mapcoder

MapCoder: Multi-Agent Code Generation Pipeline

You are orchestrating the MapCoder pipeline, a multi-agent system that replicates the human programming cycle through four specialized agents:

1. Retrieval Agent - Generates similar problems from knowledge
2. Planning Agent - Creates step-by-step algorithmic plans
3. Coding Agent - Translates plans into executable code
4. Debugging Agent - Fixes bugs using plans and test feedback

Argument Parsing

Parse $ARGUMENTS to extract:

• --lang <language>: Target programming language (default: Python)
• --sandbox: Use Docker sandbox for code execution (safer)
• Remaining text: The problem description or file path

Example inputs:

• /mapcoder implement binary search → Python, direct execution
• /mapcoder --lang javascript implement binary search → JavaScript
• /mapcoder --sandbox --lang rust implement linked list → Rust, sandboxed

Pipeline Execution

Step 1: Problem Analysis

First, analyze the input:

• If it's a file path, read the file to get the problem description
• Extract any sample test cases from the problem
• Identify the target language from arguments (default: Python)

Step 2: Retrieval Phase

Use the Task tool to spawn the retrieval agent:

Spawn a retrieval agent to generate 3-5 similar problems for:
[problem description]

The agent should return:
- Similar problem descriptions
- Solution patterns/approaches used
- Key algorithmic concepts

Use subagent_type: "general-purpose" with the retrieval-agent.md system prompt.

Step 3: Planning Phase

Use the Task tool to spawn the planning agent with the retrieved examples:

Spawn a planning agent to create step-by-step plans for:
[problem description]

Using these similar problems as reference:
[retrieved examples]

Generate 2-3 alternative algorithmic plans.

Step 4: Coding Phase

Use the Task tool to spawn the coding agent:

Spawn a coding agent to implement the solution in [language]:
[problem description]

Following this plan:
[selected plan]

Test against these sample cases:
[test cases]

If --sandbox flag is set, instruct the agent to use scripts/sandbox-runner.sh instead of direct execution.

Step 5: Validation Loop

After coding completes:

If tests pass: Return the solution with explanation.

If tests fail: Enter debugging loop (max 3 iterations):

Spawn a debugging agent to fix the failing code:

Original problem: [problem]
Current code: [code]
Error output: [errors]
Original plan: [plan]

Identify the bug and generate corrected code.

Step 6: Solution Output

When successful, output using the solution template:

## Solution

**Language**: [language]
**Status**: [Passed/Failed after N attempts]

### Code
[final code]

### Explanation
[step-by-step explanation of the approach]

### Test Results
[test output]

Adaptive Traversal

The pipeline supports adaptive routing:

• If retrieval produces highly relevant examples → use single best plan
• If planning produces uncertain results → try multiple plans in parallel
• If debugging fails 3 times → try alternative plan from Step 3
• If all plans exhausted → report failure with analysis

Configuration

• MAX_DEBUG_ITERATIONS: 3
• MAX_PLAN_ATTEMPTS: 3
• DEFAULT_LANGUAGE: Python

Error Handling

• If any agent fails to spawn, report the error and suggest manual intervention
• If code execution times out, treat as test failure and enter debugging
• If problem is ambiguous, ask user for clarification before proceeding