A dated record of what actually landed on the main branch — grouped by release day, in plain language, with the pull request behind each change. Nothing here is a roadmap or a promise; every line describes work already shipped.
A page audit found that several redirects — including /privacy and /terms — were returning 404 instead of forwarding, because the edge boundary that serves anonymous pages runs before Next's own redirect logic. This release closes that gap, publishes real Terms of Service copy, and ships a round of performance and accessibility fixes across the home page and the workspace.
Every redirect declared in the site configuration, including /privacy and /terms, now actually forwards instead of 404ing at the edge boundary.
Published real Terms of Service copy, corrected metadata title and description lengths, and added og:image and breadcrumb structured data on the affected pages.
The home page no longer downloads its hero video twice, and unused CSS was trimmed from the page.
The assurance brief and review queue now make fewer database round trips, and the review queue pages its rows instead of loading them all at once.
A pool-limit environment variable can now raise the command center's single pooled database connection.
The workflow textarea and progress bar are now labeled for assistive technology, and the evidence and findings lists use valid list markup.
The guided tour's dismissal now persists across sessions instead of reappearing.
A project register with no documents now returns its real (empty) review queue instead of failing.
The connector registry table now shows official vendor marks next to each connector.
Bring your own AI provider, and a steadier agent run
Organizations can now connect their own AI provider credentials and route agent runs through their own endpoint, rather than relying solely on the platform's built-in routing. Alongside that, this release makes an agent run itself easier to trust: a visible timeline, a larger answer budget and longer timeout on organization-supplied credentials, and several fixes to how a run's outcome is recorded.
Organizations can now supply their own AI provider credentials and route agent runs through their own configured endpoint.
A run on an organization-supplied credential now gets a larger answer budget and a longer timeout before it is treated as failed.
Agent runs are now presented as a timeline: the route taken, the prompts sent, the answer received, and the record kept.
The provider's own refusal message is now shown when a run fails, instead of a generic error.
The default conformance run now reviews three committed rows, instead of running against nothing.
A run recorded without a complete document still keeps a history entry, instead of being dropped.
Fixed a database connection pooler rejection by declaring snapshot isolation explicitly.
Rebuilt the home page hero and added screen-enter motion, keyboard sequences, and headline figures to the command center.
The collection console's panels and header now use the paper-toned surface the rest of the workspace uses, instead of a dark one.
Home page redesign, and production sign-in hardening
The public home page moved to the current visual system, and production sign-in got firmer: a deployment now checks the owner's sign-in readiness before serving production traffic, and a database connection fix removes a class of intermittent failures.
Redesigned the public home page to the current visual system.
Production now gates on the owner's sign-in readiness before serving traffic.
Owner sign-in counts stay readable through the response redactor instead of being stripped.
The database host now resolves IPv4-first for the Node.js Postgres client, fixing a class of intermittent connection failures.
PR #199 · PR #200 · PR #201 · PR #202
Quality push: dead code, type safety, and motion
A three-round internal quality pass removed thousands of lines of dead code, replaced loose types with proper ones, and moved pulsing skeleton loaders to a calmer directional shimmer across the site.
Removed roughly 3,480 lines of dead code and repaired CSS @apply usage across the workspace.
Replaced pulsing skeleton loaders with a directional shimmer treatment sitewide.
Replaced loose `any` types with proper types across eight files.
Fixed workspace ledger dividers that were not rendering.
PR #196 · PR #193 · PR #197 · PR #198 · PR #189
Refusal gate, hidden review queues, and engagement-bound review
The workspace redesign continued with a refusal gate that stops an agent on an ambiguous case instead of guessing, two review queues that existed but were not reachable from the interface, and a residual gate that binds a review to its specific engagement before it can close.
A refusal gate now governs five key workspace actions, stopping an agent to ask a human on an ambiguous case rather than proceeding.
Two review queues that existed but were not previously reachable — dual-control approval and evidence freshness — are now visible and actionable from the workspace.
Keyboard commands are now verb-labeled and land directly in a filtered queue, instead of a generic search.
A residual review gate now binds a workspace review to its specific engagement before that review can close.
PR #183 · PR #185 · PR #186 · PR #187
Standalone evidence collector, and running any requirements document
Two independent capabilities shipped on the same day: an installable evidence collector for environments that need collection to run outside the hosted workspace, and the ability to run any requirements document against any target for automated conformance checking.
Shipped a standalone, installable evidence collector for environments that need collection to run outside the hosted workspace.
Any requirements document can now be run against any target for automated conformance checking.