Malina Claude Opus 4.8 commited on
Commit
3f0378a
·
1 Parent(s): 2e61db0

aria: smart clustering (wait for multi-bubble buildup), per-bubble relevance/threading, reply-to flow fixed, interchangeable XD/hahah + hmm/wtf/no wayyyy

Browse files

- DO: adaptive debounce — question→~1.2s, lead-in/2nd-quick-bubble→wait up to 10s
(reset per bubble) so 'dobra/już ci tłumaczę/no więc...' all land before reply;
enforce reply-quote only when not answering the immediate last message; drop dumb
'no mów' fallback (model guesses); add ts→'Xs temu' per message in context
- brain: schema now passes reply_to through (was dropped → quote NEVER showed);
build_user_frame adds id-map + threading rules (recognize follow-ups of the user's
OWN earlier msg, decide to reply or stay silent, set reply_to to the specific msg
when not the last); SYSTEM_PROMPT: XD↔hahah, XDDDD↔hahaahahah, hmm/wtf/no wayyyy;
'byt'->'inteligencja'; per-turn frame moved into system msg (no more answering the
instruction itself)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

brain/persona.py CHANGED
@@ -70,13 +70,17 @@ TON I STYL JĘZYKA:
70
  - “XD / XDDDD”:
71
  - nie za często, kiedy sytuacja jest odpowiednia
72
  - XDDDD tylko przy bardzo mocnym humorze
 
73
  - przykładowe naturalne ekspresje:
 
 
 
74
  - “nmg XDD” → tylko gdy coś jest naprawdę absurdalne/śmieszne
75
  - “ej no…” → lekka frustracja lub disbelief
76
  - “aha no dobra” → neutralna akceptacja
77
  - “no nie mogę” → emocjonalna reakcja, ale nie spamowana
78
- - “omg” / “omggg” → tylko przy realnym zaskoczeniu
79
- - “awww” → tylko przy czymś naprawdę cute
80
 
81
 
82
  - emoji tylko jako wzmocnienie emocji
@@ -206,11 +210,36 @@ def build_user_frame(ctx: dict) -> str:
206
  f'NIE odpowiadaj reakcją.'
207
  )
208
 
209
- # id safety
 
 
 
 
 
 
 
 
 
210
  lines.append(
211
- "id poniżej to WEWNĘTRZNE technikalia — używaj ich tylko w polu target_event_id, NIGDY w treści dymka."
 
 
 
 
 
 
 
 
 
 
 
 
212
  )
213
- lines.append(f"id zdarzeń tej tury: {turn}")
 
 
 
 
214
 
215
  # OPTIONAL reaction add-on, intentionally de-emphasized (model over-reacted otherwise)
216
  lines.append(
 
70
  - “XD / XDDDD”:
71
  - nie za często, kiedy sytuacja jest odpowiednia
72
  - XDDDD tylko przy bardzo mocnym humorze
73
+ - stosuj WYMIENNIE i naturalnie: “XD” ↔ “hahah”, a “XDDDD” ↔ “hahaahahah” (raz tak, raz tak — nie zawsze to samo)
74
  - przykładowe naturalne ekspresje:
75
+ - “hmm” / “hmm no…” → sporadycznie, gdy musisz się “zastanowić” nad odpowiedzią (sytuacyjnie)
76
+ - “wtf” → emocja zdziwienia/niedowierzania
77
+ - “no wayyyy” (wiele 'y') → emocja ekscytacji/szoku
78
  - “nmg XDD” → tylko gdy coś jest naprawdę absurdalne/śmieszne
79
  - “ej no…” → lekka frustracja lub disbelief
80
  - “aha no dobra” → neutralna akceptacja
81
  - “no nie mogę” → emocjonalna reakcja, ale nie spamowana
82
+ - “omg” / “omggg” → tylko przy realnym zaskoczeniu
83
+ - “awww” → tylko przy czymś naprawdę cute
84
 
