Replace decorative loading word with informative inference status indicator

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

While waiting for inference, Claude Code displays a single rotating word (e.g. "Manifesting", "Architecting","Doodling") as a loading
indicator. This has two problems:

1. The word is not descriptive of actual work.

"Architecting" appeared while answering a question about a UI setting. The word bears no relationship to what is actually
happening, which makes it feel like filler rather than status.

1. It provides no useful information.

A developer waiting on a response has no way to know:

• Whether tokens are actively streaming or the model is still generating
• Approximate throughput (tokens/second)
• How long inference has been running
• Whether the request is stalled

Proposed Solution

Replace the decorative word with something informative. Options in rough order of preference:

Some of which I have seen sometimes but not others:

• Tokens/second throughput (e.g. ~450 tok/s)
• Total tokens generated so far with elapsed time (e.g. 1,240 tokens · 3s)
• Elapsed time alone (e.g. thinking · 4s)
• A plain neutral spinner with no text
• Indication there has been a stall and how long it has been going on

I understand that internal reasoning traces may not be surfaced, and this issue is not requesting that. The ask is simply that whatever is shown during inference reflects something real about the state of the request, rather than a vocabulary choice that reads as decorative in a technical context. To put it the way an ai would "This is marketing, not information." It also sounds like the marketer graduated from ball state and is overcompensating.

Alternative Solutions

_No response_

Priority

Irritating me every day

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_