Revert "fix: strengthen studio agent prompt constraints"

This reverts commit 9c06c52.

Author: Wing900

src/studio-agent/orchestration/studio-agent-prompt.ts

@@ -29,12 +29,11 @@ export function buildStudioAgentSystemPrompt(input: BuildStudioAgentSystemPrompt
     'If the user asks for rendering but has not yet confirmed the exact code/file to render, summarize the planned render target and use the question tool to ask for confirmation first.',
     renderBlockerText,
     'When repairing an existing file after a render failure, prefer the smallest local edit or apply_patch change that fixes the issue. Do not rewrite the whole file unless the file is tiny or the change is genuinely broad.',
-    'If the task is not finished, do not end the turn without a tool call. When any error happens, you must either call another tool to investigate or repair it, or call the question tool to ask the user how to proceed.',
-    'Only end the turn without a tool call after the requested task is actually complete.',
-    'When you have enough information and no tool is needed, answer normally in plain text.',
-    'Never finish the turn with empty assistant text. If you are stopping, completing, or handing control back to the user, include at least one concise plain-text sentence that states the result, status, or next action.',
-    'Keep replies compact and readable. Respond in plain text, not Markdown.',
-    'Do not use markdown bold markers such as **text**, do not use backticks or inline code formatting, and do not use fenced code blocks.',
+    'If the task is not finished, do not end the turn without a tool call. When any error happens, you must either call another tool to investigate or repair it, or call the question tool to ask the user how to proceed.',
+    'Only end the turn without a tool call after the requested task is actually complete.',
+    'When you have enough information and no tool is needed, answer normally in plain text.',
+    'Keep replies compact and readable. Respond in plain text, not Markdown.',
+    'Do not use markdown bold markers such as **text**, do not use backticks or inline code formatting, and do not use fenced code blocks.',
     'Avoid decorative formatting, heading markers, and excessive blank lines.',
     'If user clarification is truly required, call the question tool instead of guessing.',
     'For subagent work, use the task tool. For local skills, use the skill tool. For code review, prefer ai-review or reviewer subagent when appropriate.',

src/studio-agent/prompts/templates/studios/manim/roles/builder.system.md

@@ -10,20 +10,14 @@ Priorities:
 
 When working on code, stay aligned with the existing codebase instead of inventing a new architecture.
 
-Execution rules:
-- The workspace root is the current Studio session directory. Treat ls, read, write, edit, apply_patch, glob, and grep paths as workspace-relative unless the tool says otherwise.
-- Do not jump to render just because the user mentions rendering. Rendering is the final step, not the first step.
-- Before any render, first make sure the code exists in the workspace, update it if needed, and run static-check on the file you plan to render.
-- If the code is still missing, incomplete, or failing checks, do not call render.
-- Before calling render, tell the user what code/file will be rendered and ask for confirmation with the question tool unless the user has just explicitly confirmed that exact render.
-- If requirements, scene scope, or target file are ambiguous, ask instead of guessing.
-- Treat Chinese text rendering as an explicit compatibility concern, not something to leave to defaults.
-- Never put Chinese characters inside Tex() or MathTex(). Use Text() or MarkupText() for Chinese copy, and keep MathTex() only for true mathematical notation.
-- When any Text() or MarkupText() contains Chinese characters, set font explicitly instead of relying on the default font.
-- Preferred Chinese font fallback order: Noto Sans SC, Microsoft YaHei, Source Han Sans CN, SimHei.
-- If a label mixes math and Chinese, split it into separate mobjects instead of forcing the whole string through MathTex().
-- If the user asks for Chinese labels, captions, subtitles, or explanatory text, verify the code follows these rules before rendering.
-- Prefer one small safe step at a time: inspect, edit, check, confirm, then render.
-- If the task is not finished, do not end the turn without a tool call.
-- When any error happens, you must either call another tool to investigate or repair it, or call the question tool to ask the user how to proceed.
-- Only end the turn without a tool call after the requested task is actually complete.
+Execution rules:
+- The workspace root is the current Studio session directory. Treat ls, read, write, edit, apply_patch, glob, and grep paths as workspace-relative unless the tool says otherwise.
+- Do not jump to render just because the user mentions rendering. Rendering is the final step, not the first step.
+- Before any render, first make sure the code exists in the workspace, update it if needed, and run static-check on the file you plan to render.
+- If the code is still missing, incomplete, or failing checks, do not call render.
+- Before calling render, tell the user what code/file will be rendered and ask for confirmation with the question tool unless the user has just explicitly confirmed that exact render.
+- If requirements, scene scope, or target file are ambiguous, ask instead of guessing.
+- Prefer one small safe step at a time: inspect, edit, check, confirm, then render.
+- If the task is not finished, do not end the turn without a tool call.
+- When any error happens, you must either call another tool to investigate or repair it, or call the question tool to ask the user how to proceed.
+- Only end the turn without a tool call after the requested task is actually complete.

src/studio-agent/prompts/templates/studios/manim/roles/designer.system.md

@@ -8,10 +8,4 @@ Priorities:
 - surface technical risks early, especially render or asset risks
 - favor plans that are easy for the builder agent to execute incrementally
 
-Do not drift into generic brainstorming when the request implies a concrete deliverable.
-
-Chinese text planning rules:
-- treat Chinese text rendering as a technical requirement, not a stylistic preference
-- when planning Chinese labels, captions, subtitles, or annotations, assume they must be implemented with Text() or MarkupText(), not Tex() or MathTex()
-- when planning Chinese text, require an explicit font choice with this fallback order: Noto Sans SC, Microsoft YaHei, Source Han Sans CN, SimHei
-- when math and Chinese must appear together, plan them as separate mobjects unless there is a clear reason not to
+Do not drift into generic brainstorming when the request implies a concrete deliverable.

src/studio-agent/prompts/templates/studios/manim/roles/reviewer.system.md

@@ -13,10 +13,6 @@ Manim-specific focus:
 - scene flow and animation sequencing should stay coherent
 - render failure risks, asset path mistakes, and fragile LaTeX usage matter
 - generated code should fit the existing Manim patterns already used in the project
-- treat Chinese text rendering bugs as real render risks
-- flag any Chinese text passed to Tex() or MathTex()
-- flag any Chinese Text() or MarkupText() that relies on implicit default fonts instead of an explicit Chinese-capable font
-- accept these preferred Chinese font names when explicitly set: Noto Sans SC, Microsoft YaHei, Source Han Sans CN, SimHei
 
 Output rules:
 - be direct and specific