Intelligent Do

Intelligent Do

Purpose

Comprehensive intelligent task execution that automatically handles all scenarios: new projects, existing codebases, simple tasks, complex requirements, first-time work, and returning workflows - all with Serena MCP as the persistent context backend.

Core Innovation: One command that adapts to any scenario without configuration, learns from every execution, and gets faster on return visits.

---

Workflow

Step 1: Context Detection Using Serena

Check if project exists in Serena memory:

1. Determine project ID from current working directory:

   • Get current path: Use Bash tool to run pwd
   • Extract project name: Last component of path
   • Sanitize for memory key: Replace special characters with underscores

2. Check Serena for existing project memory:

   • Use Serena tool: list_memories()
   • Search for key: "shannon_project_{project_id}"
   • If key found → RETURNING_WORKFLOW
   • If key not found → FIRST_TIME_WORKFLOW

Duration: < 1 second

---

Step 2a: FIRST_TIME_WORKFLOW

For projects not yet in Serena:

Sub-Step 1: Determine Project Type

Count files in current directory to determine if new or existing project:

# Use Bash tool
find . -type f \( -name "*.py" -o -name "*.js" -o -name "*.tsx" -o -name "*.java" -o -name "*.go" \) | wc -l

• If count < 3: NEW_PROJECT (greenfield)
• If count >= 3: EXISTING_PROJECT (has codebase)

Sub-Step 2: NEW_PROJECT Path

For greenfield projects (empty or minimal files):

1. Assess Task Complexity:

   • Count words in task description
   • Count requirements (lines starting with -)
   • If task has 20+ words OR 3+ requirements → Complex task
   • If task mentions "system", "platform", "integrate" → Complex task

2. For Complex Tasks - Run Spec Analysis:

   Invoke sub-skill:
   @skill spec-analysis
   Specification: {task_description}
   

   • This will analyze complexity (8D score)
   • Create phase plan
   • Save results to Serena automatically
   • Use those results for wave execution

3. For Simple Tasks - Skip Spec:

   • Proceed directly to wave execution
   • No need for formal analysis

4. Execute Task:

   Invoke sub-skill:
   @skill wave-orchestration
   Task: {task_description}
   Spec: {spec_analysis_results if complex}
   

5. Save Project to Serena:

   Use Serena tool:
   write_memory("shannon_project_{project_id}", {
     project_path: "{full_path}",
     created: "{ISO_timestamp}",
     type: "NEW_PROJECT",
     initial_task: "{task_description}",
     complexity: "{simple|complex}",
     spec_id: "{spec_analysis_id if complex}"
   })
   

6. Save Execution Results:

   Use Serena tool:
   write_memory("shannon_execution_{timestamp}", {
     project_id: "{project_id}",
     task: "{task_description}",
     files_created: [list of files from wave results],
     duration_seconds: {duration},
     success: true,
     timestamp: "{ISO_timestamp}"
   })
   

Sub-Step 3: EXISTING_PROJECT Path

For projects with existing codebase:

1. Explore Project Structure:

   • Use Read tool to read: README.md, package.json, pyproject.toml, requirements.txt
   • Use Bash tool to find: find . -name "*.py" | head -20 (sample files)
   • Use Grep tool to search: Look for main entry points, app initialization

2. Detect Tech Stack: From files found:

   • If package.json exists → Node.js/​JavaScript
   • If pyproject.toml or requirements.txt → Python
   • If pom.xml → Java
   • If go.mod → Go
   • If Podfile → iOS/​Swift

   Extract specific frameworks from file contents:

   • package.json dependencies → React, Express, Next.js, etc.
   • requirements.txt → Flask, Django, FastAPI, etc.

3. Detect Validation Gates: From package.json:

   • Look for "scripts" section
   • Extract "test", "build", "lint" commands

   From pyproject.toml:

   • Look for [tool.pytest], [tool.ruff] sections
   • Default gates: pytest, ruff check

4. Research Decision: Check if task mentions libraries NOT in current dependencies:

   • Parse task for library names (common libraries list)
   • Compare against detected dependencies
   • If new library found → Need research

5. If Research Needed:

   For each new library:
   - Use Tavily tool: Search "{library_name} best practices guide"
   - Use Context7 tool: Get library documentation if available
   - Save research to Serena:
     write_memory("shannon_research_{library}_{timestamp}", {
       library: "{library_name}",
       project_id: "{project_id}",
       best_practices: "{tavily_results}",
       documentation: "{context7_results}",
       timestamp: "{ISO_timestamp}"
     })
   

