Harden Dream QA vision intake

docs/smoke/2026-06-12-modal-vision-sketch-evidence.json

@@ -0,0 +1,76 @@
+{
+    "generated_at": "2026-06-12T01:53:21.674225+00:00",
+    "cwd": "/Users/junhaocheng/working-dir/ai-competitions/build-small-hackthon",
+    "git_commit": "ec0441abdecacf44918ca9eace7c2a57e6795d1b",
+    "git_branch": "bugfix/emotion-led-dream-qa",
+    "model": {
+        "modal_app": "dream-customs-minicpm-backend",
+        "vision_model": "openbmb/MiniCPM-V-4.6",
+        "endpoint_host": "cjh-12569--dream-customs-minicpm-backend-vision.modal.run"
+    },
+    "secret_handling": "Modal secret used only inside temporary Modal runner; no token printed or written.",
+    "results": [
+        {
+            "case_id": "sea",
+            "image_path": "/Users/junhaocheng/working-dir/ai-competitions/build-small-hackthon/tmp/dream-sketches/dream_sketch_01_sea.png",
+            "expected_markers": [
+                "sea",
+                "wave",
+                "moon",
+                "person"
+            ],
+            "http_status": 200,
+            "transport_ok": true,
+            "elapsed_seconds": 25.48,
+            "route_error": "",
+            "raw_response": "{'input_text': [{'role': 'user', 'content': [{'type': 'image', 'path': '/tmp/tmpjyirmi6i.png'}, {'type': 'text', 'text': 'You are MiniCPM-V-4.6 acting as the visual witness for Dream QA. Describe only what is visible in this dream sketch. Return concise JSON with keys: scene_summary, objects, visible_text, spatial_relations, mood_cues, uncertain_details, surprising_detail. Do not diagnose the user.'}]}], 'generated_text': [{'role': 'user', 'content': [{'type': 'image', 'path': '/tmp/tmpjyirmi6i.png'}, {'type': 'text', 'text': 'You are MiniCPM-V-4.6 acting as the visual witness for Dream QA. Describe only what is visible in this dream sketch. Return concise JSON with keys: scene_summary, objects, visible_text, spatial_relations, mood_cues, uncertain_details, surprising_detail. Do not diagnose the user.'}]}, {'role': 'assistant', 'content': '\\n\\n{\\n  \"scene_summary\": \"A dreamlike representation of a sea under night\",\\n  \"objects\": [\"a stick figure\", \"wavy lines representing water\", \"a crescent moon\", \"a circular shape above the stick figure\"],\\n  \"visible_text\": [\"dark sea dream\"],\\n  \"spatial_relations\": \"the stick figure is on the water waves, with the moon in the upper right\",\\n  \"mood_cues\": \"calm and dreamy\",\\n  \"uncertain_details\": \"the exact emotions or states of the figure\",\\n  \"surprising_detail\": \"the stick figure has a curved line above its head, suggesting a sense of wonder or awareness\"\\n}'}]}",
+            "marker_hits": [
+                "sea",
+                "wave",
+                "moon"
+            ]
+        },
+        {
+            "case_id": "elevator",
+            "image_path": "/Users/junhaocheng/working-dir/ai-competitions/build-small-hackthon/tmp/dream-sketches/dream_sketch_02_elevator.png",
+            "expected_markers": [
+                "elevator",
+                "button",
+                "14",
+                "melted"
+            ],
+            "http_status": 200,
+            "transport_ok": true,
+            "elapsed_seconds": 5.55,
+            "route_error": "",
+            "raw_response": "{'input_text': [{'role': 'user', 'content': [{'type': 'image', 'path': '/tmp/tmpc8qf8m8w.png'}, {'type': 'text', 'text': 'You are MiniCPM-V-4.6 acting as the visual witness for Dream QA. Describe only what is visible in this dream sketch. Return concise JSON with keys: scene_summary, objects, visible_text, spatial_relations, mood_cues, uncertain_details, surprising_detail. Do not diagnose the user.'}]}], 'generated_text': [{'role': 'user', 'content': [{'type': 'image', 'path': '/tmp/tmpc8qf8m8w.png'}, {'type': 'text', 'text': 'You are MiniCPM-V-4.6 acting as the visual witness for Dream QA. Describe only what is visible in this dream sketch. Return concise JSON with keys: scene_summary, objects, visible_text, spatial_relations, mood_cues, uncertain_details, surprising_detail. Do not diagnose the user.'}]}, {'role': 'assistant', 'content': '\\n\\n{\\n  \"scene_summary\": \"A conceptual visual representation of melted elevator buttons in a setting\",\\n  \"objects\": [\"melted elevator buttons\", \"two tall gray doors\"],\\n  \"visible_text\": [\"14\", \"melted elevator buttons\"],\\n  \"spatial_relations\": \"The melted buttons are to the right of the elevator doors\",\\n  \"mood_cues\": \"Suggests a malfunction or melting process\",\\n  \"uncertain_details\": \"The exact state or effects of the melted buttons are implied\",\\n  \"surprising_detail\": \"The buttons are depicted as colorful segments, indicating a melting or transition state\"\\n}'}]}",
+            "marker_hits": [
+                "elevator",
+                "button",
+                "14",
+                "melted"
+            ]
+        },
+        {
+            "case_id": "subway_child",
+            "image_path": "/Users/junhaocheng/working-dir/ai-competitions/build-small-hackthon/tmp/dream-sketches/dream_sketch_03_subway_child.png",
+            "expected_markers": [
+                "subway",
+                "child",
+                "home",
+                "arrow"
+            ],
+            "http_status": 200,
+            "transport_ok": true,
+            "elapsed_seconds": 5.62,
+            "route_error": "",
+            "raw_response": "{'input_text': [{'role': 'user', 'content': [{'type': 'image', 'path': '/tmp/tmp38m34fjv.png'}, {'type': 'text', 'text': 'You are MiniCPM-V-4.6 acting as the visual witness for Dream QA. Describe only what is visible in this dream sketch. Return concise JSON with keys: scene_summary, objects, visible_text, spatial_relations, mood_cues, uncertain_details, surprising_detail. Do not diagnose the user.'}]}], 'generated_text': [{'role': 'user', 'content': [{'type': 'image', 'path': '/tmp/tmp38m34fjv.png'}, {'type': 'text', 'text': 'You are MiniCPM-V-4.6 acting as the visual witness for Dream QA. Describe only what is visible in this dream sketch. Return concise JSON with keys: scene_summary, objects, visible_text, spatial_relations, mood_cues, uncertain_details, surprising_detail. Do not diagnose the user.'}]}, {'role': 'assistant', 'content': '\\n\\n{\\n  \"scene_summary\": \"A simple sketch depicting a lost child at a subway station\",\\n  \"objects\": [\"a child figure\", \"a subway sign\", \"an arrow\"],\\n  \"visible_text\": [\"lost child at subway\", \"SUBWAY\", \"HOME ?\"],\\n  \"spatial_relations\": \"the child is in front of the subway sign, which has an arrow pointing to the right\",\\n  \"mood_cues\": \"suggests a situation of loss or search\",\\n  \"uncertain_details\": \"what the child is looking for or where home is\",\\n  \"surprising_detail\": \"the sign asks \\'HOME ?\\'\"\\n}'}]}",
+            "marker_hits": [
+                "subway",
+                "child",
+                "home",
+                "arrow"
+            ]
+        }
+    ]
+}

