agentus 0.8.1 → 0.10.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentus",
-  "version": "0.8.1",
+  "version": "0.10.0",
   "description": "Multi-session web cockpit for ACP-speaking coding agents (Qoder CLI / Hermes). UI has no intelligence — the agent lives in spawned CLI subprocesses.",
   "license": "Apache-2.0",
   "repository": {
@@ -83,7 +83,9 @@
     "resume-tail": "tsx scripts/qa/resume-tail.mts",
     "render-lag-sweep": "node scripts/qa/render-lag-sweep.mjs",
     "replay-window": "node scripts/qa/replay-window.mjs",
-    "think-cadence": "node scripts/qa/think-cadence.mjs"
+    "think-cadence": "node scripts/qa/think-cadence.mjs",
+    "vad": "tsx scripts/qa/vad.mts",
+    "vad-silero": "tsx scripts/qa/vad-silero.mts"
   },
   "engines": {
     "node": ">=22.5"

package/packages/server/src/acp/session-manager.ts

@@ -49,7 +49,7 @@ import type {
 } from "../../../shared/src/index.js";
 import { acceptsToolWrite, acceptsWriteFrom, clampExplanation, clampTitle, demoteInProgress, foldPlanUpdate, hasUnfinished, normalizeItems, v2PlanEnabled } from "../plan/plan.js";
 import { PLAN_GUIDANCE_ACTIVE, PLAN_GUIDANCE_INTRO, PLAN_GUIDANCE_NATIVE } from "../../../shared/src/index.js";
-import { PREFS_DEFAULT, agentSettings, type PlanHintSettings } from "../settings.js";
+import { AGENT_DEFAULT, PREFS_DEFAULT, agentSettings, type PlanHintSettings } from "../settings.js";
 
 // The cockpit CLI an agent is handed (issue #15). Resolved from THIS module's location, so it points
 // at the same tree the server runs from — a checkout and an installed package alike (`bin/` is in the
@@ -1334,12 +1334,18 @@ export class SessionManager {
    *  these three. Rendering it into a prompt, an `echo`, or a `ps` line is the one thing the CLI's own
    *  `capabilities` forbids, and the token is worth nothing once this process is gone. */
   #cockpitEnv(env: NodeJS.ProcessEnv, sessionId: string, token: string): NodeJS.ProcessEnv {
+    // 召回条目的两个上限跟着设置一起下发给 agent（issue #37 的容量）：provider 读这两个环境变量，
+    // 与 server 的 `statusItemsOf` 同源。只在这一处传，是因为「一轮发多少字」是**宿主**的存储决定，
+    // 不该要求用户去改 agent 自己的 config.json，也不该让两边各有一套默认值。
+    const agent = agentSettings();
     return {
       ...env,
       AGENTUS_COCKPIT_ENDPOINT: `http://127.0.0.1:${process.env.AGENTUS_PORT ?? 8788}`,
       AGENTUS_COCKPIT_TOKEN: token,
       AGENTUS_COCKPIT_SESSION: sessionId,
       AGENTUS_COCKPIT_BIN: COCKPIT_BIN,
+      HINDSIGHT_RECALL_ITEM_CHARS: String(agent.recallItemChars || AGENT_DEFAULT.recallItemChars),
+      HINDSIGHT_RECALL_TOTAL_CHARS: String(agent.recallTotalChars || AGENT_DEFAULT.recallTotalChars),
     };
   }
 
