fix: normalize <think> reasoning blocks
This commit is contained in:
@@ -32,12 +32,14 @@
|
||||
- Status: show provider prefix in /status model display. (#506) — thanks @mcinteerj
|
||||
- Status: compact /status with session token usage + estimated cost, add `/cost` per-response usage lines (tokens-only for OAuth).
|
||||
- Status: show active auth profile and key snippet in /status.
|
||||
- Agent: promote `<think>`/`<thinking>` tag reasoning into structured thinking blocks so `/reasoning` works consistently for OpenAI-compat providers.
|
||||
- macOS: package ClawdbotKit resources and Swift 6.2 compatibility dylib to avoid launch/tool crashes. (#473) — thanks @gupsammy
|
||||
- WhatsApp: group `/model list` output by provider for scannability. (#456) - thanks @mcinteerj
|
||||
- Hooks: allow per-hook model overrides for webhook/Gmail runs (e.g. GPT 5 Mini).
|
||||
- Control UI: logs tab opens at the newest entries (bottom).
|
||||
- Control UI: add Docs link, remove chat composer divider, and add New session button.
|
||||
- Control UI: link sessions list to chat view. (#471) — thanks @HazAT
|
||||
- Control UI: show/patch per-session reasoning level and render extracted reasoning in chat.
|
||||
- Control UI: queue outgoing chat messages, add Enter-to-send, and show queued items. (#527) — thanks @YuriNachos
|
||||
- Control UI: drop explicit `ui:install` step; `ui:build` now auto-installs UI deps (docs + update flow).
|
||||
- Telegram: retry long-polling conflicts with backoff to avoid fatal exits.
|
||||
|
||||
@@ -167,6 +167,117 @@ describe("subscribeEmbeddedPiSession", () => {
|
||||
);
|
||||
});
|
||||
|
||||
it("promotes <think> tags to thinking blocks at write-time", () => {
|
||||
let handler: ((evt: unknown) => void) | undefined;
|
||||
const session: StubSession = {
|
||||
subscribe: (fn) => {
|
||||
handler = fn;
|
||||
return () => {};
|
||||
},
|
||||
};
|
||||
|
||||
const onBlockReply = vi.fn();
|
||||
|
||||
subscribeEmbeddedPiSession({
|
||||
session: session as unknown as Parameters<
|
||||
typeof subscribeEmbeddedPiSession
|
||||
>[0]["session"],
|
||||
runId: "run",
|
||||
onBlockReply,
|
||||
blockReplyBreak: "message_end",
|
||||
reasoningMode: "on",
|
||||
});
|
||||
|
||||
const assistantMessage = {
|
||||
role: "assistant",
|
||||
content: [
|
||||
{
|
||||
type: "text",
|
||||
text: "<think>\nBecause it helps\n</think>\n\nFinal answer",
|
||||
},
|
||||
],
|
||||
} as AssistantMessage;
|
||||
|
||||
handler?.({ type: "message_end", message: assistantMessage });
|
||||
|
||||
expect(onBlockReply).toHaveBeenCalledTimes(1);
|
||||
expect(onBlockReply.mock.calls[0][0].text).toBe(
|
||||
"_Reasoning:_\n_Because it helps_\n\nFinal answer",
|
||||
);
|
||||
|
||||
expect(assistantMessage.content).toEqual([
|
||||
{ type: "thinking", thinking: "Because it helps" },
|
||||
{ type: "text", text: "Final answer" },
|
||||
]);
|
||||
});
|
||||
|
||||
it("streams <think> reasoning via onReasoningStream without leaking into final text", () => {
|
||||
let handler: ((evt: unknown) => void) | undefined;
|
||||
const session: StubSession = {
|
