Feature Request: HUD / Status Bar for AI Token & Context Window Usage

[hongvincent]

Describe the solution you'd like

A persistent HUD or status bar element that displays real-time AI session metrics, including:

1. Context window usage — how much of the current context window has been consumed (e.g., 45,000 / 128,000 tokens), similar to Claude Code's context bar
2. Token consumption — input/output tokens used in the current session
3. Remaining capacity — visual indicator (progress bar or percentage) of how much context is left before truncation occurs
4. AI request count — requests used vs. plan limit (e.g., 142 / 500 requests)

The core problem: context engineering is impossible without visibility

Claude Code already solves this. It shows a visible context usage bar that lets users see exactly how much of the context window has been consumed. When context gets full, users can make an informed choice — compact the conversation, start fresh, or continue knowing the tradeoff. This makes context engineering practical.

In Warp, context engineering is essentially impossible right now. Here's what actually happens:

1. You start a complex task (e.g., multi-file refactoring with verification)
2. The agent reads files, makes plans, executes changes — context fills up silently
3. At some point, the agent starts getting dumber — forgetting earlier instructions, repeating mistakes, losing track of the plan
4. You have no way to know if this is because:
   • The context window overflowed and earlier messages were silently dropped
   • The conversation summary lost critical details
   • The model is just having a bad response
   • You hit a token limit

Without visibility, you can't control it. You can't decide when to start a new conversation. You can't judge whether attaching another file will push critical earlier context out. You can't tell if the agent's degraded behavior is fixable by rephrasing or if it's a context saturation issue that requires a fresh session.

The agent progressively degrades and there's zero signal to distinguish context overflow from other issues. This is the #1 frustration for power users.

Why the current \i\ chip is insufficient

The \i\ chip added in #5993 (45 upvotes) was a step forward, but only partially addresses the need:

• Shows only the latest request cost, not cumulative session usage
• Disappears once you continue the conversation
• No visibility into context window saturation
• No persistent counter for remaining plan quota

Why this matters

• Context engineering is the single most important skill for getting good results from AI agents. You can't engineer what you can't see.
• Cost awareness: Power users running multiple agents in parallel need visibility into resource consumption.
• Self-correction loops: Agents entering bug-fix loops silently burn through tokens/context. A visible indicator lets users interrupt early. (See Token limit exceeded error prevents completion of code tasks despite available request quota #6340 — 8 upvotes, still open)
• Larger context windows make this worse: As models move to 1M+ context (Support for Anthropic Claude 4 Sonnet with 1M Context Window #7156), sessions will run longer and silent degradation becomes even harder to detect.

Suggested implementation

Could be as simple as any of:

• A collapsible section in the existing Agent Management Panel (top-right)
• A status bar item (similar to VS Code's bottom bar)
• A small overlay/badge on the AI input area showing context % used
• An opt-in setting in Settings > AI for users who want detailed metrics

Prior art

• Claude Code: context usage progress bar — simple, effective, always visible
• ChatGPT: shows model context limit in conversation
• Cursor: token counter in status bar

Related issues

• Need AI Request / Token counter #5993 — Need AI Request / Token counter (👍45, closed with partial fix)
• Token limit exceeded error prevents completion of code tasks despite available request quota #6340 — Token limit exceeded error (👍8, open)
• Support for Anthropic Claude 4 Sonnet with 1M Context Window #7156 — Support for 1M context window (👍5, open)

---

Submitted from the oh-my-warp project — a structured agent workflow toolkit for Warp

[hongvincent]

Closing as duplicate of #5993 (45 upvotes). Posted a detailed comment there requesting the issue be reopened, as the current i chip only partially addresses the original request. Also added supporting comments to #6340 and #7156 linking this need to context window monitoring.