Deep Research: provide a validated local PDF fallback for Codex CLI

What variant of Codex are you using?

Codex CLI 0.153.4, macOS 26.6.2, GPT-6 Astra with xhigh reasoning; deep-research-work@openai-curated-remote 0.1.14.

What feature would you like to see?

Provide a supported PDF-generation path for Deep Research sessions where no native artifact tool or matching artifact skill is exposed. Reuse the maintained PDF capability when available; otherwise document a minimal local recipe and its prerequisites.

The skill requires a professionally formatted artifact when creation and delivery are available, and directs the agent to native Work tools or an applicable artifact skill. In the tested CLI session, neither PDF-specific route was exposed, while Pandoc, XeLaTeX and Poppler were installed. The agent could complete a PDF through local tools, but had to improvise the template and renderer setup.

A small supported fallback should cover:

• A preflight that tests the actual PDF engine and selected fonts. Fontconfig reporting a font does not establish that XeLaTeX can load it.
• A minimal working template with correct metadata initialization and citation links.
• Page-count, text/link read-back and rendered-layout checks, with revisions limited to observed defects.
• Clear capability and accessibility limits when the available local tools cannot meet the requested output.

This is an enhancement request based on an observed workflow. It is not a claim that every Deep Research PDF task fails, or that a native PDF service was promised in this CLI session.

Additional information

A single visible test agent completed this bounded request on September 6, 2026:

Use $deep-research to assess whether two Linux hosts can concurrently use one SQLite WAL database over NFS. Explain concurrency, checkpoints, NORMAL versus FULL durability and version caveats. Use primary SQLite sources and deliver a concise three-to-four-page PDF with clickable citations.

The test explicitly allowed installed local artifact tools and requested full visual review. Observed recovery work:

1. The generated template requested TeX Gyre Pagella, which XeLaTeX could not load.
2. A replacement selected Inter; fontconfig found it, but XeLaTeX still failed. Georgia, Arial and Menlo worked.
3. Generated metadata initialization called hypersetup too early and failed. Deferring it fixed the build.
4. The first successful PDF had six pages. Pagination, overflowing source notes and a literal formatting marker required subsequent corrections.

The final PDF had four A4 pages, 55 link annotations and 20 distinct targets. The agent and parent visually checked the report and rechecked changed pages. The PDF remained untagged; no accessibility certification is claimed.

The research turn took 670.001 seconds. Six measured retrieval batches totalled 13.801 seconds, including one batch with parallel local preflight. These are one-run client observations, not a benchmark, and do not attribute all remaining time to PDF rendering.

The fonts and faulty LaTeX were selected/generated by the agent, not supplied by the Deep Research package. A validated recipe or clearer artifact-capability routing could avoid this class of setup/rebuild work. Refreshing the plugin returned the same 0.1.14 files.

The related font-runtime request #32968 concerns bundled LibreOffice/Poppler host-font configuration. This exercise used locally installed Pandoc/XeLaTeX, so that issue does not establish the cause of these failures. The planner default-off documentation gap is tracked separately in #42365.

Targeted open/closed searches for Deep Research, PDF generation/templates and font issues did not identify an exact duplicate of this fallback request.

Start at the Deep Research artifact-routing path in Codex CLI and inspect how native Work tools or artifact skills are selected. Reproduce the local Pandoc/XeLaTeX/Poppler workflow, including font and PDF checks, then define completion against the requested fallback, capability limits, and accessibility disclosures; no repository file or test is named in the issue.