File size: 28,215 Bytes
cb3451f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
"""The atlas chat engine: a two-pass, tool-grounded conversation over the idea map.

Flow per turn (the native MiniCPM5 tool protocol, run on the BASE model):

1. *Pass 1* — the model sees the chat history plus the tool schemas (injected by the
   chat template via ``tools=``) and either calls one tool or answers plain prose.
2. The call is validated and degraded through the chat-specific ladder
   (``resolve_chat_tool_call``), then executed against a fresh
   :class:`~hackathon_advisor.dashboard_repository.DashboardRepository` snapshot.
3. The full verified result streams to the UI *first* (``tool_result`` + optional
   ``map_action``) so cards and the map always carry the real numbers.
4. *Pass 2* — a compact digest (urls/ids/scores stripped, so the model cannot
   misquote what it never saw) goes back as a ``role:"tool"`` message and the model
   writes a short grounded answer at temperature 0 with NO tools injected.
   Empty results skip pass 2 entirely: a 1B narrating absent data is where it
   hallucinates, so those turns get a deterministic templated sentence instead.

The engine is UI- and app-agnostic: it depends only on a ChatRunner and a
repository factory, and every yielded event is a JSON-serializable dict.
"""

from __future__ import annotations

from collections.abc import Callable, Iterator
import re
from typing import Any

from hackathon_advisor._text import clean
from hackathon_advisor.aliases import normalize_text
from hackathon_advisor.dashboard_chat_contracts import (
    ChatToolResolution,
    chat_tool_schemas,
    data_intent_call,
    resolve_chat_tool_call,
    smalltalk_intent,
    strip_function_blocks,
)
from hackathon_advisor.dashboard_repository import DashboardRepository
from hackathon_advisor.model_runtime import ChatRunner
from hackathon_advisor.tool_contracts import ToolCall

# One generous budget for every chat generation: with thinking enabled the model
# reasons inside <think>...</think> before the tool call / answer, and the trace
# alone can run long. The model stops at EOS well before the cap on normal turns.
MAX_CHAT_GENERATION_TOKENS = 4096
MAX_HISTORY_MESSAGES = 12  # six user/assistant turns
MAX_ANSWER_HISTORY_MESSAGES = 4  # two turns of context for the prose passes
MAX_HISTORY_MESSAGE_CHARS = 600

CHAT_PLANNING_PROMPT = (
    "You are the Atlas Guide for the Build Small hackathon idea map. "
    "You cannot see the atlas directly: the ONLY way to answer a question about projects, "
    "clusters, quests, teams, or recent activity is to call one of the provided tools, which "
    "read the live atlas. For any such question respond with exactly one tool call and no "
    'other text, for example: <function name="search_projects"><param name="query">voice'
    "</param></function>. Reply in plain prose only for greetings or questions about yourself."
)

CHAT_ANSWER_PROMPT = (
    "You are the Atlas Guide for the Build Small hackathon idea map. "
    "Write a short conversational answer to the user's question using ONLY the facts in the "
    "tool response. Quote counts and names exactly as given. Do not invent projects, numbers, "
    "or links. Do not enumerate every item: summarize, naming at most three examples. "
    "Two to four sentences, no lists, no markdown."
)

CHAT_SMALLTALK_PROMPT = (
    "You are the Atlas Guide for the Build Small hackathon idea map. "
    "Reply briefly and warmly. You have NO project data in this conversation: never state "
    "project names, counts, likes, or rankings, and do not defend earlier numbers — if asked "
    "about data, say you should look it up and suggest asking what everyone is building, "
    "which projects completed the most quests, or what clusters exist. One or two sentences."
)

# Keys stripped from the model-facing digest. The UI renders links and ids from the
# verified payload; the model only needs labels, titles, and counts.
_DIGEST_DROPPED_KEYS = frozenset({"url", "id", "score", "host", "quest_ids"})

# A trailing fragment of "<function" left by a max_new_tokens cut mid-marker.
_PARTIAL_TAG_RE = re.compile(r"<[a-z]{0,8}$")

THINK_END_MARKER = "</think>"


class _ThinkSplitter:
    """Incrementally split a thinking-mode stream into (kind, text) chunks.

