auto-paper-improvement-loop · diff

git:20260923.2013b81 to git:20260927.efd5131

16 added, 207 removed. Audit A to A.

---

name: auto-paper-improvement-loop

description: "论文自动改进循环（2 轮）：先由外部审稿模型评审，再按要求改稿，最后重编译验证。触发词：改论文、improve paper、论文润色循环、auto improve、迭代打磨论文。"

argument-hint: [paper-directory]

allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, Agent

---

- # Auto Paper Improvement Loop: Review → Fix → Recompile

- Autonomously improve the paper at: **$ARGUMENTS**

- ## Context

- Runs after `/paper-write` + `/paper-compile`. Iterates on writing quality (not research).

- ## Constants

- - **MAX_ROUNDS = 2** — Round 1 catches structural issues, Round 2 catches presentation issues.

- - **REVIEWER_SCRIPT** — External reviewer script. Thread via `_reviewer_thread.json`.

- - **HUMAN_CHECKPOINT = false** — When true, pause after each review for user input.

- - **LATEX_ENGINE** — Auto-detect: ctex/xelatex in main.tex → xelatex, otherwise pdflatex.

- - **CUSTOM_REQUIREMENTS** — Highest priority.

- ## State Persistence

- Writes `PAPER_IMPROVEMENT_STATE.json` after each round for crash recovery:

- ```json

- {"current_round": 1, "last_score": 6, "status": "in_progress", "timestamp": "..."}

- ```

- On startup: if exists + in_progress + <24h → resume. Otherwise start fresh.

- ## Workflow

- ### Step 0: Preserve Original

- ```bash

- mkdir -p paper/_improvement_rounds

- cp paper/main.pdf paper/_improvement_rounds/round0_original.pdf

- ```

- ### Step 1: Collect Paper Text

- Concatenate all `paper/sections/*.tex` for review.

- ### Step 2: Round 1 Review

- Send the full paper text to the external reviewer via `reviewer_client.py`:

- ```bash

- mkdir -p _tmp

- cat << 'REVIEW_EOF' > _tmp/_review_prompt.txt

- You are reviewing an academic paper. Please provide a detailed, structured review.

- ## Full Paper Text:

- REVIEW_EOF

- cat _tmp/_paper_full_text.tex >> _tmp/_review_prompt.txt

- cat << 'REVIEW_EOF' >> _tmp/_review_prompt.txt

- ## Review Instructions

- Please act as a senior reviewer. Provide:

- 1. **Overall Score** (1-10, where 6 = weak accept, 7 = accept)

- 2. **Summary** (2-3 sentences)

- 3. **Strengths** (3-5 bullet points)

- 4. **Weaknesses** (categorized as CRITICAL / MAJOR / MINOR, each with specific location and actionable fix)

- - Check for "figure-as-subject" AI pattern: paragraphs starting with "图X展示了"/"As shown in Figure X" — flag as MAJOR if ≥3 occurrences

- 5. **Verdict**: "Accept" / "Almost" / "Reject"

- 6. **Actionable Fixes** (ordered by priority, each with exact section and what to change)

- REVIEW_EOF

- PYTHON=""; for _c in "$MH_PYTHON" python python3; do [ -z "$_c" ] && continue; if $_c -c "import sys" >/dev/null 2>&1; then PYTHON="$_c"; break; fi; done; [ -z "$PYTHON" ] && PYTHON=python

- $PYTHON "$REVIEWER_SCRIPT" --prompt-file _tmp/_review_prompt.txt --thread-file _tmp/_reviewer_thread.json 2>&1 | tee _tmp/_round1_review.txt

- ```

- If the reviewer script fails (API key not configured or network error), perform the review using self-analysis — act as a senior reviewer, score honestly, and proceed to Step 3. Cross-model review is preferred for objectivity but not required.

- ### Step 2b: Human Checkpoint (if enabled)

- Present score + weaknesses. Wait for "go" / custom instructions / "skip N" / "stop".

- Non-interactive mode: auto-proceed.

- ### Step 3: Implement Round 1 Fixes

- Priority: CRITICAL → MAJOR → MINOR.

