Claude Code 2.1.319 adds a permissions.denyInsteadOfPrompting setting, background handling of MCP tasks, and a /claude-api latency-optimize command. It also fixes several permission bypasses (Bash env prefix, sed, worktree sandbox escape).
Hooks: Bash matchers now also fire for Monitor commands on PreToolUse and PermissionRequest (deny only, no allow). PreCompact hook stdout now becomes instructions for the compaction summary. There is a new auth_storage_failure notification type.
Claude Code no longer installs or updates the IDE extension when claude starts in a VS Code, Cursor, or Devin Desktop terminal (v2.1.318). Only /ide from an external terminal installs it.
Code Review: subdirectory REVIEW.md files are ignored under test, example, generated, and vendor-style directories. The remaining ones share a 150,000-byte limit.
Claude Projects gains an "Auto-continue when usage limits reset" setting. Claude apps gateway docs add load-test mode notes and spend-limit management in Claude Desktop.
Added the permissions.denyInsteadOfPrompting setting: a tool call that would wait on a permission prompt is denied and the turn continues (terminal sessions, subagents, teammates)
MCP tasks now work in the terminal: a tool call that runs as a task moves to the background right away, and Claude is told the result when it finishes. Set CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND to wait inside the tool call, or CLAUDE_CODE_DISABLE_MCP_TASKS=1 to turn tasks off
Added /claude-api latency-optimize to find where your Claude API code makes people wait and try faster settings against your eval
Added the CLAUDE_CODE_DISABLE_FEATURE_FLAGS environment variable, which stops feature-flag fetching without turning off telemetry
Added an "Earlier in this session" row to /tasks listing finished subagents, so you can reopen and message them
Added a warning and a startup line when Claude Code runs on an ant auth login sign-in that bills to a Claude Console organization
Claude apps gateway: added admin.pooled_conflict_rule and admin.pooled_priority_order, an optional login_notice, CLAUDE_GATEWAY_LOG_FORMAT=json, spend for today, this week and this month in /usage, and several new startup warnings
Existing feature improvements
Auto mode behind LLM gateways or with telemetry off now asks or denies calls the server can't check instead of using the built-in classifier. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep the old behavior
/model picker groups older models under a "More models" row
/context keeps the table Claude reads short and no longer re-sends its visual grid every turn
PostToolUse hooks now run, with up to 10 seconds, for tool calls that finish after a new message interrupts the turn
Monitor's WebSocket watch now follows deny and ask rules on every Bash command
Self-paced /loop ends when a turn finishes without rescheduling. Set CLAUDE_CODE_LOOP_KEEPALIVE=1 for the old behavior
Bash now asks before running which with flags outside a lookup/display set
Nested skills under .claude/skills below the start folder load only if you start or /cd there when the Bash sandbox is on
Removed shortcuts_list and shortcuts_execute from the built-in Claude in Chrome tools
Improved contrast of faint secondary text, config save reliability on Windows, and Claude in Chrome reconnect behavior
Major bug fixes
Fixed a Bash permission bypass where a crafted environment prefix could hide part of a command from permission checks
Fixed a sed command skipping the permission check in accept-edits and auto mode
Fixed Bash deny and ask rules missing commands with quoted words
Fixed a sandbox escape when creating a worktree
Fixed /rewind restoring or deleting a file through a directory symlink
Fixed claude -p exiting 0 in auto mode when actions were refused for lack of a safety check
Fixed ~/.claude.json, settings, and edited files ending up empty or truncated when the disk fills during a save
Fixed MCP servers on the older handshake failing to connect or returning empty results
Fixed slash commands typed after a leading space, tab, or newline being sent to Claude as plain text
Fixed /diff failing when a repository has a file or folder named HEAD, and opening blank after 64 or more edit turns
Fixed /clear killing running commands of background agents, teammates, and workflows
