Code Review gained a new Manual trigger mode and documented @claude review comment behavior, allowing teams to opt specific PRs into review on demand
MCP tool search is now enabled by default for all sessions (previously activated only at 10% context threshold), with automatic disabling for non-first-party API hosts
Subagents can now scope MCP servers to only that subagent via the mcpServers field, keeping server tool descriptions out of the main conversation entirely
SessionEnd hooks now have a 1.5-second default timeout configurable via CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS
The commands/ directory format is now legacy in the Agent SDK — skills/ is the recommended format for new plugins
Added minimum version requirement: agent teams require Claude Code v2.1.32 or later. [line 8] [Source]
Added clarifying sentence distinguishing subagents from agent teams: "Subagents only report results back to the main agent and never talk to each other. In agent teams, teammates share a task list, claim work, and communicate directly with each other." [line 34] [Source]
Removed the "Safe autonomous mode" section, which documented --dangerously-skip-permissions, its risks, and guidance about running it in a sandboxed container. The flag itself is still documented in the CLI reference.
Added a new Manual review trigger mode: reviews only start when someone comments @claude review on a PR; subsequent pushes to that PR are then reviewed automatically. [line 79] [Source]
Added a new "Manually trigger reviews" section documenting that @claude review must be a top-level PR comment starting with that phrase, commenter needs owner/member/collaborator access, and the PR must be open and non-draft. [lines 86-92] [Source]
Code Review usage is now explicitly documented as billed via "extra usage" and does not count against a plan's included usage. [line 162] [Source]
The repository config UI dropdown is now named Review Behavior. [line 75] [Source]
SessionEnd hooks now have a documented default timeout of 1.5 seconds (applies to both session exit and /clear). Users can increase this via the CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS environment variable. Per-hook timeout values are capped by this budget. [line 1806] [Source]
/context command description expanded: now shows "optimization suggestions for context-heavy tools, memory bloat, and capacity warnings" in addition to the colored grid. [line 86] [Source]
Removed the note that setting CLAUDE_CODE_ENABLE_TASKS=false reverts to the previous TODO list.
Added minimum version requirement: customizable keyboard shortcuts require Claude Code v2.1.18 or later. [line 3] [Source]
Added Ctrl+P as a default keybinding for "move up in list" (messageSelector:up) and Ctrl+N for "move down in list" (messageSelector:down). [lines 226-227] [Source]
MCP tool search is now enabled by default for all sessions. Previously it activated automatically at 10% context threshold (auto mode). Now: enabled by default, but automatically disabled when ANTHROPIC_BASE_URL points to a non-first-party host. Setting ENABLE_TOOL_SEARCH=true forces it on even for proxies. [lines 1785-1791] [Source]
Added minimum version requirement: auto memory requires Claude Code v2.1.59 or later. [line 284] [Source]
Added documentation for the new autoMemoryDirectory setting: users can configure a custom directory for auto memory storage. This setting is not accepted from project settings to prevent shared repos from redirecting memory writes to sensitive locations. [lines 307-317] [Source]
Relative plugin paths (starting with ./) must resolve within the marketplace root; using ../ to climb out is explicitly disallowed. [lines 267-300] [Source]
The url field no longer requires a .git suffix — Azure DevOps and AWS CodeCommit URLs without the suffix now work. [line 386] [Source]
Clarified that strictKnownMarketplaces restricts what users can add but does not pre-register marketplaces; use it with extraKnownMarketplaces to both restrict and pre-register. [line 892] [Source]
When --plugin-dir and a marketplace plugin share the same name, the local --plugin-dir copy now takes precedence for that session. Exception: marketplace plugins force-enabled by managed settings cannot be overridden. [line 354] [Source]
cleanupPeriodDays: 0 is now documented to delete all existing transcripts at startup and disable session persistence entirely — no new .jsonl files are written, /resume shows no conversations, and hooks receive an empty transcript_path. [line 143] [Source]
New autoMemoryDirectory setting documented: custom directory for auto memory storage, accepts ~/-expanded paths, not accepted from project settings. [line 142] [Source]
New CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS environment variable: sets maximum time in milliseconds for SessionEnd hooks (default: 1500ms). [line 993] [Source]
CLAUDE_CODE_ENABLE_TASKS behavior changed: setting to true now enables the task system in non-interactive mode (-p flag); tasks are on by default in interactive mode. Previously false reverted to the old TODO list. [line 980] [Source]
ENABLE_TOOL_SEARCH documentation updated to reflect new default (enabled unless ANTHROPIC_BASE_URL is a non-first-party host), plus true to force on for proxies. [line 1021] [Source]
Added a managed-settings.json example combining strictKnownMarketplaces and extraKnownMarketplaces to both restrict and pre-register a marketplace. [line 892] [Source]
The model field now accepts full model IDs (e.g., claude-opus-4-6 or claude-sonnet-4-6) in addition to short aliases; explicitly noted to accept the same values as the --model flag. [line 233] [Source]
New section: Scope MCP servers to a subagent — the mcpServers field in subagent definitions allows scoping MCP servers to a specific subagent. Inline servers connect when the subagent starts and disconnect when it finishes; string references reuse the parent session's connection. Defining a server inline keeps it out of the main conversation's context entirely. [lines 311-340] [Source]
The commands/ directory is now documented as a legacy format; skills/ is the recommended format for new plugins. Both formats continue to work for backward compatibility. [line 16] [Source]
Plugin directory structure updated to show skills/ first and label commands/ as "Legacy: use skills/ instead." [line 150] [Source]
Troubleshooting section renamed from "Commands not available" to "Skills not appearing"; instructions updated to require skill files in subdirectories under skills/ (e.g., skills/my-skill/SKILL.md). [lines 205-209] [Source]
Four new MCP servers added to the directory: Tavily (https://mcp.tavily.com/mcp), Calendly (https://mcp.calendly.com), Metaview (https://mcp.metaview.ai/mcp), and Brex (https://api.brex.com/mcp). [lines 1057-1111] [Source]
Added a note that .claude/commands/ is the legacy format; the recommended format is now .claude/skills/<name>/SKILL.md, which supports both slash-command invocation and autonomous invocation by Claude. [line 95] [Source]
Personal and project command directory descriptions updated to note "(legacy; prefer .claude/skills/)". [lines 101-102] [Source]