fix(HEL-623): split dual-prompt text encoders; positive wiring wins; detect duration via expression

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: multica-agent <github@multica.ai>

calliope-backend/src/calliope/comfyui/parser.py

@@ -78,6 +78,7 @@ _SAMPLER_CLASSES = frozenset({"KSampler", "KSamplerAdvanced"})
 _NEGATIVE_WORDS = ("negative", "负面", "反向", "反面")
 _SYSTEM_WORDS = ("system", "系统")
 _PROMPT_WORDS = ("prompt", "positive", "提示", "正向")
+_NEGATIVE_TEXT_FIELDS = ("negative_prompt", "negative_text", "negative")
 
 # Title keyword → (role, Chinese label) for untagged numeric primitives.
 _NUMBER_TITLE_ROLES: tuple[tuple[tuple[str, ...], str, str], ...] = (
@@ -121,22 +122,66 @@ def _consumers(workflow: dict[str, Any]) -> dict[str, list[tuple[str, str]]]:
 def _feeds_negative(
     node_id: str, consumers: dict[str, list[tuple[str, str]]], depth: int = 4
 ) -> bool:
-    """True when this node's output reaches a sampler/guider ``negative`` input."""
+    """True when this node's output reaches a sampler/guider ``negative`` input
+    and never a ``positive`` one.
+
+    A text encoder wired into both (one conditioning reused for positive and
+    negative) is treated as the positive prompt; only a negative-sounding
+    title makes it negative (see ``_is_negative``).
+    """
     frontier = [node_id]
     seen = {node_id}
+    negative = False
     for _ in range(depth):
         nxt: list[str] = []
         for nid in frontier:
             for cid, name in consumers.get(nid, []):
-                if name == "negative":
-                    return True
                 if name == "positive":
+                    return False
+                if name == "negative":
+                    negative = True
                     continue
                 if cid not in seen:
                     seen.add(cid)
                     nxt.append(cid)
         frontier = nxt
-    return False
+    return negative
+
+
+_FRAME_INPUTS = frozenset({"length", "num_frames", "frames", "video_length", "frame_count"})
+_SECONDS_INPUTS = frozenset({"duration", "seconds", "duration_seconds"})
+_EXPRESSION_HINTS = ("expression", "math")
+
+
+def _duration_hint(
+    node_id: str,
+    workflow: dict[str, Any],
+    consumers: dict[str, list[tuple[str, str]]],
+    depth: int = 3,
+) -> str | None:
+    """How an untitled number feeds the video length downstream.
+
+    ``"seconds"`` — it reaches a seconds input, or a frame-count input only
+    after passing through a math/expression node (seconds × fps → frames);
+    ``"frames"`` — it is wired straight into a frame-count input.
+    """
+    frontier = [(node_id, False)]
+    seen = {node_id}
+    for _ in range(depth):
+        nxt: list[tuple[str, bool]] = []
+        for nid, via_expr in frontier:
+            for cid, name in consumers.get(nid, []):
+                if name in _SECONDS_INPUTS:
+                    return "seconds"
+                if name in _FRAME_INPUTS:
+                    return "seconds" if via_expr else "frames"
+                if cid in seen:
+                    continue
+                seen.add(cid)
+                ct = str((workflow.get(cid) or {}).get("class_type") or "").lower()
+                nxt.append((cid, via_expr or any(h in ct for h in _EXPRESSION_HINTS)))
+        frontier = nxt
+    return None
 
 
 def _entry(
@@ -224,7 +269,15 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
             continue
         if ct == "CLIPTextEncode" or "TextEncode" in ct:
             field = next((f for f in ("text", "prompt") if isinstance(inputs.get(f), str)), None)
-            if field:
+            neg_field = next(
+                (f for f in _NEGATIVE_TEXT_FIELDS if isinstance(inputs.get(f), str)), None
+            )
+            if field and neg_field:
+                # One encoder carrying both prompts (TextEncodeQwenImage21…):
+                # its outputs feed positive *and* negative, so wiring says nothing.
+                texts.append((nid, node, field, False))
+                texts.append((nid, node, neg_field, True))
+            elif field:
                 texts.append((nid, node, field, _is_negative(nid, lowered, consumers)))
             continue
 