On v2.1.314 or later, Bash matcher callbacks also run on PreToolUse and PermissionRequest for Monitor commands, which they can deny but not approve. [line 222] [Source]
New "Reload plugins in a running session" section: TypeScript Query.reloadPlugins(), or send /reload-plugins as a prompt in Python. [lines 67-70] [Source]
SandboxIgnoreViolations is now a dict[str, list[str]] mapping command substrings (or *) to violation text to hide, instead of file/network lists. [lines 3616-3621] [Source]
The result message's origin is now documented as the origin Claude Code assigned, not a copy of the one you sent. It is absent when none was assigned. [lines 1930-1943] [Source]
Added guidance on which origin kinds an app may set (human, scheduled-trigger). To send text the user didn't write, use isSynthetic: true with client_composed: true. [lines 2464-2476] [Source]
Updated the human, peer, and unclassified origin descriptions. [lines 2475-2505] [Source]
The Chat-tab opt-in paragraph was removed from "Connect Claude Desktop", and the Desktop feature row was shortened. [lines 349-522] [Source]
The Admin UI is now "Spend limits only": members of admin.admin_groups can manage spend limits from Claude Desktop 4.23118.0 or later. [line 537] [Source]
Load-test mode doesn't cover the wait at each replica's concurrent upstream request limit. A replica still logs the client requests are open warning. [line 1650] [Source]
New project setting "Auto-continue when usage limits reset" (on by default). It controls whether threads and the project conversation continue on their own after a usage limit. [line 387] [Source]
The usage-limit section was rewritten around that setting. With it off, send a message after the reset to continue. [lines 403-486] [Source]
New-project goal field renamed "What do you want to work on?", and "Setup script failed" recovery now asks the project conversation to run the task in a new thread. [line 94] [line 509] [Source]
Secret form fields updated: the Name field was removed, "Header name" and "Token" fields are described, and non-Authorization headers need the Header changed. [lines 118-135] [Source]
A failing setup script can now either fail the session or leave it started without what the script failed to install. [line 442] [Source]
Subdirectory REVIEW.md files only apply on a protected branch. [line 24] [Source]
New section "When a subdirectory REVIEW.md doesn't apply": it lists excluded directory names (tests, examples, generated, vendor and third-party). The remaining files share a 150,000-byte limit. [lines 45-57] [Source]
Sessions no longer get a separate simulator each. Claude uses an already-running simulator unless the prompt names another. [line 87] [Source]
If you decline Claude's request, the device doesn't attach. New steps (Attach simulator, Continue, Cancel) let you use it yourself. [lines 101-107] [Source]
Permission guidance revised: "always allow" skips most prompts, and an "always allow" on a file edit moves only that session to Accept edits. [lines 159-161] [Source]
A task can change its own schedule with update_scheduled_task only after you click Run now and approve the call each time. [line 178] [Source]
The single "didn't send your claude.ai login" message was replaced by three: "Your organization's settings need an update", "Couldn't load your organization's settings", and "Restart Claude Code to finish signing in". Each has its own remedy. [lines 60-62] [lines 1116-1134] [Source]
Setting fastMode: true on an unsupported model leaves you on that model with fast mode off, and persistence applies only on a supported model. [lines 34-36] [Source]
The out-of-credits notification text changed to "Fast mode disabled ยท Out of usage credits". [line 220] [Source]
New "Match Monitor commands" section (v2.1.314+). A Bash matcher also runs on Monitor commands for PreToolUse and PermissionRequest, but allow and updatedInput have no effect. Use a matcher that includes Monitor, such as Bash|Monitor, to allow or rewrite. [lines 384-392] [Source]
New auth_storage_failure Notification type (v2.1.288+): sign-in completes but credentials can't be saved. [line 293] [line 2492] [Source]
PreCompact stdout now becomes instructions for the compaction summary (capped at 10,000 characters per hook) and is exempt from the output-file spill. [lines 819-822] [line 3340] [Source]
