[FEATURE] Cowork/claude.ai: Visible context window usage indicator and working /context command
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 and claude.ai provide no visibility into context window usage during a conversation. Users cannot see how much context has been consumed, how much remains, or when auto-compaction is approaching — until it happens silently mid-task.
This makes it impossible to budget work within a session. Complex editorial or development workflows that span multiple tool calls and file operations accumulate context silently. The user has no way to know whether they can fit one more task into the current thread or should start a new one.
Concrete observation: In claude.ai (March 2026), typing /context in the input box triggers an autocomplete suggestion — "Show what's using your context window" — but the command cannot actually be invoked. The UI presents it as available functionality that doesn't work.
Prior history: Issue #20041 requested this exact feature for Claude Desktop but was closed as invalid with the note that it "doesn't seem to be related to Claude Code." At the time, there was no area:cowork label and no established path for Cowork-specific feature requests. That label now exists and is actively used (see #38466, #38468, #37413, #34832, #24677). This submission covers the same gap — context visibility for non-CLI interfaces — now that there is a proper intake path.
Related CLI issues: Claude Code CLI has extensive coverage of this need (#5547, #5720, #5739, #6577, #15247, #12833, #28962). The CLI has a /context command and statusline data. Cowork and claude.ai have neither.
Proposed Solution
- Make the
/contextcommand work in Cowork and claude.ai. The autocomplete already surfaces it. Connect it to the same data the CLI uses — token usage by category, remaining capacity, percentage consumed.
- Add a persistent, lightweight context indicator to the UI. A percentage or bar visible in the conversation header or status area (e.g., "Context: 47% used" or "~53K tokens remaining"). This should be always visible, not just at critical thresholds.
- Surface threshold warnings at meaningful breakpoints. For example, a subtle notification at 50% ("consider scoping remaining work"), a moderate one at 75% ("good time to wrap up or start a new thread"), and a strong one at 90% ("compaction imminent").
Alternative Solutions
- The
/contextcommand alone (without a persistent indicator) would be a meaningful improvement — users could at least check on demand. - A notification only at compaction time (e.g., "Context was compacted — this is compaction #2 in this session") would be a lighter-weight alternative, though it's reactive rather than proactive.
- Exposing context data via the API response headers or metadata would let power users and third-party tools build their own indicators, similar to how #6577 proposes statusline hooks for the CLI.
What doesn't work today: There is no workaround. The model itself cannot report its own context usage in claude.ai or Cowork. Third-party tools like ccusage don't read Desktop/Cowork session data. The only signal users get is when compaction happens or when the conversation silently degrades.
Priority
High - Significant impact on productivity
Feature Category
Interactive mode (TUI)
Use Case Example
- I'm running a multi-step editorial workflow in Cowork — reviewing a draft, applying style corrections, cross-referencing a lexicon, and producing a final document.
- The session involves multiple file reads, skill invocations, and iterative revisions — all consuming context silently.
- Partway through step 4, compaction fires without warning. The model loses track of earlier decisions and starts contradicting its own revisions.
- If I could see "78% used — ~44K tokens remaining," I'd know to wrap up the current task and start a new session rather than losing work to silent compaction.
- Even a simple
/contextreadout on demand would let me check before starting a large sub-task: "Do I have room for this, or should I start fresh?"
Additional Context
Related issues (CLI-focused — this request is for Cowork/claude.ai):
- #20041 — Same request for Claude Desktop, closed as
invalid("doesn't seem to be related to Claude Code"). Thearea:coworklabel now exists and is actively used. - #6577 — Context data in statusline hooks (CLI), 23+ upvotes
- #5547, #5720, #5739, #15247, #12833, #28962 — CLI context visibility requests
Platform: Windows, Claude Desktop with Cowork, Opus 4.6
Companion request: A separate feature request for compaction event visibility (has compaction occurred, how many times) will be filed as a complementary issue — that covers after compaction, this covers before.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