Changelog

Notable drops only, newest first. Dates and version numbers from the official Claude Code changelog; the full per-release list lives there. Maintained by the /content-sweep Lane C.

2026-10-09: Roundup, 2.1.280 to 2.1.296

The rest of three busy weeks, grouped.

Defaults and modes. Auto mode is now the starting permission mode for every interactive terminal and VS Code session on every plan and provider when none is configured (2.1.284), and for claude -p and Python Agent SDK runs on third-party providers or with telemetry off (2.1.285). Ultracode became its own toggle in /effort (Tab, or /effort ultracode [on|off]): it no longer forces xhigh and stays on at any level (2.1.284). See Permissions & Settings and Workflows.

Admin controls. deniedModels blocks named models, and availableModelsMatch: "exact" makes each availableModels entry allow only the version it names, so new releases wait until listed (2.1.283). allowedProviders limits which API providers a machine may use (2.1.285). All three are managed-only.

Settings and commands. "attribution": false hides all commit and PR attribution (2.1.281; older versions skip a settings file that holds it). maxProseWidth caps prose width in wide terminals while tables and code keep full width (2.1.282). /doctor prompt-audit checks your CLAUDE.md, skills, agents and commands for instructions written for older models (2.1.283; see Memory & CLAUDE.md). claude --desktop opens the desktop app on the current directory, or on a session with --continue or --resume <id> (2.1.285). /code-review --max-findings <n>|all reports more or fewer findings, and the choice sticks until --max-findings default (2.1.288). claude attach and claude logs accept part of a session name in place of the id (2.1.290).

Limits that moved. Background Bash and PowerShell commands now stop after a time limit (default 30 minutes, max 2 hours, 2.1.285), but since 2.1.288 only in unattended sessions (-p, Agent SDK, CI, cloud); terminal, desktop and VS Code sessions have none. The interactive WebSearch budget refills at 100 calls an hour instead of ending after 200 (CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR, 2.1.290). The default cap on MCP tool descriptions and server instructions doubled from 2,048 to 4,096 characters (2.1.296). The Read tool gained allow_large, so Claude can read a text file past the usual size limits in one call when the context has room (2.1.296).

Subagents. The Agent tool takes an effort parameter (2.1.292). Subagent frontmatter gains autoCompactWindow to compact earlier than the main conversation (2.1.296). CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL runs every workflow agent on one model (2.1.296). Subagents now preload at most 32 skills from their skills field (2.1.295). See Subagents.

Hooks. "onFailure": "block" on a command or HTTP hook makes a hook that can't start, times out or exits unexpectedly block the action instead of letting it through (2.1.295). /hooks opens on one list of your hooks grouped by event (2.1.286). See Hooks.

2026-10-07: The 5.5 models become the defaults (2.1.280, 2.1.284, 2.1.293)

Claude Opus 5.5 (claude-opus-5-5, 2.1.280, September 22), Sonnet 5.5 (claude-sonnet-5-5, 2.1.284, September 28) and Haiku 5.5 (claude-haiku-5-5, 2.1.293, October 7) each became the default for their tier. All have a 1M context window. Per million tokens: Opus 5.5 $4/$20 with $0.20 cache reads, Sonnet 5.5 $2/$10 with $0.10 cache reads, Haiku 5.5 $0.10/$0.50 for prompts up to 100K tokens and $0.50/$2.50 above that. Fast mode on Opus 5.5 is $8/$40.

2.1.280 also moved Pro and Team Standard from Sonnet to Opus as the default model, so every plan now starts on Opus 5.5. In Claude Code all three 5.5 models default to medium effort, and a level saved before effort became per-model no longer carries over to them. See Model Selection and Models & Cost.

2026-10-01: Claude Mods, and You should know (2.1.287)

Plugins can now carry mods: JavaScript or TypeScript that runs inside Claude Code and can watch, change or answer tool calls, prompts and model requests, add panes and commands, and restyle the screen. Mods are unsandboxed, run with your permissions, and can approve tool calls your ask rules or your own PreToolUse hooks would stop. Inspect one with claude plugin validate; stop them with --safe-mode (one session) or "disableAllHooks": true (every session). The same release adds You should know, a built-in mod where a side agent flags things you or Claude might miss; turn it on with /plugin enable cc-plugin-you-should-know@builtin. See Mods.