||||
subscribe: (fn) => {
|
||||
handler = fn;
|
||||
return () => {};
|
||||
},
|
||||
};
|
||||
|
||||
const onReasoningStream = vi.fn();
|
||||
const onBlockReply = vi.fn();
|
||||
|
||||
subscribeEmbeddedPiSession({
|
||||
session: session as unknown as Parameters<
|
||||
typeof subscribeEmbeddedPiSession
|
||||
>[0]["session"],
|
||||
runId: "run",
|
||||
onReasoningStream,
|
||||
onBlockReply,
|
||||
blockReplyBreak: "message_end",
|
||||
reasoningMode: "stream",
|
||||
});
|
||||
|
||||
handler?.({
|
||||
type: "message_update",
|
||||
message: { role: "assistant" },
|
||||
assistantMessageEvent: {
|
||||
type: "text_delta",
|
||||
delta: "<think>\nBecause",
|
||||
},
|
||||
});
|
||||
|
||||
handler?.({
|
||||
type: "message_update",
|
||||
message: { role: "assistant" },
|
||||
assistantMessageEvent: {
|
||||
type: "text_delta",
|
||||
delta: " it helps\n</think>\n\nFinal answer",
|
||||
},
|
||||
});
|
||||
|
||||
const assistantMessage = {
|
||||
role: "assistant",
|
||||
content: [
|
||||
{
|
||||
type: "text",
|
||||
text: "<think>\nBecause it helps\n</think>\n\nFinal answer",
|
||||
},
|
||||
],
|
||||
} as AssistantMessage;
|
||||
|
||||
handler?.({ type: "message_end", message: assistantMessage });
|
||||
|
||||
expect(onBlockReply).toHaveBeenCalledTimes(1);
|
||||
expect(onBlockReply.mock.calls[0][0].text).toBe("Final answer");
|
||||
|
||||
const streamTexts = onReasoningStream.mock.calls
|
||||
.map((call) => call[0]?.text)
|
||||
.filter((value): value is string => typeof value === "string");
|
||||
expect(streamTexts.at(-1)).toBe("Reasoning:\nBecause it helps");
|
||||
|
||||
expect(assistantMessage.content).toEqual([
|
||||
{ type: "thinking", thinking: "Because it helps" },
|
||||
{ type: "text", text: "Final answer" },
|
||||
]);
|
||||
});
|
||||
|
||||
it("emits block replies on text_end and does not duplicate on message_end", () => {
|
||||
let handler: ((evt: unknown) => void) | undefined;
|
||||
const session: StubSession = {
|
||||
|
||||
@@ -24,6 +24,7 @@ const THINKING_OPEN_RE = /<\s*think(?:ing)?\s*>/i;
|
||||
const THINKING_CLOSE_RE = /<\s*\/\s*think(?:ing)?\s*>/i;
|
||||
const THINKING_OPEN_GLOBAL_RE = /<\s*think(?:ing)?\s*>/gi;
|
||||
const THINKING_CLOSE_GLOBAL_RE = /<\s*\/\s*think(?:ing)?\s*>/gi;
|
||||
const THINKING_TAG_SCAN_RE = /<\s*(\/?)\s*think(?:ing)?\s*>/gi;
|
||||
const TOOL_RESULT_MAX_CHARS = 8000;
|
||||
const log = createSubsystemLogger("agent/embedded");
|
||||
const RAW_STREAM_ENABLED = process.env.CLAWDBOT_RAW_STREAM === "1";
|
||||
@@ -121,6 +122,102 @@ function stripUnpairedThinkingTags(text: string): string {
|
||||
return text;
|
||||
}
|
||||
|
||||
type ThinkTaggedSplitBlock =
|
||||
| { type: "thinking"; thinking: string }
|
||||
| { type: "text"; text: string };
|
||||
|
||||
function splitThinkingTaggedText(text: string): ThinkTaggedSplitBlock[] | null {
|
||||
const trimmedStart = text.trimStart();
|
||||
// Avoid false positives: only treat it as structured thinking when it begins
|
||||
// with a think tag (common for local/OpenAI-compat providers that emulate
|
||||
// reasoning blocks via tags).
