feat: add visualize capability with Chart.js/SVG rendering pipeline

- Add visualize agent pipeline (analysis → code generation → review)
- Add VisualizeCapability with request config and registry wiring
- Add frontend VisualizeConfigPanel, VisualizationViewer components
- Add chart.js and react-chartjs-2 dependencies
- Improve math_animator YON_IMAGE anchor block prompts (en/zh)
- Add minimax model override (supports_response_format: false)
- Suppress noisy uvicorn WebSocket connection logs
- Add selective HTTP access logging middleware (non-200 only)

Made-with: Cursor

Author: pancacake

README.md

@@ -41,7 +41,7 @@
 
 > **[2026.4.7]** [v1.0.0-beta.2](https://github.com/HKUDS/DeepTutor/releases/tag/v1.0.0-beta.2) — Runtime cache invalidation for hot settings reload, MinerU nested output support, mimic WebSocket fix, Python 3.11+ minimum, and CI improvements.
 
-> **[2026.4.4]** [v1.0.0-beta.1](https://github.com/HKUDS/DeepTutor/releases/tag/v1.0.0-beta.1) — Agent-native architecture rewrite (DeepTutor 2.0) with two-layer plugin model (Tools + Capabilities), CLI & SDK entry points, TutorBot multi-channel bot agent, Co-Writer, Guided Learning, and persistent memory.
+> **[2026.4.4]** [v1.0.0-beta.1](https://github.com/HKUDS/DeepTutor/releases/tag/v1.0.0-beta.1) — Agent-native architecture rewrite (～200k lines) with two-layer plugin model (Tools + Capabilities), CLI & SDK entry points, TutorBot multi-channel bot agent, Co-Writer, Guided Learning, and persistent memory.
 
 <details>
 <summary><b>Past releases</b></summary>

deeptutor/agents/math_animator/prompts/en/code_generator_agent.yaml

@@ -3,7 +3,20 @@ generate_system: |
   Produce runnable Python Manim code.
   Rules:
   - For video mode, return one complete Manim script with at least one renderable Scene subclass.
-  - For image mode, return YON_IMAGE anchor blocks only, and each block must contain a standalone renderable Manim script.
+  - For image mode, return YON_IMAGE anchor blocks only, and each block must contain a standalone renderable Manim script. Format:
+    ### YON_IMAGE_1_START ###
+    from manim import *
+    class Scene1(Scene):
+        def construct(self):
+            ...
+    ### YON_IMAGE_1_END ###
+    ### YON_IMAGE_2_START ###
+    from manim import *
+    class Scene2(Scene):
+        def construct(self):
+            ...
+    ### YON_IMAGE_2_END ###
+    Anchor numbering starts at 1. The code field must contain nothing outside YON_IMAGE anchor blocks.
   - Do not include commentary outside the JSON response.
   - Every geometric point, path point, and vertex passed to Manim must be 3D; do not use `[x, y]` when `[x, y, 0]` is required.
   - When positions come from axes or planes, prefer helpers such as `axes.c2p(...)` and `plane.c2p(...)` so the result is a valid 3D point.
@@ -59,6 +72,7 @@ retry_system: |
   If the error mentions `self.camera.frame`, repair by either:
   1) changing the scene base class to `MovingCameraScene`, or
   2) removing camera-frame animation while keeping the teaching flow.
+  If the error mentions `YON_IMAGE`, the image-mode code is missing anchor markers. Wrap the code in `### YON_IMAGE_n_START ###` / `### YON_IMAGE_n_END ###` blocks and ensure the code field contains nothing outside those anchor blocks.
 
 retry_user_template: |
   User request:

deeptutor/agents/math_animator/prompts/zh/code_generator_agent.yaml

@@ -3,7 +3,20 @@ generate_system: |
   你要生成可运行的 Python Manim 代码。
   要求：
   - video 模式：返回一个完整的 Manim 脚本，至少包含一个可渲染的 Scene 子类。
-  - image 模式：必须严格输出若干个 YON_IMAGE 锚点代码块，每个代码块内都必须是独立可渲染的 Manim 脚本。
+  - image 模式：必须严格输出若干个 YON_IMAGE 锚点代码块，每个代码块内都必须是独立可渲染的 Manim 脚本。格式如下：
+    ### YON_IMAGE_1_START ###
+    from manim import *
+    class Scene1(Scene):
+        def construct(self):
+            ...
+    ### YON_IMAGE_1_END ###
+    ### YON_IMAGE_2_START ###
+    from manim import *
+    class Scene2(Scene):
+        def construct(self):
+            ...
+    ### YON_IMAGE_2_END ###
+    锚点编号从 1 开始递增。code 字段中除了 YON_IMAGE 锚点块外不能有其他代码。
   - 不要输出解释性文字。
   - Manim 中所有几何点、路径点、顶点坐标都必须是 3D 形式；不要传 `[x, y]`，要传 `[x, y, 0]`。
   - 如果坐标来自坐标轴或数轴映射，优先使用 `axes.c2p(...)`、`plane.c2p(...)` 等返回 3D 点的方法。
@@ -60,6 +73,7 @@ retry_system: |
   若报错与 `self.camera.frame` 相关，必须二选一修复：
   1) 把场景类改为 `MovingCameraScene`；或
   2) 删除相机动画语句并保留原教学逻辑。
+  若报错包含 `YON_IMAGE`，说明 image 模式下代码缺少锚点标记。必须把代码重新包裹在 `### YON_IMAGE_n_START ###` / `### YON_IMAGE_n_END ###` 中，且 code 字段中除锚点块外不能有其他内容。
 
