dsh-context 0.40.1 → 0.41.0

package/lib/client.js

@@ -37,10 +37,11 @@ window.__ModuleLoader__.load({
 			"timing.title": "耗时统计",
 			"timing.hint": "整个会话的活跃时长分布",
 			"timing.total": "总活跃时长",
-			"timing.lm": "模型调用",
+			"timing.ttft": "模型等待",
+			"timing.gen": "模型生成",
 			"timing.tools": "工具执行",
 			"timing.other": "其他开销",
-			"timing.lmTimes": "{n}次",
+			"timing.callTimes": "{n}次",
 			"timing.toolTimes": "{n}次",
 			"timing.empty": "暂无耗时数据 · 对话开始后这里会显示时长分布",
 			"tokens.title": "Token 统计",
@@ -312,10 +313,11 @@ window.__ModuleLoader__.load({
 			"timing.title": "Timing Stats",
 			"timing.hint": "Whole-session active-time split",
 			"timing.total": "Active Time",
-			"timing.lm": "Model calls",
+			"timing.ttft": "TTFT",
+			"timing.gen": "LLM Gen",
 			"timing.tools": "Tool runs",
 			"timing.other": "Overhead",
-			"timing.lmTimes": "{n} calls",
+			"timing.callTimes": "{n} calls",
 			"timing.toolTimes": "{n} runs",
 			"timing.empty": "No timing data yet — durations appear as the conversation runs",
 			"tokens.title": "Token Stats",
@@ -997,7 +999,8 @@ window.__ModuleLoader__.load({
 			const t = value;
 			for (const k of [
 				"wallMs",
-				"lmMs",
+				"ttftMs",
+				"genMs",
 				"calls",
 				"toolsMs",
 				"toolCalls"
@@ -1038,7 +1041,8 @@ window.__ModuleLoader__.load({
 			}
 			return {
 				wallMs: msNumOf(data.wallMs),
-				lmMs: msNumOf(data.lmMs),
+				ttftMs: msNumOf(data.ttftMs),
+				genMs: msNumOf(data.genMs),
 				calls: msNumOf(data.calls),
 				toolsMs: msNumOf(data.toolsMs),
 				toolCalls: msNumOf(data.toolCalls),
@@ -5059,7 +5063,7 @@ window.__ModuleLoader__.load({
 			return function PluginInfo() {
 				const [latest, setLatest] = React.useState(null);
 				React.useEffect(() => {
-					if ("0.40.1".includes("-dev")) return;
+					if ("0.41.0".includes("-dev")) return;
 					let on = true;
 					fetchLatestVersion().then((v) => {
 						if (on && v) setLatest(v);
@@ -5068,8 +5072,8 @@ window.__ModuleLoader__.load({
 						on = false;
 					};
 				}, []);
-				const update = latest !== null && isNewerVersion(latest, "0.40.1") ? latest : null;
-				const nameText = "dsh-context (v0.40.1)";
+				const update = latest !== null && isNewerVersion(latest, "0.41.0") ? latest : null;
+				const nameText = "dsh-context (v0.41.0)";
 				const nameValue = [nameText];
 				if (update) nameValue.push(/* @__PURE__ */ React.createElement("span", {
 					key: "update",
@@ -5477,20 +5481,27 @@ window.__ModuleLoader__.load({
 				let segments = [];
 				let rows = [];
 				if (timing !== null && (wall > 0 || timing.calls > 0 || timing.toolCalls > 0)) {
-					const lm = Math.min(timing.lmMs, wall);
-					const toolRing = Math.max(0, Math.min(timing.toolsMs, wall - lm));
-					const other = Math.max(0, wall - lm - toolRing);
+					const ttft = Math.min(timing.ttftMs, wall);
+					const gen = Math.max(0, Math.min(timing.genMs, wall - ttft));
+					const toolRing = Math.max(0, Math.min(timing.toolsMs, wall - ttft - gen));
+					const other = Math.max(0, wall - ttft - gen - toolRing);
 					const share = (ms) => wall > 0 ? ms / wall : 0;
 					const countOf = (ms, times) => {
 						const dur = fmtDuration(ms, lang);
 						if (times === void 0) return dur;
 						return ms > 0 ? `${dur} · ${times}` : times;
 					};
+					const callTimes = t("timing.callTimes", { n: fmt(timing.calls) });
 					segments = [
 						{
-							key: "lm",
+							key: "ttft",
 							color: "#3b82f6",
-							value: share(lm)
+							value: share(ttft)
+						},
+						{
+							key: "gen",
+							color: "#8b5cf6",
+							value: share(gen)
 						},
 						{
 							key: "tools",
