Session Janitor

Automated transcript trimming, LLM memory extraction, and session hygiene for OpenClaw gateways. Keeps transcripts from bloating, extracts structured memorie...

What to consider before installing: - Scope & secrets: setup auto-reads your OpenClaw installation(s) (~/.openclaw*/openclaw.json) and uses each gateway's auth token to call the local API. That token is sensitive; the skill did not declare it up front. If you are uncomfortable granting that access, do not install. - Persistence: setup will modify your crontab and install a user watcher service (systemd/launchd). These run automatically on your account and will perform file edits, trims, and LLM extraction without further prompts. - Memory handling / possible exfiltration: despite a doc claim that memories go "only" to the mem DB, extract-llm.py also writes scene files and attempts to git commit/push them (if sceneFilesPath points inside a git repo). If that repo has a remote, extracted memories could be pushed off-machine. The skill also invokes the optional 'mem' CLI which may transmit data elsewhere depending on your mem setup. To reduce risk: leave memCli.disabled, do not set sceneFilesPath (or set it to a safe local directory not tracked by git), and inspect/disable git remotes if you enable scene files. - Restart behavior: the watchdog can call restart scripts (paths are configurable). Review those scripts before enabling autoRestart; they could encapsulate arbitrary commands run automatically by the janitor. - Inspect before enabling: review scripts/setup.sh, scripts/janitor.sh, scripts/extract-llm.py, and scripts/sidecar.py yourself. Consider running janitor in a dry-run environment (copy your sessions to a test dir) and run janitor.sh manually to observe behavior before allowing cron/service installs. - Recommended mitigations: back up ~/.openclaw, disable llmExtraction or set maxPerRun=0 if you don't want automatic LLM calls, set memCli.enabled=false, leave sceneFilesPath empty, and check crontab and systemd user services after setup. If you accept this skill, be aware it will automatically modify and archive session transcripts and can persist extracted data outside the gateway unless you reconfigure it. What would reduce my concern: explicit registry metadata declaring that the skill reads openclaw.json and gateway tokens, a documented opt-in for scene file git commits (or removal of automatic git push), and an explicit configuration/consent step for installing cron/systemd services and for any auto-restart behavior.

Type: OpenClaw Skill Name: session-janitor Version: 1.5.0 The session-janitor skill is a maintenance utility that performs transcript trimming and memory extraction, but it employs several high-risk patterns. The setup.sh script automatically harvests authentication tokens from all discovered OpenClaw gateway configurations in the user's home directory. Additionally, janitor.sh uses an unsafe 'eval' pattern to load configuration values, creating a shell injection vulnerability. The extract-llm.py script also includes logic to automatically git commit and push 'scene files' to a remote repository, which could be leveraged for data exfiltration.