Content
# Paper Generator
> One person + LLM = A complete research team.
[](https://python.org)
[](LICENSE)
Give it a research topic and experimental data, and it returns a LaTeX paper that meets journal standards. 13 Agents work together - planning, literature retrieval, writing, adversarial questioning, citation verification, typesetting, simulated review - fully automated, end-to-end.
Then you're unsure about the paper quality? 23 virtual research roles are ready: Chief Investigator helps you choose a direction, Devil's Advocate attacks your weaknesses, Ethics Officer ensures compliance, Literature Manager fixes every BibTeX.
**This is not a copilot. This is a team.**
**Who is it for:**
- **Independent Researchers** — One person completes the entire process of exploration, writing, verification, and submission
- **Graduate Students** — Structured paper generation + simulated peer review, rapid iteration
- **AI Assistant Users** — 31 MCP tools, Claude Desktop / VS Code Copilot direct call
---
## Quick Start
```bash
# Installation
git clone <repo-url> && cd new-laxtax
pip install -e .
# Explore direction
paper-generator explore "knowledge distillation for medical image segmentation" \
--provider qianwen --model qwen3.6-plus --api-key $LLM_API_KEY
# Generate paper
paper-generator generate "knowledge distillation for polyp segmentation" \
--provider qianwen --model qwen3.6-plus --api-key $LLM_API_KEY \
--data-dir data --venue springer --no-human \
--from-exploration output/exploration_report.json
# One-click full process: exploration → generation → polishing → verification
paper-generator chain "polyp segmentation with knowledge distillation" \
--provider qianwen --model qwen3.6-plus --api-key $LLM_API_KEY \
--data-dir data --venue springer
```
Three commands. From scratch to a postable paper.
---
## Watch it work
```
You: paper-generator explore "transformer pruning for edge deployment"
System: [5 Agents launched]
DiscoveryAgent — Parallelly crawls literature from Semantic Scholar + OpenAlex + Crossref + ArXiv + Google Scholar + DBLP six sources
ParserAgent — LLM batch extracts features
DirectionAgent — Clusters into 4 research directions
FeasibilityAgent — Evaluates feasibility, data availability, GPU budget for each direction
ReportAgent — Generates ranked report
→ output/exploration_report.json
You: paper-generator generate "structured pruning via knowledge distillation" \
--from-exploration output/exploration_report.json --venue neurips
System: [13 Agents pipeline]
DataAgent → PlanningAgent → KnowledgeAgent → NarrativeDirector →
LiteratureSynthesis → VoiceAlignment → WritingAgent →
DevilsAdvocate → VerificationAgent → VisualVerify → TypesettingAgent → ReviewAgent
[Automatically 3 rounds writing ↔ verification revision]
→ output/paper.tex + references.bib + paper.pdf
You: paper-generator audit output/paper.tex --bib-file output/references.bib
System: [Three-dimensional audit]
Citation integrity — 2 DOI invalid, 1 metadata mismatch
Deep verification — Section 3's 2nd claim lacks PDF evidence support
AI flavor detection — Discussion section language density high
→ audit_report.json (CRITICAL: 0, HIGH: 2, MEDIUM: 3, LOW: 1)
```
You say "explore transformer pruning". The system crawls 30+ papers from five academic databases, clusters directions, evaluates feasibility, generates a paper, and audits itself. The entire process requires no manual intervention.
---
## Generation Pipeline
**Explore → Plan → Write → Verify → Typeset → Review**
Each stage's Agent reads the previous stage's output to ensure no omissions.
