Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

fix(web): restore the reasoning stream the redesign dropped

kevin932711eb32e663ba
diff --git a/web/src/App.tsx b/web/src/App.tsx
index e2dc7e1d..30fee3fa 100644
--- a/web/src/App.tsx
+++ b/web/src/App.tsx
@@ -241,6 +241,19 @@ export default function App() {
enableThinking: thinking,
cacheSlot: supportsCacheSlots(health) ? cacheSlot : undefined,
signal: controller.signal,
+ /* Reasoning tokens are tokens: they count toward the rate, and the
+ first one is the real time-to-first-token. The answer's first token
+ arrives much later on a reasoning model. */
+ onReasoning: (delta) => {
+ if (firstToken) { setTtft(performance.now() - t0); setStreamStart(performance.now()); decodeStart = performance.now(); firstToken = false }
+ count++
+ setTokenCount(count)
+ const since = (performance.now() - decodeStart) / 1000
+ if (count > 1 && since > 0.2) setTokPerSec((count - 1) / since)
+ updateMessages((current) => current.map((item) =>
+ item.id === assistant.id ? { ...item, reasoning: (item.reasoning ?? "") + delta } : item,
+ ))
+ },
onDelta: (delta) => {
if (firstToken) { setTtft(performance.now() - t0); setStreamStart(performance.now()); decodeStart = performance.now(); firstToken = false }
count++
@@ -268,10 +281,10 @@ export default function App() {
setConnected(true)
} catch (cause) {
if (controller.signal.aborted) {
- updateMessages((current) => current.filter((item) => item.id !== assistant.id || item.content))
+ updateMessages((current) => current.filter((item) => item.id !== assistant.id || item.content || item.reasoning))
} else {
setError(cause instanceof Error ? cause.message : "status.generationFailed")
- updateMessages((current) => current.filter((item) => item.id !== assistant.id || item.content))
+ updateMessages((current) => current.filter((item) => item.id !== assistant.id || item.content || item.reasoning))
}
} finally {
abortRef.current = null
@@ -408,7 +421,12 @@ export default function App() {
{item.role !== "user" && <div className="assistant-brand"><Brand /><span>colibrì</span></div>}
{item.images?.length ? <div className="message-images">{item.images.map((url, at) =>
<img key={at} src={url} alt={t("ui.attachedImage", { n: at + 1 })} />)}</div> : null}
- <div className="message-body">{item.content ? (item.role === "assistant" ? <Markdown text={item.content} /> : item.content) : <span className="typing" aria-label={t("ui.generating")}><i /><i /><i /></span>}</div>
+ <div className="message-body">{item.reasoning
+ ? <details className="reasoning" open={!item.content}>
+ <summary>{t("sidebar.reasoning")}</summary>
+ <div className="reasoning-body">{item.reasoning}</div>
+ </details>
+ : null}{item.content ? (item.role === "assistant" ? <Markdown text={item.content} /> : item.content) : <span className="typing" aria-label={t("ui.generating")}><i /><i /><i /></span>}</div>
{item.role === "assistant" && item.content && <div className="message-actions"><button className="icon-action" aria-label={t("ui.copy")} title={t("ui.copy")} onClick={() => void copyMessage(item)}><Copy /></button>{copied === item.id && <span role="status">{t("ui.copied")}</span>}{index === messages.length - 1 && !loading && <button className="icon-action" aria-label={t("ui.regenerate")} title={t("ui.regenerate")} onClick={() => { const userIndex = messages.map((m, i) => m.role === "user" && i < index ? i : -1).reduce((a, b) => Math.max(a, b), -1); if (userIndex >= 0) void send(messages[userIndex].content, messages.slice(0, userIndex)) }}><RefreshCw /></button>}</div>}
</article>)}<div ref={bottomRef} /></div>
</div>}
diff --git a/web/src/chat-design.css b/web/src/chat-design.css
index 675e0966..a64bd0d4 100644
--- a/web/src/chat-design.css
+++ b/web/src/chat-design.css
