Exec Skill: Autonomous Task Execution

Purpose

Execute tasks autonomously with automatic library discovery, 3-tier functional validation, and atomic git commits. Provides a complete autonomous execution workflow that:

1. Discovers existing libraries (don't reinvent wheels)
2. Validates changes functionally (not just compilation)
3. Commits atomically (only validated code enters git history)
4. Retries on failure (with rollback and research)

Key Innovation: Combines Shannon Framework's multi-agent orchestration (/shannon:wave) with Shannon CLI's validation and git automation for truly autonomous, validated execution.

---

When to Use This Skill

Primary Use Cases

MANDATORY when user says:

• "Execute this task autonomously"
• "Add [feature] with validation and commit"
• "Build [component] using existing libraries"
• "Implement [functionality] and ensure it works"

RECOMMENDED when:

• Task requires code generation AND validation
• Library discovery would save time
• Git automation desired
• Quality gates needed before commit

DO NOT USE when:

• Just analyzing (use spec-analysis instead)
• Just planning (use phase-planning instead)
• Just generating (use wave-orchestration directly)
• Manual validation preferred

---

Workflow

Phase 1: Context Preparation (30-60s)

Action: Invoke /shannon:prime for session setup

@skill context-preservation
  operation: prepare
  task_focused: true

Purpose:

• Discover available skills and tools
• Verify MCP connections (Serena required)
• Load project context
• Restore any previous session state

Output: Session ready, skills discovered, context loaded

Phase 2: Library Discovery (5-30s)

Action: Call Shannon CLI library discoverer

# Execute CLI command to search registries
shannon discover-libs "[feature extracted from task]" --category [ui|auth|networking|data|forms] --json

Purpose:

• Search package registries (npm, PyPI, CocoaPods, Maven, crates.io)
• Rank by quality (stars, maintenance, downloads, license)
• Cache results in Serena MCP (7-day TTL)
• Return top 5 recommendations

Output: List of LibraryRecommendation objects with install commands and reasoning

Example:

{
  "libraries": [
    {
      "name": "next-auth",
      "score": 95.0,
      "why_recommended": "15k+ stars, maintained 4 days ago, MIT license",
      "install_command": "npm install next-auth"
    }
  ]
}

Integration with Phase 4: Library recommendations inject into wave execution context

Phase 3: Task Analysis (Optional, 10-30s)

Action: Invoke /shannon:analyze for complexity assessment

@skill spec-analysis
  specification: {task description with context}
  include_mcps: true

Purpose:

• Assess task complexity (0.00-1.00)
• Identify domains involved
• Recommend execution strategy
• Estimate timeline

Output: Complexity score, domain breakdown, MCP recommendations

Usage Decision:

• Skip for simple tasks (quick execution desired)
• Include for complex tasks (benefits from planning)

Phase 4: Execution Planning (5-15s)

Action: Build execution plan with library context

For Simple Tasks (complexity <0.30):

// Single-step execution
const plan = {
  steps: [{
    number: 1,
    description: task,
    libraries: discoveredLibraries.map(l => l.name),
    validation: {tier1: ["build"], tier2: ["test"], tier3: ["functional"]}
  }]
}

For Complex Tasks (complexity ≥0.30):

// Use sequential-thinking MCP for multi-step planning
@mcp sequential-thinking
  task: Break down {task} into execution steps
  context: Recommended libraries: {libraries}
  format: [{step: "", files: [], validation: {}}]

Output: ExecutionPlan with steps, each having validation criteria

Phase 5: Execution Loop (Per Step)

FOR EACH step in plan:

Step 5a: Execute via /shannon:wave (30-180s per step)

Action: Invoke wave with enhanced prompts

@skill wave-orchestration
  task: {step.description}
  enhanced_context: |
    RECOMMENDED LIBRARIES: {step.libraries}

    CRITICAL INSTRUCTIONS:
    - Use recommended libraries (don't build custom)
    - Make minimal focused changes
    - Follow project conventions
    - Include error handling

    [Full enhanced prompts from Shannon CLI]

Purpose:

