Skip to content

CLAUDE.md

Latest commit

 

History

History
628 lines (500 loc) · 26.6 KB

CLAUDE.md

File metadata and controls

628 lines (500 loc) · 26.6 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Repository Overview

This is the Ralph for Claude Code repository - an autonomous AI development loop system that enables continuous development cycles with intelligent exit detection and rate limiting.

Version: v0.9.8 | Tests: 276 passing (100% pass rate) | CI/CD: GitHub Actions

Core Architecture

The system consists of four main bash scripts and a modular library system:

Main Scripts

  1. ralph_loop.sh - The main autonomous loop that executes Claude Code repeatedly
  2. ralph_monitor.sh - Live monitoring dashboard for tracking loop status
  3. setup.sh - Project initialization script for new Ralph projects
  4. create_files.sh - Bootstrap script that creates the entire Ralph system
  5. ralph_import.sh - PRD/specification import tool that converts documents to Ralph format
    • Uses modern Claude Code CLI with --output-format json for structured responses
    • Implements detect_response_format() and parse_conversion_response() for JSON parsing
    • Backward compatible with older CLI versions (automatic text fallback)

Library Components (lib/)

The system uses a modular architecture with reusable components in the lib/ directory:

  1. lib/circuit_breaker.sh - Circuit breaker pattern implementation

    • Prevents runaway loops by detecting stagnation
    • Three states: CLOSED (normal), HALF_OPEN (monitoring), OPEN (halted)
    • Configurable thresholds for no-progress and error detection
    • Automatic state transitions and recovery

  2. lib/response_analyzer.sh - Intelligent response analysis

    • Analyzes Claude Code output for completion signals
    • JSON output format detection and parsing (with text fallback)
    • Supports both flat JSON format and Claude CLI format (result, sessionId, metadata)
    • Extracts structured fields: status, exit_signal, work_type, files_modified
    • Session management: store_session_id(), get_last_session_id(), should_resume_session()
    • Automatic session persistence to .claude_session_id file with 24-hour expiration
    • Session lifecycle: get_session_id(), reset_session(), log_session_transition(), init_session_tracking()
    • Session history tracked in .ralph_session_history (last 50 transitions)
    • Session auto-reset on: circuit breaker open, manual interrupt, project completion
    • Detects test-only loops and stuck error patterns
    • Two-stage error filtering to eliminate false positives
    • Multi-line error matching for accurate stuck loop detection
    • Confidence scoring for exit decisions

  3. lib/date_utils.sh - Cross-platform date utilities

    • ISO timestamp generation for logging
    • Epoch time calculations for rate limiting

Key Commands

Installation

# Install Ralph globally (run once)
./install.sh

# Uninstall Ralph
./install.sh uninstall

Setting Up a New Project

# Create a new Ralph-managed project (run from anywhere)
ralph-setup my-project-name
cd my-project-name

Running the Ralph Loop

# Start with integrated tmux monitoring (recommended)
ralph --monitor

# Start without monitoring
ralph

# With custom parameters and monitoring
ralph --monitor --calls 50 --prompt my_custom_prompt.md

# Check current status
ralph --status

# Circuit breaker management
ralph --reset-circuit
ralph --circuit-status

# Session management
ralph --reset-session    # Reset session state manually

Monitoring

# Integrated tmux monitoring (recommended)
ralph --monitor

# Manual monitoring in separate terminal
ralph-monitor

# tmux session management
tmux list-sessions
tmux attach -t <session-name>

Running Tests

# Run all tests (165 tests)
npm test

# Run specific test suites
npm run test:unit
npm run test:integration

# Run individual test files
bats tests/unit/test_cli_parsing.bats
bats tests/unit/test_json_parsing.bats
bats tests/unit/test_cli_modern.bats

Ralph Loop Configuration

The loop is controlled by several key files and environment variables:

  • PROMPT.md - Main prompt file that drives each loop iteration
  • @fix_plan.md - Prioritized task list that Ralph follows
  • @AGENT.md - Build and run instructions maintained by Ralph
  • status.json - Real-time status tracking (JSON format)
  • logs/ - Execution logs for each loop iteration