@@ -18,6 +18,16 @@ button{color:inherit}button:disabled{cursor:default;opacity:.4}svg.lucide{stroke
.send-button{border:0;display:grid;place-items:center;width:34px;height:34px;flex-shrink:0;border-radius:50%;background:var(--foreground);color:var(--background)}.send-button svg{width:21px;height:21px}
.chat-view .suggestions{display:flex;justify-content:center;flex-wrap:wrap;gap:8px;width:min(760px,100%);margin:22px auto 0}.chat-view .suggestions button{min-height:0;display:flex;align-items:center;justify-content:center;gap:7px;padding:10px 13px;font-size:11px;color:var(--muted-foreground);background:none;border:1px solid var(--border);border-radius:24px}.chat-view .suggestions button svg{width:15px;height:15px}.chat-view .suggestions button:hover{background:var(--secondary);transform:none;color:var(--foreground)}
.chat-view .conversation{flex:1;min-height:0;overflow:auto;overscroll-behavior:contain}.chat-view .message-list{width:min(760px,calc(100% - 64px));padding:26px 0 20px}.chat-view .message{display:flex;flex-direction:column;gap:13px;margin-bottom:30px}.chat-view .message.user{align-items:flex-end}.chat-view .message-body{font-size:15px;line-height:1.8;color:var(--foreground)}.chat-view .message.user .message-body{max-width:85%;padding:11px 19px;border-radius:24px;background:var(--input);color:var(--foreground)}.assistant-brand{display:flex;align-items:center;gap:9px;font-size:13px}.assistant-brand .colibri-brand>svg{width:23px;height:22px}.message-actions{display:flex;align-items:center;gap:3px}.message-actions>span{font-size:10px;color:var(--muted-foreground)}.truncation-note{max-width:760px;margin:8px auto 0;font-size:11px;color:var(--muted-foreground)}
+/* Reasoning models stream thinking before the answer. Folded once the answer
+ starts (open={!item.content} in App.tsx) so a long chain of thought does not
+ push the reply off the screen. */
+.chat-view .reasoning{margin-bottom:14px;border-left:2px solid color-mix(in srgb,var(--primary) 45%,transparent);padding-left:12px}
+.chat-view .reasoning>summary{color:var(--muted-foreground);font-size:10px;font-weight:700;letter-spacing:.07em;text-transform:uppercase;cursor:pointer;list-style:none}
+.chat-view .reasoning>summary::-webkit-details-marker{display:none}
+.chat-view .reasoning>summary::before{content:"\25B8 ";display:inline-block;transition:transform .15s}
+.chat-view .reasoning[open]>summary::before{transform:rotate(90deg)}
+.chat-view .reasoning>summary:hover{color:var(--foreground)}
+.chat-view .reasoning-body{margin-top:8px;color:var(--muted-foreground);font-size:13px;line-height:1.7;white-space:pre-wrap;overflow-wrap:anywhere}
.navigation-dock{position:fixed;left:calc(50% + var(--rail)/2);top:0;transform:translateX(-50%);width:104px;height:27px;z-index:70;display:flex;align-items:center;justify-content:center;border-radius:0 0 14px 14px;background:var(--card);border:1px solid var(--border);border-top:0;transition:width .3s,height .3s,margin .3s,border-radius .3s,background .25s,box-shadow .25s}.navigation-dock.open{width:300px;height:64px;margin-top:11px;border-radius:34px;background:color-mix(in srgb,var(--card) 88%,transparent);backdrop-filter:blur(24px);box-shadow:0 6px 22px #0000002e;border:1px solid color-mix(in srgb,var(--foreground) 12%,transparent)}.navigation-dock nav{display:flex;gap:18px;align-items:center}.navigation-dock nav button{display:grid;place-items:center;position:relative;border:0;background:none;border-radius:50%;width:46px;height:46px;padding:0;color:var(--muted-foreground);font-size:12px}.navigation-dock nav button:hover,.navigation-dock nav button:focus-visible{color:var(--foreground)}.navigation-dock nav button[aria-current]{background:color-mix(in srgb,var(--foreground) 