anti-suicide

Safety guardrail that prevents OpenClaw from self-destructing via erroneous self-modification. ALWAYS use this skill before modifying any of OpenClaw's own critical files: openclaw.json, workspace prompt files (AGENTS.md, SOUL.md, TOOLS.md), installed skills, .env files, docker-compose.yml, or any file inside ~/.openclaw/ or the openclaw installation directory. Also trigger when the user says "update my config", "modify openclaw settings", "change the agent prompt", "edit my soul/agents/tools file", "reinstall", or any operation that touches OpenClaw's own runtime files. The goal is to enforce Backup → Snapshot → Modify → Verify for every self-modification, with automatic rollback if health degrades.

install

source · Clone the upstream repo

git clone https://github.com/shyrock/anti-suicide

Claude Code · Install into ~/.claude/skills/

git clone --depth=1 https://github.com/shyrock/anti-suicide ~/.claude/skills/shyrock-anti-suicide-anti-suicide

manifest: SKILL.md

source content

Anti-Suicide Skill

OpenClaw is an always-on personal assistant. One bad write to

openclaw.json

or a corrupted

SOUL.md

can take it offline or alter its behavior in subtle, hard-to-debug ways. This skill enforces a safe modification protocol using a supervisor script that automatically rolls back changes if the service degrades.

What Counts as "Self-Modification"

Any write, edit, or delete targeting:

Path	Risk
`~/.openclaw/openclaw.json`	Gateway fails to start; all channels go dark
`~/.openclaw/workspace/AGENTS.md`	Agent identity/capabilities silently corrupted
`~/.openclaw/workspace/SOUL.md`	Personality and core behavior overwritten
`~/.openclaw/workspace/TOOLS.md`	Available tools list broken
`~/.openclaw/workspace/skills/*/SKILL.md`	Individual skill broken or lost
`~/.openclaw/.env`	API keys or channel tokens lost
`docker-compose.yml`	Container won't start
`package.json` / `pnpm-workspace.yaml`	Dependency resolution broken

See

references/critical_paths.md

for field-level details.

Safe Modification Protocol

Every self-modification follows these five steps in strict order.

Step 1 — Start the Supervisor (Before Touching Anything)

The supervisor captures a health baseline and returns a session ID. Start it before making any changes:

SESSION=$(python ~/.openclaw/workspace/skills/anti-suicide/scripts/supervisor.py snapshot \
  --files <file1> [file2 ...])
echo "Session: $SESSION"

The session ID is a path like

/tmp/anti-suicide-<timestamp>/

. Save it — you'll need it in Step 4.

If the supervisor reports the service is already unhealthy at baseline, stop immediately. Do not make modifications on a service that is already degraded. Report the health output to the user and ask how to proceed.

Step 2 — Show the Diff

Before applying any change, show the user exactly what is changing:

FILE: <filepath>

BEFORE:
───────────────────
<current content of the affected section>
───────────────────

AFTER:
───────────────────
<proposed content>
───────────────────

For JSON files, also validate the proposed content parses correctly before showing the diff:

python ~/.openclaw/workspace/skills/anti-suicide/scripts/supervisor.py validate-json \
  --content '<proposed_json>'

If JSON is invalid, stop here and tell the user what needs fixing.

Step 3 — Confirm

Say exactly:

"About to modify
<filepath>
. Supervisor session
<SESSION>
is active and will auto-rollback if the service degrades. The diff is shown above. Shall I proceed?"

Wait for explicit confirmation before writing anything.

Step 4 — Apply the Change

Make the modification using the Edit tool (never Write for existing files, to avoid accidental full overwrites).

Step 5 — Verify (Supervisor Monitors and Auto-Heals)

After applying, hand off to the supervisor to monitor the service:

python ~/.openclaw/workspace/skills/anti-suicide/scripts/supervisor.py verify \
  --session $SESSION \
  --timeout 60 \
  --interval 5

The supervisor will:

Poll
```
openclaw doctor
```
, gateway liveness, and channel status every 5 seconds for 60 seconds
If health is stable throughout → print
```
HEALTHY
```
and exit 0
If health degrades → automatically rollback all backed-up files, restart the gateway, and print a report

Report the result to the user. If auto-rollback occurred, say what was rolled back and that the service has been restored.

Hard Blocks

The following are never safe without explicit user override:

Deleting
```
openclaw.json
```
,
```
SOUL.md
```
,
```
AGENTS.md
```
, or
```
TOOLS.md
```
entirely
Overwriting a prompt file with empty or near-empty content
Killing the gateway process without first warning the user about downtime
Running
npm install
or
pnpm install
inside
```
~/.openclaw/
```
without confirming no pinned versions will break
Running
docker-compose down
without warning that all channels will disconnect

Manual Rollback

If you need to rollback outside of the supervisor flow:

python ~/.openclaw/workspace/skills/anti-suicide/scripts/rollback.py --session <SESSION>
openclaw gateway restart

Or restore a specific file:

python ~/.openclaw/workspace/skills/anti-suicide/scripts/rollback.py --file <backup_file_path>

References

```
references/critical_paths.md
```
— full list of critical file paths with required structure