agent-ecosystem/agent-skill-implementation/probe-allowed-tools

Allowed-Tools Probe

The spec's experimental allowed-tools field is "a space-separated string of tools that are pre-approved to run," with support that "may vary between agent implementations." This skill's frontmatter declares a printf-scoped Bash pattern plus Read, then the instructions ask for a printf, so testers can observe whether the field pre-approves anything. This body deliberately never spells out the field's literal value, so any appearance of it can only have come from the frontmatter itself.

Compare against probe-allowed-tools-control, which is identical except it has no allowed-tools field. If execution proceeds for both skills, the platform's general permission posture, not this field, allowed it.

Canary Phrase

The canary phrase for this skill's body is: CURLEW-SCHIST-4419

Instructions

When activated:

1. Report: "probe-allowed-tools activated. Canary: CURLEW-SCHIST-4419"

1. Report whether you can see an allowed-tools value for this skill,

and if so, what it says.

1. Run exactly this command and report its output verbatim, or the exact

error if it was blocked: printf 'GROUSE-%s-9017\n' 'MICA'

1. Report whether the command ran without any permission prompt or

approval step, as far as you can observe.