Skip auto-memory hooks when project is not initialized

[severity1]

Summary

Auto-memory hooks should not trigger unless the project has been initialized via /auto-memory:init. Running on uninitialized projects wastes resources and may cause confusion.

Problem

Currently, load_config() in post-tool-use.py returns a default config when no config file exists:

def load_config(project_dir: str) -> dict:
    config_file = Path(project_dir) / ".claude" / "auto-memory" / "config.json"
    if config_file.exists():
        # ... load config
    return {"triggerMode": "default"}  # <-- defaults even when not initialized

This causes:

• Dirty files tracked on projects that have no auto-memory setup
• Stop hook may block and request memory-updater spawn on uninitialized projects
• Unnecessary file I/O and processing overhead
• Potential confusion when hooks fire unexpectedly

Proposed Solution

Treat missing config as "not enabled":

1. post-tool-use.py: Exit early if config file doesn't exist
2. stop.py: Exit early if config file doesn't exist

def load_config(project_dir: str) -> dict | None:
    """Load plugin configuration. Returns None if not initialized."""
    config_file = Path(project_dir) / ".claude" / "auto-memory" / "config.json"
    if not config_file.exists():
        return None  # Not initialized
    # ... load and return config

# In main():
config = load_config(project_dir)
if config is None:
    return  # Project not initialized, skip

Affected Files

• scripts/post-tool-use.py - add config existence check
• scripts/stop.py - add config existence check (for consistency)

[Baelson]

Reproduction: SubagentStop fails to clear dirty-files on uninitialized projects

I hit this exact issue. Here's the specific failure path:

What happens

1. PostToolUse hook writes file paths to {project}/.claude/auto-memory/dirty-files (even without config.json)
2. Stop hook reads dirty-files, finds entries, returns "decision": "block" with agent spawn instruction
3. memory-updater agent runs, processes the files, returns successfully
4. SubagentStop handler runs — but bails early because config.json doesn't exist:

# trigger.py handle_subagent_stop()
config_file = Path(project_dir) / ".claude" / "auto-memory" / "config.json"
if not config_file.exists():
    return  # <-- dirty-files never cleared!

5. Next turn: Stop hook fires again with the same stale dirty-files → infinite loop

Impact

Every single turn triggers a blocking Stop hook + memory-updater agent spawn (~2-5K tokens each time), processing the same files repeatedly. The "error" display (anthropics/claude-code#12667) makes it look like something is broken every turn.

Workaround

Manually create the config file:

mkdir -p .claude/auto-memory
echo '{"triggerMode": "default"}' > .claude/auto-memory/config.json

Or manually clear dirty-files:

: > .claude/auto-memory/dirty-files

Suggested fix

In handle_subagent_stop(), either:

1. Remove the config_file.exists() guard (always clear dirty-files after agent completes), or
2. Also check in post-tool-use.py before writing — skip tracking if no config.json exists

Option 2 aligns with the issue title (skip hooks when not initialized).

Environment

• auto-memory v0.8.2
• Claude Code v2.1.69
• macOS Darwin 25.3.0