Skip to main content

MoE Task Execution Workflow

MoE Task Execution Workflow

Overview

This skill defines the standard workflow for task discovery, agent selection, invocation creation, and coordinated execution using Mixture-of-Experts (MoE) orchestration.

When to Use This Skill

Use this skill when:

  • Starting a new session and need to orient quickly
  • Popping the next task from the tasklist to execute
  • Coordinating multiple agents for complex tasks
  • Ensuring consistent task execution methodology
  • Updating tasklists with explicit agent invocations

Don't use this skill when:

  • Simple single-step tasks (just execute directly)
  • Non-agentic work (manual file edits)
  • Research/exploration without task execution

The 5-Step Methodology

Step 1: Find Project Plan & Checkpoints

# Find pilot plan and related documents
Glob("**/PILOT*.md")
Glob("**/CHECKPOINT*.md")

# Read latest checkpoint
Read("/path/to/docs/CHECKPOINT-YYYY-MM-DD.md")

# Read the active tasklist
Read("internal/project/plans/PILOT-PARALLEL-EXECUTION-PLAN.md")

Run multiple /cxq searches to build complete context:

# Search 1: Find the documents
/cxq --recall "pilot project-plan tasklist"

# Search 2: Get current status
/cxq --recent 200

# Search 3: Find recent work (use date constraints)
/cxq --since "1 day ago" --limit 50

# Search 4: Identify next steps
/cxq --decisions --limit 20

Step 3: Agent Discovery with /which

Find the right agents for your tasks:

# Discover agents for specific task types
/which moe-agents orchestrator project-organizer

# Get agent recommendations
/which analyze pilot tasklist
/which implement security hardening
/which create documentation

Agent Type Reference:

Task TypePrimary AgentSupporting Agents
Securitysecurity-specialisttesting-specialist
Frontendfrontend-react-typescript-expert-
Backendsenior-architectdatabase-architect
Testingtesting-specialist-
DevOpsdevops-engineercloud-architect
Documentationcodi-documentation-writer-
Orchestrationorchestratorproject-organizer

Step 4: Create Invocations & Update Tasklist

For each task, create explicit invocations and add to the tasklist:

Format:

- [ ] D.1.1: Rate limiting middleware
- **Agent:** `Task(subagent_type="security-specialist", prompt="Create api/middleware/rate_limit.py with slowapi rate limiting")`

Section-Level Invocation:

### D.1: P0 Security

**Agent Invocation:**
\`\`\`bash
Task(subagent_type="security-specialist", prompt="Implement Track D.1 P0 Security: (1) rate limiting; (2) security headers; (3) login protection")
\`\`\`

Step 5: Execute with MoE Orchestration

Option A: Full Track Orchestration

Task(subagent_type="orchestrator", prompt="Execute Track D.1 tasks per PILOT-PARALLEL-EXECUTION-PLAN.md using security-specialist agents")

Option B: Specific Task Execution

# Copy invocation from tasklist
Task(subagent_type="security-specialist", prompt="Implement Track D.1 P0 Security...")

Option C: MoE Consensus (Multiple Agents)

# Run orchestrator + project-organizer in parallel
Task(subagent_type="orchestrator", prompt="Coordinate Track D implementation")
Task(subagent_type="project-organizer", prompt="Verify file organization for Track D deliverables")

Quick Reference Commands

# Start session
/orient

# Find agents
/which <task description>

# Execute task
Task(subagent_type="<agent>", prompt="<from tasklist>")

# Mark complete
# Edit tasklist: [ ] → [x]

Invocation Templates

Security Tasks

Task(subagent_type="security-specialist", prompt="Implement security hardening: rate limiting, security headers, input validation, audit logging per PILOT-PARALLEL-EXECUTION-PLAN.md Track D")

Testing Tasks

Task(subagent_type="testing-specialist", prompt="Implement E2E testing: user flow tests, security tests, platform tests per PILOT-PARALLEL-EXECUTION-PLAN.md Track E")

Documentation Tasks

/agent codi-documentation-writer "Create user documentation: installation guide, user guide, FAQ per PILOT-PARALLEL-EXECUTION-PLAN.md Track F"

Frontend Tasks

Task(subagent_type="frontend-react-typescript-expert", prompt="Implement entitlement display: product access badges, subdomain links, upgrade CTAs per PILOT-PARALLEL-EXECUTION-PLAN.md B.3.4")

DevOps Tasks

Task(subagent_type="devops-engineer", prompt="Deploy infrastructure: GKE deployment, DNS configuration, SSL certificates per PILOT-PARALLEL-EXECUTION-PLAN.md Track C")

