Claude Code v2.0.28

What: Marketplaces can now be installed from specific branches, tags, or commit refs instead of only the default branch.

How to use:

# Install from a specific branch
claude marketplace add owner/repo#develop

# Install from a tag
claude marketplace add owner/repo#v1.2.3

# Install from a commit hash
claude marketplace add [email protected]:owner/repo.git#abc123

Details:

• Use #ref syntax after the repository name or URL to specify a branch, tag, or commit
• Updates to existing marketplaces now respect the specified ref
• Git clone operations include the --branch flag when a ref is specified (Xd8() at line 264036, b5Q() at line 264043, Cd8() at line 264049)
• Git update operations fetch and checkout the specified ref (Bd8() at line 263476)
• Clone and update logging now shows which ref is being used
• Evidence: New functions Xd8() at line 264036, Bd8() at line 263476, b5Q() at line 264043, Cd8() at line 264049; marketplace parsing extracts ref from URLs at line 263078-263135

What: Claude now tests SSH authentication before attempting to clone GitHub repositories, avoiding unnecessary failed clone attempts for users without SSH keys configured.

How it works:

# When you add a marketplace, Claude now checks SSH first
claude marketplace add owner/repo

# If SSH is configured: "Cloning via SSH: [email protected]:owner/repo.git"
# If SSH is NOT configured: "SSH not configured, cloning via HTTPS: https://github.com/owner/repo"

Details:

• Executes ssh -T [email protected] to verify authentication before cloning
• Uses BatchMode=yes to prevent interactive prompts
• 2-second connect timeout and 3-second overall timeout for fast checks
• Automatically falls back to HTTPS without a failed clone attempt
• Provides clearer logging about SSH status
• Evidence: New function Qd8() at line 263519; usage in marketplace cloning logic at lines 263717-263750

What: Agents can now resume from previous execution transcripts, allowing you to continue where a previous agent left off.

How to use:

# The Agent tool now accepts a 'resume' parameter
# When invoking an agent programmatically:
{
  "subagent_type": "investigator",
  "prompt": "Continue analyzing the codebase",
  "resume": "agent-abc123"  // ID of previous agent execution
}

Details:

• New resume parameter in Agent tool input schema
• Loads previous agent's message chain from transcript file
• Filters messages by agent ID and reconstructs conversation history
• Prepends previous messages to new prompt for continuity
• Previous messages are not re-recorded to session storage
• Evidence: New function b_2() at line 465741; resume parameter schema at line 449027-449032; resume logic at lines 449114-449130

What: Commands ending with --help now bypass expensive API calls for prefix detection, significantly improving response time.

Details:

• Validates help commands using pattern matching before API calls
• Checks that command ends with --help and contains only valid flags
• Ensures no quotes are present (which would complicate parsing)
• Validates all tokens are alphanumeric or the --help flag
• Immediately returns the full command as its own prefix
• Evidence: New function Yq8() at line 204259; usage at line 204549

What: Removing a marketplace now automatically cleans up all references from settings files, preventing orphaned configuration entries.

How it works:

# When you remove a marketplace:
claude marketplace remove my-marketplace

# Now also removes:
# - extraKnownMarketplaces.my-marketplace from all settings files
# - All enabledPlugins entries like "plugin-name@my-marketplace"
# - Cleans userSettings, projectSettings, and localSettings

Details:

• Iterates through all settings files (user, project, local)
• Removes marketplace from extraKnownMarketplaces
• Removes all plugins associated with the marketplace from enabledPlugins
• Logs cleanup actions for each settings file
• Handles errors gracefully if settings write fails
• Evidence: Enhanced function ruA() at line 263832 (replaced suA() from v2.0.27); cleanup loop at lines 263842-263868

What: Enhanced security check now detects paths with 3+ consecutive dots (e.g., ..., ....) as suspicious Windows path patterns.

Why this matters: Attackers sometimes use unusual dot sequences to bypass basic .. filters. While ... isn't standard, blocking it prevents potential exploits in path normalization edge cases.

Details:

• Adds regex check /\.{3,}/ to Windows path validation
• Blocks paths like C:\Users\....\file.txt or path/to/.../folder
• Triggers manual approval prompt: "contains a suspicious Windows path pattern"
• Complements existing checks for reserved names, alternate data streams, and other Windows-specific attacks
• Evidence: Modified function tf2() at line 465855 (renamed from mf2()); new check at line 465867

What: Git clone and update operations now use 30-second timeouts instead of 5 seconds, improving reliability for large repositories or slow connections.

