[FEATURE] Visible context window usage indicator in Cowork

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

Cowork has no visible context window usage indicator. Users cannot see how much of the context window has been consumed or how close compaction is.

Claude Code solves this with a persistent status line showing context usage as a percentage. Cowork users get nothing — they fly blind until compaction arrives as a surprise mid-task with no warning, at the worst possible moment.

Why this matters in practice:

I have 50+ Cowork sessions running a retail business through Claude collaboration. Context compaction has silently reset my session mid-task multiple times. There is no yellow zone warning, no red zone alert, no indication that compaction is imminent. When it hits, critical context disappears without any summary of what was preserved versus what was lost.

You cannot fly a plane without knowing how much fuel is in the tank. You cannot run a productive Cowork session without knowing how much context remains.

Related issues that confirm this is widespread:

• #24677 — Compaction death spiral in Cowork: 6 compactions in 3.5min due to system context consuming 86.5% of window
• #13171 — Context loss without warning breaks trust and productivity
• #34363 — Autocompact buffer exceeds 200k context window — instant compaction loop
• #20041 — Context usage indicator for Claude Desktop (closed as "not about Claude Code" — but the need applies to all surfaces)
• #35374 — Persistent status line showing context usage in CLI

Proposed Solution

A persistent, always-visible context usage indicator in the Cowork UI. Specifically:

1. Persistent indicator — always visible, not hidden behind a menu. Shows percentage of context window consumed, updated in real time as the session progresses.

1. Yellow zone — visual warning when approaching the compaction threshold (e.g., 75%). This gives users time to save important context, wrap up a task, or start a new session proactively.

1. Red zone — urgent warning when compaction is imminent (e.g., 90%+). At this point the user should be able to choose: compact now with a summary, or start fresh.

1. Compaction transparency — when compaction occurs, surface it clearly with a summary of what was preserved and what was dropped. Currently compaction is silent — the session just quietly degrades. Users deserve to know what happened.

This matches what Claude Code already provides via the status line percentage, extended to the Cowork surface where compaction impact is arguably worse because Cowork sessions tend to be longer and more complex.

Alternative Solutions

I currently build custom CLAUDE.md files and session-reload skills to survive compaction by re-reading critical context from disk after every compaction event. This works but is a manual workaround that shouldn't be necessary. The platform should tell me when compaction is coming so I can manage my session proactively instead of building defensive infrastructure around a silent failure mode.

Priority

High - Significant impact on productivity

Feature Category

CLI commands and flags

Use Case Example

Real scenario from my workflow:

1. I open a Cowork session to build a complex kiosk data pipeline for my retail business. The session involves reading database schema files, designing queries, writing Python scripts, and testing against a remote NAS server via SSH.

1. 45 minutes in, I've built up significant context — Claude understands my database schema, the SSH connection details, the file paths on the NAS, and the current state of the pipeline code.

1. Without any warning, compaction fires. Claude suddenly doesn't remember the SSH connection string, the database table relationships, or what we were building. I have to re-explain everything.

1. With a visible context indicator, I would have seen the yellow zone at 75% and proactively saved critical context to a CLAUDE.md file or started a fresh session with a handoff summary. Instead, I lost 15 minutes reconstructing context that silently vanished.

Additional Context

_No response_