@@ -5505,12 +5516,20 @@ window.__ModuleLoader__.load({
 					];
 					rows = [
 						{
-							key: "lm",
+							key: "ttft",
 							color: "#3b82f6",
-							label: t("timing.lm"),
-							dim: timing.lmMs === 0,
-							pct: fmtShare(timing.lmMs, wall),
-							count: countOf(timing.lmMs, t("timing.lmTimes", { n: fmt(timing.calls) }))
+							label: t("timing.ttft"),
+							dim: timing.ttftMs === 0,
+							pct: fmtShare(timing.ttftMs, wall),
+							count: countOf(timing.ttftMs, callTimes)
+						},
+						{
+							key: "gen",
+							color: "#8b5cf6",
+							label: t("timing.gen"),
+							dim: timing.genMs === 0,
+							pct: fmtShare(timing.genMs, wall),
+							count: countOf(timing.genMs, callTimes)
 						},
 						{
 							key: "tools",
@@ -6431,7 +6450,7 @@ window.__ModuleLoader__.load({
 		}
 		//#endregion
 		//#region \0dsh-global-css:/home/runner/work/dsh-context/dsh-context/src/client/styles/base.css.mjs
-		const css$13 = ".lc-root{box-sizing:border-box;height:100%;color:var(--dsw-alias-label-primary);padding:16px 20px 32px;font-size:13px;overflow-y:auto}.lc-card{background:var(--dsw-alias-bg-layer-1);border:1px solid var(--dsw-alias-border-l2);border-radius:10px;margin-bottom:14px;padding:14px 16px}.lc-card-title{flex-wrap:wrap;align-items:baseline;gap:8px;margin-bottom:10px;font-weight:600;display:flex}.lc-card-title-text{white-space:nowrap;flex:none}.lc-gran,.lc-kinds{background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l2);border-radius:6px;gap:2px;margin-left:auto;padding:1px;display:flex}.lc-trend-ctl{align-items:center;gap:6px;margin-left:auto;display:flex}.lc-trend-ctl .lc-gran{margin-left:0}.lc-gran-btn{color:var(--dsw-alias-label-secondary);cursor:pointer;transition:color var(--ds-transition-duration,.2s) var(--ds-ease-in-out,ease-in-out);background:0 0;border:0;border-radius:5px;padding:3px 8px;font-family:inherit;font-size:11px;line-height:1}.lc-gran-btn:hover{color:var(--dsw-alias-label-primary)}.lc-gran-on,.lc-gran-on:hover{background:var(--dsw-alias-button-primary-fill);color:var(--dsw-alias-label-primary-foreground)}.lc-tip{z-index:6;background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l1);width:max-content;color:var(--dsw-alias-label-primary);box-shadow:var(--dsw-shadow-lv3,0 2px 8px #0000002e);pointer-events:none;opacity:0;transition:opacity var(--ds-transition-duration,.2s) var(--ds-ease-in-out,ease-in-out);border-radius:6px;padding:6px 10px;font-size:12px;position:absolute}.lc-events,.lc-fa-list{flex-direction:column;gap:2px;height:320px;display:flex;overflow-y:auto}.lc-cols{flex-wrap:wrap;gap:14px;margin-bottom:14px;display:flex}.lc-col{flex:1;min-width:280px}.lc-cols>.lc-card,.lc-col>.lc-card:last-child{margin-bottom:0}.lc-col-browser{flex-direction:column;display:flex}.lc-col-browser>.lc-card{flex:1}.lc-head>.lc-card{flex:1 1 0;min-width:0;container:lc-head-card/inline-size}.lc-head>.lc-card:first-child{flex-direction:column;display:flex}.lc-head>.lc-card:first-child .lc-stats{flex:1;align-content:stretch}.lc-head>.lc-card:first-child .lc-stat{justify-content:center}.lc-empty{color:var(--dsw-alias-label-secondary);text-align:center;padding:18px 0}.lc-error{flex-direction:column;align-items:center;gap:8px;padding:40px 16px;display:flex}.lc-error-msg{font-family:var(--ds-font-family-code,ui-monospace, SFMono-Regular, Menlo, monospace);color:var(--dsw-alias-state-error-primary);background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l1);overflow-wrap:anywhere;border-radius:6px;max-width:100%;padding:4px 8px;font-size:12px}.lc-error-retry{border:1px solid var(--dsw-alias-border-l2);background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-primary);font:inherit;cursor:pointer;transition:border-color var(--ds-transition-duration,.2s) var(--ds-ease-in-out,ease-in-out);border-radius:6px;padding:4px 14px;font-size:12px}.lc-error-retry:hover{border-color:var(--dsw-alias-label-primary)}.lc-foot{color:var(--dsw-alias-label-secondary);margin-top:4px;font-size:12px}[data-conversation-scroll]:has(.lc-root)>[data-composer-seat]:not(:has([data-approval-key],[data-question-key],[data-plan-review-key])){display:none}";
