v6.2.0 — Direct PaperBanana API + Qualitative Engine

What's New

Version 6.2.0 fixes PaperBanana figure generation reliability and adds qualitative data analysis.

Direct PaperBanana Python API (Reliability Fix)

The MCP stdio transport layer was unreliable for PaperBanana's long-running figure generation (~35–45 seconds) — timeouts and silent failures. v6.2.0 bypasses MCP entirely by calling PaperBanana's Python API directly via asyncio.run().

New reliability chain: Direct Python API (primary) → MCP tools (fallback) → matplotlib/seaborn (last resort).

The wrapper script scripts/paperbanana_direct.py supports all PaperBanana features:

• diagram — methodology diagrams from text descriptions
• plot — statistical plots from JSON data
• evaluate — compare generated vs. reference diagrams
• download-references — fetch 294 exemplar diagrams (~257 MB, one-time)
• --aspect-ratio, --optimize, --auto-refine, --save-prompts flags

Qualitative Engine (/analyze-interviews)

New 15th skill engine for qualitative data analysis:

• Summary-first transcript processing (context-window-safe)
• Thematic coding: Gioia, Mayring, Braun & Clarke
• Cross-case analysis with evidence tables
• Structured outputs: codebook, theme map, cross-case matrix

API Key Management

New global key option — no need to copy .env into every project:

echo 'GOOGLE_API_KEY="your-key"' > ~/.paperbanana.env

Loading priority: Environment variable → ~/.paperbanana.env → project .env.

Updated PaperBanana

Pulled latest from llmsresearch/paperbanana:

• Anthropic Claude VLM provider
• AWS Bedrock provider
• Exemplar retrieval for better in-context learning
• Dynamic aspect ratio support
• Prompt saving in run artifacts

Summary

Component	Count
Skill engines	15
Slash commands	16
Scientific skills	24
MCP servers	2

Download

The attached zip contains the complete plugin ready for installation:

# Option 1: From GitHub marketplace
/plugin marketplace add TobiasBlask/open-paper-machine
/plugin install open-academic-paper-machine@open-paper-machine

# Option 2: From ZIP (Cowork or manual)
# Download open-paper-machine-v6.2.0.zip → upload via Plugins panel

v6.1.0 — 24 Curated Scientific Skills from K-Dense AI

What's New

Version 6.1.0 integrates 24 curated scientific skills from K-Dense AI's claude-scientific-skills (MIT License) as a complementary layer to the existing 14 skill engines.

Added Skills (8 categories)

Category	Skills
Research Ideation	hypothesis-generation, scientific-brainstorming, scientific-critical-thinking, consciousness-council, what-if-oracle
Writing	scientific-writing, citation-management, markdown-mermaid-writing
Literature & Review	literature-review, peer-review, scholar-evaluation
Statistics	statistical-analysis, exploratory-data-analysis, statsmodels
Visualization	matplotlib, seaborn, plotly, scientific-visualization, scientific-schematics
Submission	venue-templates, research-grants
Reference Management	pyzotero
Utility	get-available-resources, networkx

How It Works

Skills are stored in scientific-skills/ and follow the Agent Skills standard. Each skill auto-activates based on context — no configuration needed.

Download

The attached zip contains the complete plugin ready for installation.

v6.0.0 — Extended Capability Engines + Paper Revision

What's New in v6.0.0

6 New Skill Engines

• Screening Engine (383 lines) — PRISMA-compliant SLR screening pipeline with quality assessment
• Peer Review Engine (506 lines) — Simulated double-blind review (2 independent reviewers, venue-calibrated)
• Positioning Engine (286 lines) — Differentiation matrix and unique positioning analysis
• Submission Engine (389 lines) — Venue-specific formatting, anonymization, cover letter, reviewer suggestions
• Presentation Engine (273 lines) — Conference slide generation with speaker notes (Marp-compatible)
• Co-Author Engine (179 lines) — CRediT contribution tracking, human-AI labor documentation

7 New Slash Commands

/review-paper, /screen-papers, /analyze-positioning, /analyze-writing, /prepare-submission, /generate-slides, /generate-plot

Extended Existing Engines

• Literature Engine (222→346 lines) — monitoring mode for living literature reviews
• Writing Engine (348→450 lines) — style analysis mode with 8 quality metrics
• Figure Engine — added statistical plot generation via PaperBanana

Pipeline Agent Enhancements

• Proactive tool suggestions at relevant pipeline stages
• Post-production suite: screening after Phase 1, positioning after Phase 2, full suite after Phase 6

Paper (v6.0.0)

• 19-page working paper with revised design principles grounded in theory
• Structured validity discussion (internal, construct, external, conclusion)
• 51 references including Parnas (1972), Flower & Hayes (1981), Swales (1990), Hyland (2005), Vygotsky (1978), Denzin (1978)
• Simulated peer review + full revision round implemented
• Writing quality analysis included

System Stats

• 14 skill engines (~5,300 lines of domain knowledge)
• 15 slash commands
• 2 MCP servers (academic-search, PaperBanana)
• 8-phase pipeline with human-in-the-loop checkpoints
• 13 research method templates in the method engine

v5.4.0 — Review Engine: Automated Revision Loops

What's New in v5.4.0

Review Engine (Phase 8) — the big one

The plugin now automates the entire co-author/reviewer revision process. Derived from 4 actual revision rounds on the From Creator to Orchestrator paper.

