Fixes
- Chroma search works again while the embedding model downloads (#6399, closes the "Invalid port" part of #6365 reported by @FengJie2004, root cause found by @kevincommits). httpx 0.28 misread a bracketed IPv6 entry such as
[::1]inNO_PROXYand failed withInvalid port: ':1]', so chroma-mcp could not add documents and search fell back to FTS5. The chroma-mcp child now gets the same host without brackets (::1). Entries with a port, and every other entry, are unchanged. - Context injection keeps observations whose concepts are an old plain string (#6393, thanks @jiangjiang248, closes #6389 reported by @soarkit). Keyword search matched a legacy concept such as
缓存, but context injection dropped the same observation. Both now read concepts through one shared SQL expression, so they cannot drift apart again. - OpenCode Go answers My Work refresh and corpus steps (#6395, thanks @HemanRamlal). The API agent now sends the
x-opencode-sessionheader OpenCode Go requires (one id per run), and turns reasoning off withreasoning_effort: none, which Go honors. Before, every step failed with400 MissingSessionID, or ran out of its answer budget while thinking. - Expired live observer inputs are counted and reported (#6394, thanks @Zhong-fan, closes #6391 reported by @abhiroopb). The live buffer still drops an input nobody claims within five minutes, and still never replays it. The WARN line now carries a running total, and
/api/processing-statusreturns it asexpiredObserverInputs, so the loss is visible without reading logs. /api/tool-usesrefuses a bad date instead of ignoring it (#6387, thanks @kevincommits). AdateStartordateEndthat was not a number (an ISO date, a typo, or the parameter given twice) was dropped, and the listing returned rows from every date. It now answers 400INVALID_DATE. Epoch-millisecond values work as before./api/dedup/candidatesrefuses a fractional limit instead of answering 500 (#6388, thanks @REI-330).limit=2.5failed withdatatype mismatch. A limit that is not a whole number now answers 400limit must be a whole number, and the store clamps what it binds so a negative limit cannot read every row./api/advisor-callscaps its offset like every other listing (#6396, thanks @DeryFerd). A hugeoffsetmade SQLite scan and discard rows on the worker's shared connection. It now uses the shared listing cap, and a viewer that keeps paging stops withoffsetCapReached.
Hosted sync service
- The sync API checks the projector URL before sending its secret (#6397, thanks @DeryFerd).
INTERNAL_PROJECTOR_URLnow follows the same rules asFORWARD_ORIGIN(absolute https, plain http only for loopback, no credentials, query or fragment) and fails at startup. Before, a typo or anhttp://address sent the internal bearer secret in cleartext. This changesservices/sync-api, which is not part of the npm package: it reaches users only once that service is deployed.
Internal
- One helper now decides whether a listing offset is past the cap (#6398, thanks @DeryFerd). Behavior is unchanged.
- The 16.0.3 changelog now tells OMP users to run
npx claude-mem install --ide omponce to get the server-runtime fix.
Full changelog: v16.0.3...v16.0.4