care-voice v0.2.0

This release makes care-voice installable without a Python toolchain and hardens how it is built and shipped. The check-in engine, extractors, and alert rules behave exactly as in 0.1.0.

Added

• Published on PyPI: pip install care-voice.
• Standalone executables for Linux x64 and arm64, macOS arm64 and x64, and Windows x64, attached to every GitHub Release.
• install.sh, a curl | sh installer that downloads the right executable for your system and checks it against SHA256SUMS.
• Multi-arch container image ghcr.io/superintelligenceco/care-voice (linux/amd64, linux/arm64), signed with cosign, and a release Compose file that runs it.
• SPDX SBOM, checksums, and build provenance attestations for every release file and the image.
• Documentation site at https://superintelligenceco.github.io/care-voice/ with a quickstart, concepts, CLI and Python API reference, FAQ, architecture diagrams, and decision records.
• Demo GIF recorded from a real terminal session.
• Property-based tests for the parsers and alert rules, a test that runs the README commands and compares their output, a benchmark gate against a committed baseline, a nightly full-suite workflow, and weekly mutation testing.
• OpenSSF Scorecard, dependency review, Trivy image scanning, actionlint, and Markdown link checking in CI.
• Makefile, pre-commit hooks, dev container for GitHub Codespaces, editor settings, CITATION.cff, and llms.txt.

Changed

• Releases start from a pushed v* tag; release-please is gone.

Fixed

• On Windows, pip install care-voice now pulls in tzdata, so time zones such as UTC resolve without a system time zone database.

Install

pip install care-voice==0.2.0
curl -fsSL https://raw.githubusercontent.com/superintelligenceco/care-voice/main/install.sh | CARE_VOICE_VERSION=v0.2.0 sh
docker run --rm -p 127.0.0.1:8080:8080 ghcr.io/superintelligenceco/care-voice:v0.2.0

Verify a download with gh attestation verify <file> -R superintelligenceco/care-voice.

care-voice v0.1.0

The first public release. care-voice runs a daily check-in conversation, stores the results locally, and alerts caregivers when replies look concerning. It is not a medical device and not an emergency service.

Added

• Conversation engine driven by a YAML check-in script, with yes_no, scale, day_of_week, and text questions, follow-ups, re-prompts for unclear answers, silence handling, and call retries.
• Bundled morning check-in script and an example evening script.
• Offline rule-based extractor that understands common ways of saying yes, no, numbers, feelings, and weekdays, and flags emergency words, falls, pain, and disorientation.
• Optional LLM extractor for any OpenAI-compatible chat completions endpoint. The rule-based extractor always runs underneath it as a safety net and as the fallback when the call fails.
• Risk-rules engine with twelve alert codes: EMERGENCY_WORDS, NO_ANSWER, CHECKIN_INCOMPLETE, FALL_REPORTED, MISSED_MEDS, MEDS_UNCONFIRMED, PAIN_REPORTED, POSSIBLE_CONFUSION, MOOD_DROP, LOW_MOOD, NOT_EATEN, and POOR_SLEEP. Every alert lists the replies that triggered it.
• SQLite history, used for mood baselines and for escalating consecutive missed calls.
• Console, webhook (with optional HMAC-SHA256 signatures), and SMTP email notifiers, each with its own minimum severity.
• care-voice command with simulate, history, alerts, validate, serve, and call subcommands.
• Caregiver dashboard and JSON API, with optional HTTP basic authentication.
• Experimental telephony: a VoiceAdapter interface and a Twilio Programmable Voice adapter with webhook signature checks and retries.
• Dockerfile and Docker Compose file for the dashboard.
• CI for lint, format, typecheck, and tests on Linux and macOS with Python 3.11 to 3.13, plus CodeQL, a Docker build check, and release automation.