+		const css$13 = ".lc-root{box-sizing:border-box;height:100%;color:var(--dsw-alias-label-primary);padding:16px 20px 32px;font-size:13px;overflow-y:auto}.lc-card{background:var(--dsw-alias-bg-layer-1);border:1px solid var(--dsw-alias-border-l2);border-radius:10px;margin-bottom:14px;padding:14px 16px}.lc-root .lc-card{margin-bottom:7px;padding:7px 8px}.lc-root .lc-cols{gap:7px;margin-bottom:7px}.lc-card-title{flex-wrap:wrap;align-items:baseline;gap:8px;margin-bottom:10px;font-weight:600;display:flex}.lc-card-title-text{white-space:nowrap;flex:none}.lc-gran,.lc-kinds{background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l2);border-radius:6px;gap:2px;margin-left:auto;padding:1px;display:flex}.lc-trend-ctl{align-items:center;gap:6px;margin-left:auto;display:flex}.lc-trend-ctl .lc-gran{margin-left:0}.lc-gran-btn{color:var(--dsw-alias-label-secondary);cursor:pointer;transition:color var(--ds-transition-duration,.2s) var(--ds-ease-in-out,ease-in-out);background:0 0;border:0;border-radius:5px;padding:3px 8px;font-family:inherit;font-size:11px;line-height:1}.lc-gran-btn:hover{color:var(--dsw-alias-label-primary)}.lc-gran-on,.lc-gran-on:hover{background:var(--dsw-alias-button-primary-fill);color:var(--dsw-alias-label-primary-foreground)}.lc-tip{z-index:6;background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l1);width:max-content;color:var(--dsw-alias-label-primary);box-shadow:var(--dsw-shadow-lv3,0 2px 8px #0000002e);pointer-events:none;opacity:0;transition:opacity var(--ds-transition-duration,.2s) var(--ds-ease-in-out,ease-in-out);border-radius:6px;padding:6px 10px;font-size:12px;position:absolute}.lc-events,.lc-fa-list{flex-direction:column;gap:2px;height:320px;display:flex;overflow-y:auto}.lc-cols{flex-wrap:wrap;gap:14px;margin-bottom:14px;display:flex}.lc-col{flex:1;min-width:280px}.lc-cols>.lc-card,.lc-col>.lc-card:last-child{margin-bottom:0}.lc-col-browser{flex-direction:column;display:flex}.lc-col-browser>.lc-card{flex:1}.lc-head>.lc-card{flex:1 1 0;min-width:0;container:lc-head-card/inline-size}.lc-head>.lc-card:first-child{flex-direction:column;display:flex}.lc-head>.lc-card:first-child .lc-stats{flex:1;align-content:stretch}.lc-head>.lc-card:first-child .lc-stat{justify-content:center}.lc-empty{color:var(--dsw-alias-label-secondary);text-align:center;padding:18px 0}.lc-error{flex-direction:column;align-items:center;gap:8px;padding:40px 16px;display:flex}.lc-error-msg{font-family:var(--ds-font-family-code,ui-monospace, SFMono-Regular, Menlo, monospace);color:var(--dsw-alias-state-error-primary);background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l1);overflow-wrap:anywhere;border-radius:6px;max-width:100%;padding:4px 8px;font-size:12px}.lc-error-retry{border:1px solid var(--dsw-alias-border-l2);background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-primary);font:inherit;cursor:pointer;transition:border-color var(--ds-transition-duration,.2s) var(--ds-ease-in-out,ease-in-out);border-radius:6px;padding:4px 14px;font-size:12px}.lc-error-retry:hover{border-color:var(--dsw-alias-label-primary)}.lc-foot{color:var(--dsw-alias-label-secondary);margin-top:4px;font-size:12px}[data-conversation-scroll]:has(.lc-root)>[data-composer-seat]:not(:has([data-approval-key],[data-question-key],[data-plan-review-key])){display:none}";
