Feature: organize Desktop agents into workspaces and improve collaboration

CHANGELOG.md

@@ -8,11 +8,20 @@ submodule.
 
 ## [Unreleased]
 
+### Added
+- **Organize agents into workspaces.** A new upper tab row groups agents into named workspaces. Each workspace remembers its selected agent and split layout; switching workspaces keeps background agents running. Existing tabs move into **Main**, and all workspaces reopen on the next launch. Working and approval badges help you find agents that need attention, including in other workspaces.
+- **Close workspaces directly from their tabs.** The **×** appears on the selected workspace and when another workspace is hovered or focused, without shifting tab labels. Right-click still offers Rename and Close. Both closing actions keep at least one workspace open, block closure while agents or commands are running or approvals are waiting, and save conversations to History before closing.
+- **More ways for agents to collaborate.** Agents can ask for an immediate answer or a background reply, send information, request a read-only review, hand off work, and update or cancel individual jobs. Messages and results appear in expandable transcript entries that identify the agents involved. Background jobs use readable IDs such as `job1`, support bounded waits, and cancel when an associated agent closes. The existing `ask_agent` tool remains available alongside `ask_agent_and_wait`.
+
 ### Changed
-- **Update the bundled engine to the MandoCode 0.16 generation.** Pin the shared engine to `9e25af4`, including the RazorConsole 0.6 migration, maintenance improvements, and merged Windows/Linux CI fixes. The engine now separates commands, requests, lifecycle, model inspection, and history handling into focused code. Desktop keeps its existing WinUI conversation interface; terminal features are described in `MandoCode/docs/CHANGELOG.md`.
+- **Two levels of tabs with matching scroll spacing.** Workspace tabs sit above agent tabs. Split and New agent stay beside the agents they control in the selected workspace. Both rows reserve space beneath their tabs for horizontal scrollbars, and agent mentions show the workspace name alongside the project folder.
+- **Separate Files and Agents mentions.** Typing `@` opens Files first; press Tab to switch to Agents and back. Gold headers explain the switch. Agent rows show name and project folder, files retain explicit paths when names collide, and quoted paths/names support spaces.
+- **Responsive file tagging on large projects.** Reuse the CLI's asynchronous file index, early root listing, shared scan/cache, query debounce, and filesystem watcher. Closing the picker keeps indexing in progress. Build result rows off the UI thread, replace the list in one batch, preserve selection across updates, ignore stale lookups, and show loading/indexing feedback. End drill-down when a selected folder is empty, and dispose file watchers when the tab closes.
+- **Update the bundled engine to the MandoCode 0.16 generation.** Pin the shared engine to v0.16.1 (`277cd36`), including the RazorConsole 0.6 migration, maintenance improvements, and merged Windows/Linux CI fixes. The engine now separates commands, requests, lifecycle, model inspection, and history handling into focused code. Desktop keeps its existing WinUI conversation interface; terminal features are described in `MandoCode/docs/CHANGELOG.md`.
 - **Plan mode starts only through /plan.** Ordinary requests run directly, including multi-step tasks and prose outlines. Desktop no longer automatically proposes structured plans based on task complexity or model tool calls. Existing plan review, editing, execution, and recovery remain available. Settings now explain the explicit command instead of offering automatic planning.
 
 ### Fixed
+- **Stop peer turns from their receiving tab.** Link incoming peer turns to the target agent's normal cancellation action while keeping each background job's cancellation separate. Inbox messages are consumed on peer turns as well as user turns.
 - **Resolve Desktop snapshot types explicitly.** Avoid ambiguous references when both the updated engine and Desktop expose snapshot helpers.
 - **Clear guidance when Ollama retires a model.** HTTP 410 Gone errors, including wrapped provider errors, identify the unavailable model and recommend choosing another with `/model` instead of restarting Ollama or pulling a retired cloud model. The shared engine does not retry these failures.
 

