This issue is the log for the unattended nightly maintenance pass on this repository.

A scheduled agent runs once a night and posts one comment per run here: what it landed, what it captured for a human decision, what it was blocked on, and what it skipped. It is the single place to read to find out what happened while nobody was watching.

What a pass does

1. Preflight — abort if either working tree is dirty or main's CI is red (fixing red CI preempts everything).
2. Triage each open issue with no agent:* label: reproduce against a real build, post a diagnosis naming the file and function, and say whether the cause is in this repo or upstream in data-explorer-core.
3. Fix at most one issue per run, and only one that clears a strict confidence gate — reproduced by a test it wrote, root cause named, fix local, correct behaviour unambiguous, full suite green with no existing test modified.
4. Qualify anything customer-visible in a real browser against the shipped dist/webview bundle, because the unit suite composites no alpha and lays nothing out.
5. Refresh dependencies and the data-explorer-core pin, as its own PR.
6. Changelog and version bump, then a vX.Y.Z tag.

What it deliberately does not do

• It never publishes to the Marketplace. That stays the maintainer's manual step; a run only says whether a publish looks warranted.
• It never decides a design, UX, wording, or contrast question. A 2am run has nobody to ask, so anything with two defensible answers is captured, not attempted: labelled agent:needs-discussion, with a comment giving the measurement and two or three concrete options with a recommendation. Capturing well is a successful outcome.
• It never pushes to main directly — every change lands via a PR with green CI.

Labels it uses

Comments posted by the agent end with an invisible marker, which is how a later run tells its own words from a human's. If you reply to an agent:blocked issue, the next run picks the reply up and re-triages with it.