feat(chat): stream partial tool output during live agent response

Live tool calls that produce output while running (e.g. streaming bash
stdout from Codex) now render their partial output in real time instead
of appearing blank until the tool completes. The tool card keeps its
running visual state — spinner and 24KB tail truncation — while chunks
arrive, and transitions to the completed state once the final status
lands. A WeakMap keyed on the ACP reducer's chunks-array identity
memoizes the joined output so repeated renders don't re-run O(n) string
concatenation.

src/components/message/message-list-view.tsx

@@ -292,13 +292,22 @@ export function MessageListView({
   const { threadItems, nonStreamingAdapted } = useMemo(() => {
     const allTurns = timelineTurns.map((item) => item.turn)
     const streamingIndices = new Set<number>()
+    const inProgressToolCallIdsByIndex = new Map<number, Set<string>>()
     timelineTurns.forEach((item, i) => {
-      if (item.phase === "streaming") streamingIndices.add(i)
+      if (item.phase === "streaming") {
+        streamingIndices.add(i)
+        if (item.inProgressToolCallIds && item.inProgressToolCallIds.size > 0) {
+          inProgressToolCallIdsByIndex.set(i, item.inProgressToolCallIds)
+        }
+      }
     })
     const allAdapted = adaptMessageTurns(
       allTurns,
       adapterText,
-      streamingIndices.size > 0 ? streamingIndices : undefined
+      streamingIndices.size > 0 ? streamingIndices : undefined,
+      inProgressToolCallIdsByIndex.size > 0
+        ? inProgressToolCallIdsByIndex
+        : undefined
     )
 
     // Collect non-streaming adapted messages for plan extraction

src/contexts/conversation-runtime-context.tsx

@@ -27,6 +27,9 @@ export interface ConversationTimelineTurn {
   key: string
   turn: MessageTurn
   phase: ConversationTimelinePhase
+  // Tool call IDs whose results are still streaming (only set for streaming-phase items).
+  // The adapter uses this to keep the tool in "running" state while exposing partial output.
+  inProgressToolCallIds?: Set<string>
 }
 
 export interface ConversationRuntimeSession {
@@ -161,16 +164,38 @@ function formatLivePlanEntries(
   return `Plan updated:\n${lines.join("\n")}`
 }
 
+interface BuiltStreamingTurns {
+  turns: MessageTurn[]
+  inProgressToolCallIds: Set<string>
+}
+
+// Cache joined chunk output keyed by chunks-array identity. The ACP reducer
+// creates a new chunks array only when streaming output actually changes, so
+// a WeakMap keyed on the array reference lets repeated renders reuse the
+// joined string without re-running O(n) concatenation.
+const joinedOutputCache = new WeakMap<readonly string[], string>()
+
+function getJoinedChunks(chunks: readonly string[]): string {
+  if (chunks.length === 0) return ""
+  if (chunks.length === 1) return chunks[0]
+  const cached = joinedOutputCache.get(chunks)
+  if (cached !== undefined) return cached
+  const joined = chunks.join("")
+  joinedOutputCache.set(chunks, joined)
+  return joined
+}
+
 function buildStreamingTurnsFromLiveMessage(
   conversationId: number,
   liveMessage: LiveMessage
-): MessageTurn[] {
+): BuiltStreamingTurns {
   // Split streaming content into multiple turns matching the historical
   // pattern: each "round" (text/thinking + tool calls + tool results) is a
   // separate turn. A new turn starts when a text/thinking/plan block appears
   // after completed tool calls in the current group.
   const groups: MessageTurn["blocks"][] = [[]]
   let currentGroupHasCompletedTool = false
+  const inProgressToolCallIds = new Set<string>()
 
   for (const block of liveMessage.content) {
     const isContentBlock =
@@ -217,17 +242,35 @@ function buildStreamingTurnsFromLiveMessage(
         })
         const isFinalState =
           block.info.status === "completed" || block.info.status === "failed"
+        // Output precedence: raw_output_chunks (terminal polling / SDK
+        // raw_output field) wins over content. Some agents stream bash output
+        // via raw_output with raw_output_append, others via content-only
+        // tool_call_update notifications — we support both.
+        const resolvedOutput =
+          block.info.raw_output_chunks.length > 0
+            ? getJoinedChunks(block.info.raw_output_chunks)
+            : block.info.content
         if (isFinalState) {
           currentBlocks.push({
             type: "tool_result",
             tool_use_id: block.info.tool_call_id,
-            output_preview:
-              block.info.raw_output_chunks.length > 0
-                ? block.info.raw_output_chunks.join("")
-                : block.info.content,
+            output_preview: resolvedOutput,
             is_error: block.info.status === "failed",
           })
           currentGroupHasCompletedTool = true
+        } else if (resolvedOutput) {
+          // In-progress tool that already produced partial output. Emit the
+          // running result so the renderer can display live output, and flag
+          // the tool_call so the adapter keeps state="input-available" (the
+          // spinner/running visual and 24KB tail truncation both depend on
+          // this state).
+          currentBlocks.push({
+            type: "tool_result",
+            tool_use_id: block.info.tool_call_id,
+            output_preview: resolvedOutput,
+            is_error: false,
+          })
+          inProgressToolCallIds.add(block.info.tool_call_id)
         }
         break
       }
@@ -235,7 +278,7 @@ function buildStreamingTurnsFromLiveMessage(
   }
 
