v0.1.4

Docs: credit TurboAgent (the authors' official implementation of the LLM-as-a-Verifier PPT) + a humble note on why agent-ultramode exists (same-model, agent-agnostic, reasoned verifier + practical adaptations). No code changes.

v0.1.3

Docs release. Roadmap accuracy (Claude Code and cline are verified, so they leave the 'first-class integrations' TODO), first-person voice in Credits, and cline added to the run-a-different-agent list. No code changes since v0.1.0.

v0.1.2

Docs release. Demo link now points to the uploaded video asset (was the committed-file page), and cline is added to the verified-end-to-end agents. No code changes since v0.1.0.

v0.1.1

Docs release. Refreshes the README so npm reflects the current copy: removes the stale GitHub video-instructions note and adds a 'Watch the demo' link where npm strips the inline

v0.1.0

agent-ultramode v0.1.0 — first release.

Best-of-N for coding agents with a same-model verifier. Runs your task N times in isolated git worktrees, verifies the diffs with the same model (probabilistic-pivot tournament, LLM-as-a-Verifier), then applies the confident/majority winner — or hands you the top candidates when they diverge.

• opencode plugin: /ultra <task> (also shows up as a Tab agent mode)
• standalone CLI: npx agent-ultramode "<task>" — works with any agent (opencode by default; Claude Code, cline, etc. via --agent)
• spread attempts across several models in one pass (round-robin --agent)

Verified end to end with OpenCode and Claude Code. Benchmarked on Terminal-Bench with DeepSeek V4 Flash 0731 as both agent and verifier.

npm: https://www.npmjs.com/package/agent-ultramode · MIT