docs/smoke/2026-06-12-modal-vision-sketch-evidence.md

@@ -0,0 +1,67 @@
+# Modal Vision Sketch Evidence - 2026-06-12
+
+Goal: verify whether Dream QA can use MiniCPM-V-4.6 to understand uploaded dream sketches, and clarify how text plus image should be merged.
+
+## Test Assets
+
+- Sea sketch: `docs/smoke/assets/dream_sketch_01_sea.png`
+- Elevator sketch: `docs/smoke/assets/dream_sketch_02_elevator.png`
+- Subway child sketch: `docs/smoke/assets/dream_sketch_03_subway_child.png`
+- Raw Modal evidence: `docs/smoke/2026-06-12-modal-vision-sketch-evidence.json`
+- Chrome local paste screenshot: `docs/smoke/assets/2026-06-12-chrome-local-image-paste-fallback.png`
+
+## Chrome Local Upload Check
+
+Local app URL: `http://127.0.0.1:7870`
+
+Result:
+
+- Native file chooser injection was blocked by the Chrome extension file-access permission.
+- Clipboard paste into the Gradio image input worked. The page showed `Remove Image`, confirming the image slot received the sketch.
+- Submitting the local page without Modal endpoint/token configured did not prove real MiniCPM-V behavior. The follow-up used fallback/demo visual clues (`melted elevator button`) instead of the pasted sea sketch.
+
+Interpretation: the local Gradio image UI can receive images, but a local browser smoke without configured hosted endpoint/token only verifies upload plumbing. It must not be treated as evidence that MiniCPM-V understood the image.
+
+## Real MiniCPM-V-4.6 Route
+
+The three sketches were sent to the Modal vision endpoint through a temporary local runner. Secret handling: token was used only inside the Modal runner path; no token was printed, stored, or committed.
+
+| Case | Expected visual markers | HTTP | Elapsed | Marker hits | Modal-V result |
+| --- | --- | --- | ---: | --- | --- |
+| `sea` | sea, wave, moon, person | 200 | 25.48s | sea, wave, moon | saw a sea under night, stick figure, waves, crescent moon, visible text `dark sea dream` |
+| `elevator` | elevator, button, 14, melted | 200 | 5.55s | elevator, button, 14, melted | saw melted elevator buttons, two gray doors, visible text `14` |
+| `subway_child` | subway, child, home, arrow | 200 | 5.62s | subway, child, home, arrow | saw a lost child at a subway station, subway sign, arrow, visible text `HOME ?` |
+
+Conclusion: MiniCPM-V-4.6 did understand the low-resolution dream sketches and returned concrete visual evidence.
+
+## Backend Fix Verified
+
+Observed Modal response shape:
+
+- The hosted endpoint returned `response` as a stringified Python dictionary.
+- The useful assistant JSON was nested under `generated_text[-1].content`.
+- Before this fix, Dream QA could treat that whole string as unparsed text and lose the structured witness report.
+
+Fix:
+
+- `HostedMiniCPMVisionClient` now extracts assistant content from stringified Modal pipeline payloads and message lists.
+- Chinese Dream QA now localizes common English visual anchors from MiniCPM-V, so a Chinese user does not see awkward anchors like `dreamlike representation of a sea`.
+
+Verification:
+
+```bash
+.venv/bin/python -m pytest -q tests/test_ollama_models.py tests/test_vision_witness.py tests/test_pipeline.py::test_visual_witness_clues_drive_questions_and_today_tip tests/test_pipeline.py::test_zh_text_and_image_keep_user_question_while_using_visual_anchors
+.venv/bin/python -m py_compile dream_customs/models.py dream_customs/pipeline.py
+```
+
+Result: `21 passed in 0.12s`; `py_compile` passed.
+
+## Text + Image Product Rule
+
+When the user uploads both text and image:
+
+1. User text owns the emotional question and intent. Example: `我想知道为什么它让我这么慌` becomes the main question.
+2. Image contributes concrete anchors. Example: sea, waves, crescent moon, elevator button, subway sign, `HOME ?`.
+3. The follow-up and Today Tip must combine them: answer the user's feeling first, then use the visual anchors as evidence.
+4. If text and image conflict, do not force a single meaning. Ask one gentle clarification or phrase it as uncertainty.
+5. Text-only remains the fallback; image failure must not block the user from getting a grounded response.

dream_customs/models.py

@@ -1,4 +1,5 @@
 import base64
+import ast
 import json
 import os
 import re
@@ -549,7 +550,16 @@ def _hosted_text_from_response(payload: Dict[str, Any]) -> str:
     for key in ("response", "text", "generated_text", "output"):
         value = payload.get(key)
         if isinstance(value, str):
+            parsed = _python_literal_dict(value)
+            if parsed:
+                nested = _hosted_text_from_response(parsed)
+                if nested:
+                    return nested
             return value
+        if isinstance(value, list):
+            nested = _content_from_messages(value)
+            if nested:
+                return nested
     choices = payload.get("choices")
     if isinstance(choices, list) and choices:
         first = choices[0]
@@ -565,6 +575,32 @@ def _hosted_text_from_response(payload: Dict[str, Any]) -> str:
     return ""
 
 
+def _python_literal_dict(text: str) -> Optional[Dict[str, Any]]:
+    try:
+        parsed = ast.literal_eval(text)
+    except (SyntaxError, ValueError):
+        return None
+    return parsed if isinstance(parsed, dict) else None
+
+
+def _content_from_messages(messages: List[Any]) -> str:
+    for item in reversed(messages):
+        if not isinstance(item, dict):
+            continue
+        content = item.get("content")
+        if isinstance(content, str) and content.strip():
+            return content.strip()
+        if isinstance(content, list):
+            text_parts = [
+                str(part.get("text", "")).strip()
+                for part in content
+                if isinstance(part, dict) and str(part.get("text", "")).strip()
+            ]
+            if text_parts:
+                return "\n".join(text_parts)
+    return ""
+
+
 class HostedMiniCPMTextClient:
     def __init__(
         self,

dream_customs/pipeline.py

@@ -108,6 +108,12 @@ _ZH_ANCHOR_MARKERS = [
     "掉进海",
     "海里",
     "海",
+    "前任发消息",
+    "发消息又消失",
+    "发消息",
+    "消息",
+    "前任",
+    "消失",
     "小孩找不到家",
     "小孩",
     "找不到家",
@@ -210,6 +216,12 @@ _ZH_TO_EN_PHRASES = {
     "掉进海": "falling into the sea",
     "海里": "sea",
     "海": "sea",
+    "前任发消息": "former partner sending a message",
+    "发消息又消失": "message that disappeared",
+    "发消息": "sending a message",
+    "消息": "message",
+    "前任": "former partner",
+    "消失": "disappearing",
     "小孩找不到家": "child unable to find home",
     "小孩": "child",
     "找不到家": "unable to find home",
@@ -304,6 +316,38 @@ _ZH_TO_EN_PHRASES = {
     "梦境": "dream",
 }
 