@@ -253,6 +306,12 @@ def detect_untagged_inputs(workflow: dict[str, Any]) -> list[dict[str, Any]]:
                 if any(w in lowered for w in words):
                     role, label = r, (title if title and title != ct else zh)
                     break
+            if role is None and (not title or title == ct):
+                hint = _duration_hint(nid, workflow, consumers)
+                if hint == "seconds":
+                    role, label = "duration", "时长（秒）"
+                elif hint == "frames":
+                    label = "视频帧数"
             if role and role in taken_roles:
                 role = None
             if role:

calliope-backend/tests/fixtures/comfy_native/wen_sheng_shipin_API.json

@@ -0,0 +1,301 @@
+{
+  "1": {
+    "inputs": {
+      "value": "PASTE YOUR PROMPT HERE. Use the MiniMax-H3 prompt contract (integrated_multimodal_description / overall_soundscape / non_diegetic_music). The previews in this repo were made with a ten-shot subway prompt posted by a community member; it is not redistributed here."
+    },
+    "class_type": "PrimitiveStringMultiline",
+    "_meta": {
+      "title": "PROMPT (feeds every stage)"
+    }
+  },
+  "92": {
+    "inputs": {
+      "filename_prefix": "video/01_reference_4step_sla",
+      "format": "auto",
+      "format.codec": "auto",
+      "codec": "auto",
+      "video": [
+        "130",
+        0
+      ]
+    },
+    "class_type": "SaveVideo",
+    "_meta": {
+      "title": "REFERENCE 2026-08-29: fused ref-delta bake + SLA 0.9/64, res_multistep/simple/4, shift 12/3, 1152x640, 243 f, seed 42 -> subway_bake_sla_00001_.mp4"
+    }
+  },
+  "115": {
+    "inputs": {
+      "aspect_ratio": "16:9 (Widescreen)",
+      "megapixels": 0.7,
+      "multiple": 32
+    },
+    "class_type": "ResolutionSelector",
+    "_meta": {
+      "title": "ResolutionSelector"
+    }
+  },
+  "116": {
+    "inputs": {
+      "expression": "a"
+    },
+    "class_type": "ComfyMathExpression",
+    "_meta": {
+      "title": "ComfyMathExpression"
+    }
+  },
+  "119": {
+    "inputs": {
+      "vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "VAELoader"
+    }
+  },
+  "120": {
+    "inputs": {
+      "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "VAELoader"
+    }
+  },
+  "121": {
+    "inputs": {
+      "samples": [
+        "125",
+        0
+      ],
+      "vae": [
+        "120",
+        0
+      ]
+    },
+    "class_type": "VAEDecodeAudio",
+    "_meta": {
+      "title": "VAEDecodeAudio"
+    }
+  },
+  "122": {
+    "inputs": {
+      "samples": [
+        "125",
+        0
+      ],
+      "vae": [
+        "119",
+        0
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAEDecode"
+    }
+  },
+  "124": {
+    "inputs": {
+      "scheduler": "simple",
+      "steps": 4,
+      "denoise": 1,
+      "model": [
+        "142",
+        0
+      ]
+    },
+    "class_type": "BasicScheduler",
+    "_meta": {
+      "title": "BasicScheduler"
+    }
+  },
+  "125": {
+    "inputs": {
+      "noise": [
+        "129",
+        0
+      ],
+      "guider": [
+        "126",
+        0
+      ],
+      "sampler": [
+        "135",
+        0
+      ],
+      "sigmas": [