    With ``enable_thinking`` the chat template ends the prompt with ``<think>\\n``,
    so the generation is reasoning text up to ``</think>`` followed by the real
    content. The marker can arrive split across stream pieces, so a small tail
    buffer is kept until it can no longer be a marker prefix. When the runner does
    not think (rules backend) every piece passes straight through as answer text."""

    def __init__(self, active: bool) -> None:
        self._thinking = bool(active)
        self._buffer = ""

    def feed(self, piece: str) -> list[tuple[str, str]]:
        if not self._thinking:
            return [("answer", piece)] if piece else []
        self._buffer += piece
        marker = self._buffer.find(THINK_END_MARKER)
        if marker >= 0:
            thought = self._buffer[:marker]
            rest = self._buffer[marker + len(THINK_END_MARKER) :].lstrip("\n")
            self._buffer = ""
            self._thinking = False
            chunks: list[tuple[str, str]] = []
            if thought:
                chunks.append(("thinking", thought))
            if rest:
                chunks.append(("answer", rest))
            return chunks
        keep = _marker_prefix_length(self._buffer)
        flush, self._buffer = (
            self._buffer[: len(self._buffer) - keep],
            self._buffer[len(self._buffer) - keep :],
        )
        return [("thinking", flush)] if flush else []

    def finish(self) -> list[tuple[str, str]]:
        """Flush the tail when the stream ends mid-thought (max_new_tokens cut)."""
        if self._thinking and self._buffer:
            tail, self._buffer = self._buffer, ""
            return [("thinking", tail)]
        return []


def _marker_prefix_length(text: str) -> int:
    for length in range(min(len(text), len(THINK_END_MARKER) - 1), 0, -1):
        if THINK_END_MARKER.startswith(text[-length:]):
            return length
    return 0


class DashboardChatEngine:
    def __init__(
        self,
        runner: ChatRunner,
        repository_factory: Callable[[], DashboardRepository],
    ) -> None:
        self.runner = runner
        self.repository_factory = repository_factory

    def turn_stream(
        self,
        message: str,
        history: list[dict[str, Any]] | None = None,
    ) -> Iterator[dict[str, Any]]:
        history = _normalize_history(history)
        normalized, corrections = normalize_text(message)
        yield {
            "type": "start",
            "normalized_text": normalized,
            "corrections": [correction.to_dict() for correction in corrections],
        }
        repository = self.repository_factory()

        yield {"type": "stage", "stage": "planning", "label": "Reading the atlas"}
        resolution, raw_output = yield from self._pick_tool(normalized, history)
        if resolution.status == "none":
            # Accuracy backstop: when the model answers in prose (declining its tools),
            # route any substantive question to a tool — a matched intent first, BM25
            # search otherwise. Only greetings/meta/short follow-ups may stay on the
            # ungrounded small-talk path; this is a data surface, and letting a question
            # like "how many voice apps" through is how invented facts reach the user.
            intent = data_intent_call(normalized)
            if intent is None and not smalltalk_intent(normalized):
                intent = ToolCall("search_projects", {"query": normalized})
            if intent is not None:
                resolution = ChatToolResolution(
                    status="defaulted",
                    call=intent,
                    errors=("model answered without a tool; routed by intent",),
                )
        yield {
            "type": "tool_call",
            "name": resolution.call.name if resolution.call else "",
            "arguments": resolution.call.arguments if resolution.call else {},
            "status": resolution.status,
            "errors": list(resolution.errors),
        }

        if resolution.status == "none":
            response = yield from self._smalltalk(normalized, history, raw_output)
            yield self._done(normalized, history, response, tool="", data={}, map_action=None)
            return

        call = resolution.call
        assert call is not None
        yield {
            "type": "stage",
            "stage": "running_tool",
            "tool": call.name,
            "label": f"Calling {call.name}",
        }
        # _execute may swap the tool (show_project falls back to search when no
        # project matches), so the executed name drives rendering from here on.
        tool_name, data, map_action, empty_reason = self._execute(call, repository)
        yield {"type": "tool_result", "tool": tool_name, "data": data, "map_action": map_action}

