Wes/core architecture client rework

.agents/skills/writing-reference-docs/SKILL.md

@@ -0,0 +1,110 @@
+---
+name: writing-reference-docs
+description: >-
+  How to write a function/hook/action reference section: plain-language
+  signature, real arguments, escalating examples grounded in a real app,
+  trimmed prose with no em dashes or semicolons. Use when writing or editing a
+  packages/core/docs/content reference page.
+scope: dev
+metadata:
+  internal: true
+---
+
+# Writing Reference Docs
+
+This came out of rewriting `client-data.mdx`'s hook sections
+(`useActionQuery`, `useActionMutation`, `callAction`, `useDbSync`) section by
+section with the user. It captures the shape that emerged so the next
+reference section starts from it instead of reinventing it.
+
+## Rule
+
+Document each function, hook, or action reference as: a plain-language
+purpose statement, a bulleted argument list matching the real signature, two
+or more escalating runnable examples grounded in a real app, and a one or two
+sentence closing behavior note. Never a bare one-line description with a
+single toy snippet.
+
+## Why
+
+The original `client-data.mdx` gave each hook one sentence and one minimal
+snippet with no options argument shown at all. That hid real, common needs
+(conditional fetching via `enabled`, a post-mutation side effect via
+`onSuccess`) that readers would only discover by reading the source. It also
+leaned on a hypothetical `leads` domain (`get-lead`, `create-lead`,
+`archive-lead`) that reads as unconvincing next to an example grounded in a
+real app with a real schema and real access rules.
+
+## How
+
+1. **Match heading level to role.** Hooks that solve the same kind of problem
+   get grouped under one parent heading (e.g. `### Action hooks`) with each
+   hook as a child heading below it. A utility that is not a hook (like
+   `callAction`) is a sibling heading at the same level as the group, not a
+   child of it, even if it lives right next to the group.
+2. **Open with plain-language purpose, then link out.** State what the
+   function is for in one or two sentences ("This hook is intended to be used
+   for..."). If it wraps another library's hook (React Query's `useQuery`,
+   `useMutation`), link its reference page instead of re-documenting fields
+   you don't own.
+3. **List real arguments as a short bullet list**, matching the actual
+   exported signature in `packages/core/src/client/`, not a paraphrase. Read
+   the source before writing the list. Skip the bullet-list format entirely
+   for a function that takes only one argument. Fold that into a sentence
+   instead:
+
+   ```md
+   `useActionQuery()` accepts three arguments:
+
+   - **actionName**: the action's registered name.
+   - **params**: the action's input, typed from its `defineAction()` schema.
+   - **options**: everything from the [useQuery options](...) except
+     `queryKey` and `queryFn`, which the hook sets itself.
+   ```
+
+4. **Give two or more escalating examples.** The first is the simplest
+   possible call. The second demonstrates one real, common option (a
+   conditional fetch, a success callback, a longer timeout), with one
+   sentence before it explaining what's different and why it matters.
+5. **Ground every example in a real reference app's real action**, not an
+   invented one, e.g. `get-ticket`, `send-ticket-reply`, `update-ticket` from
+   a real ticket-support example app, rather than `get-lead`/`create-lead`.
+   Exception: if the page already threads a hypothetical domain across
+   multiple sibling pages (a running example), keep using that domain in this
+   page too. Enrich its snippets with the missing option; don't swap the
+   domain out from under the other pages.
+6. **Close with the one behavior fact a reader needs**, in one or two
+   sentences: cache key, invalidation trigger, timeout.
+7. **Prose rules, every sentence:** no em dashes, no semicolons (this
+   includes table cells), short sentences. Split into two sentences instead
+   of joining with either. Reference every function by name with parens,
+   `useActionQuery()` not `useActionQuery`.
+8. **Example values must mean something to a reader with zero app context.**
+   Don't reuse an in-app sentinel value (like a `"me"` string a real action
+   resolves specially) without explaining it. Use a literal, self-explanatory
+   value instead, like a real-looking email address.
