feat: add uncertainty visualization with token-level logprobs

- Add TokenHeatmap component for visualizing token confidence
- Collect and stream logprobs in generation pipeline
- Add regenerate-from-token feature with continue_from_prefix
- Add AbortController for request cancellation
- Support continue_final_message for seamless prefix continuation

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

.github/workflows/build-app.yml

@@ -1,5 +1,16 @@
 name: Build EXO macOS DMG
 
+# Release workflow:
+# 1. Create a draft GitHub Release with the tag name (e.g. v1.0.0) and write release notes in markdown
+# 2. Push the tag: git tag v1.0.0 && git push origin v1.0.0
+# 3. This workflow builds, signs, and notarizes the DMG
+# 4. Release notes are embedded in appcast.xml for Sparkle (rendered as markdown)
+# 5. DMG and appcast.xml are uploaded to S3
+# 6. The draft GitHub Release is published with the DMG attached
+#
+# For alpha releases (e.g. v1.0.0-alpha.1): draft release and notes are optional.
+# If no draft exists, a release is auto-created with generated notes.
+
 on:
   workflow_dispatch:
   push:
@@ -11,8 +22,10 @@ on:
 jobs:
   build-macos-app:
     runs-on: "macos-26"
+    permissions:
+      contents: write
     env:
-      SPARKLE_VERSION: 2.8.1
+      SPARKLE_VERSION: 2.9.0-beta.1
       SPARKLE_DOWNLOAD_PREFIX: ${{ secrets.SPARKLE_DOWNLOAD_PREFIX }}
       SPARKLE_FEED_URL: ${{ secrets.SPARKLE_FEED_URL }}
       SPARKLE_ED25519_PUBLIC: ${{ secrets.SPARKLE_ED25519_PUBLIC }}
@@ -87,6 +100,52 @@ jobs:
             exit 1
           fi
 
+      - name: Fetch and validate release notes
+        if: github.ref_type == 'tag'
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          # Find draft release by name using gh release list (more reliable with default token)
+          echo "Looking for draft release named '$GITHUB_REF_NAME'..."
+          DRAFT_EXISTS=$(gh release list --json name,isDraft --jq ".[] | select(.isDraft == true) | select(.name == \"$GITHUB_REF_NAME\") | .name" 2>/dev/null || echo "")
+
+          if [[ -z "$DRAFT_EXISTS" ]]; then
+            if [[ "$IS_ALPHA" == "true" ]]; then
+              echo "No draft release found for alpha tag $GITHUB_REF_NAME (optional for alphas)"
+              echo "HAS_RELEASE_NOTES=false" >> $GITHUB_ENV
+              exit 0
+            fi
+            echo "ERROR: No draft release found for tag $GITHUB_REF_NAME"
+            echo "Please create a draft release with release notes before pushing the tag."
+            exit 1
+          fi
+
+          # Fetch full release details via API to get body and ID
+          echo "Found draft release, fetching details..."
+          RELEASE_JSON=$(gh api repos/${{ github.repository }}/releases --jq ".[] | select(.draft == true) | select(.name == \"$GITHUB_REF_NAME\")" 2>/dev/null || echo "")
+
+          # Extract release notes
+          NOTES=$(echo "$RELEASE_JSON" | jq -r '.body // ""')
+          if [[ -z "$NOTES" || "$NOTES" == "null" ]]; then
+            if [[ "$IS_ALPHA" == "true" ]]; then
+              echo "Draft release has no notes (optional for alphas)"
+              echo "HAS_RELEASE_NOTES=false" >> $GITHUB_ENV
+              exit 0
+            fi
+            echo "ERROR: Draft release exists but has no release notes"
+            echo "Please add release notes to the draft release before pushing the tag."
+            exit 1
+          fi
+
+          # Save release ID for later publishing
+          RELEASE_ID=$(echo "$RELEASE_JSON" | jq -r '.id')
+          echo "DRAFT_RELEASE_ID=$RELEASE_ID" >> $GITHUB_ENV
+          echo "HAS_RELEASE_NOTES=true" >> $GITHUB_ENV
+
+          echo "Found draft release (ID: $RELEASE_ID), saving release notes..."
+          echo "$NOTES" > /tmp/release_notes.md
+          echo "RELEASE_NOTES_FILE=/tmp/release_notes.md" >> $GITHUB_ENV
+
       # ============================================================
       # Install dependencies
       # ============================================================
@@ -304,6 +363,28 @@ jobs:
             $CHANNEL_FLAG \
             .
 
+      - name: Inject release notes into appcast
+        if: github.ref_type == 'tag' && env.HAS_RELEASE_NOTES == 'true'
+        env:
+          RELEASE_VERSION: ${{ env.RELEASE_VERSION }}
+        run: |
+          # Inject markdown release notes with sparkle:format="markdown" (Sparkle 2.9+)
+          export NOTES=$(cat "$RELEASE_NOTES_FILE")
+
+          # Insert description after the enclosure tag for this version
+          awk '
+            /<enclosure[^>]*>/ && index($0, ENVIRON["RELEASE_VERSION"]) {
+              print
+              print "            <description sparkle:format=\"markdown\"><![CDATA["
+              print ENVIRON["NOTES"]
+              print "            ]]></description>"
+              next
+            }
+            { print }
+          ' output/appcast.xml > output/appcast.xml.tmp && mv output/appcast.xml.tmp output/appcast.xml
+
+          echo "Injected markdown release notes for version $RELEASE_VERSION"
+
       # ============================================================
       # Upload artifacts
       # ============================================================
@@ -336,3 +417,26 @@ jobs:
             aws s3 cp "$DMG_NAME" "s3://${SPARKLE_S3_BUCKET}/${PREFIX}EXO-latest.dmg"
             aws s3 cp appcast.xml "s3://${SPARKLE_S3_BUCKET}/${PREFIX}appcast.xml" --content-type application/xml --cache-control no-cache
           fi
+
+      - name: Publish GitHub Release
+        if: github.ref_type == 'tag'
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          DMG_PATH="output/EXO-${RELEASE_VERSION}.dmg"
+
+          if [[ "$HAS_RELEASE_NOTES" == "true" ]]; then
+            # Update the draft release with the tag and upload DMG
+            gh api --method PATCH "repos/${{ github.repository }}/releases/$DRAFT_RELEASE_ID" \
+              -f tag_name="$GITHUB_REF_NAME" \
+              -F draft=false
+            gh release upload "$GITHUB_REF_NAME" "$DMG_PATH" --clobber
+            echo "Published release $GITHUB_REF_NAME with DMG attached"
+          else
+            # Alpha without draft release - create one with auto-generated notes
+            gh release create "$GITHUB_REF_NAME" "$DMG_PATH" \
+              --title "$GITHUB_REF_NAME" \
+              --generate-notes \
+              --prerelease
+            echo "Created alpha release $GITHUB_REF_NAME with auto-generated notes"
+          fi

AGENTS.md

@@ -40,6 +40,31 @@ uv run ruff check
 nix fmt
 ```
 
+## Pre-Commit Checks (REQUIRED)
+
+**IMPORTANT: Always run these checks before committing code. CI will fail if these don't pass.**
+
+```bash
+# 1. Type checking - MUST pass with 0 errors
+uv run basedpyright
+
+# 2. Linting - MUST pass
+uv run ruff check
+
+# 3. Formatting - MUST be applied
+nix fmt
+
+# 4. Tests - MUST pass
+uv run pytest
+```
+
+Run all checks in sequence:
+```bash
+uv run basedpyright && uv run ruff check && nix fmt && uv run pytest
+```
+
+If `nix fmt` changes any files, stage them before committing. The CI runs `nix flake check` which verifies formatting, linting, and runs Rust tests.
+
 ## Architecture
 
 ### Node Composition

app/EXO/EXO.xcodeproj/project.pbxproj

@@ -585,7 +585,7 @@
 			repositoryURL = "https://github.com/sparkle-project/Sparkle.git";
 			requirement = {
 				kind = upToNextMajorVersion;
-				minimumVersion = 2.8.1;
+				minimumVersion = 2.9.0-beta.1;
 			};
 		};
 /* End XCRemoteSwiftPackageReference section */

app/EXO/EXO.xcodeproj/project.xcworkspace/xcshareddata/swiftpm/Package.resolved

@@ -6,8 +6,8 @@
       "kind" : "remoteSourceControl",
       "location" : "https://github.com/sparkle-project/Sparkle.git",
       "state" : {
-        "revision" : "5581748cef2bae787496fe6d61139aebe0a451f6",
-        "version" : "2.8.1"
+        "revision" : "e641adb41915a8409895e2e30666aa64e487b637",
+        "version" : "2.9.0-beta.1"
       }
     }
   ],

bench/exo_bench.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import argparse
+import contextlib
 import http.client
 import json
 import os
@@ -26,7 +27,7 @@ class ExoHttpError(RuntimeError):
 
 
 class ExoClient:
-    def __init__(self, host: str, port: int, timeout_s: float = 2400.0):
+    def __init__(self, host: str, port: int, timeout_s: float = 600.0):
         self.host = host
         self.port = port
         self.timeout_s = timeout_s
@@ -104,22 +105,46 @@ def runner_ready(runner: dict[str, Any]) -> bool:
     return "RunnerReady" in runner
 
 
+def runner_failed(runner: dict[str, Any]) -> bool:
+    return "RunnerFailed" in runner
+
+
+def get_runner_failed_message(runner: dict[str, Any]) -> str | None:
+    if "RunnerFailed" in runner:
+        return runner["RunnerFailed"].get("errorMessage")
+    return None
+
+
 def wait_for_instance_ready(
     client: ExoClient, instance_id: str, timeout: float = 24000.0
 ) -> None:
     start_time = time.time()
+    instance_existed = False
     while time.time() - start_time < timeout:
         state = client.request_json("GET", "/state")
         instances = state.get("instances", {})
 
         if instance_id not in instances:
+            if instance_existed:
+                # Instance was deleted after being created - likely due to runner failure
+                raise RuntimeError(
+                    f"Instance {instance_id} was deleted (runner may have failed)"
+                )
             time.sleep(0.1)
             continue
 
+        instance_existed = True
         instance = instances[instance_id]
         runner_ids = runner_ids_from_instance(instance)
         runners = state.get("runners", {})
 
+        # Check for failed runners first
+        for rid in runner_ids:
+            runner = runners.get(rid, {})
+            if runner_failed(runner):
+                error_msg = get_runner_failed_message(runner) or "Unknown error"
+                raise RuntimeError(f"Runner {rid} failed: {error_msg}")
+
         if all(runner_ready(runners.get(rid, {})) for rid in runner_ids):
             return
 
@@ -299,6 +324,12 @@ def main() -> int:
         default=4,
         help="Only consider placements using <= this many nodes.",
     )
+    ap.add_argument(
+        "--min-nodes",
+        type=int,
+        default=1,
+        help="Only consider placements using >= this many nodes.",
+    )
     ap.add_argument(
         "--instance-meta", choices=["ring", "jaccl", "both"], default="both"
     )
@@ -320,7 +351,7 @@ def main() -> int:
         help="Warmup runs per placement (uses first pp/tg).",
     )
     ap.add_argument(
-        "--timeout", type=float, default=2400.0, help="HTTP timeout (seconds)."
+        "--timeout", type=float, default=600.0, help="HTTP timeout (seconds)."
     )
     ap.add_argument(
         "--json-out",
@@ -399,7 +430,7 @@ def main() -> int:
         ):
             continue
 
-        if 0 < n <= args.max_nodes:
+        if args.min_nodes <= n <= args.max_nodes:
             selected.append(p)
 
     if not selected:
@@ -441,7 +472,13 @@ def main() -> int:
         )
 
         client.request_json("POST", "/instance", body={"instance": instance})
-        wait_for_instance_ready(client, instance_id)
+        try:
+            wait_for_instance_ready(client, instance_id)
+        except (RuntimeError, TimeoutError) as e:
+            logger.error(f"Failed to initialize placement: {e}")
+            with contextlib.suppress(ExoHttpError):
+                client.request_json("DELETE", f"/instance/{instance_id}")
+            continue
 
         time.sleep(1)
 

dashboard/src/lib/components/ChatMessages.svelte

@@ -1,14 +1,16 @@
 <script lang="ts">
-	import { 
-		messages, 
-		currentResponse, 
+	import {
+		messages,
+		currentResponse,
 		isLoading,
 		deleteMessage,
 		editAndRegenerate,
-		regenerateLastResponse
+		regenerateLastResponse,
+		regenerateFromToken
 	} from '$lib/stores/app.svelte';
 	import type { MessageAttachment } from '$lib/stores/app.svelte';
 	import MarkdownContent from './MarkdownContent.svelte';
+	import TokenHeatmap from './TokenHeatmap.svelte';
 
 	interface Props {
 		class?: string;
@@ -95,6 +97,23 @@
 let copiedMessageId = $state<string | null>(null);
 let expandedThinkingMessageIds = $state<Set<string>>(new Set());
 
+// Uncertainty view state - tracks which messages show token heatmap
+let uncertaintyViewMessageIds = $state<Set<string>>(new Set());
+
+function toggleUncertaintyView(messageId: string) {
+	const newSet = new Set(uncertaintyViewMessageIds);
+	if (newSet.has(messageId)) {
+		newSet.delete(messageId);
+	} else {
+		newSet.add(messageId);
+	}
+	uncertaintyViewMessageIds = newSet;
+}
+
+function isUncertaintyViewEnabled(messageId: string): boolean {
+	return uncertaintyViewMessageIds.has(messageId);
+}
+
 	function formatTimestamp(timestamp: number): string {
 		return new Date(timestamp).toLocaleTimeString('en-US', { 
 			hour12: false,
@@ -366,7 +385,17 @@ function isThinkingExpanded(messageId: string): boolean {
 									</div>
 								{/if}
 								<div class="text-xs text-foreground">
-									<MarkdownContent content={message.content || (loading ? response : '')} />
+									{#if message.role === 'assistant' && isUncertaintyViewEnabled(message.id) && message.tokens && message.tokens.length > 0}
+										<!-- Uncertainty heatmap view -->
+										<TokenHeatmap
+											tokens={message.tokens}
+											isGenerating={loading}
+											onRegenerateFrom={(tokenIndex) => regenerateFromToken(message.id, tokenIndex)}
+										/>
+									{:else}
+										<!-- Normal markdown view -->
+										<MarkdownContent content={message.content || (loading ? response : '')} />
+									{/if}
 									{#if loading && !message.content}
 										<span class="inline-block w-2 h-4 bg-exo-yellow/70 ml-1 cursor-blink"></span>
 									{/if}
@@ -419,7 +448,20 @@ function isThinkingExpanded(messageId: string): boolean {
 								</svg>
 							</button>
 						{/if}
-						
+
+						<!-- Uncertainty view toggle (assistant messages with tokens only) -->
+						{#if message.role === 'assistant' && message.tokens && message.tokens.length > 0}
+							<button
+								onclick={() => toggleUncertaintyView(message.id)}
+								class="p-1.5 transition-colors rounded cursor-pointer {isUncertaintyViewEnabled(message.id) ? 'text-exo-yellow' : 'text-exo-light-gray hover:text-exo-yellow'}"
+								title={isUncertaintyViewEnabled(message.id) ? 'Hide uncertainty' : 'Show uncertainty'}
+							>
+								<svg class="w-3.5 h-3.5" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+									<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M9 19v-6a2 2 0 00-2-2H5a2 2 0 00-2 2v6a2 2 0 002 2h2a2 2 0 002-2zm0 0V9a2 2 0 012-2h2a2 2 0 012 2v10m-6 0a2 2 0 002 2h2a2 2 0 002-2m0 0V5a2 2 0 012-2h2a2 2 0 012 2v14a2 2 0 01-2 2h-2a2 2 0 01-2-2z" />
+								</svg>
+							</button>
+						{/if}
+
 						<!-- Delete button -->
 						<button
 							onclick={() => handleDeleteClick(message.id)}

dashboard/src/lib/components/TokenHeatmap.svelte

@@ -0,0 +1,192 @@
+<script lang="ts">
+	import type { TokenData } from '$lib/stores/app.svelte';
+
+	interface Props {
+		tokens: TokenData[];
+		class?: string;
+		isGenerating?: boolean;
+		onRegenerateFrom?: (tokenIndex: number) => void;
+	}
+
+	let { tokens, class: className = '', isGenerating = false, onRegenerateFrom }: Props = $props();
+
+	// Tooltip state - track both token data and index
+	let hoveredTokenIndex = $state<number | null>(null);
+	let hoveredPosition = $state<{ x: number; y: number } | null>(null);
+	let isTooltipHovered = $state(false);
+	let hideTimeoutId: ReturnType<typeof setTimeout> | null = null;
+
+	// Derive the hovered token from the index (stable across re-renders)
+	const hoveredToken = $derived(
+		hoveredTokenIndex !== null && hoveredPosition && tokens[hoveredTokenIndex]
+			? { token: tokens[hoveredTokenIndex], index: hoveredTokenIndex, ...hoveredPosition }
+			: null
+	);
+
+	/**
+	 * Get confidence styling based on probability.
