OAuth profile: now requires bothaccount and organization fields before applying a fetched profile; previously any non-null result was accepted, potentially leaving org context empty. (search for "l?.account && l.organization")
Bash parse error recovery: when the tree-sitter parser returns an ERROR node wrapping a program node at the root level, the inner program node is now unwrapped before redirect analysis, preventing false-positive dangerous-redirect detections. (search for "n.type === \"ERROR\" && n.children[0]?.type === \"program\"")
PowerShell quoting: shell-quoting checks on Windows now use the full Windows quoting rules (backslash at end of string requiring special treatment), not just dollar-sign detection. (search for "SCo(c)" in Windows quoting path)
Background bash shell memory pressure reaping: Claude Code now registers a memoryPressure event listener on background shell processes when not running as a subagent, allowing the OS to signal when memory is tight. Idle, unnoticed shells are then reaped. (search for "task_local_shell_pressure_reap")
MCP needs-auth count is now surfaced in the app state (mcpNeedsAuthCount), enabling the UI to reflect pending MCP authentication requirements. (search for "mcpNeedsAuthCount")
Model picker now correctly filters and replaces deprecated model aliases with their first-party equivalents when the model is confirmed available. (search for "Ia(r.aliasModel)" in model picker)
--mcp-debug flag removed; functionality is covered by --debug. (search for removed "--mcp-debug" in string diff)
Background agent spawn: when resuming a previous session with --resume, the transcript path is now correctly passed through to the background agent launch parameters. (search for "resumeTranscriptPath")
HTML content extraction: the web fetch tool now also extracts <meta name="description"> content and <h1> headings for better URL titles when the <title> tag is missing or generic. (search for "<meta[^>]+name=" in HTML parser)
Error telemetry: errors from background spawn failures and console method failures now include a telemetryMessage field (via gh()) so they're correctly categorized in error tracking rather than grouped under generic labels. (search for "background spawn failed" or "console.${s} failed")
