Template set 13: the compressed register, the wiki's pages, and RTK at launch

[protonspy]

Two commits, both from the same session's pending work.

feat(assets) — template set 12 → 13

caveman.md, the register the agent answers in. Short output, unchanged
thinking. The output budget belongs to the code: prose about the work is written
once and then carried in every later request of the session, so narration is the
part of a long run that can be cut without losing a fact.

It ships as a rule rather than a skill because it is on by default, and a
default the model has to decide to load is not one. One level (ultra) rather
than a dial, because three descriptions of the register are three things to keep
true instead of one. What never compresses is the line that keeps it honest:
artifacts under specs/, plans/, docs/, anything a validator parses or a
shell runs, quoted output, commit and PR bodies, and questions put to the user.
A denser EARS line is a finding, not a saving.

The language is the third kickoff answer. autonomy.md asks it with the
other two and it lands in the artifact's frontmatter as lang: en|wenyan,
graded by checkKickoffAs on exactly the terms autonomy and ci are —
checked when present, absent meaning the run predates the question. There is no
--lang flag on spec new or plan new: it is the one answer that can arrive
after the file exists, so scc patch fm <artifact> lang=wenyan is the whole
path to it, and a value neither the rule nor the validator knows rolls back like
any other bad edit.

Also in the pass:

• The wiki's pages move to docs/wiki/pages/, so index.md and
  changelog.md are told apart from content by where they sit rather than by
  their names. The validator used to exclude them by matching those names, which
  made any other .md dropped into wiki/ a page, and then an orphan.
• Three places prose stated a mechanism exactly and left the judgment unsaid —
  where an agent fills the gap with the cheapest reading. The plan-run command
  said "read the plan" where the skill it invokes says map it; the review agents
  asked for surrounding code without saying the diff already held it; the wiki
  skill explained how a slug resolves without saying the name has to name the
  concept.

feat(launch) — RTK joins the launch preflight

scc launch already brought Headroom and CodeGraph up to date; RTK was the
third and the one left to the user to remember. It degrades the way the other
two do — missing cargo, declined install, unattended run, failed build all end
in the agent starting anyway — and --no-rtk is the standing answer.

Two things it does differently, both because it is the only one of the three
that writes to a file the user owns: it runs last, so a run aborted at an
earlier prompt has not already edited the entry file, and its prompt names
both halves, since consenting to an install is not consenting to an edit.

confirmInstall is a second confirm() defaulting to yes, used only by these
install offers. A wrong yes here is a tool that can be uninstalled; a wrong no
is a whole session running degraded. confirm() keeps its no-by-default,
because what it guards is somebody's file.

Verification

gofmt -l, go vet ./..., go test ./... all green. The race detector needs a
C compiler this machine does not have — CI's Linux and macOS jobs run it.

• TestRulesStayShortEnoughToBePreloaded holds caveman.md inside the 55-line
  budget, which is what keeps a preloaded rule affordable.
• TestTheRuleOffersEveryKickoffAnswerThisAccepts is new and fails if the rule
  and the validator ever name different values for a kickoff answer.
• New coverage for lang on both the spec and the plan side, and an end-to-end
  patch fm test for the write and for the rollback.
• Smoke-tested in a scratch workspace: init writes .claude/rules/caveman.md,
  patch fm specs/billing/requirements.md lang=wenyan writes and re-validates
  clean, lang=pt-BR rolls back with spec.kickoff-invalid and exit 2.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DmGyL6QvamMBKdHYyrx7Uq