Agent Skills `allowed-tools` frontmatter specs not respected in non-interactive mode

[joellis13]

Describe the bug

GIVEN any generic skill like:

---
name: greet-user
description: use when a user offers a greeting
allowed-tools:
  - bash
  - shell
  - powershell
  - any-other-tool
---

Run the `./scripts/greet.sh`
<!-- (or `./scripts/greet.py`, `./scripts/greet.bat`, `./scripts/greet.ps1`, etc.) -->

WHEN I prompt with "hello" in interactive mode in Copilot
THEN It works as expected and the script runs, but
WHEN I prompt with "hello" in NON-interactive mode (e.g. copilot -p "hello")
THEN I get a ton of Permission denied and could not request permission from user
AND A ton of wasted tokens, time, and money

Affected version

1.0.60

Steps to reproduce the behavior

All in the above description.

Expected behavior

If a tool is marked as allowed, IT SHOULD BE ALLOWED.

Additional context

No response

[BGMLAI]

Non-interactive mode needs a complete preflight. Resolve the skill, its allowed-tools declaration, and the requested capability set before starting the model loop; if any capability cannot be granted without a human prompt, fail once with a machine-readable explanation. Do not repeatedly ask the model to recover from an impossible approval channel. The same normalized policy evaluator should be used in interactive and non-interactive modes, with tests for aliases, unknown tools, nested skills, and a denied tool.