Core Concepts
Ralph Workflow is built around a few key concepts that make AI-assisted development reliable and predictable.
The Pipeline
Every Ralph workflow runs through a structured pipeline with distinct phases:
PROMPT.md -> Planning -> Development -> Review -> Commit
This structure ensures that:
- Work is planned before execution begins
- Changes are reviewed before committing
- Each phase can be checkpointed and resumed
Pipeline Phases
1. Planning Phase
Ralph analyzes your PROMPT.md and generates a structured plan in PLAN.md. This plan includes:
- Steps to accomplish the task
- Files that will be modified
- Potential risks or considerations
Review the Plan
You can review PLAN.md before proceeding. If the plan looks wrong, edit your prompt and regenerate.
2. Development Phase
AI agents execute the plan iteratively. Each iteration:
- Analyzes the current state
- Makes targeted changes
- Verifies the changes work
- Decides if more iterations are needed
Ralph limits iterations to prevent runaway execution (configurable via max_dev_continuations).
3. Review Phase
An independent AI review checks the changes for:
- Bugs or logic errors
- Security issues
- Code style violations
- Missing tests or documentation
Issues are recorded in ISSUES.md. If issues are found, Ralph can attempt automatic fixes.
4. Commit Phase
Ralph generates an appropriate commit message based on the changes made. The message follows conventional commit format and summarizes what was done.
Checkpoints
Ralph saves progress at key points:
| Checkpoint | Saved After |
|---|---|
planning_complete |
Plan generated |
iteration_N |
Each development iteration |
review_complete |
Review finished |
commit_ready |
Commit message generated |
Checkpoints are stored in .agent/checkpoint.json. To resume:
$ ralph --resume
Agent Orchestration
Ralph coordinates multiple AI agents with fallback logic:
[developer]
fallback_chain = ["opencode", "ccs"]
[reviewer]
fallback_chain = ["ccs", "opencode"]
If the first agent fails (rate limit, auth error, etc.), Ralph automatically tries the next agent in the chain.
Supported Agents
| Agent | Role | Notes |
|---|---|---|
| opencode | Developer, Reviewer | General-purpose coding agent |
| ccs | Developer, Reviewer | Claude Code Selector |
Artifacts
Ralph creates several artifacts during execution:
.agent/
checkpoint.json # Pipeline state
tmp/
plan.xml # Structured plan data
issues.xml # Review findings
commit_message.xml # Generated commit message
logs/
run-2026-02-11/ # Per-run logs
These artifacts help with debugging and auditing workflow runs.
Next Steps
- Configuration - Customize Ralph's behavior
- CLI Commands - Full command reference