The workflow I keep coming back to: write a permissive policy, let the agent run for a week, then tighten the rules and replay the old sessions to see what would have been blocked. Much better than guessing at policy upfront, and it’s the part of the tool I didn’t expect to use as much as I do.

Every gated decision gets written to jsonl, so you can grep, diff, or feed traces back through a stricter policy without re-running the agent. There’s also a TUI for browsing sessions, inspecting individual gate decisions, and stepping through replays interactively, which makes it easier to spot patterns across runs.

Currently works with Claude Code and MCP-based clients like Codex.

Still a WIP and mostly a project for myself, but figured others experimenting with coding agents might find it interesting.

GitHub: https://github.com/chocks/agentctl

Comments URL: https://news.ycombinator.com/item?id=48057567

Points: 3

# Comments: 0

We’re helping businesses get off the ground by enabling them to focus on running their business. At JustWorks, we tackle complex HR challenges so our customers don’t have to. We’re a data-driven team that’s constantly iterating and improving, and we care deeply about building meaningful products.

If you’re excited about working in a supportive, entrepreneurial environment and want to have fun while building something impactful, we’d love to hear from you.

Apply here: https://job-boards.greenhouse.io/justworks/jobs/7721539?gh_j...

So, I built Sleuther, a Oh My Zsh plugin that asks a local LLM via Ollama to explain and debug errors directly in the terminal. Right now I am using Qwen2.5-Coder.

At a previous job, I used warp.dev extensively and was missing its agent mode. Sleuther is inspired by warp and zsh-copilot.

No copy paste, no cloud, no context switching.

Also, has a small trick: if the local LLM cannot help after a couple attempts, it suggests a sanitized version of the error, stripping usernames, paths, and hostnames before sharing with ChatGPT or Claude. The sanitization is rather rudimentary right now but something I want to improve.

Curious if anyone else is experimenting with local LLMs for dev workflows?

https://github.com/chocks/sleuther

Comments URL: https://news.ycombinator.com/item?id=47491133

Points: 2

# Comments: 0

also just posted a Show HN thread: https://news.ycombinator.com/item?id=47310543

To help answer that question, I started building Locode, a open source CLI that tries this approach.

The idea is: • run simple tasks locally • route complex reasoning to Claude • reduce inference cost and latency • keep the workflow local first

This project is still very early and mostly a fun learning experiment for me. The entire project was built using Claude Code (not vibe coded). I really love the workflow and it inspired a lot of the design. I’m also a huge fan of Ruff, so I took some inspirations from that as well (no rust yet though).

There is a short demo video in the README if you want to see it in action.

Please take it for a spin if you are interested and let me know what you think and/or if you have experience with cli tools and suggestion on improving Locode, I’m happy to learn.

Cheers! Chocks

Comments URL: https://news.ycombinator.com/item?id=47310543

Points: 4

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=42328578

Points: 2

# Comments: 2

Comments URL: https://news.ycombinator.com/item?id=40599804

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=36176637

Points: 6

# Comments: 0