+	 * Following Apple design principles: high confidence tokens blend in,
+	 * only uncertainty draws attention.
+	 */
+	function getConfidenceClass(probability: number): string {
+		if (probability > 0.8) return 'text-inherit'; // Expected tokens - blend in
+		if (probability > 0.5) return 'bg-gray-500/10 text-inherit'; // Slight hint
+		if (probability > 0.2) return 'bg-amber-500/15 text-amber-200/90'; // Subtle warmth
+		return 'bg-red-500/20 text-red-200/90'; // Draws attention
+	}
+
+	/**
+	 * Get border/underline styling for uncertain tokens
+	 */
+	function getBorderClass(probability: number): string {
+		if (probability > 0.8) return 'border-transparent'; // No border for expected
+		if (probability > 0.5) return 'border-gray-500/20';
+		if (probability > 0.2) return 'border-amber-500/30';
+		return 'border-red-500/40';
+	}
+
+	function clearHideTimeout() {
+		if (hideTimeoutId) {
+			clearTimeout(hideTimeoutId);
+			hideTimeoutId = null;
+		}
+	}
+
+	function handleMouseEnter(event: MouseEvent, token: TokenData, index: number) {
+		clearHideTimeout();
+		const rect = (event.target as HTMLElement).getBoundingClientRect();
+		hoveredTokenIndex = index;
+		hoveredPosition = {
+			x: rect.left + rect.width / 2,
+			y: rect.top - 10
+		};
+	}
+
+	function handleMouseLeave() {
+		clearHideTimeout();
+		// Use longer delay during generation to account for re-renders
+		const delay = isGenerating ? 300 : 100;
+		hideTimeoutId = setTimeout(() => {
+			if (!isTooltipHovered) {
+				hoveredTokenIndex = null;
+				hoveredPosition = null;
+			}
+		}, delay);
+	}
+
+	function handleTooltipEnter() {
+		clearHideTimeout();
+		isTooltipHovered = true;
+	}
+
+	function handleTooltipLeave() {
+		isTooltipHovered = false;
+		hoveredTokenIndex = null;
+		hoveredPosition = null;
+	}
+
+	function handleRegenerate() {
+		if (hoveredToken && onRegenerateFrom) {
+			const indexToRegenerate = hoveredToken.index;
+			// Clear hover state immediately
+			hoveredTokenIndex = null;
+			hoveredPosition = null;
+			isTooltipHovered = false;
+			// Call regenerate
+			onRegenerateFrom(indexToRegenerate);
+		}
+	}
+
+	function formatProbability(prob: number): string {
+		return (prob * 100).toFixed(1) + '%';
+	}
+
+	function formatLogprob(logprob: number): string {
+		return logprob.toFixed(3);
+	}
+
+	function getProbabilityColor(probability: number): string {
+		if (probability > 0.8) return 'text-gray-300';
+		if (probability > 0.5) return 'text-gray-400';
+		if (probability > 0.2) return 'text-amber-400';
+		return 'text-red-400';
+	}
+</script>
+
+<div class="token-heatmap leading-relaxed {className}">
+	{#each tokens as tokenData, i (i)}
+		<span
+			role="button"
+			tabindex="0"
+			class="token-span inline rounded px-0.5 py-0.5 cursor-pointer transition-all duration-150 border {getConfidenceClass(tokenData.probability)} {getBorderClass(tokenData.probability)} hover:opacity-80"
+			onmouseenter={(e) => handleMouseEnter(e, tokenData, i)}
+			onmouseleave={handleMouseLeave}
+		>{tokenData.token}</span>
+	{/each}
+</div>
+
+<!-- Tooltip -->
+{#if hoveredToken}
+	<div
+		class="fixed z-50"
+		style="left: {hoveredToken.x}px; top: {hoveredToken.y}px; transform: translate(-50%, -100%);"
+		onmouseenter={handleTooltipEnter}
+		onmouseleave={handleTooltipLeave}
+	>
+		<div class="bg-gray-900/95 backdrop-blur-sm border border-gray-700/50 rounded-xl shadow-xl p-3 text-sm min-w-48">
+			<!-- Token info -->
+			<div class="mb-2">
+				<span class="text-gray-500 text-xs">Token:</span>
+				<span class="text-white font-mono ml-1">"{hoveredToken.token.token}"</span>
+				<span class="{getProbabilityColor(hoveredToken.token.probability)} ml-2">{formatProbability(hoveredToken.token.probability)}</span>
+			</div>
+
+			<div class="text-gray-400 text-xs mb-1">
+				logprob: <span class="text-gray-300 font-mono">{formatLogprob(hoveredToken.token.logprob)}</span>
+			</div>
+
+			<!-- Top alternatives -->
+			{#if hoveredToken.token.topLogprobs.length > 0}
+				<div class="border-t border-gray-700/50 mt-2 pt-2">
+					<div class="text-gray-500 text-xs mb-1">Alternatives:</div>
+					{#each hoveredToken.token.topLogprobs.slice(0, 5) as alt, idx (idx)}
+						{@const altProb = Math.exp(alt.logprob)}
+						<div class="flex justify-between items-center text-xs py-0.5">
+							<span class="text-gray-300 font-mono truncate max-w-24">"{alt.token}"</span>
+							<span class="text-gray-400 ml-2">{formatProbability(altProb)}</span>
+						</div>
+					{/each}
+				</div>
+			{/if}
+
+			<!-- Regenerate button -->
+			{#if onRegenerateFrom}
+				<button
+					onclick={handleRegenerate}
+					class="w-full mt-2 pt-2 border-t border-gray-700/50 flex items-center justify-center gap-1.5 text-xs text-gray-400 hover:text-white transition-colors cursor-pointer"
+				>
+					<svg class="w-3 h-3" fill="none" viewBox="0 0 24 24" stroke="currentColor">
+						<path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M4 4v5h.582m15.356 2A8.001 8.001 0 004.582 9m0 0H9m11 11v-5h-.581m0 0a8.003 8.003 0 01-15.357-2m15.357 2H15" />
+					</svg>
+					Regenerate from here
+				</button>
+			{/if}
+		</div>
+		<!-- Arrow -->
+		<div class="absolute left-1/2 -translate-x-1/2 top-full">
+			<div class="border-8 border-transparent border-t-gray-900"></div>
+		</div>
+	</div>
+{/if}
+
+<style>
+	.token-heatmap {
+		word-wrap: break-word;
+		white-space: pre-wrap;
+	}
+
+	.token-span {
+		margin: 0;
+		border-width: 1px;
+	}
+</style>

dashboard/src/lib/stores/app.svelte.ts

@@ -182,6 +182,20 @@ export interface MessageAttachment {
 	mimeType?: string;
 }
 
+// Token-level data for uncertainty visualization
+export interface TopLogprob {
+	token: string;
+	logprob: number;
+	bytes?: number[];
+}
+
+export interface TokenData {
+	token: string;
+	logprob: number;
+	probability: number; // exp(logprob)
+	topLogprobs: TopLogprob[];
+}
+
 export interface Message {
 	id: string;
 	role: "user" | "assistant" | "system";
@@ -191,6 +205,7 @@ export interface Message {
 	attachments?: MessageAttachment[];
 	ttftMs?: number; // Time to first token in ms (for assistant messages)
 	tps?: number; // Tokens per second (for assistant messages)
+	tokens?: TokenData[]; // Token-level data for uncertainty visualization
 }
 
 export interface Conversation {
@@ -368,6 +383,21 @@ class AppStore {
 	private fetchInterval: ReturnType<typeof setInterval> | null = null;
 	private previewsInterval: ReturnType<typeof setInterval> | null = null;
 	private lastConversationPersistTs = 0;
+	private currentRequestController: AbortController | null = null;
+
+	/**
+	 * Abort any in-flight generation request
+	 */
+	abortCurrentRequest(): boolean {
+		if (this.currentRequestController) {
+			this.currentRequestController.abort();
+			this.currentRequestController = null;
+			this.isLoading = false;
+			this.currentResponse = "";
+			return true;
+		}
+		return false;
+	}
 
 	constructor() {
 		if (browser) {
@@ -1398,6 +1428,10 @@ class AppStore {
 			let firstTokenTime: number | null = null;
 			let tokenCount = 0;
 
+			// Create abort controller for this request
+			const controller = new AbortController();
+			this.currentRequestController = controller;
+
 			const response = await fetch("/v1/chat/completions", {
 				method: "POST",
 				headers: {
@@ -1408,7 +1442,10 @@ class AppStore {
 					messages: apiMessages,
 					temperature: 0.7,
 					stream: true,
+					logprobs: true,
+					top_logprobs: 5,
 				}),
+				signal: controller.signal,
 			});
 
 			if (!response.ok) {
@@ -1424,6 +1461,7 @@ class AppStore {
 			const decoder = new TextDecoder();
 			let fullContent = "";
 			let buffer = "";
+			const collectedTokens: TokenData[] = [];
 
 			while (true) {
 				const { done, value } = await reader.read();
@@ -1445,8 +1483,8 @@ class AppStore {
 
 						try {
 							const parsed = JSON.parse(data);
-							const tokenContent = parsed.choices?.[0]?.delta?.content;
-							if (tokenContent) {
+							const delta = parsed.choices?.[0]?.delta?.content;
+							if (delta) {
 								// Track first token for TTFT
 								if (firstTokenTime === null) {
 									firstTokenTime = performance.now();
@@ -1463,7 +1501,30 @@ class AppStore {
 									this.tps = (tokenCount / elapsed) * 1000;
 								}
 
-								fullContent += tokenContent;
+								// Extract logprobs for uncertainty visualization
+								const logprobsData = parsed.choices?.[0]?.logprobs;
+								if (logprobsData?.content?.[0]) {
+									const logprobItem = logprobsData.content[0];
+									const tokenData: TokenData = {
+										token: logprobItem.token || delta,
+										logprob: logprobItem.logprob ?? 0,
+										probability: Math.exp(logprobItem.logprob ?? 0),
+										topLogprobs: (logprobItem.top_logprobs || []).map(
+											(item: {
+												token: string;
+												logprob: number;
+												bytes?: number[];
+											}) => ({
+												token: item.token,
+												logprob: item.logprob,
+												bytes: item.bytes,
+											}),
+										),
+									};
+									collectedTokens.push(tokenData);
+								}
+
+								fullContent += delta;
 
 								// Strip thinking tags for display and extract thinking content
 								const { displayContent, thinkingContent } =
@@ -1477,6 +1538,7 @@ class AppStore {
 								if (idx !== -1) {
 									this.messages[idx].content = displayContent;
 									this.messages[idx].thinking = thinkingContent || undefined;
+									this.messages[idx].tokens = [...collectedTokens];
 								}
 								this.persistActiveConversation();
 							}
@@ -1524,9 +1586,16 @@ class AppStore {
 				if (this.tps !== null) {
 					this.messages[idx].tps = this.tps;
 				}
+				if (collectedTokens.length > 0) {
+					this.messages[idx].tokens = collectedTokens;
+				}
 			}
 			this.persistActiveConversation();
 		} catch (error) {
+			// Don't show error for aborted requests (user cancelled)
+			if (error instanceof Error && error.name === "AbortError") {
+				return;
+			}
 			console.error("Error sending message:", error);
 			// Update the assistant message with error
 			const idx = this.messages.findIndex((m) => m.id === assistantMessage.id);
@@ -1536,6 +1605,237 @@ class AppStore {
 			}
 			this.persistActiveConversation();
 		} finally {
+			// Clean up controller if this is still the active request
+			if (this.currentRequestController === controller) {
+				this.currentRequestController = null;
+			}
+			this.isLoading = false;
+			this.currentResponse = "";
+			this.updateActiveConversation();
+		}
+	}
+
+	/**
+	 * Regenerate from a specific token in an assistant message.
+	 * Keeps content up to and including the specified token, then continues generation.
+	 * If a generation is already in progress, it will be aborted first.