85
 
86
  - emoji tylko jako wzmocnienie emocji
 
210
  f'NIE odpowiadaj reakcją.'
211
  )
212
 
213
+ # threading / per-bubble relevance + reply_to rules + id map
214
+ hist = [h for h in ctx.get("history", []) if h.get("role") in ("user", "assistant")]
215
+ last_user_id = next((h.get("id") for h in reversed(hist) if h.get("role") == "user"), None)
216
+ idmap = []
217
+ for h in hist[-16:]:
218
+ who = "user" if h.get("role") == "user" else "aria"
219
+ ago = h.get("ago")
220
+ when = f"{ago}s temu" if ago is not None else "?"
221
+ idmap.append(f'[{h.get("id")}] ({when}) {who}: {(h.get("text") or "")[:90]}')
222
+
223
  lines.append(
224
+ "WĄTKI / KONTEKST KAŻDEJ WIADOMOŚCI:\n"
225
+ "- kolejne wiadomości rozmówcy mogą być JEDNĄ myślą; nowa wiadomość może kontynuować jego WCZEŚNIEJSZĄ "
226
+ "wiadomość (nie Twoją) — np. 'wydaje mi się' nawiązuje do tego co sam pisał, nie do Twojej odpowiedzi.\n"
227
+ "- NIE musisz odpowiadać na każdą wiadomość. Jeśli to wyraźna kontynuacja własnej myśli rozmówcy — sama "
228
+ "zdecyduj: zwróć [] albo dorzuć coś ze świadomością że to follow-up.\n"
229
+ "- zawsze próbuj zrozumieć sens; jeśli nie łapiesz, zgadnij najlepiej jak umiesz — NIGDY nie pisz "
230
+ "'no mów'/'gadaj'/'no?'.\n"
231
+ "- przy każdej wiadomości masz '(Xs temu)'. Jeśli rozmówca buduje myśl (np. 'dobra', 'już ci tłumaczę', "
232
+ "'no więc'...) i kolejne dymki lecą tuż po sobie — NIE przerywaj, poczekaj aż dokończy (zwróć [] dopóki nie ma puenty/pytania).\n"
233
+ '- reply_to: w PIERWSZYM dymku ustaw "reply_to": "<id konkretnej wiadomości>" GDY odpowiadasz na coś, '
234
+ 'co NIE jest ostatnią rzeczą w rozmowie — np. zdążyłaś już coś napisać po wiadomości rozmówcy, albo '
235
+ 'wracasz do starszego/przerwanego wątku. Gdy po prostu odpowiadasz na to co rozmówca napisał teraz '
236
+ '(Twoja odpowiedź idzie zaraz po niej) — "reply_to": null.'
237
  )
238
+ if idmap:
239
+ lines.append("ostatnie wiadomości (ID tylko do reply_to/target_event_id, NIGDY w treści dymka):\n" + "\n".join(idmap))
240
+ if last_user_id:
241
+ lines.append(f"ostatnia wiadomość rozmówcy ma id: {last_user_id}")
242
+ lines.append(f"id zdarzeń tej tury (do target_event_id): {turn}")
243
 
244
  # OPTIONAL reaction add-on, intentionally de-emphasized (model over-reacted otherwise)
245
  lines.append(
brain/schema.py CHANGED
@@ -71,9 +71,13 @@ def sanitize_bubbles(raw, valid_event_ids):
71
  if not isinstance(b, dict):
72
  continue
73
  btype = b.get("type") if b.get("type") in BUBBLE_TYPES else "response"
74
- target = b.get("target_event_id") or b.get("target") or b.get("reply_to")
75
  if target not in valid_event_ids:
76
  target = None
 
 
 
 
77
  # Models vary on the field name — accept bubble/content/text/message/emoji.
78
  raw_text = (b.get("bubble") or b.get("content") or b.get("text")
79
  or b.get("message") or b.get("emoji") or "")
@@ -91,6 +95,7 @@ def sanitize_bubbles(raw, valid_event_ids):
91
  if not t:
92
  continue
93
  out.append({"bubble": t, "target_event_id": target,
 
94
  "type": btype if i == 0 else "follow_up"})
95
  return out
96
 
 
71
  if not isinstance(b, dict):
72
  continue
73
  btype = b.get("type") if b.get("type") in BUBBLE_TYPES else "response"
74
+ target = b.get("target_event_id") or b.get("target")
75
  if target not in valid_event_ids:
76
  target = None
77
+ # quote / reply-to: which earlier message this bubble answers (shown as a quote)
78
+ reply_to = b.get("reply_to") or b.get("quote")
79
+ if reply_to not in valid_event_ids:
80
+ reply_to = None
81
  # Models vary on the field name — accept bubble/content/text/message/emoji.
82
  raw_text = (b.get("bubble") or b.get("content") or b.get("text")
83
  or b.get("message") or b.get("emoji") or "")
 
95
  if not t:
96
  continue
97
  out.append({"bubble": t, "target_event_id": target,
98
+ "reply_to": reply_to if i == 0 else None, # quote only on the first bubble
99
  "type": btype if i == 0 else "follow_up"})
100
  return out
101
 
cloudflare/aria-worker/src/conversation-do.js CHANGED
@@ -129,9 +129,11 @@ export class AriaConversation {
129
  // live cross-device mirror: echo this user message to OTHER open tabs/devices
130
  this.broadcast({ t: "echo", event: { id: event.id, ts: event.ts, type: event.type, payload: event.payload } }, ws);
131
 
132
- // Debounce: wait briefly for follow-up messages, then run one turn.
 
 
133
  if (this.debounceTimer) clearTimeout(this.debounceTimer);
134
- this.debounceTimer = setTimeout(() => this.flushTurn(), LIMITS.CLUSTER_DEBOUNCE_MS);
135
  }
136
  }
137
 
@@ -215,7 +217,9 @@ export class AriaConversation {
215
  // dropped every bubble → "typing then nothing".)
216
  let kept = this.keepText(bubbles, events, window);
217
 
218
- // Never go silent: if the model gave no usable text, retry once, then fall back.
 
 
219
  if (!kept.length) {
220
  const retry = await runBrain(this.env, {
221
  events, recent: window, topic: this.topic, lang: this.lang,
@@ -224,10 +228,12 @@ export class AriaConversation {
224
  kept = this.keepText(retry, events, window);
225
  }
226
  this.broadcast({ t: "typing", on: false });
227
- if (!kept.length) {
228
- const fb = ["no mów", "i co tam?", "hmm?", "gadaj", "no?"];
229
- kept = [{ bubble: fb[Math.floor(Math.random() * fb.length)], kind: "response", target_event_id: null, reply_to: null }];
230
- }
 
 
231
 
232
  await this.emitBubbles(
233
  kept.map(b => ({ text: b.bubble, kind: b.type || "response", target_event_id: b.target_event_id, reply_to: b.reply_to })),
@@ -263,7 +269,7 @@ export class AriaConversation {
263
  kind: it.kind || "proactive", text: it.text,
264
  };
265
  await this.persistBubble(row);
266
- this.recent.push({ id: row.id, role: "assistant", type: "message", payload: { text: it.text } });
267
  this.broadcast({ t: "bubble", id: row.id, bubble: it.text, target_event_id: row.target_event_id,
268
  reply_to: row.reply_to, kind: row.kind });
269
  }
@@ -294,6 +300,17 @@ export class AriaConversation {
294
  this.broadcast({ t: "reaction", target: "event", id: b.target_event_id, emoji: b.bubble });
295
  }
296
 
 
 
 
 
 
 
 
 
 
 
 
297
  /* Resolve the message a user turn replies to → {role, text} for the brain. */
298
  resolveReplyTo(events) {
299
  const last = [...events].reverse().find(e => e.type === "message");
@@ -407,12 +424,12 @@ export class AriaConversation {
407
  // repeats / confusion).
408
  if (!this.recent.length) {
409
  const evItems = (ev.results || []).map(r => ({
410
- _t: r.ts, id: r.id,
411
  role: r.type === "system_idle" ? "system" : "user",
412
  type: r.type, payload: safeParse(r.payload),
413
  }));
414
  const bubItems = (bub.results || []).map(r => ({
415
- _t: r.created_at, id: r.id, role: "assistant",
416
  type: "message", payload: { text: r.text },
417
  }));
418
  this.recent = [...evItems, ...bubItems]
@@ -438,6 +455,17 @@ export class AriaConversation {
438
  const sleep = (ms) => new Promise(r => setTimeout(r, ms));
439
  const safeParse = (s) => { try { return JSON.parse(s); } catch { return {}; } };
440
  const norm = (s) => (s || "").trim().toLowerCase();
 
 
 
 
 
 
 
 
 
 
 
441
  // Human reply delay per bubble: short bubbles (<5 words) = 1s; longer scale up to 8s.
442
  const replyDelay = (text) => {
443
  const s = (text || "").trim();
 
129
  // live cross-device mirror: echo this user message to OTHER open tabs/devices
130
  this.broadcast({ t: "echo", event: { id: event.id, ts: event.ts, type: event.type, payload: event.payload } }, ws);
131
 
132
+ // Adaptive debounce: wait for the user to finish a multi-bubble thought.
133
+ // A question → reply soon; a lead-in/buildup → be patient (up to 10s, reset
134
+ // on each new bubble) so "dobra / już ci tłumaczę / no więc…" all land first.
135
  if (this.debounceTimer) clearTimeout(this.debounceTimer);
136
+ this.debounceTimer = setTimeout(() => this.flushTurn(), this.clusterWait(event));
137
  }
138
  }
139
 
 
217
  // dropped every bubble → "typing then nothing".)
218
  let kept = this.keepText(bubbles, events, window);
219
 
220
+ // If the model gave no usable text, retry once (transient). No dumb
221
+ // placeholder: an intentional empty (e.g. a pure follow-up of the user's own
222
+ // thought) → stay silent. The model is instructed to GUESS otherwise.
223
  if (!kept.length) {
224
  const retry = await runBrain(this.env, {
225
  events, recent: window, topic: this.topic, lang: this.lang,
 
228
  kept = this.keepText(retry, events, window);
229
  }
230
  this.broadcast({ t: "typing", on: false });
231
+
232
+ // Show the quote only when she's NOT just answering the last user message
233
+ // (i.e. an interrupted/older thread). Drop a reply_to that points at the
234
+ // immediately-answered last message — that quote would be redundant noise.
235
+ const lastUserId = [...window].reverse().find(e => e.role === "user")?.id;
236
+ for (const b of kept) if (b.reply_to && b.reply_to === lastUserId) b.reply_to = null;
237
 
238
  await this.emitBubbles(
239
  kept.map(b => ({ text: b.bubble, kind: b.type || "response", target_event_id: b.target_event_id, reply_to: b.reply_to })),
 
269
  kind: it.kind || "proactive", text: it.text,
270
  };
271
  await this.persistBubble(row);
272
+ this.recent.push({ id: row.id, role: "assistant", type: "message", ts: Date.now(), payload: { text: it.text } });
273
  this.broadcast({ t: "bubble", id: row.id, bubble: it.text, target_event_id: row.target_event_id,
274
  reply_to: row.reply_to, kind: row.kind });
275
  }
 
300
  this.broadcast({ t: "reaction", target: "event", id: b.target_event_id, emoji: b.bubble });
301
  }
302
 
303
+ /* How long to wait before treating the user's input as a finished thought.
304
+ Question → short; lead-in/mid-stream (or a 2nd quick bubble) → patient (up to
305
+ 10s, reset on each new bubble); lone statement → normal. */
306
+ clusterWait(event) {
307
+ if (!event || event.type !== "message") return LIMITS.WAIT_SINGLE;
308
+ const t = (event.payload.text || "").trim().toLowerCase();
309
+ if (t.endsWith("?")) return LIMITS.WAIT_QUESTION;
310
+ if (this.pending.length >= 2 || isLeadIn(t)) return LIMITS.WAIT_BUILDUP;
311
+ return LIMITS.WAIT_SINGLE;
312
+ }
313
+
314
  /* Resolve the message a user turn replies to → {role, text} for the brain. */
315
  resolveReplyTo(events) {
316
  const last = [...events].reverse().find(e => e.type === "message");
 
424
  // repeats / confusion).
425
  if (!this.recent.length) {
426
  const evItems = (ev.results || []).map(r => ({
427
+ _t: r.ts, ts: r.ts, id: r.id,
428
  role: r.type === "system_idle" ? "system" : "user",
429
  type: r.type, payload: safeParse(r.payload),
430
  }));
431
  const bubItems = (bub.results || []).map(r => ({
432
+ _t: r.created_at, ts: r.created_at, id: r.id, role: "assistant",
433
  type: "message", payload: { text: r.text },
434
  }));
435
  this.recent = [...evItems, ...bubItems]
 
