frankbria/ralph-claude-code

Autonomous AI development loop with intelligent exit detection and rate limiting

Ralph is an implementation of the Geoffrey Huntley's technique for Claude Code that enables continuous autonomous development cycles he named after Ralph Wiggum. It enables continuous autonomous development cycles where Claude Code iteratively improves your project until completion, with built-in safeguards to prevent infinite loops and API overuse.

Install once, use everywhere - Ralph becomes a global command available in any directory.

Version: v0.11.5 - Active Development Core Features: Working and tested Test Coverage: 784 tests, 100% pass rate

• Autonomous development loops with intelligent exit detection
• Dual-condition exit gate: Requires BOTH completion indicators AND explicit EXIT_SIGNAL
• Rate limiting with hourly reset (100 calls/hour, configurable)
• Circuit breaker with advanced error detection (prevents runaway loops)
• Response analyzer with semantic understanding and two-stage error filtering
• JSON output format support with automatic fallback to text parsing
• Session continuity with --resume flag for context preservation (no session hijacking)
• Session expiration with configurable timeout (default: 24 hours)
• Modern CLI flags: --output-format, --allowed-tools, --no-continue
• Interactive project enablement with ralph-enable wizard
• .ralphrc configuration file for project settings
• Live streaming output with --live flag for real-time Claude Code visibility
• Log rotation: ralph.log rotates at 10MB, keeping 4 archived files
• Dry-run mode (--dry-run) to simulate loops without API calls
• Metrics tracking with ralph-stats analytics command (JSON Lines per-loop metrics)
• Desktop notifications (--notify) for key loop events (macOS/Linux/terminal-bell)
• Automatic git backup branches (--backup) with --rollback restore
• Multi-line error matching for accurate stuck loop detection
• 5-hour API limit handling with user prompts
• tmux integration for live monitoring
• PRD import functionality
• GitHub issue import: ralph-import --github-issue plus metadata filters (labels, title, assignee, milestone, state) with first/interactive/priority selection and --dry-run preview
• CI/CD pipeline with GitHub Actions
• Dedicated uninstall script for clean removal

Unreleased - Security: .ralphrc is parsed as data (#346)

• .ralphrc is no longer executed with source: only KEY=VALUE lines for known keys are applied, with literal values. Shell syntax ($VAR, $(...), chained commands) is rejected with a warning — breaking for .ralphrc files that relied on it
• CLAUDE_CODE_CMD, RALPH_SHELL_INIT_FILE, SANDBOX_DOCKER_IMAGE, SANDBOX_E2B_TEMPLATE and SANDBOX_DOCKER_NETWORK (no host) accept only stock values from .ralphrc; set custom values as environment variables (or --sandbox-* flags). Applies to ralph, ralph-import and ralph-queue. See docs/CLI_OPTIONS.md

v0.11.5 - Community Bug Fixes (latest)

• Fixed API limit false positive: Timeout (exit code 124) no longer misidentified as API 5-hour limit (#183)
• Three-layer API limit detection: timeout guard → structural JSON (rate_limit_event) → filtered text fallback
• Unattended mode: API limit prompt now auto-waits on timeout instead of exiting
• Fixed bash 3.x compatibility: ${,,} lowercase substitution replaced with POSIX tr (#187)
• Added 8 new tests for API limit detection (548 → 566 tests)

v0.11.4 - Bug Fixes & Compatibility

• Fixed progress detection: Git commits within a loop now count as progress (#141)
• Fixed checkbox regex: Date entries [2026-01-29] no longer counted as checkboxes (#144)
• Fixed session hijacking: Use --resume <session_id> instead of --continue (#151)
• Fixed EXIT_SIGNAL override: STATUS: COMPLETE with EXIT_SIGNAL: false now continues working (#146)
• Fixed ralph-import hanging indefinitely (added --print flag for non-interactive mode)
• Fixed ralph-import absolute path handling
• Fixed cross-platform date commands for macOS with Homebrew coreutils
• Added configurable circuit breaker thresholds via environment variables (#99)
• Added tmux support for non-zero base-index configurations
• Added 13 new regression tests for progress detection and checkbox regex

v0.11.3 - Live Streaming & Beads Fix

• Added live streaming output mode with --live flag for real-time Claude Code visibility (#125)
• Fixed beads task import using correct bd list arguments (#150)
• Applied CodeRabbit review fixes: camelCase variables, status-respecting fallback, jq guards
• Added 12 new tests for live streaming and beads import improvements

v0.11.2 - Setup Permissions Fix

• Fixed issue #136: ralph-setup now creates .ralphrc with consistent tool permissions
• Updated default ALLOWED_TOOLS to include Edit, Bash(npm *), and Bash(pytest)
• Both ralph-setup and ralph-enable now create identical .ralphrc configurations
• Monitor now forwards all CLI parameters to inner ralph loop (#126)
• Added 16 new tests for permissions and parameter forwarding

v0.11.1 - Completion Indicators Fix

• Fixed premature exit after exactly 5 loops in JSON output mode
• completion_indicators now only accumulates when EXIT_SIGNAL: true
• Aligns with documented dual-condition exit gate behavior

v0.11.0 - Ralph Enable Wizard

• Added ralph-enable interactive wizard for enabling Ralph in existing projects
• 5-phase wizard: Environment Detection → Task Source Selection → Configuration → File Generation → Verification
• Auto-detects project type (TypeScript, Python, Rust, Go) and framework (Next.js, FastAPI, Django)
• Imports tasks from beads, GitHub Issues, or PRD documents
• Added ralph-enable-ci non-interactive version for CI/automation
• New library components: enable_core.sh, wizard_utils.sh, task_sources.sh

v0.10.1 - Bug Fixes & Monitor Path Corrections

• Fixed ralph_monitor.sh hardcoded paths for v0.10.0 compatibility
• Fixed EXIT_SIGNAL parsing in JSON format
• Added safety circuit breaker (force exit after 5 consecutive completion indicators)
• Fixed checkbox parsing for indented markdown

v0.10.0 - .ralph/ Subfolder Structure (BREAKING CHANGE)

