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.md -> BOARD.md -> active LOG.md tail -> human_note (if
set), executes next_action, and resumes work in under a minute -- no
rebriefing, any vendor, any day.
One command. Zero dependencies. Zero amnesia.
Fast keys. A shortcut is the entire message, never a prefix. cc keeps an active Goal Mode run moving, sss reports status without touching code, and ss checkpoints and stops. See the full 13-key map. Cyrillic twins work too: сс, ссс, аа, ее, еее, рр.
Reply language. The agent answers in Estonian by default — that is a setting, not a quirk, and nothing else about SAIPEN is Estonian. Change it in one place: the reply_language: line at the top of saipen/STYLE.md. et Estonian, en English, ru Russian, auto picks from the message you sent. The protocol, the code, the commits and every document stay English at every value.
v7.177.0 | Spec | Guide | RFC | Style | UI | Conformance | MIT
User -> /saipen continue
Agent -> reads STATE.md (phase, task, next_action, mode, human_note)
Agent -> reads BOARD.md (DOING / TODO / DONE / BLOCKED tickets)
Agent -> reads active LOG.md tail (recent events)
Agent -> reads human_note (if set, one-time nudge)
Agent -> executes next_action (command) immediately
Agent -> loads phase doc only when rules are needed
Agent -> Works.
Project state beats model memory. Memory lives in the project, not in a
model's head. Project -> Memory -> LLM becomes Project -> SAIPEN State -> LLM.
- Core state machine —
INIT → PLAN → SCOUT → BUILD → VERIFY → REVIEW → SHIP → DONE | BLOCKED - Zero-prompt autonomy — board halted (no workable
TODO, nothing inDOING) and notBLOCKED? Auto-transitionsHUNT(scan bugs) →ADD(evolve features) →HUNT, zero questions asked. A session sitting atBLOCKEDnever auto-hunts; it waits for the human to resolve the blocker (RFC § 2.1). - Strict reliability — batch input parsed into surgical one-by-one tickets,
dirty-tree adoption that never wipes uncommitted work, secret redaction
(
sk-***).
The whole surface is 16 commands; full detail in RFC § 1.10.
| Command | Does |
|---|---|
/saipen set |
Adopt a project |
/saipen continue |
Resume exactly where it stopped |
/saipen plan |
Turn a request or raw backlog into tickets |
/saipen goal <text> |
Autonomous wave execution against a new objective |
/saipen hunt |
Force the defect/improvement sweep now |
/saipen ship |
Version bump, changelog, tag, push |
/saipen clean |
Repo scrub |
/saipen validate |
Conformance check |
/saipen markhunt |
Dry uncapped audit, records only |
/saipen translate |
Isolated translation factory |
/saipen prepare |
Package work for handoff |
/saipen collect |
Integrate a ready package |
/saipen status |
Read-only report |
/saipen stop |
Checkpoint and halt |
saipen init and saipen sub complete the sixteen; both are invoked by the
protocol rather than typed day to day.
Package keys. ee/qq prepare complete translation/wiki packages without
integrating; eee/qqq accept only ready packages, then integrate, verify,
review, and push.
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 live
testing, not yet verified end to end — see extensions/subs/crew.md.
| 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).
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 / LinuxWhat 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.
2. Start a project -- open an agent in your folder, type:
saipen set
No install? Paste one line to any agent:
Read <clone>/saipen/BOOT.md first (cold-start kernel), then <clone>/saipen/RFC.md + <clone>/saipen/STYLE.md and follow them.
Platform not in the list above (DeepSeek, Qwen, standalone OpenAI, etc.)?
Per-platform notes live in extensions/adapters/.
Changed your mind? One command puts it back:
powershell -ExecutionPolicy Bypass -File .\bootstrap\uninstall.ps1 # Windows
bash bootstrap/uninstall.sh # macOS / LinuxIt strips exactly the marked block (leaving the rest of your file alone), saves
a .uninstalled.bak copy first, and removes the skill folders.
| Document | What it is |
|---|---|
| SPEC.md | Formal architecture, design goals, litmus test |
| RFC.md | Normative specification executed by agents |
| GUIDE.md | Human tutorial and ELI5 guides |
| STYLE.md | Agent communication style and voice |
| UI.md | Vintage Golden UI design guidelines |
| CONFORMANCE.md | Behavioral test scenarios and validator rules |
All 33 translated 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
⚡ FastPrompter — high-performance prompt management tool built around the SAIPEN memory protocol.
Click to expand
