Skip to content

Build your personal knowledge base with Trilium Notes

feat(llm): show each thought in stream order inside the reply

Elian Doran669d9190bfa8
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.css b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.css
index 787362f3ff..0566945568 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.css
+++ b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.css
@@ -204,20 +204,52 @@ body.background-effects.theme-supports-background-effects .llm-chat-message-user
width: 100%;
max-width: 100%;
border-style: dashed;
- margin-right: 0;
- margin: 0;
}
-.llm-chat-thinking-card .expandable-section-summary {
+.llm-chat-thinking-card .expandable-section-summary,
+.llm-chat-thinking-header {
color: var(--muted-text-color);
}
+.llm-chat-thinking-header {
+ display: flex;
+ align-items: center;
+ gap: 0.4rem;
+ padding: 0.5rem 0.75rem;
+ font-weight: 500;
+
+ > .bx {
+ font-size: 1rem;
+ }
+}
+
+.llm-chat-thinking-title {
+ min-width: 0;
+ overflow: hidden;
+ text-overflow: ellipsis;
+ white-space: nowrap;
+}
+
.llm-chat-thinking-content {
padding: 0.5rem 0.75rem;
font-size: 0.9rem;
color: var(--muted-text-color);
}
+/* Aligning the content to the end clips the overflow at the top, keeping the newest lines in view. */
+.llm-chat-thinking-live .llm-chat-thinking-content {
+ display: flex;
+ flex-direction: column;
+ justify-content: flex-end;
+ max-height: calc(4 * 1.6em + 1rem);
+ padding-top: 0;
+ overflow: hidden;
+
+ > * {
+ flex-shrink: 0;
+ }
+}
+
/* Markdown styles */
.llm-chat-markdown {
line-height: 1.6;
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.spec.tsx b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.spec.tsx
index 84cb40cdc7..c8b90a0bd4 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.spec.tsx
+++ b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.spec.tsx
@@ -60,20 +60,57 @@ describe("ChatMessage footer", () => {
});
describe("ChatMessage thinking", () => {
- it("renders the thought process as Markdown, as Codex writes its reasoning summaries", () => {
- host = document.body.appendChild(document.createElement("div"));
+ function renderMessage(content: StoredMessage["content"], options: { type?: StoredMessage["type"]; isStreaming?: boolean } = {}) {
+ host ??= document.body.appendChild(document.createElement("div"));
const target = host;
- const message: StoredMessage = {
- id: "t1", role: "assistant", type: "thinking", createdAt: "2026-01-01T00:00:00.000Z",
- content: "**Retrieving PC hostname with command**\n\nI'll read `/etc/hostname`."
- };
- act(() => render(<ChatMessage message={message} />, target));
+ const message: StoredMessage = { id: "m1", role: "assistant", type: options.type, createdAt: "2026-01-01T00:00:00.000Z", content };
+ act(() => render(<ChatMessage message={message} isStreaming={options.isStreaming} />, target));
+ return target;
+ }
+
+ it("folds each finished thought to its title, in stream order between the tool calls", () => {
+ const target = renderMessage([
+ { type: "thinking", content: "**Retrieving PC hostname with command**\n\nI'll read `/etc/hostname`." },
+ { type: "tool_call", toolCall: { id: "c1", toolName: "shell", input: {}, result: "pc" } },
+ { type: "thinking", content: "The file holds the name." },
+ { type: "text", content: "Your PC is called pc." }
+ ]);
+
+ const blocks = [...(target.querySelector(".llm-chat-message-content")?.children ?? [])];
+ expect(blocks.map(el => el.classList.contains("llm-chat-thinking-card"))).toEqual([true, false, true, false]);
+
+ const [titled, untitled] = target.querySelectorAll(".llm-chat-thinking-card");
+ const details = titled.querySelector("details");
+ expect(details).not.toBeNull();
+ expect(details?.open).toBe(false);
+ expect(titled.querySelector(".expandable-section-label")?.textContent).toBe("Retrieving PC hostname with command");
+ const body = titled.querySelector(".llm-chat-thinking-content .markdown-stub")?.textContent;
+ expect(body).toContain("<code>/etc/hostname</code>");
+ expect(body).not.toContain("Retrieving PC hostname");
+ expect(untitled.querySelector(".expandable-section-label")?.textContent).toBe("llm_chat.thought_process");
+ });
+