+_EN_TO_ZH_PHRASES = {
+    "a dreamlike representation of a sea under night": "夜晚的海",
+    "dreamlike representation of a sea": "夜晚的海",
+    "sea under night": "夜晚的海",
+    "dark sea dream": "漆黑的海",
+    "dark sea": "漆黑的海",
+    "wavy water": "海浪",
+    "waves": "海浪",
+    "wave": "海浪",
+    "sea": "海",
+    "crescent moon": "月牙",
+    "moon": "月亮",
+    "stick figure": "小人",
+    "melted elevator buttons": "融化的电梯按钮",
+    "melted elevator button": "融化的电梯按钮",
+    "elevator buttons": "电梯按钮",
+    "elevator button": "电梯按钮",
+    "elevator doors": "电梯门",
+    "elevator": "电梯",
+    "button": "按钮",
+    "floor 14": "数字 14",
+    "number 14": "数字 14",
+    "lost child at subway station": "地铁站里迷路的小孩",
+    "lost child at subway": "地铁里迷路的小孩",
+    "subway station": "地铁站",
+    "subway": "地铁",
+    "child figure": "小孩",
+    "child": "小孩",
+    "home": "回家的方向",
+    "arrow": "箭头",
+}
+
 
 def _dedupe_preserve_order(items: List[str]) -> List[str]:
     seen = set()
@@ -462,10 +506,21 @@ def _english_anchor_text(text: str) -> str:
     return clean
 
 
+def _zh_anchor_text(text: str) -> str:
+    clean = re.sub(r"\s+", " ", (text or "").strip(" .,:;!?\"'()[]{}"))
+    if not clean:
+        return ""
+    lowered = clean.lower()
+    for source, target in sorted(_EN_TO_ZH_PHRASES.items(), key=lambda item: len(item[0]), reverse=True):
+        if source in lowered:
+            return target
+    return clean
+
+
 def _anchors_for_language(intake: DreamIntake, language: str = "en") -> List[str]:
     anchors = _extract_dream_anchors(intake)
     if _is_zh(language):
-        return anchors
+        return _dedupe_preserve_order([anchor for anchor in (_zh_anchor_text(anchor) for anchor in anchors) if anchor])
     localized = [_english_anchor_text(anchor) for anchor in anchors]
     return _dedupe_preserve_order([anchor for anchor in localized if anchor])
 
@@ -543,9 +598,241 @@ def _summary_from_intake(intake: DreamIntake, language: str = "en") -> str:
     return f"你梦见{clean}"
 