Rate Limiting

  • Default: 100 API calls per hour (configurable via --calls flag)
  • Automatic hourly reset with countdown display
  • Call tracking persists across script restarts

Modern CLI Configuration (Phase 1.1)

Ralph uses modern Claude Code CLI flags for structured communication:

Configuration Variables:

CLAUDE_OUTPUT_FORMAT="json"           # Output format: json (default) or text
CLAUDE_ALLOWED_TOOLS="Write,Bash(git *),Read"  # Allowed tool permissions
CLAUDE_USE_CONTINUE=true              # Enable session continuity
CLAUDE_MIN_VERSION="2.0.76"           # Minimum Claude CLI version

CLI Options:

  • --output-format json|text - Set Claude output format (default: json)
  • --allowed-tools "Write,Read,Bash(git *)" - Restrict allowed tools
  • --no-continue - Disable session continuity, start fresh each loop

Loop Context: Each loop iteration injects context via build_loop_context():

  • Current loop number
  • Remaining tasks from @fix_plan.md
  • Circuit breaker state (if not CLOSED)
  • Previous loop work summary

Session Continuity:

  • Sessions are preserved in .claude_session_id
  • Use --continue flag to maintain context across loops
  • Disable with --no-continue for isolated iterations

Intelligent Exit Detection

The loop uses a dual-condition check to prevent premature exits during productive iterations:

Exit requires BOTH conditions:

  1. recent_completion_indicators >= 2 (heuristic-based detection from natural language patterns)
  2. Claude's explicit EXIT_SIGNAL: true in the RALPH_STATUS block

The EXIT_SIGNAL value is read from .response_analysis (at .analysis.exit_signal) which is populated by response_analyzer.sh from Claude's RALPH_STATUS output block.

Other exit conditions (checked before completion indicators):

  • Multiple consecutive "done" signals from Claude Code (done_signals >= 2)
  • Too many test-only loops indicating feature completeness (test_loops >= 3)
  • All items in @fix_plan.md marked as completed

Example behavior when EXIT_SIGNAL is false:

Loop 5: Claude outputs "Phase complete, moving to next feature"
        → completion_indicators: 3 (high confidence from patterns)
        → EXIT_SIGNAL: false (Claude explicitly says more work needed)
        → Result: CONTINUE (respects Claude's explicit intent)

Loop 8: Claude outputs "All tasks complete, project ready"
        → completion_indicators: 4
        → EXIT_SIGNAL: true (Claude confirms project is done)
        → Result: EXIT with "project_complete"

Rationale: Natural language patterns like "done" or "complete" can trigger false positives during productive work (e.g., "feature done, moving to tests"). By requiring Claude's explicit EXIT_SIGNAL confirmation, Ralph avoids exiting mid-iteration when Claude is still working.

CI/CD Pipeline

Ralph uses GitHub Actions for continuous integration:

Workflows (.github/workflows/)

  1. test.yml - Main test suite

    • Runs on push to main/develop and PRs to main
    • Executes unit, integration, and E2E tests
    • Coverage reporting with kcov (informational only)
    • Uploads coverage artifacts

  2. claude.yml - Claude Code GitHub Actions integration

    • Automated code review capabilities

  3. claude-code-review.yml - PR code review workflow

    • Automated review on pull requests

Coverage Note

Bash code coverage measurement with kcov has fundamental limitations when tracing subprocess executions. The COVERAGE_THRESHOLD is set to 0 (disabled) because kcov cannot instrument subprocesses spawned by bats. Test pass rate (100%) is the quality gate. See bats-core#15 for details.

Project Structure for Ralph-Managed Projects

Each project created with ./setup.sh follows this structure:

project-name/
├── PROMPT.md          # Main development instructions
├── @fix_plan.md       # Prioritized TODO list
├── @AGENT.md          # Build/run instructions
├── specs/             # Project specifications
├── src/               # Source code
├── examples/          # Usage examples
├── logs/              # Loop execution logs
└── docs/generated/    # Auto-generated documentation

Template System

Templates in templates/ provide starting points for new projects:

  • PROMPT.md - Instructions for Ralph's autonomous behavior
  • fix_plan.md - Initial task structure
  • AGENT.md - Build system template

