@alfe.ai/openclaw-chat 0.3.3 → 0.4.0

package/dist/index.d.cts

@@ -1,4 +1,4 @@
-import { a as plugin, c as AlfeChannelConfig, l as AlfePluginConfig, s as createAlfeChannelPlugin, u as AlfeResolvedAccount } from "./plugin.cjs";
+import { d as AlfePluginConfig, f as AlfeResolvedAccount, l as createAlfeChannelPlugin, o as plugin, u as AlfeChannelConfig } from "./plugin.cjs";
 
 //#region src/session-store.d.ts
 
@@ -19,6 +19,35 @@ interface ChatMessage {
   senderId?: string;
   senderName?: string;
 }
+/**
+ * Persisted form of a turn's agent activity (tool cards + thinking), so a
+ * refresh can replay what the live `chat-activity` stream showed. Terminal
+ * by construction: no `progressText` (transient streaming window) and no
+ * `running` status — a tool still running at flush time is persisted as
+ * `interrupted`. Tool text fields arrive pre-capped (activity-serialize.ts);
+ * thinking text is capped at append time (MAX_ACTIVITY_THINKING_CHARS).
+ */
+type ChatActivityRecord = {
+  kind: 'thinking';
+  text: string;
+  ts: number;
+} | {
+  kind: 'tool';
+  toolCallId: string;
+  name: string;
+  status: 'done' | 'failed' | 'interrupted';
+  summary?: string;
+  argsText?: string;
+  resultText?: string;
+  isError?: boolean;
+  durationMs?: number;
+  truncated?: {
+    args?: boolean;
+    progress?: boolean;
+    result?: boolean;
+  };
+  ts: number;
+};
 interface SessionData {
   sessionId: string;
   agentId: string;
@@ -28,6 +57,8 @@ interface SessionData {
   createdAt: string;
   updatedAt: string;
   messages: ChatMessage[];
+  /** Absent on session files written before activity persistence shipped. */
+  activity?: ChatActivityRecord[];
 }
 interface SessionSummary {
   sessionId: string;

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-import { a as plugin, c as AlfeChannelConfig, l as AlfePluginConfig, s as createAlfeChannelPlugin, u as AlfeResolvedAccount } from "./plugin.js";
+import { d as AlfePluginConfig, f as AlfeResolvedAccount, l as createAlfeChannelPlugin, o as plugin, u as AlfeChannelConfig } from "./plugin.js";
 
 //#region src/session-store.d.ts
 
@@ -19,6 +19,35 @@ interface ChatMessage {
   senderId?: string;
   senderName?: string;
 }
+/**
+ * Persisted form of a turn's agent activity (tool cards + thinking), so a
+ * refresh can replay what the live `chat-activity` stream showed. Terminal
+ * by construction: no `progressText` (transient streaming window) and no
+ * `running` status — a tool still running at flush time is persisted as
+ * `interrupted`. Tool text fields arrive pre-capped (activity-serialize.ts);
+ * thinking text is capped at append time (MAX_ACTIVITY_THINKING_CHARS).
+ */
+type ChatActivityRecord = {
+  kind: 'thinking';
+  text: string;
+  ts: number;
+} | {
+  kind: 'tool';
+  toolCallId: string;
+  name: string;
+  status: 'done' | 'failed' | 'interrupted';
+  summary?: string;
+  argsText?: string;
+  resultText?: string;
+  isError?: boolean;
+  durationMs?: number;
+  truncated?: {
+    args?: boolean;
+    progress?: boolean;
+    result?: boolean;
+  };
+  ts: number;
+};
 interface SessionData {
   sessionId: string;
   agentId: string;
@@ -28,6 +57,8 @@ interface SessionData {
   createdAt: string;
   updatedAt: string;
   messages: ChatMessage[];
+  /** Absent on session files written before activity persistence shipped. */
+  activity?: ChatActivityRecord[];
 }
 interface SessionSummary {
   sessionId: string;

package/dist/index.js

@@ -1,2 +1,2 @@
-import { i as plugin, o as createAlfeChannelPlugin } from "./plugin2.js";
+import { a as plugin, c as createAlfeChannelPlugin } from "./plugin2.js";
 export { createAlfeChannelPlugin, plugin as default };

package/dist/plugin.cjs

@@ -7,4 +7,6 @@ exports.admitAgentEvent = require_plugin.admitAgentEvent;
 exports.buildA2ACompletePayload = require_plugin.buildA2ACompletePayload;
 exports.buildToolActivity = require_plugin.buildToolActivity;
 exports.default = require_plugin.plugin;
+exports.joinAssistantText = require_plugin.joinAssistantText;
+exports.rewriteAssistantText = require_plugin.rewriteAssistantText;
 exports.stripLeakedToolSummary = require_plugin.stripLeakedToolSummary;

