@adia-ai/web-modules 0.8.28 → 0.8.30

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog — @adia-ai/web-modules
 
+## [0.8.30] — 2026-08-07
+
+### Added
+- **`chatShell.setStreamedText(text)`** — replaces the streaming assistant message's text outright, the counterpart to `appendChunk` for a producer that revises what it already streamed.
+
+### Fixed
+- **A rewritten or withheld turn now leaves the screen agreeing with the transcript (gh#665).** `wireAgentEvents` treated every `text` event as an append, so an `@adia-ai/agent` output guardrail's re-snapshot (`{ text: '', snapshot: <rewritten> }`) appended nothing and the ORIGINAL text stayed on screen while the session recorded the rewrite — a redaction that redacted nothing visible. An empty delta is now a replace from `snapshot` (an empty snapshot clears the bubble), which is safe for any text event since the snapshot is always the authoritative text.
+
+### Maintenance
+- **`billing/` touched in this release window** (1 file(s), e.g. `invoice-history/invoice-history.class.js`) — carried by the entries above.
+- **`chat/` touched in this release window** (5 file(s), e.g. `chat/README.md`) — carried by the entries above.
+- **`dist/` bundles rebuilt** in this cut's window (2 file(s)) — regenerated from the source changes described above, not independent edits.
+
+## [0.8.29] — 2026-08-06
+
+### Added
+- **chat-surfaces renders plan-envelope chrome** — `wireAgentEvents` gains a `step` case rendering the turn's surface manifest and "what changed" synthesis as feed chrome (workflow:'plan'), riding the existing AgentEvent `step` kind — no event-vocabulary change (gh#648, PR #654).
+
+### Maintenance
+- **`dist/` bundles rebuilt** in this cut's window (1 file(s)) — regenerated from the source changes described above, not independent edits.
+
 ## [0.8.28] — 2026-08-05
 
 ### Fixed

package/chat/README.md

@@ -0,0 +1,54 @@
+# `web-modules/chat`
+
+The chat cluster: `<chat-shell>` (behavior-only orchestrator) composing
+`<chat-header>` / `<chat-thread>` / `<chat-composer>` / `<chat-status>` /
+`<chat-empty>`, plus `chat-surfaces.js` — the L3 seam that renders an
+`@adia-ai/agent` `AgentEvent` stream (or a raw `@adia-ai/llm` chunk stream)
+into the thread. See `chat-shell/chat-shell.js`'s header comment for the
+bespoke structure vocabulary and the two operating modes (built-in LLM
+streaming vs. external-drive).
+
+## `wireAgentEvents` — AgentEvent → rendering (gh#667)
+
+`wireAgentEvents(chatShell, registry, { events, statusEl? })` folds ONE
+`AsyncIterable<AgentEvent>` (`@adia-ai/agent`'s stream contract — see
+[`packages/agent/README.md`](../../agent/README.md#events)) into the thread.
+It never touches session state — `reduce(session, event)` is the caller's
+job, the same event object just also flows through this function to paint.
+
+| Event | Rendering |
+| --- | --- |
+| `text` | Appended to the streaming bubble (`chatShell.appendChunk`); an EMPTY delta REPLACES from `snapshot` (`chatShell.setStreamedText`) — the guardrail-rewrite case (see `@adia-ai/agent`'s README, "Output rewrite"). |
+| `thinking` | `<agent-reasoning-ui>.addThought(text)` — folded into the turn's collapsed-by-default reasoning viewer. |
+| `tool_use` | `<agent-reasoning-ui>.addStep({id, label})` (or `updateStep` if `tool_input_delta` already opened the step) — a running step in the reasoning viewer's timeline. |
+| `tool_input_delta` | Cheap upsert on the running step's `label` (`updateStep`, or `addStep` if this is the first fragment) — progress only, nothing executes off a partial. |
+| `tool_result` | `updateStep(id, {outcomes:[output]})` + `completeStep(id)`, or `failStep(id, output)` when `isError` — the step's status icon flips to done/error and the (truncated) output becomes the step's collapsible outcome. |
+| `step` (`workflow: 'plan'`) | Feed-chrome prose via `describePlanManifest` / `describeTurnSynthesis` (gh#648) — unchanged, NOT routed through the reasoning viewer (a plan manifest/synthesis line is thread narration, not a pipeline step). |
+| `step` (any other workflow) | One-shot attempt marker — `@adia-ai/agent`'s workflow loop emits this AFTER the step already ran, so it lands as an immediately-settled `addStep` + `completeStep`/`failStep` in the reasoning viewer. |
+| `guardrail` | `addThought('Guardrail "<name>" <blocked\|rewrote> <target> — <reason>')` — a subtle note inside the (collapsed) reasoning viewer, never a thread interruption. |
+| `progress` | `statusEl.textContent` from the code-owned `PROGRESS_LABELS` table (CHAT-HARNESS law 3 — never the raw `stage` string, never model text). |
+| `surface` | Routed to the surface-host `registry` (WCH-4 / gh#599) — a generative-UI envelope line, mounted/updated/closed via `createSurfaceRegistry`. |
+| `memory` | Render-silent — a session-state write, not a thread event. |
+| `trace` | Render-silent by default — tracing is devtools' surface (a future inspector panel reads the `trace` event / `@adia-ai/agent`'s `trace` sink), never the thread. |
+| `error` | `reasoning.fail(message)` (if a reasoning viewer is open) + an error bubble + `stopStreaming()`. |
+| `done` | `reasoning.finish()` (or `.fail('Blocked by guardrail')` when `stopReason === 'guardrail_denied'`) + `stopStreaming()`. |
+
+**Why `<agent-reasoning-ui>`, not a new primitive:** the primitive audit
+(adia-author §0) found `agent-reasoning-ui` already built for exactly this
+shape — "Agent inner monologue + pipeline viewer with steps, thoughts,
+plans, and iterations," composing `<timeline-ui>` for step rows with a
+collapsible per-step `outcomes` list (the "collapsible tool-result body"
+requirement) and a `collapsed`/status-icon summary row (the "thinking
+affordance" requirement). One viewer, mounted lazily above the streamed
+answer on the first reasoning-worthy event per turn, covers tool activity,
+workflow-step markers, thinking, and guardrail notes — no chip primitive,
+no dedicated step-marker element, and no second collapsible region were
+minted. `<chat-status>` (candidate considered) stays scoped to the header's
+connection/streaming indicator — it is a single, module-scoped element, not
+shaped for a per-event marker stream.
+
+**Demo:** `playgrounds/agent-chat/app/agent-chat.contents.js` — the *"Ask
+about a component"* prompt path drives a real `createAgent()` loop with a
+keyless deterministic client (`simulatedClient()`, used whenever
+`npm run proxy` isn't up) through `tool_use` → `tool_result` → final text,
+rendered live via this module's `wireAgentEvents`.