@@ -2765,21 +2771,38 @@ export function visibleBackgroundTasks(tasks: BackgroundTask[]): BackgroundTask[
  *
  *  形状（fork `acp_adapter/events.py:make_status_cb`）：与状态句同一条 thought 帧，
  *  `items: [{text, tags?, type?}]`、`truncated` 说明 agent 侧已经截断过。**在这里限一次量**
- *  是因为这串东西来自 agent 进程、还要落库：provider 已经按 200 字符/条、6K/轮截过，但那是
- *  agent 的承诺，不是我们的保证。上限与前端展开的容量对齐，绝不把一帧 200KB 的 memory 写进 SQLite。
- */
-export const NOTE_ITEMS_MAX = 20;
-export const NOTE_ITEM_CHARS = 300;
+ *  是因为这串东西来自 agent 进程、还要落库：provider 也会截，但那是 agent 的承诺，不是我们的保证，
+ *  上限与前端展开的容量对齐，绝不把一帧 200KB 的 memory 写进 SQLite。
+ *
+ *  闸门是**总字符数**，不是条数（操作者 2026-10-21：「条数不限制，按照 12K 总数来」）：按条数截会
+ *  在预算还没用完时就把整条短记忆丢掉，而真正花 wire 和 SQLite 的是字节。两个数都可配
+ *  （`agent.recallItemChars` / `agent.recallTotalChars`），并且以环境变量交给 agent 侧的 provider，
+ *  让两边同源 —— 否则界面说 12K、agent 按 6K 发，用户永远对不上。 */
+export const NOTE_ITEM_CHARS = AGENT_DEFAULT.recallItemChars;
+export const NOTE_TOTAL_CHARS = AGENT_DEFAULT.recallTotalChars;
 
 export function statusItemsOf(u: Record<string, unknown>): { items: StatusNoteItem[]; truncated: boolean } | undefined {
   const raw = hermesPayload(u).items;
   if (!Array.isArray(raw)) return undefined;
+  const caps = agentSettings();
+  const itemCap = caps.recallItemChars > 0 ? caps.recallItemChars : NOTE_ITEM_CHARS;
+  const totalCap = caps.recallTotalChars > 0 ? caps.recallTotalChars : NOTE_TOTAL_CHARS;
   const items: StatusNoteItem[] = [];
+  let used = 0;
+  let cut = false;
   for (const entry of raw) {
-    if (items.length >= NOTE_ITEMS_MAX) break;
     const text = typeof entry === "object" && entry ? (entry as { text?: unknown }).text : undefined;
     if (typeof text !== "string" || !text.trim()) continue;
-    const item: StatusNoteItem = { text: text.trim().slice(0, NOTE_ITEM_CHARS) };
+    let body = text.trim();
+    if (body.length > itemCap) {
+      body = body.slice(0, itemCap);
+      cut = true;
+    }
+    if (used + body.length > totalCap) {
+      cut = true;
+      break;
+    }
+    const item: StatusNoteItem = { text: body };
     const tags = (entry as { tags?: unknown }).tags;
     if (Array.isArray(tags)) {
       const clean = tags.filter((t): t is string => typeof t === "string" && !!t).slice(0, 8);
@@ -2788,9 +2811,10 @@ export function statusItemsOf(u: Record<string, unknown>): { items: StatusNoteIt
     const kind = (entry as { type?: unknown }).type;
     if (typeof kind === "string" && kind) item.type = kind;
     items.push(item);
+    used += body.length;
   }
   if (!items.length) return undefined;
-  return { items, truncated: hermesPayload(u).truncated === true || raw.length > items.length };
+  return { items, truncated: cut || hermesPayload(u).truncated === true || raw.length > items.length };
 }
 
 function statusKindOf(u: Record<string, unknown>): string | undefined {

package/packages/server/src/settings.ts

@@ -62,7 +62,7 @@ export const THEME_DEFAULT: ThemeSettings = { mode: "system", accent: "" };
  *  what a phone's own loudspeaker leaks back into its microphone is nothing like a headset
  *  in a quiet room — so the defaults are a starting point, not a truth. The call page owns
  *  the panel; the settings page reads the same numbers. */
-export type InterruptMode = "voice" | "button";
+export type InterruptMode = "voice" | "button" | "keyword";
 
 export interface CallSettings {
   /** WHO gets to interrupt the agent, and the operator picks:
@@ -85,6 +85,10 @@ export interface CallSettings {
    *  as a turn (1 = no limit). A tap on the orb always sends what is there — that gesture
    *  is an explicit instruction, not a guess about where a sentence ended. */
   minChars: number;
+  /** 关键词打断 — the phrase that must be HEARD before the floor changes hands (opt-in; empty means
+   *  the mode is inert). No shipped default on purpose: nobody should be able to interrupt this
+   *  agent by guessing a word we chose for them. */
+  wakePhrase: string;
 }
 