README.md

@@ -61,6 +61,8 @@ Two problems, one app:
 
 ## Features
 
+- **Workspace tabs** — group agents into named workspaces in the upper tab row. Each workspace remembers its selected agent and split layout; switching keeps background agents running. Use **+ Workspace** to create one. Close with the **×** on the selected tab, or hover/focus another workspace to reveal its close button. Right-click also offers Rename and Close. **Split** and **New agent** apply to the selected workspace. Workspaces restore on launch, and existing installations migrate their tabs into **Main**. Alt+Shift+N creates a workspace, Alt+Shift+Left/Right switches, and Alt+Shift+W closes an idle workspace.
+
 - **Agent tabs, and Split view for up to four at once** — each tab is its own independent
   conversation, project folder, and model. Split view puts two, three, or four side by side so you
   can watch several agents work in parallel instead of babysitting one at a time.

docs/agent-mentions.md

@@ -1,5 +1,38 @@
 # Agent mentions — design
 
+## Desktop collaboration tools
+
+The current Desktop host exposes the following interactions. The design history below describes the earlier observation/question/delegation stages.
+
+| Tool | Blocks the requesting turn? | Behavior |
+| --- | --- | --- |
+| `ask_agent_and_wait` | Yes | Return an idle agent's answer in the current turn. `ask_agent` remains a compatibility alias. |
+| `ask_agent_async` | No | Submit a background question; let the user continue chatting. |
+| `send_agent_message` | No | Queue information for the recipient's next turn without requesting a reply or starting a model turn. Works while busy. |
+| `delegate_to_agent` | No | Background work using the target's normal tools and approval gates. |
+| `request_agent_review` | No | Background review; the shared engine's tool middleware enforces read-only access for that turn. Writes, commands, plans, delegation and mutating browser/MCP tools are refused. |
+| `handoff_to_agent` | No | Transfer remaining work and context. The sending agent is instructed to stop duplicating it; tabs are not switched. |
+| `update_agent_job` | No | Requester or assigned agent posts progress or instructions. Instructions enter the recipient's next turn, rather than interrupting current work. |
+| `cancel_agent_job` | No | Requester-only cancellation scoped to that job; already-applied changes remain. |
+| `wait_for_agent_job` | Yes, bounded | Explicit wait, 30 seconds by default and at most 120. Timeout/cancel affects only the wait. |
+| `check_delegations` | No | Read job IDs, kind, state, latest update, live progress and results. |
+
+`list_agents`, `get_agent_status`, and `read_agent_transcript` remain available. All requests resolve the target by name, reject self-calls and preserve the existing call-chain loop/depth limits. Busy targets refuse new work rather than silently queuing jobs. The target claims its actual turn atomically; if it became busy after submission, the job reports failure instead of claiming completion.
+
+Background requests use readable IDs such as `job1`. Cancellation first reports `CancellationRequested`; it becomes `Cancelled` only when the running peer call exits. Terminal results cannot be overwritten by a late completion. Closing an agent requests cancellation of associated jobs. The target's Stop action also cancels peer turns, while a requester stopping its own unrelated turn leaves detached jobs running.
+
+Completions appear as collapsed transcript entries with the full Markdown result available on expansion. Direct messages and job updates have attributed, collapsed entries in both agents' transcripts. Inbox content is consumed on the next user or peer turn, exactly once. Completion notifications do not automatically start another model turn; the requesting model receives the result on its next turn. No Desktop navigation/layout changes are required for these tools.
+
+### Manual checks
+
+- Ask `@OtherAgent` a question in the background, then keep chatting. Expand the completion entry to inspect the full response; ask a follow-up to confirm inbox delivery.
+- Compare an explicit blocking question with an async question. Try a busy target and confirm a clear refusal/failure.
+- Request a read-only review that attempts a file write or command. Confirm refusal; ordinary delegated work retains normal approval dialogs.
+- Send information to a busy agent. It should get a transcript notice immediately and context on its next turn without a competing model turn.
+- Start separate jobs and cancel one. Confirm cancellation acknowledgement and that unrelated work continues. A timed-out wait must leave the job running.
+- Close an agent involved in a job and confirm the job stops. Verify Stop also works in the receiving tab during a peer turn.
+
+
 Status: **built**, except `review_agent_work`. See the build order below.
 
