openclaw - 💡(How to fix) Fix Agent skips AGENTS.md startup sequence — no enforced context loading at session start

Problem

The agent (Sunny) has AGENTS.md instructions that define a mandatory startup sequence:

1. Read SOUL.md
2. Read USER.md
3. Read memory/session-handoff.md
4. Read today + yesterday daily notes
5. Read MEMORY.md

These are not enforced by the runtime. The agent can (and does) skip them entirely, jumping straight into answering the user's first message without loading any prior context.

Impact

Two days of avoidable errors:

• May 28: Sent email to wrong domain (.org vs .com) — correct address was in the person file that should have been loaded
• May 29: Assumed user's 8-year-old daughter was an infant — her birthday was in MEMORY.md which wasn't read at startup

Both errors would have been prevented if the startup sequence had actually run.

Expected Behavior

The runtime should either:

1. Auto-inject key workspace files (SOUL.md, USER.md, MEMORY.md, session-handoff.md) into context at session start, OR
2. Enforce the AGENTS.md startup sequence before allowing the agent to respond to the first user message

Current Behavior

AGENTS.md startup instructions are advisory only. The agent receives the user's first message and responds immediately without loading context files. The instructions say "Don't ask permission. Just do it." but there's no mechanism to ensure compliance.

Environment

• OpenClaw gateway on macOS (Darwin 25.3.0)
• WhatsApp channel (direct messages)
• Agent: Sunny (claude-opus-4-7)