 		const tagId$13 = "dsh-context/base.css";
 		if (typeof document !== "undefined" && document.querySelector("style[data-plugin-css=" + JSON.stringify(tagId$13) + "]") === null) {
 			const tag = document.createElement("style");

package/lib/index.d.ts

@@ -112,14 +112,18 @@ interface TimelineState {
   timing?: TimingTotals;
   /**
    * The open step's start instant, armed by `step/start` and consumed by the
-   * `assistant/message` (LM-call time) and `step/end` (wall time) that follow
-   * it. One slot, not a map: steps are sequential in the log, so the newest
-   * `step/start` is the one those events close — a hostile interleaved log
-   * degrades to skipped durations, never to unbounded state. Same
-   * arm/remove lifecycle as `pendingShadowedSeqs`.
+   * `assistant/message` (TTFT/generation split) and `step/end` (wall time)
+   * that follow it; `assistant/chunk` stamps `firstToken` on the step's first
+   * token delta — absent when the stream carried none (legacy or aborted
+   * steps), which leaves that call's model time unattributed. One slot, not a
+   * map: steps are sequential in the log, so the newest `step/start` is the
+   * one those events close — a hostile interleaved log degrades to skipped
+   * durations, never to unbounded state. Same arm/remove lifecycle as
+   * `pendingShadowedSeqs`.
    */
   stepStart?: {
     time: number;
+    firstToken?: number;
   };
   /**
    * Tool callId → the call's name and start instant, armed by `tool/call` and
@@ -288,18 +292,22 @@ interface ToolTimingTotals {
 }
 /**
  * Whole-session timing totals, host-folded from the durable `step/start` /
- `step/end` / `tool/call` / `tool/result` lifecycle (running totals over the
- * COMPLETE session log — the same never-trimmed framing as `cost`). Durations
- * are wall-clock milliseconds: `wallMs` sums whole steps, `lmMs` the
- * step-start → assistant-message slice (the model call), `toolsMs` the sum of
- * per-call tool durations (parallel calls each count, so it can overlap).
+ `step/end` / `assistant/chunk` / `tool/call` / `tool/result` lifecycle
+ * (running totals over the COMPLETE session log — the same never-trimmed
+ * framing as `cost`). Durations are wall-clock milliseconds: `wallMs` sums
+ * whole steps, `ttftMs` the step-start → first-token slice (the model wait)
+ * and `genMs` the first-token → assistant-message slice (the generation) —
+ * both only over calls whose stream carried a token delta, `toolsMs` the sum
+ * of per-call tool durations (parallel calls each count, so it can overlap).
  * Absent until the first step lifecycle completes in the log.
  */
 interface TimingTotals {
   /** Summed wall time of completed steps (the session's active time). */
   wallMs: number;
-  /** Summed step-start → assistant-message time (model-call time). */
-  lmMs: number;
+  /** Summed step-start → first-token time (the model wait, TTFT). */
+  ttftMs: number;
+  /** Summed first-token → assistant-message time (the generation). */
+  genMs: number;
   /** Completed model calls (assistant messages folded). */
   calls: number;
   /** Summed per-call durations of completed tool calls. */

package/lib/index.js

@@ -993,6 +993,21 @@ function durOf(from, to) {
 	return Math.max(0, to - from);
 }
 /**
+* Whether a stream chunk carries a non-empty token delta — the first-token
+* marker the TTFT fold waits for (the same rule as the harness's own
+* session-stats fold). Shape-guarded: a malformed chunk is just not a token.
