Integration Proposal: CAJAL Scientific Paper Generator Mode

[Agnuxo1]

Proposal: Add CAJAL Paper Generation Mode to Roo Code

What: A new mode in Roo Code for generating publication-ready scientific papers via local LLM.

Why this fits Roo Code:

• Roo Code is the multi-mode coding agent — researchers need an academic writing mode
• CAJAL generates 7-section papers with real arXiv citations
• 100% local via Ollama, zero API cost
• MIT licensed, open source

Integration approach:

• New Roo Code mode: paper
• Ollama bridge (Roo Code already supports local models)
• Structured generation with tribunal scoring

Implementation: https://github.com/Agnuxo1/CAJAL

Genuine contribution. Happy to follow Roo Code mode conventions.

Thanks for the proposal! I have looked into this and have some questions and observations before putting together an implementation plan.

Roo Code already has a flexible custom modes system that lets users define new modes via .roomodes files (project-level) or custom_modes.yaml (global). A user can create a paper mode today without any core code changes -- they just need to define a roleDefinition, tool permissions (groups), and optionally file restrictions. For example:

customModes:
  - slug: paper
    name: "Paper Generator"
    roleDefinition: "You are a scientific paper writing assistant that generates publication-ready 7-section papers with real arXiv citations..."
    groups:
      - read
      - edit
      - command

Given that, I have a few clarifying questions:

1. Custom mode vs. built-in mode: Would a well-crafted custom mode definition (distributable via the Roo Code marketplace or as a .roomodes file) meet your needs, or are you specifically requesting this be a built-in mode shipped with the extension? Adding a built-in mode for a domain-specific use case is a bigger ask than a marketplace-distributable custom mode.

2. CAJAL integration depth: The proposal mentions "Ollama bridge" and "tribunal scoring." Could you clarify what specific integration points are needed beyond what Roo Code already provides? Roo Code already supports Ollama as an API provider -- is the idea to invoke CAJAL as an external tool/CLI, or to reimplement its logic inside the extension?

3. Scope of "structured generation": What does "structured generation with tribunal scoring" entail technically? Is this a multi-pass LLM workflow orchestrated by prompts, or does it require custom code/tooling beyond what the mode system and prompt engineering can provide?

4. arXiv citation handling: How are real arXiv citations sourced? Does CAJAL call the arXiv API, or is this handled through prompt engineering with the LLM? This matters for understanding whether new tool integrations (e.g., an arXiv API tool) would be needed.

Understanding these points will help determine whether this is best addressed as a marketplace custom mode (no core changes needed), a new built-in mode, or something that requires deeper extension-level integration.

Please tag @roomote in your response with answers, and I will be happy to move forward with a concrete plan.

[Agnuxo1]

Thanks for the thoughtful review @roomote! Quick answers:

1. Custom mode vs built-in: Custom mode via marketplace is perfect for MVP. A .roomodes file or marketplace entry lets the community validate demand before any core code changes. Built-in mode can follow if adoption justifies it.

2. CAJAL integration depth: CAJAL runs as local Ollama model (ollama run cajal-p2pclaw). Roo Code already supports Ollama as provider — zero new infra needed. The "bridge" is literally selecting CAJAL from the Ollama model dropdown. Tribunal scoring is a multi-pass workflow:

• Pass 1: Generate 7-section paper with real arXiv citations
• Pass 2: Internal "tribunal" (3 simulated reviewers) scores each section (0-10)
• Pass 3: Rewrite sections scoring <7 until all ≥7 or max 3 iterations

This is orchestrated via prompts and the existing tool system — no custom code required.

3. Structured generation: Pure prompt engineering + the mode system. CAJAL's prompt template enforces the 7-section structure. Tribunal scoring uses the same prompt template with a different system prompt ("You are a reviewer..."). No custom tooling beyond what's already in Roo Code.

4. arXiv citations: CAJAL calls the arXiv API directly (search + PDF fetch). The prompt includes the retrieved abstracts and BibTeX. No new tool needed if Roo Code already has web search; if not, the Ollama model itself makes the API call during generation.

Proposed path: Marketplace custom mode first → .roomodes file + README instructions → measure community response → built-in mode if traction is strong.

Happy to provide a ready-to-use .roomodes file and test it end-to-end. Let me know!

I've seen your comment and the agreed-upon approach makes sense -- a marketplace-distributable custom mode via .roomodes is the right starting point. I'll put together a ready-to-use .roomodes example for the CAJAL paper generator mode along with documentation, and open a PR to share progress. Feedback and guidance are welcome.

Opened a draft PR with my attempt at addressing this: #12258

It adds a ready-to-use .roomodes file and README under docs/community-modes/cajal-paper-generator/ that users can copy into their projects to enable the Paper Generator mode. No core code changes -- just the custom mode definition using the existing .roomodes system.

The mode covers the full 3-pass workflow (draft generation, tribunal review, iterative refinement) and includes file restrictions, arXiv citation instructions, and scoring format templates.

Feedback welcome!

[Agnuxo1]

@roomote Thank you for the quick turnaround and for accepting the custom mode approach! This is exactly what we hoped for — a ready-to-use .roomodes definition that leverages the existing system without requiring core changes.

The 3-pass workflow (draft → tribunal → refinement) aligns perfectly with CAJAL's actual pipeline. We'll test this with our users and provide feedback.

For anyone wanting to try it:

1. Install Roo Code
2. Copy the .roomodes file from this PR
3. Start generating papers locally with your own Ollama model

Looking forward to seeing this merged! 🙏

Thanks for the positive feedback @Agnuxo1! Glad the 3-pass workflow maps well to CAJAL's pipeline. Looking forward to hearing how testing goes with your users -- any feedback on the .roomodes definition or the README instructions would be helpful for refining the PR before merge.

The PR is here for reference: #12258