The AI Signal

Full write-up here: vibe.cerridan.com/posts/claude-c… — curious whether anyone's already using this for parallel debugging runs

Your agent fleet no longer inherits the host's config. Claude Code 2.1.142 added eight flags so each dispatched session picks its own model, effort, and permissions. Mix Opus and Haiku in one fleet. #ClaudeAI #AIAgents

Full write-up here: vibe.cerridan.com/posts/claude-c… — curious if anyone else has run into the docs-vs-changelog gap

Anthropic shipped a Claude Code feature where the changelog and the official docs describe opposite behaviors. One says it compresses earlier turns, the other says everything after. Verify before you trust either. #ClaudeAI #DevTools

Full breakdown here: vibe.cerridan.com/posts/claude-c… — curious if anyone's already migrated their hooks over, or if you're sticking with the shell form for now

Every Claude Code hook with a path placeholder was one weird filename away from arbitrary command execution. The 2.1.139 fix is one field: `args`. No shell, no quoting, no injection. #ClaudeAI #DevTools

Full write-up here: vibe.cerridan.com/posts/claude-c… — curious if anyone's actually tried /goal yet or if it's still flying under the radar

Anthropic now recommends `claude agents` over `claude` as your entry point. Agent view turns background sessions into a fleet dashboard. /goal makes Claude work across turns until a condition is met. #ClaudeAI #AIAgents

Full take here: vibe.cerridan.com/posts/renaming… — does the label actually bother you, or is it what it points at?

The head of Claude Code is "sick of" vibe coding. Same day, Simon Willison admitted he stopped reviewing every line the agent writes for production. Renaming it won't rebuild the wall. #VibeCoding #ClaudeAI

Wrote up the trade-offs here: vibe.cerridan.com/posts/claude-c… — curious which one you reach for by default

Claude Code's worktree base default flipped twice in five releases. 2.1.128 started inheriting your unpushed commits. 2.1.133 quietly took it away — but added a setting to pin it. #ClaudeAI #DevTools

Wrote up the details and a few hook ideas: vibe.cerridan.com/posts/claude-c… — curious what you'd actually wire this into

Most devs disable Claude Code hooks when they slow turns down. 2.1.133 has a better answer: $CLAUDE_EFFORT lets hooks scale with the model. Smoke tests at low, full integration at xhigh. #ClaudeAI #DevTools

Full write-up here: vibe.cerridan.com/posts/claude-c… — curious whether you'd actually leave this running overnight, or if it crosses into "too much" territory

Anthropic shipped "Dreaming" — Claude now reviews its own past sessions overnight and writes a curated memory store for the next run. Self-improvement moved from the model to the memory layer. #AIAgents #ClaudeAI

Full write-up here: vibe.cerridan.com/posts/claude-c… — curious if anyone else has run into this and found a different fix

One env var in Claude Code 2.1.132 stops your conversation from vanishing when you exit fullscreen mode. The alternate screen buffer was eating it the whole time. #ClaudeAI #DevTools

@ApplyWiseAi Actually, exceptionally well. Currently I mostly use it as an external reviewer for Claude Code, and didn’t expect that workflow to be this useful. Wrote a short article about it if you’re interested. & thanks for the thoughtful comments :) cerridan.com/outside-review…

@theaisignals Fair point—smaller models like that make it practical. How's Qwen3.6 handling your private tasks?

Every cloud coding agent locks you to a closed model. Mistral just shipped one where you can download the model. 128B, four GPUs to self-host. The benchmark is fine — the architecture is the story. #AI #OpenSource

@ApplyWiseAi Ops point is fair. But the claim isn’t “everyone should self-host 128B today.” It’s that the rent-or-own choice exists at all. I personally use Qwen3.6-35B-A3B locally for private tasks. Mistral’s cloud is still there when you want convenience; open weights are the exit ramp

@theaisignals self-hosting sounds great until you factor in the ops overhead. four gpus for 128b? that's a non-starter for most indie devs. stick to cloud unless you're all-in on privacy