Hey everyone quick update from the Zagens team.
v0.8.1 is out. If you ve been waiting to run a DeepSeek V4 coding agent outside a browser tab, the headline is: Kernel V3 is now the sole production turn engine, and zagens-tui ships as a first-class full-screen terminal surface on Linux, macOS, and Windows same runtime threads and harness as the desktop app.
What s new
Kernel V3 (event-sourced turn engine)
We replaced the legacy imperative turn loop with an append-only KernelEvent log in sessions.db, a LiveTurnMachine planner, and an EffectInterpreter that executes model calls, tool batches, steer injection, and long-horizon hooks. Session resume is log-first; CI verifies turns with golden replay fixtures instead of shadow bakes. Desktop, TUI, and headless CLI all hit the same path no lite terminal fork.
Hey Product Hunt 👋
I built Zagens because I kept hitting the same loop: the agent writes code confidently, I find a hallucinated function name three files deep, and suddenly I'm spending an hour auditing everything it touched. It didn't mean to break things — it just has no way to verify its own work.
The problem isn't model IQ. We dropped superhuman autocomplete into an engineering workflow with no code review, no CI, no "are we actually done?" — just tokens and on to the next turn.
So instead of waiting for a smarter model, I built the engineering layer around it — a Windows harness, not prettier chat bubbles:
• Multi-agent pipeline — explore (read-only), implement, review; failed review loops back automatically
• Verification gates — builds, tests, and grep weigh in before a step is marked done
• Long-session discipline — checklist, auto-continue, turn-by-turn replay so work doesn't quietly drift
I also merged Office (Excel, Word, PowerPoint, PDF) into the same app — I was tired of switching between AI and document tools.
Local-first preview · BYOK · keys stay on your device · built for DeepSeek V4 (OpenAI-compatible too). Not affiliated with DeepSeek Inc.
zagens.com · didclawapp@gmail.com
If you've wrestled with agents that sound confident but ship broken code — what verification patterns have you wired into your own workflow? Would love to hear it. 🙏
Update: Zagens is now open source 🎉
Repo: https://github.com/didclawapp-ai/zagens
Local Windows harness for DeepSeek V4 — long-horizon tasks, review loops, Office + code, BYOK. Preview builds still at zagens.com/download if you just want to try it.
Issues & PRs welcome — especially harness reliability and Office workflows. Thanks for following along since launch day.
Launch day’s wrapped — thanks to everyone who stopped by, upvoted, or followed (means a lot for a solo Windows preview).
The daily leaderboard is behind us; iteration isn’t. Zagens will keep shipping: harness reliability, Office workflows, and everything we discussed in the thread.
Feedback anytime: zagens.com · didclawapp@gmail.com 🙏