PaPoo
cover

The changelog reads like a bug diary, not a product story

What struck me first is how much of this release is about making Claude Code stop tripping over itself. That’s not glamorous, but it’s the kind of work that tells you where the product actually is: past the “look, it can code” phase and deep into “please don’t corrupt my session, my plugin install, or my prompt cache.”

The obvious headline in this entry is AGENTS.md support. I’d actually try that immediately, because it’s a small but telling quality-of-life change: if a project doesn’t have CLAUDE.md, Claude Code will look for AGENTS.md instead. That makes the tool feel a little less opinionated about its own file name and a little more willing to fit into whatever convention a repo already uses. The fact that this still isn’t on Bedrock, Vertex, or Foundry is the sort of detail that matters more than the feature itself. It reminds you that “Claude Code” is not one thing; there are platform edges everywhere.

What I find more interesting, though, is the sheer density of fixes around malformed local state in ~/.claude.json, weird crashes at launch, broken resumes, and sessions hanging or exiting at the wrong time. That pattern usually means two things. First, a lot of real users are running this in messy environments, not clean demo setups. Second, the team is paying for every assumption they made about config shape, session history, or plugin metadata. I don’t say that as a criticism so much as a reality check. Tools like this live or die on resilience.

There’s also a lot here that feels like hard-earned scar tissue around prompts and tool calls. Empty text blocks breaking conversations, escaped backslashes being misread as Unicode, null bytes sneaking through as literal escapes, giant regex failures where the user should have seen a plain “string not found” message. None of that is exciting, but it is exactly the sort of thing that makes an agent feel trustworthy instead of haunted. If you use Claude Code in anger, these are the fixes you’d notice only when they were missing.

The plugin and marketplace changes are another clue. There are fixes for reinstalling a plugin while it’s in use, for previews leaving unpacked copies behind, for metadata being recorded incorrectly, for “failed to load” rows sticking around after uninstall. That all sounds like the kind of subsystem that grew faster than its cleanup logic. Maybe that’s unfair, but the changelog gives me that impression. It doesn’t read like a polished extension ecosystem yet; it reads like one that has started meeting reality.

I’m also mildly reassured by the attention to platform-specific weirdness: Windows path length problems, winget and apk update behavior, TMPDIR expansion, proxy and NO_PROXY handling, headless sessions losing usage totals. Those are the details that rarely make launch posts and absolutely dominate support tickets. If this changelog had only feature bullets, I’d be less interested. The mess is the signal.


Reference: Claude Code changelog - Claude Code Docs

同じ著者の記事