File Naming Conventions

  • Files prefixed with @ (e.g., @fix_plan.md) are Ralph-specific control files
  • Hidden files (e.g., .call_count, .exit_signals) track loop state
  • logs/ contains timestamped execution logs
  • docs/generated/ for Ralph-created documentation
  • docs/code-review/ for code review reports

Global Installation

Ralph installs to:

  • Commands: ~/.local/bin/ (ralph, ralph-monitor, ralph-setup, ralph-import)
  • Templates: ~/.ralph/templates/
  • Scripts: ~/.ralph/ (ralph_loop.sh, ralph_monitor.sh, setup.sh, ralph_import.sh)
  • Libraries: ~/.ralph/lib/ (circuit_breaker.sh, response_analyzer.sh, date_utils.sh)

After installation, the following global commands are available:

  • ralph - Start the autonomous development loop
  • ralph-monitor - Launch the monitoring dashboard
  • ralph-setup - Create a new Ralph-managed project
  • ralph-import - Import PRD/specification documents to Ralph format

Integration Points

Ralph integrates with:

  • Claude Code CLI: Uses npx @anthropic/claude-code as the execution engine
  • tmux: Terminal multiplexer for integrated monitoring sessions
  • Git: Expects projects to be git repositories
  • jq: For JSON processing of status and exit signals
  • GitHub Actions: CI/CD pipeline for automated testing
  • Standard Unix tools: bash, grep, date, etc.

Exit Conditions and Thresholds

Ralph uses multiple mechanisms to detect when to exit:

Exit Detection Thresholds

  • MAX_CONSECUTIVE_TEST_LOOPS=3 - Exit if too many test-only iterations
  • MAX_CONSECUTIVE_DONE_SIGNALS=2 - Exit on repeated completion signals
  • TEST_PERCENTAGE_THRESHOLD=30% - Flag if testing dominates recent loops
  • Completion detection via @fix_plan.md checklist items

Completion Indicators with EXIT_SIGNAL Gate

The completion_indicators exit condition requires dual verification:

completion_indicators EXIT_SIGNAL .response_analysis Result
>= 2 true exists Exit ("project_complete")
>= 2 false exists Continue (Claude still working)
>= 2 N/A missing Continue (defaults to false)
>= 2 N/A malformed Continue (defaults to false)
< 2 true exists Continue (threshold not met)

Implementation (ralph_loop.sh:312-327):

local claude_exit_signal="false"
if [[ -f ".response_analysis" ]]; then
    claude_exit_signal=$(jq -r '.analysis.exit_signal // false' ".response_analysis" 2>/dev/null || echo "false")
fi

if [[ $recent_completion_indicators -ge 2 ]] && [[ "$claude_exit_signal" == "true" ]]; then
    echo "project_complete"
    return 0
fi

Conflict Resolution: When STATUS: COMPLETE but EXIT_SIGNAL: false in RALPH_STATUS, the explicit EXIT_SIGNAL takes precedence. This allows Claude to mark a phase complete while indicating more phases remain.

Circuit Breaker Thresholds

  • CB_NO_PROGRESS_THRESHOLD=3 - Open circuit after 3 loops with no file changes
  • CB_SAME_ERROR_THRESHOLD=5 - Open circuit after 5 loops with repeated errors
  • CB_OUTPUT_DECLINE_THRESHOLD=70% - Open circuit if output declines by >70%

Error Detection

Ralph uses advanced error detection with two-stage filtering to eliminate false positives:

Stage 1: JSON Field Filtering

  • Filters out JSON field patterns like "is_error": false that contain the word "error" but aren't actual errors
  • Pattern: grep -v '"[^"]*error[^"]*":'

Stage 2: Actual Error Detection

  • Detects real error messages in specific contexts:
    • Error prefixes: Error:, ERROR:, error:
    • Context-specific errors: ]: error, Link: error
    • Error occurrences: Error occurred, failed with error
    • Exceptions: Exception, Fatal, FATAL
  • Pattern: grep -cE '(^Error:|^ERROR:|^error:|\]: error|Link: error|Error occurred|failed with error|[Ee]xception|Fatal|FATAL)'

Multi-line Error Matching

  • Detects stuck loops by verifying ALL error lines appear in ALL recent history files
  • Uses literal fixed-string matching (grep -qF) to avoid regex edge cases
  • Prevents false negatives when multiple distinct errors occur simultaneously