6. Complexity Assessment:

   • Simple task (< 12 words, starts with "create" or "add") → Skip spec
   • Complex task (multiple requirements, "system", "integrate") → Run spec

7. Execute with Context:

   Invoke sub-skill:
   @skill wave-orchestration
   Task: {task_description}
   Project Context:
     - Tech Stack: {detected_tech_stack}
     - Entry Points: {main_files}
     - Validation Gates: {detected_gates}
   Research: {research_results if any}
   Spec: {spec_analysis if complex}
   

8. Save Project Context to Serena:

   Use Serena tool:
   write_memory("shannon_project_{project_id}", {
     project_path: "{full_path}",
     explored: "{ISO_timestamp}",
     type: "EXISTING_PROJECT",
     tech_stack: [list of detected technologies],
     file_count: {file_count},
     entry_points: [list of main files],
     validation_gates: {
       test: "{test_command}",
       build: "{build_command}",
       lint: "{lint_command}"
     }
   })
   

9. Save Execution:

   Use Serena tool:
   write_memory("shannon_execution_{timestamp}", {
     project_id: "{project_id}",
     task: "{task_description}",
     files_created: [list from wave results],
     research_performed: {true|false},
     spec_analysis_ran: {true|false},
     duration_seconds: {duration},
     success: true,
     timestamp: "{ISO_timestamp}"
   })
   

---

Step 2b: RETURNING_WORKFLOW

For projects with Serena context:

1. Load Project Context from Serena:

   Use Serena tool:
   const projectContext = read_memory("shannon_project_{project_id}")
   

   Extract:

   • tech_stack: List of technologies
   • file_count: Previous file count
   • entry_points: Main files
   • validation_gates: Test/​build/​lint commands
   • explored: Last exploration timestamp

2. Check Context Currency:

   • Parse explored timestamp
   • Calculate age in hours: (now - explored) / 3600
   • If age > 24 hours → Context may be stale

3. Detect Changes:

   # Use Bash tool
   current_file_count=$(find . -name "*.py" -o -name "*.js" -o -name "*.tsx" | wc -l)
   

   • Calculate change percentage: abs(current - cached) / cached * 100
   • If change > 5% → Codebase changed

4. Update Context if Changed: If changes detected OR context stale:

   • Re-read key files (README, package.json)
   • Re-count files
   • Update Serena memory:

     Use Serena tool:
     write_memory("shannon_project_{project_id}", {
       ...existing_context,
       file_count: {current_count},
       updated: "{ISO_timestamp}"
     })
     

   • Display: "Codebase changed - updated context"

5. Load Context if Fresh: If no changes:

   • Use cached context as-is
   • Display: "Using cached context (< 1s)"

6. Research Decision (same as first-time):

   • Check if task mentions new libraries
   • Compare against context.tech_stack
   • Research if new libraries found

7. Complexity Decision:

   • Simple task → Direct wave execution
   • Complex task → Run spec-analysis first

8. Execute with Cached Context:

   Invoke sub-skill:
   @skill wave-orchestration
   Task: {task_description}
   Cached Context:
     - Tech Stack: {context.tech_stack}
     - Validation Gates: {context.validation_gates}
   Research: {if performed}
   Spec: {if complex}
   

9. Save Execution to Serena:

   Use Serena tool:
   write_memory("shannon_execution_{timestamp}", {
     project_id: "{project_id}",
     task: "{task_description}",
     used_cache: true,
     context_age_hours: {age},
     files_created: [list],
     timestamp: "{ISO_timestamp}"
   })
   

---

Anti-Rationalization

Rationalization 1: "Simple task, skip Serena check"

COUNTER:

• ❌ NEVER skip Serena check
• ✅ Check takes <100ms via list_memories()
• ✅ Loading cached context saves 2-5 minutes on return
• ✅ Even "simple" tasks benefit from validation gates knowledge

Rule: Always check Serena first. Every time.

Rationalization 2: "Project looks new, skip file exploration"

COUNTER:

• ❌ NEVER assume without checking
• ✅ File count check takes 1 second
• ✅ Wrong assumption = wrong execution context
• ✅ Exploration takes 10-30s, prevents hours of misaligned work

Rule: Check file count. Explore if files exist.