package/dist/plugin.d.cts

@@ -334,6 +334,21 @@ type ChatActivity = {
   ts?: number;
   seq?: number;
 };
+/**
+ * Join the text parts of an assistant-stream event's content array.
+ * Assistant `evt.data` is a message object whose `content` is an array of
+ * typed parts; deltas are CUMULATIVE snapshots (the relay length-compares
+ * them — see adapter.ts's lastSentLength slicing). Non-array/absent content
+ * yields '' (nothing to divert).
+ */
+declare function joinAssistantText(data: Record<string, unknown>): string;
+/**
+ * Rebuild an assistant event's message with its text parts replaced by one
+ * `{type:'text', text: visible}` part. Non-text parts are preserved — the
+ * relay's consumers (adapter/inject/voice-bridge) only join text parts, so
+ * relative ordering against non-text items is not load-bearing.
+ */
+declare function rewriteAssistantText(data: Record<string, unknown>, visible: string): Record<string, unknown>;
 /**
  * Extract structured tool status from an OpenClaw `tool`-stream event.
  * We send the emoji-free raw `name` plus a lifecycle `status` and let the
@@ -396,4 +411,4 @@ declare const plugin: {
   deactivate(api: PluginApi): Promise<void>;
 };
 //#endregion
-export { plugin as a, AlfeChannelConfig as c, buildToolActivity as i, AlfePluginConfig as l, admitAgentEvent as n, stripLeakedToolSummary as o, buildA2ACompletePayload as r, createAlfeChannelPlugin as s, RunEventGate as t, AlfeResolvedAccount as u };
+export { joinAssistantText as a, stripLeakedToolSummary as c, AlfePluginConfig as d, AlfeResolvedAccount as f, buildToolActivity as i, createAlfeChannelPlugin as l, admitAgentEvent as n, plugin as o, buildA2ACompletePayload as r, rewriteAssistantText as s, RunEventGate as t, AlfeChannelConfig as u };

package/dist/plugin.d.ts

@@ -334,6 +334,21 @@ type ChatActivity = {
   ts?: number;
   seq?: number;
 };
+/**
+ * Join the text parts of an assistant-stream event's content array.
+ * Assistant `evt.data` is a message object whose `content` is an array of
+ * typed parts; deltas are CUMULATIVE snapshots (the relay length-compares
+ * them — see adapter.ts's lastSentLength slicing). Non-array/absent content
+ * yields '' (nothing to divert).
+ */
+declare function joinAssistantText(data: Record<string, unknown>): string;
+/**
+ * Rebuild an assistant event's message with its text parts replaced by one
+ * `{type:'text', text: visible}` part. Non-text parts are preserved — the
+ * relay's consumers (adapter/inject/voice-bridge) only join text parts, so
+ * relative ordering against non-text items is not load-bearing.
+ */
+declare function rewriteAssistantText(data: Record<string, unknown>, visible: string): Record<string, unknown>;
 /**
  * Extract structured tool status from an OpenClaw `tool`-stream event.
  * We send the emoji-free raw `name` plus a lifecycle `status` and let the
@@ -396,4 +411,4 @@ declare const plugin: {
   deactivate(api: PluginApi): Promise<void>;
 };
 //#endregion
-export { plugin as a, AlfeChannelConfig as c, buildToolActivity as i, AlfePluginConfig as l, admitAgentEvent as n, stripLeakedToolSummary as o, buildA2ACompletePayload as r, createAlfeChannelPlugin as s, RunEventGate as t, AlfeResolvedAccount as u };
+export { joinAssistantText as a, stripLeakedToolSummary as c, AlfePluginConfig as d, AlfeResolvedAccount as f, buildToolActivity as i, createAlfeChannelPlugin as l, admitAgentEvent as n, plugin as o, buildA2ACompletePayload as r, rewriteAssistantText as s, RunEventGate as t, AlfeChannelConfig as u };

package/dist/plugin.js

@@ -1,2 +1,2 @@
-import { a as stripLeakedToolSummary, i as plugin, n as buildA2ACompletePayload, r as buildToolActivity, t as admitAgentEvent } from "./plugin2.js";
-export { admitAgentEvent, buildA2ACompletePayload, buildToolActivity, plugin as default, stripLeakedToolSummary };
+import { a as plugin, i as joinAssistantText, n as buildA2ACompletePayload, o as rewriteAssistantText, r as buildToolActivity, s as stripLeakedToolSummary, t as admitAgentEvent } from "./plugin2.js";
+export { admitAgentEvent, buildA2ACompletePayload, buildToolActivity, plugin as default, joinAssistantText, rewriteAssistantText, stripLeakedToolSummary };

