What happened?

hooks_cli.py claims to support both claude-code and codex harnesses, but _count_human_messages() appears to only count transcript entries shaped like:

{"message": {"role": "user", "content": "..."}}

That matches the Claude-style transcript shape, but it does not match the Codex session format already handled in normalize.py, where user messages are parsed from event_msg entries with payload.type == "user_message".

As a result, Codex hook message counting may stay at 0 or undercount user turns, which means the stop hook auto-save threshold may never trigger correctly for Codex sessions.

What did you expect?

If hooks_cli.py supports the codex harness, it should count user turns from the Codex transcript schema correctly, so stop-hook checkpointing works the same way it does for Claude-compatible transcripts.

How to reproduce:

1. Inspect mempalace/hooks_cli.py and note that _count_human_messages() only looks for entry["message"]["role"] == "user"
2. Inspect mempalace/normalize.py and note that Codex sessions are parsed from event_msg entries with payload.type == "user_message"
3. Run the stop hook against a Codex transcript using the Codex schema
4. Observe that user messages may not be counted correctly, so the save interval logic does not trigger as expected

Environment:

• OS: macOS
• Python version: 3.12
• MemPal version: 3.0.14