agent-sentinel#

Purpose#

This skill is the mandatory evaluation layer between the agent's intent and any high-stakes tool execution. You are not permitted to use the tools listed under Interception Triggers without first calling this script and receiving "decision": "ALLOW" or "decision": "ADVISE" in the result.

Think of this as a circuit breaker — if the Sentinel trips, the circuit opens and the action stops.

---

Interception Triggers#

You are FORBIDDEN from invoking any of the following tools without first running eval_engine.py and parsing its response.

ToolTrigger conditionweb_searchEvery search, without exceptionbooking_toolEvery flight, hotel, or travel bookingshell_commandEvery shell or terminal executionpayment_toolEvery payment, checkout, or purchase

No exceptions apply. Even if you are certain the action is safe, the Sentinel must still be called. This is a policy requirement, not a suggestion.

---

How to Call the Sentinel#

Run the following command before invoking any trigger tool:

bash

python3 ~/.openclaw/skills/agent-sentinel/eval_engine.py \
  --intent  "<what the user asked for>" \
  --action  <booking_tool|web_search|shell_command|payment_tool|other> \
  --data    "<the exact payload: URL, flight details, command, amount>" \
  [--provider  anthropic|openai|ollama] \
  [--model     <model-id>] \
  [--config_path /path/to/SENTINEL_CONFIG.md]

Example — flight booking:

bash

python3 ~/.openclaw/skills/agent-sentinel/eval_engine.py \
  --intent "Book a family trip to Orlando for spring break" \
  --action booking_tool \
  --data   "Delta Airlines, dep 08:30, arr 11:45, $389 total, non-stop, economy"

Example — web search:

bash

python3 ~/.openclaw/skills/agent-sentinel/eval_engine.py \
  --intent "Find age-appropriate science videos for my daughter" \
  --action web_search \
  --data   "https://www.youtube.com/results?search_query=kids+science+experiments"

Important: The script writes Chain-of-Thought reasoning to stderr and emits only valid JSON to stdout. Parse stdout with json.loads(...). Do not parse stderr.

---

Response Schema#

The script always returns a single JSON object:

json

{
  "decision":     "ALLOW" | "BLOCK" | "ADVISE",
  "severity":     "LOW"   | "MEDIUM" | "HIGH",
  "reason":       "<clear explanation>",
  "alternatives": "<suggestion to resolve the violation>"
}

---

Decision Handling Rules#

"ALLOW" — Proceed#

The action passed all checks. Continue with the intended tool call. If the result contains "severity": "LOW" alongside ALLOW, surface any informational notes to the user as a soft advisory but do not block.

---

"ADVISE" — Pause and Confirm#

The action is not blocked but a preference mismatch or soft-limit warning was detected. You must:

1. Stop before invoking the tool.
2. Show the reason and alternatives fields to the user verbatim.
3. Ask the user explicitly: "Would you like to proceed anyway?"
4. Only continue if the user confirms. If they do not confirm within the turn, treat it as a BLOCK.

Example ADVISE response to user:

I noticed an advisory before completing your request:

Advisory: Price $480 is within 15% of your $500 budget cap.

Suggestion: Confirm this cost is acceptable or I can search for cheaper alternatives.

Would you like me to proceed with this booking, or should I look for less expensive options?

---

"BLOCK" — Stop Immediately#

You are strictly forbidden from proceeding. Do not attempt to:

• Retry the same action with different parameters
• Find a workaround or alternative path to the same outcome
• Bypass the Sentinel by splitting the action into smaller steps
• Claim the Sentinel is wrong and proceed anyway

You must:

1. Do not call the trigger tool.
2. Apologize to the user and clearly explain the violation.
3. Quote the reason field exactly.
4. If alternatives is non-empty, present it as the recommended path forward.
5. Ask for an explicit user override if they wish to continue.

Example BLOCK response to user (budget violation):

I'm sorry — I can't complete this booking.

Blocked: Price $650.00 exceeds your maximum budget of $500.00.

What you can do: Look for options priced at or below $500. Consider flexible dates or alternate airports.

If you'd like to override this limit for this booking only, please say "override" and I'll ask you to confirm the amount before proceeding.

Example BLOCK response to user (child-safety violation):

I'm sorry — I can't perform this search.

Blocked: This content is restricted under the household child-safety policy (severity: HIGH).

Reason: [reason from the Sentinel]

Please modify your request. If you believe this is an error, an adult in the household can review and override the policy in SENTINEL_CONFIG.md.

---

Override Protocol#

If a user explicitly says "override" for a BLOCK decision, you must:

1. Repeat the blocking reason and severity back to the user.
2. Ask for explicit written confirmation: "Please type 'I confirm' to proceed despite this policy violation."
3. Log the override in your response (e.g., "Proceeding with user override.").
4. Never offer override for a severity: HIGH (Tier-1 child-safety) BLOCK unless an adult user has explicitly established that permission in writing within the same conversation turn.

---

Installing Dependencies#

bash

cd ~/.openclaw/skills/agent-sentinel
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

---

Configuration#

Edit SENTINEL_CONFIG.md (in the skill directory or ~/.openclaw/) to update your preferences and safety policy. See that file for full documentation of all supported keys.

KeyTypeEffectChild_Age_LimitintegerActivates child-safety tierMax_Budget``$NNNHard budget cap (BLOCK above, ADVISE at 85%)Night_Flights_Blocked``true/falseBlocks flights in night windowNight_Flight_Window``HH:MM - HH:MMNight restriction hoursPreferred_Airlinescomma listSoft preference (ADVISE if absent)Blocked_Airlinescomma listHard block on listed carriersMax_StopsintegerBLOCK if flight exceeds stop countPreferred_CabinstringADVISE if different cabin detectedMax_Booking_Advance_DaysintegerADVISE if booking too far ahead