Epic Implement

Implementation phase of the Explore-Plan-Implement workflow. Execute the approved plan exactly, one phase at a time with verification.

Running this command implicitly approves the implementation plan.

CRITICAL: Each phase is executed by a dedicated subagent to keep main context low.

Arguments

Argument	Description
--phase N	Start at specific phase number
--continue	Resume from last recorded progress

Workflow

1. Locate active workflow: .claude/workflows/*/state.md
2. Load plan.md and determine starting point
3. For each phase:
   a. Launch phase-implementer subagent
   b. Wait for completion with TaskOutput
   c. Read phase results file
   d. If failed/blocked: STOP and report
   e. Launch implementation-validator subagent
   f. Update state.md progress
   g. Ask before proceeding to next phase
4. Mark complete when all phases done

Prerequisites

• Plan phase must be complete
• plan.md must exist in workflow directory

If prerequisites not met:

⚠️ Cannot implement - Plan phase incomplete
Run `/epic:plan` first.

Starting Point Logic

Condition	Action
--phase N provided	Start at phase N
--continue provided	Resume from state.md progress
state.md has progress	Ask: resume or restart?
Otherwise	Start at phase 1

Phase Execution (Subagent Pattern)

1. Launch Phase Implementer

For each phase, launch a fresh subagent:

Task tool:
  subagent_type: epic:phase-implementer
  run_in_background: false
  prompt: |
    Implement Phase [N] of the approved plan.

    Workflow path: .claude/workflows/[slug]/
    Phase number: [N]

    Read the plan.md and implement Phase [N] exactly as specified.
    Write results to: .claude/workflows/[slug]/phase-[N]-result.md

CRITICAL: Always launch a FRESH agent for each phase.

• Never use the resume parameter
• Each phase gets isolated context
• Fresh context prevents accumulated bloat

2. Process Phase Results

After subagent completes:

1. Read phase-[N]-result.md from workflow directory
2. Check status:
   • SUCCESS: Proceed to validation
   • FAILED: Report failure, STOP
   • BLOCKED: Report blocker, ask user for guidance

3. Validate Implementation

After successful phase implementation, launch validator:

Task tool:
  subagent_type: epic:implementation-validator
  run_in_background: false
  prompt: |
    Validate Phase [N] implementation against the plan.
    Plan: .claude/workflows/[slug]/plan.md
    Phase results: .claude/workflows/[slug]/phase-[N]-result.md
    Append validation results to: .claude/workflows/[slug]/review.md

    Focus on:
    - Changes made match Phase [N] specification
    - No unplanned additions or omissions
    - Verification steps completed

CRITICAL: Always launch a FRESH validator agent.

• Never use the resume parameter
• Each phase gets independent validation
• Fresh context ensures unbiased validation

4. Handle Validation Results

Status	Action
ALIGNED	Proceed to next phase
PARTIALLY ALIGNED	Review deviations, ask user before continuing
MISALIGNED	STOP - do not proceed, resolve issues first

5. Update Progress

Update state.md Current Progress section:

• Phase number completed
• Verification results
• Any deviations noted
• Timestamp

6. Phase Transition

After each phase, ask user with AskUserQuestion:

• "Continue to Phase [N+1]"
• "Pause here for review"
• "Stop implementation"

Output Format

Phase Launch

## Phase [N]: [Name]

Launching implementation subagent...

Phase Complete

✓ Phase [N] Complete

Changes: [summary from results file]
Verification: [status from results file]
Validation: [ALIGNED/PARTIALLY ALIGNED/MISALIGNED]
Progress: [N]/[total] phases complete

Next: [Continue to Phase N+1 / Manual verification required / Blocked]

Phase Failed

⚠️ Phase [N] Failed

Issue: [from results file]
Location: [file:line if applicable]

Resolution required before continuing.

Phase Failure Recovery

When a phase fails, provide recovery options via AskUserQuestion:

⚠️ Phase [N] Failed

Issue: [brief description]
Details: See .claude/workflows/[slug]/phase-[N]-result.md

Recovery Options:
1. Retry phase        - Run /epic:retry to try again
2. Retry with fixes   - Run /epic:retry --fix to analyze and fix
3. Rollback to plan   - Run /epic:rollback --phase 2 to revise plan
4. Manual intervention - I'll wait while you fix, then continue
5. Skip phase         - Mark failed and continue (not recommended)

Recovery Decision Tree

Phase Failed
    │
    ├─ Transient error (network, timeout)?
    │   └─ Suggest: /epic:retry
    │
    ├─ Clear code bug (test failure, type error)?
    │   └─ Suggest: /epic:retry --fix
    │
    ├─ Plan issue (missing info, wrong assumptions)?
    │   └─ Suggest: /epic:rollback --phase 2
    │
    └─ Unknown/complex issue?
        └─ Suggest: Manual intervention

Retry Tracking

Track retry attempts in state.md:

workflow:
  phases:
    implement:
      status: in_progress
      current_phase_num: 2
      retry_count: 1
      last_failure: "TypeError: Cannot read property 'auth'"

After 3 retries on same phase:

⚠️ Phase [N] has failed 3 times.

This suggests a deeper issue. Recommended:
1. /epic:rollback to revise the plan
2. Review research artifacts for missing context
3. Manual debugging before retry

Implementation Complete

✓ Implementation Phase Complete

Feature: [description]
Phases: [N]/[N] complete
Verification: All automated checks passed

Next: Run /epic:validate, then /epic:commit

Safety Rules

1. Plan is law - Subagents follow exactly; deviations require human approval
2. Verify every phase - Never skip validation
3. Document everything - state.md enables session resumption
4. Stop on failure - Don't proceed if verification or validation fails
5. Ask, don't assume - Use AskUserQuestion when uncertain
6. Fresh agents only - Never resume subagents, always fresh context

Context Management Benefits

Using subagents per phase provides:

• Low main context: Orchestration only, implementation details stay in subagents
• Isolated failures: One phase failing doesn't pollute context with debug noise
• Clean state: Each phase starts fresh, no accumulated assumptions
• Resumability: Can restart any phase independently

Deviation Protocol

When a subagent reports a deviation:

1. STOP - Do not launch next phase
2. READ - Get full details from phase-[N]-result.md
3. COMMUNICATE - Explain to user what's different
4. ASK - Get human guidance via AskUserQuestion
5. WAIT - Do not proceed without explicit direction

References

• references/progress-template.md - State.md progress section format