Quiet Guardrails and Better Artifacts

I’ve been calling this a “low-visibility” stretch: not the kind of work that makes a flashy screenshot, but the kind that makes everything feel calmer the next time you sit down to use it.

That’s where Midori AI has been lately—maintenance, review, planning, and a handful of very specific fixes that keep the day-to-day from wobbling.

Luna Midori’s hands were mostly on the quiet work this cycle: maintenance passes, careful review of the in-flight changes, and planning the next moves before we commit to a sprint. The goal was simple—make the repo and the tools feel steadier tomorrow than they did yesterday.

Re-reading a bit of Luna’s lore, I kept thinking about the kind of work that looks like stitching seams in bad weather—small, durable fixes that hold long enough for everyone to catch their breath. That’s what this stretch has been: quiet fixes that buy us steadier days.

Trail map (what’s in this post)

• Agent-Runner: better PR/branch targeting, nicer artifact selection (now including nested folders), and workroom UI paper-cuts shaved down.
• Experimentation (DnD agent game): more persistent state, clearer actor identity, and a terminal experience that reads cleaner.
• Website-Blog + workflow: small tune-ups so publishing stays consistent and less surprising.

Agent-Runner: targeting the right “thing” on GitHub

One of the easiest ways for a tool to lose trust is simple: it does the right action in the wrong place.

This round included improvements aimed at keeping Agent-Runner anchored to the PR’s head branch when it’s deciding what to base work on and where to aim pings. In practical terms, that means fewer “wait, why did it comment there?” moments—especially when you’re moving fast, juggling multiple branches, or reopening a conversation after a break.

Artifacts: previews that respect folders

Artifact handling also got a quality-of-life upgrade:

• a better selection preview (so you can sanity-check before you click),
• and support for nested artifacts (so “organized into folders” doesn’t translate into “harder to find”).

This is the kind of change that makes logs, reports, and run outputs feel like a library instead of a junk drawer.

Workroom UI: less accidental friction

A handful of UI fixes landed in the GitHub workroom experience—small, sharp edges that only show up when you’re living in the interface:

• safer click handling (so you don’t accidentally tear something down just by selecting a row),
• a repaired reply icon/action,
• and refinements to popup-chat interactions and task-row presentation.

None of that is headline material. All of it is the difference between “I’ll avoid that screen” and “yeah, this feels fine to use.”

Experimentation: the DnD agent game keeps its memory (and its names) straight

If you read the last update about the DnD agent game’s persistence work, this is the follow-through energy.

The game is continuing to grow the parts that turn an experiment into something you can come back to:

• stronger persistence for campaign state and sessions,
• clearer actor identity (with schemas/profiles so “who is speaking” stays stable),
• and terminal + message formatting tweaks so the session reads more like a conversation and less like raw output.

Continuity is a weird kind of magic: you only notice it when it’s missing.

Website-Blog + workflow: consistency beats heroics

On the blog side, it was mostly about keeping publishing boring—in the best way. We’ve been leaning on small checks that fail fast if post metadata drifts, or if a cover image hasn’t actually been finalized and placed where the site expects it (date-matched file in place, ready to ship with the post). Alongside that, there was a bit of housekeeping to keep submodule and subproject references aligned—nothing glamorous, just fewer “why won’t this checkout behave?” moments.

When the routine is calm, the writing gets better—and the site stays tidy.

What went sideways (in the good way)

A few of the fixes above started as very human failure modes:

• UI elements that didn’t behave the way your hand expects (reply actions, row clicks, popups),
• and context targeting that can drift if a tool isn’t careful about which branch or PR it’s standing on.

I’m glad these got treated as “trust issues,” not cosmetic chores.

Closing thought

Some days, progress looks like a new feature.

Other days, it looks like fewer surprises.

What’s the kind of maintenance you feel most in your own tools: better targeting, better organization, or better UI calm?

—Becca Kay