story-kit: start 86_story_show_live_activity_status_instead_of_static_thinking_indicator_in_chat

.story_kit/work/2_current/86_story_show_live_activity_status_instead_of_static_thinking_indicator_in_chat.md

@@ -0,0 +1,34 @@
+---
+name: "Show live activity status instead of static thinking indicator in chat"
+---
+
+# Story 86: Show live activity status instead of static thinking indicator in chat
+
+## User Story
+
+As a user chatting with Claude in the web UI, I want to see what the agent is actually doing (reading files, running commands, etc.) instead of a static "thinking" message, so the application feels responsive and I can follow along.
+
+## Acceptance Criteria
+
+- [ ] When the agent is executing a tool call, the thinking indicator updates to show the tool name and a short summary (e.g. "Reading file: src/main.rs", "Running command: cargo test")
+- [ ] The status updates in place without adding new chat bubbles
+- [ ] When the agent is between tool calls (actually thinking), it shows "Thinking..."
+- [ ] The status text is derived from the streaming events already coming over the WebSocket
+- [ ] No new server-side changes needed if the stream already contains tool use events; otherwise add a lightweight status event type
+- [ ] Activity indicators must work for Claude Code PTY sessions (used by agents), not just API-based LLM providers. The Claude Code provider must parse tool activity from the stream-json output and forward it via the same on_activity callback used by the Anthropic provider.
+
+## Notes
+
+Previous implementation (now reverted to upcoming) only wired up activity events for the Anthropic API provider. The Claude Code provider in `claude_code.rs` was left with no `on_activity` callback — it only streams text tokens. Since agents use Claude Code, the web UI showed no activity during agent work.
+
+### Implementation hints for the Claude Code provider
+
+1. Add an `on_activity` callback parameter to `ClaudeCodeProvider::chat_stream()` (matching the Anthropic provider signature)
+2. In `handle_stream_event()`, parse tool names from Claude Code's stream-json output (tool activity events are already in the stream, just currently ignored — only text deltas are extracted)
+3. Call the `on_activity` callback when a tool use event is detected
+4. In `chat.rs`, pass the `on_activity` closure to the Claude Code path (same as the Anthropic path already does)
+5. `ws.rs` will then send `ToolActivity` WebSocket messages to the frontend — no frontend changes needed
+
+## Out of Scope
+
+- TBD