Phase 21: Intent-first pipeline, visible ack, KV-stable system prompt

Pipeline changes (generation_task.py, intent.py):
- Remove optimistic streaming queue/race (_drain_queue deleted)
- Remove _generate_acknowledgment — ack now embedded in intent JSON
- Round 0: await intent (~400ms), stream ack immediately as TTFT,
  then execute tool sequentially; chat-only streams directly
- IntentResult.ack: one-sentence acknowledgment, intent max_tokens 200→350
- _parse_intent extracts and trims ack field

KV cache stability (llm.py, generation_buffer.py, generation_task.py):
- build_context: replace cached_note_ids with include_note_ids
- Auto-found notes populate context_meta["auto_notes"] for sidebar but
  are NOT injected into system prompt (--- Related Notes --- removed)
- Explicitly included notes injected as --- Included Notes ---
- _conv_note_cache dict + get/set/clear functions removed from generation_buffer.py
- All clear_conv_note_cache() calls removed

Cold model retry (llm.py):
- generate_completion (used by classify_intent) retries on HTTP 500:
  3 attempts with 3s/6s delays — prevents intent failure during cold load

API + frontend (routes/chat.py, stores/chat.ts, views/ChatView.vue, components/ChatPanel.vue):
- exclude_note_ids → include_note_ids throughout
- ChatView sidebar: Suggested (auto-found, + to include) + In Context (× to remove)
- ChatPanel: remove exclude button from context pills; no IDs passed to sendMessage

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

summary.md

@@ -12,7 +12,7 @@
 >    Include file-level details in the commit body when the change is non-trivial.
 
 ## Last Updated
-2026-02-26 — Phase 20: Dedicated tag field (chip input), tags no longer extracted from body
+2026-02-26 — Phase 21: Intent-first pipeline, visible acknowledgment, KV-stable system prompt
 
