DevFlow v2.0

Universal AI Agent Development Workflow CLI with TOML-based progressive disclosure.

Overview

DevFlow v2.0 is a development workflow tool designed for AI-assisted software development. It uses TOML configuration to define workflows with progressive disclosure - AI agents see one step at a time, advancing only when gates are satisfied.

Key Features

• TOML Workflow Definition - Configure workflows in .devflow/workflows/*.toml
• Progressive Disclosure - One step at a time via CLI
• Gate System - Objective verification before advancement
• Fail Routes - Automatic routing on gate failure with escalation thresholds
• Cross-Workflow References - Seamless transitions between workflows (e.g., MODE-B → MODE-A)
• Workflow Inheritance - Extend existing workflows
• Prompt Reuse - File references for complex prompts
• Config Variables - Reference {test_command}, {lint_command} etc. from config
• Universal Design - Works with any programming language
• Ralph Loop - Autonomous execution: devflow run drives the workflow automatically
• VCS Abstraction - Git/SVN/no-VCS checkpointing via unified driver interface

Installation

pip install agent-devflow

Or from source:

git clone https://github.com/Dqz00116/devflow
cd devflow
pip install -e .

Quick Start

If .devflow/ already exists in the project, skip init and use it directly.

devflow init copies workflows and prompts from the DevFlow repository's own .devflow/ as the template.

# Only run init if .devflow/ does NOT already exist
devflow init --language python --name "My Project"

# List available workflows
devflow list-workflows

# Select a workflow (e.g., MODE-A for feature development)
devflow select-workflow MODE-A

# Follow the displayed instruction
devflow current

# When done, mark step complete and advance
devflow done

# Repeat until workflow complete

Workflow Commands

devflow list-workflows              # List all available workflows
devflow select-workflow MODE-A      # Select and start a workflow
devflow current                     # Show current step instruction
devflow done                        # Check gates and advance (or route on failure)
devflow back                        # Go back to previous step
devflow workflow-status             # Show workflow status
devflow approve ITEM                # Mark item as user-approved
devflow set KEY VALUE               # Set a state variable

# Ralph Loop — Autonomous Execution
devflow run [--tool local] [--max-iterations 10]   # Start autonomous loop
devflow loop-status                                  # Show loop status and remaining tasks
devflow sync-backlog                                 # Generate backlog from current workflow
devflow loop-reset                                   # Reset loop progress (preserves history)

Workflow Structure

Workflows are defined in TOML format:

# .devflow/workflows/my-workflow.toml
[workflow]
id = "MY-WORKFLOW"
name = "Feature Development"
extends = []  # Inheritance support

[[steps]]
id = "req-create"
name = "Create Requirement"
prompt = """
Your instruction here...
"""
gates = ["file_exists:docs/requirements/REQ-{workflow_run_id}.md"]
next = "next-step"

[[steps]]
id = "complex-step"
name = "Complex Step"
prompt_file = "prompts/complex.md"  # File reference
gates = ["command_success:{test_command}"]
next = ""

[[steps.fail_route]]
min_fails = 1
max_fails = 2
target = "fallback-step"    # Fails 1-2 times → go to fallback

[[steps.fail_route]]
min_fails = 3
target = "OTHER-WORKFLOW:escalation"  # Fails 3+ times → cross-workflow

Config Variables

Reference project commands from .devflow/config.toml in gates and prompts:

• {test_command} - Test command (e.g., pytest, go test ./...)
• {test_unit_command} - Unit test command
• {test_integration_command} - Integration test command
• {lint_command} - Lint command
• {build_command} - Build command
• {workflow_run_id} - Current workflow run ID (auto-generated)

Gate Conditions

Gates verify step completion before advancing:

• file_exists:path/to/file - File must exist
• file_exists_pattern:docs/REQ-*.md - Any file matching glob
• file_contains:path/file.md:content - File must contain text
• file_contains_pattern:docs/REQ-*.md:content - Any matching file contains text
• user_approved:ITEM-001 - User approval required (via devflow approve)
• command_success:pytest - Command must exit 0
• state_set:var_name - State variable must be set

Fail Routes

When gates fail, steps can define automatic routing based on failure count. This implements branching without requiring agent decisions.