package/chat/chat-shell/chat-shell.js

@@ -184,6 +184,25 @@ class ChatShell extends UIElement {
     this.#scrollToBottom();
   }
 
+  /**
+   * Replace the streaming assistant message's text outright, instead of
+   * appending to it. The counterpart to `appendChunk` for the case where a
+   * producer revises what it already streamed — an `@adia-ai/agent` output
+   * guardrail rewriting or withholding the turn's text is the real one.
+   * Appending cannot express that: what is on screen has to come off.
+   *
+   * @param {string} text — the full replacement text ('' clears the bubble).
+   */
+  setStreamedText(text) {
+    const last = this.#messages[this.#messages.length - 1];
+    if (!last || last.role !== 'assistant') return;
+    last.content = text;
+
+    const contentEl = this.#messagesEl?.querySelector('[data-role]:last-child [data-content]');
+    if (contentEl) contentEl.textContent = text;
+    this.#scrollToBottom();
+  }
+
   deleteMessage(id) {
     const idx = this.#messages.findIndex(m => m.id === id);
     if (idx === -1) return;

package/chat/chat-surfaces/chat-surfaces.d.ts

@@ -39,9 +39,28 @@ export function createSurfaceRegistry(opts: {
 /** CHAT-HARNESS law 3 — code-owned progress labels, never model text. */
 export const PROGRESS_LABELS: Record<string, string>;
 
+/** Plan-envelope feed chrome (gh#648) — code-owned sentence around a
+ *  `PlanEnvelope`'s per-surface rationale. Ruling 3: feed chrome, never a canvas. */
+export function describePlanManifest(plan: {
+  turnId: string;
+  surfaces: { surfaceId: string; rationale: string }[];
+}): string;
+
+/** Plan-envelope feed chrome (gh#648) — code-owned sentence around a
+ *  `TurnSynthesis`'s new/updated/closed surface lists. */
+export function describeTurnSynthesis(synthesis: {
+  turnId: string;
+  new: string[];
+  updated: string[];
+  closed: string[];
+}): string;
+
 export function wireAgentEvents(
   chatShell: Element & {
     appendChunk?(text: string): void;
+    /** Replace the streaming text outright — used when a text event carries
+     *  an empty delta, which is a producer revising what it streamed. */
+    setStreamedText?(text: string): void;
     appendMessage?(msg: { role: string; content?: string; render?: boolean }): Element;
     startStreaming?(): void;
     stopStreaming?(): void;

