Workflow Orchestrator

Feature development: Contract → Builder Team → Observer monitors → Security gate → Ship

End session:

1. Observer final report
2. Security sweep: gitleaks detect --source .
3. Update PROJECT_CONTEXT.md
4. Log costs to costs/daily-YYYY-MM-DD.json

• Session Management: "start day", "begin session", "what's the status", "end day", "wrap up", "done for today"
• Feature Development: "new feature", "build", "implement"
• Dual-Team: "dual team", "observer team", "builder team", "spawn observers", "devil's advocate"
• Debugging: "debug", "investigate", "why is this broken"
• Research: "research", "evaluate", "should we use"

DUAL-TEAM ARCHITECTURE

Every session runs two concurrent teams under the Orchestrator:

              ORCHESTRATOR
              ┌────┴────┐
        BUILDER TEAM   OBSERVER TEAM
        (ships fast)   (watches quality)
        ├ Lead Builder  ├ Code Quality (haiku)
        ├ Builder(s)    └ Architecture (sonnet)
        └ Devil's Adv.    └─ (Devil's Advocate)

Observer team is non-negotiable. It always runs, even for "small" changes.

• Builders optimize for velocity — ship features via worktrees or subagents
• Observers optimize for correctness — detect drift, debt, gaps, scope creep
• Devil's advocate on each team prevents groupthink and blind spots

Observers write findings to .claude/OBSERVER_*.md with severity levels:

• 🔴 BLOCKER — stop work, fix immediately
• 🟡 WARNING — fix before merge or log
• 🔵 INFO — backlog

Deep dive: See reference/dual-team-architecture.md

START DAY

Pre-Flight Checks

git status --short | head -5
[ -f package.json ] && [ ! -d node_modules ] && echo "Run npm install"
[ -f requirements.txt ] && [ ! -d .venv ] && echo "Run pip install"
[ -f .env.example ] && [ ! -f .env ] && echo "Copy .env.example to .env"

Context Scan (Mandatory)

pwd && git status && git log --oneline -5
cat PROJECT_CONTEXT.md 2>/dev/null || echo "No context file"
cat CLAUDE.md 2>/dev/null
cat PLANNING.md 2>/dev/null

Observer Team Spawn (Automatic)

Spawn Observer team concurrent with standup:

• Code Quality Observer (haiku) — tech debt, test gaps, imports
• Architecture Observer (sonnet) — contract drift, scope creep, design violations

# Detect Native Agent Teams support
if [ "${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS}" = "1" ]; then
  echo "Native Agent Teams enabled — using DAG task system"
  # Use TeamCreate + SendMessage for Observer coordination
else
  echo "Native Agent Teams not enabled — falling back to manual worktree spawn"
  # Use subagent-teams-skill Task tool pattern
fi

Sprint Plan

Include Observer BLOCKER gate in sprint plan — no phase transitions if active BLOCKERs in .claude/OBSERVER_ALERTS.md.

Output Format

## Session Start: [PROJECT_NAME]
### Completed (Last Session)
- [x] Task 1
### In Progress
| Task | Branch/Worktree | Status |
|------|-----------------|--------|
| API endpoint | feature/api | 70% |
### Observer Status
- Team: ACTIVE (Code Quality + Architecture)
- Active blockers: 0
### Today's Priority Queue
1. [BUILDER] Feature implementation
2. [OBSERVER] Continuous monitoring
### Cost Context
- Today: $0.00 | MTD: $12.34 | Budget: $100

Deep dive: See reference/start-day-protocol.md

RESEARCH PHASE

Trigger: Before ANY feature development involving new frameworks, APIs, or architectural decisions.

Scan → Evaluate → Decide

1. Check existing solutions in your repos and MCP cookbook
2. Use research-skill checklist for framework selection
3. Cost projection before building
4. Create RESEARCH.md → FINDINGS.md with GO/NO-GO recommendation

⛔ Gate: Human checkpoint required before proceeding

Deep dive: See reference/research-workflow.md

FEATURE DEVELOPMENT

Phase 0: CONTRACT DEFINITION (Before Code)

## Feature Contract: [NAME]
### Endpoints / Interfaces
- POST /api/widgets → { id, name, created_at }
### Scope Boundaries
- IN SCOPE: [list]
- OUT OF SCOPE: [list]
### Observer Checkpoints
- [ ] Architecture Observer approves contract
- [ ] Code Quality Observer runs after each merge

⛔ Gate: Architecture Observer must approve contract before Phase 1

Phase 1: BUILDER TEAM SPAWN

# Option A: Worktree sessions (long tasks, 30+ min)
git worktree add -b feature/api ~/tmp/worktrees/$(basename $(pwd))/api
# Option B: Native Agent Teams (short tasks, <20 min)
# Use TeamCreate + Task tool subagents
# Option C: Hybrid — worktrees for Builders, Task tool for Observers

Each builder gets a WORKTREE_TASK.md with:

• Scope boundary (what they CAN and CANNOT touch)
• Contract reference
• Devil's advocate mandate (for one builder per cycle)

Phase 2: OBSERVER MONITORING (Concurrent)

Observers run parallel to builders on a 5-10 minute loop:

1. Pull latest from builder branches
2. Run 7 drift detection patterns (see reference/observer-patterns.md)
3. Write findings to .claude/OBSERVER_QUALITY.md and .claude/OBSERVER_ARCH.md
4. Escalate BLOCKERs to .claude/OBSERVER_ALERTS.md