• Breaking: Moved all Ralph-specific files to .ralph/ subfolder
• Project root stays clean: only src/, README.md, and user files remain
• Added ralph-migrate command for upgrading existing projects

• Expanding test coverage
• Automated badge updates
• Multi-provider agent abstraction — decoupling Ralph from claude so any headless coding CLI (Codex, Gemini, OpenCode, Droid, Kilocode, Copilot) can drive the loop (multi-provider epic; see also the agent adapter contract)

Timeline to v1.0: final polish underway | Full roadmap | Contributions welcome!

• Autonomous Development Loop - Continuously executes Claude Code with your project requirements
• Intelligent Exit Detection - Dual-condition check requiring BOTH completion indicators AND explicit EXIT_SIGNAL
• Session Continuity - Preserves context across loop iterations with automatic session management
• Session Expiration - Configurable timeout (default: 24 hours) with automatic session reset
• Rate Limiting - Built-in API call management with hourly limits and countdown timers
• 5-Hour API Limit Handling - Three-layer detection (timeout guard, JSON parsing, filtered text) with auto-wait for unattended mode
• Live Monitoring - Real-time dashboard showing loop status, progress, and logs
• Task Management - Structured approach with prioritized task lists and progress tracking
• Project Templates - Quick setup for new projects with best-practice structure
• Interactive Project Setup - ralph-enable wizard for existing projects with task import
• Configuration Files - .ralphrc for project-specific settings and tool permissions
• Comprehensive Logging - Detailed execution logs with timestamps and status tracking
• Configurable Timeouts - Set execution timeout for Claude Code operations (1-120 minutes)
• Verbose Progress Mode - Optional detailed progress updates during execution
• Response Analyzer - AI-powered analysis of Claude Code responses with semantic understanding
• Circuit Breaker - Advanced error detection with two-stage filtering, multi-line error matching, and automatic recovery
• CI/CD Integration - GitHub Actions workflow with automated testing
• Clean Uninstall - Dedicated uninstall script for complete removal
• Live Streaming Output - Real-time visibility into Claude Code execution with --live flag
• Docker Sandbox Execution - Run Claude Code in an isolated container with --sandbox docker (resource limits, network policy, secure credential handoff)
• E2B Cloud Sandbox Execution - Run Claude Code in an E2B cloud sandbox with --sandbox e2b (file sync, session recovery, cost tracking with --sandbox-max-cost)

Ralph has two phases: one-time installation and per-project setup.

INSTALL ONCE              USE MANY TIMES
+-----------------+          +----------------------+
| ./install.sh    |    ->    | ralph-setup project1 |
|                 |          | ralph-enable         |
| Adds global     |          | ralph-import prd.md  |
| commands        |          | ...                  |
+-----------------+          +----------------------+

Install Ralph globally on your system:

git clone https://github.com/frankbria/ralph-claude-code.git
cd ralph-claude-code
./install.sh

This adds ralph, ralph-monitor, ralph-setup, ralph-import, ralph-queue, ralph-migrate, ralph-enable, and ralph-enable-ci commands to your PATH.

Note: You only need to do this once per system. After installation, you can delete the cloned repository if desired.

cd my-existing-project

# Interactive wizard - auto-detects project type and imports tasks
ralph-enable

# Or with specific task source
ralph-enable --from beads
ralph-enable --from github --label "sprint-1"
ralph-enable --from prd ./docs/requirements.md

# Start autonomous development
ralph --monitor

# Convert existing PRD/specs to Ralph format
ralph-import my-requirements.md my-project
cd my-project

# Review and adjust the generated files:
# - .ralph/PROMPT.md (Ralph instructions)
# - .ralph/fix_plan.md (task priorities)
# - .ralph/specs/requirements.md (technical specs)

# Start autonomous development
ralph --monitor

# Create blank Ralph project
ralph-setup my-awesome-project
cd my-awesome-project

# Configure your project requirements manually
# Edit .ralph/PROMPT.md with your project goals
# Edit .ralph/specs/ with detailed specifications
# Edit .ralph/fix_plan.md with initial priorities

# Start autonomous development
ralph --monitor

Once Ralph is installed and your project is initialized:

# Navigate to any Ralph project and run:
ralph --monitor              # Integrated tmux monitoring (recommended)

# Or use separate terminals:
ralph                        # Terminal 1: Ralph loop
ralph-monitor               # Terminal 2: Live monitor dashboard

To completely remove Ralph from your system:

# Run the uninstall script
./uninstall.sh

# Or if you deleted the repo, download and run:
curl -sL https://raw.githubusercontent.com/frankbria/ralph-claude-code/main/uninstall.sh | bash

After running ralph-enable or ralph-import, you'll have a .ralph/ directory with several files. Here's what each file does and whether you need to edit it:

PROMPT.md (high-level goals)
    ↓
specs/ (detailed requirements when needed)
    ↓
fix_plan.md (specific tasks Ralph executes)
    ↓
AGENT.md (build/test commands - auto-maintained)

• Simple projects: PROMPT.md + fix_plan.md is usually enough
• Complex features: Add specs/feature-name.md for detailed requirements
• Team conventions: Add specs/stdlib/convention-name.md for reusable patterns

See the User Guide for detailed explanations and the examples/ directory for realistic project configurations.