package/chat/chat-surfaces/chat-surfaces.js

@@ -24,7 +24,22 @@
  * specifier. The HOST (a demo page, a consumer app) builds the renderer
  * (`createRenderer` + its own `WidgetAdapter`) and hands it in; this module
  * only ever calls the five methods `Renderer` exposes.
+ *
+ * `@adia-ai/web-components` primitives ARE a declared dependency edge
+ * (unlike `@genui/*`), so `wireAgentEvents` imports `<agent-reasoning-ui>`
+ * directly (gh#667) — the "agent inner monologue + pipeline viewer with
+ * steps, thoughts, plans, and iterations" primitive already built for
+ * exactly this shape (primitive audit, adia-author §0 — no new primitive
+ * was minted). One reasoning viewer is lazily mounted per assistant turn
+ * and renders `thinking` (→ `addThought`), `tool_use`/`tool_input_delta`/
+ * `tool_result` (→ `addStep`/`updateStep`/`completeStep`/`failStep`, with
+ * the tool output riding the step's collapsible `outcomes`), non-'plan'
+ * `step` events (→ the same step API — one-shot attempt markers), and
+ * `guardrail` (→ `addThought`, a subtle note). `memory` and `trace` stay
+ * render-silent (trace is devtools' surface, not the thread's — see the
+ * event → rendering table in `packages/web-modules/chat/README.md`).
  */
+import '../../../web-components/components/agent-reasoning/agent-reasoning.js';
 
 const CLOSED_ANNOTATION_TEXT = 'Closed.';
 const PROTOCOL_VERSION = 'v1.0';
@@ -163,6 +178,38 @@ export const PROGRESS_LABELS = {
   done: 'Done',
 };
 
+/**
+ * Plan-envelope feed chrome (gh#648, docs/ORCHESTRATION.md §Plan envelope
+ * ruling 3 — "what changed" is feed chrome, never a canvas). These render
+ * from the `{ type: 'step', workflow: 'plan', step: 'manifest'|'synthesis' }`
+ * AgentEvent — the EXISTING `step` kind (`@adia-ai/agent`'s `AgentEvent`
+ * union already carries `workflow`/`step`/`data`), so no closed-vocabulary
+ * `type` amendment is needed to carry this: `workflow: 'plan'` is just a
+ * new value inside an already-open string field, the same way
+ * `packages/agent/src/workflow.ts`'s own step events use `workflow` for a
+ * different named workflow.
+ *
+ * Code-owned text (CHAT-HARNESS law 3's spirit, applied to `step` the way
+ * PROGRESS_LABELS applies it to `progress`): the rationale strings inside
+ * a PlanEnvelope ARE planner-authored copy (not raw model text — R-O10's
+ * `rationale` is a structured field, one line per surface), so they render
+ * verbatim; the surrounding sentence is code-owned.
+ */
+export function describePlanManifest(plan) {
+  const n = plan.surfaces.length;
+  const list = plan.surfaces.map((s) => `${s.surfaceId} — ${s.rationale}`).join('; ');
+  return `Planning ${n} surface${n === 1 ? '' : 's'}: ${list}`;
+}
+
+const SYNTHESIS_LABELS = { new: 'New', updated: 'Updated', closed: 'Closed' };
+
+export function describeTurnSynthesis(synthesis) {
+  const parts = ['new', 'updated', 'closed']
+    .filter((key) => synthesis[key].length > 0)
+    .map((key) => `${SYNTHESIS_LABELS[key]}: ${synthesis[key].join(', ')}`);
+  return parts.length ? `What changed — ${parts.join(' · ')}` : 'What changed — nothing this turn';
+}
+
 /**
  * Consumes ONE `AsyncIterable<AgentEvent>` for the turn already opened on
  * `chatShell` (the caller has already appended the user message and an
@@ -188,10 +235,40 @@ export async function wireAgentEvents(chatShell, registry, opts) {
     return thread?.querySelector('[data-role="assistant"]:last-child [data-bubble]') ?? null;
   };
 