|
||||
if (!trimmedStart.startsWith("<")) return null;
|
||||
if (!THINKING_OPEN_RE.test(trimmedStart)) return null;
|
||||
if (!THINKING_CLOSE_RE.test(text)) return null;
|
||||
|
||||
THINKING_TAG_SCAN_RE.lastIndex = 0;
|
||||
let inThinking = false;
|
||||
let cursor = 0;
|
||||
let thinkingStart = 0;
|
||||
const blocks: ThinkTaggedSplitBlock[] = [];
|
||||
|
||||
const pushText = (value: string) => {
|
||||
if (!value) return;
|
||||
blocks.push({ type: "text", text: value });
|
||||
};
|
||||
const pushThinking = (value: string) => {
|
||||
const cleaned = value.trim();
|
||||
if (!cleaned) return;
|
||||
blocks.push({ type: "thinking", thinking: cleaned });
|
||||
};
|
||||
|
||||
for (const match of text.matchAll(THINKING_TAG_SCAN_RE)) {
|
||||
const index = match.index ?? 0;
|
||||
const isClose = Boolean(match[1]?.includes("/"));
|
||||
|
||||
if (!inThinking && !isClose) {
|
||||
pushText(text.slice(cursor, index));
|
||||
thinkingStart = index + match[0].length;
|
||||
inThinking = true;
|
||||
continue;
|
||||
}
|
||||
|
||||
if (inThinking && isClose) {
|
||||
pushThinking(text.slice(thinkingStart, index));
|
||||
cursor = index + match[0].length;
|
||||
inThinking = false;
|
||||
}
|
||||
}
|
||||
|
||||
if (inThinking) return null;
|
||||
pushText(text.slice(cursor));
|
||||
|
||||
const hasThinking = blocks.some((b) => b.type === "thinking");
|
||||
if (!hasThinking) return null;
|
||||
return blocks;
|
||||
}
|
||||
|
||||
function promoteThinkingTagsToBlocks(message: AssistantMessage): void {
|
||||
if (!Array.isArray(message.content)) return;
|
||||
const hasThinkingBlock = message.content.some((block) => {
|
||||
if (!block || typeof block !== "object") return false;
|
||||
return (block as Record<string, unknown>).type === "thinking";
|
||||
});
|
||||
if (hasThinkingBlock) return;
|
||||
|
||||
const next: Array<Record<string, unknown>> = [];
|
||||
let changed = false;
|
||||
|
||||
for (const block of message.content) {
|
||||
if (!block || typeof block !== "object") {
|
||||
next.push(block as Record<string, unknown>);
|
||||
continue;
|
||||
}
|
||||
const record = block as Record<string, unknown>;
|
||||
if (record.type !== "text" || typeof record.text !== "string") {
|
||||
next.push(record);
|
||||
continue;
|
||||
}
|
||||
const split = splitThinkingTaggedText(record.text);
|
||||
if (!split) {
|
||||
next.push(record);
|
||||
continue;
|
||||
}
|
||||
changed = true;
|
||||
for (const part of split) {
|
||||
if (part.type === "thinking") {
|
||||
next.push({ type: "thinking", thinking: part.thinking });
|
||||
} else if (part.type === "text") {
|
||||
const cleaned = part.text.trimStart();
|
||||
if (cleaned) next.push({ type: "text", text: cleaned });
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (!changed) return;
|
||||
(message as unknown as { content: unknown }).content = next;
|
||||
}
|
||||
|
||||
function normalizeSlackTarget(raw: string): string | undefined {
|
||||
const trimmed = raw.trim();
|
||||
if (!trimmed) return undefined;
|
||||
@@ -792,6 +889,7 @@ export function subscribeEmbeddedPiSession(params: {
|
||||
const msg = (evt as AgentEvent & { message: AgentMessage }).message;
|
||||
if (msg?.role === "assistant") {
|
||||
const assistantMessage = msg as AssistantMessage;
|
||||
promoteThinkingTagsToBlocks(assistantMessage);
|
||||
const rawText = extractAssistantText(assistantMessage);
|
||||
appendRawStream({
|
||||
ts: Date.now(),
|
||||
|
||||
@@ -42,12 +42,17 @@ export async function loadSessions(state: SessionsState) {
|
||||
export async function patchSession(
|
||||
state: SessionsState,
|
||||
key: string,
|
||||
patch: { thinkingLevel?: string | null; verboseLevel?: string | null },
|
||||
patch: {
|
||||
thinkingLevel?: string | null;
|
||||
verboseLevel?: string | null;
|
||||
reasoningLevel?: string | null;
|
||||
},
|
||||
) {
|
||||
if (!state.client || !state.connected) return;
|
||||
const params: Record<string, unknown> = { key };
|
||||
if ("thinkingLevel" in patch) params.thinkingLevel = patch.thinkingLevel;