2026-09-19: Auto mode classifier runs server-side (2.1.278)

For Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, auto mode now defaults to the server-side classifier, which does not charge for classifier overhead. Claude Code warns if it falls back to the billed path. CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out on Bedrock, Vertex, Foundry and gateways, and /status gains an "Auto mode server" row.

2026-09-18: AGENTS.md support (2.1.277)

In a project with no CLAUDE.md, Claude Code now reads AGENTS.md instead; change it under "Project instructions" in /config. Not yet on Bedrock, Vertex or Foundry. Thariq's announcement says it is built on an upcoming harness customization layer, "Claude Code mods", which has not shipped. See Memory & CLAUDE.md.

Update, October 2026: mods shipped in 2.1.287, and AGENTS.md loading is now the built-in mod cc-plugin-agents-md. See Mods.

2026-09-17: claude.ai skills and plugins follow you to the terminal (2.1.275)

Skills and plugins enabled on your claude.ai account now sync to terminal sessions signed in with it; opt out with syncClaudeAiSkills: false or syncClaudeAiPlugins: false. Also new: a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn and sends all queued messages at once.

2026-09-14: Subagents without CLAUDE.md, per-command network access (2.1.271)

omitClaudeMd in agent frontmatter (and --agents JSON) lets a custom or plugin subagent run without your user, project and local CLAUDE.md files; managed policy files still load. In auto mode with sandboxing, Bash, PowerShell and Monitor commands can carry per-command allowed_domains: the hosts a command needs are reviewed with it and opened for it alone. Fast mode now works in Remote sessions where your organization allows it.

2026-09-11: claude plugin eval (2.1.269)

Run a plugin's eval suite against Claude Code and get scored, reproducible results as JSON plus an HTML report. Same release adds /output-style [name] to list and switch output styles (including over Remote Control and in headless sessions), a diff of files changed by a Bash command (bashEditDiffEnabled), and CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS (1 to 256) for the Workflow tool's per-run agent limit.

2026-09-09: Cap the effort level (2.1.267)

maxEffortLevel, set at the top level or per model under modelSettings, caps the effort level on every provider including Bedrock, Vertex and Foundry. Users can still pick a lower level.

2026-09-04: /skill-doctor (2.1.261)

Shows which loaded skills go unused and what they cost in context, so you can prune them. Also bashOutputMaxChars and taskOutputMaxChars to raise how much command and background-task output Claude receives inline before it is saved to a file (up to 128K characters).

2026-09-03: Live diff panel (2.1.260)

In fullscreen mode a diff panel opens beside the conversation and shows your uncommitted changes as Claude edits; toggle it with /diff. /cost and the status line's prompt_cache field now name a likely cause for prompt-cache misses (tool definitions or system prompt changed, idle past the TTL).

2026-09-02: Managed MCP servers and promptless headless hosts (2.1.259)

managedMcpServers lets an organization provide HTTP/SSE MCP servers to every user (entries that name a command to run are skipped). --permission-prompts none is for unattended hosts: anything that would prompt is denied automatically while the active permission mode, including auto mode, keeps deciding.

2026-09-01: Fable 5.1 becomes the default Fable (2.1.257)

Claude Fable 5.1 (claude-fable-5-1) is now the default Fable model: 1M context, $10/$50 per Mtok with $0.25/Mtok cache reads. Auto mode gains a Containment Escape rule, so cloud metadata-credential fetches, egress evasion and cross-tenant reach are no longer auto-approved unless your environment marks them expected, plus a one-time prompt before the first file read outside the working directories (permissions.blockReadsOutsideWorkingDirectories). CLAUDE_CODE_SUBAGENT_MODEL_FORCE applies one model to every subagent, and s in /effort changes effort for the current session only.

2026-08-28: Model-switch hooks and a cache line in /cost (2.1.251)

New PreModelSwitch and PostModelSwitch hook events can block, confirm or annotate a model switch, and SessionStart resume hooks now receive session staleness and the estimated re-cache cost. /cost shows a per-session prompt-cache line (hit ratio, misses, tokens re-cached, warm/cold) with a matching prompt_cache object for status line scripts.

2026-08-27: --restricted mode (2.1.248)

--restricted (or CLAUDE_CODE_RESTRICTED=1) removes the built-in tools that run commands or code, and WebFetch, unless named in --tools; keeps file tools inside the working directory; and refuses bypassPermissions. Agent frontmatter gains experimental.cacheTtl ("5m" or "1h") for a per-agent prompt cache TTL.