+  // One <agent-reasoning-ui> per turn (gh#667) — lazily mounted above the
+  // streamed answer on the first reasoning-worthy event (thinking, tool
+  // activity, a non-plan step, or a guardrail note), and re-used for the
+  // rest of the turn so tool steps/thoughts accumulate in ONE collapsible
+  // region instead of one element per event. `null` until then; a fresh
+  // `wireAgentEvents` call (one per turn, per this function's own contract
+  // above) starts a fresh viewer.
+  let reasoning = null;
+  const ensureReasoning = () => {
+    if (reasoning) return reasoning;
+    const bubble = currentBubble();
+    if (!bubble) return null;
+    reasoning = document.createElement('agent-reasoning-ui');
+    // Collapsed BY DEFAULT (gh#667 scope — thinking is a collapsed-by-default
+    // affordance, not a start-expanded one; the primitive's own default is
+    // expanded, so this is set explicitly).
+    reasoning.collapsed = true;
+    // Above the streamed text — the "showed its work, then answered" order.
+    bubble.insertBefore(reasoning, bubble.firstChild);
+    return reasoning;
+  };
+  const findStepEntry = (id) => reasoning?.entries?.find((e) => e.kind === 'step' && e.id === id);
+
   for await (const event of events) {
     switch (event.type) {
       case 'text':
-        chatShell.appendChunk?.(event.text);
+        // An EMPTY delta is a replace, not a no-op: the snapshot is the
+        // authoritative text, and a producer sends this shape when it
+        // revises what it already streamed (an @adia-ai/agent output
+        // guardrail rewriting or withholding the turn). Resetting from the
+        // snapshot is safe for any text event — appending is only the
+        // cheaper path for the ordinary incremental case.
+        if (event.text === '') chatShell.setStreamedText?.(event.snapshot ?? '');
+        else chatShell.appendChunk?.(event.text);
         break;
 
       case 'progress': {
@@ -206,30 +283,98 @@ export async function wireAgentEvents(chatShell, registry, opts) {
         break;
       }
 
-      case 'tool_use':
-        chatShell.appendMessage?.({
-          role: 'assistant',
-          content: `🔧 \`${event.name}(${JSON.stringify(event.input)})\``,
-          render: true,
-        });
-        chatShell.appendMessage?.({ role: 'assistant', content: '' });
+      case 'step': {
+        // Plan-envelope chrome (gh#648, ruling 1 — harness-level progress,
+        // never wire; ruling 3 — feed chrome, never a canvas) keeps its own
+        // rendering — a manifest/synthesis summary is feed prose, not a
+        // pipeline step. Every OTHER workflow's step is a one-shot attempt
+        // marker (`@adia-ai/agent`'s workflow.ts emits it once, after the
+        // step already ran) — rendered as an immediately-settled step in
+        // the turn's reasoning viewer (gh#667).
+        if (event.workflow === 'plan' && event.step === 'manifest') {
+          chatShell.appendMessage?.({ role: 'assistant', content: describePlanManifest(event.data), render: true });
+          chatShell.appendMessage?.({ role: 'assistant', content: '' });
+        } else if (event.workflow === 'plan' && event.step === 'synthesis') {
+          chatShell.appendMessage?.({ role: 'assistant', content: describeTurnSynthesis(event.data), render: true });
+          chatShell.appendMessage?.({ role: 'assistant', content: '' });
+        } else {
+          const r = ensureReasoning();
+          if (r) {
+            const id = `${event.workflow}:${event.step}:${event.data?.round ?? 1}`;
+            const label = `${event.workflow}: ${event.step}`;
+            r.addStep({ id, label });
+            if (event.data?.accepted) r.completeStep(id);
+            else r.failStep(id, event.data?.reason ?? 'fell through');
+          }
+        }
+        break;
+      }
+
+      case 'tool_use': {
+        const r = ensureReasoning();
+        if (r) {
+          const label = `${event.name}(${JSON.stringify(event.input)})`;
+          if (findStepEntry(event.id)) r.updateStep(event.id, { label });
+          else r.addStep({ id: event.id, label });
+        }
         break;
+      }
 
