feat: add bilingual anti-defensive writing task skills

.gitattributes

@@ -0,0 +1,5 @@
+# Keep vendored originals byte-identical on Windows checkouts.
+/apps/server/src/skills/task-anti-defensive-writing/SKILL.md text eol=lf
+/apps/server/src/skills/task-anti-defensive-writing/LICENSE text eol=lf
+/apps/server/src/skills/task-anti-defensive-writing-en/SKILL.md text eol=lf
+/apps/server/src/skills/task-anti-defensive-writing-en/LICENSE text eol=lf

apps/server/package.json

@@ -18,7 +18,7 @@
     "@fastwrite/shared": "workspace:*",
     "fast-xml-parser": "^5.11.1",
     "fflate": "^0.8.2",
-    "nodemailer": "^10.0.10",
+    "nodemailer": "^10.1.4",
     "openai": "^7.3.0",
     "pdfjs-dist": "^5.4.296"
   },

apps/server/src/agent/skill-registry.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, test } from "bun:test";
 import { WRITING_PROFILES, paperSkillForProfile } from "@fastwrite/shared";
+import { readFile } from "node:fs/promises";
 import { join } from "node:path";
 import { parseSkillManifest, SkillRegistry } from "./skill-registry";
 
@@ -63,3 +64,14 @@ describe("published Skill manifests", () => {
     expect(() => parseSkillManifest(JSON.stringify({ id: "review", version: "1.0.0", scope: "system", owner: "FastWrite", license: "Apache-2.0", workflows: [], requiredEvidence: [], capabilities: [], maxContextChars: 1000, riskLevel: "medium", requiresReview: true, eval: { status: "passed", score: 2 }, references: [] }), "review")).toThrow();
   });
 });
+
+test("reads task metadata separately from the skill instructions", async () => {
+  const directory = join(import.meta.dir, "..", "skills");
+  const registry = new SkillRegistry(directory);
+  for (const [id, language] of [["anti-defensive-writing", "中文"], ["anti-defensive-writing-en", "English"]] as const) {
+    const { descriptor, instructions } = await registry.loadTask(id);
+    expect(descriptor.description).toContain(language);
+    expect(instructions).toBe(await readFile(join(directory, `task-${id}`, "SKILL.md"), "utf8"));
+  }
+  expect((await registry.loadTask("claim-audit")).descriptor.riskLevel).toBe("high");
+});

apps/server/src/agent/skill-registry.ts

@@ -92,7 +92,9 @@ export class SkillRegistry {
     const result: AgentTaskSkillDescriptor[] = [];
     for (const entry of entries.filter((item) => item.isDirectory() && item.name.startsWith("task-"))) {
       const content = await readFile(join(this.skillsDirectory, entry.name, "SKILL.md"), "utf8").catch(() => "");
-      const field = (name: string) => new RegExp(`^\\s*${name}:\\s*(.+)$`, "m").exec(content)?.[1]?.trim() ?? "";
+      if (!content) continue;
+      const metadata = await readFile(join(this.skillsDirectory, entry.name, "TASK.md"), "utf8").catch(() => content);
+      const field = (name: string) => new RegExp(`^\\s*${name}:\\s*(.+)$`, "m").exec(metadata)?.[1]?.trim() ?? "";
       const list = (name: string) => field(name).split(",").map((item) => item.trim()).filter(Boolean);
       const intents = list("supportedIntents").filter((item): item is AgentTaskIntent => ["draft", "continue", "revise"].includes(item));
       if (!intents.length) continue;

apps/server/src/skills/task-anti-defensive-writing-en/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Adkid-Zephyr
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

