koinon 26.11.12

Follow uploads without refreshing the page

Library availability and the Uploads panel keep their observations current, including processing that lasts more than a minute and uploads that appear after opening an empty list. Temporary read failures retry automatically. The last known state stays marked as outdated until a new authorized observation succeeds; checks never repeat an upload or processing request.

Status indicators combine text and an icon with consistent colors: blue for work in progress, green for a confirmed successful state, amber for review or uncertainty and red for failure. Transfer completion remains separate from processing completion and searchable knowledge. The shared indicators also distinguish an email service accepting a message from confirmed delivery. See Library for the upload journey and recovery actions.

This change adds automatic observation to Library and Uploads. Administrative job, index and platform-health views retain their existing refresh behavior; their automatic observation is a separate follow-up.

Keep indexing calls outside database transactions

Document indexing releases its database transaction before waiting for the embedding provider. Before publishing an index, the worker checks the current source version, chunks, provider configuration, permissions and cancellation state in a new transaction. A provider response does not authorize publication on its own.

Installation model checks follow the same sequence. The worker records its intent, calls the provider outside the database transaction and rechecks the installation lease and configuration before recording the result. This reduces the time a slow provider occupies a database connection without weakening publication checks.

Cancelling a parent indexing operation prevents a recovery worker from starting a new index switch. Recovery still reports work that needs reconciliation; an uncertain outcome is not silently replayed or reported as a completed cancellation. See knowledge indexes for the worker lifecycle and recovery rules.

Measure file intake with real Office documents

PowerPoint extraction accounts for accumulated text incrementally instead of repeatedly summing all preceding slides. Content order and configured extraction limits remain unchanged.

The knowledge performance study includes measured extraction time, CPU time and memory for genuine Word, PowerPoint, Excel and CSV fixtures. The reported sample sizes and environment define the scope of those results. They do not establish production capacity, large-file limits or cloud processing costs.

Update the installation

This release requires PyFly 26.11.7 or later within the 26.11 series. Source installs pin the immutable commit of the official v26.11.07 framework release. Wheel-based installs must install the framework wheel separately, as described in the installation guide.

Use 26.11.12 consistently across the runtime packages, SDKs, management interface and desktop extension. Apply the normal migration and role checks, preserve the existing installation identity and run the first-start verification after updating the API and workers.

This patch does not add a database migration beyond the histories required by 26.11.11. It preserves uploaded originals, knowledge versions, answer evidence and communication history. Verify ingestion and an evidence-backed answer on the updated deployment. Updating the runtime does not replay welcome emails or establish that a recipient received a previous message.

SharePoint, OneDrive and Google Drive synchronization and the governed RAG/RLM transition workflow are separate work in progress. They are not activated by this patch. Existing desktop OAuth and personal-token connection behavior is unchanged; see the desktop installation guide.