Noisy "unknown field ignored" warnings on startup from plugin skills and commands

[MrFixit96]

Describe the feature or problem you'd like to solve

Every time I start Copilot CLI, I get 20+ "unknown field ignored" warnings from skills and commands in installed plugins. Most of these fields are either defined in the Agent Skills spec (like license) or widely used across the plugin ecosystem (like argument-hint in commands). The warnings are noisy and there's nothing I can do about them as a user — the fields come from published plugins.

Proposed solution

Unknown frontmatter fields in SKILL.md and command .md files should be silently ignored (or logged at --log-level debug only), not warned to the user on every startup. This is how VS Code Copilot Chat already handles them (see #951).

Specifically:

• Spec-defined fields like license, metadata, and compatibility should be recognized without warnings per the Agent Skills spec. (allowed-tools is already handled correctly — nice work!)
• Command frontmatter fields like argument-hint and hide-from-slash-command-tool — used by claude-plugins-official itself — should be recognized.
• Any other unknown field should be silently skipped for forward-compatibility as the spec and plugin ecosystem evolve.

This is related to #894 (license required when spec says optional) and #951 (metadata as last field breaks skill discovery).

Example prompts or workflows

1. Install claude-plugins-official and hashicorp-agent-skills plugins. Start copilot. Observe 20+ warning lines before the prompt appears — tools, version, license, mcpServers from skills and argument-hint from 15+ command files.

2. A skill with metadata as the last frontmatter field isn't discovered at all (Skills with metadata as last frontmatter field are not discovered #951):

   ---
   name: my-skill
   description: A skill.
   metadata:
     author: me
   ---

3. A command using argument-hint (from GitHub's own claude-plugins-official):

   ---
   description: Create a new SDK app
   argument-hint: [project-name]
   ---

   Warns "unknown field ignored: argument-hint" on every startup.

4. Skills shared across Copilot CLI, Claude Code, and Gemini CLI use the Agent Skills spec as a common format. Harness-specific fields (e.g. tools, mcpServers) from other tools shouldn't produce warnings in Copilot CLI.

5. Plugin authors shouldn't need to maintain separate skill/command files per harness to avoid warnings.

Additional context

• Copilot CLI version: 0.0.414
• OS: Windows 11 (x86_64)
• Related issues: Skill validation requires 'license' in SKILL.md frontmatter (spec says optional) #894, Skills with metadata as last frontmatter field are not discovered #951
• Plugins that trigger warnings: claude-plugins-official (tools, version, license, argument-hint, hide-from-slash-command-tool), hashicorp-agent-skills (mcpServers), leefowlercu-agent-kit (argument-hint in 8+ commands)
• allowed-tools in SKILL.md is now working correctly (observed via thrashr888-agent-kit) — only the command-level variant still warns

[MrFixit96]

Full field analysis

Here's a detailed inventory of every frontmatter field triggering warnings, mapped against the Agent Skills spec and the plugins that use them.

Agent Skills spec field coverage

Field	Spec Status	Copilot CLI Behavior
name	Required	✅ Supported
description	Required	✅ Supported
allowed-tools	Optional (experimental)	✅ Supported in SKILL.md; ⚠️ warned in command .md
license	Optional	⚠️ Warned; previously required (#894)
compatibility	Optional	⚠️ Untested / unknown
metadata	Optional	⚠️ Warned; breaks skill discovery when last field (#951)

Non-spec fields in community use

Field	Used In	Origin
tools	SKILL.md	Claude Code convention — lists allowed tool names
version	SKILL.md	Semantic versioning for skills
mcpServers	SKILL.md	Declares MCP server dependencies
argument-hint	command .md	Shows usage hint for slash commands
hide-from-slash-command-tool	command .md	Hides command from the slash command picker

Complete warning inventory (v0.0.414)

SKILL.md warnings:

Plugin	Skill	Field
claude-plugins-official	claude-automation-recommender	tools
claude-plugins-official	claude-md-improver	tools
claude-plugins-official	example-skill	version
claude-plugins-official	writing-rules (hookify)	version
claude-plugins-official	frontend-design	license
hashicorp-agent-skills	blog	mcpServers

Command .md warnings:

Plugin	Command	Field(s)
claude-plugins-official	new-sdk-app	argument-hint
claude-plugins-official	example-command	argument-hint
claude-plugins-official	feature-dev	argument-hint
claude-plugins-official	hookify	argument-hint
claude-plugins-official	review-pr	argument-hint
claude-plugins-official	cancel-ralph	hide-from-slash-command-tool
claude-plugins-official	ralph-loop	argument-hint, hide-from-slash-command-tool
leefowlercu-agent-kit	gtasks-lists	argument-hint
leefowlercu-agent-kit	gtasks-project-init	argument-hint
leefowlercu-agent-kit	gtasks-project-status	argument-hint
leefowlercu-agent-kit	gtasks-summary	argument-hint
leefowlercu-agent-kit	gtasks-todo-add	argument-hint
leefowlercu-agent-kit	gtasks-todo-complete	argument-hint
leefowlercu-agent-kit	gtasks-todo-list	argument-hint
leefowlercu-agent-kit	gtasks-todo-today	argument-hint
leefowlercu-agent-kit	subsystem-md	argument-hint

No warnings (correctly handled):

thrashr888-agent-kit uses allowed-tools in 9 of 10 SKILL.md files with zero warnings, confirming that field is now supported in skill frontmatter.

Observed progress

Worth noting that the warning message itself improved — it now says "unknown field ignored" rather than just "unknown field", which means the CLI is at least gracefully continuing rather than failing. The remaining ask is to suppress these to debug-level logging so they don't clutter the user's startup experience.

[charris-msft]

Just chiming in to say the lack of support for fields that are in the skills spec is creating a bunch of noise for the Azure skills: