@agent-native/core 0.93.0 → 0.94.0

package/corpus/core/src/shared/reasoning-effort.ts

@@ -11,6 +11,13 @@ export const REASONING_EFFORTS = [
 
 export type ReasoningEffort = (typeof REASONING_EFFORTS)[number];
 
+/**
+ * Shared chat always chooses an explicit reasoning tier. Keep `auto` in the
+ * accepted type only so older persisted selections and external callers can
+ * migrate cleanly; new UI and engine defaults resolve it to Medium.
+ */
+export const DEFAULT_REASONING_EFFORT: ReasoningEffort = "medium";
+
 export const REASONING_EFFORT_LABELS: Record<ReasoningEffort, string> = {
   auto: "Auto",
   none: "None",
@@ -22,12 +29,7 @@ export const REASONING_EFFORT_LABELS: Record<ReasoningEffort, string> = {
   max: "Max",
 };
 
-const VISIBLE_STANDARD_EFFORTS: ReasoningEffort[] = [
-  "auto",
-  "low",
-  "medium",
-  "high",
-];
+const VISIBLE_STANDARD_EFFORTS: ReasoningEffort[] = ["low", "medium", "high"];
 
 const VISIBLE_GPT_EFFORTS: ReasoningEffort[] = [
   ...VISIBLE_STANDARD_EFFORTS,
@@ -73,10 +75,9 @@ export function normalizeReasoningEffortForModel(
   model: string | undefined,
   effort: ReasoningEffort | undefined,
 ): ReasoningEffort | undefined {
-  if (!model || !effort || effort === "auto") {
-    return undefined;
-  }
-  let normalized = effort;
+  if (!model) return undefined;
+  let normalized =
+    !effort || effort === "auto" ? DEFAULT_REASONING_EFFORT : effort;
   if (
     normalized === "xhigh" &&
     isClaudeReasoningModel(model) &&
@@ -94,13 +95,47 @@ export function normalizeReasoningEffortForModel(
   return normalized;
 }
 
+/**
+ * Normalize a chat request before it reaches an engine. Explicit off/minimal
+ * sentinels must survive this layer so the engine can distinguish them from a
+ * missing selection, which now means the Medium default.
+ */
+export function normalizeReasoningEffortForRequest(
+  model: string | undefined,
+  effort: ReasoningEffort | undefined,
+): ReasoningEffort | undefined {
+  if (effort === "none" || effort === "minimal") return effort;
+  return normalizeReasoningEffortForModel(model, effort);
+}
+
 export function reasoningEffortLabel(effort: ReasoningEffort | undefined) {
-  return REASONING_EFFORT_LABELS[effort ?? "auto"];
+  return REASONING_EFFORT_LABELS[
+    !effort || effort === "auto" ? DEFAULT_REASONING_EFFORT : effort
+  ];
+}
+
+/**
+ * Resolve a user-facing selection for a model. Legacy `auto`, missing values,
+ * and tiers unsupported by the newly selected model all become Medium.
+ * Non-reasoning models still retain Medium in persisted chat state so moving
+ * back to a reasoning model has a predictable default; their engines omit the
+ * effort through `normalizeReasoningEffortForModel`.
+ */
+export function resolveReasoningEffortSelection(
+  model: string | undefined,
+  effort: ReasoningEffort | undefined,
+): ReasoningEffort {
+  const requested =
+    !effort || effort === "auto" ? DEFAULT_REASONING_EFFORT : effort;
+  const options = getReasoningEffortOptionsForModel(model);
+  return options.length === 0 || options.includes(requested)
+    ? requested
+    : DEFAULT_REASONING_EFFORT;
 }
 
 /**
- * One tier down from each effort, stopping at "minimal" — "none"/"auto"
- * (not really "tiers") and "minimal" itself are left unchanged. Used by the
+ * One tier down from each effort, stopping at "minimal" — legacy `auto`,
+ * "none", and "minimal" itself are left unchanged. Used by the
  * empty-final-response retry so a retried turn asks for meaningfully less
  * reasoning instead of repeating the exact request that came back empty.
  */
@@ -122,24 +157,31 @@ export function stepDownReasoningEffort(
 }
 
 function isGPTReasoningModel(model: string) {
-  return /^gpt-5/.test(model) || /^o\d/.test(model);
+  const id = model.toLowerCase().replace(/^openai\//, "");
+  return /^gpt-5/.test(id) || /^o\d/.test(id);
 }
 
 function isClaudeReasoningModel(model: string) {
-  return /^claude-/.test(model);
+  const id = model.toLowerCase().replace(/^anthropic\//, "");
+  if (id.includes("fable-5") || id.includes("mythos-5")) return true;
+  if (id.includes("sonnet-5") || id.includes("sonnet-4-6")) return true;
+  if (id.includes("haiku-4-5")) return true;
+  const opusMatch = id.match(/opus-4[-.](\d+)/);
+  return opusMatch ? parseInt(opusMatch[1], 10) >= 6 : false;
 }
 
 function supportsClaudeXHigh(model: string) {
+  const id = model.toLowerCase().replace(/^anthropic\//, "");
   // Models that support the xhigh effort tier (built-in extended thinking via
   // output_config.effort). Keep this version-aware so any future Claude model
   // with a higher patch/minor number is automatically included rather than
   // silently falling back to the lower "high" tier.
   // claude-fable-5 is a Mythos-class model and also supports xhigh.
-  if (model.includes("fable-5")) return true;
+  if (id.includes("fable-5")) return true;
   // Sonnet 5 supports the expanded effort ladder through Builder/Anthropic.
-  if (model.includes("sonnet-5")) return true;
+  if (id.includes("sonnet-5")) return true;
   // opus-4-7 introduced xhigh; all opus-4.x successors (4-8, 4-9…) should too.
-  const opusMatch = model.match(/opus-4-(\d+)/);
+  const opusMatch = id.match(/opus-4[-.](\d+)/);
   if (opusMatch) {
     return parseInt(opusMatch[1], 10) >= 7;
   }
@@ -147,5 +189,5 @@ function supportsClaudeXHigh(model: string) {
 }
 
 function isGeminiReasoningModel(model: string) {
-  return /^gemini-/.test(model);
+  return /^gemini-/.test(model.toLowerCase().replace(/^google\//, ""));
 }

package/corpus/core/src/styles/agent-conversation.css

@@ -476,6 +476,23 @@
   }
 }
 
+/* ─── ReasoningCell streaming tail clamp ────────────────────────────────────── */
+/* While a reasoning cell is actively streaming and open, clamp the body to
+   roughly the last ~6 lines instead of letting it grow and push the rest of
+   the transcript around (the Claude Code "tail view" behavior). The reversed
+   column anchors the box to its bottom edge, so overflow is clipped off the
+   top while the newest lines stay pinned in view — pure CSS, no scroll/rAF
+   tracking needed. A soft fade masks the clipped edge in both themes since
+   it only touches alpha, not color. */
+.reasoning-cell-tail {
+  display: flex;
+  flex-direction: column-reverse;
+  overflow: hidden;
+  max-height: 8rem;
+  mask-image: linear-gradient(to bottom, transparent, black 1.5rem);
+  -webkit-mask-image: linear-gradient(to bottom, transparent, black 1.5rem);
+}
+
 /* ─── Agent-Markdown (AssistantChat react-markdown render surface) ─────────── */
 /* These classes are applied by the SmoothMarkdownText/MarkdownText components  */
 /* in the AssistantChat/AssistantChatInner tree. The shiki dual-theme rules      */

package/corpus/core/src/templates/workspace-core/.agents/skills/voice-transcription/SKILL.md

@@ -13,17 +13,21 @@ metadata:
 
 # Voice Transcription
 
-Click-to-toggle microphone inside the sidebar composer turns speech into
-text. Users configure live transcription separately from AI cleanup in
-Settings → Voice Transcription. The feature is available in every template
-that renders `TiptapComposer`.
+The microphone inside the sidebar composer offers two distinct paths:
+editable dictation and an opt-in realtime speech-to-speech agent session.
+Users configure dictation separately from AI cleanup in Settings → Voice
+Transcription. Both paths are available in every template that renders the
+shared agent sidebar.
 
 ## UX rules
 
 - **Always show the mic alongside the send button.** Cursor replaces send
   with mic when the composer is empty; their users complain. We keep both
   visible — Lovable does the same.
-- **Click-to-toggle, not push-to-talk.** More forgiving in a sidebar, avoids
+- **Ask before switching modes.** The mic opens a shadcn popover offering
+  **Start voice mode** and **Keep dictating**; never silently replace the
+  established dictation behavior.
+- **Dictation is click-to-toggle, not push-to-talk.** More forgiving in a sidebar, avoids
   host-app hotkey clashes. Keyboard shortcut is `Cmd/Ctrl+Shift+M` and
   `Escape` cancels mid-recording.
 - **Transcript lands in the composer, editable, never auto-sent.** Insert at
@@ -34,6 +38,31 @@ that renders `TiptapComposer`.
   Never use a sparkle or robot icon.
 - **Errors via inline alert or toast, never `window.alert`.**
 