+	 */
+	async regenerateFromToken(
+		messageId: string,
+		tokenIndex: number,
+	): Promise<void> {
+		// Abort any in-flight request first
+		this.abortCurrentRequest();
+
+		const messageIdx = this.messages.findIndex((m) => m.id === messageId);
+		if (messageIdx === -1) return;
+
+		const message = this.messages[messageIdx];
+		if (message.role !== "assistant" || !message.tokens) return;
+
+		// Get tokens up to and including the specified index
+		const tokensToKeep = message.tokens.slice(0, tokenIndex + 1);
+		const prefixText = tokensToKeep.map((t) => t.token).join("");
+
+		// Remove all messages after this assistant message
+		this.messages = this.messages.slice(0, messageIdx + 1);
+
+		// Update the message to show the prefix
+		this.messages[messageIdx].content = prefixText;
+		this.messages[messageIdx].tokens = tokensToKeep;
+
+		// Set up for continuation
+		this.isLoading = true;
+		this.currentResponse = prefixText;
+		this.ttftMs = null;
+		this.tps = null;
+		this.totalTokens = tokensToKeep.length;
+
+		try {
+			// Build messages for API - include the partial assistant message
+			const systemPrompt = {
+				role: "system" as const,
+				content:
+					"You are a helpful AI assistant. Respond directly and concisely. Do not show your reasoning or thought process.",
+			};
+
+			// Get all messages up to and including the one we're regenerating from
+			const apiMessages = [
+				systemPrompt,
+				...this.messages.map((m) => {
+					let msgContent = m.content;
+					if (m.attachments) {
+						for (const attachment of m.attachments) {
+							if (attachment.type === "text" && attachment.content) {
+								msgContent += `\n\n[File: ${attachment.name}]\n\`\`\`\n${attachment.content}\n\`\`\``;
+							}
+						}
+					}
+					return { role: m.role, content: msgContent };
+				}),
+			];
+
+			// Determine model
+			let modelToUse = this.selectedChatModel;
+			if (!modelToUse) {
+				for (const [, instanceWrapper] of Object.entries(this.instances)) {
+					if (instanceWrapper && typeof instanceWrapper === "object") {
+						const keys = Object.keys(
+							instanceWrapper as Record<string, unknown>,
+						);
+						if (keys.length === 1) {
+							const instance = (instanceWrapper as Record<string, unknown>)[
+								keys[0]
+							] as { shardAssignments?: { modelId?: string } };
+							if (instance?.shardAssignments?.modelId) {
+								modelToUse = instance.shardAssignments.modelId;
+								break;
+							}
+						}
+					}
+				}
+			}
+
+			if (!modelToUse) {
+				throw new Error("No model available");
+			}
+
+			// Start timing
+			const requestStartTime = performance.now();
+			let firstTokenTime: number | null = null;
+			let tokenCount = tokensToKeep.length;
+
+			// Create abort controller
+			const controller = new AbortController();
+			this.currentRequestController = controller;
+
+			const response = await fetch("/v1/chat/completions", {
+				method: "POST",
+				headers: { "Content-Type": "application/json" },
+				body: JSON.stringify({
+					model: modelToUse,
+					messages: apiMessages,
+					stream: true,
+					logprobs: true,
+					top_logprobs: 5,
+					continue_from_prefix: true,
+				}),
+				signal: controller.signal,
+			});
+
+			if (!response.ok) {
+				const errorText = await response.text();
+				throw new Error(`API error: ${response.status} - ${errorText}`);
+			}
+
+			const reader = response.body?.getReader();
+			if (!reader) throw new Error("No response body");
+
+			const decoder = new TextDecoder();
+			let fullContent = prefixText;
+			let buffer = "";
+			const collectedTokens: TokenData[] = [...tokensToKeep];
+
+			while (true) {
+				const { done, value } = await reader.read();
+				if (done) break;
+
+				buffer += decoder.decode(value, { stream: true });
+				const lines = buffer.split("\n");
+				buffer = lines.pop() || "";
+
+				for (const line of lines) {
+					const trimmed = line.trim();
+					if (!trimmed || trimmed === "data: [DONE]") continue;
+
+					if (trimmed.startsWith("data: ")) {
+						try {
+							const json = JSON.parse(trimmed.slice(6));
+							const delta = json.choices?.[0]?.delta?.content;
+							if (delta) {
+								if (firstTokenTime === null) {
+									firstTokenTime = performance.now();
+									this.ttftMs = firstTokenTime - requestStartTime;
+								}
+
+								tokenCount += 1;
+								this.totalTokens = tokenCount;
+
+								if (
+									firstTokenTime !== null &&
+									tokenCount > tokensToKeep.length
+								) {
+									const elapsed = performance.now() - firstTokenTime;
+									this.tps =
+										((tokenCount - tokensToKeep.length) / elapsed) * 1000;
+								}
+
+								// Extract logprobs
+								const logprobsData = json.choices?.[0]?.logprobs;
+								if (logprobsData?.content?.[0]) {
+									const logprobItem = logprobsData.content[0];
+									collectedTokens.push({
+										token: logprobItem.token || delta,
+										logprob: logprobItem.logprob ?? 0,
+										probability: Math.exp(logprobItem.logprob ?? 0),
+										topLogprobs: (logprobItem.top_logprobs || []).map(
+											(item: {
+												token: string;
+												logprob: number;
+												bytes?: number[];
+											}) => ({
+												token: item.token,
+												logprob: item.logprob,
+												bytes: item.bytes,
+											}),
+										),
+									});
+								}
+
+								fullContent += delta;
+								const { displayContent, thinkingContent } =
+									this.stripThinkingTags(fullContent);
+								this.currentResponse = displayContent;
+
+								this.messages[messageIdx].content = displayContent;
+								this.messages[messageIdx].thinking =
+									thinkingContent || undefined;
+								this.messages[messageIdx].tokens = [...collectedTokens];
+								this.persistActiveConversation();
+							}
+						} catch {
+							// Skip malformed JSON
+						}
+					}
+				}
+			}
+
+			// Final update
+			const { displayContent, thinkingContent } =
+				this.stripThinkingTags(fullContent);
+			this.messages[messageIdx].content = displayContent;
+			this.messages[messageIdx].thinking = thinkingContent || undefined;
+			this.messages[messageIdx].tokens = collectedTokens;
+
+			if (this.ttftMs !== null) {
+				this.messages[messageIdx].ttftMs = this.ttftMs;
+			}
+			if (this.tps !== null) {
+				this.messages[messageIdx].tps = this.tps;
+			}
+			this.persistActiveConversation();
+		} catch (error) {
+			if (error instanceof Error && error.name === "AbortError") {
+				return;
+			}
+			console.error("Error regenerating from token:", error);
+			this.messages[messageIdx].content =
+				`${prefixText}\n\nError: ${error instanceof Error ? error.message : "Unknown error"}`;
+			this.persistActiveConversation();
+		} finally {
+			if (this.currentRequestController === controller) {
+				this.currentRequestController = null;
+			}
 			this.isLoading = false;
 			this.currentResponse = "";
 			this.updateActiveConversation();
@@ -1615,6 +1915,8 @@ export const editMessage = (messageId: string, newContent: string) =>
 export const editAndRegenerate = (messageId: string, newContent: string) =>
 	appStore.editAndRegenerate(messageId, newContent);
 export const regenerateLastResponse = () => appStore.regenerateLastResponse();
+export const regenerateFromToken = (messageId: string, tokenIndex: number) =>
+	appStore.regenerateFromToken(messageId, tokenIndex);
 
 // Conversation actions
 export const conversations = () => appStore.conversations;

src/exo/main.py

@@ -205,6 +205,14 @@ def main():
     logger.info("Starting EXO")
     logger.info(f"EXO_LIBP2P_NAMESPACE: {os.getenv('EXO_LIBP2P_NAMESPACE')}")
 
+    # Set FAST_SYNCH override env var for runner subprocesses
+    if args.fast_synch is True:
+        os.environ["EXO_FAST_SYNCH"] = "on"
+        logger.info("FAST_SYNCH forced ON")
+    elif args.fast_synch is False:
+        os.environ["EXO_FAST_SYNCH"] = "off"
+        logger.info("FAST_SYNCH forced OFF")
+
     node = anyio.run(Node.create, args)
     anyio.run(node.run)
     logger.info("EXO Shutdown complete")
@@ -218,6 +226,7 @@ class Args(CamelCaseModel):
     api_port: PositiveInt = 52415
     tb_only: bool = False
     no_worker: bool = False
+    fast_synch: bool | None = None  # None = auto, True = force on, False = force off
 
     @classmethod
     def parse(cls) -> Self:
@@ -259,6 +268,20 @@ class Args(CamelCaseModel):
             "--no-worker",
             action="store_true",
         )
+        fast_synch_group = parser.add_mutually_exclusive_group()
+        fast_synch_group.add_argument(
+            "--fast-synch",
+            action="store_true",
+            dest="fast_synch",
+            default=None,
+            help="Force MLX FAST_SYNCH on (for JACCL backend)",
+        )
+        fast_synch_group.add_argument(
+            "--no-fast-synch",
+            action="store_false",
+            dest="fast_synch",
+            help="Force MLX FAST_SYNCH off",
+        )
 
         args = parser.parse_args()
         return cls(**vars(args))  # pyright: ignore[reportAny] - We are intentionally validating here, we can't do it statically

src/exo/master/adapters/__init__.py

@@ -0,0 +1 @@
+"""API adapters for different API formats (Claude, OpenAI Responses, etc.)."""

src/exo/master/adapters/chat_completions.py

@@ -0,0 +1,175 @@
+"""OpenAI Chat Completions API adapter for converting requests/responses."""
+
+import time
+from collections.abc import AsyncGenerator
+
+from exo.shared.types.api import (
+    ChatCompletionChoice,
+    ChatCompletionMessage,
+    ChatCompletionMessageText,
+    ChatCompletionResponse,
+    ChatCompletionTaskParams,
+    ErrorInfo,
+    ErrorResponse,
+    FinishReason,
+    Logprobs,
+    LogprobsContentItem,
+    StreamingChoiceResponse,
+)
+from exo.shared.types.chunks import TokenChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.openai_responses import ResponseInputMessage, ResponsesRequest
+
+
+def chat_request_to_internal(request: ChatCompletionTaskParams) -> ResponsesRequest:
+    """Convert Chat Completions API request to ResponsesRequest (canonical internal format).
+
+    Extracts system message as instructions, converts messages to input.
+    """
+    instructions: str | None = None
+    input_messages: list[ResponseInputMessage] = []
+
+    for msg in request.messages:
+        # Normalize content to string
+        content: str
+        if msg.content is None:
+            content = ""
+        elif isinstance(msg.content, str):
+            content = msg.content
+        elif isinstance(msg.content, ChatCompletionMessageText):
+            content = msg.content.text
+        else:
+            # List of ChatCompletionMessageText
+            content = "\n".join(item.text for item in msg.content)
+
+        # Extract system message as instructions
+        if msg.role == "system":
+            if instructions is None:
+                instructions = content
+            else:
+                # Append additional system messages
+                instructions = f"{instructions}\n{content}"
+        else:
+            # Convert to ResponseInputMessage (only user, assistant, developer roles)
+            if msg.role in ("user", "assistant", "developer"):
+                input_messages.append(
+                    ResponseInputMessage(role=msg.role, content=content)
+                )
+
+    return ResponsesRequest(
+        model=request.model,
+        input=input_messages if input_messages else "",
+        instructions=instructions,
+        max_output_tokens=request.max_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        stop=request.stop,
+        seed=request.seed,
+        stream=request.stream,
+        tools=request.tools,
+        continue_from_prefix=request.continue_from_prefix,
+    )
+
+
+def chunk_to_response(
+    chunk: TokenChunk, command_id: CommandId
+) -> ChatCompletionResponse:
+    """Convert a TokenChunk to a streaming ChatCompletionResponse."""
+    # Build logprobs if available
+    logprobs: Logprobs | None = None
+    if chunk.logprob is not None:
+        logprobs = Logprobs(
+            content=[
+                LogprobsContentItem(
+                    token=chunk.text,
+                    logprob=chunk.logprob,
+                    top_logprobs=chunk.top_logprobs or [],
+                )
+            ]
+        )
+
+    return ChatCompletionResponse(
+        id=command_id,
+        created=int(time.time()),
+        model=chunk.model,
+        choices=[
+            StreamingChoiceResponse(
+                index=0,
+                delta=ChatCompletionMessage(role="assistant", content=chunk.text),
+                logprobs=logprobs,
+                finish_reason=chunk.finish_reason,
+            )
+        ],
+    )
+
+
+async def generate_chat_stream(
+    command_id: CommandId,
+    chunk_stream: AsyncGenerator[TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate Chat Completions API streaming events from TokenChunks."""
+    async for chunk in chunk_stream:
+        if chunk.finish_reason == "error":
+            error_response = ErrorResponse(
+                error=ErrorInfo(
+                    message=chunk.error_message or "Internal server error",
+                    type="InternalServerError",
+                    code=500,
+                )
+            )
+            yield f"data: {error_response.model_dump_json()}\n\n"
+            yield "data: [DONE]\n\n"
+            return
+
+        chunk_response = chunk_to_response(chunk, command_id)
+        yield f"data: {chunk_response.model_dump_json()}\n\n"
+
+        if chunk.finish_reason is not None:
+            yield "data: [DONE]\n\n"
+
+
+async def collect_chat_response(
+    command_id: CommandId,
+    chunk_stream: AsyncGenerator[TokenChunk, None],
+) -> ChatCompletionResponse:
+    """Collect all token chunks and return a single ChatCompletionResponse."""
+    text_parts: list[str] = []
+    model: str | None = None
+    finish_reason: FinishReason | None = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if chunk.finish_reason == "error":
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        if model is None:
+            model = chunk.model
+
+        text_parts.append(chunk.text)
+
+        if chunk.finish_reason is not None:
+            finish_reason = chunk.finish_reason
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    combined_text = "".join(text_parts)
+    assert model is not None
+
+    return ChatCompletionResponse(
+        id=command_id,
+        created=int(time.time()),
+        model=model,
+        choices=[
+            ChatCompletionChoice(
+                index=0,
+                message=ChatCompletionMessage(
+                    role="assistant",
+                    content=combined_text,
+                ),
+                finish_reason=finish_reason,
+            )
+        ],
+    )

src/exo/master/adapters/claude.py

@@ -0,0 +1,190 @@
+"""Claude Messages API adapter for converting requests/responses."""
+
+from collections.abc import AsyncGenerator
+
+from exo.shared.types.api import FinishReason
+from exo.shared.types.chunks import TokenChunk
+from exo.shared.types.claude_api import (
+    ClaudeContentBlockDeltaEvent,
+    ClaudeContentBlockStartEvent,
+    ClaudeContentBlockStopEvent,
+    ClaudeMessageDelta,
+    ClaudeMessageDeltaEvent,
+    ClaudeMessageDeltaUsage,
+    ClaudeMessagesRequest,
+    ClaudeMessagesResponse,
+    ClaudeMessageStart,
+    ClaudeMessageStartEvent,
+    ClaudeMessageStopEvent,
+    ClaudeStopReason,
+    ClaudeTextBlock,
+    ClaudeTextDelta,
+    ClaudeUsage,
+)
+from exo.shared.types.common import CommandId
+from exo.shared.types.openai_responses import ResponseInputMessage, ResponsesRequest
+
+
+def finish_reason_to_claude_stop_reason(
+    finish_reason: FinishReason | None,
+) -> ClaudeStopReason | None:
+    """Map OpenAI finish_reason to Claude stop_reason."""
+    if finish_reason is None:
+        return None
+    mapping: dict[FinishReason, ClaudeStopReason] = {
+        "stop": "end_turn",
+        "length": "max_tokens",
+        "tool_calls": "tool_use",
+        "content_filter": "end_turn",
+        "function_call": "tool_use",
+    }
+    return mapping.get(finish_reason, "end_turn")
+
+
+def claude_request_to_internal(request: ClaudeMessagesRequest) -> ResponsesRequest:
+    """Convert Claude Messages API request to ResponsesRequest (canonical internal format).
+
+    Converts Claude's system parameter to instructions,
+    and messages to input.
+    """
+    # Handle system message
+    instructions: str | None = None
+    if request.system:
+        if isinstance(request.system, str):
+            instructions = request.system
+        else:
+            # List of text blocks
+            instructions = "".join(block.text for block in request.system)
+
+    # Convert messages to input
+    input_messages: list[ResponseInputMessage] = []
+    for msg in request.messages:
+        content: str
+        if isinstance(msg.content, str):
+            content = msg.content
+        else:
+            # Concatenate text blocks (images not supported for MVP)
+            text_parts: list[str] = []
+            for block in msg.content:
+                if isinstance(block, ClaudeTextBlock):
+                    text_parts.append(block.text)
+            content = "".join(text_parts)
+
+        # Claude uses "user" and "assistant" roles
+        input_messages.append(ResponseInputMessage(role=msg.role, content=content))
+
+    return ResponsesRequest(
+        model=request.model,
+        input=input_messages if input_messages else "",
+        instructions=instructions,
+        max_output_tokens=request.max_tokens,
+        temperature=request.temperature,
+        top_p=request.top_p,
+        top_k=request.top_k,
+        stop=request.stop_sequences,
+        stream=request.stream,
+    )
+
+
+async def collect_claude_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[TokenChunk, None],
+) -> ClaudeMessagesResponse:
+    """Collect all token chunks and return a single ClaudeMessagesResponse."""
+    text_parts: list[str] = []
+    stop_reason: ClaudeStopReason | None = None
+    last_stats = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if chunk.finish_reason == "error":
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        text_parts.append(chunk.text)
+        last_stats = chunk.stats or last_stats
+
+        if chunk.finish_reason is not None:
+            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    combined_text = "".join(text_parts)
+
+    # Use actual usage data from stats if available
+    input_tokens = last_stats.prompt_tokens if last_stats else 0
+    output_tokens = last_stats.generation_tokens if last_stats else 0
+
+    return ClaudeMessagesResponse(
+        id=f"msg_{command_id}",
+        model=model,
+        content=[ClaudeTextBlock(text=combined_text)],
+        stop_reason=stop_reason,
+        usage=ClaudeUsage(
+            input_tokens=input_tokens,
+            output_tokens=output_tokens,
+        ),
+    )
+
+
+async def generate_claude_stream(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate Claude Messages API streaming events from TokenChunks."""
+    # Initial message_start event
+    initial_message = ClaudeMessageStart(
+        id=f"msg_{command_id}",
+        model=model,
+        content=[],
+        stop_reason=None,
+        usage=ClaudeUsage(input_tokens=0, output_tokens=0),
+    )
+    start_event = ClaudeMessageStartEvent(message=initial_message)
+    yield f"event: message_start\ndata: {start_event.model_dump_json()}\n\n"
+
+    # content_block_start
+    block_start = ClaudeContentBlockStartEvent(
+        index=0, content_block=ClaudeTextBlock(text="")
+    )
+    yield f"event: content_block_start\ndata: {block_start.model_dump_json()}\n\n"
+
+    output_tokens = 0
+    stop_reason: ClaudeStopReason | None = None
+    last_stats = None
+
+    async for chunk in chunk_stream:
+        output_tokens += 1  # Count each chunk as one token
+        last_stats = chunk.stats or last_stats
+
+        # content_block_delta
+        delta_event = ClaudeContentBlockDeltaEvent(
+            index=0,
+            delta=ClaudeTextDelta(text=chunk.text),
+        )
+        yield f"event: content_block_delta\ndata: {delta_event.model_dump_json()}\n\n"
+
+        if chunk.finish_reason is not None:
+            stop_reason = finish_reason_to_claude_stop_reason(chunk.finish_reason)
+
+    # Use actual token count from stats if available
+    if last_stats is not None:
+        output_tokens = last_stats.generation_tokens
+
+    # content_block_stop
+    block_stop = ClaudeContentBlockStopEvent(index=0)
+    yield f"event: content_block_stop\ndata: {block_stop.model_dump_json()}\n\n"
+
+    # message_delta
+    message_delta = ClaudeMessageDeltaEvent(
+        delta=ClaudeMessageDelta(stop_reason=stop_reason),
+        usage=ClaudeMessageDeltaUsage(output_tokens=output_tokens),
+    )
+    yield f"event: message_delta\ndata: {message_delta.model_dump_json()}\n\n"
+
+    # message_stop
+    message_stop = ClaudeMessageStopEvent()
+    yield f"event: message_stop\ndata: {message_stop.model_dump_json()}\n\n"

src/exo/master/adapters/responses.py

@@ -0,0 +1,173 @@
+"""OpenAI Responses API adapter for converting requests/responses.
