Tangent

Need to ask something unrelated mid-task, like looking up an API limit, debugging a side error, or exploring an alternative, without polluting your main thread? /tangent lets you branch into a side-conversation that inherits everything discussed so far, explore freely, then jump back to exactly where you left off.

While in a tangent, a ↯ name chip in the footer shows where you are.

Each tangent is a full session: it starts with your conversation's history up to the branch point, but everything after stays isolated. Nothing you do inside a tangent touches the parent conversation's context. Because a tangent branches from your conversation's history, you can't branch an empty chat — send a message first.

Tangents can also have tangents. Nest as deep as you need — /tangent ls shows the whole tree, and /tangent root gets you home from anywhere.

1. Quick lookups without derailing. Mid-refactor and need to check DynamoDB item size limits? /tangent, ask, get the answer, /tangent back. Your main thread never sees the detour.
2. Compare alternatives in parallel. Branch /tangent approach-a and /tangent approach-b off the same design discussion, explore each with full shared context, then return to the root and commit to the winner.
3. Contain noisy debugging. A failing test needs 15 turns of log-dumping? Do it in a tangent. The noise stays out of your main conversation's context, keeping it focused and cheap.
4. Ask the "dumb question" freely. Want to ask something basic without the model anchoring on it for the rest of the session? Tangent it.
5. Nest as deep as you need. A tangent mid-tangent is fine — the tree keeps track for you.

• CLI 3.0 overview for the unified agent harness and how to opt in
• New features in 3.0 for what else is new in the v3 engine
• Tangent mode (experimental) for the earlier single-checkpoint behavior