• Wave analyzes step complexity
• Spawns appropriate number of agents
• Agents execute in parallel
• Files created/modified via Write/Edit tools

Output: Files modified/created

Step 5b: Parse Wave Results (Instant)

Action: Extract file changes from wave execution

const filesChanged = parseWaveMessages(messages).filter(m =>
  m.type === 'ToolUseBlock' && ['Write', 'Edit'].includes(m.name)
).map(m => m.input.file_path)

Output: List of files that were modified

Step 5c: Validate Changes (10-120s)

Action: Call Shannon CLI validator

# Tier 1: Static validation
shannon validate --tier 1 --json

# If Tier 1 passes, Tier 2: Tests
shannon validate --tier 2 --json

# If Tier 2 passes, Tier 3: Functional
shannon validate --tier 3 --json

Validation Tiers:

Tier 1 - Static (~10s):

• Build: npm run build or cargo build or xcodebuild
• Type Check: tsc --noEmit (if TypeScript) or mypy . (if Python)
• Lint: eslint or ruff check or swiftlint

Tier 2 - Tests (~1-5min):

• Unit/Integration: npm test, pytest tests/, xcodebuild test
• Parse results for pass/fail count
• Verify: All tests pass (no regressions)

Tier 3 - Functional (~2-10min):

• Node.js: npm run dev, wait for server, curl health endpoint, verify 200 OK
• Python: uvicorn main:app, curl endpoint, verify response correct
• iOS: Boot simulator, run UI tests, verify element visible

Output: ValidationResult(tier1_passed, tier2_passed, tier3_passed, all_passed)

Step 5d: Decision Point

IF validation.all_passed:

# Commit validated changes
shannon git-commit \
  --step "{step.description}" \
  --files "{filesChanged.join(',')}" \
  --validation-json "{validation.toJSON()}"

Creates commit:

feat: {step description}

VALIDATION:
- Build: PASS
- Tests: 12/12 PASS
- Functional: Feature works in browser

Files: package.json, src/auth/login.tsx, ...

Result: Proceed to next step

ELSE validation failed:

# Rollback changes
shannon git-rollback

# Research solution (if research enabled)
@mcp firecrawl
  search: "{validation.failures[0]} solution"

# Replan step with research findings

# Retry (attempt 2 of 3)

Result: Rollback → Research → Retry (max 3 attempts)

IF all attempts fail:

• Document failure reason
• Return partial results
• Recommend manual intervention

Phase 6: Execution Report

Action: Generate comprehensive report

const report = {
  success: true,
  task: taskDescription,
  steps_completed: completedSteps.length,
  steps_total: plan.steps.length,
  commits_created: commits.map(c => c.hash),
  branch_name: branchName,
  duration_seconds: totalDuration,
  libraries_used: uniqueLibraries,
  validations: {
    passed: validationsPassed,
    failed: validationsFailed
  }
}

Display:

✅ AUTONOMOUS EXECUTION COMPLETE

Task: {taskDescription}
Branch: {branchName}
Commits: {commits.length}
Duration: {duration}s
Libraries: {libraries.join(', ')}

Ready for: git push origin {branchName}

Save to Serena:

@mcp serena
  write_memory: exec_result_{timestamp}
  content: {report.toJSON()}

---

Integration Points

Shannon Framework Skills Used

1. /shannon:prime - Phase 1 (context preparation)
2. /shannon:analyze - Phase 3 (optional complexity analysis)
3. /shannon:wave - Phase 5 (code generation per step)

Shannon CLI Modules Used

1. shannon discover-libs - Phase 2 (library search)
2. shannon validate - Phase 5c (3-tier validation)
3. shannon git-commit - Phase 5d (atomic commits)
4. shannon git-rollback - Phase 5d (rollback on failure)

MCPs Used

1. Serena - Required (cache libraries, save execution state)
2. Sequential-thinking - Optional (multi-step planning)
3. Firecrawl - Optional (research on failures)

---

Success Criteria

Technical:

• All 6 phases execute without errors
• Library discovery returns relevant packages
• Validation runs all 3 tiers correctly
• Only validated code commits to git
• Failed validations trigger rollback
• Retry logic functions (max 3 attempts)