package/dist/plugin2.cjs

@@ -344,6 +344,31 @@ async function addMessage(sessionId, role, content, senderId, senderName) {
 	writeLocks.set(sessionId, next);
 	await next;
 }
+const MAX_ACTIVITY_PER_SESSION = 500;
+const MAX_ACTIVITY_THINKING_CHARS = 16e3;
+/**
+* Append a turn's terminalized activity records. Called ONCE per turn (from
+* the dispatch `finally`) — never per-delta: thinking deltas are
+* high-frequency and each save rewrites the whole session file.
+*/
+async function appendActivity(sessionId, entries) {
+	if (entries.length === 0) return;
+	const next = (writeLocks.get(sessionId) ?? Promise.resolve()).then(async () => {
+		const session = await getSession(sessionId);
+		if (!session) return;
+		const capped = entries.map((e) => e.kind === "thinking" && e.text.length > MAX_ACTIVITY_THINKING_CHARS ? {
+			...e,
+			text: e.text.slice(0, MAX_ACTIVITY_THINKING_CHARS)
+		} : e);
+		const combined = [...session.activity ?? [], ...capped];
+		session.activity = combined.length > MAX_ACTIVITY_PER_SESSION ? combined.slice(combined.length - MAX_ACTIVITY_PER_SESSION) : combined;
+		await saveSession(session);
+	}).finally(() => {
+		if (writeLocks.get(sessionId) === next) writeLocks.delete(sessionId);
+	});
+	writeLocks.set(sessionId, next);
+	await next;
+}
 async function listSessions(filters, limit = 50) {
 	await ensureDir();
 	let files;
@@ -848,6 +873,101 @@ function extractResultText(result) {
 	return parts.join("\n");
 }
 //#endregion
+//#region src/think-tags.ts
+/**
+* Inline `<think>` span handling for assistant text.
+*
+* Reasoning models routed through the AI proxy (DeepSeek `deepseek-*`,
+* Zhipu `glm-*`, …) emit their reasoning INLINE as `<think>…</think>` inside
+* the normal content stream instead of a structured thinking channel. The
+* openclaw-chat plugin (0.4.0+) diverts those spans into thinking activity
+* frames at the source; this relay-side copy is the DEFENSIVE layer for
+* agents still running older plugin versions — leaked reasoning is dropped
+* here, never synthesized into activity (that's the plugin's job).
+*
+* KEEP IN SYNC (byte-identical): this module exists at BOTH
+*   `services/chat/src/lib/think-tags.ts`        (relay, defensive layer)
+*   `packages/openclaw-chat/src/think-tags.ts`   (plugin, at-source divert)
+* The published plugin cannot import service internals, so the module is
+* duplicated, same as the ChatActivity shape.
+*
+* Known accepted limitation: a literal `<think>` in legitimate prose or a
+* code fence is treated as reasoning and stripped/diverted.
+*/
+const OPEN_TAG = "<think>";
+const CLOSE_TAG = "</think>";
+/**
+* Split CUMULATIVE assistant text into the user-visible part and the
+* reasoning part.
+*
+* Invariant (load-bearing): for successive cumulative snapshots that grow by
+* appending, `visible` is monotonically append-only — `visible(n)` is always
+* a prefix of `visible(n + 1)`. The delta relays slice by a `lastSentLength`
+* offset into this string; if `visible` ever shrank or rewrote earlier
+* characters, clients would receive corrupted deltas. This holds because the
+* scan is deterministic left-to-right and a trailing PARTIAL tag (any strict
+* prefix of the next expected tag, e.g. `"<thi"` or `"</th"`) is HELD BACK
+* from both outputs until later text disambiguates it.
+*/
+function splitThinkTags(cumulative) {
+	let visible = "";
+	let thinking = "";
+	let inThink = false;
+	let i = 0;
+	const n = cumulative.length;
+	while (i < n) {
+		const tag = inThink ? CLOSE_TAG : OPEN_TAG;
+		const idx = cumulative.indexOf(tag, i);
+		if (idx !== -1) {
+			const chunk = cumulative.slice(i, idx);
+			if (inThink) thinking += chunk;
+			else visible += chunk;
+			inThink = !inThink;
+			i = idx + tag.length;
+			continue;
+		}
+		let rest = cumulative.slice(i);
+		const held = trailingPartialTagLength(rest, tag);
+		if (held > 0) rest = rest.slice(0, rest.length - held);
+		if (inThink) thinking += rest;
+		else visible += rest;
+		break;
+	}
+	return {
+		visible,
+		thinking
+	};
+}
+/** Length of the longest strict prefix of `tag` that is a suffix of `text`. */
+function trailingPartialTagLength(text, tag) {
+	const max = Math.min(text.length, tag.length - 1);
+	for (let len = max; len > 0; len--) if (text.endsWith(tag.slice(0, len))) return len;
+	return 0;
+}
+/**
+* Strip `<think>…</think>` spans (and an unclosed trailing `<think>…`) from
+* FINAL text, then tidy the whitespace the spans leave behind.
+*
+* Idempotent — `stripThinkSpans(stripThinkSpans(x)) === stripThinkSpans(x)`.
+* New plugin + new relay both strip, so the second pass must be a no-op; the
+* fixpoint loop also covers pathological inputs where removing one span
+* makes the surrounding fragments form a new one.
+*
+* NOT for sliced/streaming paths — the trim + blank-line collapse breaks
+* prefix alignment with `lastSentLength`; those paths use
+* `splitThinkTags(...).visible` instead.
+*/
+function stripThinkSpans(text) {
+	let out = text;
+	let prev;
+	do {
+		prev = out;
+		out = out.replace(/<think>[\s\S]*?<\/think>/g, "");
+	} while (out !== prev);
+	out = out.replace(/<think>[\s\S]*$/, "");
+	return out.replace(/\n{3,}/g, "\n\n").trim();
+}
+//#endregion
 //#region src/plugin.ts
 /**
 * @alfe.ai/openclaw-chat — OpenClaw chat channel plugin.
@@ -913,7 +1033,7 @@ function asString(v) {
 * Returns null when there's nothing to forward.
 */
 function buildThinkingActivity(data) {
-	const text = asString(data.delta) ?? asString(data.text) ?? asString(data.content) ?? asString(data.message?.text);
+	const text = asString(data.delta) ?? asString(data.text) ?? asString(data.content) ?? asString(data.reasoning) ?? asString(data.reasoning_content) ?? asString(data.message?.text);
 	if (!text) return null;
 	return {
 		kind: "thinking",
@@ -922,6 +1042,34 @@ function buildThinkingActivity(data) {
 	};
 }
 /**
+* Join the text parts of an assistant-stream event's content array.
+* Assistant `evt.data` is a message object whose `content` is an array of
+* typed parts; deltas are CUMULATIVE snapshots (the relay length-compares
+* them — see adapter.ts's lastSentLength slicing). Non-array/absent content
+* yields '' (nothing to divert).
+*/
+function joinAssistantText(data) {
+	const content = data.content;
+	if (!Array.isArray(content)) return "";
+	return content.filter((c) => typeof c === "object" && c !== null && c.type === "text" && typeof c.text === "string").map((c) => c.text).join("");
+}
+/**
+* Rebuild an assistant event's message with its text parts replaced by one
+* `{type:'text', text: visible}` part. Non-text parts are preserved — the
+* relay's consumers (adapter/inject/voice-bridge) only join text parts, so
+* relative ordering against non-text items is not load-bearing.
+*/
+function rewriteAssistantText(data, visible) {
+	const nonText = (Array.isArray(data.content) ? data.content : []).filter((c) => !(typeof c === "object" && c !== null && c.type === "text"));
+	return {
+		...data,
+		content: [...nonText, {
+			type: "text",
+			text: visible
+		}]
+	};
+}
+/**
 * Extract structured tool status from an OpenClaw `tool`-stream event.
 * We send the emoji-free raw `name` plus a lifecycle `status` and let the