+        "124",
+        0
+      ],
+      "latent_image": [
+        "131",
+        1
+      ]
+    },
+    "class_type": "SamplerCustomAdvanced",
+    "_meta": {
+      "title": "SamplerCustomAdvanced"
+    }
+  },
+  "126": {
+    "inputs": {
+      "model": [
+        "142",
+        0
+      ],
+      "conditioning": [
+        "131",
+        0
+      ]
+    },
+    "class_type": "BasicGuider",
+    "_meta": {
+      "title": "BasicGuider"
+    }
+  },
+  "127": {
+    "inputs": {
+      "unet_name": "minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot.safetensors",
+      "weight_dtype": "default"
+    },
+    "class_type": "UNETLoader",
+    "_meta": {
+      "title": "UNETLoader"
+    }
+  },
+  "128": {
+    "inputs": {
+      "clip_name": "qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors",
+      "type": "minimax",
+      "device": "default"
+    },
+    "class_type": "CLIPLoader",
+    "_meta": {
+      "title": "CLIPLoader"
+    }
+  },
+  "129": {
+    "inputs": {
+      "noise_seed": 42
+    },
+    "class_type": "RandomNoise",
+    "_meta": {
+      "title": "RandomNoise"
+    }
+  },
+  "130": {
+    "inputs": {
+      "fps": 24,
+      "bit_depth": 8,
+      "color_space": "sRGB",
+      "codec": "none",
+      "images": [
+        "122",
+        0
+      ],
+      "audio": [
+        "121",
+        0
+      ]
+    },
+    "class_type": "CreateVideo",
+    "_meta": {
+      "title": "CreateVideo"
+    }
+  },
+  "131": {
+    "inputs": {
+      "prompt": [
+        "1",
+        0
+      ],
+      "width": [
+        "115",
+        0
+      ],
+      "height": [
+        "115",
+        1
+      ],
+      "length": [
+        "132",
+        1
+      ],
+      "clip": [
+        "128",
+        0
+      ],
+      "vae": [
+        "119",
+        0
+      ]
+    },
+    "class_type": "MiniMaxH3ImageToVideo",
+    "_meta": {
+      "title": "MiniMaxH3ImageToVideo"
+    }
+  },
+  "132": {
+    "inputs": {
+      "expression": "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17",
+      "values.a": [
+        "133",
+        0
+      ]
+    },
+    "class_type": "ComfyMathExpression",
+    "_meta": {
+      "title": "ComfyMathExpression"
+    }
+  },
+  "133": {
+    "inputs": {
+      "value": 10
+    },
+    "class_type": "PrimitiveFloat",
+    "_meta": {
+      "title": "PrimitiveFloat"
+    }
+  },
+  "135": {
+    "inputs": {
+      "sampler_name": "res_multistep"
+    },
+    "class_type": "KSamplerSelect",
+    "_meta": {
+      "title": "KSamplerSelect"
+    }
+  },
+  "142": {
+    "inputs": {
+      "shift_video": 12,
+      "shift_audio": 3,
+      "model": [
+        "900",
+        0
+      ]
+    },
+    "class_type": "MiniMaxH3SigmaShift",
+    "_meta": {
+      "title": "MiniMaxH3SigmaShift"
+    }
+  },
+  "900": {
+    "inputs": {
+      "sparsity_ratio": 0.9,
+      "block_size": "64",
+      "min_seq_len": 8192,
+      "dense_last_steps": 0,
+      "protect_audio": true,
+      "enabled": true,
+      "model": [
+        "127",
+        0
+      ]
+    },
+    "class_type": "H3SLAAttention",
+    "_meta": {
+      "title": "H3 SLA Attention 0.9/64 (their graph)"
+    }
+  }
+}