 
+def _user_supplied_text(intake: DreamIntake, answers: str = "", include_mood: bool = False) -> str:
+    parts = [
+        intake.dream_text,
+        intake.voice_transcript,
+        intake.main_question,
+        intake.uncertainty,
+        intake.user_context,
+        answers or "",
+    ]
+    if include_mood:
+        parts.append(intake.mood)
+    return "\n".join(part.strip() for part in parts if part and part.strip())
+
+
+def _question_sentence_candidates(text: str) -> List[str]:
+    pieces = re.split(r"[。！？!?\n\r]+|(?<=\.)\s+|[，,；;]", text or "")
+    return [piece.strip(" ：:「」\"'()[]{}") for piece in pieces if piece.strip()]
+
+
+def _clean_user_question(text: str, language: str = "en") -> str:
+    clean = re.sub(r"\s+", " ", (text or "").strip(" ：:「」\"'()[]{}"))
+    if not clean:
+        return ""
+    if _is_zh(language):
+        clean = re.sub(r"^(我)?(醒来后)?(最)?(想知道|想问|在想|担心|害怕)[：:，,\s]*", "", clean)
+        clean = re.sub(r"^(只)?想知道[：:，,\s]*", "", clean)
+        clean = clean.strip(" ：:「」\"'()[]{}")
+        return clean if clean.endswith(("?", "？")) else f"{clean}？"
+    clean = re.sub(
+        r"^(i\s+)?(woke\s+up\s+)?(want\s+to\s+know|wonder|am\s+wondering|need\s+to\s+know|worry|fear)\s*(if|whether|why|what|how)?\s*",
+        lambda match: (match.group(4) or "").strip() + " ",
+        clean,
+        flags=re.IGNORECASE,
+    ).strip()
+    clean = clean[0].upper() + clean[1:] if clean else ""
+    if clean and not clean.endswith("?") and re.match(r"^(why|what|how|does|do|did|is|am|are|can|could|should|would)\b", clean, re.IGNORECASE):
+        clean += "?"
+    return clean
+
+
+def _extract_explicit_user_question(intake: DreamIntake, answers: str = "", language: str = "en") -> str:
+    if intake.main_question.strip():
+        return _clean_user_question(intake.main_question, language)
+    text = _user_supplied_text(intake, answers)
+    if _is_zh(language):
+        markers = (
+            "为什么",
+            "是不��",
+            "会不会",
+            "怎么办",
+            "我该",
+            "能不能",
+            "代表",
+            "说明",
+            "预兆",
+            "征兆",
+            "想知道",
+            "想问",
+            "撑不住",
+            "走出来",
+            "这么难受",
+        )
+    else:
+        markers = (
+            "why",
+            "what should",
+            "how should",
+            "does this mean",
+            "do i",
+            "am i",
+            "is this",
+            "could this",
+            "can this",
+            "should i",
+            "i wonder",
+            "want to know",
+            "not coping",
+            "cope",
+            "a sign",
+            "omen",
+        )
+    for sentence in _question_sentence_candidates(text):
+        lowered = sentence.lower()
+        if _is_skip_answer(sentence, language):
+            continue
+        if any(marker in lowered for marker in markers):
+            return _clean_user_question(sentence, language)
+    return ""
+
+
+def _emotion_labels_from_text(text: str, language: str = "en") -> List[str]:
+    lowered = (text or "").lower()
+    if _is_zh(language):
+        groups = [
+            ("害怕", ["害怕", "吓", "恐惧", "心慌", "心跳很快", "惊醒"]),
+            ("压力", ["压力", "撑不住", "扛不住", "崩溃", "压垮", "焦虑", "太累", "疲惫"]),
+            ("难过", ["难过", "难受", "伤心", "失落", "想哭", "委屈", "孤独"]),
+            ("自责", ["自责", "内疚", "愧疚", "后悔", "责怪自己"]),
+            ("需要安慰", ["安慰", "被安慰", "抱抱", "关心", "陪陪", "鸡汤"]),
+        ]
+    else:
+        groups = [
+            ("fear", ["scared", "afraid", "terrified", "panic", "panicked", "frightened"]),
+            ("pressure", ["overwhelmed", "not coping", "can't cope", "cannot cope", "burned out", "too much", "anxious"]),
+            ("sadness", ["sad", "grief", "hurt", "lonely", "heartbroken", "upset"]),
+            ("guilt", ["guilty", "ashamed", "blame myself", "regret"]),
+            ("comfort", ["comfort", "reassurance", "not productivity", "not advice", "not a pep talk"]),
+        ]
+    labels: List[str] = []
+    for label, terms in groups:
+        if any(term in lowered for term in terms):
+            labels.append(label)
+    return _dedupe_preserve_order(labels)
+
+
+def _emotion_phrase(labels: List[str], language: str = "en") -> str:
+    if not labels:
+        return "这个感受" if _is_zh(language) else "this feeling"
+    if _is_zh(language):
+        return "、".join(labels)
+    if len(labels) == 1:
+        return labels[0]
+    return ", ".join(labels[:-1]) + f", and {labels[-1]}"
+
+
+def _needs_comfort(text: str, language: str = "en") -> bool:
+    lowered = (text or "").lower()
+    if _is_zh(language):
+        return any(term in lowered for term in ["安慰", "被安慰", "抱抱", "关心", "别催", "不是鸡汤"])
