CC Version History

A curated reference of Claude Code releases. Each entry highlights features that still matter in the current 2.1.121 line. Superseded features are noted once with their replacement.

For the definitive changelog, run /release-notes inside CC or see claude update output.

Latest

2.1.121 (current)

alwaysLoad: true MCP server config — opt a server out of ENABLE_TOOL_SEARCH deferral so its tools are always loaded. Use sparingly; each always-loaded tool consumes context every turn. Per-tool opt-in also available via _meta.anthropic/alwaysLoad: true in the tool definition.
claude plugin prune (alias autoremove) removes auto-installed plugin dependencies that no longer have a parent. Flags: --dry-run (preview), --scope user|project|local (default user), -y (skip confirm — required when stdin is not a TTY). claude plugin uninstall --prune cascades the cleanup.
PostToolUse hookSpecificOutput.updatedToolOutput now works for all tools (previously MCP-only). Hooks can rewrite or redact Bash, Read, Write, etc. output before it reaches the model.
/skills gains a type-to-filter search box for long lists.
Fullscreen mode: typing into the prompt no longer jumps scroll back to the bottom after you scrolled up.
Dialogs that overflow the terminal are now scrollable with arrow keys, PgUp/PgDn, home/end, mouse wheel — both fullscreen and non-fullscreen.
Long URLs that wrap across rows in fullscreen are clickable on any wrapped line.
CLAUDE_CODE_FORK_SUBAGENT=1 works in non-interactive -p and SDK sessions.
--dangerously-skip-permissions no longer prompts for writes to .claude/skills/, .claude/agents/, .claude/commands/.
MCP servers that hit a transient error during startup auto-retry up to 3 times instead of staying disconnected.
Vertex AI: support X.509 certificate-based Workload Identity Federation (mTLS ADC).
LSP diagnostic summaries expand on click / Ctrl+O.
SDK mcp_authenticate accepts redirectUri for custom scheme completion and claude.ai connectors.
OpenTelemetry: LLM request spans add stop_reason, gen_ai.response.finish_reasons, and user_system_prompt (gated behind OTEL_LOG_USER_PROMPTS).
Memory fixes: unbounded RSS growth on image-heavy sessions, up to ~2 GB leak in /usage on machines with large transcript histories, leak when long-running tools fail to emit progress events.
Fixed Bash tool becoming permanently unusable when the directory CC was started in is deleted/moved mid-session.
Fixed --resume failing on large/corrupt session transcripts; corrupt lines are now skipped.
Fixed scrollback duplication on tmux, GNOME Terminal, Windows Terminal, Konsole when redrawing in non-fullscreen.

2.1.120

Windows: Git for Windows (Git Bash) no longer required — when absent, CC uses PowerShell as the shell tool. Pairs with the PowerShell auto-approve rules added in 2.1.119.
claude ultrareview [target] non-interactive CLI subcommand. Runs cloud multi-agent review on the current branch, a PR number, or a base branch. Flags: --json (raw bugs.json payload), --timeout <minutes> (default 30). Exits 0 on completion, 1 on failure. Suitable for CI gating; billed, so wire to manual triggers, not pre-push.
${CLAUDE_EFFORT} substitution in skill content — skills can reference the active effort level (low|medium|high|xhigh|max) to weight their guidance.
AI_AGENT env var set on subprocesses spawned by the Bash tool, so gh and similar tools can attribute traffic to Claude Code.
Spinner tips that recommend installing the desktop app or creating skills/agents are hidden when you already have them.
“Use PgUp/PgDn to scroll” hint shown when the terminal sends arrow keys instead of scroll events.
Faster session start when many claude.ai connectors are configured but not authorized.
claude plugin validate accepts $schema, version, description at the top level of marketplace.json and $schema in plugin.json.
Fixed pressing Esc during a stdio MCP tool call closing the entire server connection (regression from 2.1.105).
Fixed /rewind and other interactive overlays not responding to keyboard input after claude --resume.
Fixed DISABLE_TELEMETRY / CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC not suppressing usage metrics for API/enterprise users.
Fixed false-positive “Dangerous rm operation” prompts in auto mode for multi-line bash commands containing both a pipe and a redirect.
Fixed find in the Bash tool exhausting open file descriptors on large trees, causing host-wide crashes (macOS/Linux native builds).

2.1.119

/config settings persist to ~/.claude/settings.json (theme, editor mode, verbose, etc.) and now participate in project/local/policy override precedence. Project .claude/settings.json still wins.
prUrlTemplate setting points the footer PR badge at a custom code-review URL instead of github.com. Supports {host}, {owner}, {repo}, {number}, {url} placeholders.
CLAUDE_CODE_HIDE_CWD env var hides the working directory in the startup logo (privacy/screencast). Does NOT affect statusline, OTEL spans, or hook stdin paths.
--from-pr now accepts GitLab merge-request, Bitbucket pull-request, and GitHub Enterprise PR URLs. Auth inherits per-provider CLI credentials.
--print mode honors the agent’s tools: / disallowedTools: frontmatter, matching interactive-mode behavior.
--agent <name> honors the agent definition’s permissionMode for built-in agents.
PowerShell tool commands now auto-approvable in permission mode (same rules as Bash). Patterns like PowerShell(Get-ChildItem:*) allow, PowerShell(Remove-Item:*) deny.
Hooks: PostToolUse and PostToolUseFailure receive duration_ms in their JSON input (tool execution time only — excludes permission prompts and PreToolUse hooks).
Subagent and SDK MCP server reconfiguration connect servers in parallel instead of serially.
Plugins pinned by another plugin’s version constraint auto-update to the highest satisfying git tag.
Vim mode: Esc in INSERT no longer pulls a queued message back into the input; press Esc again to interrupt.
Slash command picker highlights matching characters and wraps long descriptions onto a second line.
owner/repo#N shorthand links now use your git remote’s host (not always github.com).
Security: blockedMarketplaces correctly enforces hostPattern and pathPattern entries.
OpenTelemetry: tool_result and tool_decision events include tool_use_id; tool_result adds tool_input_size_bytes.
Statusline stdin JSON includes effort.level and thinking.enabled.
Many UX bug fixes: pasting CRLF inserting blank lines, kitty keyboard losing newlines on multi-line paste, Glob/Grep disappearing when Bash denied (native builds), MCP HTTP OAuth on non-JSON discovery responses, Rewind showing “(no prompt)” for image messages, async PostToolUse hook empty-payload writes.

2.1.118

Hooks can invoke MCP tools directly via type: "mcp_tool" with ${tool_input.*} / ${cwd} templating. Replaces subprocess plumbing for MCP-writing hooks. Non-blocking on server disconnect; 60s default timeout; avoid on SessionStart (MCP servers typically not yet connected).
/cost and /stats merged into /usage. Both remain as typing shortcuts that open the relevant tab.
Custom named themes: create from /theme or hand-edit ~/.claude/themes/*.json. Plugins can ship themes via a themes/ directory.
DISABLE_UPDATES env var blocks all update paths including manual claude update — stricter than DISABLE_AUTOUPDATER.
WSL on Windows can inherit Windows-side managed settings via the wslInheritsWindowsSettings policy key.
Auto mode "$defaults" placeholder — include in autoMode.allow / soft_deny / environment to extend the built-in list instead of replacing it.
claude plugin tag creates release git tags for plugins with version validation.
Vim visual mode (v) and visual-line mode (V) with selection, operators, and visual feedback.
--continue / --resume finds sessions that added the current directory via /add-dir.
Fixed /fork writing the full parent conversation to disk per fork (now writes a pointer and hydrates on read).
Fixed agent-type hooks failing with “Messages are required” for events other than Stop / SubagentStop.

2.1.117

Native builds on macOS and Linux: embedded bfs and ugrep replace Glob and Grep tools (available through the Bash tool). Faster searches without a separate tool round-trip. Windows and npm-installed builds unchanged.
Opus 4.7 /context 1M-window bug fixed — CC was computing /context % against a 200K window, inflating usage and triggering early autocompact. If you measured context baselines during 2.1.111–2.1.116 on Opus 4.7, re-measure.
cleanupPeriodDays retention sweep extended to ~/.claude/tasks/, ~/.claude/shell-snapshots/, and ~/.claude/backups/ (previously only todos/).
CLAUDE_CODE_FORK_SUBAGENT=1 enables forked subagents on external builds.
Agent frontmatter mcpServers loads for main-thread agent sessions via --agent.
/resume offers to summarize stale, large sessions before re-reading them.
Faster MCP startup when both local and claude.ai servers are configured (concurrent connect now default).
Default effort for Pro/Max subscribers on Opus 4.6 and Sonnet 4.6 is now high (was medium).
OpenTelemetry: user_prompt events include command_name and command_source for slash commands. cost.usage, token.usage, api_request, api_error include an effort attribute when the model supports effort levels. Custom/MCP command names redacted unless OTEL_LOG_TOOL_DETAILS=1.

