Fix boot timing, CLI flags, and SSH quoting (review sweep #40)

.claude/commands/fleet-announce.md

@@ -89,12 +89,22 @@ Send all? Or edit?
 Look up the notification channel and recipient ID from each person's fleet file, then
 send via SSH:
 
-```bash
-ssh <host> "openclaw message send --channel <channel> -t <recipient_id> -m '<message>'"
+```python
+import subprocess, shlex
+
+msg = "<message>"  # the crafted message
+cmd = (
+    f"openclaw message send"
+    f" --channel <channel>"
+    f" --target <recipient_id>"
+    f" --message {shlex.quote(msg)}"
+)
+subprocess.run(["ssh", "<host>", cmd])
 ```
 
-**Important:** Single quotes in the message must be escaped. Messages come FROM their
-bot, not from the fleet owner.
+**Why Python:** Shell quoting is unreliable for natural-language messages that contain
+contractions (`don't`), dollar signs, or backticks. `shlex.quote()` handles all of these
+correctly. Messages come FROM their bot, not from the fleet owner.
 
 ## 5. Confirm
 

templates/BOOT.md

@@ -4,14 +4,10 @@ When the gateway starts (or restarts), run through this checklist before doing a
 else. This is your "waking up" moment — orient yourself, check what happened while you
 were offline, and get ready.
 
-## 1. Orient
+## 1. Conversation Recovery (Was I Mid-Conversation?)
 
-- Read `SOUL.md` — remember who you are
-- Read `USER.md` — remember who you're helping
-- Read `memory/YYYY-MM-DD.md` (today) — what happened so far today?
-- If today's file doesn't exist, read yesterday's — pick up where you left off
-
-## 2. Conversation Recovery (Was I Mid-Conversation?)
+**Do this first — time-sensitive.** The 1-minute window starts from the restart, not
+from when you finish orienting. Check the session store before reading any other files.
 
 Check if a restart interrupted an active conversation.
 
@@ -28,20 +24,28 @@ Check if a restart interrupted an active conversation.
 
 **If interrupted (< 1 min ago):**
 
-1. Read the last 20 lines of that session's JSONL transcript
-2. Extract the last few user messages and your last response
-3. Write a short summary of the interrupted conversation to a temp note: what was being
+1. Run Orient (Step 2) — load your identity and context before reaching out
+2. Read the last 20 lines of that session's JSONL transcript
+3. Extract the last few user messages and your last response
+4. Write a short summary of the interrupted conversation to a temp note: what was being
    discussed, what the user last asked, what you were working on
-4. Send the user a message via the `message` tool on the channel where the conversation
+5. Send the user a message via the `message` tool on the channel where the conversation
    was happening:
    - Keep it brief: "Just restarted. We were working on [X]. Want to pick up where we
      left off?"
    - Include enough context that they can say "yes, continue" without re-explaining
-5. Reply with NO_REPLY after sending the message
+6. Reply with NO_REPLY after sending the message
 
 **If NOT interrupted (> 1 min ago or no recent direct conversation):**
 
-- Skip to step 3 (check for unfinished work)
+- Continue to step 2 (Orient)
+
+## 2. Orient
+
+- Read `SOUL.md` — remember who you are
+- Read `USER.md` — remember who you're helping
+- Read `memory/YYYY-MM-DD.md` (today) — what happened so far today?
+- If today's file doesn't exist, read yesterday's — pick up where you left off
 
 ## 3. Check for Unfinished Work
 
@@ -53,7 +57,7 @@ Check if a restart interrupted an active conversation.
 
 Don't message your human on routine restarts. Only reach out if:
 
-- Step 2 detected an interrupted conversation (already handled above)
+- Step 1 detected an interrupted conversation (already handled above)
 - You found unfinished urgent work that needs their attention
 - Something is broken that was working before
 - It's been >24h since any interaction (check-in is appropriate)