 * web render the `🛠️ <Label>: <summary> ✓/✗` row — never OpenClaw's raw
@@ -1171,6 +1319,43 @@ async function handleAgentRequest(request, log) {
 		runId: null,
 		sessionKey: null
 	};
+	let lastThinkingLen = 0;
+	const turnThinking = {
+		text: "",
+		ts: 0
+	};
+	const turnTools = /* @__PURE__ */ new Map();
+	const recordThinking = (text, ts) => {
+		if (!turnThinking.text) turnThinking.ts = ts;
+		turnThinking.text += text;
+	};
+	const recordTool = (activity, ts) => {
+		const prev = turnTools.get(activity.toolCallId);
+		turnTools.set(activity.toolCallId, {
+			kind: "tool",
+			toolCallId: activity.toolCallId,
+			name: activity.name,
+			status: activity.status === "running" ? prev?.status ?? "interrupted" : activity.status,
+			...activity.summary ?? prev?.summary ? { summary: activity.summary ?? prev?.summary } : {},
+			...prev?.argsText ?? activity.argsText ? { argsText: prev?.argsText ?? activity.argsText } : {},
+			...activity.resultText ?? prev?.resultText ? { resultText: activity.resultText ?? prev?.resultText } : {},
+			...(activity.isError ?? prev?.isError) !== void 0 ? { isError: activity.isError ?? prev?.isError } : {},
+			...(activity.durationMs ?? prev?.durationMs) !== void 0 ? { durationMs: activity.durationMs ?? prev?.durationMs } : {},
+			...activity.truncated ?? prev?.truncated ? { truncated: activity.truncated ?? prev?.truncated } : {},
+			ts: prev?.ts ?? ts
+		});
+	};
+	const flushTurnActivity = async () => {
+		const entries = [];
+		if (turnThinking.text) entries.push({
+			kind: "thinking",
+			text: turnThinking.text,
+			ts: turnThinking.ts
+		});
+		for (const tool of turnTools.values()) entries.push(tool);
+		entries.sort((a, b) => a.ts - b.ts);
+		await appendActivity(sessionId, entries);
+	};
 	const UPDATE_THROTTLE_MS = 250;
 	const pendingUpdates = /* @__PURE__ */ new Map();
 	const flushToolUpdate = (toolCallId) => {
@@ -1187,33 +1372,58 @@ async function handleAgentRequest(request, log) {
 	const unsubscribe = runtime.events.onAgentEvent((evt) => {
 		if (!admitAgentEvent(runGate, evt)) return;
 		if (evt.stream === "assistant") {
+			const raw = joinAssistantText(evt.data);
+			const { visible, thinking } = splitThinkTags(raw);
+			if (thinking.length > lastThinkingLen) {
+				const thinkingDelta = thinking.slice(lastThinkingLen);
+				chatClient?.sendEvent("chat-activity", {
+					runId: evt.runId,
+					sessionKey: legacySessionKey,
+					seq: evt.seq,
+					state: "activity",
+					activity: {
+						kind: "thinking",
+						text: thinkingDelta,
+						delta: true,
+						ts: evt.ts,
+						seq: evt.seq
+					}
+				});
+				recordThinking(thinkingDelta, evt.ts);
+				lastThinkingLen = thinking.length;
+			}
+			const message = visible === raw && thinking.length === 0 ? evt.data : rewriteAssistantText(evt.data, visible);
 			chatClient?.sendEvent("chat", {
 				runId: evt.runId,
 				sessionKey: legacySessionKey,
 				seq: evt.seq,
 				state: "delta",
-				message: evt.data
+				message
 			});
 			return;
 		}
 		if (evt.stream === "thinking") {
 			const activity = buildThinkingActivity(evt.data);
-			if (activity) chatClient?.sendEvent("chat-activity", {
-				runId: evt.runId,
-				sessionKey: legacySessionKey,
-				seq: evt.seq,
-				state: "activity",
-				activity: {
-					...activity,
-					ts: evt.ts,
-					seq: evt.seq
-				}
-			});
+			if (activity) {
+				chatClient?.sendEvent("chat-activity", {
+					runId: evt.runId,
+					sessionKey: legacySessionKey,
+					seq: evt.seq,
+					state: "activity",
+					activity: {
+						...activity,
+						ts: evt.ts,
+						seq: evt.seq
+					}
+				});
+				if (activity.kind === "thinking") recordThinking(activity.text, evt.ts);
+			}
 			return;
 		}
 		if (evt.stream === "tool") {
 			const activity = buildToolActivity(evt.data);
 			if (activity?.kind !== "tool") return;
+			recordTool(activity, evt.ts);
 			const toolCallId = activity.toolCallId;
 			const send = () => {
 				chatClient?.sendEvent("chat-activity", {
@@ -1309,10 +1519,10 @@ async function handleAgentRequest(request, log) {
 				} : {}
 			},
 			deliver: async (payload) => {
-				const responseText = stripLeakedToolSummary(payload.text ?? "");
+				const responseText = stripLeakedToolSummary(stripThinkSpans(payload.text ?? ""));
 				const mediaUrls = [...payload.mediaUrl ? [payload.mediaUrl] : [], ...payload.mediaUrls ?? []];
 				await addMessage(sessionId, "assistant", responseText);
-				if (isA2A) a2aResponseBuffer += responseText;
+				if (isA2A) a2aResponseBuffer += (a2aResponseBuffer && responseText ? "\n\n" : "") + responseText;
 				chatClient?.notify("agent-message", {
 					conversationId: conversationId ?? legacySessionKey,
 					text: responseText,
@@ -1346,6 +1556,11 @@ async function handleAgentRequest(request, log) {
 	} finally {
 		unsubscribe();
 		clearAllToolUpdates();
+		try {
+			await flushTurnActivity();
+		} catch (err) {
+			log.error(`Activity persistence failed: ${err instanceof Error ? err.message : String(err)}`);
+		}
 		if (isA2A) disarmA2AEndSignal();
 	}
 }
@@ -1371,14 +1586,16 @@ async function handleSessionsGet(request, log) {
 		if (!session) {
 			chatClient?.sendResponse(request.id, true, {
 				ok: false,
-				error: "Session not found"
+				error: "Session not found",
+				messages: []
 			});
 			return;
 		}
 		if (params.userId && session.userId && session.userId !== params.userId) {
 			chatClient?.sendResponse(request.id, true, {
 				ok: false,
-				error: "Session not found"
+				error: "Session not found",
+				messages: []
 			});
 			return;
 		}
@@ -1392,7 +1609,8 @@ async function handleSessionsGet(request, log) {
 				role: m.role,
 				content: m.content,
 				timestamp: m.timestamp
-			}))
+			})),
+			activity: (session.activity ?? []).map((a) => ({ ...a }))
 		});
 	} catch (err) {
 		const errMsg = err instanceof Error ? err.message : String(err);
@@ -1500,11 +1718,13 @@ const plugin = {
 				const session = await getSession(params.sessionId);
 				if (!session) return {
 					ok: false,
-					error: "Session not found"
+					error: "Session not found",
+					messages: []
 				};
 				if (params.userId && session.userId && session.userId !== params.userId) return {
 					ok: false,
-					error: "Session not found"
+					error: "Session not found",
+					messages: []
 				};
 				return {
 					sessionId: session.sessionId,
@@ -1516,7 +1736,8 @@ const plugin = {
 						role: m.role,
 						content: m.content,
 						timestamp: m.timestamp
-					}))
+					})),
+					activity: (session.activity ?? []).map((a) => ({ ...a }))
 				};
 			});
 			gw.__alfeChatGatewayMethodsRegistered = true;
@@ -1598,12 +1819,24 @@ Object.defineProperty(exports, "createAlfeChannelPlugin", {
 		return createAlfeChannelPlugin;
 	}
 });
+Object.defineProperty(exports, "joinAssistantText", {
+	enumerable: true,
+	get: function() {
+		return joinAssistantText;
+	}
+});
 Object.defineProperty(exports, "plugin", {
 	enumerable: true,
 	get: function() {
 		return plugin;
 	}
 });
+Object.defineProperty(exports, "rewriteAssistantText", {
+	enumerable: true,
+	get: function() {
+		return rewriteAssistantText;
+	}
+});
 Object.defineProperty(exports, "stripLeakedToolSummary", {
 	enumerable: true,
 	get: function() {

