Generate a paper in the dashboard

After reviewed research results are recorded in your project, open Manuscript → Generate paper. Enter a target venue (include the year when it matters), choose a model and reasoning effort, and click Generate paper. New Codex paper generations default to GPT-5.6-Luna · High when that model is available. Subsequent generations retain your chosen settings. Writing and automatic repair share a 3-hour execution budget; cancellation remains available throughout. The project’s internal coding agent writes the article, creates figures from saved evidence, checks references, uses an official venue template when available, and builds the PDF.

The page shows human-readable progress and Cancel generation. You can switch pages or close the browser while the server continues working. Return to Manuscript for the latest status. Once generated, Paper appears beside Manuscript in the sidebar, with the browser’s built-in PDF viewer, Download PDF, Download source, and Review notes. Keep the dashboard server running. A server restart interrupts unfinished generation; the page offers a retry. When the target and frozen evidence are unchanged, retries reuse any existing draft and figures as starting material. The agent rechecks content completeness and expands underdeveloped methods, analysis and figures using available evidence. It records the review in qa/content-review.md, separating writing omissions from research gaps. A venue template alone is not a full conference article; page limits are ceilings, and missing research cannot be repaired by padding. After an interruption, a same-evidence retry also carries forward the last diagnostics and prioritizes unfinished compilation and checks; completed work and already verified records are reused. For an unchanged snapshot, the agent starts with the existing draft and concrete issues, retaining valid checks and rerunning those affected by its edits.

The source ZIP includes LaTeX, plotting code, the evidence snapshot, bibliography, and build/review notes. A generated paper is a draft for human review, not an approved submission. An unavailable requested-year template must be disclosed; an older style must never be renamed to look current.

Setup contract for the coding agent

This section is part of one-instruction setup. The setup agent prepares the tools once; paper generation subsequently happens inside the product. Do not manually write a paper as a substitute for testing the button.

1. Reuse the authenticated Codex or Claude Code installation configured for the dashboard. No separate model or image API is required. Research is processed by the selected provider using its existing login; public citation/template lookup requires network access.

2. Find Python 3.12+, Git, latexmk, a LaTeX installation including common conference packages, and Poppler (pdfinfo, pdffonts, pdftoppm, pdftotext). Reuse existing tools. For missing tools use the platform’s normal installer, preserving existing environments. On macOS, MacTeX supplies LaTeX and brew install poppler supplies previews. On Debian/Ubuntu, the usual packages are latexmk texlive-latex-extra texlive-fonts-recommended poppler-utils. For Claude Code on Linux/WSL2 also install its sandbox prerequisites (bubblewrap and socat). Claude paper generation requires its native sandbox; on Windows run that backend in WSL2. On Windows with Codex use a supported TeX distribution and Poppler on PATH. Verify the installed commands rather than assuming the package installation succeeded.

3. From the source checkout, run the product’s installer with that Python:

   python3 templates/default/ui/paper_tools.py --setup
   

   For an installed generated project the same command is python3 ui/paper_tools.py --setup. Set COAUTO_PAPER_TOOLS_DIR only if a custom location is required; use the same environment for setup and the dashboard. The default is under ~/.co-auto-research/paper-tools/.

4. Require the command to report "smoke_test": "passed". It installs the four pinned skills, creates an isolated Python environment, checks the skill helper commands, compiles a tiny LaTeX document and renders its page. It normalizes child-process locale settings, including on macOS. A missing executable, failed download or failed compile is an incomplete paper setup; resolve it and rerun the command. Dashboard setup can be reported separately.

5. Open Manuscript and confirm the target/model/reasoning controls appear. Do not start an experiment or generate a research paper merely to check installation. When the user asks for a paper, start it with the UI button.

The installer uses a sparse Git checkout of just these four skills from K-Dense-AI/scientific-agent-skills, pinned to commit 9cf7d9aea7d84754db4c167ab04b299d33c444bc:

Skill	Responsibility
scientific-writing	Evidence-bound prose, consistency checks and unresolved human review
scientific-visualization	Reproducible numeric figures, legible styles and provenance
citation-management	Retrieve, enrich, deduplicate and validate reference metadata
venue-templates	Verify actual venue rules, template provenance and PDF format

They are copied into each isolated writing workspace and explicitly named in the internal agent prompt. No global skill installation or new coding-agent session is needed. The upstream license and citation metadata remain in the tool cache; the setup receipt records source hashes. Python dependencies are pinned in ui/paper_tools.py and installed in that cache’s virtual environment, separate from both the dashboard and research experiment environments.

Evidence and execution boundaries

Paper generation requires reviewed v2 results recorded in the project and waits until project agents are idle. It snapshots the current manuscript, findings, published trial reports and small saved scientific artifacts. Unpublished stages, chat sessions, secret files, raw dataset/model arrays and transient caches are excluded. Oversized inputs fail with an actionable message rather than being silently omitted.

The internal agent works in a separate service-owned paper workspace. It must read the frozen snapshot, use all four skills, and generate figures from existing metrics. It must not train models, rerun held-out evaluations or edit the original research. The service checks snapshot hashes after writing. Recording new scientific results remains a separate autoresearch operation.

Writing and compilation run as supervised processes with time limits and a cancellation path. The service recompiles with shell escape disabled, checks for unresolved references/citations, overflow, font problems and colliding figure labels, returns concrete errors to the writing agent for at most two repair passes, renders every PDF page and releases a separate artifact copy only after these steps pass. The browser exposes only named artifacts belonging to the current project. Successful compilation is not proof of scientific accuracy: inspect the preview and review notes before using the paper. Figure text checks catch overlapping words and near-touching numeric labels; they do not establish that annotations are clear of plotted data. The writing agent must also inspect the rendered figures and pages visually.

The manuscript must preserve evidence limits, failed experiments and uncertain technical causes. It must not manufacture results, authors, bibliographic records, novelty claims, ethics approvals or human verification. If the K-Dense skills materially contribute, their own citation guidance applies and the agent records that use alongside the paper’s other tools.

The Paper entry appears alongside Manuscript in the project sidebar only after a paper has been generated. Start the first generation and follow its progress in Manuscript. Its spinner tracks generation even while you use another page. A green check means a draft was generated, not that it passed scientific or submission review. The PDF opens in the browser’s built-in reader, with its native zoom, scrolling, search, text selection and print controls where supported. Open PDF opens it in a separate tab for more reading space or browsers that cannot display an embedded PDF. No manuscript regeneration is needed for existing papers. Generate new draft keeps the previous PDF available until its replacement succeeds, including after cancellation, failure, or a server restart. When research advances, the paper shows its older evidence revision and offers generation of a new draft.