couple changes

Author: AG9898

.codex/skills/ralph-loop/SKILL.md:Zone.Identifier

@@ -172,12 +172,16 @@ Primary assistant surface for conversation and note authoring.
 - The page may show a persisted conversation list so an older chat can be resumed without leaving `/chat`.
 - Starting a new chat does not require deleting older conversations.
 - The composer remains a single unified entry point rather than a primary mode switcher.
+- The initial chat mode defaults to inference-first `Auto`.
 - The UI may expose compact create/update override controls near the composer, but inference is the default routing path.
+- The chat surface may render those override controls and the model picker inside the composer chrome rather than as a separate top toolbar.
 - `/chat` is the sole note-authoring entry point for new notes; there is no dedicated new-note page.
 - If the assistant detects a strong match to an existing note, the page may surface that note inline and offer research or review actions without forcing an immediate update flow.
+- For topic-learning prompts in `Auto` mode that do not strongly match an existing saved note, the assistant may return both a conversational answer and a `create_note` proposal so the user can add the topic to notes immediately.
 - The page may render create/update proposals as editable inline draft panels and delete proposals as explicit confirmation cards.
 - The page submits full conversation state to `POST /api/assistant/respond`.
 - Provider/model options come from the server-side registry in `src/lib/server/ai/models.ts`.
+- The UI should continue to initialize provider/model state from those server-supplied defaults rather than introducing visual-only overrides.
 - Respond-time prompt grounding also includes the shared canonical note-category list and a bounded deterministic list of existing lower-case note tags so create/update drafts reuse established taxonomy when possible.
 - Assistant proposals must follow the standard note skeleton from `docs/NOTES.md`: `Overview`, `Description`, `Key Concepts`, `Connections`, and `Resources`, with only the approved optional sections allowed between `Key Concepts` and `Connections`.
 - Assistant prompts also keep `Overview` brief, treat `Description` as the main explanatory section, prefer evergreen explanation over release-churn unless `Version Notes` is warranted, and ban deprecated default headings like `Current Status`, `Notable Features`, `Quick Examples`, and `Industry Usage`.
@@ -187,7 +191,8 @@ Primary assistant surface for conversation and note authoring.
 - Assistant responses may include a structured proposal for `create_note`, `update_note`, or `delete_note`.
 - Create/update proposals render as editable draft panels in chat before save.
 - Delete proposals render as explicit confirmation UI.
-- Live web citations may be shown in chat review, but are not persisted as dedicated source metadata.
+- Live web citations may be shown in chat review as a collapsed sources disclosure, but are not persisted as dedicated source metadata.
+- The chat UI should cap visible source links to a small primary set rather than dumping the full research result list into the thread.
 - Persisted chat history stores the app-owned transcript, not provider-managed hidden conversation state.
 - Resuming a conversation reconstructs the `messages` payload from saved chat history before calling the assistant again.
 - The response payload also includes `routing`, which exposes the resolved mode, matched note, target note, and override source so the chat UI can show the selected branch without inferring it client-side.

.codex/skills/start-task/SKILL.md:Zone.Identifier

@@ -164,6 +164,7 @@ The dedicated `/notes/new` route has been removed; new-note authoring stays insi
 - resolving current note context for update flows before prompting the model
 - deriving an explicit delete target only when the user clearly asks to delete a selected or strongly matched saved note
 - deciding whether the assistant should answer conversationally only or also return a structured proposal
+- allowing eligible topic-learning prompts in inference-first chat mode to return a conversational answer plus a create-note proposal when no strong saved-note match exists
 - normalizing proposal payloads before they return to the UI
 
 Target request-time sequence:

.codex/skills/workboard-query/SKILL.md:Zone.Identifier

@@ -167,12 +167,13 @@ Rules:
 Chat is now both the conversation surface and the primary authoring surface.
 
 ```
-[ history list / drawer ........ ][Conversation column.....]
-[ recent chats / new chat ...... ][ assistant toolbar: provider | model | compact override controls ]
-[ low-noise resume affordance ...][ note picker row: visible only when Update is active    ]
-[ compact metadata ..............][ conversation area ....................................... ]
-[ ................................][ assistant proposal panel appears inline beneath message  ]
-[ ................................][ composer ................................................ ]
+[ history list / drawer ........ ][ centered conversation column ............................ ]
+[ recent chats / new chat ...... ][ minimal brand/title in empty state ...................... ]
+[ low-noise resume affordance ...][ composer chrome: inline model | skill-like mode chips .. ]
+[ compact notebook index ........][ note picker row: visible only when Update is active .... ]
+[ ................................][ conversation area without a dashboard wrapper .......... ]
+[ ................................][ assistant proposal panel appears inline beneath message ]
+[ ................................][ composer remains the primary entry point ............... ]
 ```
 
 The assistant must be able to:
@@ -182,6 +183,18 @@ The assistant must be able to:
 - request confirmation for deletion
 - reopen a saved conversation transcript
 