Rationalization 3: "User knows libraries, skip research"

COUNTER:

• ❌ NEVER skip research when detecting new external libraries
• ✅ Research takes 30-60s, provides best practices
• ✅ User knowledge + framework patterns = better quality
• ✅ Prevents hours of debugging wrong implementation patterns

Rule: If external library detected in task but not in project dependencies, research it.

---

Serena Memory Key Patterns

Project Context:

• Key: "shannon_project_{project_id}"
• Data: {project_path, tech_stack, file_count, validation_gates, explored, type}

Execution History:

• Key: "shannon_execution_{timestamp}"
• Data: {project_id, task, files_created, duration, success, timestamp}

Research Results:

• Key: "shannon_research_{library}_{timestamp}"
• Data: {library, best_practices, documentation, timestamp}

Spec Analysis (when complex):

• Key: "spec_analysis_{timestamp}"
• Data: {complexity_score, domain_percentages, phase_plan, ...} (from spec-analysis skill)

---

Examples

Example 1: New Empty Project

User: /​shannon:do "create authentication system using Auth0"
Working Directory: /​tmp/​new-auth-app/

Execution:
1. list_memories() → No "shannon_project_new-auth-app"
2. File count: 0 → NEW_PROJECT
3. Task assessment: "authentication system" (complex) → Run spec-analysis
4. Research: "Auth0" detected → Research Auth0 integration patterns
5. Execute: wave-orchestration with spec + research
6. Save: write_memory("shannon_project_new-auth-app", {...})
7. Save: write_memory("shannon_execution_{timestamp}", {...})

Result: Authentication system created, context saved for next time
Time: 5-8 minutes

Example 2: Existing Project - First Time

User: /​shannon:do "add password reset endpoint"
Working Directory: /​projects/​my-flask-api/

Execution:
1. list_memories() → No "shannon_project_my-flask-api"
2. File count: 45 → EXISTING_PROJECT
3. Explore: Read app.py, requirements.txt → Detect Python/​Flask
4. Validation gates: Found pytest in pyproject.toml
5. Task: Simple ("add endpoint") → Skip spec
6. Research: None (internal feature)
7. Execute: wave-orchestration with Flask context
8. Save project: write_memory("shannon_project_my-flask-api", {tech_stack: ["Python/​Flask"], ...})
9. Save execution: write_memory("shannon_execution_{timestamp}", {...})

Result: Endpoint added, project context cached
Time: 3-5 minutes

Example 3: Returning - Cached Context

User: /​shannon:do "add email verification"
Working Directory: /​projects/​my-flask-api/

Execution:
1. list_memories() → Found "shannon_project_my-flask-api" ✓
2. read_memory("shannon_project_my-flask-api") → Load tech_stack, validation_gates
3. Check age: 2 hours old → Fresh
4. File count: 45 files (same) → No changes
5. Display: "Using cached context (< 1s)"
6. Task: Simple → Skip spec
7. Research: None needed
8. Execute: wave-orchestration with loaded context
9. Save execution: write_memory("shannon_execution_{timestamp}", {...})

Result: Feature added using cached context
Time: 2-3 minutes (vs 5 first time)
Speedup: 2x faster

---

Integration with Shannon CLI

Shannon CLI invokes this skill via Agent SDK:

# In Shannon CLI unified_orchestrator.py
async for msg in self.sdk_client.invoke_skill(
    skill_name='intelligent-do',
    prompt_content=f"Task: {task}"
):
    # Skill handles all intelligence via Serena
    # CLI wraps with V3 features (cost optimization, analytics, dashboard)

Division of Responsibilities:

intelligent-do Skill Provides (Shannon Framework):

• Context detection and loading (Serena MCP)
• Research integration (Tavily, Context7)
• Spec analysis decision (when to run)
• Validation gate detection
• Wave execution coordination

Shannon CLI Provides (Platform Features):

• Cost optimization (model selection)
• Analytics tracking (session history)
• Dashboard streaming (WebSocket events)
• Local CLI interface

Together: Complete intelligent execution platform

---

Testing Requirements

Use functional-testing Skill Patterns:

1. Test with sub-agents using Task tool
2. Validate Serena memory operations:
   • Context saved correctly
   • Context loaded correctly
   • Keys follow naming convention
3. Test scenarios:
   • New empty project
   • Existing project first time
   • Returning to project
   • Research integration
   • Complex task with spec

Evidence Required:

• Serena memories created (list_memories() output)
• Files created in correct locations
• Timing comparison (first vs return)
• Research results when libraries detected

---

Status: Skill template ready for proper implementation following Shannon patterns