login-probe-selfheallisted
Install: claude install-skill Tommyapi/ai-employee-skills
# Login probe and self-heal
Most "the agent failed" incidents are a dependency that died hours earlier: a browser that quit, a session that expired, a tunnel that dropped. Probe first, heal if you can, and tell the agent the truth.
## Steps
1. Write `probes.json`; `check` and `heal` are argv lists (no shell):
```json
{"probes": [
{"name": "browser", "check": ["curl", "-sf", "--max-time", "5", "http://127.0.0.1:9222/json/version"],
"heal": ["./start-browser.sh"], "retries": 1, "cooldown_min": 30, "timeout_s": 20},
{"name": "api-key", "check": ["./check-key.sh"]}
]}
```
2. Run: `python scripts/probe_heal.py --config probes.json --status status.json`
3. Exit 0 means all healthy (some may have been healed, shown as `(healed)`); exit 1 means at least one is down. The agent reads `status.json` and skips work that needs a down dependency, reporting it instead of looping.
## Behaviour
- Healthy = check exits 0. A timeout or a missing binary counts as down, never as a crash.
- Heal runs at most once per `cooldown_min` per probe, so a broken thing is not hammered (and an account is not flagged by a restart loop).
- `--no-heal` probes only.
## Rules
- Identity checks (a human must scan a QR code, solve a face check) are not healable: the heal command should notify the owner with the one concrete step, not retry.
- Keep secrets out of `probes.json`; have the check script read them from your own secret store.