+Chat layout rules:
+- The initial chat state defaults to inference-first `Auto`.
+- Empty chat should use a centered, minimal entry state rather than a large explanatory card.
+- Remove non-essential top-of-page copy and status pills; the page should not open with a heavy header block.
+- The composer should sit in a centered narrow column rather than stretching like a full-width dashboard dock.
+- The conversation text bar should be materially smaller than the current wide dock treatment: reduced width, reduced height, reduced padding, and a calmer send affordance.
+- The model selector should stay visible in the composer chrome, while provider switching becomes a quieter secondary control.
+- `Auto`, `Create`, and `Update` should read like compact skill toggles attached to the composer rather than a separate toolbar, and should be visually smaller than the main prompt surface.
+- The update target picker should appear only when `Update` is active, anchored directly under the composer.
+- The main chat surface should not be wrapped in a generic dashboard-style container component; the conversation area itself stays visually open.
+- Chat should follow the modern AI-reference layout rhythm from `references/UI/chat-example.png`, translated through Techy's tokens and type system rather than copied literally.
+
 ### Assistant Proposal Panel
 
 For `create_note` and `update_note` proposals, render an editable draft panel inline in chat.
@@ -255,6 +268,9 @@ The top nav has been replaced with a collapsible left rail (`<nav class="rail">`
 - In Update mode, a full-width note-picker `<select>` appears below the toolbar for choosing the note to review
 - Send is disabled in Update mode until a note is selected
 - Add provider/model selection controls
+- Composer chrome is constrained to a centered narrow width on desktop rather than spanning the full conversation column
+- Model remains the primary visible selector; provider stays visible but smaller and quieter
+- The composer helper copy is minimized; in `Auto` mode the composer does not need an always-visible explanatory sentence
 - Render assistant citations and proposal panels inline
 - Use Melt for selectors, confirmation affordances, and disclosure-style interaction where it improves accessibility
 - Selecting a saved conversation restores its transcript; resuming it should not depend on provider-side hidden memory

docs/API.md

@@ -136,11 +136,19 @@ Rules:
 - keep citations assistive and low-noise
 - inferred create/update behavior must not break normal conversation
 - persisted chat history should feel like a quiet notebook index, not a busy support inbox
+- default the chat entry state to inference-first `Auto`
+- reduce page-header copy to the minimum needed; avoid stacked title, eyebrow, lede, and status-pill clutter
+- keep the empty state centered and composer-led rather than explaining the product through cards
+- keep the model control inline with the composer chrome and demote provider switching to a quieter secondary control
+- style `Auto`, `Create`, and `Update` as compact skill-like toggles attached to the composer
+- avoid wrapping the main chat surface in a large generic card or dashboard shell
+- let the transcript feel flatter and more editorial than the current rounded-component treatment
 
 Recommended desktop arrangement:
 - a restrained history rail or drawer for recent chats
 - centered conversation column
-- assistant controls near the composer
+- minimal brand/title above the composer in the empty state
+- assistant controls integrated into the composer chrome
 - inline editable proposal panel for create/update
 - optional secondary context region later, but not required for the first pass
 
@@ -195,19 +203,26 @@ Rules:
 - keep the composer persistent and obvious
 - keep the chat surface unified and inference-first
 - expose create/update as compact override controls rather than the primary mental model
-- show provider/model controls without turning the toolbar into a cockpit; keep the controls compact and adjacent to the composer
+- keep the composer visually narrower and calmer than the full route width; it should feel like a centered prompt instrument, not a dashboard dock
+- show provider/model controls without turning the composer into a cockpit; keep the controls compact, smaller than the prompt field, and adjacent to it
 - let users resume prior conversations without making chat history the dominant visual element
 - render note proposals inline as editable review surfaces
 - when a strong existing-note match is found, surface it inline and offer further research or review without abruptly switching the user into an edit flow
-- show citations during review without overwhelming the draft itself
+- show citations during review without overwhelming the draft itself; prefer a collapsed low-noise sources disclosure over inline source dumps
 - keep delete confirmation compact and deliberate
 - store and present the app-owned transcript as the canonical history view; do not expose provider-specific hidden memory concepts in the UI
+- favor a modern AI-chat composition with a centered entry state and composer-led control surface, translated into Techy's typography, tokens, and spacing system
+- keep the initial selected mode on `Auto`; explicit create/update remain available as hard overrides, not as the default mental model
+- keep the model default sourced from the provider registry; do not introduce hardcoded visual-only model defaults in the chat page
+- keep the chat shell in the route surface for this pass; do not introduce a wrapper component that re-boxes the full chat experience
+- avoid persistent helper copy in `Auto` mode when the compact composer layout already communicates the interaction model
 
 Avoid:
 - making the chat UI look like a separate product
 - forcing ambiguous prompts into create/update behavior without a clear signal or fallback
 - turning assistant proposals into a wall of utility panels
 - turning saved conversation history into a dense email-style list that competes with the current conversation
+- recreating the current heavy top-of-page chat header or the `No conversation yet` explainer card
 
 ---
 

docs/ARCHITECTURE.md

@@ -26,6 +26,9 @@ The current unit tests focus on fast, deterministic modules that do not require
   - checks required section order, approved optional-section placement, deprecated-heading rejection, and heading normalization
 - `src/lib/server/ai/prompts.test.ts`
   - verifies create/update assistant prompts both inject the same shared note-structure contract
+- `src/lib/server/assistant/routing.test.ts`
+  - verifies inference-first routing keeps plain learning prompts conversational
+  - verifies follow-up asks like `add it to notes` or `save this to my notes` route into note creation
 
 ## Shared Contract