Test Suite

Test Files (265 tests total)

File Tests Description
test_cli_parsing.bats 27 CLI argument parsing for all 12 flags
test_cli_modern.bats 29 Modern CLI commands (Phase 1.1) + build_claude_command fix
test_json_parsing.bats 36 JSON output format parsing + Claude CLI format + session management
test_session_continuity.bats 26 Session lifecycle management + circuit breaker integration
test_exit_detection.bats 20 Exit signal detection
test_rate_limiting.bats 15 Rate limiting behavior
test_loop_execution.bats 20 Integration tests
test_edge_cases.bats 20 Edge case handling
test_installation.bats 14 Global installation/uninstall workflows
test_project_setup.bats 36 Project setup (setup.sh) validation
test_prd_import.bats 33 PRD import (ralph_import.sh) workflows + modern CLI tests

Running Tests

# All tests
npm test

# Unit tests only
npm run test:unit

# Specific test file
bats tests/unit/test_cli_parsing.bats

Recent Improvements

Modern CLI for PRD Import (v0.9.8)

  • Modernized ralph_import.sh to use Claude Code CLI JSON output format
    • Added --output-format json flag for structured responses
    • Implemented detect_response_format() for JSON vs text detection
    • Implemented parse_conversion_response() for extracting JSON fields
  • Enhanced error handling with structured JSON error messages
    • Extracts error_message and error_code from JSON metadata
    • Provides specific, actionable feedback on conversion failures
  • Improved file verification with JSON-derived status information
    • Reports files created vs missing based on JSON metadata
    • Logs session ID for potential conversion continuation
  • Backward compatibility with older CLI versions
    • Automatic fallback to text-based parsing when JSON unavailable
    • Version detection with check_claude_version() function
  • Enhanced logging with modern CLI awareness
    • Reports which CLI mode is being used
    • Detailed file creation status reporting
  • Added 11 new tests for modern CLI features (tests 23-33)
  • Test count: 276 (up from 265)

Session Lifecycle Management (v0.9.7)

  • Added complete session lifecycle management with automatic reset triggers:
    • get_session_id() - Retrieves current session from .ralph_session
    • reset_session(reason) - Clears session with reason logging
    • log_session_transition() - Records transitions to .ralph_session_history
    • init_session_tracking() - Initializes session file with validation
  • Session auto-reset integration points:
    • Circuit breaker open events (stagnation detection)
    • Manual interrupt (Ctrl+C / SIGINT)
    • Project completion (graceful exit)
    • Manual circuit breaker reset (--reset-circuit)
  • Added --reset-session CLI flag for manual session reset
  • Session history tracking (last 50 transitions) for debugging
  • New configuration constants: RALPH_SESSION_FILE, RALPH_SESSION_HISTORY_FILE
  • Added 26 new tests for session continuity features
  • Test count: 265 (up from 239)

JSON Output & Session Management (v0.9.6)

  • Extended parse_json_response() to support Claude Code CLI JSON format
    • Supports result, sessionId, and metadata fields alongside existing flat format
    • Extracts metadata.files_changed, metadata.has_errors, metadata.completion_status
    • Parses metadata.progress_indicators array for confidence boosting
  • Added session management functions for continuity tracking:
    • store_session_id() - Persists session with timestamp
    • get_last_session_id() - Retrieves stored session ID
    • should_resume_session() - Checks session validity (24-hour expiration)
  • Added get_epoch_seconds() to date_utils.sh for cross-platform epoch time
  • Auto-persists sessionId to .claude_session_id file during response analysis
  • Added 16 new tests covering Claude CLI format and session management
  • Test count: 239 (up from 223)

PRD Import Tests (v0.9.5)

  • Added 22 comprehensive tests for ralph_import.sh PRD conversion script
  • Tests cover: file format support (.md, .txt, .json), output file creation, project naming
  • Mock infrastructure for ralph-setup and Claude Code CLI isolation
  • Output file validation: PROMPT.md, @fix_plan.md, specs/requirements.md creation
  • Project naming tests: custom names, auto-detection from filename, path handling
  • Error handling tests: missing source file, missing ralph-setup, conversion failures
  • Help and usage tests: --help flag, no arguments behavior
  • Full workflow integration: complete project structure validation
  • Edge cases: hyphens in names, uppercase filenames, subdirectory paths
  • Test helper: added create_sample_prd_txt() fixture function
  • Test count: 223 (up from 201)