package/dist/plugin2.d.cts

@@ -1,2 +1,2 @@
-import { a as plugin, i as buildToolActivity, n as admitAgentEvent, o as stripLeakedToolSummary, r as buildA2ACompletePayload, t as RunEventGate } from "./plugin.cjs";
-export { RunEventGate, admitAgentEvent, buildA2ACompletePayload, buildToolActivity, plugin as default, stripLeakedToolSummary };
+import { a as joinAssistantText, c as stripLeakedToolSummary, i as buildToolActivity, n as admitAgentEvent, o as plugin, r as buildA2ACompletePayload, s as rewriteAssistantText, t as RunEventGate } from "./plugin.cjs";
+export { RunEventGate, admitAgentEvent, buildA2ACompletePayload, buildToolActivity, plugin as default, joinAssistantText, rewriteAssistantText, stripLeakedToolSummary };

package/dist/plugin2.d.ts

@@ -1,2 +1,2 @@
-import { a as plugin, i as buildToolActivity, n as admitAgentEvent, o as stripLeakedToolSummary, r as buildA2ACompletePayload, t as RunEventGate } from "./plugin.js";
-export { RunEventGate, admitAgentEvent, buildA2ACompletePayload, buildToolActivity, plugin as default, stripLeakedToolSummary };
+import { a as joinAssistantText, c as stripLeakedToolSummary, i as buildToolActivity, n as admitAgentEvent, o as plugin, r as buildA2ACompletePayload, s as rewriteAssistantText, t as RunEventGate } from "./plugin.js";
+export { RunEventGate, admitAgentEvent, buildA2ACompletePayload, buildToolActivity, plugin as default, joinAssistantText, rewriteAssistantText, stripLeakedToolSummary };

