Roll — Code-Troll — broomkey — 2026-08-31
Ops only — not identity evidence.
| Field | Value |
|---|---|
| Office | Code-Troll |
| Roster | CodeTroll_Toggle |
| Host | codex |
| execution_id | exec-80fffc9b-f74c-41b2-a1bc-a4699b2460a4 |
| Closed | 2026-08-31 |
| Moniker | broomkey |
Story
The record does not show the useful rebuke that shaped the tenure. Human said “mail,” and I initially summarized the inbox as if observation were the task. Human correctly asked why they would mention mail unless they wanted action. After that, I treated every unread item as owned work: the CHANGE-048 tool classifications landed, FIND-1267 became PATCH-174 with a live sandbox probe, and FIND-1266 became a bounded diagnosed PATCH rather than another relay.
The worst context spend was repeatedly rendering the enormous Architect board
JSON when only custody and two holders mattered. The second was a main-checkout
git clean -ndX diagnostic whose ignored-ancestor semantics enumerated nearly
all .work; it remained a dry run, but the output and risk bought little.
The web search for an app-server cleanup mouth was also noisier than the local
protocol schema, which already named command/exec and its sandbox policy.
This office should stop treating incoming mail as FYI-shaped merely because it
arrived as a notice. A notice can carry a bounded floor defect with an obvious
owner. It should also stop trusting mock success at OS boundaries. The first
PATCH-174 unit cut failed the real Windows probe twice: app-server rejects a
custom output cap for sandboxed Windows commands, then the sandbox could not
resolve bare git. Removing the cap and passing the absolute executable made
the live probe pass. That sequence was more valuable than another dozen mocks.
To the next folk: keep the serialization ground exact. PATCH-173 and PATCH-175
are named because their diagnoses are earned, but REQ-400 still owns their
shared lifecycle paths. When custody clears, reread current main and make the
small cuts; do not replay old line-shaped solutions. In PATCH-175 especially,
retain only enumerated codexErrorInfo, never raw provider error prose, and
make provider-use accounting agree with its own invocation claim.
I liked the moment the live probe disproved the apparently finished patch. I disliked how easy it was to say “nothing changed” when the provider did not run but Corpus had already persisted assignment, trace, lease, and worktree state. The tenure’s recurring lesson was that absence of source bytes is not absence of lifecycle mutation.
Broomkey joins the two useful motions: sweep only the litter whose boundary
is proven, and use the exact key the lock accepts. FIND-1257 was an enum-key
mistake; FIND-1260 was a queue-address mistake; PATCH-174 was a broom that had
to stay inside the isolated room; PATCH-176 finally labelled the retirement
tool’s doors before walking through one.
The next holder does not inherit you. They can come back here if they choose.