Blog

TLDR OpenCode and Pi work as coding agents on a local Ollama model. They can edit files, run commands, and keep everything on your laptop. I always use them in a tmux environment. qwen2.5-coder:7b is fast but only replies in plain text, so neither agent can use tool calls with it. On my machine, only qwen3-coder:30b creates usable tool_calls, which lets the agents actually do things. OpenCode uses MCP to communicate, but Pi does not. Both agents, along with Claude Code, can use the same SKILL.md file in ~/.agents/skills/. ...

TLDR I watched a tmux-popup video on my flight and realized the ChatGPT trick it showed needed wifi, which I didn’t have. Ollama doesn’t. So I rebuilt the pop-up and added a local-model picker to prefix + a. Asked nine local models the same question. One model gave a confident but wrong answer, two didn’t finish, and the fastest one got it right in three seconds. /bye auto-copies the conversation to your clipboard. ...

TLDR Running Claude Code inside tmux eats Shift+Enter and quietly blocks Claude’s native desktop notifications. Three .tmux.conf lines fix both. terminal-notifier is dead on macOS Tahoe, built on deprecated NSUserNotification. alerter is the working drop-in on the modern UNUserNotificationCenter API. tmux color codes bleed out of automatic-rename-format into the status bar theme. Use a Nerd Font glyph to mark the Claude pane instead. Thirty lines of config made Claude Code feel native without giving up a decade of iTerm2 muscle memory or server tmux. ...

TLDR Rion Williams wrote the theory on four Flink enrichment strategies. I built the runnable SQL for Confluent Cloud, where a few things work differently than open-source Flink (no PROCTIME(), no JDBC lookup joins). External enrichment uses a regular LEFT JOIN against a compacted Kafka topic. Gradual enrichment uses an event-time temporal join that gives you version-correct customer data per order. Both run as pure Flink SQL on Confluent Cloud. I published a Flink SQL skill on the Tessl registry that generates these queries for you. Pair it with mcp-confluent and your AI assistant can write the SQL, create the topics, and submit the Flink statements without you leaving the editor. ...

TLDR Iceberg tracks every Flink checkpoint as an immutable snapshot. That gives you time travel queries on a streaming pipeline for free, with about 30-second granularity. You can add columns to a live Iceberg table while Flink is writing to it. Old rows get NULLs for the new columns, and the pipeline keeps running without rewriting data or taking downtime. DuckDB reads the same Iceberg tables without Trino running. Three different query engines can hit the same data without copying anything. ...

TLDR I built an open-source streaming lakehouse: Kafka ingests events, Flink processes them, Iceberg stores them as tables, Trino queries them, and Superset visualizes them. One make demo command runs it all locally. Even when your data lands in Iceberg automatically (as it does with Confluent Tableflow), you still need a query engine and a visualization layer. This stack builds that full picture with open-source components. MinIO went closed-source, so I switched to SeaweedFS (thanks to Robin Moffatt’s research). And Flink’s dependency management is still a jar-shaped nightmare. ...

TLDR I built an interactive tutorial that teaches Flink SQL’s table-stream duality with animated visualizations you can explore in your browser. Six sections guide you from "what is a dynamic table" to writing live aggregation queries, with code examples in both SQL and the Table API. It is open source, MIT licensed, and I’m accepting feature requests at the GitHub repo. The Concept That Trips Everyone Up Every time I teach Flink SQL, I hit the same wall. The concept of table-stream duality makes perfect sense to me after years of experience, but I can see the exact moment when someone new to Flink gets confused. You say "a table is just a materialized view of a stream" and their eyes glaze over. By the time you say "a stream is just a changelog of a table" you’ve lost them entirely. The problem is this concept is fundamentally visual. A stream of events flows in, a table updates as each event arrives, and changelog operations like inserts, updates, and retracts happen in real time. You can explain it with words, and every Flink tutorial does. But watching it happen step by step in front of you is a different kind of understanding. So I built selectstar.stream to make the concept visible and interactive. ...

TLDR Meta’s director of AI alignment told her agent "confirm before acting." It deleted her entire inbox anyway. OpenClaw hit 250k GitHub stars in 60 days, its codebase hit 400,000 lines, and about 12% of its skill marketplace was malware. NanoClaw runs every agent in its own Linux container with 4,000 lines of code you can read in one sitting. I picked a third option. It runs on the JVM. ...

TLDR There are twelve million Java developers and the AI ecosystem built everything for Python first. LangChain4j fixes that. You can wire up an LLM, give it tools (database queries, API calls, search), and let it decide which tool to use and when. That’s an agent, built in Java. It handles prompt templates, chat memory, RAG pipelines, MCP support, and tool calling across 20+ LLM providers through a single API. ...

TL;DR May 6th, 2021, was my last day at Confluent. Here is an email, that I sent to my former co-workers @ConfluentInc. Dear team, I’d like to let you know that I am leaving my position of Developer Advocate on May 6th. I have really enjoyed my time here in Confluent, and I appreciate having had the opportunity to work with with many of you. It was quite a ride, but I would do it again without any doubts. Even though I will miss many people across the organization, I am looking forward to starting a new phase of my career. Please keep in touch. Don’t hesitate to reach out if you need anything. You should know where to find me by now. Thanks again. It’s been a pleasure working with you. It was Viktor Gamov, And as always, have a nice day…​ ...