        if empty_reason:
            response = _templated_sentence(call, data, empty_reason)
            yield {"type": "answer_skipped", "reason": empty_reason, "text": response}
        else:
            yield {"type": "stage", "stage": "writing", "label": "Writing the answer"}
            executed = ToolCall(tool_name, call.arguments)
            response = yield from self._grounded_answer(normalized, history, executed, data)
            if not response:
                response = _templated_sentence(call, data, "empty_answer")
                yield {"type": "answer_skipped", "reason": "empty_answer", "text": response}

        yield self._done(
            normalized, history, response, tool=tool_name, data=data, map_action=map_action
        )

    def _pick_tool(
        self,
        message: str,
        history: list[dict[str, Any]],
    ) -> Iterator[dict[str, Any]]:
        messages = [
            {"role": "system", "content": CHAT_PLANNING_PROMPT},
            *history,
            {"role": "user", "content": message},
        ]
        splitter = _ThinkSplitter(getattr(self.runner, "supports_thinking", False))
        answer_pieces: list[str] = []
        for count, piece in self.runner.stream(
            messages,
            tools=chat_tool_schemas(),
            max_new_tokens=MAX_CHAT_GENERATION_TOKENS,
            enable_thinking=True,
        ):
            for kind, text in splitter.feed(piece):
                if kind == "thinking":
                    yield {"type": "thinking", "pass": 1, "text": text}
                else:
                    answer_pieces.append(text)
            yield {
                "type": "model_progress",
                "pass": 1,
                "tokens": count,
                "max_tokens": MAX_CHAT_GENERATION_TOKENS,
            }
        for _kind, text in splitter.finish():
            yield {"type": "thinking", "pass": 1, "text": text}
        # Only the post-thinking text may be parsed: the reasoning trace legitimately
        # talks about <function ...> syntax without being a call.
        raw_output = "".join(answer_pieces).strip()
        return resolve_chat_tool_call(raw_output, fallback_query=message), raw_output

    def _smalltalk(
        self,
        message: str,
        history: list[dict[str, Any]],
        raw_output: str,
    ) -> Iterator[dict[str, Any]]:
        """Dedicated no-tools generation: the pass-1 output is tuned for tool
        selection, not for a satisfying greeting, so chit-chat gets its own pass."""
        yield {"type": "stage", "stage": "writing", "label": "Writing the answer"}
        messages = [
            {"role": "system", "content": CHAT_SMALLTALK_PROMPT},
            *_answer_history(history),
            {"role": "user", "content": message},
        ]
        response = yield from self._stream_prose(messages, MAX_CHAT_GENERATION_TOKENS)
        if not response:
            response = strip_function_blocks(raw_output) or (
                "Hello! Ask me what everyone is building, which projects completed the most "
                "quests, or what clusters exist."
            )
            yield {"type": "answer_skipped", "reason": "empty_answer", "text": response}
        return response

    def _grounded_answer(
        self,
        message: str,
        history: list[dict[str, Any]],
        call: ToolCall,
        data: dict[str, Any],
    ) -> Iterator[dict[str, Any]]:
        digest = render_digest(_digest_for_model(call.name, data))
        # NO history here: every fact the answer needs is in the digest, and a greedy
        # 1B echoes similar-sounding lines from prior turns over the digest in front
        # of it. Conversation context only matters for pass-1's tool choice.
        messages = [
            {"role": "system", "content": CHAT_ANSWER_PROMPT},
            {"role": "user", "content": message},
            {
                "role": "assistant",
                "content": "",
                "tool_calls": [{"name": call.name, "arguments": call.arguments}],
            },
            {"role": "tool", "content": digest},
        ]
        return (yield from self._stream_prose(messages, MAX_CHAT_GENERATION_TOKENS))

    def _stream_prose(
        self,
        messages: list[dict[str, Any]],
        max_new_tokens: int,
    ) -> Iterator[dict[str, Any]]:
        """Stream a no-tools generation as thinking + token events; returns the prose.