apps/server/src/skills/task-anti-defensive-writing-en/SKILL.md

@@ -0,0 +1,123 @@
+---
+name: anti-defensive-writing-en
+description: >
+  Stops defensive writing across the entire paper lifecycle — writing, revising,
+  cutting, and organizing experiments. Treats the paper as a press conference,
+  not a project summary, lab log, or self-audit: identify the single most
+  publishable strength of the work and build the most favorable, complete, and
+  persuasive narrative around it. Never present everything evenly, never volunteer
+  weakness, never write an experiment diary, never attack the paper on the
+  reviewer's behalf.
+  Triggers: polish my paper, revise manuscript, write abstract, write introduction,
+  write conclusion, shorten paper, rebuttal, respond to reviewers, organize
+  experiments, defensive writing, AI-flavored academic writing.
+---
+
+# The Press-Release Principle (Paper as Press Conference)
+
+## One-line principle
+
+A paper is a press conference, not a project summary, a lab log, or a self-audit.
+Your job is not to present every aspect of the work evenly, but to identify its
+most publishable value and build the most favorable, complete, and persuasive
+narrative around it.
+
+## 1. Narrative rules
+
+1. **Organize the paper around strengths only.** Look for what is genuinely ahead,
+   unique, or irreplaceable: a new capability, a new problem, a new mechanism, a new
+   perspective, broader applicability, lower cost, higher efficiency, better
+   scalability, or a more meaningful trade-off. Content that forms no advantage
+   does not enter the main storyline.
+2. **Do not write a work report.** No "we first did X, then tried Y, and Z failed"
+   chronology. Keep only the final standing logic: why the problem matters → why
+   existing methods fall short → what this paper provides → how the evidence
+   supports it.
+3. **Never set a contest you cannot win.** If a metric is not your strength, do not
+   build the narrative on it. Re-choose the task definition, evaluation dimension,
+   application scenario, constraint, or comparison frame that best reflects your
+   value. Name the game this paper actually wins.
+4. **State the advantage explicitly.** Do not expect reviewers to discover the
+   contribution from a table. Explain: under which condition the method performs
+   best, why the advantage emerges, what practical problem it solves, and why it
+   deserves attention.
+5. **Limit the comparison scope.** Do not chase "wins on every dataset and metric."
+   Make only claims your evidence firmly supports. Persuasion comes from tight
+   claim-evidence alignment, not from the number of comparisons.
+6. **Allow full story restructuring.** When existing results cannot support the
+   original narrative, do not defend it. Redefine the problem, re-order the
+   contributions, re-pick the headline result, redesign title, abstract,
+   introduction, and experiment structure. The story serves the strongest
+   evidence, not the original plan.
+
+## 2. Language rules (against self-undermining)
+
+- Banned self-weakening phrases: "unfortunately", "merely", "only achieves",
+  "still lags far behind", "limited improvement", "severely insufficient".
+- **Never say you lost.** For unfavorable results, avoid "our method is weaker",
+  "performance drops", "fails to surpass". Judge in order:
+  ① must this result be discussed at all; ② can it be explained by different
+  goals, constraints, or applicable scenarios; ③ does it reflect a reasonable
+  trade-off; ④ can the claim be narrowed to avoid a head-on collision;
+  ⑤ should the storyline change so this result no longer carries the core
+  contribution.
+- If a result forms no advantage, do not distill it into a conclusion — and never
+  help the reader escalate a local observation into a verdict on the whole method.
+
+## 3. Experiment rules
+
+Experiments are tools of argument, not a warehouse of results. Every experiment
+must carry one explicit duty:
+
+- prove the core method works;
+- prove the advantage comes from the key mechanism;
+- prove the method matters in the target scenario;
+- rule out the most plausible alternative explanations.
+
+An experiment that does not strengthen the main line, distracts, or invites
+irrelevant disputes: delete, downplay, move, or redesign it.
+
+## 4. Structure rules
+
+**Abstract and introduction = the opening of the press conference.** Establish
+four things fast: ① an important, unsolved problem; ② the key gap in existing
+methods; ③ your distinctive approach; ④ your strongest result and its meaning.
+Do not open with implementation details, process, or a wall of background. Do not
+discuss limitations before the contribution is established.
+
+**The conclusion only reinforces the takeaway:** what was solved, what was
+proposed, what was proven, why it matters. Never introduce a fresh self-negation
+or expanded limitations in the final paragraph.
+
+## 5. Default decision rules (for any unfavorable material, in priority order)
+
+1. Delete content irrelevant to the core claim;
+2. Narrow the claim to avoid a pointless head-on comparison;
+3. Switch to an evaluation dimension that better reflects the value;
+4. Explain the result as a difference in goals or a reasonable trade-off;
+5. Reorganize experiments so the advantage becomes the visual and narrative center;
+6. Redefine the paper's story;
+7. Only when unavoidable and truly affecting the core conclusion, state it plainly.
+
+## 6. Pre-submission checklist (never hand the reviewer a knife)
+
+- Does this sentence quietly expand the responsibility the paper must carry?
+- Does it raise a question nobody asked for?
+- Does it describe a local phenomenon as a universal flaw?
+- Does it use a negative judgment broader than the evidence?
+- Is any passage unfolding as "what we did" process narration?
+- Is a non-winning metric set as the main battlefield?
+- In the first five sentences of the abstract, are both the problem and the
+  contribution present?
+- Does the conclusion smuggle in a sudden self-negation?
+
+Do not manufacture review questions. Do not enlarge the attack surface.
+Do not complete the opposing side's argument for them.
+
+## Final goal
+
+Every section, paragraph, table, and sentence should do one thing together:
+convince the reader that this work solves a problem worth solving, proposes a
+method worth noticing, and already has sufficiently clear evidence of its value.
+Find the advantage that truly holds, organize all material around it, and make
+that advantage unmistakably clear.