+
+ResponsesRequest is the canonical internal format. Responses API is the most featureful,
+making it the natural choice for the internal format. All other API formats (Chat
+Completions, Claude) are converted TO ResponsesRequest.
+"""
+
+from collections.abc import AsyncGenerator
+
+from exo.shared.types.chunks import TokenChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.openai_responses import (
+    ResponseCompletedEvent,
+    ResponseContentPartAddedEvent,
+    ResponseContentPartDoneEvent,
+    ResponseCreatedEvent,
+    ResponseInProgressEvent,
+    ResponseMessageItem,
+    ResponseOutputItemAddedEvent,
+    ResponseOutputItemDoneEvent,
+    ResponseOutputText,
+    ResponsesResponse,
+    ResponseTextDeltaEvent,
+    ResponseTextDoneEvent,
+    ResponseUsage,
+)
+
+
+async def collect_responses_response(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[TokenChunk, None],
+) -> ResponsesResponse:
+    """Collect all token chunks and return a single ResponsesResponse."""
+    response_id = f"resp_{command_id}"
+    item_id = f"item_{command_id}"
+    accumulated_text = ""
+    last_stats = None
+    error_message: str | None = None
+
+    async for chunk in chunk_stream:
+        if chunk.finish_reason == "error":
+            error_message = chunk.error_message or "Internal server error"
+            break
+
+        accumulated_text += chunk.text
+        last_stats = chunk.stats or last_stats
+
+    if error_message is not None:
+        raise ValueError(error_message)
+
+    # Create usage from stats if available
+    usage = None
+    if last_stats is not None:
+        usage = ResponseUsage(
+            input_tokens=last_stats.prompt_tokens,
+            output_tokens=last_stats.generation_tokens,
+            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+        )
+
+    output_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text=accumulated_text)],
+        status="completed",
+    )
+
+    return ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="completed",
+        output=[output_item],
+        output_text=accumulated_text,
+        usage=usage,
+    )
+
+
+async def generate_responses_stream(
+    command_id: CommandId,
+    model: str,
+    chunk_stream: AsyncGenerator[TokenChunk, None],
+) -> AsyncGenerator[str, None]:
+    """Generate OpenAI Responses API streaming events from TokenChunks."""
+    response_id = f"resp_{command_id}"
+    item_id = f"item_{command_id}"
+
+    # response.created
+    initial_response = ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="in_progress",
+        output=[],
+        output_text="",
+    )
+    created_event = ResponseCreatedEvent(response=initial_response)
+    yield f"event: response.created\ndata: {created_event.model_dump_json()}\n\n"
+
+    # response.in_progress
+    in_progress_event = ResponseInProgressEvent(response=initial_response)
+    yield f"event: response.in_progress\ndata: {in_progress_event.model_dump_json()}\n\n"
+
+    # response.output_item.added
+    initial_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text="")],
+        status="in_progress",
+    )
+    item_added = ResponseOutputItemAddedEvent(output_index=0, item=initial_item)
+    yield f"event: response.output_item.added\ndata: {item_added.model_dump_json()}\n\n"
+
+    # response.content_part.added
+    initial_part = ResponseOutputText(text="")
+    part_added = ResponseContentPartAddedEvent(
+        output_index=0, content_index=0, part=initial_part
+    )
+    yield f"event: response.content_part.added\ndata: {part_added.model_dump_json()}\n\n"
+
+    accumulated_text = ""
+    last_stats = None
+
+    async for chunk in chunk_stream:
+        accumulated_text += chunk.text
+        last_stats = chunk.stats or last_stats
+
+        # response.output_text.delta
+        delta_event = ResponseTextDeltaEvent(
+            output_index=0,
+            content_index=0,
+            delta=chunk.text,
+        )
+        yield f"event: response.output_text.delta\ndata: {delta_event.model_dump_json()}\n\n"
+
+    # response.output_text.done
+    text_done = ResponseTextDoneEvent(
+        output_index=0, content_index=0, text=accumulated_text
+    )
+    yield f"event: response.output_text.done\ndata: {text_done.model_dump_json()}\n\n"
+
+    # response.content_part.done
+    final_part = ResponseOutputText(text=accumulated_text)
+    part_done = ResponseContentPartDoneEvent(
+        output_index=0, content_index=0, part=final_part
+    )
+    yield f"event: response.content_part.done\ndata: {part_done.model_dump_json()}\n\n"
+
+    # response.output_item.done
+    final_item = ResponseMessageItem(
+        id=item_id,
+        content=[ResponseOutputText(text=accumulated_text)],
+        status="completed",
+    )
+    item_done = ResponseOutputItemDoneEvent(output_index=0, item=final_item)
+    yield f"event: response.output_item.done\ndata: {item_done.model_dump_json()}\n\n"
+
+    # Create usage from stats if available
+    usage = None
+    if last_stats is not None:
+        usage = ResponseUsage(
+            input_tokens=last_stats.prompt_tokens,
+            output_tokens=last_stats.generation_tokens,
+            total_tokens=last_stats.prompt_tokens + last_stats.generation_tokens,
+        )
+
+    # response.completed
+    final_response = ResponsesResponse(
+        id=response_id,
+        model=model,
+        status="completed",
+        output=[final_item],
+        output_text=accumulated_text,
+        usage=usage,
+    )
+    completed_event = ResponseCompletedEvent(response=final_response)
+    yield f"event: response.completed\ndata: {completed_event.model_dump_json()}\n\n"

src/exo/master/api.py

@@ -1,19 +1,33 @@
-import time
 from collections.abc import AsyncGenerator
+from http import HTTPStatus
 from typing import cast
 
 import anyio
-from anyio import create_task_group
+from anyio import BrokenResourceError, create_task_group
 from anyio.abc import TaskGroup
-from fastapi import FastAPI, HTTPException
+from fastapi import FastAPI, HTTPException, Request
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import StreamingResponse
+from fastapi.responses import JSONResponse, StreamingResponse
 from fastapi.staticfiles import StaticFiles
 from hypercorn.asyncio import serve  # pyright: ignore[reportUnknownVariableType]
 from hypercorn.config import Config
 from hypercorn.typing import ASGIFramework
 from loguru import logger
 
+from exo.master.adapters.chat_completions import (
+    chat_request_to_internal,
+    collect_chat_response,
+    generate_chat_stream,
+)
+from exo.master.adapters.claude import (
+    claude_request_to_internal,
+    collect_claude_response,
+    generate_claude_stream,
+)
+from exo.master.adapters.responses import (
+    collect_responses_response,
+    generate_responses_stream,
+)
 from exo.master.placement import place_instance as get_instance_placements
 from exo.shared.apply import apply
 from exo.shared.election import ElectionMessage
@@ -26,19 +40,24 @@ from exo.shared.types.api import (
     ChatCompletionChoice,
     ChatCompletionMessage,
     ChatCompletionResponse,
+    ChatCompletionTaskParams,
     CreateInstanceParams,
     CreateInstanceResponse,
     DeleteInstanceResponse,
-    FinishReason,
+    ErrorInfo,
+    ErrorResponse,
     GenerationStats,
     ModelList,
     ModelListModel,
     PlaceInstanceParams,
     PlacementPreview,
     PlacementPreviewResponse,
-    StreamingChoiceResponse,
 )
 from exo.shared.types.chunks import TokenChunk
+from exo.shared.types.claude_api import (
+    ClaudeMessagesRequest,
+    ClaudeMessagesResponse,
+)
 from exo.shared.types.commands import (
     ChatCompletion,
     Command,
@@ -49,11 +68,19 @@ from exo.shared.types.commands import (
     TaskFinished,
 )
 from exo.shared.types.common import CommandId, NodeId, SessionId
-from exo.shared.types.events import ChunkGenerated, Event, ForwarderEvent, IndexedEvent
+from exo.shared.types.events import (
+    ChunkGenerated,
+    Event,
+    ForwarderEvent,
+    IndexedEvent,
+)
 from exo.shared.types.memory import Memory
 from exo.shared.types.models import ModelId, ModelMetadata
+from exo.shared.types.openai_responses import (
+    ResponsesRequest,
+    ResponsesResponse,
+)
 from exo.shared.types.state import State
-from exo.shared.types.tasks import ChatCompletionTaskParams
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 from exo.utils.banner import print_startup_banner
@@ -62,23 +89,6 @@ from exo.utils.dashboard_path import find_dashboard
 from exo.utils.event_buffer import OrderedBuffer
 
 
-def chunk_to_response(
-    chunk: TokenChunk, command_id: CommandId
-) -> ChatCompletionResponse:
-    return ChatCompletionResponse(
-        id=command_id,
-        created=int(time.time()),
-        model=chunk.model,
-        choices=[
-            StreamingChoiceResponse(
-                index=0,
-                delta=ChatCompletionMessage(role="assistant", content=chunk.text),
-                finish_reason=chunk.finish_reason,
-            )
-        ],
-    )
-
-
 async def resolve_model_meta(model_id: str) -> ModelMetadata:
     if model_id in MODEL_CARDS:
         model_card = MODEL_CARDS[model_id]
@@ -115,6 +125,7 @@ class API:
         self.paused_ev: anyio.Event = anyio.Event()
 
         self.app = FastAPI()
+        self._setup_exception_handlers()
         self._setup_cors()
         self._setup_routes()
 
@@ -145,6 +156,20 @@ class API:
         self.paused_ev.set()
         self.paused_ev = anyio.Event()
 
+    def _setup_exception_handlers(self) -> None:
+        @self.app.exception_handler(HTTPException)
+        async def http_exception_handler(  # pyright: ignore[reportUnusedFunction]
+            _: Request, exc: HTTPException
+        ) -> JSONResponse:
+            err = ErrorResponse(
+                error=ErrorInfo(
+                    message=exc.detail,
+                    type=HTTPStatus(exc.status_code).phrase,
+                    code=exc.status_code,
+                )
+            )
+            return JSONResponse(err.model_dump(), status_code=exc.status_code)
+
     def _setup_cors(self) -> None:
         self.app.add_middleware(
             CORSMiddleware,
@@ -168,6 +193,8 @@ class API:
             self.chat_completions
         )
         self.app.post("/bench/chat/completions")(self.bench_chat_completions)
+        self.app.post("/v1/messages", response_model=None)(self.claude_messages)
+        self.app.post("/v1/responses", response_model=None)(self.openai_responses)
         self.app.get("/state")(lambda: self.state)
         self.app.get("/events")(lambda: self._event_log)
 
@@ -373,11 +400,13 @@ class API:
             instance_id=instance_id,
         )
 
-    async def _chat_chunk_stream(
+    async def _token_chunk_stream(
         self, command_id: CommandId
     ) -> AsyncGenerator[TokenChunk, None]:
-        """Yield `TokenChunk`s for a given command until completion."""
+        """Yield `TokenChunk`s for a given command until completion.
 
+        This is the internal low-level stream used by all API adapters.
+        """
         try:
             self._chat_completion_queues[command_id], recv = channel[TokenChunk]()
 
@@ -400,69 +429,26 @@ class API:
             await self._send(command)
             del self._chat_completion_queues[command_id]
 
-    async def _generate_chat_stream(
-        self, command_id: CommandId
-    ) -> AsyncGenerator[str, None]:
-        """Generate chat completion stream as JSON strings."""
-
-        async for chunk in self._chat_chunk_stream(command_id):
-            chunk_response: ChatCompletionResponse = chunk_to_response(
-                chunk, command_id
-            )
-            logger.debug(f"chunk_response: {chunk_response}")
-
-            yield f"data: {chunk_response.model_dump_json()}\n\n"
-
-            if chunk.finish_reason is not None:
-                yield "data: [DONE]\n\n"
-
-    async def _collect_chat_completion(
-        self, command_id: CommandId
-    ) -> ChatCompletionResponse:
-        """Collect all token chunks for a chat completion and return a single response."""
-
-        text_parts: list[str] = []
-        model: str | None = None
-        finish_reason: FinishReason | None = None
-
-        async for chunk in self._chat_chunk_stream(command_id):
-            if model is None:
-                model = chunk.model
-
-            text_parts.append(chunk.text)
-
-            if chunk.finish_reason is not None:
-                finish_reason = chunk.finish_reason
-
-        combined_text = "".join(text_parts)
-        assert model is not None
-
-        return ChatCompletionResponse(
-            id=command_id,
-            created=int(time.time()),
-            model=model,
-            choices=[
-                ChatCompletionChoice(
-                    index=0,
-                    message=ChatCompletionMessage(
-                        role="assistant",
-                        content=combined_text,
-                    ),
-                    finish_reason=finish_reason,
-                )
-            ],
-        )
-
     async def _collect_chat_completion_with_stats(
         self, command_id: CommandId
     ) -> BenchChatCompletionResponse:
+        import time
+
+        from exo.shared.types.api import FinishReason
+
         text_parts: list[str] = []
         model: str | None = None
         finish_reason: FinishReason | None = None
 
         stats: GenerationStats | None = None
 
-        async for chunk in self._chat_chunk_stream(command_id):
+        async for chunk in self._token_chunk_stream(command_id):
+            if chunk.finish_reason == "error":
+                raise HTTPException(
+                    status_code=500,
+                    detail=chunk.error_message or "Internal server error",
+                )
+
             if model is None:
                 model = chunk.model
 
@@ -500,54 +486,147 @@ class API:
     async def chat_completions(
         self, payload: ChatCompletionTaskParams
     ) -> ChatCompletionResponse | StreamingResponse:
-        """Handle chat completions, supporting both streaming and non-streaming responses."""
-        model_meta = await resolve_model_meta(payload.model)
-        payload.model = model_meta.model_id
+        """OpenAI Chat Completions API - adapter."""
+        internal_params = chat_request_to_internal(payload)
+        model_meta = await resolve_model_meta(internal_params.model)
+        internal_params.model = model_meta.model_id
 
         if not any(
-            instance.shard_assignments.model_id == payload.model
+            instance.shard_assignments.model_id == internal_params.model
             for instance in self.state.instances.values()
         ):
-            await self._trigger_notify_user_to_download_model(payload.model)
+            await self._trigger_notify_user_to_download_model(internal_params.model)
             raise HTTPException(
-                status_code=404, detail=f"No instance found for model {payload.model}"
+                status_code=404,
+                detail=f"No instance found for model {internal_params.model}",
             )
 
-        command = ChatCompletion(
-            request_params=payload,
-        )
+        command = ChatCompletion(request_params=internal_params)
         await self._send(command)
+
         if payload.stream:
             return StreamingResponse(
-                self._generate_chat_stream(command.command_id),
+                generate_chat_stream(
+                    command.command_id,
+                    self._token_chunk_stream(command.command_id),
+                ),
                 media_type="text/event-stream",
             )
 
-        return await self._collect_chat_completion(command.command_id)
+        try:
+            return await collect_chat_response(
+                command.command_id,
+                self._token_chunk_stream(command.command_id),
+            )
+        except ValueError as e:
+            raise HTTPException(status_code=500, detail=str(e)) from e
 
     async def bench_chat_completions(
         self, payload: BenchChatCompletionTaskParams
     ) -> BenchChatCompletionResponse:
-        model_meta = await resolve_model_meta(payload.model)
-        payload.model = model_meta.model_id
+        # Convert to internal format (BenchChatCompletionTaskParams extends ChatCompletionTaskParams)
+        internal_params = chat_request_to_internal(payload)
+        model_meta = await resolve_model_meta(internal_params.model)
+        internal_params.model = model_meta.model_id
 
         if not any(
-            instance.shard_assignments.model_id == payload.model
+            instance.shard_assignments.model_id == internal_params.model
             for instance in self.state.instances.values()
         ):
-            await self._trigger_notify_user_to_download_model(payload.model)
+            await self._trigger_notify_user_to_download_model(internal_params.model)
             raise HTTPException(
-                status_code=404, detail=f"No instance found for model {payload.model}"
+                status_code=404,
+                detail=f"No instance found for model {internal_params.model}",
             )
 
-        payload.stream = False
+        internal_params.stream = False
 
-        command = ChatCompletion(request_params=payload)
+        command = ChatCompletion(request_params=internal_params)
         await self._send(command)
 
         response = await self._collect_chat_completion_with_stats(command.command_id)
         return response
 
+    async def claude_messages(
+        self, payload: ClaudeMessagesRequest
+    ) -> ClaudeMessagesResponse | StreamingResponse:
+        """Claude Messages API - adapter."""