2026-08-26: Feedback drafts and API cost profiling (2.1.247)

The SendFeedback tool lets Claude draft a feedback report when something goes wrong, for you to review and send from /feedback (turn off with feedbackDrafts). /claude-api cost-optimize profiles an existing project's Claude API spend and works through cost levers (caching, token hygiene, batch, effort, model choice) one measured change at a time.

2026-08-25: Auto mode rules in /permissions (2.1.246)

A new Auto mode tab in /permissions shows and edits auto mode classifier rules. Startup now warns about Bash allow rules with a wildcard before the subcommand (such as Bash(git * main)), since they also match options inserted before the subcommand.

2026-08-21: /claude-api upgrade and honest cost math (2.1.239)

New /claude-api upgrade migrates Python projects from anthropic 0.x to 1.x. Cost surfaces (/cost, the status line, --max-budget-usd) now include the 1.1x US-only-inference premium for data-residency workspaces, so caps reflect what you actually pay. Fullscreen renderer now also offered on Bedrock, Vertex, and Foundry setups.

2026-08-20: Concise output style ships built in (2.1.237)

A new "Concise" style under Output style in /config: Claude leads with results and skips preamble and narration while doing the work just as thoroughly. Same release fixes prompt caching for LLM-gateway and custom-base-URL sessions. 2.1.238 adds keybindingFlavor: "readline" (Bash-style Ctrl+W) and plugin-marketplace headersHelper for minting short-lived auth headers.

2026-08-19: Default-model env var and idle notifications (2.1.236)

ANTHROPIC_DEFAULT_MODEL sets the model new sessions start on while /model picks still win. Cross-session SendMessage gains notify_when_idle: ask another session on the machine to ping you once when it next goes idle, no polling. macOS sandbox wildcard read-deny rules (like **/.env) now hold inside allowed regions and survive file renames.

2026-08-17: Sessions continue themselves at limit reset (2.1.234)

Claude Code now continues your session automatically when a claude.ai usage limit resets (toggle in /config). GitLab lands as a first-class remote: MR badge in the footer and status line (2.1.234), plus --worktree MR support and opt-in Bash memory cgroups on Linux (2.1.233).

2026-08-14: Auto mode becomes the default (Pro, Max, Team)

New sessions start in auto mode: a classifier judges each action and blocks the dangerous ones instead of prompting you for everything. Anthropic's numbers: the classifier blocked 89% of dangerous commands in a 1,053-tester study versus 13.6% caught by manual approval, and production-severity unintended actions fell from 6.3% of manually approved sessions to 2.4%. Shift+Tab still cycles modes; admins can pin defaultMode or set disableAutoMode. Enterprise and cloud-platform rollout follows. Classifier overhead is no longer billed on Pro/Max/Team.

2026-08-13: Subagent forking on by default (2.1.232)

subagent_type: "fork" subagents inherit the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions now run in the background by default. Type @ in the prompt to mention another live Claude session by name and message it directly. Sessions on one machine now keep unique names.

2026-08-07: Self-hosted runners (2.1.224)

claude self-hosted-runner turns your own machines or containers into a place Claude Code web, mobile, and desktop sessions can run (Team and Enterprise). Plugins gain an archive source: install from a zip over HTTPS with optional SHA-256 pinning, no git or npm. 2.1.229 documents claude remote-control --continue and adds marketplace command sources.

2026-08-04: Sandbox credential masking and permission hardening (2.1.221)

Sandbox credential files get mode: "mask" on Linux/WSL: commands read a sentinel while the proxy substitutes the real value on egress. VSCode gains the Focus view (Ctrl+Alt+F). A new prompt-audit subcommand ships. The two follow-ups matter for safety: 2.1.222 stops worktree-isolated sessions running destructive git against the main checkout, and 2.1.223 closes a Bash permission bypass where crafted or invisible-Unicode commands could hide from permission checks.

2026-07-24: Claude Opus 5 becomes the default Opus (2.1.219)

Opus 5 ships as the default Opus model, with a 1M context window and fast mode at $10/$50 per Mtok. Subagents can spawn nested subagents up to depth 3 again, after 2.1.217 cut the default to

1. New sandbox.network.strictAllowlist denies non-allowlisted hosts for sandboxed commands without prompting, and a workflowSizeGuideline settings key sets the dynamic workflow size guideline from any settings file. Opus 4.7 loses fast mode, so /fast now covers Opus 5 and 4.8.

