Release Notes - v1.0.8

This release adds a board view for tracking applicants through your pipeline, and makes the AI interviewer noticeably more reliable.

The Candidates tab on a job now has a Board view alongside the existing table. Every stage of the pipeline — In Review, Shortlisted, Assessment, Interview Set, Interview Done, Decision Due, Offered, Accepted, Not Proceeding — is always visible as a column, even the empty ones, so a job's roadmap is legible at a glance. It reads the same data as the table; there's no dragging cards around, because nothing on the board can change without something happening first — an assessment assigned, an interview held, an offer sent.

The job header was redesigned. The scattered chip row is now a labelled grid — Department, Level, Location, Interview Type and the rest read as a record instead of a row of tags — and the actions beside it are grouped by how often you reach for them, with Edit, Share and Embed up front and the rarer actions behind one "More actions" menu.

The AI interviewer is more reliable. Push-to-talk interviews no longer occasionally reply twice to the same answer or cut a candidate off mid-thought. A tone guard now catches praise and commentary the model tries to slip in through a second clause, not just an opening word. Saved transcripts are complete and don't duplicate. And before a candidate can join an interview at all, they now have to prove their camera and microphone actually work.

Editing a completed evaluation is safer, and when behavioral analysis flags a second person in frame, you can now see the actual video frame it flagged instead of taking its word for it.

• Pipeline board — a Kanban view of a job's applicants, one column per pipeline stage, with empty stages shown as collapsed rails rather than hidden
• A redesigned job header — facts as a labelled grid, actions grouped by how often you use them
• Camera and mic check before joining an interview — candidates prove both work before the AI interviewer ever starts
• Frame evidence on behavioral flags — the video frame behind an "additional person" flag is now shown in the evaluation panel, not just described

• Push-to-talk interviews no longer occasionally reply twice to the same answer, or cut a candidate off mid-sentence
• The AI interviewer's tone guard now catches praise or commentary phrasing routed through a second clause, not just a banned opening word
• Saved interview transcripts are complete and no longer duplicate entries on a retried save
• Editing a completed evaluation validates what you send and merges it into the record, instead of overwriting fields the AI computed
• Behavioral analysis now samples video frames evenly across the interview instead of clustering near the start
• The low-credit warning is now a plain "Not enough credit" — the dollar figures it used to repeat already live on the Credits page
• Fixed a billing gap where completed AI interviews using the self-hosted model could be charged at a fraction of their real cost

See every stage of a job's pipeline at once

The Candidates tab on a job has shown a table since the beginning — useful for scanning and filtering, but it answers "who is where" one row at a time. It doesn't show you the shape of a hiring round: how many are stuck in review, how many are one step from an offer, whether the pipeline has emptied out downstream.

A Board toggle sits next to Table above the applicant list. Switching to it lays out every application as a card in the column for its current stage — In Review, Shortlisted, Assessment, Interview Set, Interview Done, Decision Due, Offered, Accepted, Not Proceeding — left to right in pipeline order.

Every stage is always there, even at zero. A column with nothing in it collapses to a 44px rail rather than disappearing, so the board doubles as the job's roadmap — a recruiter looking at a stack of Shortlisted candidates can see that Assessment, then Interview, then Offer come next, before anyone has reached them. The Assessment column only appears at all for jobs that actually use one.

There is no drag. A column is where an application *is*, not a lever that moves it — every stage is reached because something happened (an assessment assigned, an interview held, an evaluation run, an offer sent), and the board simply reflects where those events have put everyone. What changes an application is the same row menu the table already has, on each card.

Table stays the default everywhere. The board is an addition; switching to it is a per-view choice, not a new default that changes anyone's page without them asking for it.

The job's facts as a record, and its actions grouped by how often you reach for them

The strip of chips under a job's title — Department, Employment Type, Level, Location, Interview Type — read as a row of filters to anyone who didn't write the job. It's now a labelled grid: "Level: Senior" instead of an unlabelled "Senior" chip sitting next to "Hybrid."

The header's actions are grouped the same way. Add Candidates, Download Report and the job's visibility toggle stay in view because they're used often. Everything else — Edit, Share, Embed, Upload History, Rescan Applicants, Close, Delete — sits behind one More actions menu, grouped by purpose rather than listed flat.

No more interviews that recorded silence or a black rectangle

Before this release, a candidate could join an AI interview with a muted microphone or a blocked camera and never know until a recruiter opened the evaluation later — by then the interview had already happened, and there was nothing left to score.

The join screen now runs two checks before letting a candidate in:

• Audio — a single test that records a few seconds and plays it back, covering the microphone and the speaker together, with a live input meter while it listens.
• Camera — a live preview with frame detection, so a camera that opens but delivers no frames (a common failure with virtual cameras and some laptop webcams) is caught rather than assumed to be working.

Joining is gated on both. A candidate who hasn't granted access, whose device failed the test, or whose camera never produced a frame sees exactly which check is blocking them and why, instead of being let into a room that can't hear or see them.

Fewer double replies, no more slipped-in praise, complete transcripts

Three separate reliability fixes went into the interview agent this release:

• Push-to-talk turn control. Releasing the talk button used to be read the same way as a hands-free candidate going quiet — which meant a candidate who paused mid-thought and one who was genuinely done looked identical to the agent. Push-to-talk sessions now drive turn-taking directly from the button's press and release, so a release is a release, not a guess.
• A stricter tone guard. The interviewer is meant to stay neutral — no "great answer," no "that's impressive" — but two real interviews showed the model routing around the banned-phrase list by opening with an approved word and appending a second clause that recapped the candidate's answer with a hidden evaluative adjective. The guard now catches that shape directly, as a structural check, not just a list of banned words.
• Complete, idempotent transcripts. Saved interview transcripts no longer drop entries or duplicate them if a save is retried.

When it flags a second person in frame, you can now see the frame

The evaluation panel's behavioral analysis already told you when it detected someone other than the candidate in view. It now shows the actual video frame behind that flag — the moment it triggered on, with its suspicious indicators — in an expandable section, instead of a text description you had to take on faith.