-      case 'tool_result':
-        chatShell.appendMessage?.({
-          role: event.isError ? 'error' : 'assistant',
-          content: `↩ \`${event.name}\` → ${String(event.output).slice(0, 300)}`,
-          render: true,
-        });
-        chatShell.appendMessage?.({ role: 'assistant', content: '' });
+      case 'tool_input_delta': {
+        // Partial args — render progress on the running chip only; never
+        // executes off it (README's own rule — the matching tool_use is
+        // the call). Cheap upsert since addStep/updateStep already exist.
+        const r = ensureReasoning();
+        if (r) {
+          const label = `${event.name}(${event.partial})`;
+          if (findStepEntry(event.id)) r.updateStep(event.id, { label });
+          else r.addStep({ id: event.id, label });
+        }
         break;
+      }
+
+      case 'tool_result': {
+        const r = ensureReasoning();
+        if (r) {
+          const outcome = String(event.output).slice(0, 300);
+          if (event.isError) {
+            r.failStep(event.id, outcome);
+          } else {
+            r.updateStep(event.id, { outcomes: [outcome] });
+            r.completeStep(event.id);
+          }
+        }
+        break;
+      }
+
+      case 'guardrail': {
+        const r = ensureReasoning();
+        const verb = event.action === 'deny' ? 'blocked' : 'rewrote';
+        const target = event.toolName ? `tool call "${event.toolName}"` : 'the response';
+        r?.addThought?.(`Guardrail "${event.name}" ${verb} ${target}${event.reason ? ` — ${event.reason}` : ''}`);
+        break;
+      }
+
+      case 'thinking': {
+        const r = ensureReasoning();
+        r?.addThought?.(event.text);
+        break;
+      }
+
+      // 'memory' and 'trace' stay render-silent (see this module's header
+      // comment + the chat README's event table) — they fall to `default`.
 
       case 'error':
+        reasoning?.fail?.(event.error?.message ?? String(event.error));
         chatShell.appendMessage?.({ role: 'error', content: event.error?.message ?? String(event.error) });
         chatShell.stopStreaming?.();
         break;
 
       case 'done':
+        if (reasoning) {
+          if (event.stopReason === 'guardrail_denied') reasoning.fail('Blocked by guardrail');
+          else reasoning.finish();
+        }
         chatShell.stopStreaming?.();
         break;
 

