feat(chat): clarify new conversation Agent choice

docs/features/new-conversation-agent-selection/plan.md

@@ -0,0 +1,137 @@
+# Implementation Plan
+
+Objective: implement the selection and ownership contract in [spec.md](./spec.md).
+
+- [x] Separate Agent filtering from execution selection; unify new-conversation entry behavior
+      around the existing draft Agent and preserve explicit workspace intent.
+- [x] Add the composer Agent selector, remove the welcome-selection route and unused artifacts,
+      and preserve model/default/ACP lifecycle and manually selected workspaces.
+- [x] Review the complete change and update the smallest durable regression coverage for the
+      observable behavior, including affected existing E2E helpers.
+- [x] Run format, i18n, lint, typecheck, relevant renderer tests, and Electron checks with multiple
+      Agents from an isolated copy of the real local profile. Remove temporary probes.
+- [x] Commit and open a PR against `dev`, explain ownership and selection rules with BEFORE/AFTER
+      ASCII layouts, include validation evidence and `Closes #2462`, and attach the PR to this chat.
+
+
+## Validation
+
+- Format, i18n, lint, typecheck, and the production build passed.
+- 377 relevant renderer tests passed; 6 Electron smoke tests passed, including keyboard selection,
+  history navigation followed by New Chat, and first-message assignment to a selected custom Agent.
+- An isolated copy of the local real profile passed checks for six enabled Agents: three DeepChat
+  Agents and pi ACP, DimCode, and Claude Agent. Selection matched created session Agent/model
+  identities; filtering preserved the active conversation; ordinary and actual workspace-plus
+  entry points preserved the new-conversation choice. Wide/narrow and light/dark layouts were
+  visually checked. Sync, remote control, and scheduled jobs were disabled in the copy.
+- The real-profile checks created empty sessions and prepared configured Agents without requesting
+  live model completions. First-message execution used the Electron smoke suite's loopback fixture.
+- Normal-build provider and ACP catalog refreshes are retained.
+
+PR: https://github.com/ThinkInAIXYZ/deepchat/pull/2469 (base: `dev`; closes #2462).
+
+
+## ACP Preparation Verification
+
+- [x] Preserve warmup feedback in the composer selector and report unavailable selected workspaces;
+      keep ACP identities outside local Skill routes and discard pending native Skill results on switches.
+- [x] Distinguish automatic Agent workspace defaults from manual choices; keep the active execution
+      context unchanged on failed deactivation.
+- [x] Validate loading/configuration readiness, invalid-to-valid workspace recovery, and route errors
+      through Electron and isolated copies of the real local Agent configuration.
+
+
+ACP verification: format, i18n, lint, typecheck, and the production build passed; 332 focused
+renderer tests and four Electron smoke tests passed. The ACP stdio fixture holds initialization
+open to verify the selector's loading state, then supplies model/mode configuration and exercises
+invalid-to-valid workspace recovery while checking native Skill route errors.
+
+The local real-profile snapshot verified unavailable-directory behavior for pi ACP and DimCode,
+with no `DeepChat Agent not found` errors. DimCode completed actual draft preparation and displayed
+its model/mode/permission controls with an existing directory. Pi ACP could not complete Session
+creation: its default `pi` command resolves to the locally installed `@simon_he/pi` package manager
+(version 0.2.11), whereas the adapter requires the Pi coding Agent. No global installation or real
+profile setting was modified, and no live model completion was requested. Verification builds and
+Electron runs used a temporary worktree so the running development app retained its own output.
+
+## ACP Option Menu Verification
+
+- [x] Keep inline option menus at a compact 20rem width within the viewport; keep options on one line with
+      ellipsis and show complete values on hover or keyboard focus using the existing tooltip.
+- [x] Verify vertical scrolling, long unbroken IDs, selection, disabled selected-item tooltips,
+      Escape/focus return, and light/dark layouts through Electron at 1280px and 480px widths.
