Add custom agents: creation, : trigger, CLI agent compatibility, and Plan mode abort fix

[arnavv-guptaa]

1. Custom Agent Creation (/create-agent)
   New slash command registered in builtin-commands.ts
   CreateAgentForm component with multi-step wizard (name, scope, tools, model, description)
   Generates a prompt and sends it to the active chat session to create the agent file

2. ":" Trigger for Agents
   Typing : in the chat input opens the agents dropdown directly (skips file/skill categories)
   Supports search filtering (:work filters to matching agents)
   Selecting an agent inserts a mention pill and removes the trigger text
   @ followed by a space now dismisses the dropdown

3. CLI Agent Compatibility
   Added parseFrontmatterFallback() in agent-utils.ts — a regex-based line-by-line parser for when gray-matter fails on complex YAML (unquoted colons, literal \n, XML tags in descriptions)
   Added a check: if gray-matter "succeeds" but returns an empty description, fall through to the fallback parser (gray-matter has inconsistent behavior across calls)
   Tooltip height limited with line-clamp-3 for large agent descriptions

[serafimcloud]

Nice feature set! The /create-agent wizard and : trigger are well thought out.

A few observations:

1. Token retrieval duplication
getClaudeCodeToken() in agents.ts duplicates the one in claude.ts. Would it make sense to extract this to a shared utility?

2. Agent generation flow
The agents.generate mutation exists but handleCreateAgentSubmit sends a prompt to chat instead. Is there a reason for preferring the chat-based approach over the direct API call?

3. Space after @ closes dropdown
The check afterAt.startsWith(" ") will close the dropdown when typing @ — this differs from the previous behavior that allowed multi-word searches. Was this intentional?

4. Project agents in Settings
The Settings UI now only shows user agents (filtering by source === "user"). Should project agents still be visible there, perhaps in a separate section?

5. Plan mode abort removal
The comment explains the rationale well. Just confirming: with the abort removed, does the Claude process terminate naturally via planCompleted flag, or does it continue running in the background?

[arnavv-guptaa]

1. Token retrieval duplication

You're right that getClaudeCodeToken() is duplicated in both agents.ts (line 33) and claude.ts (line 109). I kept it separate to avoid touching unrelated files in this PR, but it definitely makes sense to extract it to a shared utility.

2. Agent generation flow

The chat-based approach was intentional. When the agent is created through the chat interface, the user can see exactly what was generated in the expandable tool steps section. It also allows for quick iteration - if they want to tweak the agent, they can just send a follow-up message. The direct API call would work too, but the chat flow felt more transparent and interactive.

3. Space after @ closes dropdown

Yes, this was intentional. Previously, if you typed '@' and kept typing, the dropdown would keep searching even when you weren't trying to mention a file - which felt disruptive. With the current behaviour, '@file name' still works perfectly as long as there's no immediate space after '@'. If you type '@ ' (with a space), it assumes you're just using the @ symbol normally.

4. Project agents in Settings

Showing project-specific agents in the global Settings felt a bit odd to me, so I opted to only show user-level agents there. My thinking was that project-related agents, skills, etc. could live in the Project tab in settings instead. We could also show all agents in one place, though it might get cluttered over time if users have many projects.

5. Plan mode abort removal

Yes, the Claude process terminates cleanly. When 'ExitPlanMode' completes, we set a 'planCompleted' flag and break out of the stream loop. JavaScript automatically triggers cleanup when you break from an async iterator - this is part of the language spec (the engine calls '.return()' on the iterator). The session is then removed from our tracking map in the 'finally' block.
This fix has already been addressed in PR#68.