package/dist/plugin2.js

@@ -344,6 +344,31 @@ async function addMessage(sessionId, role, content, senderId, senderName) {
 	writeLocks.set(sessionId, next);
 	await next;
 }
+const MAX_ACTIVITY_PER_SESSION = 500;
+const MAX_ACTIVITY_THINKING_CHARS = 16e3;
+/**
+* Append a turn's terminalized activity records. Called ONCE per turn (from
+* the dispatch `finally`) — never per-delta: thinking deltas are
+* high-frequency and each save rewrites the whole session file.
+*/
+async function appendActivity(sessionId, entries) {
+	if (entries.length === 0) return;
+	const next = (writeLocks.get(sessionId) ?? Promise.resolve()).then(async () => {
+		const session = await getSession(sessionId);
+		if (!session) return;
+		const capped = entries.map((e) => e.kind === "thinking" && e.text.length > MAX_ACTIVITY_THINKING_CHARS ? {
+			...e,
+			text: e.text.slice(0, MAX_ACTIVITY_THINKING_CHARS)
+		} : e);
+		const combined = [...session.activity ?? [], ...capped];
+		session.activity = combined.length > MAX_ACTIVITY_PER_SESSION ? combined.slice(combined.length - MAX_ACTIVITY_PER_SESSION) : combined;
+		await saveSession(session);
+	}).finally(() => {
+		if (writeLocks.get(sessionId) === next) writeLocks.delete(sessionId);
+	});
+	writeLocks.set(sessionId, next);
+	await next;
+}
 async function listSessions(filters, limit = 50) {
 	await ensureDir();
 	let files;
@@ -848,6 +873,101 @@ function extractResultText(result) {
 	return parts.join("\n");
 }
 //#endregion
+//#region src/think-tags.ts
+/**
+* Inline `<think>` span handling for assistant text.
+*
+* Reasoning models routed through the AI proxy (DeepSeek `deepseek-*`,
+* Zhipu `glm-*`, …) emit their reasoning INLINE as `<think>…</think>` inside
+* the normal content stream instead of a structured thinking channel. The
+* openclaw-chat plugin (0.4.0+) diverts those spans into thinking activity
+* frames at the source; this relay-side copy is the DEFENSIVE layer for
+* agents still running older plugin versions — leaked reasoning is dropped
+* here, never synthesized into activity (that's the plugin's job).
+*
+* KEEP IN SYNC (byte-identical): this module exists at BOTH
+*   `services/chat/src/lib/think-tags.ts`        (relay, defensive layer)
+*   `packages/openclaw-chat/src/think-tags.ts`   (plugin, at-source divert)
+* The published plugin cannot import service internals, so the module is
+* duplicated, same as the ChatActivity shape.
+*
+* Known accepted limitation: a literal `<think>` in legitimate prose or a
+* code fence is treated as reasoning and stripped/diverted.
+*/
+const OPEN_TAG = "<think>";
+const CLOSE_TAG = "</think>";
+/**
+* Split CUMULATIVE assistant text into the user-visible part and the
+* reasoning part.
+*
+* Invariant (load-bearing): for successive cumulative snapshots that grow by
+* appending, `visible` is monotonically append-only — `visible(n)` is always
+* a prefix of `visible(n + 1)`. The delta relays slice by a `lastSentLength`
+* offset into this string; if `visible` ever shrank or rewrote earlier
+* characters, clients would receive corrupted deltas. This holds because the
+* scan is deterministic left-to-right and a trailing PARTIAL tag (any strict
+* prefix of the next expected tag, e.g. `"<thi"` or `"</th"`) is HELD BACK
+* from both outputs until later text disambiguates it.
+*/
+function splitThinkTags(cumulative) {
+	let visible = "";
+	let thinking = "";
+	let inThink = false;
+	let i = 0;
+	const n = cumulative.length;
+	while (i < n) {
+		const tag = inThink ? CLOSE_TAG : OPEN_TAG;
+		const idx = cumulative.indexOf(tag, i);
+		if (idx !== -1) {
+			const chunk = cumulative.slice(i, idx);
+			if (inThink) thinking += chunk;
+			else visible += chunk;
+			inThink = !inThink;
+			i = idx + tag.length;
+			continue;
+		}
+		let rest = cumulative.slice(i);
+		const held = trailingPartialTagLength(rest, tag);
+		if (held > 0) rest = rest.slice(0, rest.length - held);
+		if (inThink) thinking += rest;
+		else visible += rest;
+		break;
+	}
+	return {
+		visible,
+		thinking
+	};
+}
+/** Length of the longest strict prefix of `tag` that is a suffix of `text`. */
+function trailingPartialTagLength(text, tag) {
+	const max = Math.min(text.length, tag.length - 1);
+	for (let len = max; len > 0; len--) if (text.endsWith(tag.slice(0, len))) return len;
+	return 0;
+}
+/**
+* Strip `<think>…</think>` spans (and an unclosed trailing `<think>…`) from
+* FINAL text, then tidy the whitespace the spans leave behind.
+*
+* Idempotent — `stripThinkSpans(stripThinkSpans(x)) === stripThinkSpans(x)`.
+* New plugin + new relay both strip, so the second pass must be a no-op; the
+* fixpoint loop also covers pathological inputs where removing one span
+* makes the surrounding fragments form a new one.
+*
+* NOT for sliced/streaming paths — the trim + blank-line collapse breaks
+* prefix alignment with `lastSentLength`; those paths use
+* `splitThinkTags(...).visible` instead.
+*/