+ it("shows the thought being generated under a spinner and its latest title, and folds it once the turn moves on", () => {
+ const thought = { type: "thinking" as const, content: "**Reading the hostname**\n\nFirst.\n\n**Checking the network**\n\nSecond." };
+ let target = renderMessage([thought], { isStreaming: true });
+ let card = target.querySelector(".llm-chat-thinking-card");
+ expect(card?.classList.contains("llm-chat-thinking-live")).toBe(true);
+ expect(card?.querySelector("details")).toBeNull();
+ expect(card?.querySelector(".bx-spin")).not.toBeNull();
+ expect(card?.querySelector(".llm-chat-thinking-title")?.textContent).toBe("Checking the network");
+ expect(card?.querySelector(".llm-chat-thinking-content .markdown-stub")?.textContent).toContain("Second.");
+
+ target = renderMessage([thought, { type: "tool_call", toolCall: { id: "c1", toolName: "shell", input: {} } }], { isStreaming: true });
+ card = target.querySelector(".llm-chat-thinking-card");
+ expect(card?.classList.contains("llm-chat-thinking-live")).toBe(false);
+ expect(card?.querySelector("details")?.open).toBe(false);
+ expect(card?.querySelector(".expandable-section-label")?.textContent).toBe("Reading the hostname");
+ });
- const thinking = target.querySelector(".llm-chat-thinking-content");
- expect(thinking?.textContent).toContain("Retrieving PC hostname with command");
- const content = thinking?.querySelector(".markdown-stub");
- expect(content).not.toBeNull();
- expect(content?.textContent).toContain("<strong>Retrieving PC hostname with command</strong>");
- expect(content?.textContent).toContain("<code>/etc/hostname</code>");
+ it("renders a stored thinking message from before thoughts moved into the reply", () => {
+ const target = renderMessage("**Retrieving PC hostname with command**\n\nI'll read `/etc/hostname`.", { type: "thinking" });
+ const card = target.querySelector(".llm-chat-thinking-card");
+ expect(card?.querySelector(".expandable-section-label")?.textContent).toBe("Retrieving PC hostname with command");
+ expect(card?.querySelector(".llm-chat-thinking-content .markdown-stub")?.textContent).toContain("<code>/etc/hostname</code>");
});
});
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.tsx b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.tsx
index 79a9e418b2..75c207f834 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.tsx
+++ b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessage.tsx
@@ -9,12 +9,13 @@ import { t } from "../../../services/i18n.js";
import utils from "../../../services/utils.js";
import { ExtendedAdmonition } from "../../react/Admonition.js";
import Button from "../../react/Button.js";
+import LoadingSpinner from "../../react/LoadingSpinner.js";
import { ReadOnlyTextContent } from "../text/ReadOnlyText.js";
import { formatErrorDetails } from "./chat_error.js";
import { renderMarkdown } from "./chat_markdown.js";
import { renderQuoteSourceLinks } from "./chat_quote.js";
import { ExpandableCard, ExpandableSection } from "./ExpandableCard.js";
-import { type ContentBlock, type FileBlock, getMessageText, type ImageBlock, type StoredMessage, type TextBlock, type TextFileBlock, type ToolCallBlock } from "./llm_chat_types.js";
+import { type ContentBlock, type FileBlock, getMessageText, type ImageBlock, type StoredMessage, type TextBlock, type TextFileBlock, type ThinkingBlock, type ToolCallBlock } from "./llm_chat_types.js";
import { shortModelName } from "./model_name.js";
import { SafeImage } from "./retry_image.js";
import ToolCallCard from "./ToolCallCard.js";
@@ -54,6 +55,7 @@ interface Props {
type ContentGroup =
| { type: "text"; block: TextBlock; index: number }
+ | { type: "thinking"; block: ThinkingBlock; index: number }
| { type: "tool_calls"; blocks: ToolCallBlock[]; index: number }
| { type: "image"; block: ImageBlock; index: number }
| { type: "file"; block: FileBlock; index: number }
@@ -118,6 +120,38 @@ function CitationsSection({ citations }: { citations: LlmCitation[] }) {
);
}
+/**
+ * One stretch of the model's reasoning. A finished one folds to a line named by its leading
+ * `**Title**`, the shape of Codex's reasoning summaries. The one being generated stays open,
+ * clamped to its last lines under a spinner and its latest title.
