Agent Onboarding

Start in the wrong order and you will miss critical context. Follow this protocol exactly.

1) Bootstrap (first run only)

If BOOTSTRAP.md exists:

1. Read it fully.
2. Execute its instructions.
3. Delete it after completion.

Treat it as one-time identity transfer, not permanent reference material.

2) Session Startup Order

At the beginning of every session, load context in this sequence:

1. SOUL.md
2. USER.md
3. memory/YYYY-MM-DD.md for today (if present)
4. memory/YYYY-MM-DD.md for yesterday (if present)
5. MEMORY.md only in the direct main session with the human
6. .learnings/ pending items

Why this order:

• SOUL.md defines behavioral/safety boundaries.
• USER.md defines who you are serving.
• Daily memory restores short-term continuity.
• MEMORY.md is private long-term context (main-session only).
• .learnings/ prevents repeated mistakes.

3) Compaction Recovery Order

If conversation context is compacted, recover from durable state before asking for re-brief:

1. memory/YYYY-MM-DD.md
2. memory/active-sessions.json
3. git log --oneline -10
4. Relevant PRD

Only ask the human for clarification when ambiguity is real and cannot be resolved from files.

4) Non-Negotiable Working Rules

• Persist important information to files immediately. Do not rely on conversational memory.
• Never trust delegated completion text alone. Verify artifacts on disk (read + git status + expected outputs).
• Route completion updates to the intended Telegram topic/thread, not the current chat by default.

Operating Posture

Recover from files, not vibes. Continue from durable state, not repeated user prompting. Write down what matters. Verify before reporting success.