hyperresearch: a deep research harness for Claude Code featuring a 16-step adversarial pipeline and a persistent source vault

What it solves

Hyperresearch is a deep research harness that transforms Claude Code into a sophisticated research agent. It solves the problem of "context-rot" in long AI research pipelines—where agents often forget steps or lose focus—by using a modular, 16-step pipeline that loads fresh context for each phase. It also addresses the issue of one-shot report generation by creating a persistent, searchable vault of sources that compounds over time.

How it works

The system uses a tier-adaptive pipeline (either light or full mode) to move from a canonical research query to a final report.

• Modular Pipeline: It employs a 16-step process including decomposition, width sweeps, contradiction graphing, depth investigation, and adversarial auditing.
• Subagent Roster: It orchestrates a variety of specialized subagents (using Claude Opus, Sonnet, and Haiku) for specific tasks like fetching, source analysis, and synthesis.
• Surgical Patching: To prevent the AI from completely rewriting a good draft, the system enforces a "patch, never regenerate" rule. Final edits are applied as surgical hunks using tool-locked [Read, Edit] permissions.
• The Vault: Every fetched source is stored as a markdown file with YAML frontmatter and indexed in SQLite, allowing for full-text search and provenance tracking across different research sessions.
• Academic-First Approach: The system prioritizes academic APIs (Semantic Scholar, arXiv, OpenAlex, PubMed) over general web searches to ensure high-quality, citation-ranked sources.

Who it’s for

Researchers, analysts, and power users of Claude Code who need high-fidelity, adversarially-audited reports with full source provenance and a persistent knowledge base.

Highlights

• 16-Step Adversarial Pipeline: A comprehensive workflow that includes contradiction analysis and multiple parallel adversarial critics.
• Persistent Knowledge Vault: A SQLite-indexed library of markdown notes that is version-controllable and searchable.
• Surgical Edit Hunks: Mechanical enforcement of draft modifications to prevent total regeneration.
• Authenticated Crawling: Support for authenticated sessions to fetch data from platforms like LinkedIn and Twitter.
• Academic API Integration: Built-in prioritization of scholarly literature over web commentary.