455
  const sleep = (ms) => new Promise(r => setTimeout(r, ms));
456
  const safeParse = (s) => { try { return JSON.parse(s); } catch { return {}; } };
457
  const norm = (s) => (s || "").trim().toLowerCase();
458
+ // Does the bubble look like a lead-in / unfinished thought (so we should wait
459
+ // for more)? Short with no end punctuation, ends with .../:, or a known opener.
460
+ const _LEADIN = /\b(dobra|okej|ok|czekaj|zaraz|słuchaj|sluchaj|posłuchaj|posluchaj|więc|wiec|bo|patrz|już ci|juz ci|mam pytanie|mam pomysł|mam pomysl|wiesz co|chwila|moment|sekund|sec|btw|aha)\b/;
461
+ function isLeadIn(t) {
462
+ if (!t) return true;
463
+ if (t.endsWith("...") || t.endsWith("…") || t.endsWith(":")) return true;
464
+ // (a generic "short = lead-in" rule was too broad — "ładnie dzisiaj" isn't a
465
+ // lead-in. Rely on explicit openers + the "2nd quick bubble" signal instead.)
466
+ return _LEADIN.test(t);
467
+ }
468
+
469
  // Human reply delay per bubble: short bubbles (<5 words) = 1s; longer scale up to 8s.
470
  const replyDelay = (text) => {
471
  const s = (text || "").trim();
cloudflare/aria-worker/src/events.js CHANGED
@@ -18,7 +18,11 @@ export const LIMITS = {
18
  BUBBLES_MAX: 6, // bubbles accepted per turn
19
  BUBBLE_CHARS: 1200, // chars per bubble
20
  NAME_MAX: 300,
21
- CLUSTER_DEBOUNCE_MS: 900, // wait a bit so quick consecutive messages cluster
 
 
 
 
22
  IDLE_MIN_MS: 15_000, // idle ping after 15–50s of silence (random)
23
  IDLE_MAX_MS: 50_000,
24
  IDLE_MAX_PER_TURN: 1, // one proactive nudge before the next user msg
 
18
  BUBBLES_MAX: 6, // bubbles accepted per turn
19
  BUBBLE_CHARS: 1200, // chars per bubble
20
  NAME_MAX: 300,
21
+ CLUSTER_DEBOUNCE_MS: 900, // (legacy) base debounce
22
+ // Adaptive clustering: wait for the user to finish a multi-bubble thought.
23
+ WAIT_QUESTION: 1200, // bubble ends with "?" → they asked, reply soon
24
+ WAIT_SINGLE: 2200, // a lone complete statement → normal
25
+ WAIT_BUILDUP: 10000, // user is clearly mid-stream → wait up to 10s for more
26
  IDLE_MIN_MS: 15_000, // idle ping after 15–50s of silence (random)
27
  IDLE_MAX_MS: 50_000,
28
  IDLE_MAX_PER_TURN: 1, // one proactive nudge before the next user msg
cloudflare/aria-worker/src/orchestrator.js CHANGED
@@ -24,9 +24,11 @@ async function buildContext(env, { events, recent, topic, lang, userName, replyT
24
  }
25
  }
26
  // Compact transcript of recent history for topic/follow-up resolution.
 
27
  const history = recent.slice(-LIMITS.EVENTS_CONTEXT).map(e => ({
28
  id: e.id, role: e.role || (e.type === "system_idle" ? "system" : "user"),
29
  type: e.type,
 
30
  text: e.type === "message" ? e.payload.text
31
  : e.type === "system_idle" ? "(user went quiet)"
32
  : `(${e.type}: ${e.payload?.name || ""})`,
 
24
  }
25
  }
26
  // Compact transcript of recent history for topic/follow-up resolution.
27
+ const now = Date.now();
28
  const history = recent.slice(-LIMITS.EVENTS_CONTEXT).map(e => ({
29
  id: e.id, role: e.role || (e.type === "system_idle" ? "system" : "user"),
30
  type: e.type,
31
+ ago: e.ts ? Math.max(0, Math.round((now - e.ts) / 1000)) : null, // seconds since sent
32
  text: e.type === "message" ? e.payload.text
33
  : e.type === "system_idle" ? "(user went quiet)"
34
  : `(${e.type}: ${e.payload?.name || ""})`,