SAIPEN

Continuation protocol for AI coding agents. SAIPEN keeps project memory in plain markdown, so a cold agent with no chat history runs /saipen continue, reads STATE, BOARD, and next_action, and resumes work in under a minute -- no rebriefing, any vendor, any day.

One command. Zero amnesia.

v7.72.0 | Spec | Guide | RFC | Style | UI | Conformance | plain markdown | zero deps | MIT

User  ->  /saipen continue
Agent ->  reads STATE ("What do I do right now?")
Agent ->  reads BOARD ("What task am I picking up?")
Agent ->  reads next_action (executes command)
Agent ->  Works.

Project State > Model Memory

Memory lives in the project, not in a model's head. Project -> Memory -> LLM becomes Project -> SAIPEN State -> LLM.

Key Protocol Logic & Guarantees

• Core State Machine: INIT → PLAN → SCOUT → BUILD → VERIFY → REVIEW → SHIP → DONE | BLOCKED
• Zero-Prompt Autonomy: No open to-dos left? Auto-transitions HUNT (scan bugs) → ADD (evolve features) → HUNT loop. Zero questions asked.
• Explicit Triggers: /saipen plan (turn a request or a raw backlog into tickets), /saipen ship (version bump, changelog, tag, push), /saipen clean (repo scrub), /saipen translate (isolated .saipen/saitranslate/ factory), /saipen markhunt (dry uncapped audit, records only), /saipen prepare (package work for handoff), /saipen validate (conformance check), /saipen goal (autonomous wave execution). Meta/control: /saipen status (read-only report), /saipen stop (checkpoint and halt). That plus saipen set and saipen continue is the whole surface -- twelve commands, full detail in RFC.md § 1.10.
• Strict Reliability: Batch input parsing (surgical 1-by-1 tickets), dirty-tree adoption (never wipes uncommitted work), secret redaction (sk-***).
• Experimental -- saicrew: an optional bonus layer (extensions/subs/, zero Core changes) for running a multi-agent crew -- one Core writer plus read-only saihunt/saipython workers reporting through their own OUTBOX.md. Under active live testing, not yet verified end-to-end -- see extensions/subs/crew.md.

Projects Powered by SAIPEN

• ⚡ FastPrompter — High-performance prompt management tool built around the SAIPEN memory protocol.

Two Layers

Layer	Required	Purpose
Core	✅	Continue work safely
Maintenance	On top of Core	Evolve the software with no tasking

Automated Evolution. No open to-dos left, type /saipen: HUNT audits for bugs, dead code, failing tests. Clean? ADD builds the next obvious missing capability, verifies it, hunts again. Product's mature -> stops gracefully.

GOAL Mode. /saipen goal <what you want> pivots the board (old tickets demoted, never deleted) and runs the new objective forward -- no "shall I continue?" between tickets, VERIFY/REVIEW never skipped. SHIP auto-pushes to an existing remote; a brand-new repo still asks once. Shipping the objective isn't the stopping point either -- it falls straight into autonomous HUNT/ADD maintenance until the product is mature, blocked, or the run hits its cap (3 waves / 20 tickets, then checkpoints and reports).

Quick Start

1. Install once per machine -- teaches Claude Code, Codex, Gemini, OpenCode, Aider, Antigravity, and any generic ~/.agents/skills reader (FreeBuff, etc.):

git clone https://github.com/vacterro/saipen
cd saipen
powershell -ExecutionPolicy Bypass -File .\bootstrap\inject.ps1     # Windows
bash bootstrap/inject.sh                                            # macOS / Linux

_{What that touches, so nothing is a surprise: it appends a marked <!-- SAIPEN:BEGIN -->...<!-- SAIPEN:END --> block to the agent instruction files you already have (~/.claude/CLAUDE.md, ~/.config/opencode/AGENTS.md, ~/.codex/AGENTS.md, ~/.gemini/GEMINI.md) -- backing each up to .bak first -- and copies the protocol into the matching skill folders. Nothing outside those paths, no daemon, no network calls.}

Changed your mind? One command puts it back:

powershell -ExecutionPolicy Bypass -File .\bootstrap\uninstall.ps1  # Windows
bash bootstrap/uninstall.sh                                         # macOS / Linux

It strips exactly the marked block (leaving the rest of your file alone), saves a .uninstalled.bak copy first, and removes the skill folders.

2. Start a project -- open an agent in your folder, type:

saipen set

No install? Paste one line to any agent:

Read /saipen/RFC.md + /saipen/STYLE.md and follow them.

Platform not in the list above (DeepSeek, Qwen, standalone OpenAI, etc.)? Per-platform notes live in extensions/adapters/.

Documentation & Specification Links

• SPEC.md -- formal architecture, design goals, litmus test.
• RFC.md -- normative specification executed by agents.
• GUIDE.md -- human tutorial & ELI5 guides:
  • 🇷🇺 Русский | 🇺🇸 English | 🇪🇪 Eesti | 🇯🇵 日本語 | 👴 Версия Деда
  • 🇺🇦 Українська | 🇩🇪 Deutsch | 🇫🇷 Français | 🇪🇸 Español | 🇮🇹 Italiano
  • 🇵🇹 Português | 🇳🇱 Nederlands | 🇵🇱 Polski | 🇸🇪 Svenska | 🇩🇰 Dansk
  • 🇫🇮 Suomi | 🇳🇴 Norsk | 🇨🇳 中文 | 🇰🇷 한국어 | 🇹🇭 ไทย | 🇻🇳 Tiếng Việt | 🇸🇦 العربية | 🇮🇱 עברית
  • 🇹🇷 Türkçe | 🇮🇳 हिन्दी | 🇮🇩 Bahasa Indonesia | 🇬🇷 Ελληνικά | 🇨🇿 Čeština | 🇷🇴 Română
  • 🇭🇺 Magyar | 🇧🇬 Български | 🇸🇰 Slovenčina | 🇭🇷 Hrvatski
• STYLE.md -- agent communication style & voice definition.
• UI.md -- Dark Golden Win95 UI design guidelines.
• CONFORMANCE.md -- behavioral test scenarios & validator rules.