+ */
+function ThinkingCard({ content, isLive }: { content: string; isLive?: boolean }) {
+ if (isLive) {
+ return (
+ <ExpandableCard className="llm-chat-thinking-card llm-chat-thinking-live">
+ <div className="llm-chat-thinking-header">
+ <LoadingSpinner />
+ <span className="llm-chat-thinking-title">{latestThinkingTitle(content) ?? t("llm_chat.thinking")}</span>
+ </div>
+ <div className="llm-chat-thinking-content">
+ <TextBlockContent content={content} />
+ </div>
+ </ExpandableCard>
+ );
+ }
+
+ const { title, body } = splitThinkingTitle(content);
+ return (
+ <ExpandableCard className="llm-chat-thinking-card">
+ <ExpandableSection icon="bx bx-brain" label={title ?? t("llm_chat.thought_process")}>
+ <div className="llm-chat-thinking-content">
+ <TextBlockContent content={body} />
+ </div>
+ </ExpandableSection>
+ </ExpandableCard>
+ );
+}
+
function ChatMessage({ message, isStreaming, onRetry }: Props) {
const isError = message.type === "error";
const isThinking = message.type === "thinking";
@@ -144,17 +178,11 @@ function ChatMessage({ message, isStreaming, onRetry }: Props) {
isThinking && "llm-chat-message-thinking"
].filter(Boolean).join(" ");
- // Render thinking messages in a collapsible card
+ // Chats saved before reasoning moved into the reply hold it as a message of its own.
if (isThinking) {
return (
<div className="llm-chat-message-wrapper llm-chat-message-wrapper-assistant">
- <ExpandableCard className="llm-chat-thinking-card">
- <ExpandableSection icon="bx bx-brain" label={t("llm_chat.thought_process")}>
- <div className="llm-chat-thinking-content">
- <TextBlockContent content={textContent} isStreaming={isStreaming} />
- </div>
- </ExpandableSection>
- </ExpandableCard>
+ <ThinkingCard content={textContent} />
</div>
);
}
@@ -260,6 +288,8 @@ function groupContentBlocks(blocks: ContentBlock[]): ContentGroup[] {
} else {
groups.push({ type: "tool_calls", blocks: [block], index: i });
}
+ } else if (block.type === "thinking") {
+ groups.push({ type: "thinking", block, index: i });
} else if (block.type === "image") {
groups.push({ type: "image", block, index: i });
} else if (block.type === "file") {
@@ -276,8 +306,8 @@ function groupContentBlocks(blocks: ContentBlock[]): ContentGroup[] {
function renderContentBlocks(blocks: ContentBlock[], isStreaming?: boolean) {
return groupContentBlocks(blocks).map((group) => {
+ const isLastBlock = group.index === blocks.length - 1;
if (group.type === "text") {
- const isLastBlock = group.index === blocks.length - 1;
return (
<div key={group.index}>
<TextBlockContent content={group.block.content} isStreaming={isStreaming && isLastBlock} />
@@ -285,6 +315,10 @@ function renderContentBlocks(blocks: ContentBlock[], isStreaming?: boolean) {
);
}
+ if (group.type === "thinking") {
+ return <ThinkingCard key={group.index} content={group.block.content} isLive={isStreaming && isLastBlock} />;
+ }
+
if (group.type === "image") {
return (
<a
@@ -320,3 +354,17 @@ function renderContentBlocks(blocks: ContentBlock[], isStreaming?: boolean) {
return <ToolCallCard key={group.index} toolCalls={group.blocks.map((b) => b.toolCall)} />;
});
}
+
+/** Matches a line that is only bold text, the title Codex puts on each reasoning summary. */
+const THINKING_TITLE_LINE = /^\*\*([^*\n]+)\*\*[ \t]*$/gm;
+
+/** Split a thought's leading `**Title**` line from the rest, so the title can name the folded card. */
+function splitThinkingTitle(content: string): { title?: string; body: string } {
+ const match = /^\s*\*\*([^*\n]+)\*\*[ \t]*(?:\n|$)/.exec(content);
+ return match ? { title: match[1].trim(), body: content.slice(match[0].length).trimStart() } : { body: content };
+}
+
+/** The last `**Title**` line of a thought, naming what the model is working on now. */
+function latestThinkingTitle(content: string): string | undefined {
