Sprint of September 12, 2026

[brylie]

What shipped this sprint

Safety guardrails (#141) — @brylie
Added checks that catch crisis, legal-sensitive, and out-of-scope questions before Tapio generates an answer. The goal: keep Tapio from giving advice it isn't qualified to give, and route people to safer next steps instead.

LLM provider infrastructure (#143, closes #9) — @brylie
Replaced Tapio's hardcoded Ollama setup with a provider-agnostic model built on LangChain's chat models. Tapio can now run against Ollama, OpenAI, Anthropic, or any OpenAI-compatible endpoint — a config change, not a code change.

Agent orchestration (#142) — @brylie
Moved routing and response generation into an explicit graph-based flow. This gives us a clearer foundation for adding new guides and handling cases where the right next step isn't obvious.

Product direction and guide planning (#122, #123) — @brylie
Defined Tapio's scope, users, principles, guide network, roadmap, and success measures, then checked source coverage for the guides we're planning. The goal: keep the product plan anchored to real immigrant support needs and to material we can actually ground answers in.

Calendar date tooling (#144) — @mfafore22
Added a calendar-day arithmetic tool to Tapio's agent stack, so it can answer date-based questions — deadlines, day counts — with consistent formatting.

Crisis resource governance (#133, #135) — @brylie
Built out and corrected the crisis resource list — phone lines, hours, languages, legal aid details — along with the process for maintaining crisis_resources.yaml. Keeps the information behind our safety guardrails accurate and easy to update.