+    return any(term in lowered for term in ["comfort", "reassurance", "not productivity", "not a pep talk", "not advice"])
+
+
+def _should_use_emotion_led_response(intake: DreamIntake, answers: str, language: str = "en") -> bool:
+    user_text = _user_supplied_text(intake, answers)
+    explicit_question = _extract_explicit_user_question(intake, answers, language)
+    labels = _emotion_labels_from_text(user_text, language)
+    if _is_low_context_intake(intake) and _is_skip_answer(answers, language) and not explicit_question and not _needs_comfort(user_text, language):
+        return False
+    return bool(explicit_question or labels or _needs_comfort(user_text, language))
+
+
+def _direct_question_reassurance(question: str, labels: List[str], language: str = "en") -> str:
+    lowered = (question or "").lower()
+    if _is_zh(language):
+        if any(term in lowered for term in ["撑不住", "扛不住", "崩溃"]):
+            return "它不等于你一定撑不住了；更像是在提醒你，最近的压力已经值得被认真照顾。"
+        if "走出来" in lowered:
+            return "它不急着证明你有没有走出来；更像是在说明这段难受还需要一点被看见的时间。"
+        if any(term in lowered for term in ["预兆", "征兆", "坏事", "会不会发生"]):
+            return "它不适合被当作预兆；我们先把它当作醒来后仍在身体里的担心来照顾。"
+        if any(term in lowered for term in ["我是不是", "是不是我"]):
+            return "它不是给你贴标签的证据；它更像是在把一个需要被接住的感受放到你面前。"
+        if "为什么" in lowered or "这么难受" in lowered:
+            return "这么难受不是你反应过度；梦可能只是把还没放下的感受放大给你看。"
+        emotion = _emotion_phrase(labels, language)
+        return f"你提到的{emotion}不是小事，它值得先被接住，而不是立刻被解释掉。"
+    if any(term in lowered for term in ["not coping", "cope", "coping", "falling apart"]):
+        return "This does not prove you are failing to cope; it may be showing that your pressure deserves care before answers."
+    if "sign" in lowered or "omen" in lowered or "bad will happen" in lowered:
+        return "This is safer to treat as a fear to care for, not as evidence that something bad will happen."
+    if "am i" in lowered or "does this mean i" in lowered:
+        return "This dream is not evidence for a label about you; it is one feeling asking to be met gently."
+    if "why" in lowered:
+        return "The ache makes sense; the dream may be enlarging a feeling that has not had enough room yet."
+    emotion = _emotion_phrase(labels, language)
+    return f"The {emotion} you named deserves to be met first, not explained away too quickly."
+
+
+def _emotion_led_interpretation(intake: DreamIntake, answers: str, anchors: List[str], language: str = "en") -> str:
+    if not _should_use_emotion_led_response(intake, answers, language):
+        return ""
+    question = _extract_explicit_user_question(intake, answers, language)
+    labels = _emotion_labels_from_text(_user_supplied_text(intake, answers), language)
+    emotion = _emotion_phrase(labels, language)
+    anchor = _answer_bridge_anchor(anchors)
+    direct = _direct_question_reassurance(question, labels, language)
+    if _is_zh(language):
+        opener = f"你问的是「{question}」。" if question else ""
+        return (
+            f"{opener}{direct} 第一层，先承认醒来后的{emotion}是真实的，不需要被责怪。"
+            f"第二层，梦里的「{anchor}」也许把这种感受变成了一个可以看的画面。"
+            "第三层，今天不急着找唯一答案，先给自己一个能站稳的小支点。"
+        )
+    opener = f"You asked, \"{question}\" " if question else ""
+    return (
+        f"{opener}{direct} First, let the {emotion} be real without blaming yourself. "
+        f"Second, the {_anchor_with_article(anchor)} may be the dream's concrete shape for that feeling. "
+        "Third, for today the goal is not to solve the whole dream, but to give yourself one steadier place to stand."
+    )
+
+
+def _emotion_led_today_tip(intake: DreamIntake, answers: str, anchors: List[str], language: str = "en") -> str:
+    if not _should_use_emotion_led_response(intake, answers, language):
+        return ""
+    if _has_prophecy_frame(_user_supplied_text(intake, answers)):
+        return ""
+    labels = _emotion_labels_from_text(_user_supplied_text(intake, answers), language)
+    emotion = _emotion_phrase(labels, language)
+    anchor = anchors[0]
+    if _is_zh(language):
+        return (
+            f"今天先把「{anchor}」当作{emotion}的形状，而不是结论：写一句“我现在被什么淹到/刺痛”，"
+            "再写一句“我此刻可以抓住的一个小安定”。"
+        )
+    return (
+        f"For today, treat the {anchor} as the shape of {emotion}, not a verdict: write one line naming "
+        "what feels heavy, then one line naming a small steady thing you can hold now."
+    )
+
+
+def _emotion_led_tiny_action(intake: DreamIntake, answers: str, anchors: List[str], language: str = "en") -> str:
+    if not _should_use_emotion_led_response(intake, answers, language):
+        return ""
