Feature Request

The status bar's context usage indicator (14/200K (7%)) currently only updates after an entire task completes. During a long multi-step agent loop (LLM → tool → LLM → tool...), the context usage stays frozen, giving the user no visibility into how much context window is being consumed in real-time.

Current Behavior

1. User sends a prompt → task starts
2. Agent loop runs 10+ LLM calls with tool use over several minutes
3. Status bar context usage stays at the value from the previous task
4. Task finally completes → usage updates in one jump

Desired Behavior

Context usage updates after each LLM call within the agent loop, so the user can see context consumption growing in real-time and decide whether to intervene (cancel, compact, etc.) before hitting the limit.

Implementation Plan

Daemon side (StreamingAgentHandler / StreamingAgentLoop)

After each LLM call in the agent loop, send a notification:

{"jsonrpc": "2.0", "method": "stream.usage", "params": {"inputTokens": 42000, "outputTokens": 8000, "totalTokens": 50000}}

The inputTokens here should be the cumulative input tokens from the most recent API call (which represents current context window consumption).

CLI side (TaskStreamReader + TerminalRepl)

1. Add case "stream.usage" handler in TaskStreamReader.handleNotification()
2. Update latestInputTokens on each stream.usage notification
3. Trigger requestUiRender() so the status bar refreshes immediately

Files to modify

• aceclaw-daemon/src/main/java/dev/aceclaw/daemon/StreamingAgentHandler.java — emit stream.usage after each LLM call
• aceclaw-cli/src/main/java/dev/aceclaw/cli/TaskStreamReader.java — handle stream.usage
• aceclaw-cli/src/main/java/dev/aceclaw/cli/TerminalRepl.java — update latestInputTokens from notification