Claude Code v2.1.318+ doesn't install the IDE extension when claude starts in an integrated terminal. Until you install it, /ide reports "No available IDEs detected". [ide line 821]
An uninstalled extension returns if you run /ide externally or start claude on a version before v2.1.318. Three ways to stop that are listed. [ide lines 1090-1099] [vs-code lines 1090-1099]
New "Subagents switch or fail without asking" section. With Ask each time, an Agent-tool subagent fails with Agent terminated early due to an API error. With Switch automatically, it re-runs on the fallback model. [lines 565-577] [Source]
Grep(path) joins Write and Glob among the path rules that are accepted but never match. A Grep(docs/**) deny doesn't stop searching. Use Read(...) or Edit(...) instead. [lines 354-361] [Source]
Clarified that the best-effort Read deny coverage is for Read(path) rules. [line 337] [Source]
In cloud sessions, Auto appears unless your organization turns it off. On claude.ai/code it is grayed out until you pick a supported model. [permission-modes line 252] [web-quickstart line 137]
A failing setup script can leave the session started but missing what failed to install. Start a new session, or click Run now for scheduled tasks. [web-quickstart lines 241-251]
A wrapper that rebuilds the child's environment must carry over CLAUDE_CODE_REMOTE_MCP_CONFIG, CLAUDE_CODE_REMOTE, and CLAUDE_CODE_ENVIRONMENT_KIND. [line 45] [Source]
Split into "Pin a permission mode from your wrapper" and "Pre-approve specific tools" sections. A pinned auto mode applies only until a client such as claude.ai/code sets a mode. [lines 827-854] [Source]
fastMode: only Opus 5.5, Opus 5, and Opus 4.8 support it. /fast from another model switches you to Opus, and switching to an unsupported model turns it off. [lines 445-447] [Source]
autoInstallIdeExtension: Claude Code no longer installs or updates the extension on start. Before v2.1.318 it did. [line 5828] [Source]
preferredNotifChannel"auto" sends desktop notifications only when it detects iTerm2, Ghostty, or Kitty. [line 4987] [Source]
Context gathering clarified. In threads, Claude reads your messages and its own. In channels, it reads your recent messages and its own plus the channel name, topic, and description. To share a teammate's text, paste it into the message. [lines 89-93] [Source]
Over SSH from outside tmux, set preferredNotifChannel to the local terminal's value so notifications reach your machine. [line 85] [Source]
tmux: the progress bar claim was dropped, and the sample config sends none from a pane in a window you aren't viewing. It also needs preferredNotifChannel. [lines 130-139] [Source]
The retired Claude Haiku 3.5 rows were removed from the pricing, batch, and tool-use token tables. The same removal was made in prompt caching. [line 38] [Source]
max_uses now applies to a whole turn, not one request. A turn spans requests that resume after pause_turn or that end with only tool_result blocks. [lines 309-320] [Source]
You can drop the advisor tool from tools in a follow-up request only if no advisor call is still pending. [line 739] [Source]
The "not currently available" platform caveat was removed. [line 1983] [Source]
Examples now set max_tokens. A new note explains how a mixed server and client tool response differs from pause_turn: send the assistant content plus a user message with only the tool_result blocks. [line 495] [Source]
Rate-limit tables were reorganized and now include Claude Haiku 5.5, Sonnet 5.5, and Opus 5.5 rows. Rate limits are shared across inference_geo values. The page was also reworded. [lines 119-171] [Source]
The note that group and custom-role requests don't need the ce-user-management-2026-07-13 beta header was removed. The scopes explanation was rewritten. [line 38] [Source]
Mostly editorial rewording of the Workload Identity Federation pages (reference, admin API, and the AWS, Azure, GCP, GitHub Actions, Kubernetes, Okta, and SPIFFE provider guides).
The entry on signing a Claude Enterprise member out everywhere via the Admin API (beta) was reworded. Date headings were normalized (for example "April 9th" became "April 9"). [line 18] [Source]
SDK reference pages (Go, Java, Ruby)
The Go, Java, and Ruby API reference pages under docs-md/api/api/ (messages, batches, sessions, agents, and others) were regenerated. These are mostly formatting and example churn.