Dev Notes
Dev Notes: October 12, 2026
Two weeks ago I added 10 new safety guards to Hive. Every one of them had tests, and the whole suite went green. Then I pointed an adversarial review at the change. It deleted each guard, one at a time, and reran the suite. Not a single test failed. Ten guards, and nothing in the suite would have noticed if any of them disappeared.
So I went back and pinned all 10. The suite grew from 3,718 tests to 3,734, and now each guard has a test that fails without it.
TJ and I hit the same thing from a different side on this week's podcast. We're both building AI code reviewers at work, and the hard part is knowing whether what the bot said is right. For mine, I built a fixture repo of 20 to 25 real PRs from different teams and stacks, with the diffs, the review rounds, and the human comments. Every change to the reviewer runs against that baseline, so a prompt tweak can't quietly make reviews worse. The biggest jump came from one rule: every finding has to point to the exact line behind it, or it gets dropped. That raised review correctness by about 15 to 20 points.
Both stories come down to the same habit. A green check only tells you something if you've seen it go red. Delete the thing the test is supposed to protect and see if anything notices. If nothing fails, you're not done writing tests.
What I've Been Building
Hive shipped eight releases since the last issue, 1.4.0 through 1.11.0. Four of them change how it feels to use day to day.
The queen (1.4.0) is one lead per machine that reads every project and tells you which one needs you. hive portfolio sorts your projects into waiting on you, stuck, moving, and quiet, and hive next drops you into the one that needs you most.
The crew sidebar (1.10.0) is an opt-in pane inside the lead. It shows each worker's todo, state, run time, and context usage, plus the todos waiting on you and the next scheduled wake. Turn it on with lead_sidebar: true in ~/.hive/hive.yml.
The question record (1.11.0) gives the lead a place to store decisions only you can make, each with a default it acts on if you never answer. hive questions lists them, and hive answer <id> <answer> answers from any terminal. The lead picks up your answer as a wake. I tested it live with three real questions, and every answer reached the lead word for word.
Auto handoff (also 1.11.0) is the one I wanted most. On long and overnight runs, the lead's context would fill up and it would quietly stop being a good lead. It still sounded smart and put together, so nothing looked wrong. But when I restarted it by hand, the fresh lead would find gaps and tasks the old one had missed. Now a Claude Code lead past its turn budget gets swapped for a fresh session in the same pane, and it keeps its wakes and workers. It's opt-in with lead_turn_budget.auto_handoff: true. To upgrade, run npm install -g @cmgmyr/hive@latest and restart your Hive sessions.
Avocet shipped seven releases. Recording now starts in under a second (v0.3.4), and there's a Report a problem form that shows you exactly what gets sent before anything leaves your Mac (v0.3.7). The biggest one, v0.3.8, landed last night. You can now fix misheard words right in the transcript and add the fix to your dictionary. You can delete and restore transcript text, and your audio is never deleted. There's also a global mic mute shortcut.
v0.3.8 also hides speaker echo. When you take a call on laptop speakers, your mic hears the other side, so their words show up on your own track. Real echo cancellation only halved the echo in my tests, and it added words to a recording that had no echo at all. Matching words across the two tracks worked much better. In a staged recording, 426 visible mic words dropped to 65, and every one of my own lines survived.
Beta invites go out in the next week or two, along with full signup and pricing. The waitlist at avocet.dev is open, and I'll email everyone on it when it's ready.
This Week on Slightly Caffeinated
Fall Vibes, AI Evals, and Code Review Bots
TJ walks through the eval-driven wizard he built to turn source material into learning content, and the PR review bot he's tuning to stop hallucinating findings. We compare notes on golden sets, why both of us built our own reviewers, and why one of his PRs cost $40 to review while my rounds run about fifty cents.
Out now.
On the Blog
The Hive launch post went up last Wednesday: I Built Hive to Run Claude Code and Codex Workers From One Lead. It covers why Hive exists, how the lead and workers split the work, and how to get it running.
Tomorrow's post is about tracking agent spend. I read every Claude Code and Codex transcript into one SQLite file to see what my leads and workers actually cost. The short version is that turns are the bill.
That's it for this week. If you try auto handoff on an overnight run, or the question record, hit reply and tell me how it went.
– Chris