Skill: duplicate-thread

Purpose

Clone an existing plan, task thread, or agent session so you can explore an alternative approach without losing the original. Analogous to "duplicate thread" in Copilot's mission control.

When to Use

• You want to try a risky refactor without throwing away the current plan
• A/B testing two implementation strategies in parallel
• Resuming a stalled task from a known-good checkpoint with different parameters
• Forking a plan for a different target environment (e.g., prod vs. staging)

How It Works

1. Identify the source thread — a .claude/threads/<id>/ directory, a plan file, or a PLAN.md
2. Copy it to a new thread ID with a meaningful suffix
3. Stamp it with metadata (source thread, fork reason, timestamp)
4. Continue in the new thread independently

Thread Directory Convention

.claude/threads/
  <thread-id>/
    PLAN.md          # the plan or task description
    CONTEXT.md       # accumulated context
    CHECKPOINT.md    # last known-good state (from checkpoint-strategy skill)
    meta.json        # thread metadata

Usage

Manual duplication

bash .agents/skills/duplicate-thread/duplicate.sh <source-thread-id> <reason>

Example:

bash .agents/skills/duplicate-thread/duplicate.sh plan-2025-01-15 "try-async-approach"
# Creates: .claude/threads/plan-2025-01-15--fork-try-async-approach/

In a Claude prompt

Duplicate the current thread as "alternative-auth-strategy" and in the fork,
replace the JWT approach with session cookies. Keep the original intact.

Claude will:

1. Copy the current PLAN.md / CONTEXT.md to a new thread folder
2. Apply the requested change only in the fork
3. Proceed with the forked version, noting the original thread ID in meta.json

meta.json Schema

{
  "thread_id": "plan-2025-01-15--fork-try-async-approach",
  "forked_from": "plan-2025-01-15",
  "fork_reason": "try-async-approach",
  "forked_at": "2025-01-15T10:30:00Z",
  "status": "active"
}

Merging Back

If the fork succeeds and you want to merge learnings back:

1. Compare PLAN.md diffs between fork and original
2. Update the original CONTEXT.md with lessons learned
3. Archive the fork: set "status": "merged" in meta.json

If the fork fails:

1. Set "status": "abandoned" in meta.json
2. Return to the original thread — it is untouched

Integration

Pairs well with:

• checkpoint-strategy — always checkpoint before duplicating
• parallel-agents — run both threads simultaneously and compare
• council-review — have the council evaluate both approaches and pick the winner
• cooldown-resume — resume a forked thread after a pause

Files

• SKILL.md — this file
• duplicate.sh — shell helper to copy a thread