+
+## Don't
+
+- Don't introduce a new framework concept into the primary example just
+  because it's technically correct for that hook. If most readers of this
+  page will never need to write it themselves (e.g. wiring `ignoreSource` with
+  a per-tab id), it belongs in the options table, not the main example, or on
+  a deeper page like Advanced.
+- Don't swap a page's running example domain without checking whether sibling
+  pages share it. Grep the other draft pages for the same domain terms first.
+- Don't leave a semicolon or em dash anywhere in the page, including table
+  cells. Split into two sentences.
+- Don't add a one-item options bullet list for a function that only takes a
+  single argument. State it in a sentence instead.
+- Don't paraphrase a signature from memory. Grep the actual export in
+  `packages/core/src/client/` and read its real parameter and option types.
+
+## Related Skills
+
+- **writing-agent-instructions** — the sibling guide for AGENTS.md/SKILL.md
+  prose, which this borrows its "say it once, plainly" spirit from.
+- **agent-native-docs** — how to look up the version-matched docs this skill
+  helps you write.
+- **internationalization** — localized copies under `content/locales/*` need
+  the same edit when a source doc's meaning changes (see CLAUDE.md).

packages/core/docs/content/client-advanced.mdx

@@ -0,0 +1,192 @@
+---
+title: "Advanced"
+description: "Lower-level app/ primitives and escape hatches: direct application-state access, background chat sends, custom suggestion chips, and the MCP App host bridge."
+---
+
+# Advanced
+
+Overview, Data & Sync, Agent Chat, and Routing cover the typical case: the hooks and helpers most pages need. This page covers the lower-level primitives and narrower escape hatches underneath them. Most apps never touch these directly. Reach for one when you're building something more custom than the default hooks support, such as a custom selection sync or an embedded MCP App. For syncing a raw `useQuery` with a hand-rolled key, see [Sync Internals](/docs/client-sync-internals) instead.
+
+## Application State {#application-state}
+
+`application_state` is the SQL-backed key/value store the agent reads to know what the UI is currently looking at, including navigation, selection, and other focused-object context (see [Context Awareness](/docs/context-awareness)). `useAgentRouteState` and `useSemanticNavigationState` (also covered in Context Awareness) already wrap these for the common navigation case. Reach for the primitives below directly when you're syncing something they don't cover, such as a custom selection, or writing from outside a component's render.
+
+### readClientAppState {#read-client-app-state}
+
+Read the current value for a key. Resolves to whatever was last written for that key, or `null` if the key has never been written.
+
+```ts
+import { readClientAppState } from "@agent-native/core/client/hooks";
+
+const selection = await readClientAppState("selection");
+// selection is exactly what was last written to that key, e.g.:
+// { kind: "tickets.rows", ids: ["tik_182", "tik_209"] }
+//
+// or, if nothing has written "selection" yet:
+// null
+```
+
+### writeClientAppState {#write-client-app-state}
+
+Write a value for a key. Unlike `setClientAppState` below, it always writes, even if you pass `null`. Passing `undefined` throws instead, since `undefined` isn't valid JSON. Use `setClientAppState(key, undefined)` or `deleteClientAppState()` to clear a key.
+
+```ts
+import { writeClientAppState } from "@agent-native/core/client/hooks";
+
+await writeClientAppState(
+  "selection",
+  { id: recordId },
+  { requestSource: TAB_ID },
+);
+```
+
+### setClientAppState {#set-client-app-state}
+
+The convenience wrapper most UI code should reach for: writes the value, or deletes the key when you pass `null`/`undefined`. Use this instead of `writeClientAppState` unless a call site genuinely needs "always write, never delete."
+
+```ts
+import { setClientAppState } from "@agent-native/core/client/hooks";
+
+await setClientAppState(
+  "selection",
+  { id: recordId },
+  { requestSource: TAB_ID },
+);
+await setClientAppState("selection", null); // clears the key, same as deleteClientAppState
+```
+
+### deleteClientAppState {#delete-client-app-state}
+
+Remove a key entirely.