- Common fix patterns:

- | Issue | Fix |

- |-------|-----|

- | Assumption-model mismatch | Rewrite assumption, add bridging proposition |

- | Overclaims | Soften: "validate"→"demonstrate relevance" |

- | Missing metrics | Add quantitative table with caveats |

- | Notation confusion | Rename globally, add Notation paragraph |

- | Figure-as-subject writing | Rewrite: move figure ref to parenthetical, lead with analysis point |

- | Missing references | Add to bib, cite appropriately |

- ### Step 4: Recompile Round 1

- Auto-detect engine. Manual 4-step: engine→bibtex→engine→engine. Then snapshot to subdir:

- ```bash

- cp paper/main.pdf paper/_improvement_rounds/round1.pdf

- ```

- ### Step 5-6: Round 2 Review + Fixes

- Same process. Use same `_tmp/_reviewer_thread.json` for context continuity (multi-turn dialogue is automatic):

- ```bash

- cat << 'REVIEW_EOF' > _tmp/_review_prompt.txt

- Since your last review, we have implemented the following fixes:

- [list the fixes from Step 3]

- Please re-review the updated paper. Focus on:

- 1. Were the CRITICAL/MAJOR issues adequately addressed?

- 2. Any new issues introduced by the fixes?

- 3. Updated score and verdict.

- ## Updated Paper Text:

- REVIEW_EOF

- cat _tmp/_paper_full_text.tex >> _tmp/_review_prompt.txt

- PYTHON=""; for _c in "$MH_PYTHON" python python3; do [ -z "$_c" ] && continue; if $_c -c "import sys" >/dev/null 2>&1; then PYTHON="$_c"; break; fi; done; [ -z "$PYTHON" ] && PYTHON=python

- $PYTHON "$REVIEWER_SCRIPT" --prompt-file _tmp/_review_prompt.txt --thread-file _tmp/_reviewer_thread.json 2>&1 | tee _tmp/_round2_review.txt

- ```

- If reviewer unavailable, use self-analysis for Round 2 as well.

- ### Step 7: Recompile Round 2

- Snapshot to subdir:

- ```bash

- cp paper/main.pdf paper/_improvement_rounds/round2.pdf

- ```

- `paper/main.pdf` itself stays as the final deliverable (= round2).

- ### Step 8: Format Check

- Check overfull hbox (>10pt → fix), page count vs limit, underfull warnings.

- ### Step 9: Document Results

- Create `paper/PAPER_IMPROVEMENT_LOG.md`：

- ```markdown

- # Paper Improvement Log

- ## Score Progression

- |-------|-------|---------|-------------|

- ## Round 1 Review & Fixes

- ### Review (full text)

- [Complete raw review text, verbatim]

- ### Fixes Implemented

- 1. [Fix description + which files changed]

- 2. [Fix description]

+ # 论文改进：独立评审、修订、重新编译

- ## Round 2 Review & Fixes

+ ## 范围

- ### Review (full text)

- [Complete raw review text, verbatim]

+ 针对已有稿件的论证和表达做最多两轮聚焦改进，不重新启动整个研究工作流。若问题涉及研究设计、实验有效性或缺少核心证据，应回报问题及所需工作，不靠修辞掩盖。

- ### Fixes Implemented

- 1. [Fix description]

- 2. [Fix description]

+ ## 输入与版本

- ## PDFs

- | File | Size | Description |

- |------|------|-------------|

- | main.pdf | X KB | **最终成品（= round2），编辑/下载只认这个** |

- | _improvement_rounds/round0_original.pdf | X KB | Before improvement (历史留痕) |

- | _improvement_rounds/round1.pdf | X KB | After round 1 (历史留痕) |

- | _improvement_rounds/round2.pdf | X KB | After round 2 (历史留痕) |

- ```

+ 读取当前步骤合同、稿件源文件、参考文献、已编译PDF、真实结果和之前的评审意见。保留原稿与每轮PDF，续作从当前事实恢复，不因时间超过24小时而丢掉历史。

- ## Output

+ ## 工作过程

- ```

- paper/

- ├── main.pdf # 最终成品(= round2)，编辑/下载只认这个

- ├── PAPER_IMPROVEMENT_LOG.md

- └── _improvement_rounds/ # 历史快照(只读留痕，不参与后续编辑)

- ├── round0_original.pdf

- ├── round1.pdf

- └── round2.pdf

