Claude Code v2.1.246 — Bash wildcard-allow startup warning, /permissions Auto-mode tab, completion-time echo, 60+ fixes

v2.1.246 released Aug 25 at 22:31. Added: a startup warning when a Bash allow rule has a wildcard before the subcommand (e.g. `Bash(git * main)`), since such rules also match options inserted before the subcommand; an Auto mode tab on `/permissions` for viewing and editing auto-mode classifier rules; the turn's completion time on the end-of-turn duration line (e.g. `Sautéed for 23s · done 6:05 PM`). Fixed (60+ items incl.): fullscreen blank transcript after resize-and-jump; severe transcript slowdown for very long single lines (now truncated with marker); background sessions failing to open after 45 s when CWD deleted / machine slept / slow host; background sessions failing with `EACCES` while another Claude Code was reinstalling the npm package; MCP tool calls interrupted in headless/remote sessions being reported `completed with no output` instead of an explicit interrupted error; MCP arguments being sent as JSON strings when the parameter schema is `{}` instead of their real type; plugin cache creating duplicate SHA-named directories; `/plugin:plugin:skill` doubling the prefix; `claude plugin update` failing for a bare plugin name; plugin install failing when `plugin.json` had a UTF-8 BOM; telemetry requests to Anthropic carrying the API key configured for a third-party gateway (credential now only sent to its own host); memory leak in fullscreen and Ctrl+O transcripts; Remote Control sessions keeping placeholder name in claude.ai until the second prompt; MCP tools marked `requiresUserInteraction` still offering a `don't ask again` option that wrote an ignored allow rule; Bash permission checks approving malformed commands with dangling `&&`/`||`. Improved: `/cd` so a new directory's project settings, hooks, `.mcp.json`, skills and agents take effect immediately; Bash tool latency by replaying snapshot functions without a base64 subshell per function; subagent that stops at its `maxTurns` limit now returns partial output with a `SendMessage` hint; non-interactive sessions (`-p`, SDK, cloud) auto-continue a response cut off mid-stream by server error, connection loss or stall; managed-settings consent prompts deferred to next interactive session.