Paper Reach

Finding The Repository Root

If the current working directory is not obviously the repo root, search these common locations and use the first directory that contains both AGENTS.md and paper_reach/cli.py:

for dir in \
  "." \
  "${CLAUDE_PLUGIN_ROOT:-}" \
  "${GEMINI_EXTENSION_DIR:-}" \
  "$HOME/.claude/skills/paper-reach" \
  "$HOME/.agents/skills/paper-reach" \
  "$HOME/.codex/skills/paper-reach"; do
  [ -n "$dir" ] && [ -f "$dir/AGENTS.md" ] && [ -f "$dir/paper_reach/cli.py" ] && PAPER_REACH_ROOT="$dir" && break
done

If you only need execution and the package is already installed, you can skip repository path discovery and call the CLI directly.

Execution Preference

Prefer these commands:

paper-reach doctor
paper-reach example-query
paper-reach screen --input query.json --output screen.json --high-recall --retrieval-limit 120
paper-reach review --input query.json --output review.json --local-path ./papers
paper-reach fetch-fulltext --input query.json --output review.json --download-dir ./downloads
paper-reach run --input query.json --output result.json

Fallback if the console script is unavailable:

python -m paper_reach.cli doctor
python -m paper_reach.cli screen --input query.json --output screen.json

Guardrails

• Never treat title-only relevance as enough for final selection.
• Treat abstract-only support as coarse evidence.
• Keep unsupported papers as ambiguous or need_fulltext.
• Prefer JSON outputs over free-form prose when handing results back to the user or another agent step.
• If remote retrieval fails, fall back to offline mode or local files instead of inventing evidence.