iterative-development（迭代开发）

Iterative Development Skill (Stop Hook TDD Loops)

Concept: Claude Code's Stop hook fires right before Claude finishes a response. Exit code 2 feeds stderr back to the model and continues the conversation. This creates a real TDD loop without any plugins.

How It Actually Works

Claude Code has a Stop hook that runs when Claude is about to conclude its response. If the hook script exits with code 2, its stderr is shown to the model and the conversation continues automatically.

┌─────────────────────────────────────────────────────────────┐
│  1. User asks Claude to implement a feature                 │
├─────────────────────────────────────────────────────────────┤
│  2. Claude writes tests + implementation                    │
├─────────────────────────────────────────────────────────────┤
│  3. Claude finishes its response                            │
├─────────────────────────────────────────────────────────────┤
│  4. Stop hook runs: executes tests, lint, typecheck         │
├─────────────────────────────────────────────────────────────┤
│  5a. All pass (exit 0) → Claude stops, work is done         │
│  5b. Failures (exit 2) → stderr fed back to Claude          │
├─────────────────────────────────────────────────────────────┤
│  6. Claude sees failures, fixes code, response ends         │
├─────────────────────────────────────────────────────────────┤
│  7. Stop hook runs again → repeat until green or max tries  │
└─────────────────────────────────────────────────────────────┘

Key insight: No fake plugins, no /ralph-loop command. The hook is real Claude Code infrastructure that runs automatically.

Setup: Stop Hook Configuration

Add this to your project's .claude/settings.json:

{
  "hooks": {
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "scripts/tdd-loop-check.sh",
            "timeout": 60,
            "statusMessage": "Running tests..."
          }
        ]
      }
    ]
  }
}

The TDD Loop Check Script

Create scripts/tdd-loop-check.sh in your project:

#!/bin/bash
# TDD Loop Check - runs after each Claude response
# Exit 0 = all good, Claude stops
# Exit 2 = failures, stderr fed back to Claude to fix

MAX_ITERATIONS=25
ITERATION_FILE=".claude/.tdd-iteration-count"

# Track iteration count
if [ -f "$ITERATION_FILE" ]; then
    count=$(cat "$ITERATION_FILE")
    count=$((count + 1))
else
    count=1
fi
echo "$count" > "$ITERATION_FILE"

# Safety: stop after max iterations
if [ "$count" -ge "$MAX_ITERATIONS" ]; then
    rm -f "$ITERATION_FILE"
    echo "Max iterations ($MAX_ITERATIONS) reached. Stopping loop." >&2
    exit 0
fi

# Skip if no test files exist yet
if ! find . -name "*.test.*" -o -name "*.spec.*" -o -name "test_*" 2>/dev/null | grep -q .; then
    rm -f "$ITERATION_FILE"
    exit 0
fi

# Run tests
TEST_OUTPUT=$(npm test 2>&1) || {
    echo "ITERATION $count/$MAX_ITERATIONS - Tests failing:" >&2
    echo "$TEST_OUTPUT" | tail -30 >&2
    echo "" >&2
    echo "Fix the failing tests and try again." >&2
    exit 2
}

# Run lint (if configured)
if [ -f "package.json" ] && grep -q '"lint"' package.json; then
    LINT_OUTPUT=$(npm run lint 2>&1) || {
        echo "ITERATION $count/$MAX_ITERATIONS - Lint errors:" >&2
        echo "$LINT_OUTPUT" | tail -20 >&2
        echo "" >&2
        echo "Fix lint errors and try again." >&2
        exit 2
    }
fi

# Run typecheck (if configured)
if [ -f "tsconfig.json" ]; then
    TYPE_OUTPUT=$(npx tsc --noEmit 2>&1) || {
        echo "ITERATION $count/$MAX_ITERATIONS - Type errors:" >&2
        echo "$TYPE_OUTPUT" | tail -20 >&2
        echo "" >&2
        echo "Fix type errors and try again." >&2
        exit 2
    }
fi

