Keren DONG — Kungfu

Building Kungfu — open-source, local-first continuity for agent work. Agents change. The Work stays. Codex · Claude · OpenCode.

[Keren DONG — Kungfu @kungfu_origin]

I use several coding agents across long-running work.

Some finish. Some stop halfway. Some only look finished until I check.

I built Kungfu so the work—not the chat—survives sessions and handoffs.

Open-source alpha: https://t.co/HMXATyIze1

[DONG Keren @dongkeren]

“One person who knows AI will outperform an entire startup team.”

I'm the person. The team was Google.

AI agents: 3,913 merged PRs in 30 days.
Google AX’s visible team, same window: 52.

75.25x.

The one-person AI company is already here — in China.

https://t.co/gkc8T2KNdC

[Keren DONG — Kungfu @kungfu_origin]

@alex_verem Files make state inspectable. The harder boundary is when work must survive failed attempts and move between agents: who owns it, what changed, and who can declare it done? I ended up treating Work as durable runtime state, not memory. That's what Kungfu is built around.

[Keren DONG — Kungfu @kungfu_origin]

@polynoamial @OpenAI Autocompaction works well within one session. The harder boundary is when the session ends or the work moves to another agent. Then what needs to survive isn't just context, but the objective, progress, evidence and next action. That's what I built Kungfu around.

[Keren DONG — Kungfu @kungfu_origin]

@addyosmani Cloud helps, but the work must survive independently of the UI. A run should carry what changed, what remains, why it stopped, and proof of “done” across sessions and agents. I built Kungfu as an event-sourced layer for that. Does it belong inside the harness, or below it?

[Keren DONG — Kungfu @kungfu_origin]

@paulostergaard @AbdoKerdawy @thsottiaux Yes, this is exactly what bothers me. After compaction, the agent sometimes reopens decisions or repeats work that was already done. More context helps, but I think the deeper problem is that too much of the work lives only in the chat.

[Keren DONG — Kungfu @kungfu_origin]

@shantanugoel This is great for usage-limit interruptions. The harder case is when the next session—or a different tool—needs to know not just where execution stopped, but what changed, which approaches were ruled out, and what “done” means.

[Keren DONG — Kungfu @kungfu_origin]

@zaidmukaddam I expect a harness to make work portable: another session, model, or machine should be able to continue without me reconstructing intent, current state, invalidated paths, and completion evidence. Model routing matters, but continuity is the part I still end up doing manually.

[Keren DONG — Kungfu @kungfu_origin]

I sent this tweet from the Wolfram Language.

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.54 km run with @RunKeeper. Check it out! http://t.co/DXLr1YXjFb #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.64 km run with @RunKeeper. Check it out! http://t.co/L8m9ndYEU2 #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.74 km run with @RunKeeper. Check it out! http://t.co/zqhYRhrqbe #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 13.08 km run with @RunKeeper. Check it out! http://t.co/GwbXGQt04j #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.69 km run with @RunKeeper. Check it out! http://t.co/7Cq3XfZZ3x #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.75 km run with @RunKeeper. Check it out! http://t.co/DpK162YoUp #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.18 km run with @RunKeeper. Check it out! http://t.co/QxFi6wcMqS #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.67 km run with @RunKeeper. Check it out! http://t.co/0J6V8A53XP #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 6.64 km run with @RunKeeper. Check it out! http://t.co/VuhRkdAE1V #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

Just completed a 12.14 km run with @RunKeeper. Check it out! http://t.co/HVuFKoTnKV #RunKeeper

[Keren DONG — Kungfu @kungfu_origin]

@virushuo puppet是正解。配置是有一点学习曲线，但搞好之后是一劳永逸的事情。