Titouan Mathis

I know that you hate Agent Swarm, so I created the Pi-Squad extension @badlogicgames. Multiple agents work together using Pi RPC. A planner agent breaks your task into multiple subtasks with dependencies and manages them. A QA agent verifies the work at the end; if anything fails, the squad creates a new task for the developer agent to rework, followed by another round of QA.

@kaelig @stylelint Still early but I am following github.com/csskit/csskit closely for this

Been experimenting for the last week to enable shorter feedback loops that are agent-friendly at enterprise scale. Excellent gains from Nx, pnpm, oxlint, oxfmt, vitest, tsgo, Storybook 10 🤘 Still looking for a @stylelint alternative!

@zeeg Did this for the CLI and MCP we are working on, DX is nice and agents can jump from MCP to CLI if needed.

Internet, give me some opinions: What if we had the Sentry MCP able to reuse the credentials from the new Sentry CLI? Meaning you could setup the CLI (or MCP) and have a large set of capabilities locally. This might only work for stdio but not sure yet. What I'm looking to do is solve people hard coding api tokens, and also recognizing that the remote oauth spec is still a pain in my ass. There are still some other complexities here, in that our implementation would need a simple scoping bridge (so you can choose skill selection, for example), but that is solvable.

@bentlegen Really interesting! How do you handle permissions from different users? Is it only managed by the slack allow list?

Baudbot is starting to put in work: - always on - juggles multiple repos - responds to AI code review feedback - shares Vercel preview links

This resonates a lot, I recommend reading both the shared link and thread by Mario

@kcosr @badlogicgames Nice! I’d love to test it out

@badlogicgames Weird you say? Like using Claude/Codex/other CLI as a provider? Anthropic can't ban me for using pi now! (wip)

pi extensions can now register custom providers, including oauth support and custom API support. If you have some corporate endpoint, some weird 3rd party LLM API, or any combination, you can now trivially add support via an extension. github.com/badlogic/pi-mo…

@badlogicgames I have been using a similar system (Claude diary) for the last few weeks, it really helped steering agents to follow to my workflow and habits. Highly recommend!

Pratham (on DIscord discord.com/invite/nKXTsAc…) had the idea to scrape pi session logs and let the agent come up with AGENTS.md additions, prompt templates, skills, or extensions based on the user behaviour it observed. Vibed a simple script. Intersting results! gist.github.com/badlogic/4de86…

@matthieunapoli I made the switch to pi (shittycodingagent.ai) after trying Claude Code and OpenCode, hard to go back! Plus, the migration was easy

Trying OpenCode: - incompatible with Warp, had to switch to Ghostty 👎 - migration from Claude Code sucks, nothing automatic - the UX is great 👍 (lost some Warp shortcuts though) - no Claude Code web 👎 Total migration time: ~3 hours 😢 Now testing "is it actually better than CC?" So far it seems to work faster but that's just a feeling (kept Opus 4.5)

@zeeg Curious about this too 👀 ours is very similar for now and does not feel optimized yet

Iterated a bit and expanded our internal skills: github.com/getsentry/skil… If you're doing this in your team, would love to know if you've got any tips for keeping this organized more efficiently or making it more usable for large numbers of developers.

@steipete @TimWachter I’d say Claude diary are more like metadata to the commits that can be used as documentation for future sessions

@titouanmathis @TimWachter aren't commits bascially the diary?

I see sooooo many solutions folks are building to access old coding sessions - I never once had the need for this?

@steipete @TimWachter I do small sessions and call /diary before closing, and every few day use /reflect to sum up what has happened. I find it useful to update local and global CLAUDE.md files with my or a project preferences/tools/workflows.

@titouanmathis @TimWachter When in your workflow do you use this?

@TimWachter @steipete I have been using Claude Diary for long term memory and self reflection, it’s quite good at summarizing sessions either globally or per project. github.com/rlancemartin/c…

@steipete How do you maintain that? I've been doing a large reverse engineering project and my docs directory is now has 65 markdown files with decreasing relevancy by age. Never saw tools like that, but yesterday just so happend to have the thought about a "context tool" like that

@mitsuhiko pi is quite nice: github.com/badlogic/pi-mo…

If you are using the Sonnet or Opus 4.5 for coding, but you are not using Claude Code, I would be curious what you use instead and why (eg: amp, Cursor, …)

@badlogicgames I was rethinking about this while reading the mom sandbox doc. Maybe running a different docker sandbox for each user from the slack org could help? User A would get a different container from user B, avoiding leaking tokens or other private tools.

@titouanmathis In my example, anyone in the channel can. I'm currently trying ideas how rto do ACL etc. Hard problem.

I truely wished I could show you what mom, my agentic Slack bot with a computer strapped to its back, is capable of now after I gave her access to emails, the CRM, invoices and wires. She helps resolve complex purchasing/licensing issues by plowing through old email threads, comparing data from the CRM/invoices/wires etc. All with human supervision and checks, as she surfaces all the info to the UI. Everything is collaborative, but I'm still saving an enormous amount of time manually searching for info and correlating it. Then she can help formulate a reply to the customer, detailing the entire history of a complex situation and (based on what I decided to do) a resolution offered to the customer. Finally, she creates a draft email in GMail, attaches any PDFs/files necessary for documentation and gives me a GMail web link to open and do the final check. This is truely amazing.

@zeeg There is no GitLab integration at the moment (or I misread the docs). Any plans to support it?

If you're using Sentry and you havent yet tried Seer's code review, why not? If your answer is "we already use Bugbot", consider using both - we do!

Fixing numbers alignment with CSS titouan.dev/notes/2025/09/…

An Alfred Workflow for SSH aliases titouan.dev/notes/2025/08/…

Using Oxlint LSP server in Sublime Text titouan.dev/notes/2025/08/…

Hey @ShopifySupport, we can not access our partner dashboard or any of our shop, any ongoing issue on your end?