions that persist until stopped.
tmux new-session -d -s cursor-$i ... tmux send-keys -t cursor-$i "cd $PROJECT_DIR" Enter
Monitor active tmux sessions, set clear stop conditions, and kill sessions when tasks finish or appear stuck.
Coding actions and model calls may use the user's authenticated account, project access, and paid model limits.
The skill depends on an already-authenticated local agent/Cursor account, which is expected for the purpose but means actions run with that account's privileges and quotas.
需要 `cursor-agent` skill 已配置好、`agent login` 已完成。
Use an appropriate account, verify model costs and project permissions, and avoid running the skill in repositories where the agent should not have write access.
Confidential architecture details, code snippets, or logs included in prompts may be shared with external model providers and retained in local temp files.
Council mode sends the user's question/context to multiple model backends and writes outputs to local temporary files.
agent --model claude-opus-4-6 ... agent --model claude-sonnet-4-5 ... agent --model gpt-5.2 ... tee /tmp/council-opus-output.txt
Do not include secrets or sensitive customer data in council prompts, confirm provider data policies, and clean up temporary council files when done.