Functional:

• Generated code actually works (not just compiles)
• Libraries are used (not custom implementations)
• Git history is clean (only validated commits)
• Execution is truly autonomous (no manual intervention)

User Experience:

• Real-time progress visible
• Clear success/failure messages
• Helpful error reporting
• Results save to Serena for reference

---

Common Pitfalls

Pitfall 1: Skipping Library Discovery

Symptom: Task builds custom authentication instead of using next-auth Cause: Phase 2 not executed or results not used Fix: Always run library discovery, inject results into wave context

Pitfall 2: Committing Unvalidated Code

Symptom: Commit created but code doesn't build Cause: Skipped validation or ignored failures Fix: Enforce all_passed check before commit

Pitfall 3: Not Rolling Back on Failure

Symptom: Git history has broken commits Cause: Validation failed but rollback not executed Fix: Always rollback before retry

Pitfall 4: Infinite Retry Loops

Symptom: Same validation error repeated indefinitely Cause: No max iterations or no error detection Fix: Enforce max_iterations=3, detect repeated failures

---

Examples

Example 1: Simple File Creation

Input:

@skill exec
  task: "create hello.py that prints hello world"

Execution:

• Phase 1: Context (skip for simple)
• Phase 2: Libraries (none needed)
• Phase 3: Analysis (skip for simple)
• Phase 4: Plan (single step)
• Phase 5: Wave creates hello.py → Validate → Commit
• Phase 6: Report success

Duration: ~20-30s Result: 1 commit with hello.py

Example 2: Authentication Feature

Input:

@skill exec
  task: "add authentication to Next.js app"
  interactive: true

Execution:

• Phase 1: Prime context
• Phase 2: Discover next-auth library (95/100 score)
• Phase 3: Analyze (complexity 0.42 - MODERATE)
• Phase 4: Plan:
  1. Install next-auth
  2. Configure NextAuth.js
  3. Create login page
  4. Add protected routes
  5. Test authentication
• Phase 5: Execute each step with wave → Validate → Commit (5 commits)
• Phase 6: Report (5 commits, authentication working)

Duration: ~5-8 minutes Result: 5 commits, working authentication

Example 3: Multi-Stack Feature

Input:

@skill exec
  task: "add user profile feature: FastAPI endpoint + React component"

Execution:

• Phase 2: Discover Pydantic (backend), shadcn/ui (frontend)
• Phase 4: Plan:
  1. Create Pydantic model
  2. Add FastAPI endpoint
  3. Create React ProfileCard component
  4. Connect frontend to API
• Phase 5: Each step executed, validated, committed
• Phase 6: Report (4 commits, full-stack feature working)

Duration: ~8-12 minutes Result: 4 commits, working profile feature

---

Related Skills

• wave-orchestration: Used in Phase 5 for code generation
• spec-analysis: Optional in Phase 3 for complexity assessment
• context-preservation: Automatic checkpoints between steps
• functional-testing: Philosophy enforced in Tier 3 validation

---

Related Commands

• /shannon:prime: Invoked in Phase 1
• /shannon:analyze: Optional in Phase 3
• /shannon:wave: Invoked per step in Phase 5
• shannon CLI commands: discover-libs, validate, git-commit, git-rollback

---

Version History

v5.1.0 (2025-11-15):

• Initial release
• Integration with Shannon CLI V3.5.0 executor modules
• Support for Python, Node.js, React, iOS validation
• Multi-registry library discovery (npm, PyPI, CocoaPods, Maven, crates.io)
• 3-tier validation with auto-detected test infrastructure
• Atomic git commits with validation messages
• Retry logic with rollback (max 3 attempts)

---

Requirements

Shannon CLI: Version 3.5.0 or higher (provides executor modules) MCPs: Serena (required), Sequential (recommended) Tools: Bash (for calling CLI commands), Read/Write (for file operations)

Installation:

# Ensure Shannon CLI installed
pip install shannon-cli>=3.5.0

# Verify exec modules available
shannon --version  # Should show 3.5.0+

---

Status: Production-ready for Shannon Framework V5.1.0