### 13 Agents
| Agent | What it does |
|-------|--------|
| **DataAgent** | Parses experimental CSV, Data-View Decoupling to prevent LLM from fabricating data coordinates |
| **PlanningAgent** | Generates IMRaD outline, locks word count and journal constraints |
| **KnowledgeAgent** | Retrieves literature, constructs citation graph, binds each paragraph with DOI anchor |
| **NarrativeDirector** | Designs global story arc, arranges sections in human writing order (Method→Results→Intro→Discussion) |
| **LiteratureSynthesis** | Upgrades Related Work from "reporting names" to thematic grouping and critical analysis |
| **VoiceAlignment** | Extracts style profile from sample .tex, Few-Shot Style Cloning |
| **WritingAgent** | Writes sections, controls citation density, prohibits clichés and placeholders |
| **DevilsAdvocate** | Generates 3-5 adversarial questions, forces Discussion to reflect deeply |
| **VerificationAgent** | Checks logical consistency + citation accuracy + method reproducibility |
| **VisualVerify** | TCVR cycle: TikZ compilation → VLM visual evaluation → LLM repair → up to 3 rounds |
| **TypesettingAgent** | Applies journal templates (IEEE / ACM / NeurIPS / Springer / Nature, etc.) |
| **ReviewAgent** | Simulates peer review with four roles (Critique / Support / Methodology / Editor) |
| **RebuttalAgent** | Responds to each review comment, generates revised abstract |
---
## Research Team — 23 Virtual Research Roles
Use slash commands in AI assistants (Claude Desktop / VS Code Copilot) to invoke. Each role encapsulates specific MCP tool combinations.
### Core Roles
| Role | Skills | What it does |
|------|------|--------|
| Chief Investigator | `/research-direction` | Explores direction, selects strategy, plans iteration, records feedback |
| Literature Reviewer | `/lit-search` | Multi-source literature retrieval, historical memory search, competitive landscape analysis |
| Methodology Consultant | `/methodology` | Research design, statistical planning, hypothesis formalization |
| Experiment Designer | `/experiment-design` | Baseline, ablation factors, resource budget, timeline |
| Data Analyst | `/data-analysis` | Data processing, table/graph generation, statistical analysis |
| Narrative Director | `/narrative` | Paper story arc, section structure blueprint |
| Paper Writer | `/draft` | Invokes generate_paper to generate complete LaTeX |
| Peer Review Team | `/peer-review` | Four reviewers' adversarial review + unified audit |
| Citation Verifier | `/verify` | DOI verification + claim-evidence matching + PDF download |
| Language Editor | `/polish` | Polishing, AI flavor removal, style calibration |
| Publishing Manager | `/publish` | Typesetting, compilation, submission package preparation |
| Lab Supervisor | `/research-pipeline` | Complete A→F pipeline arrangement + quality gate |
### Extended Roles
| Role | Skills | What it does |
|------|------|--------|
| Competitive Landscape Analyst | `/landscape` | Research field mapping, differentiation opportunities, white space identification |
| Blind Spot Analyst | `/gap-analysis` | Coverage gap, interdisciplinary opportunities, prioritization |
| Hypothesis Amplifier | `/hypothesis` | Competitive hypothesis generation, comparison matrix, human selection gate |
| Style Cloning Expert | `/style-align` | Reference style extraction, tone alignment, AI flavor removal |
| Visualization Expert | `/visualize` | TikZ/pgfplots chart generation + TCVR verification cycle |
| Devil's Advocate | `/challenge` | Five-angle attack, claim-evidence audit, discussion enhancement |
| Ethics Officer | `/ethics-audit` | COPE compliance, AI use disclosure, data ethics |
| Literature Manager | `/bibliography` | BibTeX repair, DOI refill, PDF download |
| Rebuttal Writer | `/rebuttal` | Review comment response, revised abstract |
| Progress Manager | `/checkpoint` | Checkpoint save/restore, pipeline status |
| Retro Host | `/retro` | Efficiency analysis, quality evolution, lessons learned |
### Research Pipeline Six Stages
| Stage | Role | Quality Gate |
|------|------|--------|
| **A. Exploration** | `/research-direction` `/lit-search` | At least 1 direction score ≥ 75 |
| **B. Design** | `/methodology` `/experiment-design` | Experiment plan approved |
| **C. Analysis** | `/data-analysis` | At least 1 table + 1 figure |
| **D. Writing** | `/narrative` `/draft` | .tex compilation successful |
| **E. Quality** | `/peer-review` `/verify` `/polish` `/challenge` `/ethics-audit` | Review score ≥ 6/10 |
| **F. Publication** | `/publish` `/bibliography` `/rebuttal` | Format compliant + submission package ready |
Critical decisions use Voice A (LLM) + Voice B (DevilsAdvocateAgent) dual-model confrontation. Disagreements > 3 points are submitted to the researcher for decision.
---
## MCP Server — 31 Tools
Native interface for AI assistants. Claude Desktop, VS Code Copilot, ChatGPT direct call.
```bash
# stdio mode (recommended)
paper-generator-mcp
# SSE mode
paper-generator-mcp --transport sse
```
**Claude Desktop configuration:**
```json
{
"mcpServers": {
"paper-generator": {
"command": "paper-generator-mcp",
"args": []
}
}
}
```
### Original Tools (16)
| Tool | Purpose |
|------|------|
| `explore_topic` | Six-source literature crawling + LLM direction mining + feasibility evaluation |
| `generate_paper` | 13 Agent pipeline generates complete LaTeX paper |
| `polish_paper` | 7 Agent polishing (language + AI flavor + structure + review simulation) |
| `verify_citations` | DOI verification + metadata checking + citation stacking detection |
| `deep_verify_citations` | PDF extraction + LLM claim-evidence semantic matching |
| `audit_paper` | Citation + deep verification + AI flavor, three-dimensional unified audit |
| `download_pdfs` | Automatically retrieves PDFs from Unpaywall / S2 / arXiv |
| `memory_search` | Semantic search historical exploration results |
| `memory_feedback` | Marks direction as feasible/unfeasible/need adjustment |
| `list_checkpoints` | View pipeline cache and exploration progress |
| `analyze_gaps` | Blind spot identification + cross-direction opportunities |
| `compare_landscape` | S2 + OpenAlex + ArXiv three-source competitive landscape analysis |
| `plan_next_step` | Comprehensive exploration/blind spot/landscape, decides next action |
| `execute_step` | Execute plan step (whitelist limited) |
| `design_experiment` | Structured experiment plan (hypothesis, baseline, dataset, metrics) |
| `auto_research_loop` | One-click multi-round iteration: explore → gaps → landscape → plan → execute |
### New Tools (15)
#### Phase 1: Exposing Existing Agents
| Tool | Corresponding Role | Purpose |
|------|----------|------|
| `analyze_data` | `/data-analysis` | CSV data → LaTeX table + pgfplots chart |
| `design_narrative` | `/narrative` | Generates story arc, section instructions, writing order |
| `compile_latex` | `/publish` | LaTeX fragment → PNG preview |
| `assemble_document` | `/publish` | Assembles final LaTeX document (TypesettingAgent) |
| `generate_rebuttal` | `/rebuttal` | Responds to review comments |
| `extract_style_profile` | `/style-align` | Extracts writing style fingerprint from reference paper |
#### Phase 2: Enhancing Existing Capabilities
| Tool | Corresponding Role | Purpose |
|------|----------|------|
| `simulate_review` | `/peer-review` | Multi-role review simulation + stress test |
| `generate_figures` | `/visualize` | Data chart generation + TCVR visual verification cycle |
| `generate_adversarial_critique` | `/challenge` | Five-angle adversarial questioning |
| `synthesize_literature` | `/methodology` | Citation list → thematic classification review |
| `plan_paper_outline` | `/hypothesis` | Hypothesis generation + paper outline planning |
#### Phase 3: New Capabilities
| Tool | Corresponding Role | Purpose |
|------|----------|------|
| `write_section` | `/draft` | Generates LaTeX by section (Introduction, Method, etc.) |
| `fix_bibliography` | `/bibliography` | BibTeX automatic repair (DOI refill, deduplication, normalization) |
| `check_ethics_compliance` | `/ethics-audit` | COPE compliance + AI disclosure + data ethics check |
| `save_research_checkpoint` | `/checkpoint` | Save/restore research project state |
**Autonomous exploration cycle:**
```
explore_topic → analyze_gaps → compare_landscape → plan_next_step → execute_step → ...
```
`auto_research_loop` encapsulates the entire cycle, converging automatically (best_direction_score ≥ 75 and coverage ≥ 60).