 /** How the operator wants to TALK to the agent. These were browser-local (localStorage) and
@@ -150,14 +154,32 @@ export const PREFS_DEFAULT: PrefsSettings = {
   planHint: { on: true, everyTurn: true, instruction: "" },
 };
 
+/** The longest wake phrase, in CODE POINTS (an emoji is one, not two). One number, one place — the
+ *  web side has its own copy (`WAKE_MAX` in callSettings.ts) and the two must keep saying 24. */
+const WAKE_MAX = 24;
+
+/** Drop half-pairs left by a client that counted UTF-16 units and cut in the middle of an emoji.
+ *  `for...of` walks code points, so a lone surrogate arrives as a one-character string and is easy
+ *  to spot; storing it would put U+FFFD in the database. */
+function stripLone(s: string): string {
+  let out = "";
+  for (const ch of s) {
+    const cp = ch.codePointAt(0) ?? 0;
+    if (cp >= 0xD800 && cp <= 0xDFFF) continue;   // a surrogate that did not pair with its other half
+    out += ch;
+  }
+  return out;
+}
+
 export const CALL_DEFAULT: CallSettings = {
   /* 说完停顿 1.5s (was 1.2): the operator reads a pause as "it stopped listening" before 1.2 s,
      and a long sentence with a breath in the middle got cut. Both ends of the range still let
      the room decide — this is only where a fresh install starts. */
   interruptMode: "voice", bargeSensitivity: 60, bargeMs: 300, silenceMs: 1500, minChars: 3,
+  wakePhrase: "",
 };
 
-export const INTERRUPT_MODES: InterruptMode[] = ["voice", "button"];
+export const INTERRUPT_MODES: InterruptMode[] = ["voice", "button", "keyword"];
 
 /** How the cockpit treats AGENT PROCESSES — as opposed to what a conversation contains. AionUi
  *  ships the same knob in 设置 → 系统 as 「Agent 空闲超时（分钟）」 (default 5, 1–60, step 5) with the
@@ -194,15 +216,26 @@ export interface AgentSettings {
    *  agent knows it is there, the operator does not need to. Turned ON, every running process
    *  shows, which is what the fork's original "a snapshot is the only witness" design wanted. */
   showSilentBackground: boolean;
+  /** 记忆召回条目：单条最多多少字符（issue #37 的容量）。These items are a per-turn status frame
+   *  that also lands in SQLite, so the number is a cost knob, not a formatting preference: it is
+   *  handed to the agent as `HINDSIGHT_RECALL_ITEM_CHARS`, and the server applies the same limit
+   *  when it re-caps what arrives. 400 = the operator's 2026-10-21 decision (was 200). */
+  recallItemChars: number;
+  /** 记忆召回条目：一轮最多多少字符。**这是唯一的条数闸门** —— 不按条数截，短记忆本来就该多
+   *  进来几条，按条数截会连预算都没用完就把整条记忆丢掉。Handed to the agent as
+   *  `HINDSIGHT_RECALL_TOTAL_CHARS`（12K = 操作者 2026-10-21 的决定，原来是 6K）。 */
+  recallTotalChars: number;
 }
 
 export const AGENT_DEFAULT: AgentSettings = {
   idleKillMin: 5, warmSlots: 5, warmTtlHours: 24, cockpitSend: false, completionWake: true,
-  showSilentBackground: false,
+  showSilentBackground: false, recallItemChars: 400, recallTotalChars: 12000,
 };
 