2.1.116

/resume up to 67% faster on 40MB+ sessions; handles sessions with many dead-fork entries more efficiently.
Faster MCP startup when multiple stdio servers are configured; resources/templates/list deferred to first @-mention.
/doctor can be opened while Claude is responding (previously had to wait for turn).
/reload-plugins and background auto-update now auto-install missing plugin dependencies from already-added marketplaces.
Bash tool surfaces a hint when gh commands hit GitHub’s API rate limit.
Sandbox auto-allow dangerous-path safety — no longer bypasses the dangerous-path check for rm / rmdir targeting /, $HOME, or other critical system directories.
Releases URL moved to https://downloads.claude.ai/claude-code-releases.

2.1.113

CLI now spawns a native Claude Code binary (via a per-platform optional dependency) instead of bundled JavaScript.
sandbox.network.deniedDomains setting blocks specific domains even when a broader allowedDomains wildcard would permit them.
Ctrl+A and Ctrl+E move to the start/end of the current logical line (readline behavior for multiline input).
Windows: Ctrl+Backspace deletes the previous word.
Long URLs stay clickable when they wrap across lines (in terminals with OSC 8 hyperlinks).
/loop: pressing Esc cancels pending wakeups. Wakeups display as “Claude resuming /loop wakeup”.
Subagents that stall mid-stream fail with a clear error after 10 minutes instead of hanging silently.
Bash security hardening:
- Multi-line commands whose first line is a comment now show the full command in the transcript (closes a UI-spoofing vector).
- Deny rules now match commands wrapped in env / sudo / watch / ionice / setsid and similar exec wrappers.
- Bash(find:*) allow rules no longer auto-approve find -exec / -delete.
- dangerouslyDisableSandbox respects permissions (no silent bypass).
- macOS: /private/{etc,var,tmp,home} paths treated as dangerous removal targets under Bash(rm:*) allow rules.

2.1.111

Opus 4.7 xhigh effort level, between high and max.
Interactive /effort slider for setting reasoning depth.
/less-permission-prompts skill: scans transcripts and proposes allowlist entries for .claude/settings.json.
/ultrareview cloud multi-agent review.
Auto-mode GA for Max on Opus 4.7 (no --enable-auto-mode flag needed).
“Auto (match terminal)” theme option.
Plan files named after the prompt instead of plan-1.md.
/skills sort-by-token toggle.
OTEL_LOG_RAW_API_BODIES env var for raw request/response debugging.
Ctrl+U clears the whole input buffer; Ctrl+Y restores it.
PowerShell tool progressive rollout via CLAUDE_CODE_USE_POWERSHELL_TOOL.

2.1.110

/tui command and tui setting for fullscreen mode.
/focus split from Ctrl+O; Ctrl+O is now verbose-only.
Push notification tool (opt-in via Remote Control).
autoScrollEnabled setting.
/doctor warns on MCP-scope conflicts.
Bash tool enforces the documented max timeout.
SDK reads TRACEPARENT and TRACESTATE from env.

2.1.108

ENABLE_PROMPT_CACHING_1H (1-hour cache window) and FORCE_PROMPT_CACHING_5M env vars.
/recap command, default-on for telemetry-disabled users.
/undo aliased to /rewind.
/model warns on cache miss.
/resume prefers the current working directory.
Built-in slashes (/init, /review, /security-review) re-implemented via the Skill tool.

2.1.105

PreCompact hooks can block compaction (exit 2 or {"decision":"block"}).
Skill description cap raised from 250 to 1536 characters.
Plugin monitors manifest entry.
WebFetch strips <script> and <style> before returning.
Stalled-stream abort after 5 minutes.
/doctor f applies auto-fixes.
/proactive aliased to /loop.
EnterWorktree accepts a path parameter.