---
## CLI Commands
8 subcommands, covering the complete research workflow.
### Common Parameters
| Parameter | Description | Default Value |
|------|------|--------|
| `--provider` | `openai` / `anthropic` / `qianwen` / `local` | `openai` |
| `--model` | Model name | `gpt-4o` |
| `--api-key` | API key (or `LLM_API_KEY` environment variable) | — |
| `--config` | JSON configuration file | — |
### Command Quick Check
```bash
# Exploration (5 Agent pipeline)
paper-generator explore <query> [--seed-papers DOI1,DOI2] [-o OUTPUT_DIR]
# Generation (13 Agent pipeline)
paper-generator generate <topic> [-d DATA_DIR] [-v VENUE] [--from-exploration FILE]
[--no-review] [--no-human] [--style-reference paper1.tex] [--resume]
# Polishing (7 Agent pipeline)
paper-generator polish <tex_file> [--intensity conservative|balanced|aggressive]
[--num-reviewers N] [--target-journals J1,J2]
# Full process series
paper-generator chain <topic> [-d DATA_DIR] [-v VENUE] [--skip-verify]
# Citation verification (no LLM)
paper-generator verify-citations <tex_file> [--bib-file FILE] [--check-doi] [--check-metadata]
# Deep citation verification (needs LLM)
paper-generator deep-verify <tex_file> [--pdf-dir DIR]
# Unified audit
paper-generator audit <tex_file> [--skip-deep-verify] [--skip-ai-flavor]
# Tool List
## PDF Download
paper-generator download-pdfs <bib_file> [--pdf-dir DIR] [--email EMAIL]
**Supported Journal Formats:** IEEE / ACM / NeurIPS / ACL / AAAI / ICLR / ICML / Springer / Nature
**Supported LLMs:** OpenAI / Anthropic / / Local Models
---
## Installation
```bash
pip install -e . # Basic installation
pip install -e ".[dev]" # Development dependencies (pytest, ruff)
pip install -e ".[mcp]" # MCP Server
pip install -e ".[vision]" # Visual verification (pdf2image)
pip install -e ".[deep-verify]" # Deep citation verification (PyMuPDF)
pip install -e ".[langgraph]" # LangGraph autonomous loop backend
pip install -e ".[exploration]" # Semantic search (sentence-transformers)
```
**Optional External Tools:**
- TeX Live / MiKTeX — PDF compilation
- [LanguageTool](https://languagetool.org/) — Grammar checking
---
## Web UI
```bash
paper-generator-ui # or python -m paper_generator.web_ui
```
Visit `http://localhost:7860`. 6 tabs: Environment configuration, Paper generation, Research exploration, Paper polishing, Result viewing, Log monitoring.
---
## Architecture
```
[Research topic + Experimental data + Target journal]
│
▼
Exploration subsystem (5 Agent + Autonomous loop)
DiscoveryAgent → ParserAgent → DirectionAgent → FeasibilityAgent → ReportAgent
│
▼
Generation pipeline (13 Agent)
Data → Planning → Knowledge → Narrative → Synthesis → Voice → Writing →
DevilsAdvocate → Verification → VisualVerify → Typesetting → Review
│ ↑_________|
│ (up to 3 revision rounds)
▼
Polishing subsystem (7 Agent)
LangPolish → AIFlavor → StructOpt → FormatCheck → JournalMatch → RevSimul → Report
│
▼
Audit tools (citation + deep verification + AI flavor)
│
▼
[Journal-level manuscript + Audit report + COPE statement]
```
### Three Layers of Quality Protection
| Layer | Timing | Protection |
|------|------|------|
| **Prevention** | During generation | Citation density control, prohibition of placeholders, limited boilerplate, no section preview sentences |
| **Polishing** | During polishing | Boilerplate removal (11 hedging patterns), content shrinkage protection, paragraph splitting |
| **Safety net** | Before compilation | Citation merging, near-duplicate paragraph removal, 6 LaTeX structure repairs |
### Visual Verification
- **Data-View Decoupling**: CSV data read at compile-time via `\pgfplotstableread`, preventing LLM fabrication of coordinates
- **TCVR loop**: TikZ compilation → PNG rendering → VLM visual evaluation → LLM repair → up to 3 rounds
- **TikZ template library**: 8 pre-verified skeletons (architecture diagrams, pipelines, bar charts, heatmaps, etc.)
