fix(tui): buffer streaming messages by runId to prevent render ordering issues

Fixes #1172

- Add per-runId message buffering in ChatLog
- Separate thinking stream from content stream handling
- Ensure proper sequencing (thinking always before content)
- Model-agnostic: works with or without thinking tokens
This commit is contained in:
Aaron
2026-01-19 10:40:51 +10:00
committed by Peter Steinberger
parent 41b696fa83
commit 476087f879
3 changed files with 289 additions and 45 deletions

View File

@@ -4,19 +4,30 @@ import { AssistantMessageComponent } from "./assistant-message.js";
import { ToolExecutionComponent } from "./tool-execution.js";
import { UserMessageComponent } from "./user-message.js";
/**
* Per-run streaming state - isolates each run's thinking and content streams.
* This enables proper sequencing regardless of network arrival order.
*/
interface StreamingRunState {
component: AssistantMessageComponent;
thinkingText: string;
contentText: string;
showThinking: boolean;
}
export class ChatLog extends Container {
private toolById = new Map<string, ToolExecutionComponent>();
private streamingAssistant: AssistantMessageComponent | null = null;
private streamingRunId: string | null = null;
private streamingText: string | null = null;
// FIXED: Replace single streaming fields with per-runId Map for proper isolation
private streamingRuns = new Map<string, StreamingRunState>();
// Keep reference to most recent run for backward compatibility
private lastStreamingRunId: string | null = null;
private toolsExpanded = false;
clearAll() {
this.clear();
this.toolById.clear();
this.streamingAssistant = null;
this.streamingRunId = null;
this.streamingText = null;
this.streamingRuns.clear();
this.lastStreamingRunId = null;
}
addSystem(text: string) {
@@ -28,48 +39,141 @@ export class ChatLog extends Container {
this.addChild(new UserMessageComponent(text));
}
/**
* Get or create streaming state for a specific runId.
*/
private getOrCreateRunState(runId: string, showThinking: boolean): StreamingRunState {
let state = this.streamingRuns.get(runId);
if (!state) {
const component = new AssistantMessageComponent("");
this.addChild(component);
state = {
component,
thinkingText: "",
contentText: "",
showThinking,
};
this.streamingRuns.set(runId, state);
this.lastStreamingRunId = runId;
}
return state;
}
/**
* Compose the final display text from thinking + content.
* FIXED: Ensures thinking always appears before content regardless of arrival order.
*/
private composeDisplayText(state: StreamingRunState): string {
const parts: string[] = [];
// Thinking comes first (if enabled and present)
if (state.showThinking && state.thinkingText.trim()) {
parts.push(`[thinking]\n${state.thinkingText}`);
}
// Content comes after thinking
if (state.contentText.trim()) {
parts.push(state.contentText);
}
return parts.join("\n\n").trim() || "";
}
startAssistant(text: string, runId?: string) {
const component = new AssistantMessageComponent(text);
this.streamingAssistant = component;
this.streamingRunId = runId ?? null;
this.streamingText = text;
if (runId) {
// Create proper streaming state for tracked runs
this.streamingRuns.set(runId, {
component,
thinkingText: "",
contentText: text,
showThinking: false,
});
this.lastStreamingRunId = runId;
}
this.addChild(component);
return component;
}
updateAssistant(text: string, runId?: string) {
if (
!this.streamingAssistant ||
(runId && this.streamingRunId && runId !== this.streamingRunId)
) {
this.startAssistant(text, runId);
return;
/**
* Update the assistant message with new streaming content.
* FIXED: Now properly isolates by runId and separates thinking/content.
*/
updateAssistant(
text: string,
runId?: string,
options?: {
thinkingText?: string;
contentText?: string;
showThinking?: boolean;
},
) {
const effectiveRunId = runId ?? "default";
const showThinking = options?.showThinking ?? false;
const state = this.getOrCreateRunState(effectiveRunId, showThinking);
// Update thinking and/or content separately if provided
if (options?.thinkingText !== undefined) {
state.thinkingText = options.thinkingText;
}
this.streamingText = text;
this.streamingAssistant.setText(text);
if (options?.contentText !== undefined) {
state.contentText = options.contentText;
}
// If only raw text provided (backward compatibility), use as content
if (options?.thinkingText === undefined && options?.contentText === undefined) {
state.contentText = text;
}
state.showThinking = showThinking;
// Recompose and render with guaranteed ordering
const displayText = this.composeDisplayText(state);
state.component.setText(displayText);
}
getStreamingText(runId?: string) {
if (!this.streamingAssistant) return null;
if (runId && this.streamingRunId && runId !== this.streamingRunId) {
return null;
}
return this.streamingText;
const effectiveRunId = runId ?? this.lastStreamingRunId;
if (!effectiveRunId) return null;
const state = this.streamingRuns.get(effectiveRunId);
if (!state) return null;
return this.composeDisplayText(state);
}
/**
* Get the raw streaming state (for diagnostics).
*/
getStreamingState(runId: string): { thinking: string; content: string } | null {
const state = this.streamingRuns.get(runId);
if (!state) return null;
return {
thinking: state.thinkingText,
content: state.contentText,
};
}
finalizeAssistant(text: string, runId?: string) {
if (
this.streamingAssistant &&
(!runId || runId === this.streamingRunId || text === this.streamingText)
) {
this.streamingText = text;
this.streamingAssistant.setText(text);
const effectiveRunId = runId ?? this.lastStreamingRunId;
const state = effectiveRunId ? this.streamingRuns.get(effectiveRunId) : null;
if (state) {
// Use the final text, or compose from existing state if final is empty
const finalText = text.trim() || this.composeDisplayText(state);
state.component.setText(finalText);
} else {
// No existing state - create a new component with final text
this.startAssistant(text, runId);
}
this.streamingAssistant = null;
this.streamingRunId = null;
this.streamingText = null;
// Clean up the streaming state for this run
if (effectiveRunId) {
this.streamingRuns.delete(effectiveRunId);
if (this.lastStreamingRunId === effectiveRunId) {
this.lastStreamingRunId = null;
}
}
}
startTool(toolCallId: string, toolName: string, args: unknown) {