Ralph operates on a simple but powerful cycle:

1. Read Instructions - Loads PROMPT.md with your project requirements
2. Execute Claude Code - Runs Claude Code with current context and priorities
3. Track Progress - Updates task lists and logs execution results
4. Evaluate Completion - Checks for exit conditions and project completion signals
5. Repeat - Continues until project is complete or limits are reached

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

Exit requires BOTH conditions:

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

Example behavior:

Loop 5: Claude outputs "Phase complete, moving to next feature"
        → completion_indicators: 3 (high confidence from patterns)
        → EXIT_SIGNAL: false (Claude 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 done)
        → Result: EXIT with "project_complete"

Other exit conditions:

• All tasks in .ralph/fix_plan.md marked complete — except unchecked items under optional sections
• Multiple consecutive "done" signals from Claude Code
• Too many test-focused loops (indicating feature completeness)
• Claude API 5-hour usage limit reached (with user prompt to wait or exit)

By default, Ralph keeps looping until every - [ ] item is checked. To mark work as genuinely optional so it does not block completion, put it under an optional section:

## High Priority
- [x] Core feature

## Optional
- [ ] Frontend integration   # does NOT block exit
- [ ] SMS notifications      # does NOT block exit

Unchecked items under Optional, Future, Future Enhancements, or Nice to Have headings (and their subsections) are ignored by the completion check. Customize the section names with OPTIONAL_SECTIONS in .ralphrc (comma-separated, case-insensitive). This resolves the deadlock where Claude treats low-priority items as skippable while Ralph waits for them.

The ralph-enable command provides an interactive wizard for adding Ralph to existing projects:

cd my-existing-project
ralph-enable

The wizard:

1. Detects Environment - Identifies project type (TypeScript, Python, etc.) and framework
2. Selects Task Sources - Choose from beads, GitHub Issues, or PRD documents
3. Configures Settings - Set tool permissions and loop parameters
4. Generates Files - Creates .ralph/ directory and .ralphrc configuration
5. Verifies Setup - Confirms all files are created correctly

Non-interactive mode for CI/automation:

ralph-enable-ci                              # Sensible defaults
ralph-enable-ci --from github               # Import from GitHub Issues
ralph-enable-ci --project-type typescript   # Override detection
ralph-enable-ci --json                      # Machine-readable output

Ralph can convert existing PRDs, specifications, or requirement documents into the proper Ralph format using Claude Code.

• Markdown (.md) - Product requirements, technical specs
• Text files (.txt) - Plain text requirements
• JSON (.json) - Structured requirement data
• Word documents (.docx) - Business requirements
• PDFs (.pdf) - Design documents, specifications
• Any text-based format - Ralph will intelligently parse the content

# Convert a markdown PRD
ralph-import product-requirements.md my-app

# Convert a text specification
ralph-import requirements.txt webapp

# Convert a JSON API spec
ralph-import api-spec.json backend-service

# Let Ralph auto-name the project from filename
ralph-import design-doc.pdf

Ralph-import creates a complete project with:

• .ralph/PROMPT.md - Converted into Ralph development instructions
• .ralph/fix_plan.md - Requirements broken down into prioritized tasks
• .ralph/specs/requirements.md - Technical specifications extracted from your document
• .ralphrc - Project configuration file with tool permissions
• Standard Ralph structure - All necessary directories and template files in .ralph/

The conversion is intelligent and preserves your original requirements while making them actionable for autonomous development.

Ralph can also import a development plan directly from a GitHub issue. The issue body and discussion comments are converted into the same Ralph format as a local PRD.

• GitHub CLI (gh) installed: brew install gh or sudo apt install gh (see https://cli.github.com)
• Authenticated: gh auth login
• jq installed (used to parse issue JSON)

# Import a specific issue by number
ralph-import --github-issue 42

# Import the oldest open issue matching a search
ralph-import --github-search "fix login timeout"

# Import the oldest open issue with a label
ralph-import --github-label "sprint-1"

# Fetch from a specific repository (default: repo of the current directory)
ralph-import --github-issue 42 --repo myorg/myrepo

# Override the auto-generated project name (slug of the issue title)
ralph-import --github-issue 42 my-project

# Also import issue comments (e.g. when a plan was posted as a comment)
ralph-import --github-issue 42 --include-comments

The issue title becomes the project name (slugified, e.g. Add User Login → add-user-login) and the issue body becomes the PRD content. --github-issue addresses an exact issue and cannot be combined with the filter flags below.

Instead of an exact issue number, issues can be queried by metadata. Filter flags are freely combinable; matches are sorted oldest-first.

# Issues carrying ALL listed labels (comma = AND)
ralph-import --github-label "bug,P0"

# High-priority bugs assigned to me
ralph-import --github-label "bug,P0" --github-assignee @me

# Unassigned issues with a label, skipping anything marked wontfix
ralph-import --github-label "good-first-issue" --github-assignee none --exclude-label wontfix

# Title pattern: * matches anything, everything else is literal (case-insensitive)
ralph-import --github-title "[P0]*"

# Milestone, including closed issues
ralph-import --github-milestone "v1.0" --github-state all

When several issues match, --select picks the winner:

ralph-import --github-label bug --select first        # oldest issue (default)
ralph-import --github-label bug --select interactive  # numbered menu to choose from
ralph-import --github-label bug --select priority     # highest-priority label

The priority strategy understands both bare P0–P9 labels and the priority: P0 form; ties go to the oldest issue, and issues without priority labels fall back to the oldest match. Non-interactive sessions using --select interactive fall back to the oldest match with a warning instead of blocking.

