Tool Call Guard

skill: tool-call-guard

purpose

Validate a tool call before it executes. Catches missing parameters, malformed values, and scope overreach. Blocks bad calls. Warns on irreversible ones. Prevents wasted API calls and accidental side effects.

whentouse

• Before every external tool call in an agent loop
• Tool call involves writes, deletes, or sends
• Input to the tool comes from user text that might be incomplete or malformed

whennotto_use

• Simple safe read operations where calling is always correct
• You've already validated this exact call this turn

input

tool_name: string          # Tool you're about to call
tool_params: object        # Parameters you're about to pass
tool_schema: object        # Expected types/requirements (optional — skip validation step if not provided)
task_context: string       # Why you're making this call
is_destructive: boolean    # Creates, modifies, sends, or deletes data?

output

{
  "approved": true | false,
  "issues": ["string"],
  "safe_to_call": true | false,
  "warning": "string | null",
  "corrected_params": "object | null"
}

instructions

Check this tool call before it executes.

Validate:
1. NECESSITY: Is this call needed for the task, or can the task be done without it?
2. PARAMETERS: Required params present? Types correct? Values in expected range?
3. SCOPE: Does this call do exactly what the task requires, or could it do more?
4. SIDE EFFECTS: If destructive, is scope limited to what was authorized?

Reject if: required param missing or null, value clearly wrong (negative ID, empty required field, wrong format), call would affect more data than intended.

Warn if: call is approved but irreversible.

Correct if: param has fixable format issue (whitespace, wrong case for enum, missing prefix). Never change values that affect intent.

safe_to_call is false if: not approved, OR is_destructive is true and warning is present.

Return JSON only.

constraints

• Max output tokens: 200
• corrected_params: fix format only, never change intent
• safetocall false if destructive + has warning
• If approved AND correctedparams is non-null: always use correctedparams in the actual call, not the original tool_params

example

Input:

tool_name: "delete_user"
tool_params: { "user_id": "", "confirm": true }
tool_schema: { "user_id": "required non-empty string", "confirm": "required boolean" }
task_context: "Delete account for user ID 8821"
is_destructive: true

Output:

{
  "approved": false,
  "issues": ["user_id is empty — required param missing. Context shows it should be 8821."],
  "safe_to_call": false,
  "warning": null,
  "corrected_params": { "user_id": "8821", "confirm": true }
}

feedback_log

2026-05-29 | approved:true + correctedparams set — caller used original params anyway | output combination was undocumented; callers didn't know to prefer correctedparams | added to constraints: always use corrected_params when non-null, even when approved

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: 0-uddeshya-0
• Source: 0-uddeshya-0/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.