Project Setup Tests (v0.9.4)

  • Added 36 comprehensive tests for setup.sh project initialization script
  • Tests cover: directory creation, subdirectory structure, template copying, git initialization
  • Template copying verification for PROMPT.md, @fix_plan.md, @AGENT.md
  • Git repository validation: .git exists, valid repo, initial commit, correct message
  • README.md creation and content verification
  • Custom and default project name handling
  • Working directory behavior (nested directories, relative paths)
  • Error handling: missing templates, missing PROMPT.md
  • Output message validation (startup, completion, next steps)
  • Edge cases: spaces in names, existing directories
  • Test count: 201 (up from 165)

Installation Tests (v0.9.3)

  • Added 14 comprehensive tests for install.sh global installation script
  • Tests cover: directory creation, command installation, template copying, lib copying
  • Dependency detection tests (jq, git, node) with mocked failures
  • PATH detection and warning system tests
  • Uninstallation cleanup verification
  • Idempotency testing (run twice without errors)
  • End-to-end installation workflow validation
  • All tests use isolated temp directories (no system modifications)
  • Test count: 165 (up from 151)

Prompt File Fix (v0.9.2)

  • Fixed critical bug: replaced non-existent --prompt-file CLI flag with -p flag
  • Modern CLI mode now correctly passes prompt content via CLAUDE_CMD_ARGS+=("-p" "$prompt_content")
  • Added error handling for missing prompt files in build_claude_command()
  • Added 6 new TDD tests for build_claude_command function
  • Maintains shell injection safety through array-based command building
  • Test count: 151 (up from 145)

CLI Parsing Tests (v0.9.1)

  • Added 27 comprehensive CLI argument parsing tests
  • Covers all 12 CLI flags with both long and short forms
  • Boundary value testing for --timeout (0, 1, 120, 121)
  • Invalid input handling and error message validation
  • Code review report: docs/code-review/2026-01-08-cli-parsing-tests-review.md

CI/CD Pipeline (v0.9.1)

  • Added GitHub Actions workflow for automated testing
  • kcov coverage measurement (informational only due to subprocess limitations)
  • Coverage artifacts uploaded for debugging
  • Codecov integration (optional)

Modern CLI Commands (v0.9.1 - Phase 1.1)

JSON Output Format Support

  • Added detect_output_format() function to identify JSON vs text output
  • Added parse_json_response() to extract structured fields from Claude's JSON output
  • Extracts: status, exit_signal, work_type, files_modified, error_count, summary
  • Automatic fallback to text parsing on malformed JSON
  • Maintains backward compatibility with traditional RALPH_STATUS format

Session Continuity Management

  • init_claude_session() - Resume or start new sessions
  • save_claude_session() - Persist session ID from Claude output
  • --continue flag for context preservation across loops
  • --no-continue option for isolated iterations

Loop Context Injection

  • build_loop_context() - Build contextual information for each loop
  • Includes: loop number, remaining tasks, circuit breaker state, previous work summary
  • Injected via --append-system-prompt for Claude awareness

Modern CLI Flags

  • --output-format json|text - Control Claude output format
  • --allowed-tools - Restrict tool permissions
  • -p with content - Pass prompt content (reads from file via command substitution)
  • Version checking with check_claude_version()

Circuit Breaker Enhancements (v0.9.0)

Multi-line Error Matching Fix

  • Fixed critical bug in detect_stuck_loop function where only the first error line was checked when multiple distinct errors occurred
  • Now verifies ALL error lines appear in ALL recent history files for accurate stuck loop detection
  • Uses nested loop checking with grep -qF for literal fixed-string matching

JSON Field False Positive Elimination

  • Implemented two-stage error filtering to avoid counting JSON field names as errors
  • Stage 1 filters out patterns like "is_error": false that contain "error" as a field name
  • Stage 2 detects actual error messages in specific contexts
  • Aligned patterns between response_analyzer.sh and ralph_loop.sh for consistent behavior

Installation Improvements

  • Added lib/ directory to installation process for modular architecture
  • Fixed issue where response_analyzer.sh and circuit_breaker.sh were not being copied during global installation
  • All library components now properly installed to ~/.ralph/lib/

Feature Development Quality Standards

CRITICAL: All new features MUST meet the following mandatory requirements before being considered complete.