--dry-run previews the matches and what would be selected without importing anything:

ralph-import --github-label bug --dry-run

Security note: issue comments are excluded by default — on public repositories anyone can comment, and comment text flows into the AI conversion prompt. Pass --include-comments only when you trust the discussion (e.g. plans posted by maintainers).

GitHub issues vary widely in detail. During import, Ralph scores the issue 0–100 for implementation readiness (acceptance criteria, task checklists, code examples, section structure, guidance keywords, length). Issues scoring below the threshold (default: 60) get an implementation plan generated by Claude Code before conversion — the plan is folded into the generated fix_plan.md tasks and preserved verbatim at .ralph/specs/implementation-plan.md.

# Default: assess the issue, generate a plan only if it lacks detail
ralph-import --github-issue 42

# Always generate a plan, with a specific model
ralph-import --github-issue 42 --generate-plan --plan-model opus

# Never generate a plan (fails if the issue scores below the threshold)
ralph-import --github-issue 42 --no-generate-plan

# Tune the threshold and skip the approval prompt (automation/CI)
ralph-import --github-issue 42 --completeness-threshold 75 --auto-approve

Generated plans are shown for approval before conversion. Non-interactive sessions (and --auto-approve) accept the plan automatically, so unattended imports never block on a prompt.

• "GitHub CLI (gh) is not installed" — install it from https://cli.github.com
• "GitHub CLI is not authenticated" — run gh auth login

Once development is underway, Ralph can close the loop on the whole GitHub workflow. Pass --github-issue <ref> (a number 69, #69, owner/repo#69, or a full issue URL) to ralph and opt into any of the lifecycle actions below. They all use the gh CLI (so the same gh auth login prerequisite applies), and every GitHub operation degrades gracefully — if a call is denied, Ralph logs a warning and keeps developing rather than crashing the loop.

# Post a progress comment to the issue every 5 loops while developing
ralph --github-issue 69 --comment-progress --comment-interval 5

# On completion: open a PR linked to the issue, comment a summary, and close it
ralph --github-issue 69 --create-pr --link-issue --close-summary --auto-close

# Add labels when closing, and open a follow-up issue for any TODO/FIXME left in the diff
ralph --github-issue 69 --auto-close --add-label completed \
      --create-followups --followup-label tech-debt

# Open the PR as a draft for manual review before merge
ralph --github-issue 69 --create-pr --draft-pr

These can also be set in .ralphrc (COMMENT_PROGRESS, AUTO_CLOSE, CREATE_PR, etc.). Lifecycle state is tracked in .ralph/.github_lifecycle_state.

For larger efforts, the ralph-queue command builds a persistent queue of work items (GitHub issues or local PRD specs) and processes them sequentially, respecting priority and dependencies. The queue is stored at .ralph/queue.json and survives restarts, so an interrupted run can be resumed. See docs/QUEUE_MANAGEMENT.md for the full guide.

# Queue issues by metadata filter (reuses the import filter flags)
ralph-queue add --github-label "bug,P0"
ralph-queue add --github-milestone "v1.0"

# Queue specific issues by number
ralph-queue add --github-issues 69,70,71

# Queue a local PRD/spec file
ralph-queue add --prd ./docs/feature.md

When a GitHub issue is queued, its priority is read from P0–P9 / priority: PN labels and its dependencies are parsed from the body (depends on #N, blocked by #N, requires #N).

ralph-queue status            # Show the queue (counts + items); --json for machine output
ralph-queue next              # Print the id of the next ready item
ralph-queue reorder           # Sort the queue by priority (P0 first)
ralph-queue validate          # Check for circular dependencies
ralph-queue remove 69         # Remove an item by issue number or id
ralph-queue clear             # Empty the queue

These are also available through ralph itself: ralph --queue-status, ralph --queue-next, ralph --queue-clear, and ralph --queue-remove <id|N>.

# Process all ready pending items in priority/dependency order
ralph --process-queue
ralph-queue process

# Stop at the first failure instead of skipping it
ralph --process-queue --halt-on-failure

# Resume after an interruption (only pending items are picked up)
ralph --resume-queue

For each ready item the processor stages the project from the issue/spec, runs the Ralph loop, commits the work as Fix #N: <title> (one commit per issue when in a git repo), then advances. Failed items are marked failed and skipped by default (or halt the run with --halt-on-failure); items whose dependencies never complete remain pending. Progress is written to .ralph/logs/queue_processing.log and shown in the ralph-monitor dashboard.

Concurrent (parallel) processing is intentionally out of scope — items are processed one at a time on a single branch.

• "Could not fetch issue #N" — check the issue number, your repo access, and the --repo value
• "No issues match the specified filters" — relax or remove some filters; only open issues are matched unless --github-state closed|all is given. --dry-run shows what a filter set matches.

Run Claude Code inside an isolated Docker container instead of directly on your machine (Issue #74). Ralph's loop, rate limiting, and monitoring stay on the host; only Claude's execution — the part that edits files and runs commands autonomously — is containerized. The project directory is bind-mounted read-write at /workspace, so changes land on the host directly and ralph-monitor works unchanged.