+ return [...content.matchAll(THINKING_TITLE_LINE)].at(-1)?.[1].trim();
+}
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.spec.tsx b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.spec.tsx
index 1f33dd0e86..56b0ef15de 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.spec.tsx
+++ b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.spec.tsx
@@ -31,7 +31,6 @@ function renderStreamingTurn(streamingBlocks: ContentBlock[], streamingStatus: s
messages: [],
isStreaming: true,
streamingStatus,
- streamingThinking: "",
streamingBlocks,
pendingCitations: [],
retryLast: () => undefined,
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.tsx b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.tsx
index 4dc5846449..e79e64bd81 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.tsx
+++ b/apps/client/src/widgets/type_widgets/llm_chat/ChatMessageList.tsx
@@ -42,16 +42,8 @@ export default function ChatMessageList({ chat, emptyStateText, className }: Cha
/>
)), [messages, isStreaming, retryLast]);
- // Stable placeholder objects: rebuilt only when their streamed content advances, so
- // renders caused by anything else let memo(ChatMessage) skip the placeholders too.
- const thinkingMessage = useMemo<StoredMessage | null>(() => chat.streamingThinking ? {
- id: "streaming-thinking",
- role: "assistant",
- content: chat.streamingThinking,
- createdAt: new Date().toISOString(),
- type: "thinking"
- } : null, [chat.streamingThinking]);
-
+ // A stable placeholder object: rebuilt only when its streamed content advances, so
+ // renders caused by anything else let memo(ChatMessage) skip the placeholder too.
const streamingMessage = useMemo<StoredMessage | null>(() => chat.streamingBlocks.length > 0 ? {
id: "streaming",
role: "assistant",
@@ -67,7 +59,7 @@ export default function ChatMessageList({ chat, emptyStateText, className }: Cha
<NoItems icon="bx bx-conversation" text={emptyStateText} />
)}
{storedMessages}
- {isStreaming && !thinkingMessage && !streamingMessage && (
+ {isStreaming && !streamingMessage && (
<div className={`chat-stream-status ${chat.streamingStatus ? "" : "chat-stream-status-waiting"}`} role="status">
<LoadingSpinner />
{chat.streamingStatus
@@ -75,12 +67,6 @@ export default function ChatMessageList({ chat, emptyStateText, className }: Cha
: t("llm_chat.stream_status.waiting_for_reply")}
</div>
)}
- {isStreaming && thinkingMessage && (
- <ChatMessage
- message={thinkingMessage}
- isStreaming
- />
- )}
{isStreaming && streamingMessage && (
<ChatMessage
message={streamingMessage}
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.spec.ts b/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.spec.ts
index e993002c51..eafa0b3438 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.spec.ts
+++ b/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.spec.ts
@@ -47,10 +47,11 @@ describe("resolveAnchorIndices", () => {
});
describe("buildProseProjection", () => {
- it("concatenates prose while skipping code, math, and non-markdown text", () => {
+ it("concatenates prose while skipping code, math, non-markdown text and thoughts", () => {
const root = contentRoot(
`<p>Hello <code>skip()</code> world</p><pre>ignored block</pre><p>bye</p>`,
- `<div class="tool-card">tool noise</div>`
+ `<div class="tool-card">tool noise</div>` +
+ `<div class="llm-chat-thinking-card"><div class="llm-chat-markdown"><p>thought</p></div></div>`
);
expect(buildProseProjection(root).text).toBe("Hello worldbye");
});
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.ts b/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.ts
index a2af6baa30..6fd47396ed 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.ts
+++ b/apps/client/src/widgets/type_widgets/llm_chat/chat_highlights_anchor.ts
@@ -13,8 +13,8 @@ import type { HighlightAnchor } from "./llm_chat_types.js";