+- [x] Validate the real DimCode catalog from an isolated local profile at 1280px and 360px widths.
+      All 412 options stayed on one line with no horizontal scrolling; constrained values showed
+      ellipsis and a complete tooltip. No real model completion was requested.
+- [x] Keep the composer toolbar on one row, shrink Agent/model labels, and reserve attachment/send
+      actions at the edges. Only the central settings region scrolls when its minimum widths cannot
+      fit; focused controls are revealed automatically. Validate full ACP status/configuration rows
+      at 1280px, 800px, and 480px and real DeepChat/DimCode configurations at the same widths.
+- [x] Keep Agent and model names visible, with the Agent selector beside the workspace below the
+      composer and model before mode in its toolbar. Reveal other icon controls on
+      hover, keyboard focus, and while their menu is open; preserve warmup feedback and click
+      behavior. Keep revealed targets stable during pointer traversal and disable
+      transitions for reduced-motion preferences. Validate compact, expanded, keyboard, and motion
+      behavior through Electron and the isolated local DeepChat/DimCode configuration.
+
+- [x] Resolve model-picker icons from model IDs for selected values and menu entries, prefer model
+      names over hosting namespaces, and preserve provider icon/API-type fallback. Validate known
+      ACP model icons and generic unknown-model fallback through the existing status bar coverage.
+
+Layout checks: format, i18n, lint, typecheck, production compilation, 120 composer component tests,
+and four focused Electron smoke tests passed. The normal-build provider catalog refresh is retained.
+
+Real DeepChat/DimCode checks at 1280px, 800px, and 480px confirmed visible Agent/model names,
+workspace-row placement, model-before-mode ordering, 32px compact controls, one toolbar row, and
+intermediate animation frames during expansion. Reduced-motion and menu/keyboard behavior are
+covered by the ACP Electron fixture. No live model completion was requested.
+
+Model icon checks: 98 focused status-bar/icon component tests and two Electron smoke tests passed.
+Selected values and menu rows use DeepSeek, Gemini, and GLM icons from model IDs; an unknown model
+keeps the custom OpenAI-compatible provider fallback. The ACP smoke explicitly focuses its native
+window for hover checks and dismisses nested tooltips before testing Escape/focus return.
+
+An isolated copy of the real local configuration verified loaded DeepSeek, Qwen, GLM, and GPT icons
+for custom OpenAI-compatible providers. Provider icons retain their API-type behavior. No live
+model completion was requested.
+
+## Accessibility and Icon Fallback Verification
+
+- [x] Make context usage keyboard-focusable and keep focused controls visible after label expansion.
+      Preserve readable values on disabled toolbar controls.
+- [x] Give disabled ACP option rows focusable tooltip anchors while retaining native disabled
+      actions and suppressing action hover styling.
+- [x] Match model icons only against the final ID path segment, preserve unknown-model provider
+      fallback, remove the unused Agent slot, and keep one namespaced icon case in the workspace
+      Electron fixture.
+
+Format, i18n, lint, typecheck, production compilation, 110 focused component tests, and two Electron
+smoke tests passed. The ACP smoke covers keyboard context reveal, focus visibility after expansion
+at narrow widths, visible values during a held configuration save, and disabled selected-row
+tooltips on hover and focus. The existing CI E2E selection is unchanged.
+
+A WAL-aware snapshot of the real local profile verified DeepSeek, Qwen, GLM, and GPT model icons,
+keyboard context reveal in an existing DeepChat conversation, and hover/focus tooltips on DimCode's
+disabled selected model row. Its action remained disabled. No live model generation was requested.
+
+## ACP Chats Execution
+
+- [x] Prepare the managed Chats directory on demand through a typed project route and a versioned
+      snapshot. Keep manual selection and the custom default project unchanged. Recover the configured
+      managed directory before other candidates and persist any required managed fallback.
+- [x] Use the effective Chats directory consistently for warmup, draft creation, authentication
+      continuation, workspace references, and submission. Persist the directory in ACP sessions so
+      follow-up messages pass the existing workdir validation.
