A port of jordan-gibbs/hyperresearch ("the most powerful deep research harness") from Claude Code to opencode.
Same deep-research machine, rebuilt on opencode primitives:
- Tier-adaptive 16-step research pipeline (light / full / dissertation)
- Parallel critic/subagent roster with model-map profiles
- Patch-only draft modification enforced via opencode agent tool locks + plugins
- Persistent vault: markdown notes as truth, SQLite index as cache (search, lint, graph, PageRank)
- Resumable budgeted runs with ship-gate lint battery
- Academic-APIs-first fetch lane with Unpaywall/Europe PMC open-access recovery and untrusted-source fencing
- Extra web sources behind one provider protocol: builtin (default floor), tavily, exa, crawl4ai (optional), parallel (api.parallel.ai) — with an ordered fallback chain (
[web] provider = ["parallel", "builtin"]) and an opt-in parallel search lane (hpr search-web) - MCP server and local web UI
Public contract preserved: hyperresearch <cmd> / hpr <cmd> CLI and /hyperresearch <prompt> inside opencode.
Planned / scaffolding. Execution follows an OpenUltraCode gauntlet loop: each piece is built by an implementer and judged blind against the upstream source by a separate adversarial critic until ours wins. Progress: HYPERRESEARCH-OPENCODE-PROGRESS.md (added during Phase 0).
Upstream: jordan-gibbs/hyperresearch by Jordan Gibbs, MIT licensed. This port adapts that code under the same license.