Agent and CI mode

Add --agent to any okibi command when a script or a coding agent runs it:

okibi --agent clis show my-cli

--agent turns on three things at once:

• No prompts. okibi never stops to ask a question. Steps that need a person in a browser, such as okibi auth publisher, return approval_required with the URL to open.
• JSON output wherever the command supports it.
• Structured errors, so a failure can be read as data instead of scraped from text.

Most listing and management commands also accept --json on its own, for JSON without the other two.

Exit codes

With --agent, a failure prints JSON to stdout:

{"ok": false, "error": "…"}

Without --agent, the error is plain text on stderr.

okibi generate --review exits 0 when the run stops for spec approval. That pause is not a failure.

Run okibi in CI

1. Create a workspace-locked key with okibi keys create --name ci --workspace acme and store it as the secret OKIBI_API_KEY. See Log in.
2. Install okibi in the job and run commands with --agent:

curl -fsSL https://okibi.ai/install | sh
export PATH="$HOME/.local/bin:$PATH"
okibi --agent generate --repo https://github.com/OWNER/REPO --name my-cli --no-pull

--no-pull skips downloading the finished CLI when the job only needs the hosted release. Set OKIBI_NO_AUTO_UPDATE=1 to pin the installed version for the job.

Retry safely

Messages to a running generation take a stable --request-id. Reuse the same id and the same text after an ambiguous failure, and okibi won't deliver it twice:

okibi --agent message <run-id> --text "Add pagination" --request-id fix-pagination-1 --wait

Things that always need a human

• Deleting a CLI. okibi clis delete refuses to run with --agent. A person must type the CLI name in a terminal.
• Logging in. The first login opens a browser or prints a device code.
• Paid plan changes. okibi billing change needs --yes, which confirms the charge.