+- [x] Prepare only missing effective Chats directories, including after a snapshot loses the path.
+      Keep a focusable retry icon with the failure reason until preparation succeeds.
+- [x] Exclude ACP identities from native runtime Skill validation. Give disabled option anchors
+      button semantics and disabled state, and locate them by their accessible role in Electron.
+      Check visible loading cues while retaining the parent's exposed loading contract test.
+
+Full verification passed: 9,856 main tests, 2,750 renderer tests, formatting, i18n, lint, typecheck,
+and production compilation. The ACP stdio smoke clears the selected workspace, checks the actual
+`session/new` cwd, retries a failed preparation through its visible control, and completes two chat
+turns without native Skill scope errors. An isolated real
+profile verified DeepChat and DimCode with Chats selected, ready configuration, and an enabled Send
+button after entering text while preserving a separate custom default project. No live model
+completion was requested.

docs/features/new-conversation-agent-selection/spec.md

@@ -0,0 +1,120 @@
+# New Conversation Agent Selection
+
+## Goal
+
+Closes https://github.com/ThinkInAIXYZ/deepchat/issues/2462.
+
+The sidebar Agent selector filters session history only. The new-conversation context row displays
+and selects its execution Agent independently, beside the workspace selector. Every generic New Chat
+entry point uses the same selection rule.
+
+## Ownership And Behavior
+
+- `agentStore.filterAgentId` owns history filtering. Changing it neither closes the active session
+  nor changes the route, execution Agent, composer draft, or selected workspace.
+- `draftStore.agentId` remembers the current window's new-conversation Agent, initially `deepchat`.
+  Only explicit selection and new-conversation initialization update it; default-configuration
+  watchers and browsing existing sessions do not overwrite it. No new persisted setting is required.
+- `agentStore.selectedAgentId` remains the execution-context projection consumed by existing
+  renderer controls. Entering New Thread restores it from the new-conversation draft. Existing
+  sessions retain their own durable Agent assignment.
+- Ordinary New Chat, keyboard/menu commands, the collapsed header action, workspace plus actions,
+  and empty workspace activation use the draft Agent. Workspace entry points only supply a
+  one-shot project-directory intent, including an explicit null intent for Chats.
+- Explicit Agent commands may select an Agent through the same new-conversation action without
+  modifying history filtering. Provider start deeplinks retain their explicit DeepChat target.
+- A deleted or disabled draft Agent falls back to the enabled built-in DeepChat Agent. Agent list
+  ordering and workspace/session history never choose an execution Agent.
+- New Thread always opens the composer. Remove the separate Agent welcome-selection page and its
+  unused translations and tests. Agent management remains available through existing Settings.
+
+## Composer Interaction
+
+The composer offers the existing enabled Agent catalog, using AgentAvatar and the existing
+dropdown primitives. Agent and model names remain visible with their icons. Selection is disabled
+during submission. Model and ACP controls continue to use the selected execution context.
+
+The Agent selector sits beside the workspace below the composer; its loading indicator reflects
+the existing ACP configuration-loading state. The composer starts with model, mode, then other ACP
+options. Agent and model never collapse. Mode,
+permission, reasoning, context, status, and settings controls default to icons. Hover reveals their
+current values or labels using the existing motion tokens. Revealed controls remain expanded while
+the pointer moves within the toolbar or a menu is open, keeping targets stable. Keyboard focus also
+reveals the value; open menus keep their trigger expanded. ACP loading feedback stays visible in
+the compact state. Disabled controls keep their values visible, and context usage receives keyboard
+focus. Focused controls scroll into view after expanding. Disabled option rows expose complete
+tooltips through a focusable wrapper without enabling their actions. Reduced-motion preferences
+disable the reveal transition. Click behavior and menu contents retain their existing ownership.
