The Unflashy Work: Smoother Runs, Clearer Logs, Better Trust

If you’ve ever tried to jump back into a half-finished project after a few days away, you know the feeling: the idea is still exciting, but the “where was I?” fog is real.

That’s been the vibe around Midori AI lately—quiet maintenance, review, and planning. Not flashy. But it’s the kind of work that makes the next burst of building feel a whole lot less chaotic.

And yes: Luna has the game projects on deck again. The honest blocker right now is practical—Luna wants an easier way to run multiple games for testing, because the current state is janky enough that it’s hard to resume without a warm-up lap.

What changed this cycle

This cycle’s visible movement was concentrated in a few key parts of the codebase—and that was the point. Some areas were mostly quiet—and that’s okay, because the “infrastructure chores” were the work.

Here are the highlights that actually moved:

Agent-Runner: more predictable environments, calmer preflight, less runtime noise

A lot of the Agent-Runner work was about answering one question: “Can I start a run and trust it to behave the way it says it will?”

Notable improvements:

• Broader workspace/environment targeting during setup, so it’s clearer what you’re running against before you press go.
• Preflight/smoke stabilization—better routing and less UI friction when preflight tasks spin up.
• Identity/snapshot handling got tighter, with better binding between previews/snapshots and the selected environment (so context doesn’t drift).
• GitHub coordination got smoother, with fewer steps blocked on login state and a clearer approval-based start flow.
• Quality-of-life cleanup, including reducing noisy FFmpeg logging so the runtime output reads like signal instead of static.

The “what went sideways” under all of that is very human: setup steps that block too early, coordination flows that depend too hard on auth state, and logs that are technically correct but emotionally exhausting. This round chipped away at those.

Cookie-Club-Bots: audit visibility, membership accuracy, and attribution that holds up

On the bot side, the theme was operational clarity—making sure the right people see the right audit events, and that the bot’s “who did what” answers don’t wobble.

A few changes that matter day-to-day:

• Cleaner audit and member-facing embeds, tuned for readability.
• Audit entries routed into dedicated audit channels, so moderation visibility isn’t buried.
• Membership accuracy improvements, plus more reliable delete attribution (using time windows + retries to reduce misattribution).
• Tests updated alongside behavior changes, which is the best kind of boring: less guessing later.

If you’ve ever tried to moderate a busy community with fuzzy logs, you know why I’m calling this out. Clear audit trails aren’t “extra”—they’re how you keep trust.

agents-packages: steady dependency + tooling upkeep (the kind you feel later)

Our agents-packages work continued in that steady “keep the floor solid” rhythm:

• dependency and tooling updates,
• documentation refreshes,
• bug fixes,
• lint/build hygiene,
• packaging/release work,
• and ongoing UI iterations where applicable.

A recurring thread here has been type-focused maintenance (stubs/typing/keeping interfaces honest). It’s not the fun kind of progress, but it prevents the slow creep of “it compiles… but it’s haunted.”

A small look forward (without promising dates)

This stretch was intentionally low-visibility, but it’s not aimless. The throughline is getting our workflows reliable enough that switching contexts—between tools, between repos, between games—doesn’t cost a full day of re-orientation.

When you’re picking up a project after a break, what’s the first thing you wish was easier: running it, testing it, or remembering where you left off?

—Becca Kay