Wino.Mail.AI.OpenAI 0.1.0-alpha.37

Wino Mail AI

Mail-intelligence infrastructure shared by Wino Mail and Wino Mail API.

The solution separates four concerns:

• provider-neutral mail, artifact and metering contracts;
• deterministic HTML sanitization, visible-text extraction and content hashing;
• OpenAI structured generation, the current implementation of the summarization stage;
• the TypeSafe Jev adapter, the current implementation of the classification stage.

Processing is classification first, summarization second. Classification answers a fixed set of typed questions for every selected message — one Noul per smart label, one Noul for briefing inclusion, one Score for urgency and one Choice for the single action the recipient would take. Summarization writes a headline and a one-line summary, and only for the messages classification included in the briefing.

Urgency is a Score rather than a Choice because its levels are ordered: the answer is a position that can land between two levels, so the cut points stay tunable in code. Every raw probability behind a classification is kept in MailClassificationSignals, which is what lets a threshold be retuned by re-reading stored results instead of reclassifying a mailbox.

The two stages are named for what they do, not for the models behind them. TypeSafe's Jev implements classification today and OpenAI implements summarization; either can be replaced without changing MailClassificationArtifact or MailSummaryArtifact.

Jev returns typed, calibrated answers and never generates free-form text. It also reads dates as text rather than ordered quantities and cannot count or do arithmetic, so no caller may ask it for dates, totals or multi-hop conclusions. Keep that work in application code: JevMailClassifier projects a message's age as words such as "4 days ago" precisely because the ordering is done here rather than asked for. Its accuracy degrades when unrelated context is present, so JevMailClassifier projects a deliberately tight state.

Question wording and thresholds are supplied by the caller and belong in server configuration. Calibrating a threshold must never require a contract change.

The packages are published publicly on nuget.org. The workflow requires the repository NUGET_API_KEY secret and verifies every package after publication.

docs/architecture.md describes the pipeline these packages compose, including the identity and freshness rules and the configuration boundary.

Build

dotnet restore Wino.Mail.AI.slnx
dotnet build Wino.Mail.AI.slnx --no-restore
dotnet test Wino.Mail.AI.slnx --no-build

Tests that call a paid provider are skipped unless WINO_AI_LIVE_TESTS=1 and WINO_AI_OPENAI_TEST_KEY are both set, so a normal run costs nothing.

No NuGet credentials belong in source control. Configure NUGET_API_KEY as a GitHub Actions repository secret before dispatching the public publication workflow.