Stray 'court' + raw <invoke> tool-call leak still present in v2.1.198 with Fable 5 (Bash/Edit/Write affected)

Summary

The known "stray court + raw <invoke> XML rendered as text" bug (#64108) is still present in the current release. Filing as a version/model data point: it now reproduces on v2.1.198 with Fable 5 (claude-fable-5), and it affects Bash and Write calls in addition to the Edit/Read calls reported earlier.

Instead of dispatching the tool, the assistant turn renders the serialized call verbatim in the transcript, always prefixed by a lone court token:

court
<invoke name="Edit">
<parameter name="replace_all">false</parameter>
<parameter name="file_path">~/GitHub/<project>/docs/architecture/core.md</parameter>
<parameter name="old_string">- HuggingFaceModelDownloader.swift - downloads model files ...</parameter>
<parameter name="new_string">- HuggingFaceModelDownloader.swift - downloads model files ...</parameter>
</invoke>

The tool does not execute. On the next turn the model notices on its own and re-issues the call correctly ("repeating the edit with a correct call"), so the session limps along, but every occurrence costs a wasted turn plus the re-serialized (often large) parameters.

Environment

• Claude Code CLI v2.1.198, macOS (Darwin 25.5)
• Model: Fable 5 (claude-fable-5), effort max, Claude Max plan
• Long session (hours, large context); one safeguards-triggered auto-switch to Opus 4.8 happened mid-session, so at least one occurrence may be attributable to Opus 4.8

Frequency / pattern in one session

4 occurrences in a single session, spread across tool types:

1. Bash (a swift test verification command; the session was restarted before it was re-issued, so the verification silently never ran)
2. Edit (docs update)
3. Edit (docs update)
4. Write (new shell script, large content parameter)

Same signature every time: a line containing only court, then the raw <invoke>...</invoke> block as plain text. No partial execution observed on these four (unlike #64108, where the tool sometimes still ran).

Impact

• Wasted turns and tokens (large Edit/Write payloads get emitted twice).
• Silent loss of a verification step when the user restarts the session before the model re-issues the call.
• Transcript noise; looks like the agent is malfunctioning mid-task.

Related

Same signature reported in #64108 (canonical, Opus, CLI), #66153, #67331, #68354 (Windows + Cowork), #71487, #72641 (Desktop, v2.1.193). This report adds: still reproducible on v2.1.198, reproducible with Fable 5, and Bash/Write are affected too.

Redacted transcript excerpts available on request.