CortexDB 0.9.6 — blob extraction repaired under the unified WAL

Patch release. One important fix: multimodal blob extraction works again on the unified-WAL write path — a regression present since v0.9.0, surfaced by an external real-data report on 2026-08-05. Plus cortexdb-connectors 0.2.16/0.2.17 with two batches of real-workspace Slack fixes. No storage format changes, no migration; 0.9.x data directories load as-is.

The fix (the headline)

Since v0.9.0, CORTEX_UNIFIED_WAL (the production default, and the architecture going forward) routed single POST /v1/experience writes through the indexer tail — which rebuilt each stored event without the API-resolved physical BlobRef. A content.kind=blob_ref experience (pdf/docx/image/audio uploaded via POST /v1/blobs) was therefore written and embedded but stamped Complete as plain text: the enrichment scanner's blob filter never matched and the configured Tika/vision/Whisper content processors were never invoked. Only text/plain (inlined at capture) passed through.

Now the indexer tail re-resolves the event's logical blob_id against the same blob metadata store instance the API layer resolves with — shared by construction — applying the same owner ACL (a blob owned by another actor stays opaque; DISC-024 parity) and the same binary-only gate. Wired into both tail apply paths (per-item and bulk) and the legacy crash-reconcile path, with regression pins so it cannot silently come back. Blob-backed experiences land as Pending, the scanner dispatches them, and extraction runs — with enrichment on or off.

Also in this area:

• Loud startup warning when content processors are configured but no LLM router is: that combination gets no extraction scanner (the scanner requires the LLM-router constructor path), and previously failed silently. Content-only jobs make zero LLM calls — configuring CORTEX_LLM_URL does not add LLM spend for extraction.
• The processor knobs — CORTEX_DOCUMENT_PROVIDER, CORTEX_IMAGE_PROVIDER, CORTEX_AUDIO_PROVIDER, CORTEX_VIDEO_PROVIDER, CORTEX_SENSOR_PROVIDER, CORTEX_FFMPEG_PATH — are now registered with the boot-time config lint, so a typo'd provider is flagged instead of silently absent.

Connectors 0.2.16 / 0.2.17 (PyPI)

Two Slack bug-fix batches driven by a real-workspace hunt:

• 0.2.16: content-versioned message keys (edits land as new versions instead of 409-poisoning every sync; display-name changes no longer permanently fail re-sync), live message_changed webhook edits ingest, reactions get their own identity key (ending a 409→500→ redelivery storm), empty webhook signing secrets fail closed (Slack and GitHub), and a bounded stale-thread scan stops incremental sync from silently dropping replies to old threads.
• 0.2.17: Unicode content gate (non-Latin messages ingest), 1 h shared-cursor lookback for multi-channel in-flight loss, webhook bodies capped at 5 MiB before auth (unauthenticated memory-DoS), invalid UTF-8 → clean 400, Slack deliveries acked within the 3 s deadline with background ingest + bounded retries, per-item dead-letter budgets that name every abandoned item, per-channel failure isolation, empty-window cursor hold, HTML unescaping, and correct private-channel visibility labeling on both paths (real member rosters via conversations.members, fail-closed channel info, webhook privacy resolution via SLACK_BOT_TOKEN).

Compatibility

• /v1 surface unchanged. No new required configuration.
• Data directories from any 0.9.x release load without migration.
• Blob-backed events ingested while the regression was live were indexed as placeholders; re-ingesting them (same blob id, new write) runs extraction. No automatic backfill is attempted.