+function stripThinkSpans(text) {
+	let out = text;
+	let prev;
+	do {
+		prev = out;
+		out = out.replace(/<think>[\s\S]*?<\/think>/g, "");
+	} while (out !== prev);
+	out = out.replace(/<think>[\s\S]*$/, "");
+	return out.replace(/\n{3,}/g, "\n\n").trim();
+}
+//#endregion
 //#region src/plugin.ts
 /**
 * @alfe.ai/openclaw-chat — OpenClaw chat channel plugin.
@@ -913,7 +1033,7 @@ function asString(v) {
 * Returns null when there's nothing to forward.
 */
 function buildThinkingActivity(data) {
-	const text = asString(data.delta) ?? asString(data.text) ?? asString(data.content) ?? asString(data.message?.text);
+	const text = asString(data.delta) ?? asString(data.text) ?? asString(data.content) ?? asString(data.reasoning) ?? asString(data.reasoning_content) ?? asString(data.message?.text);
 	if (!text) return null;
 	return {
 		kind: "thinking",
@@ -922,6 +1042,34 @@ function buildThinkingActivity(data) {
 	};
 }
 /**
+* Join the text parts of an assistant-stream event's content array.
+* Assistant `evt.data` is a message object whose `content` is an array of
+* typed parts; deltas are CUMULATIVE snapshots (the relay length-compares
+* them — see adapter.ts's lastSentLength slicing). Non-array/absent content
+* yields '' (nothing to divert).
+*/
+function joinAssistantText(data) {
+	const content = data.content;
+	if (!Array.isArray(content)) return "";
+	return content.filter((c) => typeof c === "object" && c !== null && c.type === "text" && typeof c.text === "string").map((c) => c.text).join("");
+}
+/**
+* Rebuild an assistant event's message with its text parts replaced by one
+* `{type:'text', text: visible}` part. Non-text parts are preserved — the
+* relay's consumers (adapter/inject/voice-bridge) only join text parts, so
+* relative ordering against non-text items is not load-bearing.
+*/
+function rewriteAssistantText(data, visible) {
+	const nonText = (Array.isArray(data.content) ? data.content : []).filter((c) => !(typeof c === "object" && c !== null && c.type === "text"));
+	return {
+		...data,
+		content: [...nonText, {
+			type: "text",
+			text: visible
+		}]
+	};
+}
+/**
 * Extract structured tool status from an OpenClaw `tool`-stream event.
 * We send the emoji-free raw `name` plus a lifecycle `status` and let the
 * web render the `🛠️ <Label>: <summary> ✓/✗` row — never OpenClaw's raw
@@ -1171,6 +1319,43 @@ async function handleAgentRequest(request, log) {
 		runId: null,
 		sessionKey: null
 	};
+	let lastThinkingLen = 0;
+	const turnThinking = {
+		text: "",
+		ts: 0
+	};
+	const turnTools = /* @__PURE__ */ new Map();
+	const recordThinking = (text, ts) => {
+		if (!turnThinking.text) turnThinking.ts = ts;
+		turnThinking.text += text;
+	};
+	const recordTool = (activity, ts) => {
+		const prev = turnTools.get(activity.toolCallId);
+		turnTools.set(activity.toolCallId, {
+			kind: "tool",
+			toolCallId: activity.toolCallId,
+			name: activity.name,
+			status: activity.status === "running" ? prev?.status ?? "interrupted" : activity.status,
+			...activity.summary ?? prev?.summary ? { summary: activity.summary ?? prev?.summary } : {},
+			...prev?.argsText ?? activity.argsText ? { argsText: prev?.argsText ?? activity.argsText } : {},
+			...activity.resultText ?? prev?.resultText ? { resultText: activity.resultText ?? prev?.resultText } : {},
+			...(activity.isError ?? prev?.isError) !== void 0 ? { isError: activity.isError ?? prev?.isError } : {},
+			...(activity.durationMs ?? prev?.durationMs) !== void 0 ? { durationMs: activity.durationMs ?? prev?.durationMs } : {},
+			...activity.truncated ?? prev?.truncated ? { truncated: activity.truncated ?? prev?.truncated } : {},
+			ts: prev?.ts ?? ts
+		});
+	};
+	const flushTurnActivity = async () => {
+		const entries = [];
+		if (turnThinking.text) entries.push({
+			kind: "thinking",
+			text: turnThinking.text,
+			ts: turnThinking.ts
+		});
+		for (const tool of turnTools.values()) entries.push(tool);
+		entries.sort((a, b) => a.ts - b.ts);
+		await appendActivity(sessionId, entries);
+	};
 	const UPDATE_THROTTLE_MS = 250;
 	const pendingUpdates = /* @__PURE__ */ new Map();
 	const flushToolUpdate = (toolCallId) => {
@@ -1187,33 +1372,58 @@ async function handleAgentRequest(request, log) {
 	const unsubscribe = runtime.events.onAgentEvent((evt) => {
 		if (!admitAgentEvent(runGate, evt)) return;
 		if (evt.stream === "assistant") {
+			const raw = joinAssistantText(evt.data);
+			const { visible, thinking } = splitThinkTags(raw);
+			if (thinking.length > lastThinkingLen) {
+				const thinkingDelta = thinking.slice(lastThinkingLen);
+				chatClient?.sendEvent("chat-activity", {
+					runId: evt.runId,
+					sessionKey: legacySessionKey,
+					seq: evt.seq,
+					state: "activity",
+					activity: {
+						kind: "thinking",
+						text: thinkingDelta,
+						delta: true,
+						ts: evt.ts,
+						seq: evt.seq
+					}
+				});
+				recordThinking(thinkingDelta, evt.ts);