calliope-backend/tests/fixtures/comfy_native/wen_sheng_tu_API.json

@@ -0,0 +1,137 @@
+{
+  "13": {
+    "inputs": {
+      "aspect_ratio": "1:1 (Square)",
+      "megapixels": 1,
+      "multiple": 8
+    },
+    "class_type": "ResolutionSelector",
+    "_meta": {
+      "title": "分辨率选择器"
+    }
+  },
+  "461": {
+    "inputs": {
+      "filename_prefix": "Qwen_image_2.1",
+      "format": "png",
+      "format.bit_depth": "8-bit",
+      "format.input_color_space": "sRGB",
+      "images": [
+        "459:457",
+        0
+      ]
+    },
+    "class_type": "SaveImageAdvanced",
+    "_meta": {
+      "title": "保存图像（高级）"
+    }
+  },
+  "459:451": {
+    "inputs": {
+      "unet_name": "qwen_image_2.1_int8_convrot.safetensors",
+      "weight_dtype": "default"
+    },
+    "class_type": "UNETLoader",
+    "_meta": {
+      "title": "UNet加载器"
+    }
+  },
+  "459:452": {
+    "inputs": {
+      "prompt": "Greyscale fashion editorial portrait of an avant-garde woman tilted upwards in profile, captured with striking high-contrast chiaroscuro lighting. She wears oversized, thick-rimmed circular black sunglasses with matte dark lenses, concealing her upward gaze. Her attire merges high couture with tactical abstraction: a high-necked structural dress tailored from heavyweight fabric featuring a disrupted optical camouflage pattern—interlocking oversized polka dots, pixelated stippling, and organic contour blotches in stark monochrome that mimic military disruptive patterns in a refined runway style. Around her shoulders, asymmetric pleated webbing and tailored cargo straps are integrated seamlessly into the garment’s architecture. In her gloved hand, she clutches a structured matte-leather geometric handbag accented with subtle tactical buckles, utility stitching, and dark brushed-metal hardware. Her skin texture is rendered with hyper-detailed clarity, sculpted by intense, razor-sharp studio key light and deep graphite shadows. The striking background is a vivid mixed-media graphic collage, composed of flat vibrant lime green planes, stark black and white overlapping circles, segmented semi-circles with bold zebra striping, dense micro-halftone dot matrices, and vector topographic contour lines. The clash between monochromatic disruptive camo elements on her clothing and the saturated, retro-futuristic pop-art backdrop creates an intense visual dissonance, emphasizing sharp silhouettes, paper cut-out edges, dynamic compositional balance, and clean graphic novel textures. Absolutely no text, no typography, no letters, no words, no numbers, no logos, no watermark, no captions, pure imagery only.",
+      "negative_prompt": "",
+      "resolution": 1024,
+      "clip": [
+        "459:453",
+        0
+      ]
+    },
+    "class_type": "TextEncodeQwenImage21",
+    "_meta": {
+      "title": "Text Encode Qwen Image 2.1"
+    }
+  },
+  "459:453": {
+    "inputs": {
+      "clip_name": "qwen3vl_8b_int8_convrot.safetensors",
+      "type": "qwen_image",
+      "device": "default"
+    },
+    "class_type": "CLIPLoader",
+    "_meta": {
+      "title": "加载CLIP"
+    }
+  },
+  "459:454": {
+    "inputs": {
+      "vae_name": "qwen_image_2.1_vae_bf16.safetensors"
+    },
+    "class_type": "VAELoader",
+    "_meta": {
+      "title": "加载VAE"
+    }
+  },
+  "459:456": {
+    "inputs": {
+      "width": [
+        "13",
+        0
+      ],
+      "height": [
+        "13",
+        1
+      ],
+      "batch_size": 1
+    },
+    "class_type": "EmptyLatentImage",
+    "_meta": {
+      "title": "空Latent图像"
+    }
+  },
+  "459:457": {
+    "inputs": {
+      "samples": [
+        "459:458",
+        0
+      ],
+      "vae": [
+        "459:454",
+        0
+      ]
+    },
+    "class_type": "VAEDecode",
+    "_meta": {
+      "title": "VAE解码"
+    }
+  },
+  "459:458": {
+    "inputs": {
+      "seed": 593103825222985,
+      "steps": 25,
+      "cfg": 1,
+      "sampler_name": "euler",
+      "scheduler": "simple",
+      "denoise": 1,
+      "model": [
+        "459:451",
+        0
+      ],
+      "positive": [
+        "459:452",
+        0
+      ],
+      "negative": [
+        "459:452",
+        1
+      ],
+      "latent_image": [
+        "459:456",
+        0
+      ]
+    },
+    "class_type": "KSampler",
+    "_meta": {
+      "title": "K采样器"
+    }
+  }
+}

calliope-backend/tests/test_untagged_workflows.py

@@ -70,6 +70,85 @@ def test_untagged_h3_video_detects_prompt_duration_seed():
     assert roles["seed"]["nodeId"] == "217@noise_seed"
 