 /** The numeric half of `AgentSettings` — the keys the range-checked helpers may write. */
-type AgentNumberKey = "idleKillMin" | "warmSlots" | "warmTtlHours";
+type AgentNumberKey =
+  | "idleKillMin" | "warmSlots" | "warmTtlHours" | "recallItemChars" | "recallTotalChars";
+
 
 /** What the settings page may offer: the ends of the number field. Server-side enforcement uses
  *  the same pair, so a hand-written request cannot store a nonsense value. */
@@ -212,6 +245,11 @@ export const AGENT_IDLE_RANGE: [number, number] = [0, 60];
 export const AGENT_WARM_RANGE: [number, number] = [0, 20];
 /** 0 = 不过期; a week is the longest span worth offering on the page. */
 export const AGENT_WARM_TTL_RANGE: [number, number] = [0, 168];
+/** 召回条目单条字符数：下限是为了别把一条记忆截成半句话，上限是「一整帧别超过一次 prompt 的量」。
+ *  与 provider 的默认同值（400），页面改这里，spawn 时用环境变量覆盖 agent 侧同一对常量。 */
+export const AGENT_RECALL_ITEM_RANGE: [number, number] = [80, 2000];
+/** 一轮条目的总字符数。这就是条数闸门本身，所以下限不能太小（2000 ≈ 5 条长记忆）。 */
+export const AGENT_RECALL_TOTAL_RANGE: [number, number] = [2000, 60000];
 
 /** 观察智能体 (`工作播报`, issue #16): ONE long-lived ACP session per observed session, spawned by the
  *  cockpit, in its own scratch directory, read-only, out of the rail. It is the thing that turns
@@ -353,7 +391,11 @@ function callAtShippedDefaults(r: Record<string, unknown>): boolean {
   // The mode counts too: a row that only says `interruptMode: "button"` IS a deliberate choice and
   // must not be reset along with a section that happens to still carry the old numbers.
   const modeShipped = r.interruptMode === undefined || r.interruptMode === CALL_DEFAULT.interruptMode;
-  return sens && modeShipped && Number(r.bargeMs) === 300 && Number(r.silenceMs) === 1200 && Number(r.minChars) === 1;
+  // …and a stored wake phrase is a deliberate choice too: without this, setting one and switching
+  // back to 语音打断 would make the whole section look untouched and the phrase would be dropped.
+  const wakeShipped = r.wakePhrase === undefined || r.wakePhrase === CALL_DEFAULT.wakePhrase;
+  return sens && modeShipped && wakeShipped
+    && Number(r.bargeMs) === 300 && Number(r.silenceMs) === 1200 && Number(r.minChars) === 1;
 }
 