View File

@@ -1,6 +1,12 @@
import type { TUI } from "@mariozechner/pi-tui";
import type { ChatLog } from "./components/chat-log.js";
import { asString, extractTextFromMessage, resolveFinalAssistantText } from "./tui-formatters.js";
import {
asString,
extractTextFromMessage,
extractThinkingFromMessage,
extractContentFromMessage,
resolveFinalAssistantText,
} from "./tui-formatters.js";
import type { AgentEvent, ChatEvent, TuiStateAccess } from "./tui-types.js";
type EventHandlerContext = {
@@ -11,12 +17,25 @@ type EventHandlerContext = {
refreshSessionInfo?: () => Promise<void>;
};
/**
* Per-run stream buffer for tracking thinking/content separately.
* Enables proper sequencing regardless of network arrival order.
*/
interface RunStreamBuffer {
thinkingText: string;
contentText: string;
lastUpdateMs: number;
}
export function createEventHandlers(context: EventHandlerContext) {
const { chatLog, tui, state, setActivityStatus, refreshSessionInfo } = context;
const finalizedRuns = new Map<string, number>();
// FIXED: Per-run stream buffers for proper isolation
const runBuffers = new Map<string, RunStreamBuffer>();
const noteFinalizedRun = (runId: string) => {
finalizedRuns.set(runId, Date.now());
runBuffers.delete(runId); // Clean up buffer
if (finalizedRuns.size <= 200) return;
const keepUntil = Date.now() - 10 * 60 * 1000;
for (const [key, ts] of finalizedRuns) {
@@ -31,6 +50,22 @@ export function createEventHandlers(context: EventHandlerContext) {
}
};
/**
* Get or create a stream buffer for a specific runId.
*/
const getOrCreateBuffer = (runId: string): RunStreamBuffer => {
let buffer = runBuffers.get(runId);
if (!buffer) {
buffer = {
thinkingText: "",
contentText: "",
lastUpdateMs: Date.now(),
};
runBuffers.set(runId, buffer);
}
return buffer;
};
const handleChatEvent = (payload: unknown) => {
if (!payload || typeof payload !== "object") return;
const evt = payload as ChatEvent;
@@ -40,11 +75,33 @@ export function createEventHandlers(context: EventHandlerContext) {
if (evt.state === "final") return;
}
if (evt.state === "delta") {
const text = extractTextFromMessage(evt.message, {
includeThinking: state.showThinking,
const buffer = getOrCreateBuffer(evt.runId);
// FIXED: Extract thinking and content SEPARATELY for proper sequencing
// This is model-agnostic: models without thinking blocks just return empty string
const thinkingText = extractThinkingFromMessage(evt.message);
const contentText = extractContentFromMessage(evt.message);
// Update buffer with new content
// In streaming, we typically receive the full accumulated text each time
if (thinkingText) {
buffer.thinkingText = thinkingText;
}
if (contentText) {
buffer.contentText = contentText;
}
buffer.lastUpdateMs = Date.now();
// Skip render if both are empty
if (!buffer.thinkingText && !buffer.contentText) return;
// FIXED: Pass separated streams to ChatLog for proper sequencing
chatLog.updateAssistant("", evt.runId, {
thinkingText: buffer.thinkingText,
contentText: buffer.contentText,
showThinking: state.showThinking,
});
if (!text) return;
chatLog.updateAssistant(text, evt.runId);
setActivityStatus("streaming");
}
if (evt.state === "final") {
@@ -54,11 +111,23 @@ export function createEventHandlers(context: EventHandlerContext) {
? ((evt.message as Record<string, unknown>).stopReason as string)
: ""
: "";
const text = extractTextFromMessage(evt.message, {
includeThinking: state.showThinking,
});
// FIXED: Extract final content with proper thinking handling
const thinkingText = extractThinkingFromMessage(evt.message);
const contentText = extractContentFromMessage(evt.message);
// Compose final text with proper ordering (thinking before content)
const parts: string[] = [];
if (state.showThinking && thinkingText.trim()) {
parts.push(`[thinking]\n${thinkingText}`);
}
if (contentText.trim()) {
parts.push(contentText);
}
const finalComposed = parts.join("\n\n").trim();
const finalText = resolveFinalAssistantText({
finalText: text,
finalText: finalComposed,
streamedText: chatLog.getStreamingText(evt.runId),
});
chatLog.finalizeAssistant(finalText, evt.runId);
@@ -70,12 +139,14 @@ export function createEventHandlers(context: EventHandlerContext) {
}
if (evt.state === "aborted") {
chatLog.addSystem("run aborted");
runBuffers.delete(evt.runId);
state.activeChatRunId = null;
setActivityStatus("aborted");
void refreshSessionInfo?.();
}
if (evt.state === "error") {
chatLog.addSystem(`run error: ${evt.errorMessage ?? "unknown"}`);
runBuffers.delete(evt.runId);
state.activeChatRunId = null;
setActivityStatus("error");
void refreshSessionInfo?.();