# One time: pull the official image (published on release tags)...
docker pull ghcr.io/frankbria/ralph-sandbox:latest
docker tag ghcr.io/frankbria/ralph-sandbox:latest ralph-sandbox:latest

# ...or build it yourself (from a source checkout, or ~/.ralph after install)
docker build -t ralph-sandbox .

ralph --sandbox docker                          # Default image, 4g RAM, 2 CPUs, bridge network
ralph --sandbox docker --sandbox-image node:20  # Any image with `claude` on PATH
ralph --sandbox docker --sandbox-memory 8g --sandbox-cpus 4
ralph --sandbox docker --sandbox-network none   # Full network isolation (blocks the Claude API —
                                                # only for images with their own auth/proxy setup)
ralph --monitor --sandbox docker                # Works with tmux monitoring

Equivalent .ralphrc settings: SANDBOX_PROVIDER, SANDBOX_DOCKER_IMAGE, SANDBOX_DOCKER_MEMORY, SANDBOX_DOCKER_CPUS, SANDBOX_DOCKER_NETWORK (CLI flags override).

• ANTHROPIC_API_KEY set → handed off via a 0600 env-file passed to docker run --env-file (never logged)
• Otherwise, your ~/.claude/.credentials.json is copied into a container-scoped directory mounted as the container's home — the container never touches your real ~/.claude
• Both are removed by the automatic cleanup on exit (including Ctrl+C)

A persistent container serves all loop iterations (Claude session state survives between loops); it is stopped and removed when the loop exits. If sandbox setup fails, Ralph refuses to fall back to host execution. See docs/DOCKER_SANDBOX.md for details.

Run Claude Code inside an E2B cloud sandbox instead of on your machine (Issue #75). Like the Docker provider, Ralph's loop, rate limiting, and monitoring stay host-side — but since the cloud has no bind mount, the project is uploaded once at startup and changed files are downloaded back after every iteration.

pip install e2b                       # official E2B Python SDK (the transport)
export E2B_API_KEY="e2b_..."          # or: store in ~/.ralph/e2b_api_key (chmod 600)

ralph --sandbox e2b                                   # base template; claude CLI auto-bootstrapped
ralph --sandbox e2b --sandbox-template my-template    # custom template with claude preinstalled
ralph --sandbox e2b --sandbox-timeout 7200            # 2h session timeout (expired sandboxes auto-recreate)
ralph --sandbox e2b --sandbox-max-cost 5.00 --sandbox-cost-alert 2.00   # budget controls
ralph --sandbox e2b --sandbox-keep-alive              # leave it running; reuse with --sandbox-id <id>
ralph --monitor --sandbox e2b                         # works with tmux monitoring (cost shown in the dashboard)

# Sync filtering (Issue #76): control what uploads/downloads
ralph --sandbox e2b --sync-include "src/**,tests/**,*.md" --sync-exclude "*.log,node_modules"

Equivalent .ralphrc settings: SANDBOX_E2B_TEMPLATE, SANDBOX_E2B_TIMEOUT, SANDBOX_E2B_KEEP_ALIVE, SANDBOX_E2B_MAX_COST, SANDBOX_E2B_COST_ALERT, SANDBOX_E2B_COST_PER_HOUR (CLI flags override; the API key itself never goes in .ralphrc).

Claude auth reaches the sandbox via ANTHROPIC_API_KEY (passed as a sandbox env var) or a copy of your host ~/.claude/.credentials.json. Cost is estimated as cumulative runtime × SANDBOX_E2B_COST_PER_HOUR (cost accrues across sandbox replacements, so --sandbox-max-cost spans the whole run) and surfaced in status.json, the monitor, and .ralph/logs/e2b_cost.log. File sync includes deletion propagation — files removed or renamed in the sandbox are deleted from the host after each iteration. What syncs is filterable: --sync-include/--sync-exclude patterns, a .ralphignore file (gitignore-like subset), and a large-file policy (SYNC_MAX_FILE_SIZE/SYNC_LARGE_FILE_ACTION); upload and download log file-count + size summaries. Note: commits made inside the sandbox are not synced back — changes arrive as uncommitted modifications on the host (.git is excluded from sync in both directions). Docker and E2B are the supported providers; additional providers (Daytona, Cloudflare) are not planned (#79, #80). See docs/E2B_SANDBOX.md and docs/SANDBOX_SYNC.md for details.

Each Ralph project can have a .ralphrc configuration file:

# .ralphrc - Ralph project configuration
PROJECT_NAME="my-project"
PROJECT_TYPE="typescript"

# Claude Code CLI command (auto-detected, override if needed)
CLAUDE_CODE_CMD="claude"
# CLAUDE_CODE_CMD="npx @anthropic-ai/claude-code"  # Alternative: use npx

# Custom CLI paths and shell init files are environment-only (never read from
# .ralphrc, which is repository-controlled):
#   export CLAUDE_CODE_CMD="/opt/local/bin/claude"
#   export RALPH_SHELL_INIT_FILE="$HOME/.zshrc"

# Loop settings
MAX_CALLS_PER_HOUR=100
CLAUDE_TIMEOUT_MINUTES=15
CLAUDE_OUTPUT_FORMAT="json"
# Token budget per hour (0 = disabled). One Claude call can use 100k+ tokens.
#MAX_TOKENS_PER_HOUR=500000

# Tool permissions
ALLOWED_TOOLS="Write,Read,Edit,Bash(git *),Bash(npm *),Bash(pytest)"

# Session management
SESSION_CONTINUITY=true
SESSION_EXPIRY_HOURS=24

# Circuit breaker thresholds
CB_NO_PROGRESS_THRESHOLD=3
CB_SAME_ERROR_THRESHOLD=5

Ralph includes intelligent rate limiting and circuit breaker functionality:

# Default: 100 calls per hour
ralph --calls 50

# With integrated monitoring
ralph --monitor --calls 50

# Check current usage (shows calls and tokens used this hour)
ralph --status

Rate limiting supports two independent limits — both reset hourly:

Token tracking extracts input_tokens + output_tokens from each Claude response. A single call can consume 100k+ tokens, so MAX_TOKENS_PER_HOUR provides cost control that MAX_CALLS_PER_HOUR alone cannot.