/respond-reviewers @annotated_review.pdf

7-step workflow:

1. EXTRACT — Parse PDF annotations (highlights, sticky notes, strikeouts) via PyMuPDF, or parse pasted reviewer comments
2. MAP — Locate each comment in paper.tex (line number, section, context)
3. CLASSIFY — Determine action type (DELETE, REPLACE, MOVE, RESTRUCTURE, FIX, FIGURE, SHORTEN, APPROVE, QUESTION) and priority
4. PLAN — Present structured change plan for approval (quality gate — nothing executes without your OK)
5. IMPLEMENT — Execute changes in dependency order, invoking figure-engine or writing-engine as needed
6. VERIFY — Recompile LaTeX (0 errors), generate latexdiff for visual change tracking
7. DOCUMENT — Change log, optional R&R letter, orchestration log entry, git commit

Supports: Co-author annotated PDFs, journal R&R decision letters, self-review output. Handles multilingual comments (German/English). Repeats for Round 1 → 2 → 3 → ... until acceptance.

New Files

• skills/review-engine/SKILL.md — Full skill definition with 10 action types, dependency ordering, error recovery
• skills/review-engine/review_engine_workflow.png — PaperBanana-generated workflow diagram
• scripts/extract_annotations.py — Reusable PyMuPDF-based PDF annotation extractor (CLI + importable)

Updated

• commands/respond-reviewers.md — Expanded from 3-line stub to full revision command
• agents/paper-machine.md — Added Phase 8 (Revision) as repeatable pipeline phase
• README.md — 8 engines, 8 phases, new Revision Automation section, SSRN paper link

Pipeline Overview (now 8 phases)

Phase	Engine
1. Reconnaissance	literature-engine
2. Framing	theory-engine
3. Structure	writing-engine
4. Production	writing-engine + figure-engine
5. Assembly	latex-engine
6. LaTeX & PDF	latex-engine
7. Verification (opt.)	verification-engine
8. Revision (repeatable)	review-engine

---

Full Changelog: v5.3.0...v5.4.0

v5.3.0 — PaperBanana Paper Citation + Regenerated Figures

What's New in v5.3.0

PaperBanana Academic Reference

• Added citation to the official PaperBanana paper: Zhu et al. (2026), Automating Academic Illustration for AI Scientists, arXiv:2601.23265
• Updated README, figure-engine SKILL.md, and plugin description with paper reference and BibTeX entry
• Links to official research repo alongside community MCP implementation

Documentation

• Fixed version badge (was showing v5.1.0)
• Updated figure generation section to describe the 5-agent, 2-phase pipeline architecture
• Expanded .gitignore for cleaner repository

Figures

• Regenerated pipeline and authorship figures with latest PaperBanana (3 iterations each)

Dependencies

• Pulled latest upstream from llmsresearch/paperbanana (MCP server improvements, Gemini thinking support)

---

Full Changelog: v5.2.0...v5.3.0

v5.2.0 — Orchestration Audit Trail

What's New

Orchestration Audit Trail — Every pipeline run now produces a structured, publishable orchestration_log.md.

Changes

• Agent prompt (agents/paper-machine.md): Added 67 lines of orchestration logging instructions. The agent now creates outputs/orchestration_log.md at pipeline start and appends entries at every phase checkpoint, recording timestamps, actors, outputs, quality gate decisions, and orchestrator feedback.
• All 7 skill engines: Each skill now logs its activation with timestamp, actor, input summary, and output summary to the orchestration log.
• Version: Bumped to 5.2.0

Why

The companion paper "From Creator to Orchestrator?" proposes an Orchestration Trace (§8.2) as a practical instrument for transparent human–AI accountability in research. This release implements a first version: a persistent markdown log that can be published alongside any manuscript produced by the Open Paper Machine.

Log Format

Each entry follows this structure:

### Skill Activation: Writing Engine
**Timestamp:** 2026-02-26 10:30
**Actor:** AI Agent (writing-engine)
**Input:** Draft introduction section
**Output:** 6-paragraph introduction (~800 words)

Quality gate entries additionally record the orchestrator's decision (approved / redirected / rejected) and verbatim feedback.

Full Changelog

9 files changed, 131 insertions, 1 deletion

v5.1.0...v5.2.0