v0.41.0 — "Data"
Data (Star Trek: The Next Generation, 1987, Gene Roddenberry) — the android who grows more human by patiently observing the people around him, then refining his own conduct to match, and the crew's tireless analyst when something needs explaining. This release teaches Curia both halves: it watches what you actually send and learns your voice from the edits you make, and it can now investigate its own behavior when you ask what happened.
This release is about Curia learning from what you actually do, and being able to explain itself when you ask.
The centerpiece is passive email observation. Curia now quietly watches your Sent folder, notices how you edited the drafts it wrote for you, and once a week makes a single pass to propose an updated writing-voice guide, which you approve or dismiss. Nothing is learned behind your back, and nothing rewrites your profile without a yes. The same loop closes to-dos: when an email you send finishes an open task, Curia completes the low-risk, obvious ones (with one-click undo) and asks you to confirm the rest. For messages it decided to leave for you, it now drafts what it would have sent and quietly measures that against your actual reply, so its autonomy can grow on evidence rather than guesswork. Learning items reach you directly now, not through a digest.
The other half is the new Diagnostics agent: an opt-in, read-only investigator you can ask "what happened, and why?" about Curia's own behavior. Point it at a duplicate notification, an email that never got processed, or a scheduler that fired twice, and it walks the audit log, cites the events it relied on, and hands back a grounded explanation. It can only read; it never acts.
Elsewhere: the built-in 8 AM digest gave way to an on-demand, chat-editable morning briefing. The web browser got sturdier, gaining an optional residential egress proxy, stable element references, a circuit-breaker on repeated failures, action batching for long forms, and warm sessions that survive across wakes. scheduler-update lets an agent resume, pause, or edit a job without the console. And a raft of durability fixes hardened the sent-observer, the scheduler, config merging, and the learning subsystem.
On the safety side, knowledge-graph writes from untrusted senders are now blocked, so a stranger can't seed Curia's memory with crafted "facts." On governance, this release adds a GOVERNANCE.md, a documented dependency policy, DCO sign-off enforcement, and cosign verification docs for signed release artifacts.
Heads up for operators: memory.scopes was removed (it was inert and never enforced) as a breaking change. See ADR-028 for deploy ordering.
The Sent folder read,
your voice gleaned from small edits.
It asks what happened.