A State Machine for When AI Won't Read the Spec

What it does

TaiyiForge is a plugin system that installs into four AI coding terminals and imposes a rigid nine-stage workflow on every change. The engine moves each feature through change, requirement, design, task, dev, test, review, and integration stages, stopping at key milestones until a human explicitly approves. It treats AI agents like junior developers who can write code but cannot greenlight their own designs or skip verification.

The interesting bit

The project enforces a canonical vocabulary of 28 slash commands and six umbrella concepts that behave identically across all four terminals, so switching tools does not mean relearning rituals. It also auto-generates compressed context files to survive long sessions and maintains a ChangeGraph knowledge map to track how edits ripple across the codebase.

Key highlights

• Nine-stage state machine with hard human gates at the change, design, and review phases
• Cross-terminal parity: the same commands and behavior in Claude, Codex, Cursor, and OpenCode
• Mandatory evidence gates: every acceptance criterion must include an executable verification command that exits cleanly
• Automatic token compression via CONTEXT-COMPACT.md and a ChangeGraph dependency tracker for long-running sessions
• Three workflow profiles—full, lite, and nano—scaled to the size of the task

Caveats

• Production-grade SLAs and full internationalization are explicitly listed as not yet ready
• The README is upfront that this is essentially an orchestration layer atop existing standards like Harness, OpenSpec, and GStack rather than a net-new protocol

Verdict

Teams already using Claude, Cursor, or Codex for real features but tired of context loss and skipped design phases should try this. If you prefer ad-hoc prompting and freestyle AI coding, the process constraints will feel like overhead.