View File

@@ -12,25 +12,94 @@ export function resolveFinalAssistantText(params: {
return "(no output)";
}
/**
* Extract ONLY thinking blocks from message content.
* Model-agnostic: returns empty string if no thinking blocks exist.
*/
export function extractThinkingFromMessage(message: unknown): string {
if (!message || typeof message !== "object") return "";
const record = message as Record<string, unknown>;
const content = record.content;
if (typeof content === "string") return "";
if (!Array.isArray(content)) return "";
const parts: string[] = [];
for (const block of content) {
if (!block || typeof block !== "object") continue;
const rec = block as Record<string, unknown>;
if (rec.type === "thinking" && typeof rec.thinking === "string") {
parts.push(rec.thinking);
}
}
return parts.join("\n").trim();
}
/**
* Extract ONLY text content blocks from message (excludes thinking).
* Model-agnostic: works for any model with text content blocks.
*/
export function extractContentFromMessage(message: unknown): string {
if (!message || typeof message !== "object") return "";
const record = message as Record<string, unknown>;
const content = record.content;
if (typeof content === "string") return content.trim();
if (!Array.isArray(content)) return "";
const parts: string[] = [];
for (const block of content) {
if (!block || typeof block !== "object") continue;
const rec = block as Record<string, unknown>;
if (rec.type === "text" && typeof rec.text === "string") {
parts.push(rec.text);
}
}
// If no text blocks found, check for error
if (parts.length === 0) {
const stopReason = typeof record.stopReason === "string" ? record.stopReason : "";
if (stopReason === "error") {
const errorMessage = typeof record.errorMessage === "string" ? record.errorMessage : "";
return formatRawAssistantErrorForUi(errorMessage);
}
}
return parts.join("\n").trim();
}
function extractTextBlocks(content: unknown, opts?: { includeThinking?: boolean }): string {
if (typeof content === "string") return content.trim();
if (!Array.isArray(content)) return "";
const parts: string[] = [];
// FIXED: Separate collection to ensure proper ordering (thinking before text)
const thinkingParts: string[] = [];
const textParts: string[] = [];
for (const block of content) {
if (!block || typeof block !== "object") continue;
const record = block as Record<string, unknown>;
if (record.type === "text" && typeof record.text === "string") {
parts.push(record.text);
textParts.push(record.text);
}
if (
opts?.includeThinking &&
record.type === "thinking" &&
typeof record.thinking === "string"
) {
parts.push(`[thinking]\n${record.thinking}`);
thinkingParts.push(`[thinking]\n${record.thinking}`);
}
}
return parts.join("\n").trim();
// FIXED: Always put thinking BEFORE text content for consistent ordering
const parts: string[] = [];
if (thinkingParts.length > 0) {
parts.push(...thinkingParts);
}
if (textParts.length > 0) {
parts.push(...textParts);
}
return parts.join("\n\n").trim();
}
export function extractTextFromMessage(