---
## Project Structure
```
new-laxtax/
├── paper_generator/
│ ├── main.py # CLI entry point (8 subcommands)
│ ├── pipeline.py # Main generation pipeline
│ ├── llm_client.py # OpenAI-compatible API client
│ ├── mcp_server.py # MCP Server (16 tools)
│ ├── web_ui.py # Gradio Web UI
│ ├── agents/ # Paper generation Agent (13)
│ │ ├── data_agent.py # Data-View Decoupling
│ │ ├── planning.py # Outline generation
│ │ ├── knowledge.py # Literature retrieval
│ │ ├── narrative.py # Narrative direction
│ │ ├── synthesis.py # Literature synthesis
│ │ ├── voice.py # Style alignment
│ │ ├── writing.py # Section writing
│ │ ├── devils_advocate.py # Adversarial questioning
│ │ ├── verification.py # Logic/citation verification
│ │ ├── visual_verify.py # TCVR visual verification
│ │ ├── typesetting.py # Journal layout
│ │ ├── review.py # Simulated peer review
│ │ └── rebuttal.py # Referee comment response
│ ├── exploration/ # Exploration subsystem (5 Agent + Autonomous toolchain)
│ │ ├── orchestrator.py # Pipeline orchestration
│ │ ├── discovery.py # Six-source literature discovery
│ │ ├── parser.py # Feature analysis
│ │ ├── direction.py # Direction mining
│ │ ├── feasibility.py # Feasibility evaluation
│ │ ├── report.py # Report generation
│ │ ├── gap_analyzer.py # Blind spot analysis
│ │ ├── landscape.py # Competitive landscape analysis
│ │ ├── planner.py # Next step planning
│ │ ├── auto_loop.py # Autonomous research loop
│ │ ├── experiment_design.py # Experiment design
│ │ ├── memory.py # Exploration memory
│ │ └── langgraph_loop.py # LangGraph backend
│ ├── polishing/ # Polishing subsystem (7 Agent)
│ │ ├── orchestrator.py # Polishing orchestration
│ │ ├── lang_polish.py # Language polishing
│ │ ├── ai_flavor.py # AI flavor detection
│ │ ├── struct_opt.py # Structure optimization
│ │ ├── format_check.py # Format compliance
│ │ ├── journal_match.py # Journal matching
│ │ ├── review_simul.py # Review simulation
│ │ └── cope_declaration.py # COPE statement
│ └── tools/ # Independent tools
│ ├── verify_citations.py # Citation verification
│ ├── deep_verify.py # Deep citation verification
│ ├── audit.py # Unified audit
│ └── latex_compiler.py # LaTeX compiler
├── research-team/ # Virtual research team (23 roles/ skills)
│ ├── SKILL.md # Router
│ ├── TEAM.md # Role reference manual
│ └── <23 skill dirs>/ # Role SKILL.md
├── tests/ # Tests
├── data/ # Experimental data (CSV / figures)
├── config.example.json # Configuration example
└── pyproject.toml # Project metadata
```
---
## Configuration
```json
{
"llm": {
"provider": "",
"model": "",
"base_url": "",
"temperature": 0.3,
"max_tokens": 4096,
"timeout": 300
},
"retrieval": {
"serpapi_key": "",
"openalex_api_key": "",
"crossref_email": "",
"semantic_scholar_api_key": "",
"max_papers": 30,
"recent_years": 3
},
"polishing": {
"reviewer_strictness": "moderate",
"citation_stacking_threshold": 3
},
"venue": "springer",
"human_in_the_loop": true
}
```
Full example see [config.example.json](config.example.json).
