Home

saipen

saipen (pronounced sigh-pen) is a cross-agent continuity protocol and project management system. One command. Zero dependencies. Zero amnesia. It lets any AI agent pick up exactly where another left off — across vendors, across sessions, across continents.

Why saipen?

Every AI coding agent has context windows. When the window resets, so does the agent's memory of what you were doing. SAIPEN solves this by externalizing project state into a persistent .saipen/ directory that any agent can read and write.

Problem	Without saipen	With saipen
Switch from Claude to Gemini	Lose all context, re-brief from scratch	saipen continue — agent reads .saipen/ and resumes
Session ends at 2 AM	Next morning: "what was I doing?"	STATE.md says phase: BUILD, task: T-14, pick up exactly
Agent crashes mid-flight	Stale state, lost work, manual recovery	RFC 1.5 Recovery: auto-detect staleness, reclaim, resume
3 tools, 1 project	Silos of context per tool	Shared .saipen/ — one memory, any tool
Need a bug sweep mid-feature	Pause main work, context switch	saihunt — subSaipen runs in parallel, OUTBOX collected

How it works

project-root/
  .saipen/
    STATE.md        # Current phase, task, next_action, blocker
    BOARD.md        # Ticket board: DOING / TODO / DONE / BLOCKED
    LOG.md          # Append-only event graph (E-### chain)
    kitchen/        # Scratch workspace, half-finished findings
    recovery/       # Crash-recovery snapshots
    logs/           # Sealed LOG history segments
    saitranslate/   # Translation bundle (if active)
    extensions/
      subs/         # SubSaipen sub-agents
        saiwiki/
        saihunt/
        saitranslate/
  saipen/
    RFC.md          # Full protocol constitution
    BOOT.md         # Cold-start kernel
    STYLE.md        # Communication voice
    CONFORMANCE.md  # Behavioral specification
    phases/         # Phase rule definitions
    tools/          # validate.py, install_hook.py, run_scenarios.py
    tests/          # Test fixtures and scenarios
  VERSION           # Semantic version

1. Agent reads BOOT.md -> STATE.md -> BOARD.md -> LOG.md tail
2. Executes STATE.next_action immediately
3. Transitions through phases: PLAN -> SCOUT -> BUILD -> VERIFY -> REVIEW -> SHIP -> DONE
4. After each ticket: LOG -> BOARD -> STATE checkpoint

Key features

• Agent-agnostic — works with Claude, Gemini, Codex, Aider, DeepSeek, Qwen, any agent that reads markdown
• Portable — one command, zero dependencies, one .saipen/ directory
• Phase-driven — structured, repeatable workflow from PLAN to DONE
• Goal mode — multi-wave execution with automatic scoping (3 waves / 20 ticket caps)
• SubSaipen — delegatable read-only sub-agents for parallel tasks
• Crash-recovery — auto-detect stale state, restore from recovery, reclaim tickets
• Multi-platform — Linux, macOS, Windows (PowerShell + bash bootstrap)
• CI-ready — pre-commit hook, validate.py structural checker, executable test scenarios

Learn more

• Getting Started — setup in 3 commands
• Phases — comprehensive phase reference (16 phases)
• Scenarios — 8 workflow walkthroughs
• Tutorials — 7 step-by-step guides
• Use-Cases — 10 real-world patterns
• SubSaipen — sub-agent roles and lifecycle

External docs:

• RFC — full protocol constitution
• BOOT — cold-start kernel (read this first)
• STYLE — communication voice
• GUIDE — comprehensive user guide
• SPEC — technical specification
• CONFORMANCE — behavioral spec
• CHANGELOG — full release history

Version

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