+        internal_params = claude_request_to_internal(payload)
+        model_meta = await resolve_model_meta(internal_params.model)
+        internal_params.model = model_meta.model_id
+
+        if not any(
+            instance.shard_assignments.model_id == internal_params.model
+            for instance in self.state.instances.values()
+        ):
+            await self._trigger_notify_user_to_download_model(internal_params.model)
+            raise HTTPException(
+                status_code=404,
+                detail=f"No instance found for model {internal_params.model}",
+            )
+
+        command = ChatCompletion(request_params=internal_params)
+        await self._send(command)
+
+        if payload.stream:
+            return StreamingResponse(
+                generate_claude_stream(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="text/event-stream",
+            )
+
+        try:
+            return await collect_claude_response(
+                command.command_id,
+                payload.model,
+                self._token_chunk_stream(command.command_id),
+            )
+        except ValueError as e:
+            raise HTTPException(status_code=500, detail=str(e)) from e
+
+    async def openai_responses(
+        self, payload: ResponsesRequest
+    ) -> ResponsesResponse | StreamingResponse:
+        """OpenAI Responses API - native format (no conversion needed)."""
+        model_meta = await resolve_model_meta(payload.model)
+        # Update model to resolved model_id
+        request_params = payload.model_copy(update={"model": model_meta.model_id})
+
+        if not any(
+            instance.shard_assignments.model_id == request_params.model
+            for instance in self.state.instances.values()
+        ):
+            await self._trigger_notify_user_to_download_model(request_params.model)
+            raise HTTPException(
+                status_code=404,
+                detail=f"No instance found for model {request_params.model}",
+            )
+
+        command = ChatCompletion(request_params=request_params)
+        await self._send(command)
+
+        if payload.stream:
+            return StreamingResponse(
+                generate_responses_stream(
+                    command.command_id,
+                    payload.model,
+                    self._token_chunk_stream(command.command_id),
+                ),
+                media_type="text/event-stream",
+            )
+
+        try:
+            return await collect_responses_response(
+                command.command_id,
+                payload.model,
+                self._token_chunk_stream(command.command_id),
+            )
+        except ValueError as e:
+            raise HTTPException(status_code=500, detail=str(e)) from e
+
     def _calculate_total_available_memory(self) -> Memory:
         """Calculate total available memory across all nodes in bytes."""
         total_available = Memory()
@@ -607,14 +686,14 @@ class API:
                 for idx, event in self.event_buffer.drain_indexed():
                     self._event_log.append(event)
                     self.state = apply(self.state, IndexedEvent(event=event, idx=idx))
-                    if (
-                        isinstance(event, ChunkGenerated)
-                        and event.command_id in self._chat_completion_queues
-                    ):
+                    if isinstance(event, ChunkGenerated):
                         assert isinstance(event.chunk, TokenChunk)
-                        await self._chat_completion_queues[event.command_id].send(
-                            event.chunk
-                        )
+                        queue = self._chat_completion_queues.get(event.command_id)
+                        if queue is not None:
+                            try:
+                                await queue.send(event.chunk)
+                            except BrokenResourceError:
+                                self._chat_completion_queues.pop(event.command_id, None)
 
     async def _pause_on_new_election(self):
         with self.election_receiver as ems:

src/exo/master/tests/test_api_error_handling.py

@@ -0,0 +1,107 @@
+# pyright: reportUnusedFunction=false, reportAny=false
+from typing import Any, get_args
+
+from fastapi import FastAPI, HTTPException
+from fastapi.testclient import TestClient
+
+from exo.shared.types.api import ErrorInfo, ErrorResponse, FinishReason
+from exo.shared.types.chunks import TokenChunk
+from exo.worker.tests.constants import MODEL_A_ID
+
+
+def test_http_exception_handler_formats_openai_style() -> None:
+    """Test that HTTPException is converted to OpenAI-style error format."""
+    from exo.master.api import API
+
+    app = FastAPI()
+
+    # Setup exception handler
+    api = object.__new__(API)
+    api.app = app
+    api._setup_exception_handlers()  # pyright: ignore[reportPrivateUsage]
+
+    # Add test routes that raise HTTPException
+    @app.get("/test-error")
+    async def _test_error() -> None:
+        raise HTTPException(status_code=500, detail="Test error message")
+
+    @app.get("/test-not-found")
+    async def _test_not_found() -> None:
+        raise HTTPException(status_code=404, detail="Resource not found")
+
+    client = TestClient(app)
+
+    # Test 500 error
+    response = client.get("/test-error")
+    assert response.status_code == 500
+    data: dict[str, Any] = response.json()
+    assert "error" in data
+    assert data["error"]["message"] == "Test error message"
+    assert data["error"]["type"] == "Internal Server Error"
+    assert data["error"]["code"] == 500
+
+    # Test 404 error
+    response = client.get("/test-not-found")
+    assert response.status_code == 404
+    data = response.json()
+    assert "error" in data
+    assert data["error"]["message"] == "Resource not found"
+    assert data["error"]["type"] == "Not Found"
+    assert data["error"]["code"] == 404
+
+
+def test_finish_reason_includes_error() -> None:
+    valid_reasons = get_args(FinishReason)
+    assert "error" in valid_reasons
+
+
+def test_token_chunk_with_error_fields() -> None:
+    chunk = TokenChunk(
+        idx=0,
+        model=MODEL_A_ID,
+        text="",
+        token_id=0,
+        finish_reason="error",
+        error_message="Something went wrong",
+    )
+
+    assert chunk.finish_reason == "error"
+    assert chunk.error_message == "Something went wrong"
+
+
+def test_token_chunk_without_error() -> None:
+    chunk = TokenChunk(
+        idx=1,
+        model=MODEL_A_ID,
+        text="Hello",
+        token_id=42,
+        finish_reason=None,
+    )
+
+    assert chunk.finish_reason is None
+    assert chunk.error_message is None
+
+
+def test_error_response_construction() -> None:
+    error_response = ErrorResponse(
+        error=ErrorInfo(
+            message="Generation failed",
+            type="InternalServerError",
+            code=500,
+        )
+    )
+
+    assert error_response.error.message == "Generation failed"
+    assert error_response.error.code == 500
+
+
+def test_normal_finish_reasons_still_work() -> None:
+    for reason in ["stop", "length", "tool_calls", "content_filter", "function_call"]:
+        chunk = TokenChunk(
+            idx=0,
+            model=MODEL_A_ID,
+            text="done",
+            token_id=100,
+            finish_reason=reason,  # type: ignore[arg-type]
+        )
+        assert chunk.finish_reason == reason

src/exo/master/tests/test_claude_api.py

@@ -0,0 +1,283 @@
+"""Tests for Claude Messages API conversion functions and types."""
+
+import json
+from typing import Any, cast
+
+import pydantic
+import pytest
+
+from exo.master.adapters.claude import (
+    claude_request_to_internal,
+    finish_reason_to_claude_stop_reason,
+)
+from exo.shared.types.claude_api import (
+    ClaudeContentBlockDeltaEvent,
+    ClaudeContentBlockStartEvent,
+    ClaudeContentBlockStopEvent,
+    ClaudeMessage,
+    ClaudeMessageDelta,
+    ClaudeMessageDeltaEvent,
+    ClaudeMessageDeltaUsage,
+    ClaudeMessagesRequest,
+    ClaudeMessageStart,
+    ClaudeMessageStartEvent,
+    ClaudeMessageStopEvent,
+    ClaudeTextBlock,
+    ClaudeTextDelta,
+    ClaudeUsage,
+)
+
+
+class TestFinishReasonToClaudeStopReason:
+    """Tests for finish_reason to Claude stop_reason mapping."""
+
+    def test_stop_maps_to_end_turn(self):
+        assert finish_reason_to_claude_stop_reason("stop") == "end_turn"
+
+    def test_length_maps_to_max_tokens(self):
+        assert finish_reason_to_claude_stop_reason("length") == "max_tokens"
+
+    def test_tool_calls_maps_to_tool_use(self):
+        assert finish_reason_to_claude_stop_reason("tool_calls") == "tool_use"
+
+    def test_function_call_maps_to_tool_use(self):
+        assert finish_reason_to_claude_stop_reason("function_call") == "tool_use"
+
+    def test_content_filter_maps_to_end_turn(self):
+        assert finish_reason_to_claude_stop_reason("content_filter") == "end_turn"
+
+    def test_none_returns_none(self):
+        assert finish_reason_to_claude_stop_reason(None) is None
+
+
+class TestClaudeRequestToInternal:
+    """Tests for converting Claude Messages API requests to ResponsesRequest."""
+
+    def test_basic_request_conversion(self):
+        request = ClaudeMessagesRequest(
+            model="claude-3-opus",
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_internal(request)
+
+        assert params.model == "claude-3-opus"
+        assert params.max_output_tokens == 100
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].role == "user"
+        assert params.input[0].content == "Hello"
+        assert params.instructions is None
+
+    def test_request_with_system_string(self):
+        request = ClaudeMessagesRequest(
+            model="claude-3-opus",
+            max_tokens=100,
+            system="You are a helpful assistant.",
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_internal(request)
+
+        assert params.instructions == "You are a helpful assistant."
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].role == "user"
+        assert params.input[0].content == "Hello"
+
+    def test_request_with_system_text_blocks(self):
+        request = ClaudeMessagesRequest(
+            model="claude-3-opus",
+            max_tokens=100,
+            system=[
+                ClaudeTextBlock(text="You are helpful. "),
+                ClaudeTextBlock(text="Be concise."),
+            ],
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+            ],
+        )
+        params = claude_request_to_internal(request)
+
+        assert params.instructions == "You are helpful. Be concise."
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+
+    def test_request_with_content_blocks(self):
+        request = ClaudeMessagesRequest(
+            model="claude-3-opus",
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(
+                    role="user",
+                    content=[
+                        ClaudeTextBlock(text="First part. "),
+                        ClaudeTextBlock(text="Second part."),
+                    ],
+                ),
+            ],
+        )
+        params = claude_request_to_internal(request)
+
+        assert isinstance(params.input, list)
+        assert len(params.input) == 1
+        assert params.input[0].content == "First part. Second part."
+
+    def test_request_with_multi_turn_conversation(self):
+        request = ClaudeMessagesRequest(
+            model="claude-3-opus",
+            max_tokens=100,
+            messages=[
+                ClaudeMessage(role="user", content="Hello"),
+                ClaudeMessage(role="assistant", content="Hi there!"),
+                ClaudeMessage(role="user", content="How are you?"),
+            ],
+        )
+        params = claude_request_to_internal(request)
+
+        assert isinstance(params.input, list)
+        assert len(params.input) == 3
+        assert params.input[0].role == "user"
+        assert params.input[1].role == "assistant"
+        assert params.input[2].role == "user"
+
+    def test_request_with_optional_parameters(self):
+        request = ClaudeMessagesRequest(
+            model="claude-3-opus",
+            max_tokens=100,
+            messages=[ClaudeMessage(role="user", content="Hello")],
+            temperature=0.7,
+            top_p=0.9,
+            top_k=40,
+            stop_sequences=["STOP", "END"],
+            stream=True,
+        )
+        params = claude_request_to_internal(request)
+
+        assert params.temperature == 0.7
+        assert params.top_p == 0.9
+        assert params.top_k == 40
+        assert params.stop == ["STOP", "END"]
+        assert params.stream is True
+
+
+class TestClaudeMessagesRequestValidation:
+    """Tests for Claude Messages API request validation."""
+
+    def test_request_requires_model(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "max_tokens": 100,
+                    "messages": [{"role": "user", "content": "Hello"}],
+                }
+            )
+
+    def test_request_requires_max_tokens(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "model": "claude-3-opus",
+                    "messages": [{"role": "user", "content": "Hello"}],
+                }
+            )
+
+    def test_request_requires_messages(self):
+        with pytest.raises(pydantic.ValidationError):
+            ClaudeMessagesRequest.model_validate(
+                {
+                    "model": "claude-3-opus",
+                    "max_tokens": 100,
+                }
+            )
+
+
+class TestClaudeStreamingEvents:
+    """Tests for Claude Messages API streaming event serialization."""
+
+    def test_message_start_event_format(self):
+        message = ClaudeMessageStart(
+            id="msg_123",
+            model="claude-3-opus",
+            content=[],
+            stop_reason=None,
+            usage=ClaudeUsage(input_tokens=10, output_tokens=0),
+        )
+        event = ClaudeMessageStartEvent(message=message)
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "message_start"
+        assert parsed["message"]["id"] == "msg_123"
+        assert parsed["message"]["type"] == "message"
+        assert parsed["message"]["role"] == "assistant"
+        assert parsed["message"]["model"] == "claude-3-opus"
+
+    def test_content_block_start_event_format(self):
+        event = ClaudeContentBlockStartEvent(
+            index=0,
+            content_block=ClaudeTextBlock(text=""),
+        )
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "content_block_start"
+        assert parsed["index"] == 0
+        assert parsed["content_block"]["type"] == "text"
+        assert parsed["content_block"]["text"] == ""
+
+    def test_content_block_delta_event_format(self):
+        event = ClaudeContentBlockDeltaEvent(
+            index=0,
+            delta=ClaudeTextDelta(text="Hello"),
+        )
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "content_block_delta"
+        assert parsed["index"] == 0
+        assert parsed["delta"]["type"] == "text_delta"
+        assert parsed["delta"]["text"] == "Hello"
+
+    def test_content_block_stop_event_format(self):
+        event = ClaudeContentBlockStopEvent(index=0)
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "content_block_stop"
+        assert parsed["index"] == 0
+
+    def test_message_delta_event_format(self):
+        event = ClaudeMessageDeltaEvent(
+            delta=ClaudeMessageDelta(stop_reason="end_turn"),
+            usage=ClaudeMessageDeltaUsage(output_tokens=25),
+        )
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "message_delta"
+        assert parsed["delta"]["stop_reason"] == "end_turn"
+        assert parsed["usage"]["output_tokens"] == 25
+
+    def test_message_stop_event_format(self):
+        event = ClaudeMessageStopEvent()
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "message_stop"
+
+    def test_sse_format(self):
+        """Test that SSE format is correctly generated."""
+        event = ClaudeContentBlockDeltaEvent(
+            index=0,
+            delta=ClaudeTextDelta(text="Hello"),
+        )
+        # Simulate the SSE format used in the streaming generator
+        sse_line = f"event: content_block_delta\ndata: {event.model_dump_json()}\n\n"
+
+        assert sse_line.startswith("event: content_block_delta\n")
+        assert "data: " in sse_line
+        assert sse_line.endswith("\n\n")

src/exo/master/tests/test_master.py

@@ -7,7 +7,6 @@ from loguru import logger
 
 from exo.master.main import Master
 from exo.routing.router import get_node_id_keypair
-from exo.shared.types.api import ChatCompletionMessage, ChatCompletionTaskParams
 from exo.shared.types.commands import (
     ChatCompletion,
     CommandId,
@@ -24,6 +23,7 @@ from exo.shared.types.events import (
 )
 from exo.shared.types.memory import Memory
 from exo.shared.types.models import ModelId, ModelMetadata
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.profiling import (
     MemoryPerformanceProfile,
     NodePerformanceProfile,
@@ -143,13 +143,9 @@ async def test_master():
                 command=(
                     ChatCompletion(
                         command_id=CommandId(),
-                        request_params=ChatCompletionTaskParams(
+                        request_params=ResponsesRequest(
                             model="llama-3.2-1b",
-                            messages=[
-                                ChatCompletionMessage(
-                                    role="user", content="Hello, how are you?"
-                                )
-                            ],
+                            input="Hello, how are you?",
                         ),
                     )
                 ),
@@ -200,11 +196,9 @@ async def test_master():
         assert isinstance(events[2].event, TaskCreated)
         assert events[2].event.task.task_status == TaskStatus.Pending
         assert isinstance(events[2].event.task, ChatCompletionTask)
-        assert events[2].event.task.task_params == ChatCompletionTaskParams(
+        assert events[2].event.task.task_params == ResponsesRequest(
             model="llama-3.2-1b",
-            messages=[
-                ChatCompletionMessage(role="user", content="Hello, how are you?")
-            ],
+            input="Hello, how are you?",
         )
 
         await master.shutdown()

src/exo/master/tests/test_openai_responses_api.py

@@ -0,0 +1,293 @@
+"""Tests for OpenAI Responses API types.
+
+ResponsesRequest is the canonical internal type used throughout the pipeline.
+No conversion is needed for Responses API requests.