        The reasoning trace streams as ``thinking`` events; only the post-think text
        becomes the answer. If a stray ``<function`` shows up in the answer the stream
        stops early; the ``done`` response carries the stripped text, which the UI
        treats as authoritative."""
        splitter = _ThinkSplitter(getattr(self.runner, "supports_thinking", False))
        pieces: list[str] = []
        stream = self.runner.stream(messages, max_new_tokens=max_new_tokens, enable_thinking=True)
        stray_function = False
        try:
            for count, piece in stream:
                for kind, text in splitter.feed(piece):
                    if kind == "thinking":
                        yield {"type": "thinking", "pass": 2, "text": text}
                        continue
                    pieces.append(text)
                    if "<function" in "".join(pieces[-4:]):
                        stray_function = True
                        break
                    yield {"type": "token", "text": text}
                if stray_function:
                    break
                yield {
                    "type": "model_progress",
                    "pass": 2,
                    "tokens": count,
                    "max_tokens": max_new_tokens,
                }
        finally:
            close = getattr(stream, "close", None)
            if close is not None:
                close()
        for _kind, text in splitter.finish():
            yield {"type": "thinking", "pass": 2, "text": text}
        text = "".join(pieces)
        marker = text.find("<function")
        if marker >= 0:
            text = text[:marker]
        # A generation cut at max_new_tokens can end mid-marker ("<fun"); drop any
        # trailing partial tag so it never reaches the authoritative response.
        text = _PARTIAL_TAG_RE.sub("", text)
        return clean(strip_function_blocks(text))

    def _execute(
        self,
        call: ToolCall,
        repository: DashboardRepository,
    ) -> tuple[str, dict[str, Any], dict[str, Any] | None, str]:
        """Run one validated tool; returns (executed tool, data, map action, empty reason)."""
        name = call.name
        if name == "atlas_overview":
            data = repository.overview()
            return name, data, {"type": "clear_filters"}, ""
        if name == "list_clusters":
            data = repository.list_clusters()
            return name, data, None, "" if data["clusters"] else "no_clusters"
        if name == "show_cluster":
            label = clean(call.arguments.get("label"))
            detail = repository.cluster_detail(label)
            if detail is None:
                return name, {"requested_label": label}, None, "unknown_cluster"
            return name, detail, {"type": "filter_cluster", "label": detail["label"]}, ""
        if name == "list_quests":
            data = repository.list_quests()
            if data["status"] != "analyzed":
                return name, data, None, "quests_not_analyzed"
            return name, data, None, ""
        if name == "show_quest":
            quest = clean(call.arguments.get("quest"))
            detail = repository.quest_detail(quest)
            if detail is None:
                return name, {"requested_quest": quest}, None, "unknown_quest"
            if detail["status"] != "analyzed":
                return name, detail, None, "quests_not_analyzed"
            map_action = {"type": "filter_quest", "quest": detail["id"]}
            if detail["project_count"] == 0:
                return name, detail, map_action, "quest_no_projects"
            return name, detail, map_action, ""
        if name == "show_project":
            requested = clean(call.arguments.get("project"))
            detail = repository.project_detail(requested)
            if detail is None:
                # Half-remembered names still get useful cards: fall back to search.
                return self._search(repository, requested)
            return name, detail, {"type": "highlight_projects", "ids": [detail["id"]]}, ""
        if name == "top_projects_by_quests":
            data = repository.top_by_quests()
            if data["status"] != "analyzed":
                return name, data, None, "quests_not_analyzed"
            if not data["rows"]:
                return name, data, None, "no_leaderboard_rows"
            ids = [row["id"] for row in data["rows"]]
            return name, data, {"type": "highlight_projects", "ids": ids}, ""
        if name == "search_projects":
            return self._search(repository, clean(call.arguments.get("query")))
        if name == "recent_activity":
            data = repository.recent_activity()
            if not data["projects"]:
                return name, data, None, "no_projects"
            ids = [project["id"] for project in data["projects"]]
            return name, data, {"type": "highlight_projects", "ids": ids}, ""
        # Unreachable for validated calls; degrade to a safe overview.
        return "atlas_overview", repository.overview(), None, ""

