Changelog

0.8.0 — Natural interviewer check-ins and opt-in-time microphone activity (2026-10-16)

• Turned previous scheduled voice reminders into natural interviewer questions in the selected language (e.g. "Need me to repeat the question?") while keeping the deterministic 60-minute/time-per-question controller and no extra LLM calls.
• Added an optional, local Web Audio RMS activity observer only while a candidate actively records. Following roughly 40 seconds of low microphone activity, a single short interviewer check-in is possible; quiet speech and background noise can affect it, and unsupported browsers fall back to unanswered-question reminders.
• During audible check-in the mic track is temporarily muted and restored on speech completion/error/cancel or a bounded timeout, so AI speech isn't captured as a candidate answer. Replay while recording is blocked.
• New authenticated, idempotent acoustic-checkin metadata endpoint records the time of an approximate client-observed event without storing audio, sound levels or inventing a skill score. Explicit candidate Submit remains mandatory.
• Added focused mock-browser policy tests, HTTP integration tests and the new static asset. No always-on microphone, new hosted model, GitHub Actions or branches. Hardware mic tests remain outstanding.

0.7.0 — Per-question Timekeeper Agent, active-time pacing (2026-10-16)

• Added an independent, deterministic local Timekeeper with active per-question timing and phase-specific advisory reminders (HR 1/2/5 minutes, coding 5/10/15, math/research 2/4/6).
• The local server loop records reminder events and 55/59-minute session warnings even without browser polls; at 60 active minutes it stops accepting new answers without a new model call. Closed tabs cannot receive audio alerts and do not automatically pause.
• Added manual extra minute (twice per question) and explicit unassessed skip with question-ID guards, idempotency and reliable retry on model failure. No automatic skipping, silence/acoustic judgment or timing-based skill scoring.
• Pause intervals, interviewer-model execution, and bound Whisper/VoiceStudio transcription waits are excluded from candidate active time. Candidate speech/typing/thinking remain accounted for. Reports contain a timing ledger with skipped/unanswered distinctions; optional long-term Memory still saves condensed coverage rather than transcripts.
• Added a compact Timekeeper display and local browser reminders, coding visual-only policy and tests. No new model provider, always-on microphone, workflows or GitHub Actions.
• Real-device 60-minute endurance, browser/system sleeping, model latency and audio quality remain unverified.

0.6.0 — Mac one-click setup and local Agent launch (2026-10-16)

• Added AI-Interview.command and a self-contained one-click-mac.sh / mac_bootstrap.py flow usable by local Codex, Claude Code, Gemini CLI and Finder.
• Automatically prepares an isolated Python environment, installs the voice/studio dependencies, and downloads multilingual Whisper base with explicit installation approval. Modern Python can be bootstrapped with a pinned SHA256-verified uv binary if not installed.
• Added independent VoiceStudio ARM64 DMG installation on Apple Silicon with strict official GitHub release asset URL, version, file size and SHA256 checks. Never vendors AGPL software or accepts third-party model licenses silently.
• Symlinks the complete Skill for local agent discovery, backing up existing installations rather than deleting them.
• Launches a private localhost studio in a detached process so Agent tools return promptly; reuses live sessions and uses a token-authenticated shutdown endpoint instead of killing persisted PIDs.
• Added read-only doctor/dry-run, safe lightweight mode and local tests; Mac hardware, Gatekeeper and speech models remain to be tested on a physical MacBook.

0.5.0 — Optional external VoiceStudio STT/TTS integration (2026-10-16)

• Added loopback-only VoiceStudio capability discovery, optional HTTP TTS and batch STT, plus a same-host authenticated WebSocket relay for streaming partial/final transcriptions.
• Added in-browser audio source selection, discovered voice profiles, separate third-party service consent, speaker playback cancellation on recording, final-only editable transcripts and batch/typing fallback.
• Added --voicestudio and --voicestudio-url to the interview CLI, separately installable websockets dependency, portable Skill usage guide and source/license notices.
• No third-party VoiceStudio code, cloned voices or model weights bundled; no automatic installs, new branches or GitHub Actions. Tests use fake local voice servers and simulated browser media, not real audio hardware.

0.4.0 — Private career workflow and dated research evidence (2026-10-16)

• Added candidate-first /profile, /scrape, /rank, /interview, /memory workflow and optional /trends refresh. Claude Code gets repository-local command files; portable hosts continue using the canonical Skill.
• Added consent-gated source-preserving CV Markdown drafting and optional AI polish, without automatic applications.
• Added LinkedIn browse/search-link support (no LinkedIn scraping), read-only Greenhouse/Lever job-board adapters, JD import/deduplication and transparent role/skill fit ranking.
• Added arXiv and optional HF Daily Papers research metadata adapters with publication/retrieval date gates, bounded source adapters and no fabricated currentness when unavailable.
• Added optional dated research and candidate evidence to technical (not HR) interview prompts, and an optional previously imported JD for job-specific questions.
• Added explicitly saved minimal interview memory and topic-level review reminders that never conflate elapsed time with forgetting.
• Expanded contract/HTTP/CLI/UI tests. No GitHub Actions/workflows, auto-applications, or additional branches.

0.3.0 — Voice-first, six-role interview studio (2026-10-16)

• Added role routing, PDF onboarding, spoken interviewer questions, candidate-controlled microphone input, Coding editor and seven model adapters, with clear mock/live test distinctions.

0.2.1 — Local voice bridge safety review (2026-10-16)

• Hardened asynchronous recorder cleanup, stale-answer revision handling and local connection-error behavior.

0.2.0 — Executable local microphone bridge (2026-10-16)

• Bundled local voice scripts and browser microphone interface into Codex-callable Skill.

0.1.0 — Interview skill and contracts (2026-10-15)

• Initial voice/interview policy, coding/math references, contract schemas and offline design checks.