kagi

Three Unix-style CLIs for Kagi: kagi-search, kagi-maps, and kagi-summarize. README.md covers install, the token, and usage.

Gates

task test:live calls the real Kagi service. Run it when you change the request path or a parser: src/cli.rs, src/client.rs, or src/parse.rs. It needs a session token (see README.md) and fails without one.

Keep three separate binaries. Do not add a combined command again.
Never hardcode a session token.
--sort means two things. kagi-search sends it to Kagi as the order parameter. kagi-maps gets the whole result page, sorts it locally, then cuts it to --limit.
Our machines install the CLIs with mise (github:rubas/kagi). mise picks the release archive by the OS and architecture in its name, kagi-linux-x86_64.tar.gz or kagi-macos-aarch64.tar.gz, and finds the binaries in its bin/ dir. A change to an archive name or layout breaks the install on every machine.
The repo ships no agent skill. The search skill in rubas/dotfiles covers these CLIs.
A version bump is the release trigger. On each push to main, release.yml reads version from Cargo.toml. When the tag v<version> does not exist, it tags and publishes. The tag decides, not the parent commit, so a push of many commits or a cancelled run does not lose a release. A tag that someone deletes comes back on the next push. To recover a failed build or release, run release.yml through workflow_dispatch.

ci.yml runs its checks only on a pull request from a branch of this repo, so a fork PR never reaches the incus runners. A fork PR shows no checks. That is the gate, not a broken run.