The circuit breaker automatically:

• Detects API errors and rate limit issues with advanced two-stage filtering
• Opens circuit after 3 loops with no progress or 5 loops with same errors
• Eliminates false positives from JSON fields containing "error"
• Accurately detects stuck loops with multi-line error matching
• Gradually recovers with half-open monitoring state
• Auto-recovers after cooldown period (default: 30 minutes) — OPEN → HALF_OPEN → CLOSED
• Provides detailed error tracking and logging with state history

Auto-recovery options:

# Default: 30-minute cooldown before auto-recovery attempt
CB_COOLDOWN_MINUTES=30     # Set in .ralphrc (0 = immediate)

# Auto-reset on startup (for fully unattended operation)
ralph --auto-reset-circuit
# Or set in .ralphrc: CB_AUTO_RESET=true

When Claude's 5-hour usage limit is reached, Ralph:

1. Detects the limit using three-layer verification (timeout guard → structural JSON → filtered text fallback)
2. Prompts you to choose:
   • Option 1: Wait 60 minutes for the limit to reset (with countdown timer)
   • Option 2: Exit gracefully
3. Unattended mode: Auto-waits on prompt timeout (30s) instead of exiting
4. Prevents false positives from echoed file content mentioning "5-hour limit"

# Use custom prompt file
ralph --prompt my_custom_instructions.md

# With integrated monitoring
ralph --monitor --prompt my_custom_instructions.md

# Set Claude Code execution timeout (default: 15 minutes)
ralph --timeout 30  # 30-minute timeout for complex tasks

# With monitoring and custom timeout
ralph --monitor --timeout 60  # 60-minute timeout

# Short timeout for quick iterations
ralph --verbose --timeout 5  # 5-minute timeout with progress

# Enable detailed progress updates during execution
ralph --verbose

# Combine with other options
ralph --monitor --verbose --timeout 30

# Enable real-time visibility into Claude Code execution
ralph --live

# Combine with monitoring for best experience
ralph --monitor --live

# Live output is written to .ralph/live.log
tail -f .ralph/live.log  # Watch in another terminal

Live streaming mode shows Claude Code's output in real-time as it works, providing visibility into what's happening during each loop iteration.

Ralph maintains session context across loop iterations for improved coherence:

# Sessions are enabled by default with --continue flag
ralph --monitor                 # Uses session continuity

# Start fresh without session context
ralph --no-continue             # Isolated iterations

# Reset session manually (clears context)
ralph --reset-session           # Clears current session

# Check session status
cat .ralph/.ralph_session              # View current session file
cat .ralph/.ralph_session_history      # View session transition history

Session Auto-Reset Triggers:

• Circuit breaker opens (stagnation detected)
• Manual interrupt (Ctrl+C / SIGINT)
• Project completion (graceful exit)
• Manual circuit breaker reset (--reset-circuit)
• Session expiration (default: 24 hours)

Sessions are persisted to .ralph/.ralph_session with a configurable expiration (default: 24 hours). The last 50 session transitions are logged to .ralph/.ralph_session_history for debugging.

Set these in your project's .ralphrc, or as environment variables (which take precedence over .ralphrc):

Exit Detection Thresholds:

MAX_CONSECUTIVE_TEST_LOOPS=3     # Exit after 3 test-only loops
MAX_CONSECUTIVE_DONE_SIGNALS=2   # Exit after 2 "done" signals
TEST_PERCENTAGE_THRESHOLD=30     # Flag if 30%+ loops are test-only

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%
CB_COOLDOWN_MINUTES=30           # Minutes before OPEN → HALF_OPEN auto-recovery
CB_AUTO_RESET=false              # true = reset to CLOSED on startup (bypasses cooldown)

Completion Indicators with EXIT_SIGNAL Gate:

Ralph creates a standardized structure for each project with a .ralph/ subfolder for configuration:

my-project/
├── .ralph/                 # Ralph configuration and state (hidden folder)
│   ├── PROMPT.md           # Main development instructions for Ralph
│   ├── fix_plan.md        # Prioritized task list
│   ├── AGENT.md           # Build and run instructions
│   ├── specs/              # Project specifications and requirements
│   │   └── stdlib/         # Standard library specifications
│   ├── examples/           # Usage examples and test cases
│   ├── logs/               # Ralph execution logs
│   └── docs/generated/     # Auto-generated documentation
├── .ralphrc                # Ralph configuration file (tool permissions, settings)
└── src/                    # Source code implementation (at project root)

Migration: If you have existing Ralph projects using the old flat structure, run ralph-migrate to automatically move files to the .ralph/ subfolder.

1. Be Specific - Clear requirements lead to better results
2. Prioritize - Use .ralph/fix_plan.md to guide Ralph's focus
3. Set Boundaries - Define what's in/out of scope
4. Include Examples - Show expected inputs/outputs

• Place detailed requirements in .ralph/specs/
• Use .ralph/fix_plan.md for prioritized task tracking
• Keep .ralph/AGENT.md updated with build instructions
• Document key decisions and architecture

• Use ralph-monitor for live status updates
• Check logs in .ralph/logs/ for detailed execution history
• Monitor .ralph/status.json for programmatic access
• Watch for exit condition signals

• Bash 4.0+ - For script execution
• Claude Code CLI - npm install -g @anthropic-ai/claude-code (or use npx — set CLAUDE_CODE_CMD in .ralphrc)
• tmux - Terminal multiplexer for integrated monitoring (recommended)
• jq - JSON processing for status tracking
• Git - Version control (projects are initialized as git repos)
• GNU coreutils - For the timeout command (execution timeouts)
  • Linux: Pre-installed on most distributions
  • macOS: Install via brew install coreutils (provides gtimeout)
• Standard Unix tools - grep, date, etc.

