Release on August 28, 2026

Release

Areas Affected

AI Search, Email

Highlights

• Added the ability for the AI search feature to read and understand content from spreadsheet files (CSV and Excel), making more project information searchable.
• Strengthened security on the internal tools that support email search to better protect customer data.

What Changed

• Enabled the AI search feature to read and index spreadsheet files (CSV and Excel), with follow-up improvements to handle larger files more reliably. (Ticket)
• Improved the reliability of behind-the-scenes data updates and added stronger security protections to the internal tool that powers email similarity search. (Ticket)
• Rolled back a recent reliability improvement to spreadsheet file search after an issue was found. (PR)
• Rolled back the update that enabled spreadsheet file search, temporarily removing this capability while the issue is investigated. (PR)

---

Hotfixes Since Last Release

These fixes were deployed to production independently before this release — they are listed here for a complete record of what changed since the previous release.

No hotfixes since the last release.

---

Chores

• disable DEV Textract testing flags per CDK TODO (#423) (PR)
• defer force_reprocess clear until all shard uploads succeed (#414) (PR)
• harden SSE heartbeat/pacing streaming with ContextVar-safe pump (#480) (PR)
• Reuse Ernest API HTTP client in NLS (#376) (PR)
• accept every send_emails shape the tool accepts, and alert on failure (#482) (PR)
• Batch document vector writes (#373) (PR)
• store embedding shards in a temp bucket with a 7-day lifecycle (#483) (PR)
• add NLS E2E smoke tests for service health, Bedrock, and MCP (iterations 1–3) (#383) (PR)
• move smoke-test tooling to src/smoke_tests and rewrite as pytest (#484) (PR)
• replace log-scraping embedding assertions with a diagnostics payload (#485) (PR)
• remove remaining log-based assertions from E2E embedding smoke tests (#487) (PR)
• fix stale stack comment, type feature-flag records as boolean (#488) (PR)
• Lower Sentry trace sample rate for non-prod environments (#490) (PR)