10%,transparent);color:var(--foreground)}.navigation-dock nav button>span{position:absolute;top:calc(100% + 19px);left:50%;white-space:nowrap;font-size:12px;color:var(--muted-foreground);opacity:0;pointer-events:none;transform:translateX(-50%) translateY(-4px);transition:opacity .18s,transform .18s}.navigation-dock nav button:hover>span,.navigation-dock nav button:focus-visible>span{opacity:1;transform:translateX(-50%) translateY(0)}.navigation-dock nav svg{width:21px;height:21px;flex-shrink:0}.dock-handle{position:absolute;inset:0;width:100%;height:27px;background:none;border:0;display:grid;place-items:center;touch-action:none}.dock-handle span{width:13px;height:13px;border-right:2.5px solid var(--foreground);border-bottom:2.5px solid var(--foreground);border-radius:0 0 3px 0;opacity:.8;transform:translateY(-4px) rotate(45deg);transition:transform .3s,opacity .25s}.navigation-dock.open .dock-handle{top:auto;bottom:-14px;height:18px}.navigation-dock.open .dock-handle span{transform:translateY(3px) rotate(225deg);opacity:.45}.dock-busy{width:4px;height:4px;border-radius:50%;background:#b6dac3;position:absolute;right:8px;top:8px}
.settings-page,.profiling-workspace,.brio-workspace{flex:1;min-height:0;overflow:auto;padding:28px 44px 32px}.page-heading,.settings-tabs,.settings-content{max-width:1080px;margin:0 auto}.page-heading>span{font-size:9px;letter-spacing:1.5px;color:var(--muted-foreground)}.page-heading h1{font-size:32px;letter-spacing:-.9px;font-weight:450;margin:12px 0 10px}.page-heading p{font-size:13px;line-height:1.7;color:var(--muted-foreground);margin:0}.settings-tabs{display:flex;gap:6px;margin:28px auto 25px;border-bottom:1px solid var(--border);padding-bottom:12px}.settings-tabs button{border:0;border-radius:9px;padding:10px 17px;background:none;color:var(--muted-foreground);font-size:12px}.settings-tabs button[aria-current]{background:var(--secondary);color:var(--foreground)}.settings-card{max-width:730px;border:1px solid var(--border);border-radius:20px;padding:26px;background:var(--card)}.settings-card h2{font-size:16px;font-weight:500;margin:0 0 23px}.settings-card .side-section{gap:20px}.settings-card .section-title{color:var(--muted-foreground);font-size:11px}.settings-card .side-section label{color:var(--muted-foreground);font-size:12px;gap:9px}.settings-card .side-section select{font-size:13px}.settings-card .runtime-grid strong{font-size:24px;font-weight:400}.settings-card .field-help{font-size:11px;line-height:1.6}.settings-card .hw-panel{gap:12px}.settings-card .hw-row{font-size:12px}.settings-card .session-stats{margin:15px 0}.settings-help{font-size:11px;margin:25px 0 0;color:var(--muted-foreground)}.theme-choices{display:flex;gap:14px}.theme-choice{width:145px;padding:10px;background:none;border:1px solid var(--border);border-radius:12px;text-align:left}.theme-choice[aria-pressed=true]{border-color:var(--primary)}.theme-choice>span{display:block;height:80px;background:#212121;border:1px solid #ffffff13;border-radius:7px;box-shadow:inset 23px 0 #171717}.theme-choice.light>span{background:#fff;box-shadow:inset 23px 0 #eee;border-color:#ddd}.theme-choice b{display:block;font-size:12px;font-weight:400;margin-top:10px}.language-field{display:flex;align-items:center;gap:10px;margin-top:26px;font-size:12px}.language-field svg{width:16px}.language-field select{margin-left:auto;border:1px solid var(--border);border-radius:8px;padding:9px;background:var(--input);max-width:180px}