/** Only text inside rendered markdown counts as highlightable prose. */
const PROSE_CONTAINER_SELECTOR = ".llm-chat-markdown";
-/** Non-prose subtrees skipped within the markdown: code, math (async KaTeX), diagrams. */
-const NON_PROSE_SELECTOR = "pre, code, .math-tex, .katex, .ck-math-tex, .mermaid, svg";
+/** Non-prose subtrees skipped within the markdown: code, math (async KaTeX), diagrams, thoughts. */
+const NON_PROSE_SELECTOR = "pre, code, .math-tex, .katex, .ck-math-tex, .mermaid, svg, .llm-chat-thinking-card";
/** How much surrounding prose to store as disambiguating context on each side of a highlight. */
const CONTEXT_LENGTH = 24;
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.spec.ts b/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.spec.ts
index 1d5ce97d22..6cb1cd7b95 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.spec.ts
+++ b/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.spec.ts
@@ -275,6 +275,20 @@ describe("extractChatHeadings", () => {
]);
});
+ it("skips headings inside a reply's thoughts", () => {
+ const headings = extractChatHeadings([
+ userMessage("u1", "Q"),
+ assistantMessage("a1", "Answer")
+ ], containerWith(
+ messageWrapper("u1"),
+ messageWrapper("a1", `<div class="llm-chat-thinking-card">${markdownContent("<h1>Thought</h1>")}</div>${markdownContent("<h1>Real</h1>")}`)
+ ));
+ expect(headings).toMatchObject([
+ { id: "u1", level: 1 },
+ { id: "a1:0", level: 2, text: "Real" }
+ ]);
+ });
+
it("skips error and thinking messages entirely", () => {
const headings = extractChatHeadings([
userMessage("u1", "Q"),
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.ts b/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.ts
index 4499ae21f9..0ce93794dc 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.ts
+++ b/apps/client/src/widgets/type_widgets/llm_chat/chat_toc.ts
@@ -187,6 +187,7 @@ export function extractChatHeadings(messages: StoredMessage[], container: HTMLEl
const levelOffset = hasUserHeading ? 1 : 0;
let index = 0;
for (const headingEl of messageEl.querySelectorAll<HTMLHeadingElement>(REPLY_HEADING_SELECTOR)) {
+ if (headingEl.closest(".llm-chat-thinking-card")) continue;
headings.push({
id: `${message.id}:${index}`,
level: parseInt(headingEl.tagName.substring(1), 10) + levelOffset,
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/llm_chat_types.ts b/apps/client/src/widgets/type_widgets/llm_chat/llm_chat_types.ts
index 703e6b5a4e..c540e95141 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/llm_chat_types.ts
+++ b/apps/client/src/widgets/type_widgets/llm_chat/llm_chat_types.ts
@@ -21,6 +21,15 @@ export interface TextBlock {
content: string;
}
+/**
+ * A stretch of the model's reasoning, kept in stream order among the reply's other blocks.
+ * It is shown in the timeline but never sent back to the model.
+ */
+export interface ThinkingBlock {
+ type: "thinking";
+ content: string;
+}
+
/** A tool invocation block shown inline in the message timeline. */
export interface ToolCallBlock {
type: "tool_call";
@@ -71,7 +80,7 @@ export interface TextFileBlock {
}
/** An ordered content block in a chat message. */
-export type ContentBlock = TextBlock | ToolCallBlock | ImageBlock | FileBlock | TextFileBlock;
+export type ContentBlock = TextBlock | ThinkingBlock | ToolCallBlock | ImageBlock | FileBlock | TextFileBlock;
/**
* Extract the plain text from message content (works for both legacy string and block formats).
@@ -138,7 +147,10 @@ export interface StoredMessage {
content: string | ContentBlock[];
createdAt: string;
citations?: LlmCitation[];
- /** Message type for special rendering. Defaults to "message" if omitted. */
+ /**
+ * Message type for special rendering. Defaults to "message" if omitted. `"thinking"` appears
+ * only in chats saved before reasoning moved into {@link ThinkingBlock}s.