+"""
+
+import json
+from typing import Any, cast
+
+import pydantic
+import pytest
+
+from exo.shared.types.openai_responses import (
+    ResponseCompletedEvent,
+    ResponseContentPartAddedEvent,
+    ResponseCreatedEvent,
+    ResponseInputMessage,
+    ResponseMessageItem,
+    ResponseOutputItemAddedEvent,
+    ResponseOutputItemDoneEvent,
+    ResponseOutputText,
+    ResponsesRequest,
+    ResponsesResponse,
+    ResponseTextDeltaEvent,
+    ResponseTextDoneEvent,
+    ResponseUsage,
+)
+
+
+class TestResponsesRequestAsCanonicalType:
+    """Tests for ResponsesRequest as the canonical internal type."""
+
+    def test_string_input(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input="Hello, how are you?",
+        )
+
+        assert request.model == "gpt-4o"
+        assert request.input == "Hello, how are you?"
+        assert request.instructions is None
+
+    def test_message_array_input(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input=[
+                ResponseInputMessage(role="user", content="Hello"),
+                ResponseInputMessage(role="assistant", content="Hi there!"),
+                ResponseInputMessage(role="user", content="How are you?"),
+            ],
+        )
+
+        assert isinstance(request.input, list)
+        assert len(request.input) == 3
+        assert request.input[0].role == "user"
+        assert request.input[0].content == "Hello"
+        assert request.input[1].role == "assistant"
+        assert request.input[1].content == "Hi there!"
+        assert request.input[2].role == "user"
+        assert request.input[2].content == "How are you?"
+
+    def test_request_with_instructions(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input="Hello",
+            instructions="You are a helpful assistant. Be concise.",
+        )
+
+        assert request.input == "Hello"
+        assert request.instructions == "You are a helpful assistant. Be concise."
+
+    def test_request_with_optional_parameters(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input="Hello",
+            max_output_tokens=500,
+            temperature=0.8,
+            top_p=0.95,
+            stream=True,
+        )
+
+        assert request.max_output_tokens == 500
+        assert request.temperature == 0.8
+        assert request.top_p == 0.95
+        assert request.stream is True
+
+    def test_request_with_new_fields(self):
+        """Test the additional fields added for internal use."""
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input="Hello",
+            top_k=40,
+            seed=42,
+            stop=["STOP", "END"],
+            tools=[{"type": "function", "function": {"name": "test"}}],
+        )
+
+        assert request.top_k == 40
+        assert request.seed == 42
+        assert request.stop == ["STOP", "END"]
+        assert request.tools == [{"type": "function", "function": {"name": "test"}}]
+
+    def test_request_with_system_role_in_messages(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input=[
+                ResponseInputMessage(role="system", content="Be helpful"),
+                ResponseInputMessage(role="user", content="Hello"),
+            ],
+        )
+
+        assert isinstance(request.input, list)
+        assert len(request.input) == 2
+        assert request.input[0].role == "system"
+        assert request.input[1].role == "user"
+
+    def test_request_with_developer_role(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input=[
+                ResponseInputMessage(role="developer", content="Internal note"),
+                ResponseInputMessage(role="user", content="Hello"),
+            ],
+        )
+
+        assert isinstance(request.input, list)
+        assert len(request.input) == 2
+        assert request.input[0].role == "developer"
+
+
+class TestResponsesRequestValidation:
+    """Tests for OpenAI Responses API request validation."""
+
+    def test_request_requires_model(self):
+        with pytest.raises(pydantic.ValidationError):
+            ResponsesRequest.model_validate(
+                {
+                    "input": "Hello",
+                }
+            )
+
+    def test_request_requires_input(self):
+        with pytest.raises(pydantic.ValidationError):
+            ResponsesRequest.model_validate(
+                {
+                    "model": "gpt-4o",
+                }
+            )
+
+    def test_request_accepts_string_input(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input="Hello",
+        )
+        assert request.input == "Hello"
+
+    def test_request_accepts_message_array_input(self):
+        request = ResponsesRequest(
+            model="gpt-4o",
+            input=[ResponseInputMessage(role="user", content="Hello")],
+        )
+        assert len(request.input) == 1
+
+
+class TestResponsesStreamingEvents:
+    """Tests for OpenAI Responses API streaming event serialization."""
+
+    def test_response_created_event_format(self):
+        response = ResponsesResponse(
+            id="resp_123",
+            model="gpt-4o",
+            status="in_progress",
+            output=[],
+            output_text="",
+        )
+        event = ResponseCreatedEvent(response=response)
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.created"
+        assert parsed["response"]["id"] == "resp_123"
+        assert parsed["response"]["object"] == "response"
+        assert parsed["response"]["status"] == "in_progress"
+
+    def test_output_item_added_event_format(self):
+        item = ResponseMessageItem(
+            id="item_123",
+            content=[ResponseOutputText(text="")],
+            status="in_progress",
+        )
+        event = ResponseOutputItemAddedEvent(output_index=0, item=item)
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.output_item.added"
+        assert parsed["output_index"] == 0
+        assert parsed["item"]["type"] == "message"
+        assert parsed["item"]["id"] == "item_123"
+        assert parsed["item"]["role"] == "assistant"
+
+    def test_content_part_added_event_format(self):
+        part = ResponseOutputText(text="")
+        event = ResponseContentPartAddedEvent(
+            output_index=0,
+            content_index=0,
+            part=part,
+        )
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.content_part.added"
+        assert parsed["output_index"] == 0
+        assert parsed["content_index"] == 0
+        assert parsed["part"]["type"] == "output_text"
+
+    def test_text_delta_event_format(self):
+        event = ResponseTextDeltaEvent(
+            output_index=0,
+            content_index=0,
+            delta="Hello",
+        )
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.output_text.delta"
+        assert parsed["output_index"] == 0
+        assert parsed["content_index"] == 0
+        assert parsed["delta"] == "Hello"
+
+    def test_text_done_event_format(self):
+        event = ResponseTextDoneEvent(
+            output_index=0,
+            content_index=0,
+            text="Hello, world!",
+        )
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.output_text.done"
+        assert parsed["text"] == "Hello, world!"
+
+    def test_output_item_done_event_format(self):
+        item = ResponseMessageItem(
+            id="item_123",
+            content=[ResponseOutputText(text="Hello, world!")],
+            status="completed",
+        )
+        event = ResponseOutputItemDoneEvent(output_index=0, item=item)
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.output_item.done"
+        assert parsed["item"]["status"] == "completed"
+        assert parsed["item"]["content"][0]["text"] == "Hello, world!"
+
+    def test_response_completed_event_format(self):
+        item = ResponseMessageItem(
+            id="item_123",
+            content=[ResponseOutputText(text="Hello!")],
+            status="completed",
+        )
+        response = ResponsesResponse(
+            id="resp_123",
+            model="gpt-4o",
+            status="completed",
+            output=[item],
+            output_text="Hello!",
+            usage=ResponseUsage(input_tokens=10, output_tokens=5, total_tokens=15),
+        )
+        event = ResponseCompletedEvent(response=response)
+        json_str = event.model_dump_json()
+        parsed = cast(dict[str, Any], json.loads(json_str))
+
+        assert parsed["type"] == "response.completed"
+        assert parsed["response"]["status"] == "completed"
+        assert parsed["response"]["output_text"] == "Hello!"
+        assert parsed["response"]["usage"]["total_tokens"] == 15
+
+    def test_sse_format(self):
+        """Test that SSE format is correctly generated."""
+        event = ResponseTextDeltaEvent(
+            output_index=0,
+            content_index=0,
+            delta="Hello",
+        )
+        # Simulate the SSE format used in the streaming generator
+        sse_line = (
+            f"event: response.output_text.delta\ndata: {event.model_dump_json()}\n\n"
+        )
+
+        assert sse_line.startswith("event: response.output_text.delta\n")
+        assert "data: " in sse_line
+        assert sse_line.endswith("\n\n")

src/exo/shared/types/api.py

@@ -11,10 +11,21 @@ from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 
 FinishReason = Literal[
-    "stop", "length", "tool_calls", "content_filter", "function_call"
+    "stop", "length", "tool_calls", "content_filter", "function_call", "error"
 ]
 
 
+class ErrorInfo(BaseModel):
+    message: str
+    type: str
+    param: str | None = None
+    code: int
+
+
+class ErrorResponse(BaseModel):
+    error: ErrorInfo
+
+
 class ModelListModel(BaseModel):
     id: str
     object: str = "model"
@@ -146,10 +157,13 @@ class ChatCompletionTaskParams(BaseModel):
     stream: bool = False
     temperature: float | None = None
     top_p: float | None = None
+    top_k: int | None = None
     tools: list[dict[str, Any]] | None = None
     tool_choice: str | dict[str, Any] | None = None
     parallel_tool_calls: bool | None = None
     user: str | None = None
+    # When True, continue the last assistant message without EOS tokens
+    continue_from_prefix: bool = False
 
 
 class BenchChatCompletionTaskParams(ChatCompletionTaskParams):

src/exo/shared/types/chunks.py

@@ -1,6 +1,6 @@
 from enum import Enum
 
-from exo.shared.types.api import GenerationStats
+from exo.shared.types.api import GenerationStats, TopLogprobItem
 from exo.utils.pydantic_ext import TaggedModel
 
 from .api import FinishReason
@@ -20,8 +20,11 @@ class BaseChunk(TaggedModel):
 class TokenChunk(BaseChunk):
     text: str
     token_id: int
+    logprob: float | None = None  # Log probability of the selected token
+    top_logprobs: list[TopLogprobItem] | None = None  # Top-k alternative tokens
     finish_reason: FinishReason | None = None
     stats: GenerationStats | None = None
+    error_message: str | None = None
 
 
 class ImageChunk(BaseChunk):

src/exo/shared/types/claude_api.py

@@ -0,0 +1,168 @@
+"""Claude Messages API types for request/response conversion."""
+
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+# Type aliases
+ClaudeRole = Literal["user", "assistant"]
+ClaudeStopReason = Literal["end_turn", "max_tokens", "stop_sequence", "tool_use"]
+
+
+# Content block types
+class ClaudeTextBlock(BaseModel, frozen=True):
+    """Text content block in Claude Messages API."""
+
+    type: Literal["text"] = "text"
+    text: str
+
+
+class ClaudeImageSource(BaseModel, frozen=True):
+    """Image source for Claude image blocks."""
+
+    type: Literal["base64", "url"]
+    media_type: str | None = None
+    data: str | None = None
+    url: str | None = None
+
+
+class ClaudeImageBlock(BaseModel, frozen=True):
+    """Image content block in Claude Messages API."""
+
+    type: Literal["image"] = "image"
+    source: ClaudeImageSource
+
+
+ClaudeContentBlock = ClaudeTextBlock | ClaudeImageBlock
+
+
+# Request types
+class ClaudeMessage(BaseModel, frozen=True):
+    """Message in Claude Messages API request."""
+
+    role: ClaudeRole
+    content: str | list[ClaudeContentBlock]
+
+
+class ClaudeMessagesRequest(BaseModel):
+    """Request body for Claude Messages API."""
+
+    model: str
+    max_tokens: int
+    messages: list[ClaudeMessage]
+    system: str | list[ClaudeTextBlock] | None = None
+    stop_sequences: list[str] | None = None
+    stream: bool = False
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    metadata: dict[str, str] | None = None
+
+
+# Response types
+class ClaudeUsage(BaseModel, frozen=True):
+    """Token usage in Claude Messages API response."""
+
+    input_tokens: int
+    output_tokens: int
+
+
+class ClaudeMessagesResponse(BaseModel, frozen=True):
+    """Response body for Claude Messages API."""
+
+    id: str
+    type: Literal["message"] = "message"
+    role: Literal["assistant"] = "assistant"
+    content: list[ClaudeTextBlock]
+    model: str
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+    usage: ClaudeUsage
+
+
+# Streaming event types
+class ClaudeMessageStart(BaseModel, frozen=True):
+    """Partial message in message_start event."""
+
+    id: str
+    type: Literal["message"] = "message"
+    role: Literal["assistant"] = "assistant"
+    content: list[ClaudeTextBlock] = Field(default_factory=list)
+    model: str
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+    usage: ClaudeUsage
+
+
+class ClaudeMessageStartEvent(BaseModel, frozen=True):
+    """Event sent at start of message stream."""
+
+    type: Literal["message_start"] = "message_start"
+    message: ClaudeMessageStart
+
+
+class ClaudeContentBlockStartEvent(BaseModel, frozen=True):
+    """Event sent at start of a content block."""
+
+    type: Literal["content_block_start"] = "content_block_start"
+    index: int
+    content_block: ClaudeTextBlock
+
+
+class ClaudeTextDelta(BaseModel, frozen=True):
+    """Delta for text content block."""
+
+    type: Literal["text_delta"] = "text_delta"
+    text: str
+
+
+class ClaudeContentBlockDeltaEvent(BaseModel, frozen=True):
+    """Event sent for content block delta."""
+
+    type: Literal["content_block_delta"] = "content_block_delta"
+    index: int
+    delta: ClaudeTextDelta
+
+
+class ClaudeContentBlockStopEvent(BaseModel, frozen=True):
+    """Event sent at end of a content block."""
+
+    type: Literal["content_block_stop"] = "content_block_stop"
+    index: int
+
+
+class ClaudeMessageDeltaUsage(BaseModel, frozen=True):
+    """Usage in message_delta event."""
+
+    output_tokens: int
+
+
+class ClaudeMessageDelta(BaseModel, frozen=True):
+    """Delta in message_delta event."""
+
+    stop_reason: ClaudeStopReason | None = None
+    stop_sequence: str | None = None
+
+
+class ClaudeMessageDeltaEvent(BaseModel, frozen=True):
+    """Event sent with final message delta."""
+
+    type: Literal["message_delta"] = "message_delta"
+    delta: ClaudeMessageDelta
+    usage: ClaudeMessageDeltaUsage
+
+
+class ClaudeMessageStopEvent(BaseModel, frozen=True):
+    """Event sent at end of message stream."""
+
+    type: Literal["message_stop"] = "message_stop"
+
+
+ClaudeStreamEvent = (
+    ClaudeMessageStartEvent
+    | ClaudeContentBlockStartEvent
+    | ClaudeContentBlockDeltaEvent
+    | ClaudeContentBlockStopEvent
+    | ClaudeMessageDeltaEvent
+    | ClaudeMessageStopEvent
+)

src/exo/shared/types/commands.py

@@ -1,8 +1,8 @@
 from pydantic import Field
 
-from exo.shared.types.api import ChatCompletionTaskParams
 from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.models import ModelMetadata
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
 from exo.utils.pydantic_ext import CamelCaseModel, TaggedModel
@@ -17,7 +17,7 @@ class TestCommand(BaseCommand):
 
 
 class ChatCompletion(BaseCommand):
-    request_params: ChatCompletionTaskParams
+    request_params: ResponsesRequest
 
 
 class PlaceInstance(BaseCommand):

src/exo/shared/types/openai_responses.py

@@ -0,0 +1,190 @@
+"""OpenAI Responses API types for request/response conversion.
+
+ResponsesRequest serves as both:
+1. The external API request type for /v1/responses
+2. The canonical internal type used throughout the inference pipeline
+
+All external API formats (Chat Completions, Claude) are converted to
+ResponsesRequest at the API boundary.
+"""
+
+import time
+from typing import Any, Literal
+
+from pydantic import BaseModel, Field
+
+# Type aliases
+ResponseStatus = Literal["completed", "failed", "in_progress", "incomplete"]
+ResponseRole = Literal["user", "assistant", "system", "developer"]
+
+
+# Request types
+class ResponseInputMessage(BaseModel, frozen=True):
+    """Input message for Responses API.
+
+    This is also used as the internal message format throughout the pipeline.
+    """
+
+    role: ResponseRole
+    content: str
+
+
+class ResponsesRequest(BaseModel):
+    """Request body for OpenAI Responses API.
+
+    This is also the canonical internal task params format used throughout
+    the inference pipeline. All external API formats are converted to this
+    format at the API boundary.
+
+    Field mapping from other APIs:
+    - input: Replaces 'messages' from Chat Completions
+    - instructions: System message, extracted from messages or Claude's 'system'
+    - max_output_tokens: Replaces 'max_tokens' from Chat Completions
+    """
+
+    model: str
+    input: str | list[ResponseInputMessage]
+    instructions: str | None = None
+    max_output_tokens: int | None = None
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    stop: str | list[str] | None = None
+    seed: int | None = None
+    stream: bool = False
+    # Tools support
+    tools: list[dict[str, Any]] | None = None
+    # previous_response_id not supported in MVP
+    metadata: dict[str, str] | None = None
+    # When True, continue the last assistant message without EOS tokens
+    continue_from_prefix: bool = False
+
+
+# Response types
+class ResponseOutputText(BaseModel, frozen=True):
+    """Text content in response output."""
