Everlock 0.10.0

The 0.10.0 release adds two operator tools — an on-demand view of the server's recent log and an explicit restart — makes the git host's address a live setting, names the MCP endpoint at start, and moves the embedded model to its Thinking variant. Binaries are on the download page; the container image is published as cr.everlock.sh/everlock:0.10.

Highlights

/server log shows what the server just did. The server keeps its recent log lines in memory — the same lines that reach stderr, nothing written to disk — and serves them from the admin console, MCP and the HTTP API. count= takes the last N matching lines, since= and until= bracket a window (15m, 2h, 1d back from now, or an RFC 3339 time), grep= matches case-insensitively, level= sets a floor and target= narrows to a module. Every report leads with how much of the buffer it covers, so a window that reaches past what is still held is visible rather than silently empty. The ring's size is the live log.buffer_lines setting, 2000 lines by default.
/server restart restarts on request. The same path a restart-class configuration change takes — flag, drain, re-exec, or a clean exit for the socket-activation supervisor to bring the process back — is now a verb of its own, and an MCP tool.
The git host moves without a restart. git.http.vhost is a registered server setting, and a running git host follows it the moment it is set: routing, the runner server URL, the web UI's instance name and the vhost publication that binds HTTPS, issues a certificate and announces a .local name. The settings tools on MCP and the HTTP API derive their key list from the registry, and the console completes keys from it, so a setting exists everywhere at once.
Connecting an assistant is quicker. With the MCP outlet enabled, the start log names the endpoint — MCP endpoint: http://localhost:8080/mcp — and links to the client setup page, which now carries the settings-file block and the field names Claude Code, VS Code, Cursor and oMLX each expect; the getting-started guide points there for every client beyond Claude Code and Codex.
The embedded model reasons before it answers. MiniCPM-V 4.6 ships as its Thinking variant: same architecture, same size, same vision sidecar, and the binary does not grow. The model works through a hidden block before it answers, which is what makes a caption come back in the language asked for — the previous model answered a German description prompt in English on most photos — and describes a photo more precisely, counting the objects in a frame rather than summarising them. Description prompts for German, French, Spanish, Italian, Portuguese and Chinese are now worded in that language.

Behavior changes

Captioning spends more tokens per photo. The Thinking variant uses roughly twice the tokens of its predecessor, tag lists up to three times. The background captioning job absorbs that as throughput rather than latency; nothing waits on it.
A hot-added .local host is announced at once, on both channels. The site, image, calendar, contacts, files and vault commands said mDNS would announce a new .local name after the next restart; the name is announced the moment the command returns, and the note now says so. The SSDP/UPnP announcer that lists hosts in Windows networking was the half that really did wait; it now takes every host published later, sending its alive NOTIFY right away.

Fixes

IMAP no longer answers a stray * BAD after every APPEND. The literal reader left the CRLF that ends the command line in the stream, where the next read took it for an empty command. The remainder of the line after a literal is now part of that command, as RFC 3501 has it.
Two album share links work side by side in one browser. Opening a share link used to store its guest session in the single login cookie, so the second link a person received replaced the first. Each opened share now keeps its own cookie for the lifetime of its token, the login cookies are left alone, and an admin who opens a share link keeps their login. Links already sent out keep working unchanged.
Captions no longer come back as fragments such as th>, think or a lone .. The chat template's thinking marker reached the model split into sub-word pieces instead of the single token it was trained on, and a short answer could sit entirely inside the engine's holdback buffer and be mistaken for an empty response. Both are fixed in the inference engine; the effect was most visible on non-English descriptions and on photos without text.

Upgrading

Self-updating servers pick 0.10.0 up on their daily check. For a manual upgrade, download the new binary (or pull the new image tag) and restart — stores, mail, and all versioned content carry over untouched. The full list of changes is in the changelog.