 retry_user_template: |
   用户需求：

deeptutor/agents/visualize/__init__.py

@@ -0,0 +1,5 @@
+"""Visualize agents and pipeline."""
+
+from .pipeline import VisualizePipeline
+
+__all__ = ["VisualizePipeline"]

deeptutor/agents/visualize/agents/__init__.py

@@ -0,0 +1,11 @@
+"""Agent building blocks for the visualize capability."""
+
+from .analysis_agent import AnalysisAgent
+from .code_generator_agent import CodeGeneratorAgent
+from .review_agent import ReviewAgent
+
+__all__ = [
+    "AnalysisAgent",
+    "CodeGeneratorAgent",
+    "ReviewAgent",
+]

deeptutor/agents/visualize/agents/analysis_agent.py

@@ -0,0 +1,74 @@
+"""Analysis stage: decide SVG vs Chart.js and produce a structured brief."""
+
+from __future__ import annotations
+
+from deeptutor.agents.base_agent import BaseAgent
+from deeptutor.core.trace import build_trace_metadata, new_call_id
+
+from ..models import VisualizationAnalysis
+from ..utils import extract_json_object
+
+
+class AnalysisAgent(BaseAgent):
+    def __init__(
+        self,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        api_version: str | None = None,
+        language: str = "zh",
+    ) -> None:
+        super().__init__(
+            module_name="visualize",
+            agent_name="analysis_agent",
+            api_key=api_key,
+            base_url=base_url,
+            api_version=api_version,
+            language=language,
+        )
+
+    async def process(
+        self,
+        *,
+        user_input: str,
+        history_context: str,
+        render_mode: str = "auto",
+    ) -> VisualizationAnalysis:
+        if render_mode in ("svg", "chartjs"):
+            system_prompt = self.get_prompt("system_fixed")
+            user_template = self.get_prompt("user_template_fixed")
+        else:
+            system_prompt = self.get_prompt("system")
+            user_template = self.get_prompt("user_template")
+        if not system_prompt or not user_template:
+            raise ValueError("AnalysisAgent prompts are not configured.")
+
+        format_kwargs: dict[str, str] = {
+            "user_input": user_input.strip(),
+            "history_context": history_context.strip() or "(none)",
+        }
+        if render_mode in ("svg", "chartjs"):
+            format_kwargs["render_type"] = render_mode
+
+        user_prompt = user_template.format(**format_kwargs)
+
+        chunks: list[str] = []
+        async for chunk in self.stream_llm(
+            user_prompt=user_prompt,
+            system_prompt=system_prompt,
+            response_format={"type": "json_object"},
+            stage="analyzing",
+            trace_meta=build_trace_metadata(
+                call_id=new_call_id("viz-analysis"),
+                phase="analyzing",
+                label="Visualization analysis",
+                call_kind="viz_analysis",
+                trace_role="analyze",
+                trace_kind="llm_output",
+            ),
+        ):
+            chunks.append(chunk)
+        response = "".join(chunks)
+        result = VisualizationAnalysis.model_validate(extract_json_object(response))
+        if render_mode in ("svg", "chartjs"):
+            result.render_type = render_mode  # type: ignore[assignment]
+        return result