package/dist/chat/chat-shell.min.js

@@ -47,4 +47,4 @@ var RS=Object.defineProperty;var ee=(i,e,t)=>()=>{if(t)throw t[0];try{return i&&
 `):t}async#f(){if(this.#o||this.#e)return;let e=this.#i,t=Zu(this.language);this.#o=!0;try{let n=await Promise.resolve().then(()=>(Au(),Xu)),r=null;try{r=(await n.importWithTimeout(n.languages[t],`lang-${t}`))?.extension??null}catch(s){this.dispatchEvent(new CustomEvent("language-load-error",{bubbles:!0,detail:{phase:"language",language:t,error:s}}))}if(e!==this.#i||!this.isConnected)return;this.#O(n,r,t)}catch(n){this.dispatchEvent(new CustomEvent("language-load-error",{bubbles:!0,detail:{phase:"core",error:n}})),console.warn("[code-ui] CodeMirror failed to load; staying on static fallback.",n)}finally{this.#o=!1}}#O(e,t,n){let{EditorState:r,EditorView:s,Compartment:o,lineNumbers:l,placeholder:a,syntaxHighlighting:h,history:c,historyKeymap:f,defaultKeymap:u,indentWithTab:d,keymap:O,adiaBaseTheme:m,adiaHighlightStyle:g,lintGutter:b,jsonLinter:S}=e,y=this.querySelector(":scope > pre"),C=y?y.querySelector("code")?.textContent??"":"",x=this.text||C||"";this.#l=x;let k=document.createElement("div");k.setAttribute("data-cm-mount",""),y?y.replaceWith(k):this.appendChild(k);let w=[m,h(g)];this.editable&&w.push(c(),O.of([{key:"Mod-s",run:E=>(this.#g(E),!0)},...u,...f,d]),s.updateListener.of(E=>{if(!E.docChanged)return;let j=E.state.doc.toString();this.#d(j),this.dispatchEvent(new CustomEvent("input",{bubbles:!0,detail:{value:j}}))}),s.domEventHandlers({focus:(E,j)=>(this.#n=j.state.doc.toString(),!1),blur:(E,j)=>{let N=j.state.doc.toString();return this.#n!==null&&N!==this.#n&&this.dispatchEvent(new CustomEvent("change",{bubbles:!0,detail:{value:N}})),this.#n=null,!1}})),this.#s=new o,w.push(this.#s.of(this.#u(s,r))),t&&w.push(t),this.lineNumbers&&w.push(l()),this.placeholder&&w.push(a(this.placeholder)),this.editable&&n==="json"&&w.push(S,b()),this.#e=new s({parent:k,state:r.create({doc:x,extensions:w})}),this.editable&&this.#d(x)}#u(e,t){return this.editable&&!this.disabled&&!this.readonly?[]:[e.editable.of(!1),t.readOnly.of(!0)]}async#p(){if(!this.#e||!this.#s)return;let e=await Promise.resolve().then(()=>(Au(),Xu));this.#e.dispatch({effects:this.#s.reconfigure(this.#u(e.EditorView,e.EditorState))})}get form(){return this.internals.form}get labels(){return this.internals.labels}get validity(){return this.internals.validity}get validationMessage(){return this.internals.validationMessage}get willValidate(){return this.internals.willValidate}checkValidity(){return this.internals.checkValidity()}reportValidity(){return this.internals.reportValidity()}#d(e){this.editable&&(e=e??this.value??"",this.internals.setFormValue(e),this.#m(e))}#m(e){return e=e??"",this.required&&!e.trim()?(this.internals.setValidity({valueMissing:!0},this.getAttribute("data-msg-required")||"This field is required.",this),this.setAttribute("aria-invalid","true"),!1):(this.internals.setValidity({}),this.removeAttribute("aria-invalid"),!0)}formResetCallback(){this.#e?this.#e.dispatch({changes:{from:0,to:this.#e.state.doc.length,insert:this.#l}}):this.text=this.#l}formDisabledCallback(e){this.disabled=e}formStateRestoreCallback(e,t){typeof e=="string"&&(this.#e?this.#e.dispatch({changes:{from:0,to:this.#e.state.doc.length,insert:e}}):this.text=e)}#g(e){this.dispatchEvent(new CustomEvent("save",{bubbles:!0,cancelable:!0,detail:{value:e.state.doc.toString()}}))}get value(){return this.#e?this.#e.state.doc.toString():this.querySelector(":scope > pre > code")?.textContent??this.text??""}set value(e){this.text=String(e??"")}#b(){let e=this.#e?this.#e.state.doc.toString():this.querySelector(":scope > pre > code")?.textContent??"";e&&navigator.clipboard.writeText(e).then(()=>{let t=this.querySelector(':scope > header [slot="copy"]');if(!t)return;let n=t.textContent;t.textContent="Copied!",this.#r!=null&&clearTimeout(this.#r),this.#r=setTimeout(()=>{this.#r=null,t.textContent=n},1500)})}render(){let e=this.querySelector(':scope > header [slot="label"]');if(e&&this.language&&(e.textContent=this.language),this.#p(),!this.inline&&this.text!=="")if(this.#e){if(this.editable&&this.#e.hasFocus)return;let t=this.#e.state.doc.toString();t!==this.text&&this.#e.dispatch({changes:{from:0,to:t.length,insert:this.text}})}else{let t=this.querySelector(":scope > pre > code");t&&t.textContent!