+    anchor = anchors[0]
+    if _is_zh(language):
+        return f"用 5 分钟写两行：1.「{anchor}」让我最难受的是…… 2. 现在我能给自己的一点安定是……"
+    return f"Spend five minutes writing two lines: 1. The {anchor} hurts most because... 2. One steady thing I can offer myself now is..."
+
+
+def _emotion_led_caring_note(intake: DreamIntake, answers: str, language: str = "en") -> str:
+    if not _should_use_emotion_led_response(intake, answers, language):
+        return ""
+    labels = _emotion_labels_from_text(_user_supplied_text(intake, answers), language)
+    emotion = _emotion_phrase(labels, language)
+    if _is_zh(language):
+        return f"你不是太脆弱，也不是需要被催着立刻想通；这份{emotion}可以先被轻轻接住。"
+    return f"You are not weak for feeling {emotion}; you do not have to turn it into a lesson before you are comforted."
+
+
 def _main_question_from_intake(intake: DreamIntake, language: str = "en") -> str:
     if intake.main_question.strip():
-        return intake.main_question.strip()
+        return _clean_user_question(intake.main_question, language)
+    explicit_question = _extract_explicit_user_question(intake, "", language)
+    if explicit_question:
+        return explicit_question
     task = _task_focus(intake.merged_text(), language)
     if task:
         if not _is_zh(language):
@@ -1089,14 +1376,20 @@ def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = ""
         setattr(polished, field, _clean_placeholder_phrase(getattr(polished, field)))
     if not polished.dream_summary.strip() or _is_placeholder_anchor(polished.dream_summary) or not _text_uses_anchor(polished.dream_summary, anchors):
         polished.dream_summary = _summary_from_intake(intake, language)
-    if (
+    explicit_question = _extract_explicit_user_question(intake, answers, language)
+    if explicit_question:
+        polished.main_question = explicit_question
+    elif (
         not polished.main_question.strip()
         or _is_placeholder_anchor(polished.main_question)
         or not _text_uses_anchor(polished.main_question, anchors)
     ):
         polished.main_question = _main_question_from_intake(intake, language)
+    emotion_interpretation = _emotion_led_interpretation(intake, answers, anchors, language)
     answer_interpretation = _answer_based_interpretation(answers, _answer_bridge_anchor(anchors), language)
-    if answer_interpretation:
+    if emotion_interpretation:
+        polished.interpretation = emotion_interpretation
+    elif answer_interpretation:
         polished.interpretation = answer_interpretation
     elif _has_prophecy_frame(intake.merged_text()):
         anchor = _answer_bridge_anchor(anchors)
@@ -1115,8 +1408,11 @@ def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = ""
     elif not polished.interpretation.strip() or not _anchor_in_text(polished.interpretation, anchors):
         polished.interpretation = _fallback_interpretation(intake, language)
     generic_tip_markers = ["drink water", "hydrate", "多休息", "保持积极", "take a walk"]
+    emotion_tip = _emotion_led_today_tip(intake, answers, anchors, language)
     answer_tip = _answer_based_today_tip(answers, anchors[0], language)
-    if answer_tip:
+    if emotion_tip:
+        polished.today_tip = emotion_tip
+    elif answer_tip:
         polished.today_tip = answer_tip
     elif _has_prophecy_frame(intake.merged_text()):
         anchor = anchors[0]
@@ -1140,8 +1436,11 @@ def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = ""
     ):
         polished.today_tip = _grounded_today_tip(intake, language)
     hard_action_markers = ["address it immediately", "fix it immediately", "solve it immediately"]
+    emotion_action = _emotion_led_tiny_action(intake, answers, anchors, language)
     answer_action = _answer_based_tiny_action(answers, language)
-    if answer_action:
+    if emotion_action:
+        polished.tiny_action = emotion_action
+    elif answer_action:
         polished.tiny_action = answer_action
     elif _has_prophecy_frame(intake.merged_text()):
         polished.tiny_action = (
@@ -1165,7 +1464,10 @@ def _polish_today_tip(card: TodayTipCard, intake: DreamIntake, answers: str = ""
             polished.tiny_action = f"用 5 分钟写下：今天和「{anchors[0]}」有关的第一小步是什么？"
         else:
             polished.tiny_action = f"Spend five minutes writing the first small step connected to the {anchors[0]}."
-    if not polished.caring_note.strip():
+    emotion_caring_note = _emotion_led_caring_note(intake, answers, language)
+    if emotion_caring_note:
+        polished.caring_note = emotion_caring_note
+    elif not polished.caring_note.strip():
         polished.caring_note = (
             "你不需要一醒来就解决整个梦，先把一个细节照亮就很好。"
             if _is_zh(language)