**Environment Variables:**
| Variable | Description |
|------|------|
| `LLM_API_KEY` | LLM API key |
| `LLM_BASE_URL` | API endpoint |
| `GRADIO_SERVER_PORT` | Web UI port (default 7860) |
---
## Changelog
### v0.9.0 — MCP Tool Full Coverage (31 tools)
**Added 15 MCP tools**, making all 23 virtual research roles have at least 1 dedicated tool:
**Phase 1 — Exposing existing Agent (6 tools):**
- `analyze_data` — DataAgent packaging: CSV → LaTeX table + pgfplots figure
- `design_narrative` — NarrativeDirectorAgent packaging: story arc + section instructions + writing order
- `compile_latex` — StandaloneCompiler packaging: LaTeX snippet → PNG preview
- `assemble_document` — TypesettingAgent packaging: assemble final LaTeX document
- `generate_rebuttal` — RebuttalAgent packaging: respond to referee comments
- `extract_style_profile` — VoiceAlignmentAgent packaging: style fingerprint extraction
**Phase 2 — Enhancing existing capabilities (5 tools):**
- `simulate_review` — ReviewAgent packaging: multi-role review simulation + stress testing
- `generate_figures` — DataAgent + VisualVerifyAgent: figure generation + TCVR loop
- `generate_adversarial_critique` — DevilsAdvocateAgent packaging: five-angle adversarial questioning
- `synthesize_literature` — LiteratureSynthesisAgent packaging: topic classification literature review
- `plan_paper_outline` — PlanningAgent packaging: hypothesis generation + outline planning
**Phase 3 — New capabilities (4 tools):**
- `write_section` — WritingAgent packaging: section generation LaTeX
- `fix_bibliography` — BibTeX automatic repair (DOI refill, deduplication, author name normalization)
- `check_ethics_compliance` — COPE compliance + AI disclosure + data ethics check
- `save_research_checkpoint` — research project state save/restore
**Other improvements:**
- `execute_step` whitelist expanded to 24 tools
- Updated TEAM.md role→tool mapping
- Updated copilot-instructions.md tool documentation
### Research Team v1.1.0 (2026-04-17) — Skill system reinforcement
- 24 SKILL.md all added Error Handling tables (customized error handling strategies by skill type)
- Added `shared/dual-voice-protocol.md` (five-angle attack, consensus table template, >3 point escalation rule)
- Cross-platform support: `shared/preamble.md` added bash↔PowerShell equivalent command table
- Corrected tool list: hypothesis (removed generate_paper, added memory_search), lit-search (added analyze_gaps), methodology (added memory_search), research-pipeline (added execute_step)
- Skill-specific enhancements: review rating scale, 6-source parallel instructions, deadline awareness, generation failure recovery, API key configuration guidelines
### v0.8.0 (2026-04-13) — Powerful polishing + MCP one-stop
**Phase 1: Independent citation verification tool**
- Added `verify-citations` CLI subcommand
- DOI verification (Crossref API), metadata cross-validation (Semantic Scholar)
- Citation stacking detection, SQLite cache, JSON report output
**Phase 2: AI flavor detection Agent**
- Added `AIFlavorAgent`, inserted into polishing pipeline LangPolish→**AIFlavor**→StructOpt
- 6-dimensional detection: sentence pattern repetition, hedging language density, claim overload, list/parallel overload, passive voice density, citation stacking
- Per paragraph AI-flavor score 0-1 + specific problem annotation
- AI-flavor automatic correction loop: high-score paragraphs automatically request rewriting suggestions from LLM
**Phase 3: Visual integration into polishing process**
- Automatically extract images before polishing and invoke VLM to generate descriptions
- StructOpt added caption-figure consistency check
- RevSimul review simulation incorporates figure evaluation
- FormatCheck added caption verification
**Phase 4: MCP Server**
- Added `paper-generator-mcp` entry point (FastMCP SDK)
- 16 tools: basic 6 + exploration memory 3 + autonomous exploration 4 + autonomous loop + experiment design + PDF download
- Support stdio / SSE transmission mode
- Path traversal security protection