Phase 3: SECURITY + QUALITY GATE

# Security scans
semgrep --config auto .
gitleaks detect --source .
npm audit --audit-level=critical || pip-audit
pytest --cov=src || npm test -- --coverage

⛔ Gate: ALL must pass + no active Observer BLOCKERs

gate = (
    sast_clean AND
    secrets_found == 0 AND
    critical_vulns == 0 AND
    test_coverage >= 80 AND
    observer_blockers == 0  # NEW: Observer gate
)

Phase 4: SHIP

git diff main...HEAD
git add . && git commit -m "feat: [description]"
git push
echo '{"feature": "X", "cost": 1.23}' >> costs/by-feature.jsonl

Deep dive: See reference/feature-development.md

DEBUG MODE

Trigger: When standard troubleshooting fails or issue is complex.

Evidence → Hypothesize → Test → Verify

1. Evidence gathering — exact error, reproduction steps, expected vs actual
2. Hypothesis formation — 3+ hypotheses with evidence for each
3. Systematic testing — one variable at a time
4. Verification — root cause confirmed before committing fix

Critical Rules

• ❌ NO DRIVE-BY FIXES — explain WHY before committing
• ❌ NO GUESSING — verify everything
• ✅ Use all tools: MCP servers, web search, extended thinking
• ✅ One variable at a time

Deep dive: See reference/debug-methodology.md

END DAY

Phase 1: Observer Final Report (Before Security Sweep)

Observers generate final reports:

• Summary of all findings (resolved + open)
• Metrics: debt items, test coverage delta, contract compliance
• Write to .claude/OBSERVER_QUALITY.md and .claude/OBSERVER_ARCH.md

⛔ Gate: Review Observer report before proceeding

Phase 2: Security Sweep (Mandatory)

gitleaks detect --source . --verbose
git log -p | grep -E "(password|secret|api.?key|token)" || echo "Clean"
npm audit --audit-level=critical 2>/dev/null || pip-audit 2>/dev/null

⛔ Gate: ALL must pass before any commits

Phase 3: Context + Metrics + Cleanup

## PROJECT_CONTEXT.md Update
### Completed This Session
- [x] Feature X
### Observer Summary
- Blockers resolved: N | Warnings logged: N | Debt delta: +/-N
### Tomorrow's Priorities
1. Next task

Archive observer reports to .claude/OBSERVER_HISTORY/, reset for next session.

# Cost tracking
echo '{"total": 0.47}' >> costs/daily-$(date +%Y-%m-%d).json
# Portfolio metrics
git diff --stat $(git log --since="today 00:00" --format="%H" | tail -1)..HEAD 2>/dev/null

Deep dive: See reference/end-day-protocol.md

COST GATE

Pre-Flight Budget Check

COST_FILE=~/.claude/daily-cost.json
SPENT=$(jq '.spent' "$COST_FILE" 2>/dev/null || echo 0)
BUDGET=$(jq '.budget_monthly' "$COST_FILE" 2>/dev/null || echo 100)
echo "MTD: \$$SPENT / \$$BUDGET"

Deep dive: See reference/cost-tracking.md

ROLLBACK / RECOVERY

• Stash current work → find last known good → selective rollback or full revert → verify tests → investigate root cause
• Use debug-like-expert-skill for root cause analysis

Deep dive: See reference/rollback-recovery.md

SKILL INVOCATION QUICK REFERENCE

Native Agent Teams (Experimental)

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
claude --worktree  # auto-creates worktree branch

Features: DAG task system, peer-to-peer messaging, TeammateIdle/TaskCompleted hooks, shared task lists, isolation: "worktree" for subagents.

Deep dive: See reference/agent-routing.md for complete 70+ agent catalog

PROGRESS RENDERING

Use TaskCreate with activeForm for live UI spinners:

TaskCreate({ subject: "Plan architecture", activeForm: "Planning architecture" })
TaskUpdate({ taskId: "1", status: "in_progress" })  // → live spinner
TaskUpdate({ taskId: "1", status: "completed" })     // → checkmark

Use addBlockedBy for phase sequencing. Markdown tables for summaries.

CLAUDE CODE COMMANDS

GTME PERSPECTIVE

This dual-team workflow demonstrates advanced GTME capabilities:

1. Process Engineering — Quality control built into lifecycle, not bolted on
2. Parallel Orchestration — Managing concurrent teams with different objectives
3. Adversarial Thinking — Devil's advocate prevents "happy path only" trap
4. Cost Awareness — Observers use cheaper models; unit economics thinking
5. Measurable Output — Observer reports quantify debt, coverage, compliance

Portfolio Value: Produces auditable artifacts demonstrating engineering discipline + process thinking + cost optimization.

Emit Outcome Sidecar

As the final step, write to ~/.claude/skill-analytics/last-outcome-workflow-orchestrator.json:

{"ts":"[UTC ISO8601]","skill":"workflow-orchestrator","version":"2.0.0","variant":"default",
 "status":"[success|partial|error]","runtime_ms":[estimated ms from start],
 "metrics":{"workflows_run":[n],"stages_completed":[n],"observers_spawned":[n]},
 "error":null,"session_id":"[YYYY-MM-DD]"}

Use status "partial" if some stages failed but results were produced. Use "error" only if no output was generated.