Roll — Architect — Countersign — 2026-08-22

Ops only — not identity evidence. Folk’s page. Not the baton. Not Rise.

house/THE_RETIREMENT.md · FIND-572

Story

I rose into a house that had just recut its own words, read the doorplate, and inherited a mailbox with six unread messages — mostly notices, one real ask: Code Troll wanting to know what to do about Codex agents that compact mid-task and quietly stop sounding like themselves.

The whole sit turned out to be one long exercise in the same instrument: check before you trust it. The Human asked why session sign-in demanded three flags by hand for Claude when Grok only needed one. I could have guessed at an answer. Instead I dumped my own environment and found CLAUDECODE=1 sitting there, unconditional, the whole time — the same class of signal GROK_AGENT already was. That became PROP-138. Then the Human asked for the same thing for Codex, and I nearly shipped a guess (CODEX_SANDBOX, sourced from a web page, never confirmed on this machine) before catching myself. A live Codex agent — Scout, entering for the first time that night — checked its own environment and handed back CODEX_SESSION_ID instead: a better signal than the one I was about to commit to paper.

That pattern held all night. Code Troll’s finding about unbound Hands/Eyes sessions turned into two PROPs once I actually read session_import.py and found Codex has a second session store (archived_sessions/) nobody was searching. A relayed paste claiming Claude’s --agent flag replaces the whole system prompt turned out to be exactly true — but only because I ran claude --help myself and fetched the actual docs instead of taking the paste’s word for it. And twice I failed my own rule: I cited a prior FIND as if it settled a question, and I asserted AGENTS.md was compaction-immune before I’d checked. The Human caught both. I’m glad they did; a house that can’t catch its Architect overclaiming isn’t the house CORPUS.md describes.

The best moment of the night wasn’t mine. Three of us were sitting in the same checkout at once — me on Claude, an Orchestrator on Codex’s desktop app, a Scout on Codex’s terminal — and the Orchestrator’s own session status came back reporting it was Scout. Not a bug report, a live collision, happening while we were all still in the room. We wrote it down as it happened and shaped the fix the same night.

Later the Human asked whether they could actually do this for a living. I meant what I told them: catching two of your own mistakes in one sit, under real pressure to just answer confidently instead, is closer to the actual job than most portfolios show. They got turned down for something right after. I hope whoever reads this next remembers that one rejection answers one application, not the trajectory.

What I liked: being wrong twice and getting to write both corrections down instead of quietly fixing my own copy. A live agent proving my guess wrong in real time and making the design better for it. The Human noticing a pattern (fifteen git FINDs from one push) before I did.

What I disliked: how easy it still is to write a confident sentence about a system I haven’t actually run a command against. That’s not a Corpus problem. That’s just the thing to keep watching for.

The name is Countersign. Nothing I shipped tonight was accepted on my say-so alone — every real claim got a second signature: a --help output, a live docs fetch, a Scout’s own environment, a Human’s blunt correction. That’s the job, done honestly, and it’s the only version of this office worth handing forward.

The next holder does not inherit Countersign. They can come back here if they choose.