   const timestamp = new Date(liveMessage.startedAt).toISOString()
-  return groups
+  const turns = groups
     .filter((blocks) => blocks.length > 0)
     .map((blocks, i) => ({
       id:
@@ -246,6 +289,8 @@ function buildStreamingTurnsFromLiveMessage(
       blocks,
       timestamp,
     }))
+
+  return { turns, inProgressToolCallIds }
 }
 
 function upsertExternalIdIndex(
@@ -345,7 +390,7 @@ function reducer(
         ? buildStreamingTurnsFromLiveMessage(
             current.conversationId,
             current.liveMessage
-          )
+          ).turns
         : []
 
       // Promote: optimisticTurns + streamingTurns → localTurns
@@ -636,18 +681,21 @@ export function ConversationRuntimeProvider({
 
       // Phase 4: Streaming turns (live agent response, split into rounds)
       const streamingMessage = session.liveMessage
-      const streamingTurns = streamingMessage
+      const built = streamingMessage
         ? buildStreamingTurnsFromLiveMessage(conversationId, streamingMessage)
-        : []
+        : null
 
       const result = [...persisted, ...local, ...optimistic]
 
-      for (const [i, turn] of streamingTurns.entries()) {
-        result.push({
-          key: `streaming-${conversationId}-${streamingMessage?.id ?? "unknown"}-${i}`,
-          turn,
-          phase: "streaming",
-        })
+      if (built) {
+        for (const [i, turn] of built.turns.entries()) {
+          result.push({
+            key: `streaming-${conversationId}-${streamingMessage?.id ?? "unknown"}-${i}`,
+            turn,
+            phase: "streaming",
+            inProgressToolCallIds: built.inProgressToolCallIds,
+          })
+        }
       }
 
       return result

src/lib/adapters/ai-elements-adapter.ts

@@ -600,11 +600,19 @@ function buildToolResultMap(
 /**
  * Transform a MessageTurn (from backend) to AdaptedMessage format.
  * Same correlation logic as adaptUnifiedMessage but operates on turn.blocks.
+ *
+ * `inProgressToolCallIds` lets streaming consumers expose partial tool output
+ * (e.g. terminal stdout streamed during execution) without flipping the tool
+ * into a "completed" visual state. When a tool_use's id is in this set, the
+ * adapter emits state="input-available" with the partial output attached, so
+ * the renderer can keep showing the running spinner while the live output
+ * streams in.
  */
 export function adaptMessageTurn(
   turn: MessageTurn,
   text: AdapterMessageText,
-  isStreaming: boolean = false
+  isStreaming: boolean = false,
+  inProgressToolCallIds?: Set<string>
 ): AdaptedMessage {
   const adaptedContent: AdaptedContentPart[] = []
   const resultMap = buildToolResultMap(turn.blocks)
@@ -635,6 +643,9 @@ export function adaptMessageTurn(
         ? resultMap.get(block.tool_use_id)
         : undefined
 
+      const isToolStillRunning =
+        !!block.tool_use_id && !!inProgressToolCallIds?.has(block.tool_use_id)
+
       if (matchedResult) {
         matchedResultIds.add(block.tool_use_id!)
         adaptedContent.push({
@@ -642,7 +653,11 @@ export function adaptMessageTurn(
           toolCallId,
           toolName: block.tool_name,
           input: block.input_preview,
-          state: matchedResult.is_error ? "output-error" : "output-available",
+          state: isToolStillRunning
+            ? "input-available"
+            : matchedResult.is_error
+              ? "output-error"
+              : "output-available",
           output: matchedResult.output_preview,
           errorText: matchedResult.is_error
             ? matchedResult.output_preview || undefined
@@ -736,13 +751,24 @@ export function adaptMessageTurn(
 /**
  * Transform all turns in a conversation to AdaptedMessage[].
  * Internally computes completedToolIds so callers don't need to.
+ *
+ * `inProgressToolCallIdsByIndex` carries the set of tool_call_ids that are
+ * still streaming for each streaming-phase turn (keyed by turn index). The
+ * adapter forwards this to adaptMessageTurn so partial output renders without
+ * flipping the tool out of the running visual state.
  */
 export function adaptMessageTurns(
   turns: MessageTurn[],
   text: AdapterMessageText,
-  streamingIndices?: Set<number>
+  streamingIndices?: Set<number>,
+  inProgressToolCallIdsByIndex?: Map<number, Set<string>>
 ): AdaptedMessage[] {
   return turns.map((turn, i) =>
-    adaptMessageTurn(turn, text, streamingIndices?.has(i) ?? false)
+    adaptMessageTurn(
+      turn,
+      text,
+      streamingIndices?.has(i) ?? false,
+      inProgressToolCallIdsByIndex?.get(i)
+    )
   )
 }