KodoCode：为AI编程工具打造结构化工作流的编排层

KodoCode is an open-source desktop application designed to fill the gap between raw AI coding tools (like Codex CLI and Claude Agent SDK) and over-encapsulated solutions. It retains the power of underlying tools while providing structured workflow orchestration via four intelligent modes (Ask, Plan, Code, Review), enabling developers to work in a more predictable, traceable way aligned with their thinking process.

Named after the Japanese word "古道" (meaning traditional method/right path), KodoCode aims to shift AI-assisted coding from "casual chat" to structured engineering practice. It addresses the dilemma developers face: choosing between raw CLI tools (requiring manual orchestration) or over-encapsulated tools (losing fine-grained control). As an orchestration layer, it doesn't build new AI models but lets developers use familiar Codex/Claude tools in a structured manner.

KodoCode divides development into four distinct modes:

1. Ask: Read-only exploration (e.g., query codebase, feasibility checks) without code changes.
2. Plan: Uses high-reasoning models to create structured, persistent plans (as artifacts) before implementation.
3. Code: Executes approved plans with speed/accuracy-optimized models, using structured context to avoid deviation.
4. Review: Validates changes (tests, regression checks, compliance) post-execution.

KodoCode supports both OpenAI's Codex CLI and Anthropic's Claude Agent SDK. Developers can switch engines per project/session/mode while keeping workflow semantics consistent. Each mode allows independent model parameter config:

• Ask: Lightweight models (e.g., gpt-5.4) for fast responses.
• Plan: Strong reasoning models (e.g., gpt-5.4) for deep planning.
• Code: Code-optimized models with configurable推理强度.
• Review: Independent models (e.g., claude-sonnet-4-6) for validation.

Before entering a mode, KodoCode offers three prompt enhancement levels:

• Minimal: Fixes syntax/spelling while preserving original expression.
• Balanced: Expands intent into well-structured, scoped prompts (default).
• Vibe: Full rewrite to maximize nuanced intent and context. This addresses imprecise initial expressions, letting developers choose between speed and precision.

Unlike Claude Code, Roo Code, or Cline, KodoCode acts as an orchestration layer (not replacement). Key differences:

• Explicit mode separation vs. single dialogue flow.
• Plans as persistent artifacts vs. one-time outputs.
• First-class configuration UI vs. command-line parameters.
• Dual engine support vs. locked single vendor.

KodoCode is a TypeScript-based desktop app with Git worktree isolation (avoids interfering with dev directories). Its roadmap:

• Phase1: Desktop version (completed).
• Phase2: VS Code extension (upcoming).
• Phase3: Self-hosted deployment (Docker, homelab support).

KodoCode empowers developers by structuring AI-assisted coding into exploration, planning, execution, and review stages, using appropriate models/tools for each. It's ideal for those who:

• Want stronger engine support than tools like Cline.
• Care about token efficiency and model cost trade-offs.
• Value prompt refinement.
• Prefer desktop use now with future editor integration. Currently in Phase1 (MIT licensed open source), it's worth trying for developers seeking structured AI coding workflows.