Testing Requirements

  • Test Pass Rate: 100% - all tests must pass, no exceptions
  • Test Types Required:
    • Unit tests for bash script functions (if applicable)
    • Integration tests for Ralph loop behavior
    • End-to-end tests for full development cycles
  • Test Quality: Tests must validate behavior, not just achieve coverage metrics
  • Test Documentation: Complex test scenarios must include comments explaining the test strategy

Note on Coverage: The 85% coverage threshold is aspirational for bash scripts. Due to kcov subprocess limitations, test pass rate is the enforced quality gate.

Git Workflow Requirements

Before moving to the next feature, ALL changes must be:

  1. Committed with Clear Messages:

    git add .
git commit -m "feat(module): descriptive message following conventional commits"
    • Use conventional commit format: feat:, fix:, docs:, test:, refactor:, etc.
    • Include scope when applicable: feat(loop):, fix(monitor):, test(setup):
    • Write descriptive messages that explain WHAT changed and WHY

  2. Pushed to Remote Repository:

    git push origin <branch-name>
    • Never leave completed features uncommitted
    • Push regularly to maintain backup and enable collaboration
    • Ensure CI/CD pipelines pass before considering feature complete

  3. Branch Hygiene:

    • Work on feature branches, never directly on main
    • Branch naming convention: feature/<feature-name>, fix/<issue-name>, docs/<doc-update>
    • Create pull requests for all significant changes

  4. Ralph Integration:

    • Update @fix_plan.md with new tasks before starting work
    • Mark items complete in @fix_plan.md upon completion
    • Update PROMPT.md if Ralph's behavior needs modification
    • Test Ralph loop with new features before completion

Documentation Requirements

ALL implementation documentation MUST remain synchronized with the codebase:

  1. Script Documentation:

    • Bash: Comments for all functions and complex logic
    • Update inline comments when implementation changes
    • Remove outdated comments immediately

  2. Implementation Documentation:

    • Update relevant sections in this CLAUDE.md file
    • Keep template files in templates/ current
    • Update configuration examples when defaults change
    • Document breaking changes prominently

  3. README Updates:

    • Keep feature lists current
    • Update setup instructions when commands change
    • Maintain accurate command examples
    • Update version compatibility information

  4. Template Maintenance:

    • Update template files when new patterns are introduced
    • Keep PROMPT.md template current with best practices
    • Update @AGENT.md template with new build patterns
    • Document new Ralph configuration options

  5. CLAUDE.md Maintenance:

    • Add new commands to "Key Commands" section
    • Update "Exit Conditions and Thresholds" when logic changes
    • Keep installation instructions accurate and tested
    • Document new Ralph loop behaviors or quality gates

Feature Completion Checklist

Before marking ANY feature as complete, verify:

  • All tests pass (if applicable)
  • Script functionality manually tested
  • All changes committed with conventional commit messages
  • All commits pushed to remote repository
  • CI/CD pipeline passes
  • @fix_plan.md task marked as complete
  • Implementation documentation updated
  • Inline code comments updated or added
  • CLAUDE.md updated (if new patterns introduced)
  • Template files updated (if applicable)
  • Breaking changes documented
  • Ralph loop tested with new features
  • Installation process verified (if applicable)

Rationale

These standards ensure:

  • Quality: Thorough testing prevents regressions in Ralph's autonomous behavior
  • Traceability: Git commits and @fix_plan.md provide clear history of changes
  • Maintainability: Current documentation reduces onboarding time and prevents knowledge loss
  • Collaboration: Pushed changes enable team visibility and code review
  • Reliability: Consistent quality gates maintain Ralph loop stability
  • Automation: Ralph integration ensures continuous development practices

Enforcement: AI agents should automatically apply these standards to all feature development tasks without requiring explicit instruction for each task.