* ✨ feat: add message queue for agent runtime (soft interrupt)

Implement per-context message queue that allows users to send messages
while the agent is executing. Messages are queued and consumed via two
paths: injected at step boundaries during execution (Path A), or
triggering a new sendMessage after completion (Path B).

- Add QueuedMessage type and queuedMessages state in operation store
- Add enqueue/drain/remove/clear actions and selectors
- Modify sendMessage to enqueue when execAgentRuntime is running
- Add queue checkpoint in step loop (streamingExecutor)
- Add Path B: drain remaining queue after completion → new sendMessage
- Keep input enabled during agent execution (remove isInputLoading guard)
- Add QueueTray component showing "N Queued" above ChatInput
- Add electron-testing skill for agent-browser CLI automation

Fixes LOBE-6001

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 🐛 fix: Path B deferred execution to avoid recursive internal_execAgentRuntime

Use setTimeout(0) to break out of the current execution context when
triggering a new agent runtime for queued messages after completion.
Direct recursive calls caused issues with zustand state batching.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 🐛 fix: remove premature Path A drain, fix Path B with fresh store ref

Path A (step checkpoint injection) was draining the queue before the
last LLM step, leaving nothing for Path B. For agents without tool
calls, this meant queued messages were consumed but never acted upon.

Fix: remove Path A for now (will be re-added for tool-call scenarios),
and use useChatStore.getState() in Path B setTimeout to get a fresh
store reference instead of a stale closure capture.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 📝 docs: update electron-testing skill with real-world patterns

Based on lessons learned during message queue testing:
- Must cd to apps/desktop before npx electron-vite dev
- Use polling loop for startup detection
- snapshot -i -C required for contenteditable (chat input)
- Use sleep + screenshot instead of agent-browser wait for long ops
- Access store via window.__LOBE_STORES.chat()
- Add error interceptor and store inspection patterns
- Document all gotchas (HMR, daemon blocking, fill vs type)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* ✨ feat: add Path A - early handoff to Path B at tool completion

When tools finish and queue has messages, break the step loop early
and let Path B create user message + start new operation. The new
LLM call sees full context including tool results + new user message.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 🐛 fix: Path B use sendMessage for proper message creation

Use sendMessage instead of optimisticCreateMessage + internal_execAgentRuntime.
sendMessage handles the full lifecycle correctly: creates user message
on server, creates assistant message placeholder, and triggers
internal_execAgentRuntime — ensuring both messages are visible in UI.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 💄 style: redesign QueueTray to Codex-style card layout

Each queued message shows as a card with icon, text preview,
and delete button. Uses antd CSS variables for consistent theming.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 💄 style: connect QueueTray with ChatInput as unified container

QueueTray and ChatInput now share a connected border:
- QueueTray has top-rounded corners, no bottom border
- ChatInput gets bottom-only rounded corners when queue has items
- Uses cssVar for proper theme token styling
- Zero gap between tray and input

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* ♻️ refactor: move queue check logic into GeneralChatAgent

Move the "finish early when queue has messages" decision from
streamingExecutor into GeneralChatAgent.runner(). The agent now
checks stepContext.hasQueuedMessages at tools_batch_result phase
and returns finish instruction, which is architecturally cleaner.

- Add hasQueuedMessages to RuntimeStepContext and computeStepContext
- GeneralChatAgent returns finish when tools complete + queue non-empty
- Remove Path A/B labels from comments
- streamingExecutor just passes hasQueuedMessages via stepContext

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* 🐛 fix: forward queued files in sendMessage and drain only on success

- Forward merged file attachments when replaying queued messages
  (sendMessage now receives files from merged queue)
- Move drainQueuedMessages inside the status==='done' branch so
  queued messages are preserved on error/interrupted states
- Add queued_message_interrupt to FinishReason enum
- Add hasQueuedMessages check to tool_result and tasks_batch_result
  phases in GeneralChatAgent (not just tools_batch_result)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* ♻️ refactor: use full operationContext for context key indexing

- operationsByContext index now uses messageMapKey(context) with full
  context (including threadId, scope, etc.) instead of stripped key
- Fixes key mismatch where thread/scoped contexts couldn't find
  running operations, causing overlapping generations
- Move mergeQueuedMessages from services/messageQueue.ts into
  operation/types.ts alongside QueuedMessage type
- Delete services/messageQueue.ts

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>