 ## Project Overview
 Fabled Assistant is a self-hosted note-taking and task-tracking application with
@@ -75,11 +75,15 @@ for AI-assisted features.
   reconnection support. Frontend uses `fetch()` + `ReadableStream`. Simpler than
   WebSockets; Quart supports async generators natively for SSE.
 - **Context building server-side:** Backend fetches URL content and searches notes —
-  frontend just sends the message text + optional note ID + optional exclude list.
+  frontend sends the message text + optional context note ID + optional `include_note_ids`.
   Keyword extraction uses simple word splitting with stopword filtering (no embeddings).
   `build_context()` returns `(messages, context_meta)` tuple; metadata includes
   auto-found note IDs/titles sent to frontend via SSE `context` event before streaming.
   Multi-word search splits terms into per-word ILIKE with AND logic (not adjacent match).
+  Auto-found notes populate the sidebar but are **not** injected into the system prompt
+  automatically — users must click `+` in the sidebar to include them (stable system
+  prompt prefix enables Ollama KV cache reuse). Explicitly included notes appear as
+  `--- Included Notes ---` in the system prompt.
 - **Reverse proxy required for production:** The app does not terminate TLS. A
   reverse proxy (Nginx/Traefik/Caddy) must sit in front of port 5000. Do not
   expose the app directly to the internet.
@@ -283,8 +287,8 @@ fabledassistant/
 │       │   ├── llm.py           # Ollama interaction: build_context with user_id, streaming (stream_chat + stream_chat_with_tools), ChatChunk dataclass, URL fetching; uses Config.OLLAMA_NUM_CTX for KV cache window
 │       │   ├── chat.py          # Conversation CRUD with user_id isolation, add_message, save/summarize as note (LLM-titled, chat-tagged)
 │       │   ├── generation_buffer.py # In-memory SSE event buffer with cancel_event, reconnect support, auto-cleanup; supports chat (int keys) and assist (string keys)
-│       │   ├── generation_task.py   # Background asyncio tasks: run_generation (chat, DB flush, titles, optimistic streaming + intent racing + tool loop) + run_assist_generation (lightweight, no DB); _drain_queue() async generator helper
-│       │   ├── intent.py          # Intent routing: classify_intent() makes fast non-streaming LLM call to detect tool intent before streaming
+│       │   ├── generation_task.py   # Background asyncio tasks: run_generation (chat, DB flush, titles, intent-first pipeline + tool loop) + run_assist_generation (lightweight, no DB)
+│       │   ├── intent.py          # Intent routing: classify_intent() makes fast non-streaming LLM call; IntentResult has ack field (one-sentence acknowledgment streamed as TTFT)
 │       │   ├── tools.py            # LLM tool definitions (create/delete note+task, update_note w/tag management, get_note, list_notes, search_notes w/type filter, list_tasks, full CalDAV suite incl. search_todos) + execute_tool dispatcher
 │       │   ├── tag_suggestions.py # LLM-powered tag suggestions: suggest_tags() builds prompt with existing tags, calls generate_completion, parses JSON response
 │       │   ├── caldav.py          # CalDAV integration: full event lifecycle (create/list/search/update/delete), todos (create/list/search/update/complete/delete), list_calendars, timezone (ZoneInfo), reminders (VALARM), attendees, multi-calendar search
@@ -317,7 +321,7 @@ fabledassistant/
     │   │   ├── auth.ts          # Auth state: user, isAuthenticated, isAdmin, oauthEnabled, localAuthEnabled, login/register/logout/checkAuth/checkHasUsers
     │   │   ├── notes.ts         # CRUD + tag filter, resolveTitle, convertToTask, convertToNote, fetchBacklinks, fetchAllTags (with toast errors)
     │   │   ├── tasks.ts         # CRUD + status/priority filter, patchStatus (with toast errors)
-    │   │   ├── chat.ts          # Conversation CRUD, sendMessage (SSE streaming), status polling (memory-leak-safe _pollUntilLoaded), running models, model warming, updateConversationModel (with toast errors)
+    │   │   ├── chat.ts          # Conversation CRUD, sendMessage (SSE streaming, includeNoteIds param), status polling (memory-leak-safe _pollUntilLoaded), running models, model warming, updateConversationModel (with toast errors)
     │   │   ├── settings.ts      # App settings: assistantName, defaultModel, installedModels, defaultChatModel, defaultIntentModel, pullModel, deleteModel (with toast errors)
     │   │   └── toast.ts         # Toast notification state (success/error/warning), 4s auto-dismiss, dismiss(id)
     │   ├── types/
@@ -342,7 +346,7 @@ fabledassistant/
     │   │   ├── RegisterView.vue # Register form with password confirmation; shows "closed" message when registration disabled
     │   │   ├── RegisterInviteView.vue # Invitation-based registration: validates token, creates account with pre-set email
     │   │   ├── UserManagementView.vue # Admin user management: registration toggle, invitations (send/revoke), user list with delete
-    │   │   ├── ChatView.vue     # Dedicated /chat page: responsive sidebar (overlay on mobile), bubble messages, note picker, persistent context sidebar (right panel, hidden mobile), model selector in header
+    │   │   ├── ChatView.vue     # Dedicated /chat page: responsive sidebar (overlay on mobile), bubble messages, note picker, context sidebar with “In Context” (user-included, ×) + “Suggested” (auto-found, +), model selector in header
     │   │   ├── HomeView.vue     # Chat-first dashboard: quick actions + chat widget (top, full-width), inline response panel, two-column grid (3fr tasks / 2fr notes); task sections: Overdue, Due Today, Due This Week, High Priority, In Progress, Other (capped 10, due-dated first); 8 recent notes; model warming on mount
     │   │   ├── SettingsView.vue # Settings page: assistant name, chat/intent model dropdowns (populated from installed models), email change (with password confirmation for local-auth users), change password, notifications, CalDAV, SMTP (admin), base URL (admin), data export/restore (admin)
     │   │   ├── NotesListView.vue # Note list: search, sort, tag filter pills, pagination
@@ -354,7 +358,7 @@ fabledassistant/
     │   ├── components/
     │   │   ├── LogsView.vue     # Admin log viewer: stats summary, category/search/date filters, paginated table with IP column + expandable detail rows (expands on ip_address or details)
     │   │   ├── AppHeader.vue    # Nav bar: brand, nav links (incl. admin Logs), status indicator, theme toggle, user info + logout, hamburger menu (mobile)
-    │   │   ├── ChatPanel.vue    # Slide-out chat panel (right side overlay, receives contextNoteId prop), bubble-style messages, floating dark input, note picker, context pills with promote/exclude
+    │   │   ├── ChatPanel.vue    # Slide-out chat panel (right side overlay, receives contextNoteId prop), bubble-style messages, floating dark input, note picker, context pills with promote (+) only (no exclude)
