Claude Code v2.1.73

Clipboard image detection and reading on macOS now uses a native module when available (gated by the tengu_collage_kaleidoscope flag), bypassing the slower osascript approach. When a clipboard image is detected, users see an "Image in clipboard · [key] to paste" notification. The native path reads PNG data directly, handles resizing, and falls back to the original osascript method if the native module is unavailable.

Plugin installation now resolves transitive dependencies. When installing a plugin that declares dependencies, Claude Code validates the entire dependency closure and reports issues clearly:

• Dependency cycles are detected and reported: "Dependency cycle: A → B → A"
• Cross-marketplace dependencies are blocked with an explanation
• Missing dependencies report which marketplace they expected to be in
• Installation messages now show the dependency count: "✓ Installed my-plugin (+ 2 dependencies)"
• Disabled dependencies are flagged: "disabled — enable it or remove the dependency"

The Bash command safety checker now correctly parses command wrappers that prepend the real command, preventing false safety denials when using common Unix utilities:

• timeout — parses --foreground, --kill-after, --preserve-status, --signal, and duration arguments
• stdbuf — parses -i, -o, -e buffer-mode flags
• env — parses environment variable assignments and -i, -0, -u, -v flags
• time, nohup, nice — unwrapped to find the actual command being run
• These wrappers are recursively stripped before analyzing the inner command for safety

The /output-style command has been deprecated. It now shows a message directing users to /config or the settings file. The command is hidden from help output but still works for backwards compatibility.

The effortLevel setting description was simplified from "Persisted effort level for supported models. \"max\" is session-scoped and not persisted." to just "Persisted effort level for supported models." — the session-scoped restriction on "max" has been lifted.

The system prompt for the SendUserMessage tool has been significantly condensed. The old version was a detailed multi-paragraph guide with examples. The new version is a concise description: "Send a message the user will read. Text outside this tool is visible in the detail view, but most won't open it — the answer lives here." The companion "Talking to the user" guide has similarly been streamlined to focus on the key pattern: replies go through SendUserMessage; text outside it should be assumed unread.

Team memory synchronization now supports hash-based conflict resolution with entryChecksums. Instead of pulling and re-pushing the entire memory file on conflict, the system probes the server for per-entry hashes (?view=hashes), enabling delta-based pushes. This reduces bandwidth and merge conflicts for organizations with large shared memory files.

Claude Code now queries the terminal emulator for its identity using the XTVERSION escape sequence. This can be used to tailor behavior to specific terminals. If the terminal ignores the query, it logs "XTVERSION: no reply".

Plan mode now supports multiple verbosity levels for Phase 4 (Final Plan) output, controlled by the tengu_pewter_ledger feature flag. Options include "trim" (concise), "cut" (minimal, no context section), and "cap" (hard 40-line limit, no prose). The default behavior is unchanged.

A new metadata display now shows output token usage per turn and per session: "Output tokens — turn: X · session: Y". This gives users better visibility into how many tokens each response consumes.

• Orphaned bash tasks from exiting agents are now killed automatically, preventing zombie processes from accumulating (search for "killBashTasksForAgent: killing orphaned bash task")
• Voice mode now handles early pre-transcript errors gracefully by retrying once, and ignores stale onError callbacks from previous sessions (search for "early voice_stream error (pre-transcript), retrying once", "ignoring onError from stale session")
• OAuth token expiration during bridge/REPL sessions now shows a clear message directing users to run /login (search for "OAuth token expired and could not be refreshed. Run /login to re-authenticate.")
• Bridge REPL initialization now tracks consecutive failures and stops retrying after a threshold, preventing infinite retry loops (search for "consecutive init failures, not retrying this session")
• Hook success content changed from "Success" / "Condition met" to empty string, reducing noise in hook output
• SSL certificate errors now provide actionable guidance: set NODE_EXTRA_CA_CERTS or ask IT to allowlist *.anthropic.com (search for "SSL certificate error")
• Cron job deletion errors now provide clear messages: "Cannot delete cron job '...'" (search for "Cannot delete cron job")
• Orphaned cron jobs (for agents that no longer exist) are now automatically cleaned up (search for "gone, removing orphaned cron")
• Durable cron jobs are now explicitly blocked for teammate sessions with a clear error message (search for "durable crons are not supported for teammates")
• Environment variables related to ANTHROPIC_UNIX_SOCKET connections are now stripped when applying config-level env overrides, preventing socket connection leaks across configurations (search for "ANTHROPIC_UNIX_SOCKET" in Rr6())
• Text selection in the terminal now uses word-boundary-aware double-click selection with proper anchor span tracking (search for "anchorSpan" with kind: "word")