AgentSkillsCN

fork

当对话偏离主题时，可将其分支出来，另行处理。工具会自动生成上下文摘要，并提供一条即用的命令，方便用户在新的终端会话中直接复制粘贴，继续推进工作。

SKILL.md
--- frontmatter
name: fork
description: Branch off a conversation to handle tangents. Outputs context summary and ready-to-paste command for a new terminal session.

Fork Conversation

Generate a context handoff for a new Claude session. No tool calls needed - synthesize from memory and output immediately.

Output Format

code
## Fork Ready

**Context Summary:**
- Working on: [one sentence - current task]
- Files: [comma-separated paths, max 5-7, mention "+ N others" if more]
- State: [key decisions, branch, blockers, test status - bullets if multiple]

**Paste in new terminal:**
```bash
cd [target directory] && \
claude "/mission-control [escaped prompt here]"

This conversation continues here. The fork runs independently.

code

## Building the Command

Always invoke `/mission-control` so the forked session starts in coordination mode:

/mission-control Context from forked conversation:

  • •Working on: [description]
  • •Files: [list]
  • •Key context: [important decisions, findings, state]

Continue with: [focus hint from args, or "Pick up where we left off"]

code

## Args Handling

- `/fork` - generic continuation, opens in mission-control
- `/fork debug the auth issue` - args become the "Continue with:" line

## Permission Mode Propagation

If the current session has bypass permissions enabled, propagate it to the fork:

```bash
cd [dir] && \
claude --dangerously-skip-permissions "/mission-control ..."

Check your current permission context. If you're running with elevated permissions (no confirmation prompts for file edits, bash commands, etc.), include the flag so the forked session has the same capabilities.

Shell Escaping

  • •Wrap entire prompt in double quotes
  • •Escape internal double quotes as \"
  • •Escape dollar signs as \$
  • •Escape backticks as \`

What to Include

  • •Target directory for the forked task (may be current cwd, or different if the tangent involves another repo/location)
  • •Current task (one sentence)
  • •Key decisions/findings (not speculation)
  • •Relevant file paths
  • •Important state (branch, test status, blockers)

What NOT to Include

  • •Full conversation history
  • •Implementation details already in files
  • •Secrets or credentials
  • •Plans not yet acted on
  • •Tool call history

Edge Cases

  • •Conversation just started: Minimal fork, note limited context
  • •Many files: Top 5-7 most relevant, note "and N others"
  • •Sensitive content: Omit credentials, note they exist separately

Anti-patterns

  • •Multi-paragraph summaries
  • •Waiting for confirmation
  • •Suggesting alternatives
  • •Including noise from tool outputs