+				lastThinkingLen = thinking.length;
+			}
+			const message = visible === raw && thinking.length === 0 ? evt.data : rewriteAssistantText(evt.data, visible);
 			chatClient?.sendEvent("chat", {
 				runId: evt.runId,
 				sessionKey: legacySessionKey,
 				seq: evt.seq,
 				state: "delta",
-				message: evt.data
+				message
 			});
 			return;
 		}
 		if (evt.stream === "thinking") {
 			const activity = buildThinkingActivity(evt.data);
-			if (activity) chatClient?.sendEvent("chat-activity", {
-				runId: evt.runId,
-				sessionKey: legacySessionKey,
-				seq: evt.seq,
-				state: "activity",
-				activity: {
-					...activity,
-					ts: evt.ts,
-					seq: evt.seq
-				}
-			});
+			if (activity) {
+				chatClient?.sendEvent("chat-activity", {
+					runId: evt.runId,
+					sessionKey: legacySessionKey,
+					seq: evt.seq,
+					state: "activity",
+					activity: {
+						...activity,
+						ts: evt.ts,
+						seq: evt.seq
+					}
+				});
+				if (activity.kind === "thinking") recordThinking(activity.text, evt.ts);
+			}
 			return;
 		}
 		if (evt.stream === "tool") {
 			const activity = buildToolActivity(evt.data);
 			if (activity?.kind !== "tool") return;
+			recordTool(activity, evt.ts);
 			const toolCallId = activity.toolCallId;
 			const send = () => {
 				chatClient?.sendEvent("chat-activity", {
@@ -1309,10 +1519,10 @@ async function handleAgentRequest(request, log) {
 				} : {}
 			},
 			deliver: async (payload) => {
-				const responseText = stripLeakedToolSummary(payload.text ?? "");
+				const responseText = stripLeakedToolSummary(stripThinkSpans(payload.text ?? ""));
 				const mediaUrls = [...payload.mediaUrl ? [payload.mediaUrl] : [], ...payload.mediaUrls ?? []];
 				await addMessage(sessionId, "assistant", responseText);
-				if (isA2A) a2aResponseBuffer += responseText;
+				if (isA2A) a2aResponseBuffer += (a2aResponseBuffer && responseText ? "\n\n" : "") + responseText;
 				chatClient?.notify("agent-message", {
 					conversationId: conversationId ?? legacySessionKey,
 					text: responseText,
@@ -1346,6 +1556,11 @@ async function handleAgentRequest(request, log) {
 	} finally {
 		unsubscribe();
 		clearAllToolUpdates();
+		try {
+			await flushTurnActivity();
+		} catch (err) {
+			log.error(`Activity persistence failed: ${err instanceof Error ? err.message : String(err)}`);
+		}
 		if (isA2A) disarmA2AEndSignal();
 	}
 }
@@ -1371,14 +1586,16 @@ async function handleSessionsGet(request, log) {
 		if (!session) {
 			chatClient?.sendResponse(request.id, true, {
 				ok: false,
-				error: "Session not found"
+				error: "Session not found",
+				messages: []
 			});
 			return;
 		}
 		if (params.userId && session.userId && session.userId !== params.userId) {
 			chatClient?.sendResponse(request.id, true, {
 				ok: false,
-				error: "Session not found"
+				error: "Session not found",
+				messages: []
 			});
 			return;
 		}
@@ -1392,7 +1609,8 @@ async function handleSessionsGet(request, log) {
 				role: m.role,
 				content: m.content,
 				timestamp: m.timestamp
-			}))
+			})),
+			activity: (session.activity ?? []).map((a) => ({ ...a }))
 		});
 	} catch (err) {
 		const errMsg = err instanceof Error ? err.message : String(err);
@@ -1500,11 +1718,13 @@ const plugin = {
 				const session = await getSession(params.sessionId);
 				if (!session) return {
 					ok: false,
-					error: "Session not found"
+					error: "Session not found",
+					messages: []
 				};
 				if (params.userId && session.userId && session.userId !== params.userId) return {
 					ok: false,
-					error: "Session not found"
+					error: "Session not found",
+					messages: []
 				};
 				return {
 					sessionId: session.sessionId,
@@ -1516,7 +1736,8 @@ const plugin = {
 						role: m.role,
 						content: m.content,
 						timestamp: m.timestamp
-					}))
+					})),
+					activity: (session.activity ?? []).map((a) => ({ ...a }))
 				};
 			});
 			gw.__alfeChatGatewayMethodsRegistered = true;
@@ -1574,4 +1795,4 @@ const plugin = {
 	}
 };
 //#endregion
-export { stripLeakedToolSummary as a, plugin as i, buildA2ACompletePayload as n, createAlfeChannelPlugin as o, buildToolActivity as r, admitAgentEvent as t };
+export { plugin as a, createAlfeChannelPlugin as c, joinAssistantText as i, buildA2ACompletePayload as n, rewriteAssistantText as o, buildToolActivity as r, stripLeakedToolSummary as s, admitAgentEvent as t };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@alfe.ai/openclaw-chat",
-  "version": "0.3.3",
+  "version": "0.4.0",
   "description": "OpenClaw chat plugin for Alfe — web widget and mobile app channels",
   "type": "module",
   "main": "./dist/plugin.js",
@@ -27,8 +27,8 @@
     "openclaw.plugin.json"
   ],
   "dependencies": {
-    "@alfe.ai/agent-api-client": "^0.5.0",
     "@alfe.ai/chat": "^0.0.14",
+    "@alfe.ai/agent-api-client": "^0.5.0",
     "@alfe.ai/config": "^0.3.0"
   },
   "peerDependencies": {