Integration with /orient

The /orient command uses this skill for the enhanced 5-step methodology:

# Full orientation with MoE workflow
/orient

# Deep orientation for specific topic
/orient --deep "security hardening"

# Quick status check
/orient --quick

Best Practices

  1. Always create invocations first - Before executing, add the invocation to the tasklist
  2. Use section-level invocations - Group related tasks under a single orchestrated invocation
  3. Date-constrain searches - Use --since and --until for relevant context
  4. Mark tasks immediately - Update [ ][x] as soon as task completes
  5. Update checkpoints - Create session checkpoint after significant work

Success Output

When this skill completes successfully, output:

✅ SKILL COMPLETE: moe-task-execution

Completed:
- [x] Project plan and checkpoints found and read
- [x] Multi-query context search executed (4+ searches)
- [x] Agent discovery completed with /which
- [x] Invocations created and added to tasklist
- [x] Tasks executed with appropriate agents
- [x] Tasklist updated with completed tasks

MoE Execution Summary:
- Plans found: [count]
- Context queries executed: [count]
- Agents discovered: [count]
- Invocations created: [count]
- Tasks completed: [count]

Outputs:
- Updated tasklist with invocations
- Task completion markers ([x])
- Session context captured

Completion Checklist

Before marking this skill as complete, verify:

  • Project plan found (PILOT*.md or project-plan.md)
  • Latest checkpoint read and context restored
  • At least 4 context searches executed (/cxq)
  • Agent discovery performed for each task type (/which)
  • Invocations created in correct format (Task() or /agent)
  • Invocations added to tasklist before execution
  • Tasks executed successfully with recommended agents
  • Tasklist updated with completion markers [x]
  • Session context captured with /cx

Failure Indicators

This skill has FAILED if:

  • ❌ Project plan not found or not read
  • ❌ Context search skipped or insufficient (<4 queries)
  • ❌ Agent discovery not performed (/which skipped)
  • ❌ Invocations not created before execution
  • ❌ Wrong agent used for task (didn't follow /which recommendations)
  • ❌ Tasklist not updated with invocations or completion
  • ❌ Tasks executed without proper MoE coordination
  • ❌ Context lost between sessions (no /cx)

When NOT to Use

Do NOT use this skill when:

  • Simple single-step tasks (just execute directly)
  • No tasklist exists (exploratory work)
  • Task doesn't require agent coordination (manual file edits)
  • Research/analysis without execution
  • Quick fixes or hotfixes (overhead not worth it)
  • Emergency debugging (skip coordination overhead)
  • Working outside CODITECT framework

Use instead:

  • Direct tool calls for simple tasks
  • Manual execution for quick fixes
  • Standard workflow for non-agentic work
  • Research skills for exploration
  • Debugging skills for emergency fixes

Anti-Patterns (Avoid)

Anti-PatternProblemSolution
Skip context searchStart blind, repeat past workAlways run multi-query /cxq first
Execute without /whichUse wrong agent, inefficientAlways discover agents before execution
No invocation in tasklistCan't reproduce, track, or auditAlways document invocation before running
Wrong agent usedTask fails or poor qualityFollow /which recommendations
Skip orientationLost context, inefficient sessionAlways run /orient at session start
Single context queryIncomplete context, missed infoRun 4+ queries from different angles
No completion markersCan't track progressUpdate tasklist [x] immediately after completion
Execute before creating invocationNo audit trail, can't reproduceInvocation BEFORE execution, always

Principles

This skill embodies CODITECT principles:

  • #1 Recycle → Extend → Re-Use → Create - Reuses past context, prevents duplicate work
  • #2 First Principles - Understands WHY MoE coordination improves quality
  • #3 Keep It Simple - 5 clear steps, repeatable methodology
  • #4 Separation of Concerns - Discovery separate from execution
  • #5 Eliminate Ambiguity - Explicit invocations, clear agent assignments
  • #6 Clear, Understandable, Explainable - Each step documented in tasklist
  • #8 No Assumptions - Always discover, never assume right agent
  • #9 Measurable & Traceable - Completion markers, invocation history
  • #10 Provide Context - Multi-query search ensures complete context

Full Standard: CODITECT-STANDARD-AUTOMATION.md

ComponentPurpose
/orientSession orientation with 5-step methodology
/whichAgent discovery and recommendations
/agentCODITECT agent invocation
orchestratorMulti-agent coordination
project-organizerFile/directory organization
session-analysisSession indexing and tracking

Created: December 31, 2025 Author: CODITECT Team Version: 1.0.0 Updated: 2026-01-04