Options passed to a custom skill invocation are dropped before reaching the model

Describe the bug

When invoking a custom skill using /skill-name --option value, Copilot CLI records the complete input but omits the option arguments from the prompt dispatched to the skill.

Environment:

• GitHub Copilot CLI 1.0.81-6
• Windows
• Session: 2da2d958-a8cd-4eeb-9760-241fac5b53ed
• Occurred around 2026-08-21 04:10:27 UTC

The session event log contains the complete command and arguments. However, the model-visible invocation contains only that the skill was explicitly invoked plus its SKILL.md instructions. Because the arguments are absent, the skill follows its no-arguments behavior.

Affected version

1.0.81-6

Steps to reproduce the behavior

1. Define a repository skill that supports command-line-style arguments and has different behavior when no arguments are supplied.
2. Invoke it with arguments, for example:
   /evaluate-meeting-prep --slides week-over-week,speed,tokens --current "Model=C:\path\to\cohort"
3. Observe that the skill loads successfully.
4. Observe that the skill behaves as if no arguments were supplied and displays its interactive slide-selection menu.
5. Inspect the session event log: the original user.message contains the complete command, including --slides and --current, but those arguments are absent from the prompt delivered to the model with the skill context.

Expected behavior

When a custom skill is invoked as /skill-name <arguments>, the complete argument text should be included in the model-visible skill invocation. In this example, the skill should receive --slides week-over-week,speed,tokens and the supplied cohort paths, then proceed without asking the user to select slides again.

Additional context

• Copilot CLI: 1.0.81-6
• Operating system: Windows x64
• Shell: PowerShell
• Session ID: 2da2d958-a8cd-4eeb-9760-241fac5b53ed
• Timestamp: 2026-08-21 04:10:27 UTC

Diagnostic detail: the session's raw user.message event contains the complete 521-character invocation, including --slides, --current, --prior, and --outdir. The subsequent model-visible skill invocation contains the loaded skill instructions but none of those options. This indicates the options are lost during skill dispatch rather than input parsing.

在 Windows PowerShell 上复现 GitHub Copilot CLI 中自定义 skill 的调用，然后将原始的 user.message 事件与模型可见的 skill 调用进行比较。跟踪 skill 的 dispatch 路径，找出选项文本在哪个位置被丢弃；完成的标准是完整参数（包括 --slides、--current、--prior 和 --outdir）传递到模型，并且 skill 跳过其无参数行为。