Repository navigation
v2.0.5 — Runtime heap, memory settings, compatible Check
Patch after v2.0.4: a 4 GiB Node heap in the runtime image, durable assistant-memory settings saves, and a streaming OpenAI-compatible Connectivity Check. Upgrade like any patch — pull 2.0.5 or :latest, restart, let migrations run.
Highlights
- Runtime image ships
--max-old-space-size=4096so long tool chats stop dying at the default ~2 GiB V8 cap - Assistant Memory settings keep queued plugin and memory writes instead of dropping overlapping saves
- OpenAI-compatible Connectivity Check streams First/TTFT, honors the Responses radio, and ignores SSE comments
- Settings provider pane keeps Check/Fetch and brand tiles inside the card
Upgrade
- Image:
docker.io/lqdflying/chathub:2.0.5or:latest - Database migrations run on container start
- Optional sidecars unchanged (Code Interpreter, MarkItDown)
- Docker Deployment and Upgrades
Note
The image NODE_OPTIONS is now --max-old-space-size=4096 --dns-result-order=ipv4first --use-openssl-ca. A matching one-line override in lobe.env is still valid. A second NODE_OPTIONS line in that file overwrites the first. Do not raise the heap on an ~8–10 GiB host.
Fixed
Runtime
- Production
NODE_OPTIONSno longer drops the 4 GiB heap flag when setting DNS/TLS options
Assistant memory
- Overlapping settings patches serialize instead of clobbering
- Queued plugin writes keep distinct overlay tokens and ownership
- Pending plugin selections survive queued saves
- Settings refresh after a durable
saveMemory
OpenAI-compatible / Check
- Check and Graphile chat honor the Responses API radio
- Connectivity Check streams so gateways record First/TTFT
- Empty Check waits for the SSE stop before passing
- SSE heartbeat comments are not recovered as Check text
Settings
- Provider pane no longer clips Check and Fetch
- Builtin provider tiles use vendored full-bleed avatars
- Header logos stay on the Avatar tile at 24px
Docs
Full diff: v2.0.4...v2.0.5