`skillDirectories` setting from `settings.json` is not honored in ACP mode (`copilot --acp`)

Describe the bug

~/.copilot/settings.json supports a skillDirectories array for registering
additional custom skill source directories. These directories' skills are
correctly discovered and listed when running the interactive copilot CLI,
but do not show up when the same user/config runs in ACP mode
(copilot --acp, e.g. via the Zed editor's ACP integration).

Affected version

Copilot CLI 1.0.82

Environment

• OS: Windows
• ACP client: Zed editor (copilot --acp)
• Windows Terminal: copilot.exe --acp

Steps to reproduce

1. Add a skillDirectories entry to ~/.copilot/settings.json, e.g.:

   {
     "skillDirectories": [
       "C:\\Users\\<user>\\ai_space\\skill_groups\\workflow_skills",
       "C:\\Users\\<user>\\ai_space\\skill_groups\\shorthand_skills",
       "C:\\Users\\<user>\\ai_space\\skill_groups\\ddstack_skills"
     ]
   }
   
   

Each directory contains one or more <skill-name>/SKILL.md skills.

2. Run interactive copilot, then copilot skill list (or /skills list).
   Result: skills from all skillDirectories entries appear under a
   "Custom skills:" heading, alongside project (.github/skills/) and
   personal (~/.copilot/skills/) skills. Confirmed working — e.g.
   workflow-design, shorthand-crew, principle-fix-root-causes, etc. all
   listed (43 custom skills total in our test).

3. Independently, spawn copilot.exe --acp as a subprocess and drive it
   directly over stdio with raw JSON-RPC (bypassing any client like Zed
   entirely, to rule out client-side filtering):

   • initialize → session/new (with the same cwd) → session/prompt
     asking the agent to "list every skill name available to you right now,
     grouped by category, without invoking any tool."

   Result: the agent's response only lists project (12 skills),
   user/personal (5 skills), and builtin (2 skills) categories. The
   custom category empty, none of the 43 skills sourced from skillDirectories, does
   not appear in its own self-reported system context.

Expected behavior

ACP mode should discover and expose the same skill set as interactive mode,
including all directories configured via skillDirectories in
settings.json, since both entry points read the same global config file.

Actual behavior

ACP mode does not surface any skillDirectories-sourced skills, with no
error, warning, or other indication that the setting was skipped.

Additional context

• settings.json and config.json are unmodified between the two runs —
  same file, same user, same machine.
• Personal (~/.copilot/skills/) and project (.github/skills/) skills
  do load correctly in ACP mode, so this isn't a total skill-loading
  failure — it's specific to the skillDirectories custom-path mechanism.
• Reproduced twice: once via a live Zed → copilot --acp session, and once
  via a standalone hand-rolled ACP client script, ruling out a
  client-specific (Zed) cause.

从 copilot --acp 入口开始，将其技能发现与交互式 copilot 进行比较，并使用 ~/.copilot/settings.json 中的 skillDirectories 设置作为输入。使用 initialize、session/new 和 session/prompt 进行复现；当 ACP 将已配置的自定义技能与项目技能、个人技能和内置技能一同公开时，即表示完成。