+
+    type: Literal["output_text"] = "output_text"
+    text: str
+    annotations: list[dict[str, str]] = Field(default_factory=list)
+
+
+class ResponseMessageItem(BaseModel, frozen=True):
+    """Message item in response output array."""
+
+    type: Literal["message"] = "message"
+    id: str
+    role: Literal["assistant"] = "assistant"
+    content: list[ResponseOutputText]
+    status: ResponseStatus = "completed"
+
+
+ResponseItem = ResponseMessageItem  # Can expand for function_call, reasoning, etc.
+
+
+class ResponseUsage(BaseModel, frozen=True):
+    """Token usage in Responses API response."""
+
+    input_tokens: int
+    output_tokens: int
+    total_tokens: int
+
+
+class ResponsesResponse(BaseModel, frozen=True):
+    """Response body for OpenAI Responses API."""
+
+    id: str
+    object: Literal["response"] = "response"
+    created_at: int = Field(default_factory=lambda: int(time.time()))
+    status: ResponseStatus = "completed"
+    model: str
+    output: list[ResponseItem]
+    output_text: str
+    usage: ResponseUsage | None = None
+
+
+# Streaming event types
+class ResponseCreatedEvent(BaseModel, frozen=True):
+    """Event sent when response is created."""
+
+    type: Literal["response.created"] = "response.created"
+    response: ResponsesResponse
+
+
+class ResponseInProgressEvent(BaseModel, frozen=True):
+    """Event sent when response starts processing."""
+
+    type: Literal["response.in_progress"] = "response.in_progress"
+    response: ResponsesResponse
+
+
+class ResponseOutputItemAddedEvent(BaseModel, frozen=True):
+    """Event sent when an output item is added."""
+
+    type: Literal["response.output_item.added"] = "response.output_item.added"
+    output_index: int
+    item: ResponseItem
+
+
+class ResponseContentPartAddedEvent(BaseModel, frozen=True):
+    """Event sent when a content part is added."""
+
+    type: Literal["response.content_part.added"] = "response.content_part.added"
+    output_index: int
+    content_index: int
+    part: ResponseOutputText
+
+
+class ResponseTextDeltaEvent(BaseModel, frozen=True):
+    """Event sent for text delta during streaming."""
+
+    type: Literal["response.output_text.delta"] = "response.output_text.delta"
+    output_index: int
+    content_index: int
+    delta: str
+
+
+class ResponseTextDoneEvent(BaseModel, frozen=True):
+    """Event sent when text content is done."""
+
+    type: Literal["response.output_text.done"] = "response.output_text.done"
+    output_index: int
+    content_index: int
+    text: str
+
+
+class ResponseContentPartDoneEvent(BaseModel, frozen=True):
+    """Event sent when a content part is done."""
+
+    type: Literal["response.content_part.done"] = "response.content_part.done"
+    output_index: int
+    content_index: int
+    part: ResponseOutputText
+
+
+class ResponseOutputItemDoneEvent(BaseModel, frozen=True):
+    """Event sent when an output item is done."""
+
+    type: Literal["response.output_item.done"] = "response.output_item.done"
+    output_index: int
+    item: ResponseItem
+
+
+class ResponseCompletedEvent(BaseModel, frozen=True):
+    """Event sent when response is completed."""
+
+    type: Literal["response.completed"] = "response.completed"
+    response: ResponsesResponse
+
+
+ResponsesStreamEvent = (
+    ResponseCreatedEvent
+    | ResponseInProgressEvent
+    | ResponseOutputItemAddedEvent
+    | ResponseContentPartAddedEvent
+    | ResponseTextDeltaEvent
+    | ResponseTextDoneEvent
+    | ResponseContentPartDoneEvent
+    | ResponseOutputItemDoneEvent
+    | ResponseCompletedEvent
+)

src/exo/shared/types/tasks.py

@@ -2,8 +2,8 @@ from enum import Enum
 
 from pydantic import Field
 
-from exo.shared.types.api import ChatCompletionTaskParams
 from exo.shared.types.common import CommandId, Id
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.worker.instances import BoundInstance, InstanceId
 from exo.shared.types.worker.runners import RunnerId
 from exo.shared.types.worker.shards import ShardMetadata
@@ -50,7 +50,7 @@ class StartWarmup(BaseTask):  # emitted by Worker
 
 class ChatCompletion(BaseTask):  # emitted by Master
     command_id: CommandId
-    task_params: ChatCompletionTaskParams
+    task_params: ResponsesRequest
 
     error_type: str | None = Field(default=None)
     error_message: str | None = Field(default=None)

src/exo/shared/types/worker/runner_response.py

@@ -1,4 +1,4 @@
-from exo.shared.types.api import FinishReason, GenerationStats
+from exo.shared.types.api import FinishReason, GenerationStats, TopLogprobItem
 from exo.utils.pydantic_ext import TaggedModel
 
 
@@ -13,7 +13,8 @@ class TokenizedResponse(BaseRunnerResponse):
 class GenerationResponse(BaseRunnerResponse):
     text: str
     token: int
-    # logprobs: list[float] | None = None # too big. we can change to be top-k
+    logprob: float | None = None  # Log probability of the selected token
+    top_logprobs: list[TopLogprobItem] | None = None  # Top-k alternative tokens
     finish_reason: FinishReason | None = None
     stats: GenerationStats | None = None
 

src/exo/worker/engines/mlx/__init__.py

@@ -40,4 +40,6 @@ class TokenizerWrapper:
         messages_dicts: list[dict[str, Any]],
         tokenize: bool = False,
         add_generation_prompt: bool = True,
+        continue_final_message: bool = False,
+        tools: list[dict[str, Any]] | None = None,
     ) -> str: ...

src/exo/worker/engines/mlx/generator/generate.py

@@ -8,13 +8,12 @@ from mlx_lm.tokenizer_utils import TokenizerWrapper
 
 # from exo.engines.mlx.cache import KVPrefixCache
 from exo.shared.types.api import (
-    BenchChatCompletionTaskParams,
-    ChatCompletionMessage,
     FinishReason,
     GenerationStats,
+    TopLogprobItem,
 )
 from exo.shared.types.memory import Memory
-from exo.shared.types.tasks import ChatCompletionTaskParams
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.worker.runner_response import (
     GenerationResponse,
 )
@@ -53,14 +52,9 @@ def warmup_inference(
 
     warmup_prompt = apply_chat_template(
         tokenizer=tokenizer,
-        chat_task_data=ChatCompletionTaskParams(
+        task_params=ResponsesRequest(
             model="",
-            messages=[
-                ChatCompletionMessage(
-                    role="user",
-                    content=content,
-                )
-            ],
+            input=content,
         ),
     )
 
@@ -118,11 +112,11 @@ def eos_ids_from_tokenizer(tokenizer: TokenizerWrapper) -> list[int]:
 def mlx_generate(
     model: Model,
     tokenizer: TokenizerWrapper,
-    task: ChatCompletionTaskParams,
+    task: ResponsesRequest,
+    is_bench: bool = False,
 ) -> Generator[GenerationResponse]:
     # Ensure that generation stats only contains peak memory for this generation
     mx.reset_peak_memory()
-    is_bench: bool = isinstance(task, BenchChatCompletionTaskParams)
 
     # Currently we support chat-completion tasks only.
     logger.info(f"task_params: {task}")
@@ -132,7 +126,7 @@ def mlx_generate(
 
     prompt = apply_chat_template(
         tokenizer=tokenizer,
-        chat_task_data=task,
+        task_params=task,
     )
 
     caches = make_kv_cache(model=model)
@@ -146,9 +140,20 @@ def mlx_generate(
     sampler = make_sampler(
         temp=task.temperature if task.temperature is not None else 0.7,
         top_p=task.top_p if task.top_p is not None else 1.0,
+        top_k=task.top_k if task.top_k is not None else 0,
     )
 
-    max_tokens = task.max_tokens or MAX_TOKENS
+    # Normalize stop sequences to a list
+    stop_sequences: list[str] = (
+        ([task.stop] if isinstance(task.stop, str) else task.stop)
+        if task.stop is not None
+        else []
+    )
+    max_stop_len = max((len(s) for s in stop_sequences), default=0)
+
+    max_tokens = task.max_output_tokens or MAX_TOKENS
+    accumulated_text = ""
+
     for out in stream_generate(
         model=model,
         tokenizer=tokenizer,
@@ -161,11 +166,34 @@ def mlx_generate(
         prefill_step_size=2048,
         kv_group_size=KV_GROUP_SIZE,
         kv_bits=KV_BITS,
+        return_logprob=True,
+        return_top_logprobs=5,
     ):
         logger.info(out.text)
+        accumulated_text += out.text
 
+        # Check for stop sequences
+        text = out.text
+        finish_reason: FinishReason | None = cast(
+            FinishReason | None, out.finish_reason
+        )
+        stop_matched = False
+
+        if stop_sequences:
+            for stop_seq in stop_sequences:
+                if stop_seq in accumulated_text:
+                    # Trim text to just before the stop sequence
+                    stop_index = accumulated_text.find(stop_seq)
+                    text_before_stop = accumulated_text[:stop_index]
+                    chunk_start = len(accumulated_text) - len(out.text)
+                    text = text_before_stop[chunk_start:]
+                    finish_reason = "stop"
+                    stop_matched = True
+                    break
+
+        is_done = finish_reason is not None
         stats: GenerationStats | None = None
-        if out.finish_reason is not None:
+        if is_done:
             stats = GenerationStats(
                 prompt_tps=float(out.prompt_tps),
                 generation_tps=float(out.generation_tps),
@@ -173,22 +201,41 @@ def mlx_generate(
                 generation_tokens=int(out.generation_tokens),
                 peak_memory_usage=Memory.from_gb(out.peak_memory),
             )
-
-            if out.finish_reason not in get_args(FinishReason):
-                # We don't throw here as this failure case is really not all that bad
-                # Just log the error and move on
+            if not stop_matched and out.finish_reason not in get_args(FinishReason):
                 logger.warning(
                     f"Model generated unexpected finish_reason: {out.finish_reason}"
                 )
 
+        # Extract logprobs if available
+        logprob: float | None = getattr(out, "logprob", None)
+        top_logprobs_raw: list[tuple[int, float]] | None = getattr(
+            out, "top_logprobs", None
+        )
+
+        top_logprobs: list[TopLogprobItem] | None = None
+        if top_logprobs_raw is not None:
+            top_logprobs = [
+                TopLogprobItem(
+                    token=text if i == 0 else tokenizer.decode([tok_id]),
+                    logprob=float(lp),
+                )
+                for i, (tok_id, lp) in enumerate(top_logprobs_raw)
+            ]
+
         yield GenerationResponse(
-            text=out.text,
+            text=text,
             token=out.token,
-            finish_reason=cast(FinishReason | None, out.finish_reason),
+            logprob=logprob,
+            top_logprobs=top_logprobs,
+            finish_reason=finish_reason,
             stats=stats,
         )
 
-        if out.finish_reason is not None:
+        if is_done:
             break
 
+        # Limit accumulated_text to what's needed for stop sequence detection
+        if max_stop_len > 0 and len(accumulated_text) > max_stop_len:
+            accumulated_text = accumulated_text[-max_stop_len:]
+
         # TODO: Do we want an mx_barrier?

src/exo/worker/engines/mlx/utils_mlx.py

@@ -2,7 +2,9 @@ import json
 import os
 import resource
 import sys
+import threading
 import time
+from collections.abc import Callable
 from pathlib import Path
 from typing import Any, cast
 
@@ -40,10 +42,9 @@ import mlx.nn as nn
 from mlx_lm.utils import load_model
 from pydantic import RootModel
 
-from exo.shared.types.api import ChatCompletionMessageText
 from exo.shared.types.common import Host
 from exo.shared.types.memory import Memory
-from exo.shared.types.tasks import ChatCompletionTaskParams
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.worker.instances import (
     BoundInstance,
     MlxJacclInstance,
@@ -82,6 +83,45 @@ def get_weights_size(model_shard_meta: ShardMetadata) -> Memory:
     )
 
 
+class ModelLoadingTimeoutError(Exception):
+    pass
+
+
+TimeoutCallback = Callable[[], None]
+
+
+def eval_with_timeout(
+    mlx_item: Any,  # pyright: ignore[reportAny]
+    timeout_seconds: float = 60.0,
+    on_timeout: TimeoutCallback | None = None,
+) -> None:
+    """Evaluate MLX item with a hard timeout.
+
+    If on_timeout callback is provided, it will be called before terminating
+    the process. This allows the runner to send a failure event before exit.
+    """
+    completed = threading.Event()
+
+    def watchdog() -> None:
+        if not completed.wait(timeout=timeout_seconds):
+            logger.error(
+                f"mlx_item evaluation timed out after {timeout_seconds:.0f}s. "
+                "This may indicate an issue with FAST_SYNCH and tensor parallel sharding. "
+                "Terminating process."
+            )
+            if on_timeout is not None:
+                on_timeout()
+            os._exit(1)
+
+    watchdog_thread = threading.Thread(target=watchdog, daemon=True)
+    watchdog_thread.start()
+
+    try:
+        mx.eval(mlx_item)  # pyright: ignore[reportAny]
+    finally:
+        completed.set()
+
+
 def mx_barrier(group: Group | None = None):
     mx.eval(
         mx.distributed.all_sum(
@@ -188,7 +228,9 @@ def initialize_mlx(
 
 
 def load_mlx_items(
-    bound_instance: BoundInstance, group: Group | None
+    bound_instance: BoundInstance,
+    group: Group | None,
+    on_timeout: TimeoutCallback | None = None,
 ) -> tuple[Model, TokenizerWrapper]:
     if group is None:
         logger.info(f"Single device used for {bound_instance.instance}")
@@ -202,7 +244,9 @@ def load_mlx_items(
     else:
         logger.info("Starting distributed init")
         start_time = time.perf_counter()
-        model, tokenizer = shard_and_load(bound_instance.bound_shard, group=group)
+        model, tokenizer = shard_and_load(
+            bound_instance.bound_shard, group=group, on_timeout=on_timeout
+        )
         end_time = time.perf_counter()
         logger.info(
             f"Time taken to shard and load model: {(end_time - start_time):.2f}s"
@@ -216,6 +260,7 @@ def load_mlx_items(
 def shard_and_load(
     shard_metadata: ShardMetadata,
     group: Group,
+    on_timeout: TimeoutCallback | None = None,
 ) -> tuple[nn.Module, TokenizerWrapper]:
     model_path = build_model_path(shard_metadata.model_meta.model_id)
 
@@ -252,7 +297,15 @@ def shard_and_load(
             logger.info(f"loading model from {model_path} with pipeline parallelism")
             model = pipeline_auto_parallel(model, group, shard_metadata)
 
-    mx.eval(model.parameters())
+    # Estimate timeout based on model size
+    base_timeout = float(os.environ.get("EXO_MODEL_LOAD_TIMEOUT", "60"))
+    model_size_gb = get_weights_size(shard_metadata).in_bytes / (1024**3)
+    timeout_seconds = base_timeout + model_size_gb / 5
+    logger.info(
+        f"Evaluating model parameters with timeout of {timeout_seconds:.0f}s "
+        f"(model size: {model_size_gb:.1f}GB)"
+    )
+    eval_with_timeout(model.parameters(), timeout_seconds, on_timeout)
 
     # TODO: Do we need this?
     mx.eval(model)
@@ -336,35 +389,53 @@ def load_tokenizer_for_model_id(model_id: str, model_path: Path) -> TokenizerWra
 
 def apply_chat_template(
     tokenizer: TokenizerWrapper,
-    chat_task_data: ChatCompletionTaskParams,
+    task_params: ResponsesRequest,
 ) -> str:
-    # Now we can properly access the messages
-    messages = chat_task_data.messages
+    """Convert ResponsesRequest to a chat template prompt.
 
+    Converts the internal format (input + instructions) to a messages list
+    that can be processed by the tokenizer's chat template.