 
+def test_real_qwen21_t2i_splits_single_encoder_into_both_prompts():
+    """白栖知的 wen-sheng-tu.json：一个 TextEncodeQwenImage21 同时喂 positive/negative."""
+    wf = _load("wen_sheng_tu_API.json")
+    inputs = parse_dynamic_inputs(wf)
+    roles = _by_role(inputs)
+    assert roles["prompt"]["nodeId"] == "459:452@prompt"
+    assert roles["prompt"]["label"] == "正向提示词"
+    assert roles["prompt"]["required"] is True
+    assert roles["prompt"]["defaultValue"].startswith("Greyscale fashion editorial")
+    assert roles["negative"]["nodeId"] == "459:452@negative_prompt"
+    assert roles["negative"]["label"] == "反向提示词"
+    assert roles["negative"]["required"] is False
+    assert roles["seed"]["nodeId"] == "459:458@seed"
+    assert roles["steps"]["nodeId"] == "459:458@steps"
+    assert roles["cfg"]["nodeId"] == "459:458@cfg"
+    assert roles["batch"]["nodeId"] == "459:456@batch_size"
+    # width/height are wired from the resolution selector — not exposed
+    assert "width" not in roles and "height" not in roles
+    assert [i for i in inputs if i["required"]] == [roles["prompt"]]
+
+    patched = patch_workflow(
+        wf, {roles["prompt"]["nodeId"]: "新正向", roles["negative"]["nodeId"]: "模糊"}
+    )
+    enc = patched["459:452"]["inputs"]
+    assert enc["prompt"] == "新正向"
+    assert enc["negative_prompt"] == "模糊"
+    assert enc["clip"] == ["459:453", 0]
+
+
+def test_real_h3_t2v_duration_from_expression_primitive():
+    """wen-sheng-shipin.json：无标题 PrimitiveFloat 经表达式换算帧数 → 时长（秒）."""
+    inputs = parse_dynamic_inputs(_load("wen_sheng_shipin_API.json"))
+    roles = _by_role(inputs)
+    assert roles["prompt"]["nodeId"] == "1"
+    assert roles["prompt"]["required"] is True
+    assert roles["duration"]["nodeId"] == "133"
+    assert roles["duration"]["label"] == "时长（秒）"
+    assert roles["duration"]["defaultValue"] == 10
+    assert roles["seed"]["nodeId"] == "129@noise_seed"
+    assert roles["steps"]["nodeId"] == "124@steps"
+    assert "negative" not in roles
+
+
+def _encoder_graph(encoder: dict, positive_slot: list, negative_slot: list) -> dict:
+    return {
+        "1": encoder,
+        "2": {
+            "class_type": "KSampler",
+            "inputs": {"seed": 1, "steps": 20, "cfg": 4, "positive": positive_slot,
+                       "negative": negative_slot},
+        },
+    }
+
+
+def test_custom_encoder_with_text_and_negative_prompt_is_split():
+    wf = _encoder_graph(
+        {"class_type": "MyTextEncodeDual", "inputs": {"text": "猫", "negative_prompt": "糊"},
+         "_meta": {"title": "MyTextEncodeDual"}},
+        ["1", 0],
+        ["1", 1],
+    )
+    roles = _by_role(parse_dynamic_inputs(wf))
+    assert roles["prompt"]["nodeId"] == "1"
+    assert roles["prompt"]["defaultValue"] == "猫"
+    assert roles["negative"]["nodeId"] == "1@negative_prompt"
+
+
+def test_clip_encoder_wired_to_both_positive_and_negative_stays_positive():
+    wf = _encoder_graph(
+        {"class_type": "CLIPTextEncode", "inputs": {"text": "猫"},
+         "_meta": {"title": "CLIP Text Encode"}},
+        ["1", 0],
+        ["1", 0],
+    )
+    roles = _by_role(parse_dynamic_inputs(wf))
+    assert roles["prompt"]["nodeId"] == "1"
+    assert "negative" not in roles
+
+
 def test_tagged_workflows_are_unchanged():
     for path in sorted(EXAMPLES.glob("*.json")):
         inputs = parse_dynamic_inputs(json.loads(path.read_text(encoding="utf-8")))
@@ -158,6 +237,35 @@ def test_native_workflow_full_chain(client, fake_comfy):  # noqa: F811
         assert any(j["id"] == job["id"] and j["status"] == "done" for j in history)
 
 
+def test_real_qwen21_t2i_full_chain(client, fake_comfy):  # noqa: F811
+    with _real_run(client, fake_comfy):
+        wf_id = _import_workflow(
+            client, "真实·Qwen2.1 文生图", "image", _load("wen_sheng_tu_API.json")
+        )
+        roles = _by_role(client.get(f"/api/workflows/{wf_id}").json()["input_schema"])
+        r = client.post(
+            "/api/playground/generate",
+            json={
+                "workflow_id": wf_id,
+                "input_values": {
+                    roles["prompt"]["nodeId"]: "雨夜霓虹街道",
+                    roles["negative"]["nodeId"]: "低清晰度",
+                    roles["seed"]["nodeId"]: 7,
+                },
+                "source": "studio-image",
+                "endpoint_id": settings.comfyui_active_id,
+            },
+        )
+        assert r.status_code == 200, r.text
+        job = _wait_done(client, r.json()["job"]["id"])
+        assert job["status"] == "done", job.get("error")
+        queued = fake_comfy.prompts[-1]["prompt"]
+        assert queued["459:452"]["inputs"]["prompt"] == "雨夜霓虹街道"
+        assert queued["459:452"]["inputs"]["negative_prompt"] == "低清晰度"
+        assert queued["459:458"]["inputs"]["seed"] == 7
+        assert queued["459:458"]["inputs"]["negative"] == ["459:452", 1]
+
+
 def test_comfy_rejection_is_reported_in_chinese(client, fake_comfy):  # noqa: F811
     fake_comfy.fail_prompt = True
     with _real_run(client, fake_comfy):