==this.text&&(t.textContent=this.text)}}};is("code-ui",es);function Ru(i){return i.replace(/&/g,"&amp;").replace(/</g,"&lt;").replace(/>/g,"&gt;")}var Dv=0,jl=class extends an{static properties={streaming:{type:Boolean,default:!1,reflect:!0},provider:{type:String,default:"",reflect:!0},model:{type:String,default:"",reflect:!0},system:{type:String,default:"",reflect:!0},proxyUrl:{type:String,default:"",attribute:"proxy-url",reflect:!0},thinking:{type:Boolean,default:!1,reflect:!0}};static template=()=>null;#t=[];#r=null;#e=null;#i=null;#o=null;#n=null;#l="";get messages(){return[...this.#t]}get conversation(){return this.#t.filter(e=>e.role==="user"||e.role==="assistant").map(e=>e.html?{role:e.role,content:e.content,html:!0}:{role:e.role,content:e.content})}set conversation(e){this.clear();for(let t of e)this.appendMessage({role:t.role,content:t.content,render:!t.html,html:!!t.html})}set apiKey(e){this.#l=e}connected(){this.#e=this.querySelector("chat-thread"),this.#i=this.querySelector("chat-composer"),this.#o=this.querySelector("chat-empty"),this.#n=this.querySelector("chat-status"),this.#i?.addEventListener("composer-submit",this.#a)}disconnected(){this.#i?.removeEventListener("composer-submit",this.#a),this.abort()}#s(e,t){this.dispatchEvent(new CustomEvent(e,{bubbles:!0,detail:t}))}#a=e=>{if(this.streaming)return;let{text:t,model:n}=e.detail||{};t&&(this.#i.clear(),this.#s("submit",{text:t,model:n}),(this.proxyUrl||this.#l)&&this.send(t,n?{model:n}:{}))};appendMessage({role:e,content:t="",render:n=!1,html:r=!1}){let s=`msg_${++Dv}`;this.#t.push({id:s,role:e,content:t,html:r});let o=document.createElement("div");if(o.setAttribute("data-role",e),o.setAttribute("data-id",s),e==="user")o.innerHTML=`<div data-bubble>${Ru(t)}</div>`;else if(e==="assistant"){let l=n||r,a=r?t:n&&t?ql(t):Ru(t);o.innerHTML=`
         <span data-avatar>AI</span>
         <div data-bubble><div data-content>${a}</div>${l?"":"<span data-cursor></span>"}</div>
-      `}else e==="error"&&(o.innerHTML=`<icon-ui name="warning"></icon-ui><span>${Ru(t)}</span>`);return this.#e?.appendChild(o),this.#h(),this.#s("message",{id:s,role:e,content:t}),o}appendChunk(e){let t=this.#t[this.#t.length-1];if(!t||t.role!=="assistant")return;t.content+=e;let n=this.#e?.querySelector("[data-role]:last-child [data-content]");n&&n.insertAdjacentText("beforeend",e),this.#h()}deleteMessage(e){let t=this.#t.findIndex(n=>n.id===e);t!==-1&&(this.#t.splice(t,1),this.#e?.querySelector(`[data-id="${e}"]`)?.remove())}clear(){if(this.#t.length=0,this.#e){let e=this.#o;this.#e.innerHTML="",e&&this.#e.appendChild(e)}this.#s("clear")}startStreaming(){this.streaming=!0,this.#i&&(this.#i.disabled=!0),this.#n&&(this.#n.textContent="Typing..."),this.#e?.tagName?.toLowerCase()==="chat-thread"&&(this.#e.streaming=!0)}stopStreaming(){this.streaming=!1,this.#i&&(this.#i.disabled=!1),this.#n&&(this.#n.textContent=""),this.#e?.tagName?.toLowerCase()==="chat-thread"&&(this.#e.streaming=!1),this.#e?.querySelector("[data-role]:last-child [data-cursor]")?.remove();let e=this.#t[this.#t.length-1];if(e?.role==="assistant"&&e.content&&!e.html){let t=this.#e?.querySelector("[data-role]:last-child [data-content]");t&&(t.innerHTML=ql(e.content),this.#c(t))}this.#i?.focus()}abort(){this.#r&&(this.#r.abort(),this.#r=null,this.#s("abort")),this.streaming&&this.stopStreaming()}async send(e,t={}){let n=t.model||this.model||this.#i?.model;if(!n)throw new Error("No model specified");this.appendMessage({role:"user",content:e}),this.appendMessage({role:"assistant",content:""}),this.startStreaming(),this.#r=new AbortController;try{let r={provider:this.provider||void 0,apiKey:this.#l||void 0,model:n,system:this.system||void 0,proxyUrl:this.proxyUrl||void 0,thinking:this.thinking||void 0,messages:this.conversation.slice(0,-1),signal:this.#r.signal,...t},s;for(let o of["../../../llm/index.js","../../llm/index.js"])try{({streamChat:s}=await import(new URL(o,import.meta.url).href));break}catch{}if(!s){this.appendMessage({role:"error",content:"@adia-ai/llm is not available \u2014 in a fresh checkout run the packages/llm build (see gh#527/gh#541)."}),this.#s("error",{error:new Error("llm build artifact missing")});return}for await(let o of s(r))o.type==="text"?(this.appendChunk(o.text),this.#s("chunk",{text:o.text,snapshot:o.snapshot})):o.type==="thinking"?this.#s("thinking",{text:o.text}):o.type==="done"?this.