Bug

The Stop hook in mempal_save_hook.sh (lines 87-104) counts every JSONL entry with role: "user" as a human message. In Claude Code transcripts, tool results (Bash, Read, Grep, subagent dispatches, etc.) are also sent as role: "user" with content blocks of type: "tool_result".

This inflates the exchange count significantly. In one measured session, 425 actual human messages produced a count of 1,221 — a ~2.9x inflation. Tool-heavy and subagent-heavy sessions trigger saves far more often than intended.

Current behavior

if isinstance(msg, dict) and msg.get('role') == 'user':
    content = msg.get('content', '')
    if isinstance(content, str) and '<command-message>' in content:
        continue
    count += 1

Only <command-message> strings are filtered. Tool results (where content is a list of {"type": "tool_result", ...} blocks) are counted as human messages.

Expected behavior

Only actual human text input should increment the counter. Tool results should be skipped.

Proposed fix

When content is a list, check whether all blocks are tool_result type — if so, skip:

if isinstance(msg, dict) and msg.get('role') == 'user':
    content = msg.get('content', '')
    if isinstance(content, str) and '<command-message>' in content:
        continue
    # Skip tool results — they arrive as role: "user" but aren't human input
    if isinstance(content, list) and all(
        isinstance(b, dict) and b.get('type') == 'tool_result'
        for b in content
    ):
        continue
    count += 1

This is safe for non-Claude-Code transcript formats: if another LLM uses role: "tool" for tool results (e.g. OpenAI format), those entries never match role == "user" in the first place, so this filter is a no-op.

Related

• fix: skip tool_result messages in _count_human_messages #316 attempts this fix but targets hooks_cli.py, which doesn't exist on main — the counting logic lives in inline Python within hooks/mempal_save_hook.sh
• Bug: Codex hook message counting does not match Codex transcript schema #347 / fix: count Codex user_message turns in _count_human_messages (#347) #373 address a separate Codex transcript counting issue