 /** What each knob may be. The ranges are the panel's slider ends; the server is the one
@@ -487,6 +529,12 @@ function pickCall(raw: unknown): CallSettings {
     if (callAtShippedDefaults(raw as Record<string, unknown>)) return out;
     const mode = (raw as Record<string, unknown>).interruptMode;
     if (INTERRUPT_MODES.includes(mode as InterruptMode)) out.interruptMode = mode as InterruptMode;
+    const wake = (raw as Record<string, unknown>).wakePhrase;
+    // NOT length-capped here on purpose: this is what somebody ELSE stored (curl, a script, a previous
+    // version). Cutting it on the way out would then be written straight back by the `persist()` that
+    // runs at boot — silently destroying the operator's phrase with no log and nothing on screen.
+    // Length belongs to the write path (`saveCall`), which is where a human typed it.
+    if (typeof wake === "string") out.wakePhrase = wake.replace(/\s+/g, " ").trim();
     for (const k of CALL_NUMERIC) {
       const v = Number((raw as Record<string, unknown>)[k]);
       const [lo, hi] = CALL_RANGE[k];
@@ -514,6 +562,8 @@ function pickAgent(raw: unknown): AgentSettings {
     take("idleKillMin", AGENT_IDLE_RANGE);
     take("warmSlots", AGENT_WARM_RANGE);
     take("warmTtlHours", AGENT_WARM_TTL_RANGE);
+    take("recallItemChars", AGENT_RECALL_ITEM_RANGE);
+    take("recallTotalChars", AGENT_RECALL_TOTAL_RANGE);
     if (typeof r.cockpitSend === "boolean") out.cockpitSend = r.cockpitSend;
     if (typeof r.completionWake === "boolean") out.completionWake = r.completionWake;
     if (typeof r.showSilentBackground === "boolean") out.showSilentBackground = r.showSilentBackground;
@@ -631,6 +681,8 @@ export function publicSettings(): Record<string, unknown> {
     agentIdleRange: AGENT_IDLE_RANGE,
     agentWarmRange: AGENT_WARM_RANGE,
     agentWarmTtlRange: AGENT_WARM_TTL_RANGE,
+    agentRecallItemRange: AGENT_RECALL_ITEM_RANGE,
+    agentRecallTotalRange: AGENT_RECALL_TOTAL_RANGE,
     observer: cache.observer,
     observerDefaults: OBSERVER_DEFAULT,
     observerPromptDefault: OBSERVER_PROMPT_DEFAULT,
@@ -714,10 +766,24 @@ export function saveCall(patch: Record<string, unknown>): Settings {
   if ("interruptMode" in patch) {
     const mode = patch.interruptMode;
     if (!INTERRUPT_MODES.includes(mode as InterruptMode)) {
-      throw new Error(`invalid interruptMode: ${JSON.stringify(mode)} (expected voice|button)`);
+      throw new Error(`invalid interruptMode: ${JSON.stringify(mode)} (expected voice|button|keyword)`);
     }
     c.interruptMode = mode as InterruptMode;
   }
+  if ("wakePhrase" in patch) {
+    if (typeof patch.wakePhrase !== "string") throw new Error("invalid wakePhrase: expected a string");
+    // Truncated, not rejected: this field arrives in the same body as the sliders, and refusing it
+    // would throw away the operator's OTHER changes with it (and the panel would then say "saved
+    // locally, not on the server" for a phrase nobody was trying to abuse). Same rule as pickCall.
+    const wake = stripLone(patch.wakePhrase.replace(/\s+/g, " ").trim());   // an old client can slice a surrogate in half
+    const kept = [...wake];
+    if (kept.length > WAKE_MAX) {
+      // Truncated rather than refused, but never silently: one line in the log is the difference
+      // between 「我设过这个词」 and 「它自己变短了」 when a script writes 40 characters.
+      console.warn(`[agentus] wakePhrase truncated: ${kept.length} → ${WAKE_MAX} characters`);
+    }
+    c.wakePhrase = kept.slice(0, WAKE_MAX).join("");
+  }
   for (const key of CALL_NUMERIC) {
     if (!(key in patch)) continue;
     const raw = typeof patch[key] === "number" ? patch[key] as number : Number(patch[key]);
@@ -823,6 +889,8 @@ export function saveAgent(patch: Record<string, unknown>): Settings {
   set("idleKillMin", AGENT_IDLE_RANGE, "minutes", "0 = off");
   set("warmSlots", AGENT_WARM_RANGE, "slots", "0 = off");
   set("warmTtlHours", AGENT_WARM_TTL_RANGE, "hours", "0 = never expires");
+  set("recallItemChars", AGENT_RECALL_ITEM_RANGE, "chars", "no 0 — a cap must stay a cap");
+  set("recallTotalChars", AGENT_RECALL_TOTAL_RANGE, "chars", "no 0 — a cap must stay a cap");
   // A boolean next to three numbers: spelled out rather than forced through the numeric `set`.
   if ("cockpitSend" in patch) {
     if (typeof patch.cockpitSend !== "boolean") throw new Error("invalid cockpitSend: expected true or false");