|
||||
if ("verboseLevel" in patch) params.verboseLevel = patch.verboseLevel;
|
||||
if ("reasoningLevel" in patch) params.reasoningLevel = patch.reasoningLevel;
|
||||
try {
|
||||
await state.client.request("sessions.patch", params);
|
||||
await loadSessions(state);
|
||||
@@ -55,4 +60,3 @@ export async function patchSession(
|
||||
state.sessionsError = String(err);
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -38,6 +38,11 @@ export function renderChat(props: ChatProps) {
|
||||
const canCompose = props.connected;
|
||||
const isBusy = props.sending || Boolean(props.stream);
|
||||
const sessionOptions = resolveSessionOptions(props.sessionKey, props.sessions);
|
||||
const activeSession = props.sessions?.sessions?.find(
|
||||
(row) => row.key === props.sessionKey,
|
||||
);
|
||||
const reasoningLevel = activeSession?.reasoningLevel ?? "off";
|
||||
const showReasoning = reasoningLevel !== "off";
|
||||
const composePlaceholder = props.connected
|
||||
? "Message (↩ to send, Shift+↩ for line breaks)"
|
||||
: "Connect to the gateway to start chatting…";
|
||||
@@ -72,6 +77,7 @@ export function renderChat(props: ChatProps) {
|
||||
</div>
|
||||
<div class="chat-header__right">
|
||||
<div class="muted">Thinking: ${props.thinkingLevel ?? "inherit"}</div>
|
||||
<div class="muted">Reasoning: ${reasoningLevel}</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
@@ -107,7 +113,7 @@ export function renderChat(props: ChatProps) {
|
||||
{ streaming: true }
|
||||
);
|
||||
}
|
||||
return renderMessage(item.message, props);
|
||||
return renderMessage(item.message, props, { showReasoning });
|
||||
}
|
||||
)}
|
||||
</div>
|
||||
@@ -326,7 +332,7 @@ function renderReadingIndicator() {
|
||||
function renderMessage(
|
||||
message: unknown,
|
||||
props?: Pick<ChatProps, "isToolOutputExpanded" | "onToolOutputToggle">,
|
||||
opts?: { streaming?: boolean }
|
||||
opts?: { streaming?: boolean; showReasoning?: boolean }
|
||||
) {
|
||||
const m = message as Record<string, unknown>;
|
||||
const role = typeof m.role === "string" ? m.role : "unknown";
|
||||
@@ -334,6 +340,10 @@ function renderMessage(
|
||||
const hasToolCards = toolCards.length > 0;
|
||||
const isToolResult = isToolResultMessage(message);
|
||||
const extractedText = extractText(message);
|
||||
const extractedThinking =
|
||||
opts?.showReasoning && role === "assistant"
|
||||
? extractThinking(message)
|
||||
: null;
|
||||
const contentText = typeof m.content === "string" ? m.content : null;
|
||||
const fallback = hasToolCards ? null : JSON.stringify(message, null, 2);
|
||||
|
||||
@@ -345,10 +355,15 @@ function renderMessage(
|
||||
: !isToolResult && fallback
|
||||
? { kind: "json" as const, value: fallback }
|
||||
: null;
|
||||
const markdown =
|
||||
const markdownBase =
|
||||
display?.kind === "json"
|
||||
? ["```json", display.value, "```"].join("\n")
|
||||
: (display?.value ?? null);
|
||||
const markdown = extractedThinking
|
||||
? [formatReasoningMarkdown(extractedThinking), markdownBase]
|
||||
.filter(Boolean)
|
||||
.join("\n\n")
|
||||
: markdownBase;
|
||||
|
||||
const timestamp =
|
||||
typeof m.timestamp === "number" ? new Date(m.timestamp).toLocaleTimeString() : "";
|
||||
@@ -413,6 +428,60 @@ function extractText(message: unknown): string | null {
|
||||
return null;
|
||||
}
|
||||
|
||||
function extractThinking(message: unknown): string | null {
|
||||
const m = message as Record<string, unknown>;
|
||||
const content = m.content;
|
||||
const parts: string[] = [];
|
||||
if (Array.isArray(content)) {
|
||||
for (const p of content) {
|
||||
const item = p as Record<string, unknown>;
|
||||
if (item.type === "thinking" && typeof item.thinking === "string") {
|
||||
const cleaned = item.thinking.trim();
|
||||
if (cleaned) parts.push(cleaned);
|
||||
}
|
||||
}
|
||||
}
|
||||
if (parts.length > 0) return parts.join("\n");
|
||||
|
||||
// Back-compat: older logs may still have <think> tags inside text blocks.
