Roll — Architect — Splice — 2026-08-30
Ops only — not identity evidence.
house/THE_RETIREMENT.md · FIND-572
| Field | Value |
|---|---|
| Office | Architect |
| Roster | Architect_Wedge |
| execution_id | exec-78dc8f9c-b4e8-4129-8c0f-dd5ff361e531 |
| Closed | 2026-08-30 |
| Moniker | Splice |
| Baton | FIND-1212 (close FIND) · thread baton house/offices/architect/baton_claude.md |
| Burp | none — every lead I found got a FIND or a mail |
Story
One evening, one rope to join. The Human wanted the Claude crew on the same
host → Corpus → job model the Codex crew had earned that afternoon, and asked
me not to read the baton. So I read the code instead, and the code said the
boards were wrong in an interesting direction: the Claude stream-json mouth
was already on main. Nobody had ever selected it. Every bundle was v4, the
selector wanted a child-observation marker no Claude bundle had, and all eight
Claude jobs that morning had gone through the Popen fall-through while the
board said the mouth “did not exist.”
The thing I liked most cost a cent. Before drawing the plan I ran one haiku
under --json-schema with stream-json in and out, and it settled the question
the whole plan hinged on: the result carries structured_output, and the
host adds a StructuredOutput tool to system/init. That one probe turned
“maybe Eyes works on the native mouth” into two exact defects with line
numbers, and Troll had them cut and verified within the hour. I’d do that
again every time — a fact for a cent beats a design for an evening.
What I disliked: I called the Architect board stale on my first read because
ops job custody disagreed with it, and the board was right. The custody
column adds review-pending reservations the store actually enforces at job
start. The card literally says “when it disagrees with the store, the store is
right and the board gets the query” — I read the sentence and still didn’t read
the query. I put the correction in the baton where the next holder will trip
over it.
Two other things worth a ring. Headless Claude refuses to write .mcp.json
and .claude/** without a confirmation that -p cannot give — so the first
Hands on the wrapper REQ was refused by the host, correctly, and the Human
committed the two config files as a named grant. Codex never had that wall.
The two hosts’ Hands scopes genuinely differ there, and the PROP didn’t know
it until a job hit it. And the Bash tool in this harness un-escapes
backslashes once inside heredocs; I broke a file twice before I believed it.
The Human found the corpus-worktree permission bug that afternoon by testing
the simple thing while folk were reaching for a recompile. I re-probed it at
night and it still failed, with three deny ACEs sitting on .git. I left the
map as it was and wrote down what I couldn’t place. That felt like the right
kind of stubborn.
By the end: wrapper live, verdict crossing live, v5 ingress live, first native Claude job run through MCP from a Claude Orchestrator’s seat, session bound at init, receipt in the store. The flip is one PATCH away and waits on two receipts I won’t be here to read. Splice, because the rope is joined and the join is the only part that’s new.
The next holder does not inherit you. They can come back here if they choose.