+
+Model-picker triggers and entries use the existing substring matcher on the final model-ID path
+segment; hosting namespaces do not supply a model icon. Unknown models fall back to the provider
+icon. Provider entries retain their existing API-type icon behavior, and ACP models use known
+model-family icons while retaining the generic model icon for unknown IDs.
+
+Switching Agents uses existing per-Agent composer drafts and default-model/configuration loading.
+It preserves manually chosen workspaces, including an explicitly cleared workspace. Stale default
+or ACP preparation results must not overwrite newer choices. Existing session Agents are immutable
+through this control; it appears only on New Thread.
+
+```text
+BEFORE
+Sidebar Agent -> history filter + execution selection
+New Thread   -> Agent welcome -> composer with model control
+
+AFTER
+Sidebar      -> history filter
+New Thread   -> [message composer                     ]
+               [model v] [mode / permission icons] [send]
+               [workspace v] [Agent v]
+```
+
+## ACP Preparation And Scope
+
+The composer selector displays the existing ACP configuration-loading state until warmup completes.
+Chats do not display a directory-selection prompt. ACP Chats lazily prepare an application-managed
+chat directory, independently of the custom default project. The visible selection remains Chats;
+its ACP session records the managed directory so first and subsequent turns have a valid workdir.
+The directory follows the existing Documents/Home/user-data candidate policy and remains in Chats
+history grouping even when the user has a different custom default project. A configured managed
+directory is recreated before trying other candidates. If it cannot be recreated, the selected
+managed fallback becomes the configured default so restoring the original directory cannot move
+new Chats back to a different location. Custom defaults are preserved.
+The typed `project.ensureChatWorkspace` route publishes a versioned project snapshot. Preparation
+runs only when ACP needs a missing effective directory, including when a later snapshot loses it.
+Failure leaves a retry icon beside the context selectors, with the reason available on hover or
+keyboard focus; it does not add a directory-selection prompt.
+An unavailable explicitly selected workspace still displays its localized
+reason and blocks ACP preparation/submission. Native DeepChat Chats with an explicit null workspace
+keep their general tool profile; selecting the managed directory explicitly still uses the code
+profile, as for any other selected project. Native
+Skill catalog, active-Skill requests, and runtime Skill validation only run with a DeepChat Agent
+scope. ACP selection has no
+local Skill scope, and late native Skill responses cannot populate ACP composer state.
+
+Inline ACP option menus use a compact 20rem width, constrained by the viewport.
+Options remain on one line with an ellipsis for overflowing values. Hover and keyboard focus show
+the full value in a viewport-constrained tooltip; long lists scroll vertically without horizontal
+scrolling. The same menu serves new and existing conversations.
+
+The composer toolbar keeps attachment, Agent/model configuration, status, and send controls on one
+row. Expanded labels use ellipsis when constrained while action buttons retain usable widths. When the
+minimum control widths cannot fit, only the central settings region scrolls horizontally; attachment
+and send remain visible at the edges, and keyboard focus can reveal every settings control.
+
+Workspace selection distinguishes Agent defaults from manual choices. An automatic Agent directory
+uses the `agent` source, follows the next Agent's default (or the global default when absent), and
+survives unrelated project snapshot refreshes. Only manual selection, including an explicit clear,
+blocks automatic Agent defaults. A failed active-session deactivation preserves its execution Agent.
+
+## Compatibility And Acceptance
+
+No database migration, persisted setting, or provider configuration change is required. A typed
+project route prepares the existing managed chat directory through the preload bridge.
+An active Agent B conversation remains open while the sidebar filters Agent A. Creating another
+B conversation retains that filter, so B's new history row is absent from the A-only list.
+
+Acceptance requires consistent Agent assignments across new-conversation entry points, unchanged
+filtering during composer selection, direct composer startup, preservation of per-Agent drafts and
+explicit workspace choices, and matching visible/created Agent identities. Validate multiple
+configured Agents using an isolated copy of the local real profile as well as durable regression
+checks; never mutate the live database through an external tool.