Keep one conversation going with Claude Code and Codex
Chief now holds a single ongoing session with your coding agent instead of starting over each turn, so context carries between prompts and you approve tool use as it happens.
Working with a coding agent through Chief used to mean each turn stood on its own, with the agent picking up the thread again from a session id. Chief now speaks the Agent Client Protocol, the same protocol Claude Code and Codex already understand, and holds one conversation from start to finish.
Persistent agent conversations
You open a conversation once and keep prompting it. The agent keeps what came before, so a follow-up like "now do the same for the other file" lands the way you meant it.
If a conversation's process is gone, Chief can reopen it with agents that support it, and you pick up where you left off without the earlier back and forth flooding your screen again.
Live answers and tool activity
Answers arrive in pieces as the agent writes them rather than all at once at the end. You also see the agent's own title for each tool it runs, so you can follow what it is doing to your files while it does it.
Approving tool use
When the agent wants to run a tool, it asks first. You get the tool, what kind of work it is (reading, editing, running a command, fetching), and the choices the agent offered: allow this once, allow from now on, refuse once, or refuse from now on. Your answer is what the agent is told.
Stopping a turn without losing the thread
Stopping a prompt used to mean losing the whole conversation with it. Chief now asks the agent to stop the current turn and keeps the conversation alive, so you can redirect and prompt again. An agent that ignores the request does not leave you waiting.
Real reasons when an agent fails
One of these agents exits as though nothing went wrong even when you are signed out, which made failures hard to read. Chief no longer treats a clean exit as success. When a turn fails, you get what the agent actually printed, including the sign-in error or the usage text it produced when a flag was wrong.
Agents keep their own file and shell tools and any MCP servers you configured. Chief brings none of its own to the session, and it will not read or write files on the agent's behalf, so the approval prompt stays the place where you decide.