DeepRead v0.5.4 is the current stable release of the evidence-first reading assistant for DeepSeek Harness, Codex, Claude Code, and other Agent Skills-compatible tools.
Highlights
- Five reading modes: quick, deep, knowledge map, Feynman, and whole book.
- Batch comparison for 2–10 documents, including conflicts, complementary evidence, and a synthesis report.
- Page-aware claim and citation tracing for PDFs.
- Four confidence levels and explicit “no evidence provided” markers.
- Token, model-call, and time estimates before running an expensive read.
- Live parsing and reading progress for long documents, large PDFs, and batches.
- A browser reading panel with recent-read history, budget preflight, and one-click start.
- Markdown, HTML, and FreeMind/XMind-compatible export.
- Portable Agent Skill manifests for Codex and Claude Code.
v0.5.4 improvements
- Stream PDF parsing progress page by page inside background jobs.
- Stream per-document parsing, reading, and completion progress for batch jobs.
- Select realistic cold-start throughput by model family and self-calibrate from actual calls.
- Fix the browser budget-preflight route registration order.
- Keep the official pending card and long-running job status visible throughout deep reads.
Install
Full DeepSeek Harness plugin:
dsh plugin --profile web add dsh-deepread@^0.5.4Portable Agent Skill:
npx skills@latest add xiehuan123/dsh-deepreadSee the repository README for real output examples and the differences between the full plugin and portable skill forms.