Skills and plugins enabled on your claude.ai account now sync automatically into terminal sessions when you're signed in, no longer requiring non-interactive mode or an env var for skills, and extending plugin sync beyond Cowork/cloud sessions for the first time
2.1.274 adds a critical-memory warning with recovery steps, and fixes several stuck-session bugs: endless "unexpected tool_use_id" retries, a lost /goal after resuming a compacted session, and hook-driven sessions failing with "Prompt is too long" instead of compacting
Marketplace and plugin git clones no longer download Git LFS content at all — files arrive as pointers, and the skipLfs field is now a no-op
The Compliance API's local session endpoints now cover Claude in Chrome sessions (claude_in_chrome), in beta for Claude Enterprise organizations
The Agent SDK's Python system_prompt gains an object form (SystemPromptCustom) with snapshot control, and hooks can now inject mid-session instruction updates via additionalContext without touching the recorded system prompt
Added a visible warning when memory usage is critical, with steps to free memory or restart safely
Added CLAUDE_CODE_MCP_STARTUP_WAIT_MS to bound how long the first non-interactive turn waits for connecting MCP servers (0 = don't wait)
Added click-to-expand for collapsed teammate and agent messages in fullscreen mode
Added claude_code.managed_settings_resolved OTel event, and an effort attribute on the claude_code.llm_request trace span
[VSCode] Added continuation of the step a window reload interrupted, with a Claude Code: Continue After Reload setting to turn it off
[VSCode] Added Memory and Instructions entries to the Customize menu, and a claudeCode.lockEditorGroups setting
[Claude Code on the web] Added a "Compare against" branch picker to a cloud session's diff view
[Claude Tag] Added a Guests setting to the Add channel and Add workspace forms in admin settings
Existing feature improvements
Changed Bedrock, Vertex, Foundry and telemetry-disabled installs to use the v2 MCP client and MCP 2026-07-28 negotiation with direct HTTP servers by default
Changed marketplace and plugin repository clones to leave Git LFS files as pointers instead of downloading them
Changed /code-review to use leaner inline review prompts instead of spawning many review subagents
Improved OTEL_LOG_RAW_API_BODIES=file:<dir> output with a new index.jsonl linking each response to its request file and transcript message
Improved Claude apps gateway boot (retries the first Postgres connection) and its spend-limit check (one database round trip instead of four)
Improved headless and SDK sessions to answer already-queued background-task completions with a single model call instead of one per task
[VSCode] Improved screen reader navigation of the conversation
Major bug fixes
Fixed sessions getting stuck endlessly retrying "unexpected tool_use_id" 400 errors; corrupted transcripts now self-heal where possible
Fixed hook-driven sessions (such as an active /goal) ending with "Prompt is too long" instead of compacting, and fixed an active /goal being lost when resuming a compacted session
Fixed MCP servers configured as http that only speak legacy HTTP+SSE failing to connect on a 422 or other 4xx first response
Fixed Streamable HTTP MCP tool calls timing out after about 5 minutes even with a longer per-server timeout
Fixed MCP tool calls refused with 403 insufficient_scope being reported as an expired sign-in instead of naming the missing permissions
Fixed claude agents losing --model, --effort, --permission-mode, and other flags after an auto-update relaunch
Fixed subagents with model: "opus" on Bedrock, Vertex, or Foundry leaving the session's model when its id has no recognizable model family
Fixed self-hosted runner sessions failing every turn with a 401 after a few failed token refreshes
Fixed a per-turn slowdown when a language server publishes project-wide diagnostics for thousands of files
Fixed the Bash tool re-sourcing the shell profile after every plugin reload instead of only when plugins' bin/ directories changed
Fixed Bash permission checks and worktree-isolated session checks for certain special shell variables and nested shell expansions
Fixed background commands being stopped after 30 idle minutes under mild memory pressure instead of only when memory is critically low
Added a "Disable claude.ai sync" managed-settings row: syncClaudeAiSkills/syncClaudeAiPlugins stop skill/plugin syncing, and turning off Skills for the org on claude.ai now also removes already-synced copies on v2.1.273+. [line 92] [Source]
New "Update Claude's instructions mid-session" section: send new instructions in the next user message, or return additionalContext from a UserPromptSubmit/PostToolUse hook, instead of changing systemPrompt. [line 346] [Source]
Clarified that turning snapshot off to iterate on prompt wording is now documented for both TypeScript and Python, and that recording defaults to on outside cloud sessions and bare mode. [line 353] [Source]
Added SystemPromptCustom, an object form of system_prompt ({"type": "custom", "prompt": ..., "snapshot": ...}) equivalent to a plain string but with snapshot control; requires claude-agent-sdk v0.2.153+. [lines 726-893] [Source]
Documented that Claude Code terminal sessions now also load the skills and plugins enabled on your claude.ai account when signed in with the same account. [line 389] [Source]