[[steps]]
id = "debug-fix"
gates = ["command_success:{test_command}"]
next = "debug-verify"

[[steps.fail_route]]
min_fails = 1
max_fails = 2
target = "debug-root-cause"    # Fails 1-2 times → re-investigate

[[steps.fail_route]]
min_fails = 3
target = "debug-question"      # Fails 3+ times → escalate to human

Behavior:

• min_fails: Required, integer >= 1
• max_fails: Optional, defaults to infinity
• target: Step ID or cross-workflow reference (WORKFLOW-ID:STEP-ID)
• Routes match in declaration order; first match wins
• No matching route = stay and retry
• Fail count persists across visits (only resets when gates pass), enabling escalation

Cross-Workflow References

Steps can reference steps in other workflows:

[[steps]]
id = "debug-finish"
next = "MODE-A:write-plan"    # After debug, enter MODE-A at write-plan

On transition:

• Target workflow is loaded automatically
• current_workflow switches to the target
• workflow_run_id is preserved for file naming continuity

Included Workflows

MODE-A: Feature Development

1. req-create - Create requirement document
2. req-approve - Get user approval + create FEAT document
3. brainstorm - Explore design options
4. write-plan - Decompose into tasks
5. implement-sdd - TDD implementation
6. code-review - Review changes (requires user_approved)
7. test-run - Run all tests
8. verify - Collect evidence
9. finish - Deliver and cleanup (requires REQ status: done)

MODE-B: Debug

1. debug-root-cause - Investigate root cause
2. debug-pattern - Pattern analysis
3. debug-hypothesis - Form and test hypothesis (fail → back to root-cause)
4. debug-fix - Implement fix (fails 1-2x → root-cause; fails 3x → escalate)
5. debug-question - Question architecture (requires user_approved, then → root-cause)
6. debug-verify - Verify fix
7. debug-finish - Complete debug → auto-transitions to MODE-A:write-plan

Fail route escalation:

debug-fix fails 1-2 times → debug-root-cause (re-investigate)
debug-fix fails 3+ times  → debug-question (escalate to human)
debug-question approved   → debug-root-cause (try again with guidance)
debug-finish complete     → MODE-A:write-plan (full review cycle)

Project Structure

.devflow/
├── workflows/
│   ├── MODE-A.toml          # Feature development
│   └── MODE-B.toml          # Debug workflow
├── prompts/
│   ├── brainstorm.md        # Reusable prompts
│   ├── implement-sdd.md
│   └── ...
├── config.toml              # Project config
└── state.toml               # Current workflow & step (gitignored)

docs/
├── requirements/            # REQ files
├── features/                # FEAT files (technical implementation docs)
├── superpowers/specs/       # Design docs
├── superpowers/plans/       # Implementation plans
├── evidence/                # Verification evidence
├── debug/                   # Debug analysis
└── completion/              # Workflow completion summaries

Configuration

Project configuration in .devflow/config.toml:

[project]
name = "my-project"
language = "python"
stack = "Python 3.11 + FastAPI"

[commands]
test = "pytest"
test_unit = "pytest tests/unit -v"
test_integration = "pytest tests/integration -v"
lint = "ruff check ."
build = "python -m build"

[paths]
src = "src"
tests = "tests"
docs = "docs"

Legacy CLI Commands

DevFlow includes legacy commands for direct requirement/task management. These bypass the workflow engine; prefer workflow commands (current/done) for AI agents.

devflow init --language python --name "My Project"
devflow req new REQ-001 --title "Feature"
devflow feat new FEAT-001 -r REQ-001 -t "Implementation"
devflow task new -r REQ-001 -t "Task"
devflow status

License

MIT License