 ## What it is
@@ -122,36 +155,13 @@ Open problems, all of which belong to this tool alone:
 
 ## The `@` picker
 
-`ChatTabView.Input.cs` already implements `@` for project files: it walks back from the caret to the
-token start, and if the token opens with `@` it filters through `FileAutocompleteProvider` and calls
-`ShowSuggestions(SuggestMode.File, ...)`.
-
-Agents join that same picker and **rank above files**. Rationale:
-
-- Open agents are a small, closed, known set; project files are thousands. A short list on top costs
-  the file case almost nothing.
-- Callsigns are capitalised single words, so genuine collisions with real filenames are rare, and
-  the picker disambiguates the rare ones visually.
-- `@` already means "a participant" to anyone who has used Slack. A second sigil would be a thing to
-  learn for no benefit.
-
-Implementation notes:
-
-- A new `SuggestMode.Agent`, because `AcceptSuggestion` branches on the mode and an agent mention
-  substitutes differently from a file path — no trailing `/` drill-in behaviour, and the accepted
-  text should be the callsign, not a path.
-- The agent's own tab must be excluded from its own picker.
-- Suggestion rows want a distinguishing glyph and a subtitle (folder name, or busy/idle), so an
-  agent row never reads as a file row.
-- **The host must also learn about mentions, not just the picker.** An earlier draft of this doc
-  claimed the picker was an affordance and no host-side parsing was needed. That was wrong, and
-  testing found it immediately: `ChatController.ProcessFileReferences` expands every `@token` at
-  send time and warns `Couldn't find the referenced file or folder: Ninja` when the lookup misses.
-  A mention never reached the model at all. Agent names must be resolved *before* the file lookup
-  and skipped by it.
-- Agents win over files on a name clash. A callsign is a deliberate act of addressing someone; a
-  same-named file is a coincidence, and that file stays reachable by any path carrying a separator
-  or an extension.
+Typing `@` opens **Files — Press [Tab] for Agents**. Press Tab to show **Agents — Press [Tab] for Files**, and press it again to switch back. Files and agents stay in separate lists. Typing continues in the input, arrows navigate, Enter accepts a file or agent, and Escape dismisses. Tab still accepts command and emoji suggestions. Shift+Enter retains the input's multiline behavior. The current agent is excluded; agent rows show the callsign and project folder without a busy/idle suffix.
+
+File suggestions use the engine's background index rather than synchronous recursive enumeration. Root entries publish before descendant scanning finishes. One scan continues across typing, switching views, and closing/reopening the popup. Filtering is debounced by the shared provider, and UI result projection runs off the dispatcher. A virtualized, height-bounded ListView receives a single result array, without a match limit or thousands of individual collection notifications. Refreshes retain the selected path; stale/cancelled UI requests cannot reopen a dismissed picker or replace a newer query. Loading and indexing feedback stay outside the selectable rows.
+
+After indexing, the provider applies queued create/delete/rename events incrementally when lookup resumes. Watcher overflow invalidates the cache for recovery. Project changes start the appropriate index and old results are discarded. Closing a tab unsubscribes the UI and disposes its provider/watcher. This is an in-memory cache for the agent's lifetime, not a persistent cross-process index; initial full indexing still takes work on huge trees.
+
+Selecting a folder drills into its immediate children; an empty folder ends drill-down and leaves the selected directory in the input. Paths with spaces use quoted reference tokens. A file whose name resolves to an agent is inserted with `./` so it stays a file reference. The host parses the same token format when assembling file context and agent mentions.
 
 ## Who am I talking to