- ```

+ 1. 生成独立评审任务。通过执行会话的 `review-request` 绑定当前稿件及需要的证据文件，要求评审者指出位置、严重性、依据和最小修复；只读评审不得直接改稿。

+ 2. 当前宿主提供独立上下文执行评审。通过 `review-receive` 接收实际原文与真实调用标识，程序校验稿件版本。没有独立上下文时明示未完成独立评审，不预填分数、会话或通过结论。

+ 3. 按fatal/major/minor落实修订：修正假设与方法不一致、收敛过度论断、补足真正存在的评价依据、统一符号和引用。不能编造实验数据或替用户接受新的研究方向。

+ 4. 用执行会话 `edit` 或显式 `mutates` 修改源码，使用已配置的真实编译工具重新编译。工具返回码和清单自动记录；检查编译错误、未定义引用、页数、越界和实际图件。

+ 5. 必要时对修改后的版本再评审一次。评审原文、修改位置和剩余问题写入 `paper/PAPER_IMPROVEMENT_LOG.md`。分数仅记录实际评审结果，不为达到阈值调整结论。

+ 6. `finish` 自动重验当前合同要求的源文件、参考文献、PDF和日志，不能只交变更说明。质量问题留在当前步骤返修，人工批准点必须等待真实批准，无回复不自动继续。

- ## Key Rules

+ ## 输出

- ⛔ **File writing strategy (prevent both failure modes):**

- - For short content (<150 lines): use the **Write tool** directly (atomic, reliable)

- - For long content (>150 lines): use **Write** for the first section (ensures file exists on disk), then append remaining sections with `cat << 'EOF' >> paper/PAPER_IMPROVEMENT_LOG.md`

- - **NEVER `end_turn` without producing the improved `paper/main.pdf`** — even if some rounds had issues, keep the best compiled PDF

+ 以当前步骤合同为准；通常包括 `paper/main.pdf`、源文件及 `paper/PAPER_IMPROVEMENT_LOG.md`，并保留 `paper/_improvement_rounds/` 中的旧版本。只用已接受的当前版本作为最终交付，不改旧哈希冒充重跑。

- ⛔ **MUST run output verification before ending**:

- ```bash

- PASS=true

- # 主产出: 改进后的 paper/main.pdf (引擎按此判定成败)

- [ -f paper/main.pdf ] && SZ=$(wc -c < paper/main.pdf) || SZ=0

- if [ "$SZ" -ge 100000 ]; then

- echo "✅ paper/main.pdf ($SZ bytes)"

- else

- echo "❌ paper/main.pdf missing or too small ($SZ bytes) — must keep a compiled PDF before ending"

- PASS=false

- fi

- # 改进日志

- if [ -f paper/PAPER_IMPROVEMENT_LOG.md ]; then

- echo "✅ paper/PAPER_IMPROVEMENT_LOG.md"

- else

- echo "❌ paper/PAPER_IMPROVEMENT_LOG.md missing — write the improvement log before ending"

- PASS=false

- fi

- [ "$PASS" != true ] && echo "⛔ Verification failed — must produce output before ending step"

- ```

+ ## 质量责任

- - Preserve all PDF snapshots under `paper/_improvement_rounds/` — keep `paper/` root clean (only `main.pdf` there, so the editor / download never points at a stale round snapshot)

- - Save full raw review text (don't truncate)

- - Same thread file for Round 2 (context continuity)

- - Always recompile after fixes

- - Don't fabricate experimental results

- - Soften overclaims rather than adding unsupported claims

- - Global consistency when renaming notation (all files)

- - ⛔ Main output: `paper/main.pdf` + `paper/PAPER_IMPROVEMENT_LOG.md`. Don't write extra reports (e.g. AUTO_PAPER_IMPROVEMENT_REPORT.md, AUTO_PAPER_IMPROVEMENT_FINAL_REPORT.md) to root

- - ⛔ Temp files (`_review_prompt.txt`, `_paper_full_text.tex`, `_round1_review.txt` etc.) go to `_tmp/`

- - Large files: Bash heredoc

+ 模型负责理解批评、权衡修改和形成有效论证；程序负责字节版本、执行事实、结构完整性、编译与合同检查。负面结果、未解决的主要问题和评审缺席都必须明示。达到最大轮数只能停止迭代，不能自动获得投稿就绪结论。