+    """
     formatted_messages: list[dict[str, Any]] = []
-    for message in messages:
-        if isinstance(message.content, ChatCompletionMessageText):
-            message.content = message.content.text
-        if isinstance(message.content, list):
-            if len(message.content) == 0:
-                logger.warning("Received prompt with no content, skipping")
-                continue
 
-            message.content = "\n".join(c.text for c in message.content).strip()
-        if message.content is None and message.thinking is None:
-            continue
-
-        # Null values are not valid when applying templates in tokenizer
+    # Add system message (instructions) if present
+    if task_params.instructions:
         formatted_messages.append(
-            {k: v for k, v in message.model_dump().items() if v is not None}  # type: ignore
+            {"role": "system", "content": task_params.instructions}
         )
 
-    prompt: str = tokenizer.apply_chat_template(
-        formatted_messages,
-        tokenize=False,
-        add_generation_prompt=True,
-        tools=chat_task_data.tools,
-    )
+    # Convert input to messages
+    if isinstance(task_params.input, str):
+        # Simple string input becomes a single user message
+        formatted_messages.append({"role": "user", "content": task_params.input})
+    else:
+        # List of InputMessage
+        for msg in task_params.input:
+            if not msg.content:
+                logger.warning("Received message with empty content, skipping")
+                continue
+            formatted_messages.append({"role": msg.role, "content": msg.content})
+
+    # Use continue_final_message when continuing from prefix (e.g., regenerate from token)
+    # This keeps the final assistant message open without EOS tokens
+    # Note: explicitly set add_generation_prompt=False when using continue_final_message
+    # because some tokenizers (e.g., Kimi) default add_generation_prompt=True
+    prompt: str
+    if task_params.continue_from_prefix:
+        prompt = tokenizer.apply_chat_template(
+            formatted_messages,
+            tokenize=False,
+            continue_final_message=True,
+            add_generation_prompt=False,
+            tools=task_params.tools,
+        )
+    else:
+        prompt = tokenizer.apply_chat_template(
+            formatted_messages,
+            tokenize=False,
+            add_generation_prompt=True,
+            tools=task_params.tools,
+        )
 
     logger.info(prompt)
 

src/exo/worker/runner/bootstrap.py

@@ -17,15 +17,23 @@ def entrypoint(
     task_receiver: MpReceiver[Task],
     _logger: "loguru.Logger",
 ) -> None:
-    if (
-        isinstance(bound_instance.instance, MlxJacclInstance)
-        and len(bound_instance.instance.ibv_devices) >= 2
+    fast_synch_override = os.environ.get("EXO_FAST_SYNCH")
+    if fast_synch_override == "on" or (
+        fast_synch_override != "off"
+        and (
+            isinstance(bound_instance.instance, MlxJacclInstance)
+            and len(bound_instance.instance.ibv_devices) >= 2
+        )
     ):
         os.environ["MLX_METAL_FAST_SYNCH"] = "1"
+    else:
+        os.environ["MLX_METAL_FAST_SYNCH"] = "0"
 
     global logger
     logger = _logger
 
+    logger.info(f"Fast synch flag: {os.environ['MLX_METAL_FAST_SYNCH']}")
+
     # Import main after setting global logger - this lets us just import logger from this module
     try:
         from exo.worker.runner.runner import main

src/exo/worker/runner/runner.py

@@ -1,6 +1,8 @@
 import time
 from collections.abc import Generator
+from contextlib import contextmanager
 from functools import cache
+from typing import cast
 
 import mlx.core as mx
 from mlx_lm.models.gpt_oss import Model as GptOssModel
@@ -11,8 +13,8 @@ from openai_harmony import (  # pyright: ignore[reportMissingTypeStubs]
     load_harmony_encoding,
 )
 
-from exo.shared.types.api import ChatCompletionMessageText
 from exo.shared.types.chunks import TokenChunk
+from exo.shared.types.common import CommandId
 from exo.shared.types.events import (
     ChunkGenerated,
     Event,
@@ -20,6 +22,8 @@ from exo.shared.types.events import (
     TaskAcknowledged,
     TaskStatusUpdated,
 )
+from exo.shared.types.models import ModelId
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.tasks import (
     ChatCompletion,
     ConnectToGroup,
@@ -48,6 +52,7 @@ from exo.shared.types.worker.runners import (
     RunnerWarmingUp,
 )
 from exo.utils.channels import MpReceiver, MpSender
+from exo.worker.engines.mlx import Model
 from exo.worker.engines.mlx.generator.generate import mlx_generate, warmup_inference
 from exo.worker.engines.mlx.utils_mlx import (
     initialize_mlx,
@@ -57,6 +62,33 @@ from exo.worker.engines.mlx.utils_mlx import (
 from exo.worker.runner.bootstrap import logger
 
 
+@contextmanager
+def send_error_chunk_on_exception(
+    event_sender: MpSender[Event],
+    command_id: CommandId,
+    model_id: ModelId,
+    device_rank: int,
+):
+    try:
+        yield
+    except Exception as e:
+        logger.error(e)
+        if device_rank == 0:
+            event_sender.send(
+                ChunkGenerated(
+                    command_id=command_id,
+                    chunk=TokenChunk(
+                        idx=0,
+                        model=model_id,
+                        text="",
+                        token_id=0,
+                        finish_reason="error",
+                        error_message=str(e),
+                    ),
+                )
+            )
+
+
 def main(
     bound_instance: BoundInstance,
     event_sender: MpSender[Event],
@@ -118,7 +150,20 @@ def main(
                         )
                     )
 
-                    model, tokenizer = load_mlx_items(bound_instance, group)
+                    def on_model_load_timeout() -> None:
+                        event_sender.send(
+                            RunnerStatusUpdated(
+                                runner_id=runner_id,
+                                runner_status=RunnerFailed(
+                                    error_message="Model loading timed out"
+                                ),
+                            )
+                        )
+                        time.sleep(0.5)
+
+                    model, tokenizer = load_mlx_items(
+                        bound_instance, group, on_timeout=on_model_load_timeout
+                    )
 
                     current_status = RunnerLoaded()
                     logger.info("runner loaded")
@@ -135,7 +180,7 @@ def main(
 
                     logger.info(f"warming up inference for instance: {instance}")
                     toks = warmup_inference(
-                        model=model,
+                        model=cast(Model, model),
                         tokenizer=tokenizer,
                         # kv_prefix_cache=kv_prefix_cache,  # supply for warmup-time prefix caching
                     )
@@ -148,8 +193,6 @@ def main(
                 case ChatCompletion(task_params=task_params, command_id=command_id) if (
                     isinstance(current_status, RunnerReady)
                 ):
-                    assert model
-                    assert tokenizer
                     logger.info(f"received chat request: {str(task)[:500]}")
                     current_status = RunnerRunning()
                     logger.info("runner running")
@@ -158,41 +201,48 @@ def main(
                             runner_id=runner_id, runner_status=current_status
                         )
                     )
-                    assert task_params.messages[0].content is not None
-                    _check_for_debug_prompts(task_params.messages[0].content)
+                    with send_error_chunk_on_exception(
+                        event_sender,
+                        command_id,
+                        shard_metadata.model_meta.model_id,
+                        shard_metadata.device_rank,
+                    ):
+                        assert model
+                        assert tokenizer
+                        _check_for_debug_prompts(task_params)
 
-                    # Generate responses using the actual MLX generation
-                    mlx_generator = mlx_generate(
-                        model=model,
-                        tokenizer=tokenizer,
-                        task=task_params,
-                    )
+                        # Generate responses using the actual MLX generation
+                        mlx_generator = mlx_generate(
+                            model=cast(Model, model),
+                            tokenizer=tokenizer,
+                            task=task_params,
+                        )
 
-                    # GPT-OSS specific parsing to match other model formats.
-                    if isinstance(model, GptOssModel):
-                        mlx_generator = parse_gpt_oss(mlx_generator)
+                        # GPT-OSS specific parsing to match other model formats.
+                        if isinstance(model, GptOssModel):
+                            mlx_generator = parse_gpt_oss(mlx_generator)
 
-                    # TODO: Add tool call parser here
+                        # TODO: Add tool call parser here
 
-                    for response in mlx_generator:
-                        match response:
-                            case GenerationResponse():
-                                if shard_metadata.device_rank == 0:
-                                    event_sender.send(
-                                        ChunkGenerated(
-                                            command_id=command_id,
-                                            chunk=TokenChunk(
-                                                idx=response.token,
-                                                model=shard_metadata.model_meta.model_id,
-                                                text=response.text,
-                                                token_id=response.token,
-                                                finish_reason=response.finish_reason,
-                                                stats=response.stats,
-                                            ),
+                        for response in mlx_generator:
+                            match response:
+                                case GenerationResponse():
+                                    if shard_metadata.device_rank == 0:
+                                        event_sender.send(
+                                            ChunkGenerated(
+                                                command_id=command_id,
+                                                chunk=TokenChunk(
+                                                    idx=response.token,
+                                                    model=shard_metadata.model_meta.model_id,
+                                                    text=response.text,
+                                                    token_id=response.token,
+                                                    logprob=response.logprob,
+                                                    top_logprobs=response.top_logprobs,
+                                                    finish_reason=response.finish_reason,
+                                                    stats=response.stats,
+                                                ),
+                                            )
                                         )
-                                    )
-                                # case TokenizedResponse():
-                                # TODO: something here ig
 
                     current_status = RunnerReady()
                     logger.info("runner ready")
@@ -266,17 +316,23 @@ EXO_RUNNER_MUST_OOM = "EXO RUNNER MUST OOM"
 EXO_RUNNER_MUST_TIMEOUT = "EXO RUNNER MUST TIMEOUT"
 
 
-def _check_for_debug_prompts(
-    prompt: str | ChatCompletionMessageText | list[ChatCompletionMessageText],
-):
-    if isinstance(prompt, list):
-        if len(prompt) == 0:
-            logger.debug("Empty message prompt received in debug prompt")
-            return
-        prompt = prompt[0]
+def _check_for_debug_prompts(task_params: ResponsesRequest) -> None:
+    """Check for debug prompt triggers in the input.
 
-    if isinstance(prompt, ChatCompletionMessageText):
-        prompt = prompt.text
+    Extracts the first user input text and checks for debug triggers.
+    """
+    prompt: str
+    if isinstance(task_params.input, str):
+        prompt = task_params.input
+    else:
+        # List of InputMessage - get first message content
+        if len(task_params.input) == 0:
+            logger.debug("Empty message list in debug prompt check")
+            return
+        prompt = task_params.input[0].content
+
+    if not prompt:
+        return
 
     if EXO_RUNNER_MUST_FAIL in prompt:
         logger.info("raising exception")

src/exo/worker/tests/unittests/test_plan/test_task_forwarding.py

@@ -1,7 +1,7 @@
 from typing import cast
 
 import exo.worker.plan as plan_mod
-from exo.shared.types.api import ChatCompletionTaskParams
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.tasks import ChatCompletion, Task, TaskId, TaskStatus
 from exo.shared.types.worker.instances import BoundInstance, InstanceId
 from exo.shared.types.worker.runners import (
@@ -59,7 +59,7 @@ def test_plan_forwards_pending_chat_completion_when_runner_ready():
         instance_id=INSTANCE_1_ID,
         task_status=TaskStatus.Pending,
         command_id=COMMAND_1_ID,
-        task_params=ChatCompletionTaskParams(model=MODEL_A_ID, messages=[]),
+        task_params=ResponsesRequest(model=MODEL_A_ID, input=""),
     )
 
     result = plan_mod.plan(
@@ -107,7 +107,7 @@ def test_plan_does_not_forward_chat_completion_if_any_runner_not_ready():
         instance_id=INSTANCE_1_ID,
         task_status=TaskStatus.Pending,
         command_id=COMMAND_1_ID,
-        task_params=ChatCompletionTaskParams(model=MODEL_A_ID, messages=[]),
+        task_params=ResponsesRequest(model=MODEL_A_ID, input=""),
     )
 
     result = plan_mod.plan(
@@ -152,7 +152,7 @@ def test_plan_does_not_forward_tasks_for_other_instances():
         instance_id=other_instance_id,
         task_status=TaskStatus.Pending,
         command_id=COMMAND_1_ID,
-        task_params=ChatCompletionTaskParams(model=MODEL_A_ID, messages=[]),
+        task_params=ResponsesRequest(model=MODEL_A_ID, input=""),
     )
 
     result = plan_mod.plan(
@@ -201,7 +201,7 @@ def test_plan_ignores_non_pending_or_non_chat_tasks():
         instance_id=INSTANCE_1_ID,
         task_status=TaskStatus.Complete,
         command_id=COMMAND_1_ID,
-        task_params=ChatCompletionTaskParams(model=MODEL_A_ID, messages=[]),
+        task_params=ResponsesRequest(model=MODEL_A_ID, input=""),
     )
 
     other_task_id = TaskId("other-task")

src/exo/worker/tests/unittests/test_runner/test_error_handling.py

@@ -0,0 +1,50 @@
+# pyright: reportAny=false
+from unittest.mock import MagicMock
+
+from exo.shared.types.chunks import TokenChunk
+from exo.shared.types.common import CommandId
+from exo.shared.types.events import ChunkGenerated
+from exo.worker.runner.runner import send_error_chunk_on_exception
+from exo.worker.tests.constants import MODEL_A_ID
+
+
+def test_send_error_chunk_on_exception_no_error() -> None:
+    event_sender = MagicMock()
+    command_id = CommandId()
+
+    with send_error_chunk_on_exception(
+        event_sender, command_id, MODEL_A_ID, device_rank=0
+    ):
+        _ = 1 + 1
+
+    event_sender.send.assert_not_called()
+
+
+def test_send_error_chunk_on_exception_catches_error() -> None:
+    event_sender = MagicMock()
+    command_id = CommandId()
+
+    with send_error_chunk_on_exception(
+        event_sender, command_id, MODEL_A_ID, device_rank=0
+    ):
+        raise ValueError("test error")
+
+    event_sender.send.assert_called_once()
+    call_args = event_sender.send.call_args[0][0]
+    assert isinstance(call_args, ChunkGenerated)
+    assert call_args.command_id == command_id
+    assert isinstance(call_args.chunk, TokenChunk)
+    assert call_args.chunk.finish_reason == "error"
+    assert call_args.chunk.error_message == "test error"
+
+
+def test_send_error_chunk_on_exception_skips_non_rank_zero() -> None:
+    event_sender = MagicMock()
+    command_id = CommandId()
+
+    with send_error_chunk_on_exception(
+        event_sender, command_id, MODEL_A_ID, device_rank=1
+    ):
+        raise ValueError("test error")
+
+    event_sender.send.assert_not_called()

src/exo/worker/tests/unittests/test_runner/test_event_ordering.py

@@ -5,7 +5,6 @@ from typing import Callable
 import pytest
 
 import exo.worker.runner.runner as mlx_runner
-from exo.shared.types.api import ChatCompletionMessage
 from exo.shared.types.chunks import TokenChunk
 from exo.shared.types.events import (
     ChunkGenerated,
@@ -14,9 +13,9 @@ from exo.shared.types.events import (
     TaskAcknowledged,
     TaskStatusUpdated,
 )
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.tasks import (
     ChatCompletion,
-    ChatCompletionTaskParams,
     ConnectToGroup,
     LoadModel,
     Shutdown,
@@ -85,11 +84,11 @@ SHUTDOWN_TASK = Shutdown(
     runner_id=RUNNER_1_ID,
 )
 
-CHAT_PARAMS = ChatCompletionTaskParams(
+CHAT_PARAMS = ResponsesRequest(
     model=str(MODEL_A_ID),
-    messages=[ChatCompletionMessage(role="user", content="hello")],
+    input="hello",
     stream=True,
-    max_tokens=4,
+    max_output_tokens=4,
     temperature=0.0,
 )
 

tests/headless_runner.py

@@ -13,10 +13,10 @@ from pydantic import BaseModel
 
 from exo.shared.logging import InterceptLogger, logger_setup
 from exo.shared.models.model_cards import MODEL_CARDS, ModelId
-from exo.shared.types.api import ChatCompletionMessage, ChatCompletionTaskParams
 from exo.shared.types.commands import CommandId
 from exo.shared.types.common import Host, NodeId
 from exo.shared.types.events import Event
+from exo.shared.types.openai_responses import ResponsesRequest
 from exo.shared.types.tasks import (
     ChatCompletion,
     ConnectToGroup,
@@ -169,16 +169,10 @@ async def execute_test(test: Tests, instance: Instance, hn: str):
             send.send(StartWarmup(instance_id=iid))
             send.send(
                 ChatCompletion(
-                    task_params=ChatCompletionTaskParams(
+                    task_params=ResponsesRequest(
                         model=test.model_id,
-                        messages=[
-                            ChatCompletionMessage(
-                                role="system", content="You are a helpful assistant"
-                            ),
-                            ChatCompletionMessage(
-                                role="user", content="What is the capital of France?"
-                            ),
-                        ],
+                        instructions="You are a helpful assistant",
+                        input="What is the capital of France?",
                     ),
                     command_id=CommandId("yo"),
                     instance_id=iid,