See TESTING.md for the comprehensive testing guide.

If you want to run the test suite:

# Install BATS testing framework
npm install -g bats bats-support bats-assert

# Run unit + integration tests (771 tests)
npm test

# Run end-to-end tests (13 tests; full ralph_loop.sh subprocess runs)
npm run test:e2e

# Run specific test suites
bats tests/unit/test_rate_limiting.bats
bats tests/unit/test_exit_detection.bats
bats tests/unit/test_json_parsing.bats
bats tests/unit/test_cli_modern.bats
bats tests/unit/test_cli_parsing.bats
bats tests/unit/test_session_continuity.bats
bats tests/unit/test_enable_core.bats
bats tests/unit/test_task_sources.bats
bats tests/unit/test_ralph_enable.bats
bats tests/unit/test_wizard_utils.bats
bats tests/unit/test_circuit_breaker_recovery.bats
bats tests/integration/test_loop_execution.bats
bats tests/integration/test_prd_import.bats
bats tests/integration/test_project_setup.bats
bats tests/integration/test_installation.bats

# Run error detection and circuit breaker tests
./tests/test_error_detection.sh
./tests/test_stuck_loop_detection.sh

Current test status:

• 784 tests across 34 test files
• 100% pass rate (784/784 passing)
• Comprehensive unit, integration, and end-to-end tests
• Specialized tests for JSON parsing, CLI flags, circuit breaker, EXIT_SIGNAL behavior, enable wizard, and installation workflows
• True E2E suite running ralph_loop.sh as a subprocess against a mock Claude CLI (tests/e2e/)

Note on Coverage: Bash code coverage measurement with kcov has fundamental limitations when tracing subprocess executions. Test pass rate (100%) is the quality gate. See bats-core#15 for details.

# Ubuntu/Debian
sudo apt-get install tmux

# macOS
brew install tmux

# CentOS/RHEL
sudo yum install tmux

Ralph uses the timeout command for execution timeouts. On macOS, you need to install GNU coreutils:

# Install coreutils (provides gtimeout)
brew install coreutils

# Verify installation
gtimeout --version

Ralph automatically detects and uses gtimeout on macOS. No additional configuration is required after installation.

# Integrated tmux monitoring (recommended)
ralph --monitor

# Manual monitoring in separate terminal
ralph-monitor

Shows real-time:

• Current loop count and status
• API calls used vs. limit
• Recent log entries
• Rate limit countdown

tmux Controls:

• Ctrl+B then D - Detach from session (keeps Ralph running)
• Ctrl+B then ←/→ - Switch between panes
• tmux list-sessions - View active sessions
• tmux attach -t <session-name> - Reattach to session

# JSON status output
ralph --status

# Manual log inspection
tail -f .ralph/logs/ralph.log

• Ralph exits silently on first loop - Claude Code CLI may not be installed or not in PATH. Ralph validates the command at startup and shows installation instructions. If using npx, add CLAUDE_CODE_CMD="npx @anthropic-ai/claude-code" to .ralphrc
• Rate Limits - Ralph automatically waits and displays countdown
• 5-Hour API Limit - Ralph detects and prompts for user action (wait or exit)
• Stuck Loops - Check fix_plan.md for unclear or conflicting tasks
• Early Exit - Review exit thresholds if Ralph stops too soon
• Premature Exit - Check if Claude is setting EXIT_SIGNAL: false (Ralph now respects this)
• Execution Timeouts - Increase --timeout value for complex operations
• Missing Dependencies - Ensure Claude Code CLI and tmux are installed
• tmux Session Lost - Use tmux list-sessions and tmux attach to reconnect
• Session Expired - Sessions expire after 24 hours by default; use --reset-session to start fresh
• timeout: command not found (macOS) - Install GNU coreutils: brew install coreutils
• Permission Denied - Ralph halts when Claude Code is denied permission for commands:
  1. Edit .ralphrc and update ALLOWED_TOOLS to include required tools
  2. Common patterns: Bash(npm *), Bash(git *), Bash(pytest)
  3. Run ralph --reset-session after updating .ralphrc
  4. Restart with ralph --monitor

Ralph is actively seeking contributors! We're working toward v1.0.0 with clear priorities and a detailed roadmap.

See CONTRIBUTING.md for the complete contributor guide including:

• Getting started and setup instructions
• Development workflow and commit conventions
• Code style guidelines
• Testing requirements (100% pass rate mandatory)
• Pull request process and code review guidelines
• Quality standards and checklists

# Fork and clone
git clone https://github.com/YOUR_USERNAME/ralph-claude-code.git
cd ralph-claude-code

# Install dependencies and run tests
npm install
npm test && npm run test:e2e  # All 784 tests must pass

1. Test Implementation - Help expand test coverage
2. Feature Development - Pick up an open issue
3. Documentation - Tutorials, troubleshooting guides, examples
4. Real-World Testing - Use Ralph, report bugs, share feedback

Every contribution matters - from fixing typos to implementing major features!

This project is licensed under the MIT License - see the LICENSE file for details.