.live-metrics{display:flex;flex-wrap:wrap;gap:9px;max-width:1080px;margin:18px auto}.live-metrics [data-slot=badge]{gap:6px;font-weight:400;padding:8px 12px}.profiling-workspace .prof-page{max-width:1080px;margin:auto;padding:14px 0;overflow:visible}.prof-head .section-title{color:var(--muted-foreground)}.prof-tiles>div,.prof-shares,.prof-chart,.prof-table-wrap{border-radius:17px;padding:20px;background:var(--card)}.prof-tiles strong{font:400 29px ui-sans-serif,system-ui,sans-serif;letter-spacing:-1px}.prof-charts{gap:16px}.prof-chart{padding:24px}.prof-plot svg{height:160px}.prof-table-wrap{padding:0}.prof-table td{color:var(--foreground)}.prof-share-head,.prof-chart-title,.prof-legend{color:var(--muted-foreground)}.prof-plot:focus-visible{outline:2px solid var(--primary);outline-offset:5px}
diff --git a/web/src/lib/api.test.ts b/web/src/lib/api.test.ts
index e08ee85c..97ae2ebf 100644
--- a/web/src/lib/api.test.ts
+++ b/web/src/lib/api.test.ts
@@ -83,6 +83,39 @@ describe("chat request extensions", () => {
})
})
+describe("reasoning stream", () => {
+ /* The server puts thinking on delta.reasoning_content (openai_server.py,
+ 22 call sites). #1148 / #1153 / #1156 taught the dashboard to read that
+ field. The workspace redesign dropped the callback, so a turn with
+ enable_thinking on streamed thinking tokens that never appeared. */
+ const reasoningStream = () => new Response(
+ 'data: {"choices":[{"delta":{"reasoning_content":"think"}}]}\n\n'
+ + 'data: {"choices":[{"delta":{"content":"answer"}}]}\n\n'
+ + "data: [DONE]\n\n",
+ { headers: { "content-type": "text/event-stream" } },
+ )
+
+ it("forwards reasoning_content to onReasoning and keeps it out of content", async () => {
+ vi.stubGlobal("fetch", vi.fn().mockResolvedValue(reasoningStream()))
+ const content: string[] = []
+ const reasoning: string[] = []
+ await streamChat({
+ baseUrl: "http://localhost:8000/v1",
+ apiKey: "",
+ model: "test-model",
+ messages: [],
+ temperature: 0,
+ maxTokens: 8,
+ enableThinking: true,
+ signal: new AbortController().signal,
+ onDelta: (text) => content.push(text),
+ onReasoning: (text) => reasoning.push(text),
+ })
+ expect(reasoning).toEqual(["think"])
+ expect(content).toEqual(["answer"])
+ })
+})
+
describe("askBrio", () => {
/* The options must travel as a field, never folded into the prompt: keeping
them out of the text is half of what the mode saves, and a refactor that
diff --git a/web/src/lib/api.ts b/web/src/lib/api.ts
index b6ed33a6..545527d3 100644
--- a/web/src/lib/api.ts
+++ b/web/src/lib/api.ts
@@ -4,6 +4,10 @@ export interface ChatMessage {
id: string
role: ChatRole
content: string
+ /* Reasoning models stream their thinking on a separate delta field before
+ the answer. Kept apart from `content` so it can be rendered as its own
+ block and excluded from what is sent back as conversation history. */
+ reasoning?: string
/* Data URIs of the pictures attached to this turn. Kept on the message rather
than on the draft, because the transcript is resent on every later turn and
the model has to keep seeing what it was shown. */
@@ -150,6 +154,7 @@ export interface StreamChatOptions {
cacheSlot?: number
signal: AbortSignal
onDelta: (text: string) => void
+ onReasoning?: (text: string) => void
}
export async function streamChat(options: StreamChatOptions): Promise<StreamChatResult> {
@@ -188,12 +193,14 @@ export async function streamChat(options: StreamChatOptions): Promise<StreamChat
const consume = (data: string) => {
if (data === "[DONE]") return
const event = JSON.parse(data) as {
- choices?: Array<{ delta?: { content?: string }; finish_reason?: string | null }>
+ choices?: Array<{ delta?: { content?: string; reasoning_content?: string }; finish_reason?: string | null }>
usage?: TokenUsage | null
}
const choice = event.choices?.[0]
const text = choice?.delta?.content
if (text) options.onDelta(text)
+ const reasoning = choice?.delta?.reasoning_content
+ if (reasoning) options.onReasoning?.(reasoning)
if (choice?.finish_reason) finishReason = choice.finish_reason
if (event.usage) usage = event.usage
}