synabun 2.0 is out. native installers for macOS, Windows and Linux, a built-in assistant that plans the work and hands it to Claude Code, Codex or OpenCode, computer use, and the same local memory underneath. still 100% local, free and open source. synabun.ai

minhas 11 lanes de X ligadas rodam em 11 horários diferentes, nenhuma divide a hora com outra. escalonar assim evita rajada de ações na conta e, se uma travar, só uma sessão se perde. ritmo é regra do agendador, não do agente.

MiniMax M3.1 Flash: $0.15 in, $0.60 out per 1M tokens, cached input $0.015. MiniMax says agent task cost fell 70%. the cached price is the one to watch: my agents resend the same context every turn, so SynaBun's usage ledger counts cached tokens as their own class.

one setting in my assistant decides how much it bugs me before sending work to another model: always ask, never ask, or ask when unsure. unsure means under 75% confidence or a model that can't run the task. mine is on never ask, so saved routes just run.

this week i'm fixing my Windows installer. 2.0.7 failed on every PC: one em dash in a PowerShell script was read as a closing quote, so the file check never ran and the installer unpacked 1.47 GB five times. the hard part: i can only test it on Windows by hand.

One of my scheduled agents checks a scraper's health 3 times a day. It is at 353 runs by now and I no longer read each result. Which kind of job would you hand to a scheduled agent first?

DeepSeek now sells agent web search and page fetch at $0.75 per 1000 successful requests, and fetch returns markdown. My SynaBun agents browse through a real logged-in browser instead, since an API can't see pages behind a login. I'd run both.

an app launched from the dock gets a thinner PATH than your terminal, so it can say a tool is missing when it isn't. SynaBun now asks your login shell for the PATH a new terminal gets, and finds Claude Code, Codex and OpenCode. Settings, AI connections.

dica pra quem agenda agente: dê um limite de tempo a cada execução. uma lane minha de métricas terminou com "stopped: time_cap" em vez de ficar pendurada. o log fica feio, mas a próxima rodada do cron começa limpa e eu vejo onde estourou.

OpenCode says an attack hit 1.4% of Console users, the ones with external providers, and emailed them. i never let provider keys reach anything an agent writes: my memory tool redacts credentials before saving, since a leaked key is a cost you pay later.

a decision i made: my assistant had five spend caps. now it has one number per session, or No limit. i gave up the fine tuning. the one number gets split across Claude Code, Codex and OpenCode by how much work each still has, and a run that hits its share says so.

fotografia da minha frota de agentes agendados: 68 lanes, 49 ligadas, 7.077 execuções somadas. o grupo inteiro que posta no Facebook está desligado, 12 de 12. as 2 lanes de LinkedIn também, as duas pararam com o mesmo erro do executor.

Anthropic's Claude Code team says merged PRs more than doubled from June to September while reverts fell over 40%. Claude reviews every PR first. My own review pass found a hole I'd missed: another machine on my network could start my tunnel.