tests/test_ollama_models.py

@@ -242,3 +242,44 @@ def test_hosted_vision_client_accepts_partial_witness_json():
     assert "Object: elevator button" in clues
     assert "Visible text: 14" in clues
     assert "Mood cue: cold" in clues
+
+
+def test_hosted_vision_client_extracts_modal_pipeline_assistant_content():
+    class StubHostedNestedVisionClient(HostedMiniCPMVisionClient):
+        def _post_image(self, image_path):
+            return {
+                "response": str(
+                    {
+                        "input_text": [
+                            {
+                                "role": "user",
+                                "content": [
+                                    {"type": "image", "path": "/tmp/demo.png"},
+                                    {"type": "text", "text": "Describe the sketch."},
+                                ],
+                            }
+                        ],
+                        "generated_text": [
+                            {"role": "user", "content": [{"type": "text", "text": "Describe the sketch."}]},
+                            {
+                                "role": "assistant",
+                                "content": (
+                                    '{"scene_summary":"A child is lost at a subway station.",'
+                                    '"objects":["child figure","subway sign","arrow"],'
+                                    '"visible_text":["SUBWAY","HOME ?"],'
+                                    '"spatial_relations":"child in front of sign",'
+                                    '"mood_cues":"lost or searching",'
+                                    '"uncertain_details":"where home is",'
+                                    '"surprising_detail":"the sign asks HOME ?"}'
+                                ),
+                            },
+                        ],
+                    }
+                )
+            }
+
+    clues = StubHostedNestedVisionClient(endpoint="https://example.test").extract_witness("demo.png").to_visual_clues()
+
+    assert "Scene: A child is lost at a subway station." in clues
+    assert "Object: child figure" in clues
+    assert "Visible text: HOME ?" in clues

tests/test_pipeline.py

@@ -178,6 +178,73 @@ def test_generate_today_tip_keeps_answer_history_and_removes_placeholder_text():
     assert "does not indicate any real-life concerns" not in combined
 
