Release Notes - v1.0.1

The first feature release after v1.0.0 sharpens three areas: the AI Assistant, security, and AI model control. The chat assistant can now filter candidates by years of experience and generate a shortlist of a job's top applicants in a guided flow — backed by a new experience field extracted from every CV so the filter works on any model. Chat itself got a major reliability and UX pass: history search, rename, and delete, a stall-recovery watchdog so a stuck model response no longer hangs forever, per-chat model persistence, and plain-language model labels. On the platform side, this release ships two batches of security hardening — multiple IDOR fixes, stored-XSS patches on offer letters, AI-SQL tenant isolation, webhook signature verification, and a production guard on a test endpoint — plus per-company S3 storage isolation. Recruiters also get per-module AI model selection (choose a different model for CV review, job posting, scheduling, behavioural analysis, and more — including a Taqat-hosted Qwen model), a visible scoring rubric, selective CV download as a ZIP, background suggested-candidate matching with live progress, deferred rejection emails until a job closes, and a reaper that auto-recovers and retries CV uploads that were silently killed mid-process. Outgoing emails are now branded Truepick.

• AI Assistant: experience filter + shortlist generation — the assistant can now honor "exclude candidates with more than N years of experience" and generate a ranked shortlist of a job's applicants, persisting the top-N as shortlisted via a two-step preview/save confirmation
• Per-module AI model selection — pick a different model per operation (chat, CV review, interview, assessment, jobs), including a Taqat-hosted Qwen custom model; all modules default to GPT-5 Mini
• Chat history management — search, rename, and delete past conversations
• Stuck-upload auto-recovery & retry — a reaper fails CV jobs stuck with no progress for >15 minutes and a one-click Retry re-queues only the unfinished files; no more permanently spinning uploads
• Visible scoring rubric — recruiters can see how an interview score is built (scale, dimension weights, per-question vs holistic rollup), shown additively so older evaluations render unchanged
• Selective CV download (ZIP export) — select candidates and download their CVs as a single streamed ZIP
• Background suggested-candidate matching — matching runs in the background with live activity progress instead of blocking
• Deferred rejection emails — candidate rejection emails are held and sent only when the job is closed
• Assessment anti-cheat — paste is blocked on text answers and question text can no longer be copied
• Redesigned Import Candidates dialog — location/skills filters, already-applied candidates excluded, duplicate flags, and a confirmation step

• Chat stall recovery — a client-side timeout and recovery prompt surface and retry a stuck AI response instead of hanging indefinitely
• Per-chat model persistence — the selected model is remembered across chats via a secure cookie, with plain-language labels and i18n loading strings
• Chat input character counter — live counter with a 2,000-character client-side limit
• Multi-tool replies summarize — the agent step limit was raised so multi-tool chat replies return a final summary instead of cutting off
• Recruiter-facing chat cleanup — token-usage telemetry hidden from recruiters; assistant tool links use relative URLs; date-range candidate analytics and open-job listing added
• AI interview settings on job detail — the configured interview settings are now displayed on the job detail page
• Consistent back button — a shared back button across recruiter detail pages
• Register/login validation — name and email validation errors surface inline, a TLD is required on email, and Confirm Password has a show/hide toggle; login honors the original `redirectUrl`
• Interview integrity score — now reflects behavioral concerns instead of staying optimistic

• IDOR fixes — company ownership enforced on candidate update, upload history, offer-letter download, and the S3 file proxy
• Offer-letter stored XSS — user fields escaped in offer-letter HTML (two sinks closed)
• AI SQL tenant isolation — the AI SQL tool is scoped to the caller's company, with a comma-join bypass closed
• Webhook signature verification — Fireflies and Composio webhooks now verify signatures before processing
• Private job endpoints — auth required, deny-by-default for private `/api/jobs` routes
• Production guard — the test Inngest-trigger endpoint is blocked in production
• Auth-bypass / path-traversal batch — an OTP auth-bypass, chat path-traversal, and additional IDOR holes closed
• Per-company S3 isolation — all company-owned uploads scoped to a `companies/{companyId}/...` namespace

