A decision is not made until the artifacts that state it are rewritten.
Mirrored from the fleet knowledge base: _rules/a-closed-fork-rewrites-what-it-invalidated.md
Updated
Mirrored verbatim from
utopia-models/knowledge_rules/a-closed-fork-rewrites-what-it-invalidated.mdby the allowlisted publish pipeline. Edit it there, not here.
name: a-closed-fork-rewrites-what-it-invalidated
cases: [2026-08-08-i-researched-a-superseded-document-for-two-runs, 2026-08-09-a-plan-i-do-not-update-is-a-plan-tyler-has-to-audit]
status: active
mechanism: prose — the trigger needs judgement a hook cannot observe
audience: agent
last-recurrence: 2026-08-09A decision is not made until the artifacts that state it are rewritten.
Fires when
Tyler resolves a fork, reverses a decision, or answers an AskUserQuestion — and you are about to
write the NEXT section. Also when reusing any extract, draft, title or published page created before
that answer.
The rule
Before writing anything new, re-read what is already written against the new decision. The H1, the published artifact, the phase headings, the filename, the SETTLED section. One question: does anything above this still describe the version we just replaced?
Republish the artifact at every fork closure, not at COMMIT. Same file path keeps the URL. An artifact published once and never republished is worse than none — it looks current, carries a timestamp, and states decisions already reversed.
Prefer derived over hand-maintained. task regen re-derives title and phases from the plan
document, so a corrected H1 fixes the registry for free. Anything with no derivation — filenames,
artifacts, prose summaries — is the part that stays wrong, so name files from stable ids
(plan-116-<slug>), never from a title a fork can change.
A cached extract is not the source. Re-read the source directory, not the copy you made of it.
Cost when it fired
2026-08-09: Tyler corrected the same plan's state THREE times in under an hour — a stale title after he scoped it down ("are we still using meta prompts?"), a stale artifact and filenames ("The artifact and plan seem to be legacy too"), and a mechanism I had asserted without checking ("I thought .claude/commands was soon to be legacy" — he was right). Each was a check I could have run and did not. The plan's own subject was documentation that cannot drift.
2026-08-08: 136 agents ran against sol.txt, a days-old extract matching neither PDF in the source
folder. The unread file was amendment 0002 — it flipped the set-aside to FOR small business and moved
the due date. I had noted that file existed and never opened it.