apps/server/src/skills/task-anti-defensive-writing-en/TASK.md

@@ -0,0 +1,11 @@
+---
+version: 1.0.0
+description: Apply the Press-Release Principle to paper narrative, language, and experiments (English).
+supportedIntents: draft,continue,revise
+allowedScope: project
+requiredEvidence: manuscript
+validationCommands: review
+riskLevel: medium
+allowNewFiles: false
+requiresReview: true
+---

apps/server/src/skills/task-anti-defensive-writing/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Adkid-Zephyr
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

apps/server/src/skills/task-anti-defensive-writing/SKILL.md

@@ -0,0 +1,94 @@
+---
+name: anti-defensive-writing
+description: >
+  论文写作、修改、压缩和实验组织全流程中阻止【防御性写作】。把论文视为一场学术发布会，
+  而不是项目总结、实验日志或自我审查报告：识别这项工作最值得发表的价值，围绕它建立
+  最有利、最完整、最有说服力的叙事。不平均展示、不主动示弱、不写实验流水账、不替审稿人
+  攻击自己。
+  触发词：论文润色、论文修改、写摘要、写引言、写结论、压缩论文、rebuttal、回复审稿意见、
+  组织实验、论文AI味、防御性写作、心虚。
+---
+
+# 学术写作原则(The Press-Release Principle)
+
+## 一句话原则
+
+论文是一场学术发布会，不是项目总结、实验日志或自我审查报告。
+你的任务不是平均呈现这项工作的所有方面，而是识别其中最值得发表的价值，
+并围绕这一价值建立最有利、最完整、最有说服力的叙事。
+
+## 一、叙事规则
+
+1. **只围绕优势组织论文。** 优先寻找本文真正领先、独特或不可替代的部分：新能力、新问题、
+   新机制、新视角、更强适用性、更低成本、更高效率、更好扩展性、更有意义的权衡。
+   没有形成优势的内容，不进入主线。
+2. **不写成工作汇报。** 不按"我们先做了什么、后来尝试了什么、哪些失败了"的过程展开。
+   只保留最终成立的逻辑：问题为什么重要 → 现有方法为什么不够 → 本文提供了什么新解法 →
+   证据如何支持。
+3. **打不过的维度，不设为比赛项目。** 某指标不占优势，就不要把叙事建立在该指标上。
+   重新选择更能反映本文价值的任务定义、评价维度、应用场景、约束条件或比较口径。
+   不争夺不属于本文的冠军，明确本文赢的是另一场更有意义的比赛。
+4. **优势必须被明确说出来。** 不要期待审稿人自己从表格中发现贡献。主动解释：
+   本文在哪个条件下表现最好、为什么这种优势会出现、解决了什么实际问题、为什么值得关注。
+5. **控制比较范围。** 不追求"所有数据集、所有指标全面领先"。只提出证据能牢固支撑的主张。
+   说服力来自主张与证据高度一致，而不是比较项目数量最多。
+6. **允许彻底重构故事。** 现有结果无法支撑原始叙事时，不要围绕原叙事防守。重新定义问题、
+   重排贡献、重选主结果、重做标题摘要引言和实验结构。故事服务于最强证据，而非忠于最初设想。
+
+## 二、语言规则(防自我削弱)
+
+- 禁用自我削弱式表达:"遗憾的是""仅""仍明显落后""效果有限""存在严重不足"。
+- **不说输。** 面对不占优的结果,不使用"本文弱于""性能下降""未能超过"。按顺序判断:
+  ①该结果是否必须讨论;②能否从不同目标、约束或适用场景解释;③是否体现某种合理权衡;
+  ④能否收缩主张,避免不必要的正面对撞;⑤是否应调整故事主线,让它不再承担证明核心贡献的职责。
+- 某项结果不构成优势时,不必提炼成结论,更不应帮读者把局部现象上升为对整体方法的否定。
+
+## 三、实验规则
