Ship the code. We draft the thread.

A subtle caching bug in Hono, now fixed. The ETag middleware hashed only the first chunk of a streamed response and treated that digest as the whole body. Two responses that differed only after the first chunk would produce the same ETag, and a client holding the old copy would get a 304 for content that had changed. Hashing now runs incrementally across every chunk and finalises when the stream closes. If you stream responses and rely on ETags, this is worth picking up.

## ETags are now correct for chunked responses

- The ETag middleware hashed only the first chunk of a streamed response, so two different bodies could share an ETag.
- Hashing now runs incrementally across every chunk, and the digest is finalised once the stream closes.
- A conditional request against a changed streamed body now correctly misses the cache instead of returning 304.

A small change, and the kind you feel every single time you use the app. Opening a photo in Immich used to wait for the full-resolution file before drawing anything. On a slower connection that meant a blank frame for a second or two, on every photo you opened, which is the sort of thing people experience as "this app is slow" without being able to say why. The viewer now shows a small preview immediately and sharpens it as the full image arrives. It also picks the preview size based on your connection, so a slow network is not handed a file it cannot download quickly. Perceived speed is mostly about whether something appears, not about how long the whole thing takes.

## Photos now appear the moment you open them

- The viewer used to wait for the full-resolution file before drawing anything, so a slow connection meant a blank frame on every photo you opened.
- A small preview now renders straight away and is replaced as the full image arrives, so there is always something on screen.
- The preview size is chosen from your current connection speed, so a slow network is not handed a large file it cannot download quickly.

Astro shipped a small change to content collections that removes a genuinely annoying debugging loop. A reference() pointing at a nonexistent entry used to pass the sync step and only fail at render time, often on a route unrelated to the one you were editing. The stack trace pointed at the renderer rather than the content file. Reference IDs are now validated during sync. A typo fails immediately, and the error names the id, the collection, and the file the reference came from. Shorter feedback loops are most of what good DX is.

## Broken collection references now fail at sync, not at render

- A reference() pointing at an entry that does not exist used to pass sync and only surface as an error when the page rendered.
- Reference IDs are now checked against the target collection during sync, so a typo fails immediately with the offending id and collection named.
- The error names the file the bad reference came from, so you are not searching the content tree for it.

A small fix in Hono with an outsized failure mode. The client collapsed multiple cookies into a single header value, so only the last one set actually made it into the request. If you set a session cookie and then a preference cookie, authentication quietly broke, in a way that looks like a server problem rather than a client one. Cookies are now serialised as separate pairs and all of them arrive. The bugs worth writing about are usually the ones that fail silently in the wrong direction.

## The client now sends every cookie, not just the last one

- Multiple cookies were collapsed into a single header value, so only the last one survived the request.
- Cookies are now serialised as separate pairs joined correctly, so a session cookie and a preference cookie both arrive.