#s("done",{text:o.text,usage:o.usage,stopReason:o.stopReason}):o.type==="error"&&(this.appendMessage({role:"error",content:o.error.message}),this.#s("error",{error:o.error}))}catch(r){r.name!=="AbortError"&&(this.appendMessage({role:"error",content:r.message}),this.#s("error",{error:r}))}this.#r=null,this.stopStreaming()}export(){return{messages:this.#t.map(e=>({role:e.role,content:e.content})),model:this.model,system:this.system}}import(e){e.model&&(this.model=e.model),e.system&&(this.system=e.system),e.messages&&(this.conversation=e.messages)}#h(){let e=this.#e;e&&requestAnimationFrame(()=>{e.scrollTop=e.scrollHeight})}#c(e){for(let t of e.querySelectorAll("pre")){let n=t.querySelector("code");if(!n)continue;let r=n.getAttribute("data-lang")||"",s=document.createElement("code-ui");r&&s.setAttribute("language",r),s.textContent=n.textContent,t.replaceWith(s)}}};is("chat-shell",jl);export{jl as ChatShell};
+      `}else e==="error"&&(o.innerHTML=`<icon-ui name="warning"></icon-ui><span>${Ru(t)}</span>`);return this.#e?.appendChild(o),this.#h(),this.#s("message",{id:s,role:e,content:t}),o}appendChunk(e){let t=this.#t[this.#t.length-1];if(!t||t.role!=="assistant")return;t.content+=e;let n=this.#e?.querySelector("[data-role]:last-child [data-content]");n&&n.insertAdjacentText("beforeend",e),this.#h()}setStreamedText(e){let t=this.#t[this.#t.length-1];if(!t||t.role!=="assistant")return;t.content=e;let n=this.#e?.querySelector("[data-role]:last-child [data-content]");n&&(n.textContent=e),this.#h()}deleteMessage(e){let t=this.#t.findIndex(n=>n.id===e);t!==-1&&(this.#t.splice(t,1),this.#e?.querySelector(`[data-id="${e}"]`)?.remove())}clear(){if(this.#t.length=0,this.#e){let e=this.#o;this.#e.innerHTML="",e&&this.#e.appendChild(e)}this.#s("clear")}startStreaming(){this.streaming=!0,this.#i&&(this.#i.disabled=!0),this.#n&&(this.#n.textContent="Typing..."),this.#e?.tagName?.toLowerCase()==="chat-thread"&&(this.#e.streaming=!0)}stopStreaming(){this.streaming=!1,this.#i&&(this.#i.disabled=!1),this.#n&&(this.#n.textContent=""),this.#e?.tagName?.toLowerCase()==="chat-thread"&&(this.#e.streaming=!1),this.#e?.querySelector("[data-role]:last-child [data-cursor]")?.remove();let e=this.#t[this.#t.length-1];if(e?.role==="assistant"&&e.content&&!e.html){let t=this.#e?.querySelector("[data-role]:last-child [data-content]");t&&(t.innerHTML=ql(e.content),this.#c(t))}this.#i?.focus()}abort(){this.#r&&(this.#r.abort(),this.#r=null,this.#s("abort")),this.streaming&&this.stopStreaming()}async send(e,t={}){let n=t.model||this.model||this.#i?.model;if(!n)throw new Error("No model specified");this.appendMessage({role:"user",content:e}),this.appendMessage({role:"assistant",content:""}),this.startStreaming(),this.#r=new AbortController;try{let r={provider:this.provider||void 0,apiKey:this.#l||void 0,model:n,system:this.system||void 0,proxyUrl:this.proxyUrl||void 0,thinking:this.thinking||void 0,messages:this.conversation.slice(0,-1),signal:this.#r.signal,...t},s;for(let o of["../../../llm/index.js","../../llm/index.js"])try{({streamChat:s}=await import(new URL(o,import.meta.url).href));break}catch{}if(!s){this.appendMessage({role:"error",content:"@adia-ai/llm is not available \u2014 in a fresh checkout run the packages/llm build (see gh#527/gh#541)."}),this.#s("error",{error:new Error("llm build artifact missing")});return}for await(let o of s(r))o.type==="text"?(this.appendChunk(o.text),this.#s("chunk",{text:o.text,snapshot:o.snapshot})):o.type==="thinking"?this.#s("thinking",{text:o.text}):o.type==="done"?this.#s("done",{text:o.text,usage:o.usage,stopReason:o.stopReason}):o.type==="error"&&(this.appendMessage({role:"error",content:o.error.message}),this.#s("error",{error:o.error}))}catch(r){r.name!=="AbortError"&&(this.appendMessage({role:"error",content:r.message}),this.#s("error",{error:r}))}this.#r=null,this.stopStreaming()}export(){return{messages:this.#t.map(e=>({role:e.role,content:e.content})),model:this.model,system:this.system}}import(e){e.model&&(this.model=e.model),e.system&&(this.system=e.system),e.messages&&(this.conversation=e.messages)}#h(){let e=this.#e;e&&requestAnimationFrame(()=>{e.scrollTop=e.scrollHeight})}#c(e){for(let t of e.querySelectorAll("pre")){let n=t.querySelector("code");if(!n)continue;let r=n.getAttribute("data-lang")||"",s=document.createElement("code-ui");r&&s.setAttribute("language",r),s.textContent=n.textContent,t.replaceWith(s)}}};is("chat-shell",jl);export{jl as ChatShell};