stream-json lets apps drive Claude Code, tanacode 2.0.0 shows

Stream-json is framed as a middle path between two other ways of running Claude Code: the interactive terminal screen and the one-shot claude -p command. In one-shot mode, the process ends after a single answer, so continuing a conversation requires restarting with --resume and gives no way to answer permission checks mid-turn. Stream-json instead keeps the CLI process open until standard input is closed, so a program can send successive user lines and receive assistant, result, and control_request lines as the turn progresses.

The developer's own app, tanacode, illustrates why this matters in practice. Version 1.x ran the interactive screen in the background and read its text to build chat and permission cards, a design that broke whenever Claude Code's screen changed slightly. Version 2.0.0 rebuilt the chat on stream-json and removed the screen-reading code, so permission prompts and questions become cards built from can_use_tool requests and answers flow back as control_response lines.

Operational caveats are also flagged: programs should reply to every CLI request even if unused, interrupted messages remain in the conversation, and stream-json sessions do not show up in the claude --resume list, requiring the session_id from the init line to resume. The developer notes that stream-json line formats may change with Claude Code updates, and says tanacode runs the latest Claude Code against a mock API daily to detect new line types.