+
+```ts
+import { deleteClientAppState } from "@agent-native/core/client/hooks";
+
+await deleteClientAppState("pending-selection-context");
+```
+
+### readClientAppStateMany {#read-client-app-state-many}
+
+Read several keys in one request. Returns `{ values, missing }`. `values` holds an entry for every key the server has a row for, including a key that was explicitly written as `null`. `missing` lists the keys that have never been written at all. This is how "never written" stays distinguishable from "written as null."
+
+```ts
+import { readClientAppStateMany } from "@agent-native/core/client/hooks";
+
+const { values, missing } = await readClientAppStateMany([
+  "selection",
+  "pending-selection-context",
+]);
+```
+
+Requests over 100 keys are chunked automatically. `readClientAppState()` itself is built on this: calls issued in the same tick are coalesced into one batched request behind the scenes, so calling it many times in a single render pass costs one round trip, not one per call.
+
+### Options {#application-state-options}
+
+`signal` is accepted by all five functions above, to cancel an in-flight call. `requestSource` and `keepalive` only apply to the three write functions, since they don't mean anything for a read.
+
+| Option          | Type           | Applies to                                                         | Description                                                                                  |
+| --------------- | -------------- | ------------------------------------------------------------------ | -------------------------------------------------------------------------------------------- |
+| `requestSource` | `string?`      | `writeClientAppState`, `setClientAppState`, `deleteClientAppState` | Tags the write with an id so a matching `useDbSync({ ignoreSource })` can skip refetching it |
+| `keepalive`     | `boolean?`     | `writeClientAppState`, `setClientAppState`, `deleteClientAppState` | Lets the request outlive page unload. Use for cleanup writes like clearing selection         |
+| `signal`        | `AbortSignal?` | all five functions                                                 | Cancels the request                                                                          |
+
+### Failure behavior {#failure-behavior}
+
+All five functions throw on a failed request, a non-2xx response, or a body that can't be parsed. The thrown `Error` carries a `.status` property when the server responded, so a caller can branch on the HTTP status if needed. Wrap a call in `try`/`catch`, or let it propagate to an error boundary, the same as any other network call.
+
+### What TAB_ID is {#tab-id}
+
+`requestSource` is commonly set to `TAB_ID`, a plain constant every template exports from `app/lib/tab-id.ts`: a short random string that stays the same for the life of a tab and differs between tabs. Pairing `requestSource: TAB_ID` on a write with `useDbSync({ ignoreSource: TAB_ID })` on the read side is how a tab avoids refetching its own write while still reacting to everyone else. See [Context Awareness](/docs/context-awareness#jitter-prevention) for the full walkthrough.
+
+## sendToAgentChat: Background Sends {#background-send}
+
+Use `background: true` when a UI action should kick off real agent work without
+opening or focusing the sidebar. This still creates a normal chat thread/run,
+uses the agent's tools/actions/context, and keeps the work observable through
+the runs tray. It is not a raw one-shot model call.
+
+```ts
+import { sendToAgentChat } from "@agent-native/core/client/agent-chat";
+
+const tabId = sendToAgentChat({
+  message: "Analyze this import and create any missing records",
+  context: `Import batch id: ${batchId}`,
+  submit: true,
+  newTab: true,
+  background: true,
+  openSidebar: false,
+});
+```
+
+`background` is intended to be paired with `newTab` so the hidden work does not
+overwrite the user's active conversation. Use the returned `tabId` if the UI
+needs to correlate follow-up status or deep-link into the run later.
