omnilearn-workflow

Installer for OmniLearn v2.3.0 — multi-agent learning workflows for OpenCode.

npx omnilearn-workflow

Full docs: github.com/BlackPool25/OmniLearn

What you get

Six commands installed to ~/.config/opencode/command/, plus research methodology templates under ~/.config/opencode/skills/omnilearn/references/:

• /omnilearn-init — configure your learning directory
• /omnilearn-roadmap — research + generate a learning roadmap
• /omnilearn-roadmap-edit — modify a roadmap without losing progress
• /omnilearn-start — hands-on learning with assignments and tests
• /omnilearn-refine — dig deeper into a specific concept
• /omnilearn-research — deep research engine on any topic

The research engine

/omnilearn-research runs systematic review, not ad-hoc web scraping. Every run follows the same pipeline: background literature review (academic + industry angles) → falsifiable hypotheses → adversarial evidence gathering (confirmation and falsification passes) → calibrated synthesis with explicit uncertainty.

What changed in this version:

• PRISMA-S search strategy tables — every angle logs full query strings verbatim with limits and flow counts, so a third party can rerun the search.
• Contradiction-seeking queries — each angle runs dedicated queries whose goal is to find disagreement and negative results, not confirmation.
• arXiv grounding — the academic angle extracts section-cited notes from primary sources instead of trusting abstracts.
• Hypothesis-first ordering — no retrieval runs before each sub-hypothesis has verification-method, expected-outcome, and falsification criterion.
• Relevance gate + cascade retrieval — every query must target a hypothesis; sources resolve cache → DOI → scholar connectors → web search, with a gap-pass for what is still missing.
• Field-level citation check — every source verified on title/authors/venue/year/DOI and labeled R (real), P (needs re-check), or H (removed).
• Claim lock (Step 3.5) — high-risk claims reach synthesis only with two independent domains plus counter-search plus a primary source; the rest go to an Unresolved/Refuted annex.
• Evidence-bound confidence rubric — ≥85% needs L1–2 sources, 70–85% needs L3 plus survived falsification, below is contested, never a number.
• Six quality gates — arxiv step, claim lock, contradiction floor, falsification coverage, citation labels, and rubric compliance are each enforced with a named verifier before the run concludes.

Flags

| Flag | What it does | |---|---| | --yes | Non-interactive: skills + MCPs, keeps existing omo/opencode | | --skills-only | Skills + MCPs only — never touches omo or opencode | | --skip-omo | Never touches the omo binary | | --skip-opencode | Never touches the opencode binary | | --check | Verify install health (commands, references, MCP, plugin, binaries) | | --version | Show package version | | --help | Show usage help |

Requirements

• Node.js >= 18 — no Bun needed

• OpenCode — install it yourself: curl -fsSL https://opencode.ai/install | bash

• oh-my-openagent (omo) — provides multi-agent orchestration: npx oh-my-openagent@latest install

  Providers are configured later inside OpenCode.

What the installer does

The default scope is skills + MCPs. The installer detects your existing omo and opencode installs first — stable releases and prerelease channels (-beta, -next, -rc) alike count as installed — prints the detected version, and asks before touching either binary. Keeping what you have is always the default, and non-interactive runs keep existing installs. A beta is never downgraded to stable unless you explicitly choose a reinstall. If you manage both tools yourself, --skills-only skips them entirely.

1. Detects existing omo/opencode (version + channel) — asks, defaults to keep
2. Copies the 6 command files to ~/.config/opencode/command/
3. Copies the research templates to the omnilearn skill references dir
4. Configures the Context7 MCP for documentation lookups
5. Registers the oh-my-openagent plugin entry (binaries only on opt-in)
6. Optionally configures your learning directory

Context7 MCP setup

Documentation lookups go through Context7 in remote mode — no API key, no local server. The installer writes this into your OpenCode config (opencode.json, .jsonc also accepted):

{
  "mcp": {
    "context7": {
      "type": "remote",
      "url": "https://mcp.context7.com/mcp",
      "enabled": true
    }
  }
}

If the entry already exists under any *context7*/ctx7 key, the installer leaves it alone. Run npx omnilearn-workflow --check anytime to verify.

Verify

npx omnilearn-workflow --check

Checks detected omo/opencode versions, all 6 commands, reference templates, Context7 MCP, the oh-my-openagent plugin, and your learning directory — warnings include the exact fix command.

License

MIT