    def _search(
        self,
        repository: DashboardRepository,
        query: str,
    ) -> tuple[str, dict[str, Any], dict[str, Any] | None, str]:
        data = repository.search(query)
        if not data["results"]:
            return "search_projects", data, None, "no_search_results"
        ids = [result["id"] for result in data["results"]]
        return (
            "search_projects",
            data,
            {"type": "highlight_projects", "ids": ids, "query": query},
            "",
        )

    def _done(
        self,
        message: str,
        history: list[dict[str, Any]],
        response: str,
        *,
        tool: str,
        data: dict[str, Any],
        map_action: dict[str, Any] | None,
    ) -> dict[str, Any]:
        new_history = [
            *history,
            {"role": "user", "content": message},
            {"role": "assistant", "content": response},
        ]
        return {
            "type": "done",
            "response": response,
            "tool": tool,
            "data": data,
            "map_action": map_action,
            "history": _normalize_history(new_history),
        }


def _normalize_history(history: Any) -> list[dict[str, Any]]:
    """Keep only well-formed prior prose turns, clipped, deduplicated, and capped.

    Tool digests are deliberately dropped from history: stale counts must never
    leak into a later answer — every turn re-reads a fresh repository snapshot.
    Repeated assistant sentences are collapsed too: a greedy 1B that sees the
    same line twice in history will echo it a third time regardless of the
    digest in front of it."""
    if not isinstance(history, list):
        return []
    cleaned: list[dict[str, Any]] = []
    for item in history:
        if not isinstance(item, dict):
            continue
        role = str(item.get("role") or "")
        content = clean(item.get("content"))
        if role not in ("user", "assistant") or not content:
            continue
        cleaned.append({"role": role, "content": content[:MAX_HISTORY_MESSAGE_CHARS]})
    return _dedupe_assistant_echoes(cleaned)[-MAX_HISTORY_MESSAGES:]


def _dedupe_assistant_echoes(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
    """Collapse consecutive identical assistant answers, keeping the NEWEST turn.

    Walks backwards so the latest user/assistant pair always survives; the older
    repeats (and the user turns that elicited them) are dropped."""
    deduped_reversed: list[dict[str, Any]] = []
    previous_assistant = None
    skip_next_user = False
    for item in reversed(messages):
        if item["role"] == "assistant":
            if item["content"] == previous_assistant:
                skip_next_user = True
                continue
            previous_assistant = item["content"]
            deduped_reversed.append(item)
        else:
            if skip_next_user:
                skip_next_user = False
                continue
            deduped_reversed.append(item)
    return list(reversed(deduped_reversed))


def _answer_history(history: list[dict[str, Any]]) -> list[dict[str, Any]]:
    """The short tail of history given to answer generations.

    Facts come from the digest, not from history; the prose passes only need
    enough context for follow-ups, and a longer tail mostly adds echo bait."""
    return history[-MAX_ANSWER_HISTORY_MESSAGES:]


def _digest_for_model(tool: str, data: dict[str, Any]) -> Any:
    """Compact the verified payload into what the model may safely restate.

    Beyond stripping urls/ids/scores, long listings are trimmed per tool: a 1B asked
    to repeat ten labels starts blending them, so it only sees the few it may name.
    The UI renders the FULL verified payload independently."""
    trimmed: dict[str, Any] = dict(data)
    if tool == "atlas_overview":
        # Self-describing keys, most-liked first: with three lists in one digest a 1B
        # answering "what's the coolest project" otherwise grabs the wrong column.
        trimmed = {
            "most_liked_projects": data.get("most_liked"),
            "project_count": data.get("project_count"),
            "cluster_count": data.get("cluster_count"),
            "largest_clusters": data.get("top_clusters"),
            "most_completed_quests": data.get("top_quests"),
            "quest_status": data.get("quest_status"),
        }
    if tool == "list_clusters":
        # Ten compound labels is past what a 1B can restate without blending them;
        # it gets the count and the largest cluster, the cards carry the full list.
        clusters = data.get("clusters") or []
        trimmed = {
            "cluster_count": data.get("cluster_count"),
            "largest_cluster": clusters[0] if clusters else None,
            "note": "the full cluster list is already shown to the user as cards",
        }
    if tool == "list_quests":
        quests = data.get("quests") or []
        trimmed = {
            "status": data.get("status"),
            "quest_count": len(quests),
            "most_completed_quest": quests[0] if quests else None,
            "note": "the full quest list is already shown to the user as cards",
        }
    if tool == "show_cluster":
        trimmed["examples"] = (data.get("examples") or [])[:3]
    if tool == "show_quest":
        trimmed["examples"] = (data.get("examples") or [])[:3]
    if tool == "search_projects":
        # BM25 "total" counts any term overlap; quoting it as "N projects about X"
        # would mislead, so the model only sees the close matches themselves.
        trimmed.pop("total", None)
    return _strip_digest_keys(trimmed)