     │   │   ├── ModelSelector.vue # Model dropdown (v-model pattern): fetches installed + running models, hot/cold indicators
     │   │   ├── DashboardChatInput.vue # Inline chat bar: ModelSelector + note picker + textarea + send button; emits submit event
     │   │   ├── ToolCallCard.vue # Compact card for tool call results (created/deleted task/note, note content, notes list, search results, CalDAV events/todos, errors) + suggested tag pills with apply-on-click
@@ -432,7 +436,7 @@ fabledassistant/
 | GET | `/api/chat/conversations/:id` | Get conversation with all messages |
 | DELETE | `/api/chat/conversations/:id` | Delete conversation (cascades to messages) |
 | PATCH | `/api/chat/conversations/:id` | Update conversation title or model (body: `{title?, model?}`) |
-| POST | `/api/chat/conversations/:id/messages` | Start generation: save user message, launch background task, return 202 (body: `{content, context_note_id?, exclude_note_ids?}`) |
+| POST | `/api/chat/conversations/:id/messages` | Start generation: save user message, launch background task, return 202 (body: `{content, context_note_id?, include_note_ids?}`) |
 | GET | `/api/chat/conversations/:id/generation/stream` | SSE endpoint tailing generation buffer; supports `Last-Event-ID` reconnection; emits `context`, `chunk`, `done`, `error` events |
 | POST | `/api/chat/conversations/:id/generation/cancel` | Cancel active generation (sets cancel_event, saves partial content) |
 | POST | `/api/chat/messages/:id/save-as-note` | Save assistant message as a new note (LLM-generated title, tagged `chat`) |
@@ -565,9 +569,12 @@ When adding a new migration, follow these conventions:
 - Background generation with `GenerationBuffer` (in-memory SSE fan-out, `Last-Event-ID` reconnect, 60s cleanup)
 - Stop generation with partial content preservation
 - Note-aware context building: current note + keyword search for related notes + URL fetching
-- **Persistent context sidebar:** Right panel in ChatView accumulates auto-found notes across turns.
-  Manually attached note appears with 📌 pin, clears after send. × excludes from future auto-search.
-  Hidden on mobile (≤768px). Replaces old ephemeral context pills in the message stream.
+- **Context sidebar (Phase 21):** Right panel in ChatView shows two sections: **Suggested** (auto-found
+  by semantic/keyword search — click `+` to include) and **In Context** (explicitly included by user —
+  click `×` to remove). Removing an included note moves it back to Suggested. Manually attached note
+  appears with 📌 pin in the In Context section, clears after send. Hidden on mobile (≤768px).
+  Auto-found notes are shown as sidebar candidates only — they are NOT injected into the system prompt
+  automatically, keeping the system prompt prefix stable for Ollama KV cache reuse.
 - Note picker (paperclip) in chat input; attached note title passed to store for optimistic render.
   `context_note_title` synthesised server-side at conversation load via batch `get_notes_by_ids()`;
   message badge shows note title instead of "Note #N".
@@ -614,21 +621,22 @@ When adding a new migration, follow these conventions:
     (`search_todos` keyword-filters the todo list — companion to `list_todos`)
 - **Streaming status transparency:** The backend emits `status` SSE events at each pipeline stage
   so the user always sees what's happening instead of a blank progress dot. Stages:
-  (1) `"Generating response..."` immediately after context is built (no blocking wait for intent);
-  (2) human-readable tool label (e.g. `"Creating calendar event..."`) before each tool executes
-  (both intent-routed and native); (3) `"Composing response..."` before tool follow-up rounds.
+  (1) intent ack text streamed as a `chunk` event (tool responses — TTFT ~400ms) or
+  `"Generating response..."` status event (chat-only responses);
+  (2) human-readable tool label (e.g. `"Creating calendar event..."`) before each tool executes;
+  (3) `"Composing response..."` before tool follow-up rounds.
   Frontend: `chat.ts` stores `streamingStatus` ref, cleared on first content chunk or on done/error.
   `ChatView.vue` shows a pulsing dot + italic label above the content while status is active, then
   falls back to the blinking cursor when content streams in. `HomeView.vue` dashboard panel shows
   the status label in place of `...` before any content arrives.
-- **Intent routing (optimistic streaming):** On the first round, the backend races intent
-  classification against the start of the LLM stream using `asyncio.wait(FIRST_COMPLETED)`.
-  The LLM stream is immediately started into an `asyncio.Queue` while `classify_intent()` runs
-  concurrently. If the stream produces its first token before classification completes, the intent
-  task is cancelled and the user sees tokens immediately (zero blocking for pure chat). If
-  classification finishes first and detects a tool call, the stream is cancelled and the tool
-  executes directly — bypassing the model's native (sometimes unreliable) tool calling API. Falls
-  through to normal streaming when no tool is detected or classification fails.
+- **Intent routing (intent-first pipeline, Phase 21):** On the first round, `build_context()` and
+  `classify_intent()` run concurrently. Once intent returns (~400ms), the pipeline immediately acts:
+  if a tool is detected, the intent's one-sentence `ack` field is streamed as the first chunk
+  (becoming TTFT), the tool executes, then the main model generates a follow-up response with the
+  tool result. For chat-only responses, the model streams directly with no ack prefix.
+  No optimistic streaming queue or race — eliminates wasted GPU prefill when intent won the race.
+  `IntentResult.ack` (one-sentence acknowledgment) is embedded in the intent JSON output, so no
+  additional LLM call is needed for acknowledgment. Intent model `max_tokens` 350 (was 200).
   Dedicated intent model configurable via `OLLAMA_INTENT_MODEL` env var (default `qwen2.5:1.5b`)
   or per-user `intent_model` setting — smaller/faster model for routing. Main model default
   is `qwen3:latest` (configurable via `OLLAMA_MODEL` env var or per-user `default_model` setting
@@ -641,6 +649,8 @@ When adding a new migration, follow these conventions:
   search_events), update_note vs create_note disambiguation, reminder_minutes conversion,
   delete_note vs delete_task disambiguation, get_note for "read/show me this note", list_notes for
   "browse/list notes", tag management via update_note (tag_mode add/remove), search_todos.
+  `generate_completion` (used by intent classifier) retries on HTTP 500 (3 attempts, 3s/6s delays)
+  to handle cold model loading without failing intent classification.
 - **CalDAV calendar integration:** Per-user CalDAV settings (URL, username, password, calendar name, timezone).
   LLM tools: `create_event` (all_day, recurrence, timezone, reminder_minutes, attendees, calendar_name),
   `list_events`, `search_events`, `update_event`, `delete_event`, `list_calendars`,