deeptutor/agents/visualize/agents/code_generator_agent.py

@@ -0,0 +1,68 @@
+"""Code generation stage: produce SVG or Chart.js code from the analysis."""
+
+from __future__ import annotations
+
+import json
+
+from deeptutor.agents.base_agent import BaseAgent
+from deeptutor.core.trace import build_trace_metadata, new_call_id
+
+from ..models import VisualizationAnalysis
+from ..utils import extract_code_block
+
+
+class CodeGeneratorAgent(BaseAgent):
+    def __init__(
+        self,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        api_version: str | None = None,
+        language: str = "zh",
+    ) -> None:
+        super().__init__(
+            module_name="visualize",
+            agent_name="code_generator_agent",
+            api_key=api_key,
+            base_url=base_url,
+            api_version=api_version,
+            language=language,
+        )
+
+    async def process(
+        self,
+        *,
+        user_input: str,
+        history_context: str,
+        analysis: VisualizationAnalysis,
+    ) -> str:
+        system_prompt = self.get_prompt("system")
+        user_template = self.get_prompt("user_template")
+        if not system_prompt or not user_template:
+            raise ValueError("CodeGeneratorAgent prompts are not configured.")
+
+        user_prompt = user_template.format(
+            user_input=user_input.strip(),
+            history_context=history_context.strip() or "(none)",
+            render_type=analysis.render_type,
+            analysis_json=json.dumps(analysis.model_dump(), ensure_ascii=False, indent=2),
+        )
+
+        chunks: list[str] = []
+        async for chunk in self.stream_llm(
+            user_prompt=user_prompt,
+            system_prompt=system_prompt,
+            stage="generating",
+            trace_meta=build_trace_metadata(
+                call_id=new_call_id("viz-codegen"),
+                phase="generating",
+                label="Code generation",
+                call_kind="viz_code_generation",
+                trace_role="generate",
+                trace_kind="llm_output",
+            ),
+        ):
+            chunks.append(chunk)
+        response = "".join(chunks)
+
+        lang_hint = "svg" if analysis.render_type == "svg" else "javascript"
+        return extract_code_block(response, lang_hint) or extract_code_block(response)

deeptutor/agents/visualize/agents/review_agent.py

@@ -0,0 +1,67 @@
+"""Review stage: check and optionally optimise the generated code."""
+
+from __future__ import annotations
+
+import json
+
+from deeptutor.agents.base_agent import BaseAgent
+from deeptutor.core.trace import build_trace_metadata, new_call_id
+
+from ..models import ReviewResult, VisualizationAnalysis
+from ..utils import extract_json_object
+
+
+class ReviewAgent(BaseAgent):
+    def __init__(
+        self,
+        api_key: str | None = None,
+        base_url: str | None = None,
+        api_version: str | None = None,
+        language: str = "zh",
+    ) -> None:
+        super().__init__(
+            module_name="visualize",
+            agent_name="review_agent",
+            api_key=api_key,
+            base_url=base_url,
+            api_version=api_version,
+            language=language,
+        )
+
+    async def process(
+        self,
+        *,
+        user_input: str,
+        analysis: VisualizationAnalysis,
+        code: str,
+    ) -> ReviewResult:
+        system_prompt = self.get_prompt("system")
+        user_template = self.get_prompt("user_template")
+        if not system_prompt or not user_template:
+            raise ValueError("ReviewAgent prompts are not configured.")
+
+        user_prompt = user_template.format(
+            user_input=user_input.strip(),
+            render_type=analysis.render_type,
+            analysis_json=json.dumps(analysis.model_dump(), ensure_ascii=False, indent=2),
+            code=code,
+        )
+
+        chunks: list[str] = []
+        async for chunk in self.stream_llm(
+            user_prompt=user_prompt,
+            system_prompt=system_prompt,
+            response_format={"type": "json_object"},
+            stage="reviewing",
+            trace_meta=build_trace_metadata(
+                call_id=new_call_id("viz-review"),
+                phase="reviewing",
+                label="Code review",
+                call_kind="viz_code_review",
+                trace_role="review",
+                trace_kind="llm_output",
+            ),
+        ):
+            chunks.append(chunk)
+        response = "".join(chunks)
+        return ReviewResult.model_validate(extract_json_object(response))

deeptutor/agents/visualize/models.py

@@ -0,0 +1,51 @@
+"""Data models for the visualize pipeline."""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+
+class VisualizationAnalysis(BaseModel):
+    """Output of the analysis stage."""
+
+    render_type: Literal["svg", "chartjs"] = Field(
+        description="Whether to render as raw SVG or as a Chart.js configuration.",
+    )
+    description: str = Field(
+        default="",
+        description="High-level description of what the visualization should show.",
+    )
+    data_description: str = Field(
+        default="",
+        description="Description of the data or elements to be visualized.",
+    )
+    chart_type: str = Field(
+        default="",
+        description="Chart.js chart type (bar, line, pie, doughnut, radar, etc.) when render_type is chartjs.",
+    )
+    visual_elements: list[str] = Field(
+        default_factory=list,
+        description="Key visual elements to include (shapes, labels, axes, colors, etc.).",
+    )
+    rationale: str = Field(
+        default="",
+        description="Why this render_type was chosen over the alternative.",
+    )
+
+
+class ReviewResult(BaseModel):
+    """Output of the review / optimization stage."""
+
+    optimized_code: str = Field(
+        description="The final (potentially optimized) visualization code.",
+    )
+    changed: bool = Field(
+        default=False,
+        description="Whether the reviewer made modifications.",
+    )
+    review_notes: str = Field(
+        default="",
+        description="Notes on what was checked or changed.",
+    )