User message lost when slash command fails with 'Unknown skill'

[prokesmartin]

Description

When a user types a message alongside a slash command (e.g. /myskill do something), and the slash command fails with Unknown skill: myskill, the user's original message is not preserved in the conversation history.

The message gets wrapped in a <local-command-caveat> tag and Claude is instructed to ignore it. The user's text is effectively swallowed — neither Claude nor the user can see what was originally typed.

Steps to reproduce

1. Create a new project-level skill during an active session (e.g. .claude/skills/myskill/skill.md)
2. Try to invoke it with /myskill (it won't be recognized until session restart)
3. The terminal shows Unknown skill: myskill
4. The user's message that accompanied the command is lost from the conversation
5. The user cannot scroll back to see what they typed to retry it after restarting

Expected behavior

When a slash command fails, the user's original input should be preserved and visible in the conversation, so they can:

• See what they typed
• Copy/paste it to retry after a session restart
• Or have Claude process the text portion of the message

Actual behavior

The entire message (including any user text) is hidden behind <local-command-caveat> and marked as something Claude should not respond to. The user has no way to retrieve their original input.

Impact

This is especially problematic when:

• A skill was just created and the user needs to restart to pick it up
• The user typed a long or detailed prompt alongside the slash command
• The user wants to retry the same input after fixing the issue

Environment

• Claude Code on Windows 11
• Bash shell (Git Bash)

[github-actions]

Found 3 possible duplicate issues:

1. [BUG] Stashed Messages Get Wiped During Execution #24776
2. [BUG] Claude Desktop: Send button silently fails - message disappears when context is high #18676
3. [BUG] User questions and answers lost from transcript while background subagents are running #27338

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[prokesmartin]

Update: Root cause clarified

After investigation, the Unknown skill error was caused by a skill discovery path issue, not just session caching:

• Claude Code was launched from D:\Repos\EReasoner\ (a worktree root, NOT a git repo)
• The skill was created in D:\Repos\EReasoner\main\.claude\skills\optimize\ (inside the actual git repo subdirectory)
• Other skills in that same path (clitest, review) ARE discovered — but they existed before the session started
• After restarting the session, the new skill was still not found, confirming this is NOT a hot-reload issue

Fix: Copying the skill to D:\Repos\EReasoner\.claude\skills\optimize\ (the CWD-relative path) resolved it.

The underlying UX issue remains valid though: When /myskill fails with "Unknown skill", the user's accompanying message text is completely lost from the conversation. It's wrapped in <local-command-caveat> and hidden from both Claude and the user. The user cannot scroll back to see what they typed, making it impossible to retry after fixing the issue.

---

Regarding the suggested duplicates — none of them match this issue:

• [BUG] Stashed Messages Get Wiped During Execution #24776 — stashed messages during execution (different mechanism)
• [BUG] Claude Desktop: Send button silently fails - message disappears when context is high #18676 — Claude Desktop high-context send failure (different product)
• [BUG] User questions and answers lost from transcript while background subagents are running #27338 — messages lost during background subagent runs (different mechanism)

This is specifically about slash command failure swallowing the user's input text.

[dicksontsai]

We'll work on a way to surface the swallowed text as a system message, not as something Claude should help with directly. This helps users avoid accidentally incurring additional inference costs.