• Inspired by the Ralph technique created by Geoffrey Huntley
• Built for Claude Code by Anthropic
• Community feedback and contributions

• Claude Code - The AI coding assistant that powers Ralph
• Aider - Original Ralph technique implementation

./install.sh              # Install Ralph globally
./uninstall.sh            # Remove Ralph from system (dedicated script)
./install.sh uninstall    # Alternative: Remove Ralph from system
./install.sh --help       # Show installation help
ralph-migrate             # Migrate existing project to .ralph/ structure

ralph [OPTIONS]
  -h, --help              Show help message
  -c, --calls NUM         Set max calls per hour (default: 100)
  -p, --prompt FILE       Set prompt file (default: .ralph/PROMPT.md)
  -s, --status            Show current status and exit
  -m, --monitor           Start with tmux session and live monitor
  -v, --verbose           Show detailed progress updates during execution
  -l, --live              Enable live streaming output (real-time Claude Code visibility)
  -t, --timeout MIN       Set Claude Code execution timeout in minutes (1-120, default: 15)
  --dry-run               Simulate loop execution without making actual Claude API calls
  -n, --notify            Enable desktop notifications for key events
  -b, --backup            Enable automatic git backup branch before each loop (requires git)
  --rollback [BRANCH]     Roll back to a backup branch (lists available branches if none given)
  --show-tool-args        Show tool arguments (commands, file paths) in live streaming output
  --output-format FORMAT  Set output format: json (default) or text
  --allowed-tools TOOLS   Set allowed Claude tools (default: granular git subcommands + npm/pytest)
  --no-continue           Disable session continuity (start fresh each loop)
  --session-expiry HOURS  Set session expiration time in hours (default: 24)
  --reset-circuit         Reset the circuit breaker
  --circuit-status        Show circuit breaker status
  --auto-reset-circuit    Auto-reset circuit breaker on startup (bypasses cooldown)
  --reset-session         Reset session state manually
  --queue-status          Show the issue queue and exit
  --process-queue         Process pending queued issues sequentially (--halt-on-failure to stop on first failure)
  --resume-queue          Resume processing the remaining pending issues
  --queue-next            Print the id of the next ready queued issue
  --queue-clear           Remove all items from the queue
  --queue-remove <id|N>   Remove one item from the queue

Full reference: every flag is documented in depth, with examples and .ralphrc patterns, in docs/CLI_OPTIONS.md.

ralph-setup project-name     # Create new Ralph project
ralph-enable                 # Enable Ralph in existing project (interactive)
ralph-enable-ci              # Enable Ralph in existing project (non-interactive)
ralph-import prd.md project  # Convert PRD/specs to Ralph project
ralph-queue add --github-label bug   # Build a batch queue of issues
ralph-queue status           # Show the issue queue
ralph --process-queue        # Process the queue sequentially
ralph --monitor              # Start with integrated monitoring
ralph --status               # Check current loop status
ralph --verbose              # Enable detailed progress updates
ralph --timeout 30           # Set 30-minute execution timeout
ralph --calls 50             # Limit to 50 API calls per hour
ralph --reset-session        # Reset session state manually
ralph --live                 # Enable live streaming output
ralph --dry-run              # Simulate a loop without API calls
ralph --notify               # Desktop notifications for key events
ralph --backup               # Git backup branch before each loop
ralph-stats                  # Metrics summary from .ralph/logs/metrics.jsonl
ralph-monitor                # Manual monitoring dashboard

tmux list-sessions        # View active Ralph sessions
tmux attach -t <name>     # Reattach to detached session
# Ctrl+B then D           # Detach from session (keeps running)

Ralph is under active development with a clear path to v1.0.0. See IMPLEMENTATION_PLAN.md for the complete roadmap.

What's Delivered:

• Core loop functionality with intelligent exit detection
• Dual-condition exit gate (completion indicators + EXIT_SIGNAL)
• Rate limiting (100 calls/hour) and circuit breaker pattern
• Response analyzer with semantic understanding
• 784 comprehensive tests (100% pass rate)
• Live streaming output mode for real-time Claude Code visibility
• Log rotation, dry-run mode, metrics (ralph-stats), desktop notifications, and git backup/rollback
• tmux integration and live monitoring
• PRD import functionality with modern CLI JSON parsing
• Installation system and project templates
• Modern CLI commands with JSON output support
• CI/CD pipeline with GitHub Actions
• Interactive ralph-enable wizard for existing projects
• .ralphrc configuration file support
• Session lifecycle management with auto-reset triggers
• Session expiration with configurable timeout
• Dedicated uninstall script

Test Coverage:

• 784 tests across 34 test files — unit, integration, and end-to-end (100% pass rate)
• See TESTING.md for the per-suite breakdown and how to run each category

Completed milestones: installation/setup workflow tests, tmux integration tests, monitor dashboard tests, end-to-end full-loop tests, log rotation, dry-run mode, metrics (ralph-stats), desktop notifications, and git backup/rollback.

Remaining

• Final documentation sweep and release prep
• Supply-chain hardening for CI workflows (#275)
• Real-world feedback from the open issues backlog

See IMPLEMENTATION_STATUS.md for detailed progress tracking.

Ralph is seeking contributors! See CONTRIBUTING.md for the complete guide. Priority areas:

1. Test Implementation - Help expand test coverage (see plan)
2. Feature Development - Pick up an open issue
3. Documentation - Usage examples, tutorials, troubleshooting guides
4. Bug Reports - Real-world usage feedback and edge cases

Ready to let AI build your project? Start with ./install.sh and let Ralph take it from there!