# All green - reset counter and let Claude stop
rm -f "$ITERATION_FILE"
exit 0

Python Variant

#!/bin/bash
# Python TDD Loop Check

MAX_ITERATIONS=25
ITERATION_FILE=".claude/.tdd-iteration-count"

if [ -f "$ITERATION_FILE" ]; then
    count=$(cat "$ITERATION_FILE")
    count=$((count + 1))
else
    count=1
fi
echo "$count" > "$ITERATION_FILE"

if [ "$count" -ge "$MAX_ITERATIONS" ]; then
    rm -f "$ITERATION_FILE"
    echo "Max iterations ($MAX_ITERATIONS) reached." >&2
    exit 0
fi

if ! find . -name "test_*" -o -name "*_test.py" 2>/dev/null | grep -q .; then
    rm -f "$ITERATION_FILE"
    exit 0
fi

TEST_OUTPUT=$(pytest -v 2>&1) || {
    echo "ITERATION $count/$MAX_ITERATIONS - Tests failing:" >&2
    echo "$TEST_OUTPUT" | tail -30 >&2
    exit 2
}

if command -v ruff &>/dev/null; then
    LINT_OUTPUT=$(ruff check . 2>&1) || {
        echo "ITERATION $count/$MAX_ITERATIONS - Lint errors:" >&2
        echo "$LINT_OUTPUT" | tail -20 >&2
        exit 2
    }
fi

if command -v mypy &>/dev/null; then
    TYPE_OUTPUT=$(mypy . 2>&1) || {
        echo "ITERATION $count/$MAX_ITERATIONS - Type errors:" >&2
        echo "$TYPE_OUTPUT" | tail -20 >&2
        exit 2
    }
fi

rm -f "$ITERATION_FILE"
exit 0

Additional Hooks for Quality Enforcement

PreToolUse Hook: Lint Before File Writes

Runs a linter before any Write/Edit lands:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "scripts/pre-write-lint.sh",
            "timeout": 10,
            "statusMessage": "Checking code quality..."
          }
        ]
      }
    ]
  }
}

SessionStart Hook: Auto-Inject Context

Runs at session start to inject project info:

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "echo 'TDD loop active. Tests run automatically after each response. Fix failures to continue.'",
            "statusMessage": "Loading project context..."
          }
        ]
      }
    ]
  }
}

Core Philosophy

┌─────────────────────────────────────────────────────────────┐
│  ITERATION > PERFECTION                                     │
│  ─────────────────────────────────────────────────────────  │
│  Don't aim for perfect on first try.                        │
│  Let the loop refine the work. Each iteration builds on     │
│  previous attempts visible in files and git history.        │
├─────────────────────────────────────────────────────────────┤
│  FAILURES ARE DATA                                          │
│  ─────────────────────────────────────────────────────────  │
│  Failed tests, lint errors, type mismatches are signals.    │
│  The Stop hook feeds them directly to Claude as context.    │
├─────────────────────────────────────────────────────────────┤
│  CLEAR COMPLETION CRITERIA                                  │
│  ─────────────────────────────────────────────────────────  │
│  The hook defines "done": tests pass, lint clean, types ok. │
│  No ambiguity about when to stop.                           │
└─────────────────────────────────────────────────────────────┘

Error Classification

Not all failures should loop. The hook script should distinguish:

The sample scripts above handle this — they only exit 2 for test/lint/type failures, not for environment issues.

When to Use TDD Loops

Good For

Not Good For

Disabling the Loop

To temporarily disable the TDD loop for a session:

1. Remove or rename the Stop hook in .claude/settings.json
2. Or set MAX_ITERATIONS=1 in the script
3. Or delete scripts/tdd-loop-check.sh

The hook only fires if the script exists and is configured.

Gitignore Additions

# TDD loop state
.claude/.tdd-iteration-count