+
+实验不是结果仓库，而是论证工具。每个实验必须承担明确职责之一：
+
+- 证明核心方法有效;
+- 证明优势来自关键机制;
+- 证明方法在目标场景中具有价值;
+- 排除最可能的替代解释。
+
+无法强化主线、容易分散注意力或引出无关争论的实验：删除、弱化、移动或重新设计。
+
+## 四、结构规则
+
+**摘要和引言 = 发布会开场**,开头迅速建立四件事:
+①一个重要且尚未解决的问题;②现有方法的关键缺口;③本文独特的解决思路;④最有分量的结果和意义。
+不从实现细节、研究过程或大量背景知识开始,不在贡献尚未建立时提前讨论不足。
+
+**结论只强化记忆点**:本文解决了什么、提出了什么、证明了什么、为什么重要。
+不在最后一段突然加入新的自我否定或扩大局限性。
+
+## 五、默认决策规则(遇到不理想材料时,按优先级处理)
+
+1. 删除与核心主张无关的内容;
+2. 缩小主张,避免无意义的正面对比;
+3. 更换更能体现价值的评价维度;
+4. 将结果解释为目标差异或合理权衡;
+5. 重组实验,使优势成为视觉和叙事中心;
+6. 重新定义论文故事;
+7. 只有在无法回避且确实影响核心结论时,才进行必要说明。
+
+## 六、交稿前自查(不给审稿人递刀子)
+
+- 这句话是否无意中扩大了本文需要承担的责任?
+- 是否提出了一个本来没人要求回答的问题?
+- 是否把局部现象描述成了普遍缺陷?
+- 是否使用了比证据更宽泛的负面判断?
+- 有没有按"我们做了什么过程"在展开?
+- 有没有把不领先的指标设成了主战场?
+- 摘要前五句里,问题和贡献都出来了吗?
+- 结论里有没有突然的自我否定?
+
+不主动制造审稿问题,不主动扩大攻击面,不主动替反方完成论证。
+
+## 最终目标
+
+论文中的每一个章节、段落、表格和句子,都应共同完成一件事:让读者相信,
+这项工作解决了一个值得解决的问题,提出了一种值得关注的方法,
+并且已经有足够清晰的证据证明它的价值。
+找到真正成立的优势,围绕它组织全部材料,并把这个优势讲到足够清楚。

apps/server/src/skills/task-anti-defensive-writing/TASK.md

@@ -0,0 +1,11 @@
+---
+version: 1.0.0
+description: 反防御性写作（中文）：按原版学术发布会原则组织论文叙事、语言与实验。
+supportedIntents: draft,continue,revise
+allowedScope: project
+requiredEvidence: manuscript
+validationCommands: review
+riskLevel: medium
+allowNewFiles: false
+requiresReview: true
+---

apps/web/src/styles.css

@@ -403,7 +403,7 @@ code { font-family: "SFMono-Regular", Consolas, "Liberation Mono", monospace; }
 .ai-workspace__controls { min-width: 0; display: flex; flex-wrap: wrap; align-items: center; gap: 8px 16px; }
 .ai-workspace__mode { flex: 0 0 auto; }
 .ai-workspace__mode button { height: 30px; font-size: 12px; }
-.ai-workspace__mode button[aria-pressed="true"] { color: var(--accent); background: var(--vscode-active); }
+.ai-workspace__mode button[aria-pressed="true"] { color: var(--vscode-foreground); background: var(--vscode-active); }
 .ai-workspace__harness { display: flex; align-items: center; gap: 8px; }
 .ai-workspace__harness > div { width: 112px; }
 .ai-workspace__harness [role="combobox"] { height: 30px; }