+
+A few more `AgentChatMessage` options exist beyond the ones covered in [Agent Chat](/docs/client-agent-chat#agentchatmessage), mostly for embedded or downstream-integration cases:
+
+| Option                | Type        | Description                                                                |
+| --------------------- | ----------- | -------------------------------------------------------------------------- |
+| `background`          | `boolean?`  | With `newTab`, run without focusing the tab and show the run in `RunsTray` |
+| `projectSlug`         | `string?`   | Optional project slug for structured context                               |
+| `preset`              | `string?`   | Optional preset name for downstream consumers                              |
+| `referenceImagePaths` | `string[]?` | Optional reference image paths                                             |
+
+## MCP App Host Bridge {#mcp-app-host-bridge}
+
+Routes embedded as MCP Apps should be URL-first: load the current artifact from
+path/query params, render the real React route or a focused shared component,
+and use the host bridge only for host-owned behavior. `@agent-native/core/client`
+exports the helpers embedded routes call:
+
+```ts
+import {
+  getMcpAppHostContext,
+  openMcpAppHostLink,
+  requestMcpAppDisplayMode,
+  updateMcpAppModelContext,
+  useMcpAppHostContext,
+} from "@agent-native/core/client/agent-chat";
+```
+
+`getMcpAppHostContext()` reads the latest pushed host context snapshot.
+`useMcpAppHostContext()` subscribes React components to changes. The request
+helpers (`openMcpAppHostLink`, `requestMcpAppDisplayMode`,
+`updateMcpAppModelContext`) return `false` outside an embedded MCP App frame, or
+`Promise<boolean>` inside a frame.
+
+`sendToAgentChat()` (see [Agent Chat](/docs/client-agent-chat#sendtoagentchat)) uses the same bridge for auto-submitted prompts from embedded routes. Inside an MCP App embed created with `embedApp()`, auto-submitted messages (`submit` omitted or `true`) are forwarded to the bridge, which asks the containing host to add hidden context and send the visible user turn. `context` stays model-visible without being posted as user-facing chat. `submit: false` keeps the local prefill/review behavior because MCP Apps do not define a standard draft-prefill API. Internally this is the submitted-chat path sometimes surfaced as `agentNative.submitChat`. App code should still call `sendToAgentChat()` rather than posting that event directly.
+
+The bridge itself, the `ui/*` JSON-RPC messages, the `agentNative.mcpHost.*`
+wrapper relay, transplant vs. controlled-frame rendering, host context, and
+display-mode requests, is owned by
+[External Agents](/docs/external-agents#mcp-app-bridge). See [MCP Apps](/docs/mcp-apps) for the full protocol.
+
+## Custom Suggestion Chips {#custom-suggestions}
+
+`<AgentSidebar>`'s `dynamicSuggestions` (see [Agent Chat](/docs/client-agent-chat#dynamic-suggestions)) already merges static `suggestions` with the framework's own context-aware heuristic. Pass `getSuggestions` instead when an app wants deterministic, domain-specific chips computed from the same `application_state` context rather than the framework's default heuristic.
+
+```tsx
+<AgentSidebar
+  suggestions={["Summarize my inbox"]}
+  getSuggestions={(state) =>
+    state.navigation?.view === "inbox"
+      ? [`Summarize ${state.navigation.label ?? "this"} label`]
+      : []
+  }
+>
+  <App />
+</AgentSidebar>
+```
+
+`getSuggestions` receives the same `navigation`/`selection`/`pending-selection-context` snapshot the built-in heuristic reads, so a custom implementation can react to exactly the state the default one does, just with app-specific rules instead of generic ones.
+
+## What's next
+
+- [**Overview**](/docs/client-overview) — the `app/` and `public/` directory layout these primitives live in
+- [**Data & Sync**](/docs/client-data) — the default action hooks and `useDbSync` most pages should use instead
+- [**Sync Internals**](/docs/client-sync-internals) — `useChangeVersion`/`useChangeVersions` and the two latency paths a refetch can take
+- [**Agent Chat**](/docs/client-agent-chat) — the default `sendToAgentChat`/`askUserQuestion` helpers most UI should use instead, and the staged agent chat context-state APIs
+- [**Entry Points**](/docs/client-entry-points) — `app/entry.client.tsx`/`app/entry.server.tsx` and SSR streaming customization
+- [**Context Awareness**](/docs/context-awareness) — the `navigation`/`selection` semantics built on `application_state`, and the `TAB_ID`/jitter-prevention walkthrough
+- [**Generative UI**](/docs/generative-ui) — richer inline chat controls beyond `askUserQuestion`
+- [**External Agents**](/docs/external-agents#mcp-app-bridge) — the full MCP App host bridge internals