Details:

• Clone operations: timeout increased from 5000ms to 30000ms
• Update operations: timeout increased from 5000ms to 30000ms
• Reduces false failures on slow networks or large marketplaces
• Evidence: Id8() at line 263544 uses timeout: 30000; Bd8() at line 263476 uses timeout: 30000

What: Improved detection of when terminal output needs to be redrawn, accounting for output height relative to terminal rows.

Details:

• Now considers whether output height exceeds terminal rows
• Checks both previous and current state for overflow conditions
• Prevents visual artifacts when output is taller than terminal
• Evidence: Enhanced function OL0() at line 70043 (was wL0() in v2.0.27); checks B.outputHeight >= B.rows at line 70046

What: Session state now initializes with the actual current working directory instead of placeholder values.

Details:

• Calls Uo2() (likely process.cwd()) to get actual working directory
• Normalizes the path using zo2()
• Sets both originalCwd and cwd to the correct value at session start
• Evidence: Modified function wo2() at line 3444 (was Yo2() in v2.0.27); uses zo2(Uo2()) at line 3446

What: The permission stream class now tracks pending permission requests, enabling better state management and UI updates.

Details:

• New method getPendingPermissionRequests() on the stream class
• Returns array of pending can_use_tool requests
• Filters request map to find active permission prompts
• Stores request metadata alongside resolve/reject callbacks
• Evidence: Enhanced class dzA at line 475314 (was hzA in v2.0.27); new method getPendingPermissionRequests() at line 475580; request storage at line 475650

What: Agent execution results now include the agent ID for better tracking and correlation.

Details:

• Agent tool results now return agentId field along with content and usage stats
• Enables linking results back to specific agent invocations
• Supports transcript resumption and execution tracking
• Evidence: Modified function Bv6() at line 448963 (was dx6() in v2.0.27); returns agentId: B at line 449019

What: Agent model selection now supports an additional override parameter for finer-grained control.

Details:

• New third parameter Q in model resolution function
• If override is provided, it takes precedence over agent-specified model
• Falls back to agent model, then "inherit", then default
• Environment variable CLAUDE_CODE_SUBAGENT_MODEL still has highest priority
• Evidence: Modified function akA() at line 144682 (was lkA() in v2.0.27); uses override parameter at line 144693

What: API requests now respect a token budget cap, adjusting both max_tokens and thinking budget_tokens when needed.

Details:

• New function ensures total tokens don't exceed specified budget
• Reduces thinking budget proportionally when capping is needed
• Prevents thinking budget from consuming all available tokens
• Evidence: New function pb6() at line 460609

What: Plugin SessionStart hooks now receive session context information for better initialization.

Details:

• Hooks receive session metadata when invoked
• Enables plugins to customize behavior based on session state
• Evidence: Modified function Lz() at line 317267 (was qz() in v2.0.27); passes session context parameter B at line 317350

What: Plugin commands and skills now have access to session ID during processing.

Details:

• Session ID passed through to command/skill creation functions
• Enables session-aware plugin behavior
• Supports plugin state management across invocations
• Evidence: Modified functions ZR2() at line 441792, YR2() at line 441892, yUA() at line 441803 now accept session ID parameter

What: File write permission checks now use a centralized safety validation function for consistency.

Details:

• New SL1() function provides unified safety checks for write operations
• Checks for suspicious Windows paths, sensitive files, and permission rules
• Returns structured result with safety status and message
• Used by both direct file operations and tool permission handlers
• Evidence: New function SL1() at line 465870; used in jL1() at line 466119 and xn() at line 466119

What: Added support for custom ripgrep binary location via environment variable.

Details:

• New RIPGREP_NODE_PATH environment variable supported
• If set, loads ripgrep from specified path instead of bundled version
• Useful for custom deployments or testing
• Falls back to bundled ripgrep if not set
• Evidence: Modified function To2() at line 3881 (was Eo2() in v2.0.27); checks process.env.RIPGREP_NODE_PATH at line 3882

• Session ID generation functions consolidated
• Permission stream class methods reorganized
• Skills loading function signatures updated for session context
• Git clone/update error handling refactored into shared R5Q() function at line 263494

• Removed unused stream import
• Removed unused node:child_process import
• Removed unused node:process import
• Added targeted imports: execSync, spawn, readdir, readFile, PassThrough, homedir, openSync, existsSync, readFileSync, writeFileSync, dg6, cg6, pg6
• Added process.cwd import: import { cwd as iL0 } from "node:process" at line 70611

• Path validation logic consolidated into reusable SL1() function
• Git error handling extracted to R5Q() helper
• Agent ID generation moved to dedicated HrA() function at line 393160
• Transcript path functions reorganized (bf2() at line 465227, PU() at line 465250)

Note: This changelog focuses on user-visible changes and significant internal improvements. Many variable and function names appear random due to the code being previously minified. Line numbers reference the prettified v2.0.28 source file.