Agent Reasoning Loop

The Four-Phase Execution Loop

Rather than generating speculative code strings in a single pass, December enforces a deterministic four-phase loop:

1. Context Gathering & Inspection

Before touching source code, the agent inspects project structure:

• Reads package.json, tsconfig.json, or relevant build manifests
• Searches symbols and dependencies using regex and AST indexing
• Reads git status and existing test suites

2. Implementation Planning

For multi-step edits, the agent breaks the problem down into verifiable atomic steps:

• Outlines exact modules that will be modified
• Identifies potential breaking changes and edge cases
• Formulates verification tests

3. Tool Execution & Mutation

The agent applies code changes using atomic file operations and runs shell commands:

• Edit files with precise line-range diff replacements
• Install missing dependencies
• Format code using configured linters (prettier, eslint)

4. Diagnostic Feedback & Self-Correction

If a test or build fails, the error output is piped directly into the agent context:

• Compiler error diagnostics are analyzed
• Stack traces pinpoint exact failure sites
• The agent iterates and corrects itself until the test suite passes