Explicit slash skill is redundantly reloaded through model registry and fails with “Skill not found”

[pchung13]

Describe the bug

An explicitly invoked interactive slash skill is successfully resolved and expanded into the conversation, but the assistant then attempts to load the same skill through the model-facing skill() tool. For skills excluded from model invocation, that redundant call fails with Skill not found, and the requested workflow does not start.

This differs from #4438: the interactive /skill resolver succeeds and provides the skill context. The failure happens afterward because the assistant routes the already-resolved explicit invocation back through the model-only registry.

Affected version

GitHub Copilot CLI 1.0.79

Steps to reproduce the behavior

1. Create a personal skill at %USERPROFILE%\.agents\skills\implement.

2. Configure agents\openai.yaml:

   interface:
     display_name: "Implement"
     short_description: "Build work from a spec or tickets"
   policy:
     allow_implicit_invocation: false

3. Start an interactive Copilot CLI session using GPT-5.6 Sol.

4. Invoke the skill directly:

   /implement 5514365
   

5. The CLI successfully injects an explicit invocation message and <skill-context name="implement"> containing the skill's base directory and instructions.

6. Instead of following that supplied context, the assistant calls the model-facing tool:

   skill: "implement"
   

7. The tool returns:

   Skill not found: implement
   

The same session shows that skills with allow_implicit_invocation: false are absent from the model-facing registry, while skills without that policy are present. Skills under %USERPROFILE%\.copilot\skills without this policy are also present.

Expected behavior

Once an interactive slash command has resolved a skill and supplied its <skill-context>, the assistant should execute those instructions directly. It should not call the model-facing skill() tool again.

allow_implicit_invocation: false should prevent automatic model selection without breaking a direct /skill invocation that the CLI has already resolved.

Additional context

• OS: Microsoft Windows 10.0.26200, x64
• PowerShell: 7.6.4
• Model: GPT-5.6 Sol
• Related registry/filtering issue: disable-model-invocation: true makes a skill unreachable, not manual-only (Skill not found on explicit invocation) #4438

Debug logs available on request.