+ */
type?: MessageType;
/**
* For `type: "error"` messages, the failed provider call's context (HTTP status, URL,
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.spec.tsx b/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.spec.tsx
index 24f0767807..54751cba7a 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.spec.tsx
+++ b/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.spec.tsx
@@ -302,6 +302,53 @@ describe("useLlmChat", () => {
expect(api().getContent()).toMatchObject({ selectedModel: "mini", selectedProvider: "openai", selectedProviderId: "o_1" });
});
+ it("keeps thoughts in stream order inside the reply, and never sends them back", async () => {
+ streamChatCompletionMock.mockImplementationOnce(async (_messages, _options, cb) => {
+ cb.onThinking("**Reading** the ");
+ cb.onThinking("hostname");
+ cb.onToolUse("c1", "shell", {});
+ cb.onToolResult("c1", "shell", "pc", false);
+ cb.onThinking("Got it.");
+ cb.onChunk("Your PC is pc.");
+ cb.onDone();
+ });
+ await mountChat();
+ await act(async () => {
+ api().loadFromContent({
+ version: 1,
+ messages: [
+ { id: "m1", role: "user", content: "hi", createdAt: "2026-01-01T00:00:00.000Z" },
+ { id: "m2", role: "assistant", type: "thinking", content: "old thoughts", createdAt: "2026-01-01T00:00:01.000Z" },
+ { id: "m3", role: "assistant", content: "hello", createdAt: "2026-01-01T00:00:02.000Z" }
+ ]
+ });
+ });
+
+ for (const text of ["name?", "thanks"]) {
+ await act(async () => {
+ api().setInput(text);
+ });
+ await act(async () => {
+ await api().handleSubmit(new Event("submit"));
+ });
+ }
+
+ expect(api().messages[4]).toMatchObject({
+ role: "assistant",
+ content: [
+ { type: "thinking", content: "**Reading** the hostname" },
+ { type: "tool_call", toolCall: { id: "c1", result: "pc" } },
+ { type: "thinking", content: "Got it." },
+ { type: "text", content: "Your PC is pc." }
+ ]
+ });
+ expect(api().messages.filter(m => m.type === "thinking").map(m => m.id)).toEqual(["m2"]);
+
+ const sent = (call: number) => (streamChatCompletionMock.mock.calls[call][0] as { content: unknown }[]).map(m => m.content);
+ expect(sent(0)).toEqual(["hi", "hello", "name?"]);
+ expect(sent(1)).toEqual(["hi", "hello", "name?", "Your PC is pc.", "thanks"]);
+ });
+
it("keeps what a turn waits on until the turn ends", async () => {
let callbacks: { onStatus: (status: string) => void } | undefined;
let finish: () => void = () => undefined;
diff --git a/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.ts b/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.ts
index d05870fc3c..af003a5c67 100644
--- a/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.ts
+++ b/apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.ts
@@ -40,7 +40,7 @@ function getLastUserMessageEl(container: HTMLElement): HTMLElement | null {
/**
* Flatten a stored message's content into the wire format the server expects.
* Plain string content stays as-is; block-shaped content (with images) becomes
- * an ordered array of text/image parts, with tool-call blocks stripped.
+ * an ordered array of text/image parts, with tool-call and thinking blocks stripped.
*/
function flattenToApiContent(content: string | ContentBlock[]): string | LlmMessagePart[] {
if (typeof content === "string") {
@@ -59,7 +59,8 @@ function flattenToApiContent(content: string | ContentBlock[]): string | LlmMess
}
// tool_call blocks belong to assistant history rendering only — they
// are reconstructed from the model's own tool-use turns and must not
- // be re-sent as user/assistant content.
+ // be re-sent as user/assistant content. thinking blocks are the model's
+ // reasoning, which it must not read back as something it said.
}
// Collapse to a string if there is no multimodal content — keeps backwards
// compatibility with providers/paths that haven't been touched.
