Plandex

Plandex maintains a stateful conversation context across multiple code generation steps, allowing developers to iteratively refine complex implementations without losing prior context. The system uses a plan-based architecture where each step builds on previous outputs, with automatic context summarization to manage token limits while preserving semantic continuity across long development sessions.

Plandex analyzes the existing codebase structure and automatically injects relevant file contents and context into generation prompts, enabling the AI to generate code that respects existing patterns, dependencies, and architectural conventions. The system uses file indexing and semantic matching to determine which files are relevant to a task without requiring manual context specification.

Plandex breaks down complex development tasks into discrete, sequenced steps using a plan-based reasoning approach. Each step is tracked with status (pending, in-progress, completed, failed), and developers can review, modify, or re-execute individual steps. The system maintains a structured plan representation that persists across sessions, enabling long-running projects to be paused and resumed without losing task structure.

Plandex operates as a CLI-first tool with real-time streaming output of generated code and execution logs directly to the terminal. The interface supports interactive prompts, inline code review, and immediate execution feedback without context-switching to web browsers or IDEs. The streaming architecture allows developers to see generation progress and interrupt long-running tasks mid-execution.

Plandex abstracts away provider-specific API differences through a unified interface that supports OpenAI, Anthropic, and local Ollama models. The system translates high-level generation requests into provider-specific API calls, handling differences in token counting, context window limits, and function-calling conventions. Developers can switch providers or models without changing task definitions or prompts.

Plandex integrates with Git to track all AI-generated changes as commits, enabling developers to review diffs, revert changes, and maintain a clear audit trail of AI modifications. The system uses diff-based code modification rather than full file replacement, preserving manual edits and minimizing merge conflicts. Changes are staged in Git before application, allowing selective acceptance or rejection.

Plandex can execute generated code and feed error messages, test failures, and execution logs back into the generation loop for automatic refinement. The system detects compilation errors, runtime exceptions, and test failures, then re-prompts the LLM with error context to generate fixes. This creates a feedback loop where the AI learns from execution failures and iteratively improves code until it passes.

Plandex automatically determines which files are relevant to a development task using semantic analysis and dependency tracking, then includes only relevant files in the generation context. The system uses heuristics based on import statements, file naming patterns, and code structure to avoid overwhelming the LLM with irrelevant context. Developers can manually override file selection or exclude specific files from context.