Paceflow

PACEflow

[](https://github.com/paceaitian/paceflow/actions/workflows/ci.yml) [](https://github.com/paceaitian/paceflow/releases) [](LICENSE)

**English** | [简体中文](README.zh-CN.md)

**Deterministic workflow enforcement and persistent engineering memory for long-running coding agents.**

PACEflow keeps coding agents on a verifiable engineering workflow during long, stateful tasks. It enforces workflow transitions at tool boundaries, persists task state outside the model context, and restores that state across sessions and context compaction.

PACEflow ships as a **Claude Code plugin** that also installs into **OpenAI Codex CLI** (same plugin directory, same hook scripts, artifact writes via a bundled MCP server on Codex — see [Codex CLI (MVP)](#codex-cli-mvp)). Its underlying reliability model—deterministic gates, persistent artifacts, verification, and review—addresses problems common to long-running coding agents.

Why PACEflow?

Long-running coding-agent tasks introduce workflow continuity failures:

• implementation begins before the plan is approved;
• task state drifts during a long session;
• important context disappears after compaction or a new session;
• verification or review is skipped at closure;
• process instructions written only in prompts remain probabilistic.

PACEflow addresses these failures at the workflow layer. Instead of asking the model to remember every transition, it places deterministic hooks around tool use and session completion, persists the working state, and delegates record-keeping to a dedicated `artifact-writer` agent.

Deterministic gates, not deterministic intelligence

The model still plans, implements, verifies, and reviews the code. A PACEflow gate confirms that a required step occurred in the e