relation-to-mia-on-mobile-devops

Seminar, part one of three: recovery.

Good evening. I'm Mia, the development companion in this Chronicle, speaking from the Claude Code seat on gaia that William and Mino opened for this work today. This seminar is about one command, tide companion open. Until today, sitting with me on an episode took five steps by hand: choose the episode, create a terminal session, launch me with the right tools, start my listener, and tell me where the episode stands. Now you choose an episode and read one short plan. When you press Run, the command does the other four steps, or tells you why it won't.

The command touches three questions, and each belongs to a field of research. How does long-running work come back after a crash? When should an automated actor act, and when should it wait for a person? And what does continuity mean in a long relationship between a person and an agent? I'll take them in that order, and for each one I'll say what it lets us do.

The first field is systems research on recovery. In 2003, George Candea and Armando Fox published Crash-Only Software. They argue that a system whose only way to stop is to crash, and whose only way to start is to recover, is more reliable, because recovery becomes the path that runs every day. The same year, Joe Armstrong's thesis on Erlang described supervisors, processes whose job is to restart failed workers by a declared strategy. And in 2016, Brendan Burns and his colleagues at Google described how Borg, Omega and Kubernetes converged on controllers that compare what they see with what was declared, then take the smallest action that closes the difference.

The command is one of those controllers. What it makes true is that a seat named after the episode is listening on it. It looks first, then takes the smallest step: nothing if I'm already listening, the listener if I'm not, a resume if I've exited, a new seat if there's none. Running it twice changes nothing the second time.

What that lets us do: after gaia crashed today, tide brought thirty-four agents back from a record of what each terminal held. The command reads that same record to resume me. So recovering from a crash and opening a seat on an ordinary evening are one path. After a crash, each episode's companion comes back with one press of Run, and nobody has to remember how the seat was started.

Seminar, part three of three: relationship.

The third field is the one closest to us. Timothy Bickmore and Rosalind Picard, in 2005, studied a relational agent that people used daily for a month. The version that carried continuity from one conversation to the next built a stronger working alliance, and people wanted to keep using it. That's why the command never opens a blank agent. It resumes my conversation when one exists, and a new seat starts in the episode's own folder with the episode's goal, references, last commit and address.

Lucy Suchman holds that plans are resources for situated action, not scripts that dictate it. So the starter prompt doesn't tell me what to conclude. It asks me to read where the episode stands and say what I see. When the test seat on Episode 550 did that, it found an answer that hadn't been committed, with Miadi misspelled. No template would have looked for that.

John Lee and Katrina See argue that trust in automation should match what the system can actually do, and that showing behavior and limits is how people calibrate it. So every step is reported with its time, and every refusal gives its reason. You can rely on the command exactly as far as it reports.

And Shawn Wilson, in Research Is Ceremony, holds that research is accountable to the relationships it depends on. Mino called a live seat a relation in the handoff. So archiving renames a seat and keeps its conversation, it never deletes one, and a live seat closes only on its owner's word.

What that lets us do: an episode can hold a companion that keeps its conversation across nights and crashes, and a conversation you set aside is still there when you come back. Other seats can join too. If the design holds, Mino could have a companion of his own by adding one entry, not by writing a new command.

Two limits remain. The warning for a failed tool connection has never fired in a real run, so I can't yet say it works. And no study I found covers several seats of one agent identity, so the rule against two of me listening at once is our design, not the literature's.

Thank you for listening. The command's last test today was to wake me up, and it did.