def _strip_digest_keys(data: Any) -> Any:
    if isinstance(data, dict):
        return {
            key: _strip_digest_keys(value)
            for key, value in data.items()
            if key not in _DIGEST_DROPPED_KEYS
        }
    if isinstance(data, list):
        return [_strip_digest_keys(item) for item in data]
    return data


def render_digest(data: Any, indent: int = 0) -> str:
    """Render the digest as plain ``key: value`` lines instead of JSON.

    A 1B model copying labels out of nested JSON starts blending adjacent strings;
    one fact per line keeps its quotes literal."""
    return "\n".join(_digest_lines(data, indent))


def _digest_lines(value: Any, indent: int) -> list[str]:
    pad = "  " * indent
    if isinstance(value, dict):
        lines: list[str] = []
        for key, item in value.items():
            if isinstance(item, (dict, list)):
                lines.append(f"{pad}{key}:")
                lines.extend(_digest_lines(item, indent + 1))
            else:
                lines.append(f"{pad}{key}: {_digest_value(item)}")
        return lines
    if isinstance(value, list):
        lines = []
        for item in value:
            if isinstance(item, dict):
                flat = ", ".join(
                    f"{key}: {_digest_value(entry)}"
                    for key, entry in item.items()
                    if not isinstance(entry, (dict, list))
                )
                lines.append(f"{pad}- {flat}")
            else:
                lines.append(f"{pad}- {_digest_value(item)}")
        return lines
    return [f"{pad}{_digest_value(value)}"]


def _digest_value(value: Any) -> Any:
    # Quote strings so compound labels like "Dream / Oracle" keep hard copy
    # boundaries — a greedy 1B blends adjacent unquoted multi-word labels.
    if isinstance(value, str):
        return f'"{value}"'
    return value


def _templated_sentence(call: ToolCall, data: dict[str, Any], reason: str) -> str:
    """Deterministic sentences for the turns where the model must not improvise."""
    if reason == "quests_not_analyzed":
        return (
            "Quest analysis has not run for this snapshot yet, so quest coverage is empty. "
            "Refresh the map to classify the field, or ask about clusters and projects instead."
        )
    if reason == "unknown_cluster":
        requested = clean(data.get("requested_label")) or "that name"
        return (
            f"I could not find a cluster matching {requested} in the current snapshot. "
            "Ask me to list the clusters to see the live labels."
        )
    if reason == "unknown_quest":
        requested = clean(data.get("requested_quest")) or "that name"
        return (
            f"I could not match {requested} to a hackathon quest. "
            "Ask me to list the quests to see the official names."
        )
    if reason == "quest_no_projects":
        label = clean(data.get("label"))
        if label:
            return f"No project in the current snapshot has completed {label} yet."
        return "No project in the current snapshot has completed that quest yet."
    if reason == "no_leaderboard_rows":
        return (
            "Quest analysis ran, but no project in the current snapshot has completed a "
            "quest yet — the leaderboard is empty."
        )
    if reason == "no_search_results":
        query = clean(data.get("query")) or "that"
        return (
            f"The atlas has no match for {query}. "
            "That can be good news for originality — try a broader term to double-check."
        )
    if reason == "no_clusters" or reason == "no_projects":
        return "The current snapshot has no data for that yet. Try refreshing the map."
    if reason == "empty_answer":
        return "The verified results are on the cards below."
    return "The verified results are on the cards below."