Security
v12.11.0 fixes GHSA-989p-hfjf-q97q. With a per-store write grant such as write:projectA, memory_store could write to and supersede memories in another store. Affected: 12.4.0 to 12.10.0, on the backends that confine by store (sqlite-vec and hybrid). Setting MCP_AUTO_EXTRACT_DEFAULT=false is not a mitigation, because an explicit auto_extract argument wins. Upgrading is the fix.
Special Thanks to RISCON Labs (@RISCON-Labs, https://www.risconlabs.com) for reporting it. The conflict-detection item in the same entry was found separately, by the maintainer.
Also in this release: the dashboard offers the openai-compatible quality provider (#1668, KshitijKasodkar).
Security
- A per-store write grant could write to and supersede memories in another store through
memory_store(GHSA-989p-hfjf-q97q, reported by RISCON Labs, @RISCON-Labs).
Withauto_extract=true,handle_store_memorystored the memory in the granted store but ran auto-capture without it.AutoCaptureService.capture(),SessionHarvester._try_evolve()andMemoryService.evolve_memory()took nostore, so the extracted candidates were compared with thedefaultstore, replaced a similar memory there, or were inserted there. A client holding onlywrite:projectAcould overwritedefault. The NLI contradiction check on store (MCP_NLI_ON_STORE, inline and background) had the same gap: it compared a memory with thedefaultstore whatever store it was written to, linked the two with acontradictsedge and reported the hit count to the caller. It also missed contradictions inside a named store for the same reason. Conflict detection on store, which runs on every write without any flag, compared the new memory with every store as well and tagged a near-identical memory in another storeconflict:unresolved, withcontradictsedges between the two; it is confined to the store of the write now.storeis now passed down all of these paths, andevolve_memory()refuses a hash from another store by itself. SettingMCP_AUTO_EXTRACT_DEFAULT=falsewas never a mitigation: an explicitauto_extractargument wins. Affected: 12.4.0 to 12.10.0, on backends that confine by store (sqlite-vec, hybrid). On Cloudflare and Milvus a per-store grant never authorized these calls; there, auto-capture for a named store now stores candidates in that store without evolving existing ones.
Added
- The dashboard offers the
openai-compatiblequality provider (#1668, KshitijKasodkar).
The Quality tab and the environment listing offeredlocal, groq, gemini, auto, none, so
the provider behind Ollama, vLLM, MLX-LM, LiteLLM andfm servecould not be picked, and
MCP_QUALITY_AI_BASE_URL,MCP_QUALITY_AI_MODELandMCP_QUALITY_AI_API_KEYwere not
listed. They are now, with the API key masked likeMCP_MILVUS_TOKEN. Gemini is no longer
offered while_score_with_geminiraisesNotImplementedError. Savingopenai-compatible
from the Quality tab is refused until the base URL and model are set, the same waygroq
needsGROQ_API_KEY, so the restart cannot come back to a server that fails validation.
Fixed
- Hybrid sync progress was shown nowhere on the dashboard (#2002, InlitX).
handleSyncProgress()andhandleSyncCompleted()wrote thesync_progressand
sync_completedevents into a#syncStatuselement thatindex.htmlno longer has,
a manual sync raised one toast per progress event, and every finished sync raised one
more. Hybrid sync now reports to the status island as its own "Cloud sync" source: a
rising percentage while it runs, a confirmation that steps back to "Synced" when it
ends, and a failure that stays until the next sync status. The ten-second
GET /api/sync/statuspoll mapssynced,syncing,pending,errorand paused to
the island without overwriting a measured percentage or a confirmation, and a backend
without hybrid sync reports nothing. - The session-start hook reported every install as 10.13.0 (outdated) (#2019).
readLocalVersion()inclaude-hooks/utilities/version-checker.jstook the first
__version__literal in__init__.py, which is only the fallback for a missing
_version.py. It now reads_version.pyfirst and keeps__init__.pyas the
fallback. - The dashboard status island labeled its reconnect action "Refresh" (#2010,#2022 Shivraj-G).
The island was built without catalog keys of its own, so it borrowedanalytics.refresh,
manage.ontology.column.detailandmodal.close. The reconnect action therefore read
"Refresh" instead of "Reconnect", and the connection row in the panel had no name. The
island now has its own keys (island.reconnect,island.details,island.dismiss,
island.source.connection) in all eleven catalogs, and the connection row is named
"Live updates". scripts/maintenance/sync_status.pyno longer carries Cloudflare credentials. It fell back to a hardcoded API token, account ID and D1 database ID when the environment did not set them, so the token sat in the public repository since January. The fallbacks are gone and the script exits with a clear message naming the missing variables. CI now runs gitleaks over each pull request's own commits, and the pre-commit hook scans staged changes when gitleaks is installed.time_expr="day before yesterday"now selects the day before yesterday.parse_time_expressionmatched the word "yesterday" inside the phrase and returned yesterday's window, somemory_search, the HTTP time search and recall silently returned the wrong day.extract_time_expressionalso left "day before" in the search text. Both now treat the whole phrase as two days ago, with or without a daypart.
Internal
- A request to take over an assigned issue went unnoticed, and so did a claim with a condition attached (#2020).
On #1668 a contributor wrote "I would like to take this if ..." on an issue whose
assignee had been silent for a week.Claim Pendinglogged "comment is not a claim":
the phrase had to end its clause, and an assigned issue was never labelled anyway. A
claim followed byif,unlessoroncenow counts, and a prose claim on an issue
assigned to someone else getsclaim-pending. The assignee's own comment and a bare
/claimdo not. A declined takeover leaves the label for manual removal. storage/mixins/retrieve.pyandstorage/shared.pyjoinGUARDED_MODULES(#2026, part of #1146, massimiliano1991).
Inretrieve.pythe two f-string logger calls (result count with the sanitised
query, result count with the tag list) are lazy%d/%scalls, the tag list goes
through_sanitize_log_valueasstorage/mixins/delete.pyalready does, and the
tag count and a rejected tag's type name are read into their own names before they
are logged. Inshared.pytheexcof_safe_json_loadsis wrapped: the text it
parses is themetadatacolumn read back off a row. A sample test pins the three forms.consolidation/consolidator.py,decay.py,health.py,insights.py, andscheduler.pyjoinGUARDED_MODULES(#1146).
All f-string logger calls are converted to lazy%-style; memory hashes (content_hash), exception values (e,exc,errors,msg) and config paths (db_path) are wrapped with_sanitize_log_value; ternary message selects are split into plainif/elsebranches;insights.pygains the_sanitize_log_valueimport. A sentinel test pins the wrapper requirement for content hashes, exception values, and f-string logger calls across the five modules.web/api/health.pysanitizes its logged exception (#1146).
When the embedding integrity check behind/api/health/detailedfails, the
exception is now passed through_sanitize_log_value, like the same message
inutils/health_check.py, so a newline in it cannot forge a log line. The
module is added toGUARDED_MODULES.