Skip to content

Complete Cycle Workflow

Purpose: Execute a full development cycle from research to learning

When to use: - New features (complete implementation needed) - Complex changes (multiple files, dependencies) - Architectural changes (requires research and planning) - Learning opportunities (extract patterns at end)

Token budget: 80-160k tokens across 2-4 sessions (40% per session)


Workflow Phases

Text Only
Phase 1: Discovery (20-30% context)
   ↓
Phase 2: Scope + risk (20-30% context)
   ↓
Phase 3: Implementation (20-40% context)
   ↓
Phase 4: Validate (5-10% context)
   ↓
Phase 5: Learn (5-10% context)

Key principle: Fresh context per phase (prevents degradation)


Phase 1: Research

Goal: Understand the system deeply before planning

Commands:

Bash
Read CLAUDE.md                    # Load repository primer context
/research "[topic]"               # Or /research-multi for 3x speedup
/bundle-save [topic]-research     # Compress findings

Agents used: - code-explorer: Map code structure - doc-explorer: Find relevant documentation - history-explorer: Mine git history

Output: Research bundle (500-1k tokens)

Success criteria: ✅ System understood ✅ Similar implementations found ✅ Constraints identified ✅ Approach recommended

Token budget: 40-60k tokens (20-30%)


Phase 2: Plan

Goal: Specify EVERY change with file:line precision

Commands:

Bash
# New session (fresh context)
Read CLAUDE.md                    # Reload primer context if needed
/bundle-load [topic]-research     # Load research (1k tokens)
/plan [topic]-research            # Create detailed plan
/bundle-save [topic]-plan         # Compress plan

Agents used: - spec-architect: Design detailed specifications - validation-planner: Create test strategy - risk-assessor: Identify and mitigate risks

Output: Plan bundle (1-2k tokens)

Success criteria: ✅ All files specified ✅ Exact changes detailed (file:line, before/after) ✅ Test strategy defined ✅ Implementation order clear ✅ Risks assessed and mitigated

Token budget: 40-60k tokens (20-30%)


Phase 3: Implement

Goal: Execute approved plan mechanically

Commands:

Bash
# New session (fresh context)
Read CLAUDE.md                    # Reload primer context if needed
/bundle-load [topic]-plan         # Load plan (1.5k tokens)
/implement [topic]-plan           # Execute changes

# If context approaches 40% mid-implementation:
# Auto-checkpoint: [topic]-implementation-progress.md

# Resume in next session:
/implement --resume [topic]-implementation-progress

Agents used: - change-executor: Apply changes mechanically - test-generator: Create test cases - continuous-validator: Validate continuously

Output: Working implementation + commit

Success criteria: ✅ All changes applied ✅ All validation passed ✅ Build succeeds ✅ Tests pass ✅ Ready to commit

Token budget: 40-80k tokens (20-40%), may span 2 sessions


Phase 4: Validate

Goal: Comprehensive quality check before deployment

Commands:

Bash
/vibe recent                      # Use --deep for more judges (or --quick for fast inline check)

Agents used: - continuous-validator: Run full validation suite

Output: Validation report

Success criteria: ✅ Syntax checks pass ✅ All tests pass ✅ Coverage meets threshold ✅ No security vulnerabilities ✅ Performance acceptable

Token budget: 10-20k tokens (5-10%)


Phase 5: Learn

Goal: Extract reusable patterns for institutional memory

Commands:

Bash
/retro --quick [topic]            # Extract patterns
/bundle-save [topic]-learning     # Share learnings

Agents used: - (Learning analysis - pattern extraction)

Output: Learning bundle + pattern catalog update

Success criteria: ✅ Pattern documented ✅ Evidence included ✅ Discoverable (tags, index) ✅ Reusable (generalizes)

Token budget: 10-20k tokens (5-10%)


Multi-Session Strategy

Session 1: Research

Bash
Read CLAUDE.md
/research "[topic]"
/bundle-save [topic]-research
# Context: 40-60k (20-30%)

Session 2: Plan

Bash
Read CLAUDE.md
/bundle-load [topic]-research
/plan [topic]-research
/bundle-save [topic]-plan
# Context: 40-60k (20-30%)

