loop（loop 工具）

Loop

A self-sustaining development loop with two modes: Researcher and Implementor.

Prerequisites

Full permissions required. The loop runs autonomously and needs unrestricted access.

Claude Code

claude --dangerously-skip-permissions

Codex

The script automatically uses --yolo mode (no sandbox, no prompts).

Startup Flow

When this skill is invoked, ask the user FOUR questions using AskUserQuestion, then one text prompt:

Question 1: Mode (AskUserQuestion)

• Question: "Which mode would you like to run?"
• Options:

1. Research - Explore, investigate, and plan. Creates specs for implementation. 2. Implement - Execute on specs. Write code, tests, and documentation.

Question 2: Git Workflow (AskUserQuestion)

• Question: "What should happen after each task?"
• Options:

- Push - Commit and push to current branch (default) - Commit only - Commit locally, no push - Open PR - Open PR, wait for CI (no merge) - PR and merge - Open PR, wait for CI, then auto-merge

Question 3: Context Management (AskUserQuestion)

• Question: "Should each iteration start with fresh context?"
• Options:

- Fresh context - Clear context between iterations (default, recommended) - Keep context - Maintain conversation history across iterations

Question 4: Directions (AskUserQuestion - optional)

• Question: "Any specific directions for this run?"
• Options (mode-aware, user can select "Other" for custom):

- None - No specific directions, work autonomously - For Research mode: - Focus on specs - Review and improve existing implementation specs - Explore dependencies - Research external libraries and frameworks - For Implement mode: - Fix issues first - Prioritize fixing build errors and warnings - Skip tests - Focus on implementation, skip test writing for now

If user selects "Other", they can provide custom directions like:

• "Focus on authentication patterns"
• "Search Apple developer docs for NSPanel"
• "Prioritize the archive feature"
• "Redo the task list with proper SwiftUI patterns"

The agent interprets directions intelligently (creating dots, modifying tasks, updating docs, changing priorities, etc.).

Question 5: Iterations (Text prompt - NOT AskUserQuestion)

After the AskUserQuestion completes, ask:

"How many iterations? Enter a number, 'inf' for infinite, or 'comp' for until complete:"

Parse the response:

• Number (e.g., "2", "5", "10") → exact iteration count
• "inf", "infinite", "-1" → infinite mode (-1)
• "comp", "complete", "0" → until complete mode (0)

Key distinction:

• Infinite / N iterations: Completion promise is IGNORED. Loop continues for re-analysis and improvement.
• Until complete: Only mode where completion promise stops the loop.

After collecting answers:

1. Initialize loop state by running: ./scripts/setup-loop.sh <mode> --iterations <N> --git-workflow <workflow> [--fresh-context] [--directions "..."]

- Iterations: -1=infinite, 0=until complete, N=exact count - Workflow: commit/push/pr/pr-merge - Directions: optional user guidance for the agent

1. Build the prompt by combining mode-specific + shared content:

- Read <mode>-loop.md (mode-specific sections) - Read loop-shared.md (common sections) - Concatenate: mode-specific + shared

1. Begin execution based on context mode:

Fresh context mode (default): You are the orchestrator. Run this loop in the main conversation:

iteration = 0
PROMPT = contents of <mode>-loop.md + "\n\n" + contents of loop-shared.md
DIRECTIONS = user's directions (if provided)

# Exponential backoff for infinite mode
MIN_DELAY = 5        # seconds
current_delay = MIN_DELAY

# If directions were provided, append them to the prompt
if DIRECTIONS:
    PROMPT = PROMPT + "\n\n## User Directions\n\n" + DIRECTIONS

LOOP:
    iteration += 1
    print "=== Iteration {iteration} ==="

    result = Task(prompt=PROMPT, subagent_type="general-purpose")

    # Check termination conditions
    if iterations == 0:  # "until complete" mode
        if result contains "RANDROID_LOOP_COMPLETE":
            print "Loop complete (promise found after {iteration} iterations)"
            EXIT LOOP
    elif iterations > 0:  # exact N iterations mode
        if iteration >= iterations:
            print "Completed {iteration} iterations"
            EXIT LOOP
    # iterations == -1 means infinite, continues below