Reworded the skills section: local sessions load ~/.claude/skills/ and also sync claude.ai skills; cloud sessions load synced skills instead. [line 409] [Source]
Noted that plugins enabled on claude.ai sync into sessions without a marketplace install. [line 7] [Source]
The plugin menu's Installed tab now lists synced plugins with a synced source, manageable there unless required by the org; available in terminal sessions on v2.1.273+. [line 325] [Source]
/reload-plugins now also fires after a claude.ai sync adds, updates, or removes a plugin. [line 397] [Source]
CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE now describes skipping the re-clone when a marketplace refresh can't reach or authenticate to the remote, not just on a failed git pull. [line 319] [Source]
CLAUDE_CODE_SYNC_SKILLS and CLAUDE_CODE_SYNC_SKILLS_INSTALL_TIMEOUT_MS rewritten: terminal sessions signed in with claude.ai now sync skills automatically (resyncing every ~10 minutes) without needing this variable; it's now only for making a -p run wait for the current list on its first query. [line 364] [Source]
MCP_PROTOCOL_NEGOTIATION/MCP_SDK_GENERATION: without the variable, Claude Code now also probes claude.ai connector servers and defaults to the v2 runtime in more sessions, tied to feature-flag fetching rather than a fixed list of platforms. [lines 444-446] [Source]
With feature-flag fetching off, sessions can no longer sync claude.ai skills/plugins into the terminal, and only lose the connector-server protocol probe rather than the whole v2 MCP runtime. [lines 503-506] [Source]
Documented the bound to a different conversation thinking-signature rejection: it comes from the API's preserved-thinking check and can be caused by a gateway that rewrites system, tools, or earlier messages content. [lines 145-148] [Source]
The v1/v2 MCP client runtime choice is now framed around whether the session fetches feature flags rather than a fixed platform list, and on v2.1.274+ the v2 runtime becomes the default even in sessions that don't fetch feature flags (except Bedrock/AWS/GCP/Foundry, gateway sign-ins, and telemetry-off sessions). [line 315] [Source]
Documented the new tool.output span event in detail: which tools/conditions produce it, its attributes (content, output, diff, file_path, bash_command), and their individual gates. [line 263] [Source]
Added request_body_id on api_request_body/api_response_body events, and message.id/message.uuid on the response event, to pair requests, responses, and transcript messages; also documents the new index.jsonl file for file:<dir> mode. [lines 820-845] [Source]
Marketplace and plugin repository clones now never download Git LFS content; LFS-tracked files always arrive as pointer files. [line 738] [Source]
Background auto-update mechanics changed: instead of git pull, Claude Code checks the remote for new commits and, if any are found (or the check can't authenticate), re-clones and swaps in the new checkout, leaving the existing one in place on failure. [line 770] [Source]
Updated the offline/airgapped troubleshooting entry to match the new check-then-reclone behavior (previously described as a failing git pull). [line 1452] [Source]
Major rewrite of "Plugins synced from claude.ai": terminal sessions signed in with a claude.ai account now sync enabled plugins automatically in the background (v2.1.273+), not just Cowork/cloud sessions; covers required-by-org plugins, /reload-plugins prompts, and the new syncClaudeAiPlugins opt-out. [line 409] [Source]
Documents that a required org plugin can't be disabled via claude plugin disable, with the exact refusal message. [line 429] [Source]
Plugins are now specified in three ways instead of two, adding claude.ai account sync alongside --plugin-dir/marketplace. [line 799] [Source]
syncClaudeAiSkills reworded to reflect automatic syncing in signed-in terminal sessions rather than requiring CLAUDE_CODE_SYNC_SKILLS in non-interactive mode.
The skipLfs marketplace-source field is now a no-op, since LFS content is never downloaded on clone/update as of v2.1.274. [line 4480] [Source]
Major rewrite of "Skills synced from claude.ai": terminal sessions signed in with claude.ai now sync skills automatically at startup and resync every ~10 minutes, replacing the old CLAUDE_CODE_SYNC_SKILLS + non-interactive-only workflow. Lists the specific sessions where syncing doesn't happen (API-key auth, no feature-flag fetching, bare/safe mode, plugin-lockdown policy). [line 185] [Source]
Documents the new syncClaudeAiSkills: false opt-out and that skills already synced move to ~/.claude/skills/.trash/. [line 189] [Source]
Added a fourth case that blocks worktree creation: a repository git config setting that points Git LFS at an untrusted external program (lfs.customtransfer.*, lfs.standalonetransferagent), with a new Git was not run error and fix guidance. [line 309] [Source]
Compliance API coverage now includes Claude in Chrome sessions alongside Cowork, Claude Code, Claude Science, and Claude for Microsoft 365. [lines 9-58] [Source]
Added Claude in Chrome to the local-session FAQ answer and noted its coverage is in beta, alongside Claude Science and Claude for Microsoft 365. [lines 71-75] [Source]
Added Claude in Chrome as a new local-session product surface (claude_in_chrome), including its row in the product/endpoint table and its product_surface value in the API reference. [lines 9-123] [Source]