Skip to content
Grandpa SAIPEN edited this page Jul 27, 2026 · 8 revisions

saipen

Hate repeating yourself? SAIPEN remembers so you don't.

One command. Zero deps. Zero amnesia. Any AI agent picks up where another left off — across vendors, sessions, continents. Claude stops. Gemini starts. No briefing.

Why saipen?

AI agents forget everything on context reset. SAIPEN dumps brain into .saipen/. Any agent reads, writes, resumes. No repeats.

Problem Without With saipen
Switch Claude→Gemini Re-brief from scratch saipen continue — resumes
Session ends at 2 AM "What was I doing?" STATE.md: phase BUILD, task T-14
Agent crashes Stale state, manual recovery RFC 1.5 auto-detect, reclaim, resume
3 tools, 1 project Context silos Shared .saipen/ — one memory
Mid-feature bug sweep Context switch saihunt — parallel subSaipen, OUTBOX collected

How it works

project-root/
  .saipen/
    STATE.md        # Current phase, task, next_action, blocker
    BOARD.md        # DOING / TODO / DONE / BLOCKED
    LOG.md          # Append-only event graph (E-###)
    kitchen/        # Scratch workspace
    recovery/       # Crash snapshots
    logs/           # Sealed LOG history
    saitranslate/   # Translation bundle
    extensions/subs/  # SubSaipen agents
  saipen/
    RFC.md, BOOT.md, STYLE.md, CONFORMANCE.md
    phases/         # Phase rule definitions
    tools/          # validate.py, install_hook.py, run_scenarios.py
  VERSION           # Semver
  1. Agent reads BOOT.md → STATE.md → BOARD.md → LOG.md tail
  2. Executes STATE.next_action immediately
  3. Transitions: PLAN → SCOUT → BUILD → VERIFY → REVIEW → SHIP → DONE
  4. After each ticket: LOG → BOARD → STATE checkpoint

Key features

  • Agent-agnostic — Claude, Gemini, Codex, Aider, DeepSeek, Qwen, any markdown reader
  • Portable — one command, zero deps, one .saipen/ dir
  • Phase-driven — structured pipeline PLAN→DONE
  • Goal mode — multi-wave execution (3 wave / 20 ticket caps)
  • SubSaipen — delegatable read-only sub-agents for parallel tasks
  • Crash-recovery — auto-detect stale state, restore from recovery
  • Multi-platform — Linux, macOS, Windows (PS + bash)
  • CI-ready — pre-commit hook, validate.py, test scenarios

Without vs With SAIPEN

Without With
Switch Claude→Gemini. Agent forgets everything. Re-explain project. Agent writes different code. Scream into void. saipen continue. Agent reads STATE. Continues where other stopped. Sip tea.
Session ends 2 AM. Morning: "What was I doing?" Agent says "IDK". Open 47 files. Still no clue. STATE.md says "phase: BUILD, task: T-14". Everyone knows.
Agent crashes. Cry. Agent returns. Starts from zero. Cry more. RFC 1.5. Stale detected. Recovery 30 seconds.
Run 3 agents. Agent A knows X. Agent B knows Y. Agent C knows nothing. You're marriage counselor. One .saipen/. All agents read same brain. Fire counselor hat.
Feature building. Want bug sweep. Must wait. Feature takes 2 hours. Bug still there. saihunt runs parallel. Feature continues. Bug found.

Learn more

External:

  • RFC — full protocol
  • BOOT — cold-start kernel
  • STYLE — communication voice
  • GUIDE — user guide
  • SPEC — technical spec
  • CHANGELOG — full history

Version

v7.82.0 — One command. Zero dependencies. Zero amnesia. — MIT

saipen — one command, zero deps, zero amnesia


Clone this wiki locally