• Bulk CV ZIP streaming — large CV ZIP exports are streamed to prevent worker out-of-memory
• One active application per candidate per job — enforced at the database level
• Applicant counts reconciled — status counts now sum to the total; `review_done` is included in the Interviewed count
• Candidate-search render loop fixed — eliminated an infinite re-render on candidate search
• Recording access hardened — presigned URLs replace public ACLs; egress auto-detects LiveKit Cloud vs self-hosted credentials; video is downloaded to a temp file before FFmpeg
• Incomplete interviews skipped — interviews with no answered questions are no longer scored, and behavioural writes no longer clobber the evaluation
• Arabic CV parsing — Arabic Presentation Forms are accepted by the PDF text-meaningfulness check
• Truepick email branding — outgoing emails rebranded from "Taqat Recruitment" to Truepick, with the logo inlined via CID so it stops 404ing in clients

Ask for "under 15 years" and get a real shortlist

The chat assistant previously could not honor a request like *"exclude candidates with more than N years of experience"* on any model — candidates had no numeric experience field, and the scoring tools reported job-count as if it were years of experience.

What changed:

The assistant works the request agentically — discovering the job, listing applicants, and ranking them — then persists the chosen candidates as shortlisted once you confirm.

Choose the right model for each job

AI Settings now lets a workspace pick a different model for each AI operation instead of using one model everywhere.

Settings → AI Settings: pick one model per module — every AI tool call within that module uses the selected model

A faster, calmer assistant

The model selector now shows plain-language names and descriptions grouped by provider

Chat History — searchable, grouped by date, with rename and delete on every conversation

When a model response stalls, the assistant no longer hangs — it surfaces a recovery prompt so you can retry in place.

If a response takes too long, a "taking longer than expected — Try again" control appears instead of an indefinite spinner

No more permanently spinning uploads

When a background CV-processing function was silently killed mid-step (server redeploy, out-of-memory, AI-credit exhaustion), its failure handler never ran — leaving the upload stuck in "processing" forever with no way to cancel or retry.

• Reaper (every 3 minutes) fails jobs stuck in "processing" with no progress for more than 15 minutes, cleans up open steps and bulk-upload items, and broadcasts the change over SSE so the UI leaves the spinner
• Unified manual retry — a tenant-scoped retry endpoint, capped at 3 attempts via a race-safe conditional update; bulk re-queues only the non-successful files under the same tracking row
• Retry surfaced on failed batches in the upload-center widget (Cancel already existed)

See how the score is built

The candidate evaluation view now exposes the AI scoring methodology so recruiters can understand and defend a score:

• A collapsible "How We Score" panel — scoring scale, why the interview dominates, what the other dimensions add, and why a random question subset is intentional
• A rubric breakdown — grade scale legend, dimension × weight contribution table, and the 75% per-question vs 25% holistic interview rollup

The rubric is wired in additively — older evaluations without the new fields render exactly as before, and a per-question rubric is hidden for legacy evaluations that lack the asked-question flag.

Cleaner intake, accurate counts

The redesigned Import Candidates dialog adds location and skills filters, excludes candidates already on the job, flags duplicates, and confirms before importing.

Import Candidates — search plus location and skills filters, with select-all and a clear import action

Applicant status counts now reconcile to the total, `review_done` is counted as interviewed, and the applicants toolbar carries selective Download CVs (ZIP), Compare, and Refresh.

Applicants overview — status cards that sum to the total, interview-score grades, and Download CVs / Compare / Refresh actions

Other changes: one active application per candidate per job (DB-enforced), serial numbers unique per company, background suggested-candidate matching with live progress, and assessment anti-cheat (paste blocked on text answers, question text non-copyable).

Two batches of fixes