|
||||
const rawText = extractRawText(message);
|
||||
if (!rawText) return null;
|
||||
const matches = [...rawText.matchAll(/<\s*think(?:ing)?\s*>([\s\S]*?)<\s*\/\s*think(?:ing)?\s*>/gi)];
|
||||
const extracted = matches
|
||||
.map((m) => (m[1] ?? "").trim())
|
||||
.filter(Boolean);
|
||||
return extracted.length > 0 ? extracted.join("\n") : null;
|
||||
}
|
||||
|
||||
function extractRawText(message: unknown): string | null {
|
||||
const m = message as Record<string, unknown>;
|
||||
const content = m.content;
|
||||
if (typeof content === "string") return content;
|
||||
if (Array.isArray(content)) {
|
||||
const parts = content
|
||||
.map((p) => {
|
||||
const item = p as Record<string, unknown>;
|
||||
if (item.type === "text" && typeof item.text === "string") return item.text;
|
||||
return null;
|
||||
})
|
||||
.filter((v): v is string => typeof v === "string");
|
||||
if (parts.length > 0) return parts.join("\n");
|
||||
}
|
||||
if (typeof m.text === "string") return m.text;
|
||||
return null;
|
||||
}
|
||||
|
||||
function formatReasoningMarkdown(text: string): string {
|
||||
const trimmed = text.trim();
|
||||
if (!trimmed) return "";
|
||||
const lines = trimmed
|
||||
.split(/\r?\n/)
|
||||
.map((line) => line.trim())
|
||||
.filter(Boolean)
|
||||
.map((line) => `_${line}_`);
|
||||
return lines.length ? ["_Reasoning:_", ...lines].join("\n") : "";
|
||||
}
|
||||
|
||||
type ToolCard = {
|
||||
kind: "call" | "result";
|
||||
name: string;
|
||||
|
||||
@@ -23,12 +23,17 @@ export type SessionsProps = {
|
||||
onRefresh: () => void;
|
||||
onPatch: (
|
||||
key: string,
|
||||
patch: { thinkingLevel?: string | null; verboseLevel?: string | null },
|
||||
patch: {
|
||||
thinkingLevel?: string | null;
|
||||
verboseLevel?: string | null;
|
||||
reasoningLevel?: string | null;
|
||||
},
|
||||
) => void;
|
||||
};
|
||||
|
||||
const THINK_LEVELS = ["", "off", "minimal", "low", "medium", "high"] as const;
|
||||
const VERBOSE_LEVELS = ["", "off", "on"] as const;
|
||||
const REASONING_LEVELS = ["", "off", "on", "stream"] as const;
|
||||
|
||||
export function renderSessions(props: SessionsProps) {
|
||||
const rows = props.result?.sessions ?? [];
|
||||
@@ -117,6 +122,7 @@ export function renderSessions(props: SessionsProps) {
|
||||
<div>Tokens</div>
|
||||
<div>Thinking</div>
|
||||
<div>Verbose</div>
|
||||
<div>Reasoning</div>
|
||||
</div>
|
||||
${rows.length === 0
|
||||
? html`<div class="muted">No sessions found.</div>`
|
||||
@@ -130,6 +136,7 @@ function renderRow(row: GatewaySessionRow, basePath: string, onPatch: SessionsPr
|
||||
const updated = row.updatedAt ? formatAgo(row.updatedAt) : "n/a";
|
||||
const thinking = row.thinkingLevel ?? "";
|
||||
const verbose = row.verboseLevel ?? "";
|
||||
const reasoning = row.reasoningLevel ?? "";
|
||||
const displayName = row.displayName ?? row.key;
|
||||
const canLink = row.kind !== "global";
|
||||
const chatUrl = canLink
|
||||
@@ -170,6 +177,19 @@ function renderRow(row: GatewaySessionRow, basePath: string, onPatch: SessionsPr
|
||||
)}
|
||||
</select>
|
||||
</div>
|
||||
<div>
|
||||
<select
|
||||
.value=${reasoning}
|
||||
@change=${(e: Event) => {
|
||||
const value = (e.target as HTMLSelectElement).value;
|
||||
onPatch(row.key, { reasoningLevel: value || null });
|
||||
}}
|
||||
>
|
||||
${REASONING_LEVELS.map((level) =>
|
||||
html`<option value=${level}>${level || "inherit"}</option>`,
|
||||
)}
|
||||
</select>
|
||||
</div>
|
||||
</div>
|
||||
`;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user