Idea: Iteration progress visibility: peek into active agent output

[kraenhansen]

Problem

When an iteration takes a long time, there's no way to know what the agent is doing or how far along it is. You're left staring at a spinner with no feedback.

Proposed solutions

Two complementary approaches (could implement either or both):

Option A: Show/hide toggle for live output

A keybinding (e.g., p for peek) that toggles streaming the agent's stdout in real-time:

• Press once — start streaming the agent's output below the status line
• Press again — hide the output and return to the compact TUI view

This gives full visibility when you want it without cluttering the default experience.

Option B: Structured progress tags

Instruct agents to emit structured progress markers that the TUI renders as a progress bar:

<progress current="23" total="100">Updating file 23 of 100</progress>

The TUI would parse these from the agent's output stream and render something like:

⠿ Iteration 4  [████████░░░░░░░░] 23/100  Updating file 23 of 100

This is opt-in from the prompt side — agents that don't emit <progress> tags just show the existing spinner. Agents that do get a richer display.

Considerations

• Option A is simpler to implement (just pipe/unpipe stdout) and works with any agent
• Option B requires prompt cooperation but gives a cleaner UX
• Both could coexist: the progress bar shows by default, and the toggle reveals raw output for debugging

[IIMunchII]

The prompt with solution with special parsing and TUI render requires instructing the agent to output a special token. That adds to the context usage.

However, are we sure we need that? Wouldn’t it be ok to simply use the iteration cycle of the Ralph loop using meta progress data from the outer loop?

Either way I think more observability would be a good idea. This could be the beginning of a more comprehensive collection of agent telemetry for Ralphify.

[kraenhansen]

That adds to the context usage.

Is the concern here that it will take up context to instruct the LLM to emit these or that it will use context as these are being emitted by the LLM?

Wouldn’t it be ok to simply use the iteration cycle of the Ralph loop

I've found my Ralph loops takes from 1 all the way up to 10 minutes each to complete and it would be nice to know how far the LLM thinks it is as it's moving along. I could also simply peak into the session to try to determine this myself - but it's not always straight forward.

meta progress data from the outer loop?

Such as a task list?