@@ -94,7 +95,6 @@ export interface UseLlmChatReturn {
hasInputText: boolean;
isStreaming: boolean;
streamingBlocks: ContentBlock[];
- streamingThinking: string;
/** What the streaming turn waits on before its reply starts, if the server said. */
streamingStatus: LlmStreamStatus | null;
pendingCitations: LlmCitation[];
@@ -201,7 +201,6 @@ export function useLlmChat(
// displayed `streamingBlocks` is derived from this with the trailing text
// block smoothed via useSmoothStreaming for a steady reveal cadence.
const [targetBlocks, setTargetBlocks] = useState<ContentBlock[]>([]);
- const [streamingThinking, setStreamingThinking] = useState("");
const [streamingStatus, setStreamingStatus] = useState<LlmStreamStatus | null>(null);
const { displayedText: smoothedTailText, append: smoothAppend, drain: smoothDrain, reset: smoothReset } = useSmoothStreaming();
const [pendingCitations, setPendingCitations] = useState<LlmCitation[]>([]);
@@ -557,11 +556,9 @@ export function useLlmChat(
setMessagesInternal(conversation);
setIsStreaming(true);
setTargetBlocks([]);
- setStreamingThinking("");
setStreamingStatus(null);
smoothReset();
- let thinkingContent = "";
const contentBlocks: ContentBlock[] = [];
const citations: LlmCitation[] = [];
let usage: LlmUsage | undefined;
@@ -577,7 +574,7 @@ export function useLlmChat(
return block as ContentBlock & { type: "text" };
}
- const apiMessages: LlmMessage[] = trimToFirstUserMessage(conversation).map(m => ({
+ const apiMessages: LlmMessage[] = trimToFirstUserMessage(conversation.filter(m => m.type !== "thinking")).map(m => ({
role: m.role,
content: stripQuoteSourcesFromApiContent(flattenToApiContent(m.content))
}));
@@ -634,35 +631,18 @@ export function useLlmChat(
}
}
- const finalNewMessages: StoredMessage[] = [];
-
- if (thinkingContent) {
- finalNewMessages.push({
- id: randomString(),
- role: "assistant",
- content: thinkingContent,
- createdAt: new Date().toISOString(),
- type: "thinking"
- });
- }
-
if (contentBlocks.length > 0) {
- finalNewMessages.push({
+ setMessages([...conversation, {
id: randomString(),
role: "assistant",
content: contentBlocks,
createdAt: new Date().toISOString(),
citations: citations.length > 0 ? citations : undefined,
usage
- });
- }
-
- if (finalNewMessages.length > 0) {
- setMessages([...conversation, ...finalNewMessages]);
+ }]);
}
setTargetBlocks([]);
- setStreamingThinking("");
setPendingCitations([]);
setStreamingStatus(null);
setIsStreaming(false);
@@ -689,8 +669,16 @@ export function useLlmChat(
setTargetBlocks([...contentBlocks]);
},
onThinking: (text) => {
- thinkingContent += text;
- setStreamingThinking(thinkingContent);
+ const last = contentBlocks[contentBlocks.length - 1];
+ if (last?.type === "thinking") {
+ contentBlocks[contentBlocks.length - 1] = { type: "thinking", content: last.content + text };
+ } else {
+ // The smoother animates only a trailing text block, so reveal the
+ // text before this thought in full.
+ smoothDrain();
+ contentBlocks.push({ type: "thinking", content: text });
+ }
+ setTargetBlocks([...contentBlocks]);
},
onToolInputStart: (toolCallId, toolName) => {
// Snap any pending smoothed text to its full value before
@@ -793,7 +781,6 @@ export function useLlmChat(
setMessages(finalMessages);
smoothReset();
setTargetBlocks([]);
- setStreamingThinking("");
setStreamingStatus(null);
setIsStreaming(false);
},
@@ -822,7 +809,6 @@ export function useLlmChat(
}]);
smoothReset();
setTargetBlocks([]);
- setStreamingThinking("");
setStreamingStatus(null);
setIsStreaming(false);
abortControllerRef.current = null;
@@ -928,7 +914,6 @@ export function useLlmChat(
hasInputText,
isStreaming,
streamingBlocks,
- streamingThinking,
streamingStatus,
pendingCitations,
pendingAttachments,