+## Realtime speech mode
+
+Realtime speech mode is a full-duplex OpenAI Realtime WebRTC session, separate
+from the dictation providers below. It requires the authenticated user's
+`OPENAI_API_KEY`, uses the server-owned `gpt-realtime-2.1` session
+configuration, and keeps the key out of browser code.
+
+- Starting voice mode collapses the chat into a persistent bottom-end speech
+  orb. Clicking the orb shows or hides chat without ending the session; ending
+  the session is a separate, explicit control.
+- Semantic VAD keeps listening, starts responses automatically, and supports
+  barge-in while the agent is speaking.
+- Function calls must cross the authenticated realtime tool bridge and enter
+  `executeAgentToolCall`. Never call `ActionEntry.run` directly: that bypasses
+  schema validation, approvals, audit/journal behavior, timeouts, redaction,
+  and mutation refreshes.
+- Preserve the active browser-tab id in request context so `set-url`,
+  `set-search-params`, `view-screen`, and tab-scoped application state affect
+  the app the user is actually speaking to.
+- Do not persist audio. Store only compact lifecycle and latest-transcript
+  context at `application_state["realtime-voice-session"]`; delete it when the
+  session ends.
+- Missing-key failures should open Settings focused on the user-scoped
+  `OPENAI_API_KEY` field. Never send, log, or echo the key to the browser.
+
 ## Source And Cleanup
 
 Settings must keep these as separate choices:
@@ -80,11 +109,14 @@ Default behavior:
 | --------------------------------------------------------------------- | --------------------------------------------------- |
 | `packages/core/src/client/composer/useVoiceDictation.ts`              | Provider-routing hook (MediaRecorder / Web Speech)  |
 | `packages/core/src/client/composer/VoiceButton.tsx`                   | Mic button + live amplitude + cancel overlay        |
+| `packages/core/src/client/composer/RealtimeVoiceMode.tsx`            | Opt-in popover + persistent speech orb              |
+| `packages/core/src/client/composer/useRealtimeVoiceMode.tsx`         | WebRTC lifecycle, provider events, and tool bridge  |
 | `packages/core/src/client/composer/TiptapComposer.tsx`                | Wires the hook, insertion, and keyboard shortcut    |
 | `packages/core/src/client/settings/VoiceTranscriptionSection.tsx`     | Live source + cleanup controls in sidebar settings  |
 | `packages/core/src/client/transcription/BuilderTranscriptionCta.tsx`  | CTA shown when Builder account isn't connected      |
 | `packages/core/src/client/transcription/use-live-transcription.ts`    | Web Speech live-transcription hook for recordings   |
 | `packages/core/src/server/transcribe-voice.ts`                        | Route handler (routes to Builder/Gemini/Groq/Whisper) |
+| `packages/core/src/server/realtime-voice.ts`                          | Authenticated OpenAI SDP and action-tool routes     |
 | `packages/core/src/transcription/builder-transcription.ts`            | Builder proxy transcription client                  |
 | `packages/core/src/voice/`                                            | Shared voice context pack, prompt, and replacement helpers |
 | `packages/core/src/secrets/register-framework-secrets.ts`             | Framework-level provider key registration           |
@@ -139,4 +171,6 @@ Templates can:
 - Don't call transcription providers from the client — go through `/_agent-native/transcribe-voice` so the user's secret stays server-side.
 - Don't remove the cancel affordance — mic permission abuse paranoia is real.
 - Don't auto-submit the transcript — users always edit before sending.
+- Don't route realtime tool calls around the central guarded agent executor.
+- Don't persist realtime audio or put provider keys in SDP/session payloads.
 - Don't copy Cursor's "hide send when empty" pattern — it confuses users.

package/corpus/templates/brain/app/lib/settings-navigation.ts

@@ -0,0 +1,22 @@
+export function createSettingsSectionIds(
+  appSectionIds: Iterable<string>,
+): Set<string> {
+  return new Set(["general", "team", "whats-new", ...appSectionIds]);
+}
+
+export function resolveSettingsSection(
+  section: string | null,
+  validSections: ReadonlySet<string>,
+): string {
+  return section && validSections.has(section) ? section : "general";
+}
+
+export function withSettingsSection(
+  search: URLSearchParams,
+  section: string,
+): URLSearchParams {
+  const next = new URLSearchParams(search);
+  if (section === "general") next.delete("section");
+  else next.set("section", section);
+  return next;
+}