Version 3.10.0

3.10.0 is a cleanup release in two senses. It gives you the tools to clear out a feed that has grown to thousands of jobs, and it removes the places where JobIntel showed you a spinner or a cheerful "done" instead of saying what had actually happened.

Clear out a feed that has grown too large

• Filter by age. Jobs older than a week, two weeks, a month — pick a cutoff and see only what's left. Older postings are the ones most likely to be filled already.
• Delete exactly what your filters match. Set up any combination of filters, then delete that whole set in one action instead of removing jobs one at a time. Jobs you're tracking an application against are always kept, and JobIntel tells you how many it protected.
• Sort oldest-first. The feed has always sorted newest-first; "Oldest" is now a choice, which makes clearing out a backlog far quicker.
• Your place in the list is kept. Deleting from partway down the feed no longer throws you back to the top.
• Clearing thousands of jobs finishes. On very large feeds the delete used to be able to stop partway through without saying so. Now it works through the whole set, and if it genuinely can't finish in one go it says so plainly and picks up where it left off rather than reporting success.

Scores and searches that tell the truth

• "Not scored" instead of an endless spinner. A job whose scoring failed, or a posting we couldn't extract any skills from, used to sit on "Matching…" or "Analyzing…" indefinitely. It now shows an honest Not scored badge, so you can tell a genuine problem from one that's simply still working.
• A search that hits a limit says so. Scoring and search progress have internal ceilings for very large result sets. Reaching one used to look identical to finishing normally. It's now reported as what it is.
• Slow searches name the source that was slow. When a search takes a while, the delay is attributed to the specific job source responsible rather than left unexplained.

More jobs, fewer quietly lost

• Large-radius searches no longer get cut short. A wide-radius search that matched more than a couple of thousand jobs was being truncated. That ceiling is gone.
• Source filters match what they say. Filtering the feed by where a job came from missed some listings because the filter and the stored values didn't quite line up.
• StepStone and Totaljobs remote listings now read correctly. Company name and location were being pulled from the wrong place on those two boards, so their remote jobs arrived incomplete.

Mock interviews

• A dropped connection no longer ends your session. The interview stream now resumes where it left off — closing a laptop lid, switching networks, or a brief outage no longer costs you the interview. Sessions that reach the 30-minute cap are reported as capped rather than as finished.
• Interview language follows your account. Your language choice is remembered on your account, and if a job posting you're interviewing against is in a different language, JobIntel asks which you'd rather use instead of silently choosing.
• Czech postings are now recognised, so that question actually gets asked for them.
• Find the right job or application quickly. Interview setup now has searchable, paged pickers instead of one long list.
• The debrief status badge reads correctly when your generated feedback happens to contain the same words the badge looks for.

Coaching

• One coach at a time, enforced properly. Accepting an invitation while another was already being accepted could previously leave two live coaching relationships on one account. The database now prevents it outright.
• Invitations that can't mislead you. The invitation screen states the coach's capacity before you fill anything in, shows what accepting actually grants, and lets you decline explicitly rather than leaving the invitation hanging.
• Pausing, ending, or deleting closes access at once. Whichever way a coaching relationship stops, the coach's access to your data stops with it — one decision, applied everywhere.
• Notifications in both directions. Coaches and clients each get a feed of relationship events — invitations, acceptances, declines — with read/unread state.
• "Needs Attention" counts people, not alerts. A client with three alerts counted as three; it now counts as one client.
• Applications tell you when the job has gone. If a posting you're tracking leaves the feed, the application says so instead of linking to nothing.

Signing in

• A lockout clears when you prove who you are. Mistyping your password enough times locks sign-in briefly, as it should. But successfully resetting your password now clears that lockout immediately, instead of leaving you locked out with correct credentials in hand.

Smaller fixes

• Restored a missing space after bold and italic text in some app copy.
• Blog posts and shorts now produce a proper preview card with an image when shared on social media or in chat.

Under the hood

The feed is meaningfully faster: it had been fetching several kilobytes per job that were never displayed, and a measured review of the database's indexes dropped four that were dead weight, narrowed four more, and added two that were missing. We also added caching to skill extraction and credibility scoring on the paths that actually run them.

Alongside that: storage headroom for the production database, hang detection so a stalled background worker is noticed and restarted rather than sitting quietly, and a considerable amount of work on our own test and release gates — including fixing several that were reporting failures for validation that had in fact passed, and one that was reporting success it hadn't earned.