Streaming tool_use leaks raw invoke syntax ('court' + <invoke>) into chat and hangs on 'invoking' (Desktop app, macOS, v2.1.193)

Summary

In the Claude Code desktop app (chat UI), assistant turns that emit tool calls intermittently leak the raw tool-invocation syntax into the visible chat as plain text — a stray token court immediately followed by <invoke name="..."> — and the turn then hangs showing "invoking" / a running-agent spinner but never completes and no tool actually runs.

Environment

• Claude Code desktop app on macOS (Darwin 25.3.0)
• App/CLI version: 2.1.193
• Model: Claude Opus 4.8
• Setup: many MCP connectors enabled (Gmail/Drive/Slack/Calendar/freee/davinci-resolve/computer-use/github, etc.), ~40 skills, SessionStart/UserPromptSubmit hooks injecting context each turn.

Symptom

• Visible artifact: the literal text court appears, followed by leaked <invoke ...> / <parameter ...> XML, in the rendered assistant message.
• It correlates with the boundary where the assistant transitions from prose to a tool call, and especially with Agent/Task (subagent) delegation calls.
• Two severities observed:
• (A) cosmetic — court shows but the tool still executes and returns results.
• (B) destructive — court shows, the UI stays on "invoking"/"agent running", and the turn hangs permanently with zero result (the tool never runs). This recurs repeatedly ("連発").

Reproduction notes

• Intermittent (not every turn), so a single attempt does not reliably reproduce.
• Occurs after app restart and in fresh/lighter sessions, so it is not solely caused by a full context window.

Mitigations tried (reduced frequency but did NOT eliminate)

• Set CLAUDE_CODE_MAX_OUTPUT_TOKENS=65536 in ~/.claude/settings.json env.
• Trimmed always-injected context (CLAUDE.md ~20% smaller, memory index shortened).
• Reduced permission allowlist.
• Destructive (B) hangs still occur in separate threads after these changes.

Impact

Makes the desktop app unreliable for tool-heavy / subagent-delegation workflows: the user cannot tell whether work ran, and turns silently produce nothing.

Questions

1. Is this a known client-side streaming/tool_use rendering bug (raw invoke syntax surfacing as text, then hang)?
2. Is there a setting or app channel/version that avoids it?
3. Does the terminal CLI / IDE extension render streamed tool_use through a different path that is not affected?