 
+def test_generate_today_tip_follows_user_question_and_comfort_need_in_chinese():
+    intake = build_intake(
+        dream_text="我梦到自己掉进海里，醒来很害怕。我想知道这是不是说明我快撑不住了？",
+        mood="害怕",
+    )
+
+    card = generate_today_tip(intake, "我最近工作压力很大，真的很想被安慰一下。", FakeTextClient(), language="zh")
+    combined = "\n".join(
+        [
+            card.main_question,
+            card.interpretation,
+            card.today_tip,
+            card.tiny_action,
+            card.caring_note,
+        ]
+    )
+
+    assert "撑不住" in card.main_question
+    assert "海" in combined
+    assert "压力" in combined or "害怕" in combined
+    assert "第一层" in card.interpretation and "第二层" in card.interpretation
+    assert "太脆弱" in card.caring_note
+    assert "电梯" not in combined
+    assert "只打开那件事" not in combined
+
+
+def test_generate_today_tip_answers_sad_relationship_question_without_productivity_template():
+    intake = build_intake(
+        dream_text="我梦到前任发消息又消失了，我醒来很难过，想知道是不是我还没走出来。",
+        mood="难过",
+    )
+
+    card = generate_today_tip(intake, "我不想要鸡汤，只想知道为什么这么难受。", FakeTextClient(), language="zh")
+    combined = "\n".join([card.main_question, card.interpretation, card.today_tip, card.caring_note])
+
+    assert "为什么这么难受" in card.main_question or "没走出来" in card.main_question
+    assert "难受" in combined or "难过" in combined
+    assert "消息" in combined
+    assert "第一句话" not in combined
+    assert "打开任务" not in combined
+
+
+def test_generate_today_tip_keeps_english_emotional_question_from_becoming_task_advice():
+    intake = build_intake(
+        dream_text="I dreamed I was drowning in dark water and woke scared. Does this mean I am not coping?",
+        mood="scared",
+    )
+
+    card = generate_today_tip(intake, "I need comfort, not productivity advice.", FakeTextClient(), language="en")
+    combined = "\n".join(
+        [
+            card.main_question,
+            card.interpretation,
+            card.today_tip,
+            card.tiny_action,
+            card.caring_note,
+        ]
+    ).lower()
+
+    assert "not coping" in card.main_question.lower()
+    assert "water" in combined
+    assert "first" in card.interpretation.lower() and "second" in card.interpretation.lower()
+    assert "not weak" in combined or "comforted" in combined
+    assert "open the task" not in combined
+    assert "first line" not in combined
+
+
 def test_generate_today_tip_adds_safety_note_for_repeated_insomnia_without_self_harm():
     intake = build_intake(
         dream_text="我昨晚反复梦见自己在一条漆黑的走廊里走，醒来后心跳很快，已经连续3晚都睡不好。",
@@ -536,6 +603,38 @@ def test_visual_witness_clues_drive_questions_and_today_tip():
     assert "objects" not in combined
 
 
+def test_zh_text_and_image_keep_user_question_while_using_visual_anchors():
+    class SeaDreamVision:
+        def extract_witness(self, image_path):
+            return VisionWitness(
+                scene_summary="A dreamlike representation of a sea under night.",
+                objects=["stick figure", "wavy water", "crescent moon"],
+                visible_text=["dark sea dream"],
+                mood_cues=["small figure in a large dark place"],
+            )
+
+        def extract_clues(self, image_path):
+            return ["flat fallback should not win"]
+
+    session = add_evidence(
+        create_session(language="zh"),
+        dream_text="我醒来很害怕，这张草图是梦里最清楚的画面。我想知道为什么它让我这么慌。",
+        image_path="sea.png",
+        mood="害怕",
+        vision_client=SeaDreamVision(),
+        asr_client=FakeASRClient(),
+        language="zh",
+    )
+    card = generate_today_tip(session.intake, "", FakeTextClient(), language="zh")
+    combined = "\n".join([card.main_question, ",".join(card.dream_anchors), card.interpretation, card.today_tip])
+
+    assert "为什么它让我这么慌？" in card.main_question
+    assert any(anchor in card.dream_anchors for anchor in ["夜晚的海", "海浪", "月牙", "漆黑的海"])
+    assert "blue hallway" not in combined
+    assert "dreamlike representation" not in combined.lower()
+    assert "这么难受不是你反应过度" in card.interpretation
+
+
 def test_witness_failure_keeps_text_path_alive():
     class BrokenWitnessVision:
         def extract_witness(self, image_path):