Read tool reports "pdftoppm is not installed" on Windows despite poppler being on PATH

Summary

On Windows, passing a PDF to the Read tool for page rendering fails with:

pdftoppm is not installed. Install poppler-utils (e.g. brew install poppler or apt-get install poppler-utils)

poppler is installed and reachable, so this is a detection bug on win32, not a missing binary. The suggested remedies are also macOS/Linux only.

Environment

• Claude Code 2.1.203 (native installer, ~/.local/bin/claude)
• Windows 11, Version 10.0.26100.8737 (x64)
• poppler 25.07.0, installed via winget (oschwartz10612.Poppler)
• Node v24.15.0

Evidence it is a detection bug, not a missing/unreachable binary

• pdftoppm -v runs and reports pdftoppm version 25.07.0.
• pdftoppm.exe exists under the winget poppler ...\Library\bin directory.
• The poppler bin dir is in the Claude process's raw inherited environment: bash --norc --noprofile -c 'echo "$PATH"' (no profile sourcing) still contains it.
• A freshly spawned cmd.exe /c where pdftoppm returns the full path with exit code 0.
• Claude Code was fully restarted after poppler was added to PATH, so this is not a stale-env case.
• Rendering the same PDF manually with pdftoppm -png -r 150 ... succeeds, and Read displays the resulting PNG correctly. Only Read's internal PDF path is affected.

Likely cause

Read's poppler detection does not resolve pdftoppm the way a normal PATH lookup would on Windows: it likely looks for pdftoppm without the .exe suffix / without honoring PATHEXT, checks a hardcoded set of directories rather than PATH, or assumes a Unix-style install location. The hardcoded brew/apt guidance in the error string points the same way.

Suggested fix

• Resolve the binary with a PATHEXT-aware lookup (append .exe on win32) or a cross-platform which-style resolver.
• Make the "not installed" message platform-aware (e.g. mention winget install oschwartz10612.Poppler or choco install poppler on Windows).

Workaround (no functional loss)

Shell out to poppler directly instead of relying on Read's built-in rendering:

• Text PDFs: pdftotext -enc UTF-8 -layout input.pdf out.txt, then read the text.
• Scanned / layout-sensitive PDFs: pdftoppm -png -r 150 -f N -l N input.pdf prefix, then read the PNG.