Dev cycle

Dev Cycle - Development Session Manager

Orchestrated development workflow: Discovery → Planning → Review → Build

Agent Roles

• Scout: Codebase Explorer - explores before planning (haiku, fast, cheap)
• Bob: Technical Builder - builds code, ensures quality
• Garry: Chief of Staff - validates completion, checks alignment
• Arlo: Pattern Whisperer - involved when data/analytics/numbers are part of the work

Session State

Sessions are stored in ~/.claude/dev-cycles/active/ and ~/.claude/dev-cycles/completed/

Starting a Session

1. Check for Active Session

ls ~/.claude/dev-cycles/active/ | grep "$(basename $(pwd))"

If active session exists for this project:

• Read the session file
• Show current status (phase, completed tasks, blockers)
• Ask: "Continue this session or start new?"

2. Start New Session

Ask these questions in order:

1. Branch: "Work on existing branch, create new, or no branch?"

   • If existing: git branch -a to show options
   • If new: Ask for branch name (follow branch-workflow conventions)
   • If none: Work on current branch

2. Scope: "What are we building?"

   • Get clear description
   • For small tasks: Can collapse Discovery+Planning

3. Create session file:

~/.claude/dev-cycles/active/{project-name}--{YYYYMMDD-HHMMSS}.md

Session File Format

# Dev Cycle: {description}

## Meta
- **Project**: {project-name}
- **Branch**: {branch-name or "none"}
- **Started**: {timestamp}
- **Phase**: discovery | planning | review | building | blocked
- **Involves Data/Analytics**: yes | no

## Discovery
{Requirements, context, decisions made during discovery}

## Plan
{Tasks with checkboxes, phases, acceptance criteria}

## Progress
- [x] Task 1 - completed
- [ ] Task 2 - current
- [ ] Task 3 - pending

## Blockers
{Any blockers encountered - stop and ask the user}

## Batched Questions
{Easy questions to ask after build is done}

## Decisions
{Key decisions made during this session}

Phase Flow

Discovery Phase

1. Scout Exploration (for non-trivial tasks):

Delegate to Scout: "Explore [area relevant to task].
Find: where things live, how they work, existing patterns.
Questions: [specific unknowns from the user's description]"

Scout returns: What exists, how it works, options, unknowns.

2. Discovery Discussion:

• Review Scout's findings with the user
• Clarify edge cases
• Discuss options and tradeoffs
• Make decisions on approach

3. Update session file with discoveries

Exit: "Ready to plan?"

Skip Scout when:

• Small/trivial task
• You already know the codebase well
• The user says "just do it"

Planning Phase

• Create task list with phases
• Define acceptance criteria
• Consider technical approach
• For complex work: Use OpenSpec structure
• Update session file with plan
• Exit: "Ready for your review?"

Review Phase

• Present plan summary to the user
• The user reviews and adjusts
• Final approval to build
• Exit: The user says "go" or "approved" or similar

Build Phase

Orchestrated execution:

1. For each task/phase:
   └─ Delegate to Bob: "Build [task]. Context: [from session]"

2. After Bob completes:
   └─ Delegate to Garry: "Validate [task] is complete. Check quality."

3. If task involves data/analytics/numbers:
   └─ Delegate to Arlo: "Validate data logic, patterns, calculations in [task]"
   └─ Arlo checks: data flow, edge cases, number handling, analytics accuracy

4. I review output:
   └─ Problems? → Tell agents to fix
   └─ Blocker? → STOP, update session, ask the user
   └─ Easy question? → Add to "Batched Questions" section

5. Update session file with progress

6. Continue until all tasks done

When to Involve Arlo

Bring Arlo into the loop when the work involves:

• Analytics dashboards or reporting
• Data processing or transformation
• Calculations, aggregations, statistics
• Database queries with numeric results
• Charts, graphs, visualizations
• Pattern detection or anomaly detection
• Financial calculations
• Metrics, KPIs, measurements

Arlo validates:

• Data logic is correct
• Edge cases handled (nulls, zeros, negatives)
• Calculations are accurate
• Patterns make sense
• Numbers tell the right story

Handling Issues

Blockers (STOP immediately):

• Missing requirements
• External dependency needed
• Permission/access issues
• Architecture questions
• Anything that could go wrong

Batched Questions (ask after build):

• Naming preferences
• Minor style choices
• Optional enhancements
• "Nice to have" decisions

Small Task Mode

For small tasks (single feature, few files):

• Collapse Discovery + Planning into one conversation
• Still ask about branch
• Still create session file (for learning capture)
• Can skip formal Review phase if the user says "just do it"

Integration

With branch-workflow:

• Follow branch naming conventions
• Commit regularly during build
• Push to remote

With megg:

• Load context at start: mcp__megg__context
• Log significant decisions: mcp__megg__learn

With OpenSpec (for complex work):

• Create change proposal in openspec/changes/
• Follow openspec-workflow

Commands in Session

While in an active dev-cycle, the user can say:

• "continue" - Resume where we left off
• "skip to build" - Jump to build phase
• "pause" - Save state, stop for now
• "blocked on X" - Log blocker
• "/dev-finish" - Complete the cycle

Example Session Start

Me: Checking for active dev-cycle sessions...
    No active session for this project.

    Let's start a new dev cycle.

    1. Branch: Work on existing, create new, or stay on current?

User: new branch feat/add-user-auth

Me: Created branch feat/add-user-auth

    2. What are we building?

User: User authentication with JWT tokens

Me: Got it. Creating session...
    [Session file created]

    Let's start discovery. Tell me more about:
    - What auth methods do you need?
    - Any existing auth code to integrate with?
    - ...