Session 3: Implement (Part 1)

Bash
Read CLAUDE.md
/bundle-load [topic]-plan
/implement [topic]-plan
# Context approaches 40% → auto-checkpoint
# Saved: [topic]-implementation-progress.md

Session 4: Implement (Part 2) + Validate + Learn

Bash
/implement --resume [topic]-implementation-progress
# Complete implementation
/vibe recent
/retro --quick [topic]
# Context: 40-70k total (20-35%)

Total: 4 sessions, ~160-200k tokens, sustained quality


When to Use vs Alternatives

Use Complete Cycle when:

✅ Change is complex (multiple files, dependencies) ✅ Risk is significant (critical system, production) ✅ Learning is valuable (new pattern to extract) ✅ Time is available (2-4 hours or multi-day)

Use Quick Fix instead when:

  • Change is simple (1-2 files, low risk)
  • Time is critical (need fix NOW)
  • Pattern is well-known (done 10+ times)

Use Debug Cycle instead when:

  • Problem needs investigation first
  • Root cause unknown
  • Fix approach unclear

Success Metrics

Complete cycle succeeds when: ✅ Research → comprehensive understanding ✅ Plan → detailed specification ✅ Implement → working solution ✅ Validate → all checks pass ✅ Learn → pattern extracted

Quality indicators: - Research was thorough (found constraints early) - Plan was complete (implementation felt mechanical) - Implementation was clean (few surprises) - Validation passed first time (no rework) - Learning is reusable (generalizes to other problems)


Example: Complete Cycle for Redis Caching

Session 1: Research

Bash
Read CLAUDE.md
/research "Redis connection pool exhaustion under burst traffic"
# Agent explores:
# - config/redis.yaml (current settings)
# - app/cache.go (how Redis is used)
# - Similar implementations (other services with pooling)
# - Git history (past connection issues)
#
# Findings:
# - Pool size: 10 (too small for burst)
# - Similar: auth-service increased to 100 (worked)
# - Pattern: Add health checks + circuit breaker
#
/bundle-save redis-pooling-research
# Compressed: 60k → 1.2k tokens

Session 2: Plan

Bash
Read CLAUDE.md
/bundle-load redis-pooling-research
/plan redis-pooling-research
# Agent specifies:
# 1. config/redis.yaml:15 - pool_size: 10 → 100
# 2. app/cache.go:34 - Add pool initialization
# 3. app/cache.go:89 - Add health check endpoint
# 4. tests/cache_test.go:1-50 - Add pool tests
# 5. Validation: go test ./app/... && load test
#
/bundle-save redis-pooling-plan
# Compressed: 50k → 1.5k tokens

Session 3: Implement + Validate

Bash
Read CLAUDE.md
/bundle-load redis-pooling-plan
/implement redis-pooling-plan
# Agent executes:
# ✅ Edit config/redis.yaml:15
# ✅ Edit app/cache.go:34
# ✅ Edit app/cache.go:89
# ✅ Create tests/cache_test.go
# ✅ Validate: All tests pass
#
/vibe recent
# ✅ Syntax: PASSED
# ✅ Tests: 48/48 passed, 89% coverage
# ✅ Security: No issues
# ✅ Performance: P95 500ms → 50ms (10x improvement!)
#
# Commit with full context/solution/learning/impact

Session 4: Learn

Bash
/retro --quick redis-pooling-implementation
# Agent extracts:
# Pattern: Redis Connection Pooling for Burst Traffic
# - When: High-traffic services with burst patterns
# - Solution: Increase pool size, health checks, circuit breaker
# - Evidence: 10x latency improvement, handled 3x traffic
#
/bundle-save redis-pooling-learning
# Pattern added to: .agentops/patterns/implementation/

Total time: 3 hours across 4 sessions Result: Problem solved + Pattern extracted + Institutional memory improved


  • quick-fix: Simplified version (skip research/planning)
  • debug-cycle: Research-focused version (root cause investigation)
  • knowledge-synthesis: Learning-focused version (extract from multiple sources)

Start complete cycle with: Read CLAUDE.md to load primer context → describe your task