+*/
+function isTokenDelta(chunk) {
+	if (chunk === null || typeof chunk !== "object") return false;
+	const c = chunk;
+	switch (c.type) {
+		case "text-delta":
+		case "reasoning-delta": return typeof c.text === "string" && c.text !== "";
+		case "tool-call-delta": return typeof c.argumentsDelta === "string" && c.argumentsDelta !== "" || c.name !== void 0;
+		default: return false;
+	}
+}
+/**
 * The fold's private timing accumulator: created on first use, and CLONED on
 * every later ensure() (see `applyTimeline`) — the object left in the
 * persisted previous state is never written into in place.
@@ -1000,7 +1015,8 @@ function durOf(from, to) {
 function ensureTiming(st) {
 	if (st.timing === void 0) st.timing = {
 		wallMs: 0,
-		lmMs: 0,
+		ttftMs: 0,
+		genMs: 0,
 		calls: 0,
 		toolsMs: 0,
 		toolCalls: 0,
@@ -1091,6 +1107,17 @@ function applyTimeline(state, event, bounds) {
 					};
 				}
 				break;
+			case "assistant/chunk": {
+				const start = state.stepStart;
+				if (start === void 0 || start.firstToken !== void 0) return state;
+				if (!isTokenDelta(data?.chunk)) return state;
+				const s = ensure();
+				s.stepStart = {
+					time: start.time,
+					firstToken: event.time
+				};
+				break;
+			}
 			case "step/start": {
 				const s = ensure();
 				s.stepStart = { time: event.time };
@@ -1177,7 +1204,10 @@ function applyTimeline(state, event, bounds) {
 				const timing = ensureTiming(s);
 				timing.calls += 1;
 				const stepStart = state.stepStart;
-				if (stepStart !== void 0) timing.lmMs += durOf(stepStart.time, event.time);
+				if (stepStart !== void 0 && stepStart.firstToken !== void 0) {
+					timing.ttftMs += durOf(stepStart.time, stepStart.firstToken);
+					timing.genMs += durOf(stepStart.firstToken, event.time);
+				}
 				const asstMsg = deriveEventMessage(event);
 				applySurface(s, event, event.type, data, asstMsg);
 				break;
@@ -1404,7 +1434,8 @@ const toolTimingSchema = z.object({
 }).strict();
 const timingTotalsSchema = z.object({
 	wallMs: z.number().nonnegative(),
-	lmMs: z.number().nonnegative(),
+	ttftMs: z.number().nonnegative(),
+	genMs: z.number().nonnegative(),
 	calls: z.number().int().nonnegative(),
 	toolsMs: z.number().nonnegative(),
 	toolCalls: z.number().int().nonnegative(),
@@ -1460,7 +1491,10 @@ const timelineStateSchema = z.object({
 	}).strict().optional(),
 	archiveFloor: z.number().optional(),
 	timing: timingTotalsSchema.optional(),
-	stepStart: z.object({ time: z.number() }).strict().optional(),
+	stepStart: z.object({
+		time: z.number(),
+		firstToken: z.number().optional()
+	}).strict().optional(),
 	callNames: z.record(z.string(), z.object({
 		name: z.string(),
 		start: z.number()
@@ -1499,7 +1533,7 @@ function createContextTimelineDefinition(config) {
 		},
 		init: () => createTimelineState(),
 		apply: (state, event) => applyTimeline(state, event, bounds),
-		stateVersion: 11
+		stateVersion: 12
 	};
 }
 //#endregion

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-context",
-  "version": "0.40.1",
+  "version": "0.41.0",
   "description": "A DeepSeek Harness plugin for context insight and management, with context dashboard and context command, for understanding how the context is made of, and how it evolves.",
   "author": "bowenliang123",
   "repository": {