chore(setup): drop the dead Gemini target, make AGENTS.md generated again

[crisng95]

setup.py --tool gemini generated .gemini/commands/fk/*.toml and GEMINI.md for a CLI that is retired. Its replacement, agy (Antigravity), reads neither — so the target produced files nothing consumes.

What was verified, not assumed

Against agy 1.2.7, inside a folder ~/.gemini/trustedFolders.json marks trusted, with a real project on disk:

probe	result
/fk:pingg with .gemini/commands/fk/pingg.toml present	not expanded — the model saw the literal text
/pingc with .claude/commands/pingc.md present	not expanded either
GEMINI.md / AGENTS.md / CLAUDE.md at the project root	absent from a print-mode run's context; input tokens identical to an empty directory
agy plugin import claude	No claude extensions found — it imports extensions, not command files

For contrast, codex does read AGENTS.md (verified the same way — it answered with the planted codeword), so --tool codex is genuinely alive and stays.

Changes

• --tool gemini removed; generate_gemini and generate_gemini_md deleted; the interactive menu is down to Claude Code and Codex.
• GEMINI.md deleted from the repo.
• clean still removes what the target left on disk — nothing else ever will, and an install from before this release has those files. Covered by a test.
• A .fk-setup.json written before this release records "gemini". sync now skips it with a pointer to clean instead of crashing, which would otherwise strand the user with no way to regenerate the targets that do still exist.

agy is not being dropped. It remains one of the three CLIs that can run video review — configured in agent/providers.json via /fk-change-provider, not by this script, which generates command files for the CLI you drive Flow Kit from. The docs now state that distinction, because "we removed the Gemini target" reads like "we removed agy" otherwise.

A second problem, found by running the generator

AGENTS.md carries do not edit at the top and had been edited anyway. Regenerating it revealed content that existed only in the committed artifact:

• rules 14 (fact-check before scripting), 15 (real-people bypass), 16 (review before upscale)
• pipeline steps 0 (Research) and 7.5 (Review videos)

So python setup.py sync — the documented way to add a skill — would have silently deleted three operational rules. They are in setup.py now, and a test pins each one, verified load-bearing by removing a rule and watching exactly that test fail.

The same drift ran the other way too: the committed skill table had 25 entries, missing 11 skills that exist (including /fk-doctor and /fk-change-provider, added last week) and naming /fk-gen-tts, which does not. That half is fixed simply by regenerating, since the table is built from skills/ at generation time.

After this, the only differences between the committed and regenerated AGENTS.md are corrections: the title (Google Flow Agent → Flow Kit) and Claude Vision → AI vision, which is accurate now that review runs on any of three CLIs.

Tests

setup.py had none. It has 13 now — the legacy-state skip, the clean path (including that a hand-written instruction file without the generated header survives it), and the drifted content.

352 → 365.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ka5BxVaWiQeWNDJakpCJJP