2026-07-22: /code-review moves to a background subagent (2.1.218)

Review work no longer fills your conversation. Skills with context: fork now run in the background by default; opt out per skill with background: false. Agent frontmatter hooks now require the agent file's own folder to have accepted workspace trust, and /deep-research starts only when you invoke it rather than launching on its own.

2026-07-21: Subagent concurrency and budget caps (2.1.217)

Concurrent subagents are capped at 20 by default, overridable with CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS, so a single message cannot fan out unbounded background agents. --max-budget-usd now actually halts background subagents once the cap is reached instead of letting them run on past it. Also fixes a memory leak where truncated MCP tool output kept the full untruncated result in memory for the rest of the session.

2026-07-20: Long-session slowdown fix (2.1.216)

Message normalization cost grew quadratically with the number of turns, causing multi-second stalls and slow resumes in long sessions. New sandbox.filesystem.disabled setting skips filesystem isolation while keeping network egress control. Worktree-isolated subagents can no longer redirect git into the shared checkout via git -C, --git-dir, or GIT_DIR.

2026-06-12: Conversation-language session titles (2.1.176)

Session titles are now generated in the language of your conversation. New footerLinksRegexes setting renders regex-matched link badges in the footer row; improved Bedrock credential caching.

2026-06-08: --safe-mode + /cd (2.1.169)

New --safe-mode flag for a locked-down launch; /cd moves a session to a new working directory mid-run; self-hosted runner gains a post-session lifecycle hook.

2026-06-06: Configurable fallback models (2.1.166)

fallbackModel setting tries up to three models in order when the primary is overloaded or unavailable (--fallback-model now applies to interactive sessions too). Finer thinking controls (MAX_THINKING_TOKENS=0, --thinking disabled), plus MCP and cross-session messaging hardening.

2026-06-04: /plugin list + managed version ranges (2.1.163)

Plugin listing with --enabled/--disabled filters; requiredMinimumVersion / requiredMaximumVersion managed settings to pin a Claude Code version range; Stop and SubagentStop hooks can now return additionalContext.

2026-06-03: Remote Control footer + agent-view waitingFor (2.1.162)

Remote Control shows as a persistent footer pill instead of a startup message; claude agents --json exposes what a session is blocked on; glob patterns supported in deny-rule tool names.

2026-05-29: Plugins auto-load from .claude/skills (2.1.157)

Plugins placed in .claude/skills load with no marketplace; claude plugin init <name> scaffolds a new plugin; EnterWorktree can switch between Claude-managed worktrees mid-session.

2026-05-28: Opus 4.8 + dynamic workflows (2.1.154)

Opus 4.8 ships and defaults to high effort (/effort xhigh); Fast mode on 4.8 runs ~2× the standard rate for ~2.5× the speed. Dynamic workflows arrive: ask Claude to orchestrate work across tens to hundreds of background agents and run /workflows to view runs. The trigger keyword was renamed workflow to ultracode. Background command execution via claude --bg --exec.

2026-05-27: /code-review --fix + skill controls (2.1.152)

/code-review --fix applies review findings to the working tree and /simplify invokes it; skills can declare disallowed-tools; added /reload-skills and a MessageDisplay hook.

2026-05-21: Pinned background sessions; /simplify → /code-review (2.1.147)

Pinned background sessions stay alive when idle and restart in place across updates; the review commands consolidate under /code-review; EnterWorktree branches from local HEAD.

2026-05-19: /model session-scoped + background resume (2.1.144)

/model now changes the current session only (press d to set the default for new sessions); claude --bg sessions appear in /resume, marked bg.

2026-05-11: Agent view (research preview) + /goal (2.1.139)

claude agents lists every session (running, blocked, done); /goal sets a completion condition and Claude keeps working until it's met, with a live elapsed/turns/tokens overlay.

2026-05-06: Code with Claude: Remote Control, proactive workflows, managed agents

Anthropic previews cross-device Remote Control, routines that run on cron schedules / GitHub webhooks / API endpoints, and Claude Managed Agents (sandboxed execution, checkpointing, credential scoping).