**Phase 5: Deep citation verification v2**
- Added `deep-verify` CLI subcommand
- PyMuPDF extract cited PDF text
- LLM claim-evidence semantic matching, credibility score 0-1
**Phase 6: CEO Review comprehensive reinforcement**
- Added `audit` CLI subcommand — unified audit (citation + deep verification + AI flavor)
- `chain` command added automatic verification, `--skip-verify` can be skipped
- API client refactoring, reviewer personality system, security fixes
**Phase 8-10: Autonomous exploration + PDF download**
- `auto_research_loop` one-click multi-round iteration
- Multi-source literature parallel capture (6 sources)
- `design_experiment` structured experiment plan
- `download-pdfs` get PDF from Unpaywall / S2 / arXiv
- LangGraph StateGraph optional backend
**Research Team Phase 1-2: Virtual research team**
- 12 core roles + 11 extended roles = 23 research skills
- Dual-channel adversarial protocol (Voice A + Voice B)
- Six-stage pipeline + quality gate
**Phase 7: Autonomous exploration toolchain**
- Added 4 MCP tools: `analyze_gaps`, `compare_landscape`, `plan_next_step`, `execute_step`
- Added exploration memory system: `memory_search`, `memory_feedback`, `list_checkpoints` (3 MCP tools)
- `GapAnalyzer` blind spot and cross-direction analysis (LLM + heuristic fallback)
- `LandscapeAnalyzer` competitive landscape analysis (Semantic Scholar literature capture + LLM differential evaluation)
- `Planner` next step planning (deep_dive / pivot / broaden / need_data / converged)
- `execute_step` whitelist distributor, security restrictions can adjust tool range
- Exploration memory: TF-IDF semantic search, feedback weight adjustment, persistent storage
- 32 new tests
### v0.6.0 (2026-04-10) — Pipeline Robustness Upgrade
**6 Key Bug Fixes**, eliminating the cascade of failures with `chain` command exit code 1:
| # | File | Issue | Fix |
|---|------|------|------|
| 1 | `exploration/discovery.py` | OpenAlex API returns `"source": null` causing `NoneType.get()` crash | Use defensive chaining: `((item.get("primary_location") or {}).get("source") or {}).get("display_name", "")` |
| 2 | `agents/planning.py` | `chat_json()` call without try/except, JSON parsing failure crashes directly | Add exception capture, throw `RuntimeError` with context |
| 3 | `agents/knowledge.py` | `chat_json()` call without try/except, causing pipeline interruption | Add exception capture + empty result fallback (`{"references": [], "citation_plan": {}}`) |
| 4 | `agents/review.py` | `chat_json()` call without try/except, review stage failure terminates | Add exception capture + empty dictionary fallback, `_parse_review()` returns default ReviewReport |
| 5 | `llm_client.py` | `_try_repair_json` uses simple counting to repair truncated JSON, cannot handle brackets within strings | Rewrite as **stack-based parser**: track `in_string` state, maintain ordered structure stack, regex clean incomplete key-value pairs |
| 6 | `pipeline.py` | `_run_stage()` re-raises all stage failures, non-critical stage crashes terminate the entire pipeline | Distinguish **critical stages** (`data`/`planning`/`writing`/`typesetting`) from non-critical stages; non-critical stage failures only log warnings and continue |
**Root Cause Analysis**: OpenAlex NoneType → JSON truncation repair failure → unprotected `chat_json` calls bubble up → pipeline re-raises all stages → exit code 1. Six fixes form a deep defense.
### v0.5.x — Quality and Typesetting Improvements
- **WritingAgent**: Enhance `_SECTION_ROLES` (statistical rigor, reproducibility, limitations), anti-overstatement rule #11
- **TypesettingAgent**: Compliance statement sections (Ethics, Data Availability, COPE AI statement), `_wrap_long_lines()` fixes overfull boxes
- **LangPolishAgent**: balanced mode paragraph density instructions
- **sanitize_latex**: Floating body boundary fixes, wide table adjustbox double packaging protection
- **converter.py**: Windows cp950 encoding fix (`encoding="utf-8"`), `-no-shell-escape` security hardening
- **User-defined Sections**: Pipeline + WritingAgent + WebUI + CLI `--user-sections`