    # Exponential backoff for infinite mode when no work done
    if iterations == -1:  # infinite mode
        if result contains "RANDROID_LOOP_COMPLETE":
            print "No work this iteration, backing off for {current_delay}s..."
            sleep(current_delay)
            current_delay = current_delay * 2  # No max cap
        else:
            # Meaningful work done, reset backoff
            current_delay = MIN_DELAY

    GOTO LOOP

CRITICAL: After each Task returns, YOU (the orchestrator) must:

1. Check the result for the completion promise
2. Check if iteration limit reached
3. If neither → spawn another Task for the next iteration
4. Do NOT stop just because one Task finished

Keep context mode: Run the loop directly in the current conversation. The stop hook will intercept exit and continue looping with accumulated context.

Usage

Interactive (Recommended)

/loop

Prompts for mode, iterations, and optional directions.

Aliases: /randroid, /randroid-loop

Direct (Skip Questions)

• /loop research - Research mode, prompts for iterations
• /loop implement - Implement mode, prompts for iterations
• /loop research --loop - Research mode, infinite (ignores completion)
• /loop research --until-complete - Research mode, stops on completion
• /loop implement --iterations 5 - Implement mode, exactly 5 iterations
• /loop implement --open-pr - Open PR workflow
• /loop implement --pr-and-merge - PR with auto-merge workflow
• /loop implement --commit-only - Local commits only
• /loop implement --keep-context - Keep conversation context (no fresh start)
• /loop implement --iterations 5 --open-pr - Combine options

With Directions

You can provide guidance for the agent. When prompted for directions:

• Select a preset option (mode-specific), or
• Select "Other" and type custom directions like:

- "Focus on authentication patterns" - "Prioritize the archive feature, skip tests for now"

Directions can include:

• Topics to research or questions to answer
• Priorities for which tasks to work on
• Specific implementation guidance
• Requests to redo or improve previous work
• Scope changes (add/remove/modify tasks)

Codex (External Script)

# From terminal (outside Codex)
./scripts/randroid-loop.sh
# Interactive prompts for mode, iterations, and git workflow
# Note: Wrapper always uses fresh context (each iteration is a new codex exec)

# Direct invocation:
./scripts/randroid-loop.sh research -1           # infinite
./scripts/randroid-loop.sh research 0            # until complete
./scripts/randroid-loop.sh implement 5           # exactly 5 iterations
./scripts/randroid-loop.sh implement 5 pr        # 5 iterations, open PR
./scripts/randroid-loop.sh implement 0 pr-merge  # until complete, PR + merge

Modes

Research Mode

The researcher explores, investigates, and plans. It:

• Creates research: prefixed dots to track its own exploration
• Creates implement: prefixed dots as deliverables for the implementor
• Does NOT write production code
• Outputs findings, decisions, and clear implementation specs

Implementor Mode

The implementor executes. It:

• Pulls from ready implement: dots
• Writes code, tests, and documentation
• Creates new implement: dots if scope expands
• Creates research: dots if blocked by unknowns (for next research cycle)

Loop Mechanics

Fresh Context Each Iteration

Each loop iteration starts with fresh conversational context. Only the filesystem persists:

• Modified files
• Git history and commits
• Dots system state
• Research/implementation artifacts

This prevents context bloat and allows the agent to approach each iteration with clarity.

What Persists

• All file changes from previous iterations
• Git commits and history
• .dots/ task state
• Any written documentation or specs

What Resets

• Conversation history
• In-memory state
• Token usage (fresh budget each iteration)

Loop Termination

Output <promise>RANDROID_LOOP_COMPLETE</promise> when:

• No more ready tasks for your mode
• All work is committed and pushed

The loop also stops when --iterations N limit is reached.

Architecture

loop/
├── SKILL.md              # This file
├── research-loop.md      # Research mode prompt
├── implement-loop.md     # Implementor mode prompt
├── loop-shared.md        # Shared sections (git, dots, termination)
├── LOOPING_DESIGN.md     # Technical design doc
├── hooks/
│   └── stop-hook.sh      # Claude Code stop hook
├── scripts/
│   ├── setup-loop.sh     # Initialize loop state
│   └── randroid-loop.sh  # Codex external wrapper
└── state/
    └── .gitignore        # Excludes local state files