Repository navigation
v4.1.4
A door for feedback. A finished run of cast, paint or bunshin now ends its report with a
link to a one-minute form: which skill, how it came out. The README says what is tested and what
is not. Nothing else a skill loads or does changed.
Added
- A one-minute form, "How did it come out?", first in the issue chooser. Two answers are
required: which skill, and whether the result came out better than expected, fine, plain or
generic, or wrong or broken. The request, the version, where it ran, the model, a capture and
the permission to feature it are optional. Until now the only forms asked for a bug with its
resolution error or a correction with its primary source, and nobody could simply say "it came
out plain". Output that came out wrong is as useful as output that came out well. - A finished run ends its report with the link to that form: two fixed lines, once per
conversation, on the first report of a run that wrote code into the project, in a headless run
too, and never on a run that wrote none. The link fills in the skill, and the genjutsu version,
the host and the model when the run knows them: the version is only known on a Claude Code
plugin install, where the install path carries it. Nothing from your project goes in the URL
(no request, name, path or content). The skill is told to print the link and stop there: never
to open it, file the issue or fill the form itself. Nothing is sent unless you open the link
and submit the form. The block is a new shared region,feedback, byte-identical in the three
orchestrators and guarded byscripts/check-shared-blocks.sh, which also fails when the link
names a form or a text field that does not exist.
Changed
- The README says what has been tested, and no more. The Compose and SwiftUI modules ship,
but every recorded example is web: no native run has been compiled and recorded yet. genjutsu
is developed and run on Claude Code with Claude models. The claude.ai bundle was last checked
by hand on 2026-09-28,castonly; no Cowork run is logged; bunshin has only been run on
Claude Code; other agents and other models are not tested.PLATFORM-CONTRACT.mdsaid that
Claude Code, claude.ai and Cowork "are tested by the maintainer before a release", which
docs/claude-ai-testing.mddoes not back: it now says what that log holds. The eval table
carries its date: measured on 2026-09-28, on commitef31234of the branch that became
v4.0.0, and not rerun since. - The README gives the install command to hand to an agent, says once that this is not Higgsfield
Genjutsu, the video tool, and replaces "Show what you cast" with "Tell us how it came out",
which leads with the form. - The issue forms are numbered so the chooser lists them in order:
1-output.yml,
2-pipeline-bug.yml,3-wrong-claim.yml,4-stale-version.yml. A link carrying an old
template=name no longer finds its form: update it. plugin.jsonnames Cowork next to Claude Code and claude.ai, mentions bunshin, and carries the
listing fields Claude Code documents for Anthropic's directory:documentationUrl,
supportUrlandicon.
Fixed
- The bug form listed
/genjutsu:castand/genjutsu:paintonly./genjutsu:bunshinis in the
list, and in the family list of the wrong-claim form.