2.1.101

/team-onboarding skill generates a teammate ramp-up guide from local usage patterns.
OS CA certificate trust is now the default (enterprise TLS proxies work out of the box). Set CLAUDE_CODE_CERT_STORE=bundled to restrict.
API_TIMEOUT_MS is honored (previous 5-minute hardcoded timeout removed).
Brief-mode retry on transient failures.
Remote auto cloud environment detection.

2.1.98 — 2.1.99

2.1.99

Settings resilience: an unrecognized hook event name no longer discards the whole settings.json. Only the bad entry is skipped. Audit hooks after upgrading — silently-dead hooks may start firing.
permissions.deny now overrides a PreToolUse hook’s permissionDecision: "ask".
Subagents inherit MCP tools from dynamically-injected servers (tools added mid-session via claude mcp add).
Subagents in isolated worktrees can Read/Edit files inside their own worktree.
--resume <name> accepts session titles set via /rename or --name.
/team-onboarding command (see 2.1.101).
Fixed --dangerously-skip-permissions being silently downgraded to accept-edits mode.
Fixed permissions.additionalDirectories changes not applying mid-session.
Fixed command injection in POSIX which fallback.

2.1.98

Monitor tool: streams stdout events from background scripts. Each stdout line becomes a notification. Use for long builds, deploys, or fetches. See Part V chapter on Monitor for patterns.
Bash permission hardening — 6 bypass vectors closed:
- Compound commands (echo x && killall node) now prompt in auto/bypass modes.
- Backslash-escaped flags (killall\ node) no longer auto-allowed.
- /dev/tcp and /dev/udp redirects prompt instead of auto-allowing.
- Env-var prefix (LANG=C killall node) prompts unless the var is known-safe (LANG, TZ, NO_COLOR, etc.).
- Whitespace matching: Bash(cmd:*) wildcards match extra spaces/tabs.
PID namespace isolation on Linux when CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 is set.
CLAUDE_CODE_SCRIPT_CAPS=<N> limits per-session script invocations.
OTEL tracing: Bash tool auto-injects TRACEPARENT to subprocesses. Opt-in span attributes: OTEL_LOG_USER_PROMPTS, OTEL_LOG_TOOL_DETAILS, OTEL_LOG_TOOL_CONTENT (debug-only, never in shared config).
/agents view split into Running and Library tabs.
Interactive Bedrock and Vertex AI setup wizards on the login screen.
--exclude-dynamic-system-prompt-sections flag for print mode enables cross-user prompt caching.

2.1.95 — 2.1.97

2.1.97

Focus view (Ctrl+O in CLAUDE_CODE_NO_FLICKER=1 mode): clean view showing prompt, one-line tool summary with diffstats, and final response. (Split to dedicated /focus in 2.1.110.)
Statusline refreshInterval setting re-runs the script every N seconds for live indicators.
workspace.git_worktree field added to statusline stdin JSON — shows active worktree path. Important for multi-worktree setups.
Accept Edits mode auto-approves commands prefixed with known-safe env vars (LANG=, NO_COLOR=, timeout).
/agents view shows ● N running indicators for live subagent instances.
Fixed 429 retries burning all attempts in ~13s (exponential backoff is now the minimum).
Fixed MCP HTTP/SSE connections leaking ~50 MB/hr on reconnect.
Fixed subagents with worktree isolation leaking cwd back to the parent session.

2.1.96

Fixed Bedrock 403 regression with AWS_BEARER_TOKEN_BEDROCK and CLAUDE_CODE_SKIP_BEDROCK_AUTH.

2.1.95

NO_FLICKER mode stability fixes (URL wrapping, MCP hover crash, Windows scrolling, short-terminal statusline, Korean/Japanese copy).

2.1.93 — 2.1.94

2.1.94

Default effort level changed from medium to high for API-key, Bedrock, Vertex, Foundry, Team, and Enterprise users. Use /effort low to revert per-session.
Amazon Bedrock Mantle support via CLAUDE_CODE_USE_MANTLE=1.
Plugin skill YAML hooks now fire correctly (were silently ignored before).
Plugin output styles gained keep-coding-instructions frontmatter.
Fixed agents stuck after 429 with long Retry-After headers.
Fixed Console login on macOS when the keychain is locked.
Fixed plugin hooks failing when CLAUDE_PLUGIN_ROOT is unset.

2.1.89 — 2.1.92

2.1.92

/cost per-model and cache-hit breakdown for subscription users.
/release-notes presents an interactive version picker.
apply-seccomp helper ships in npm and native builds for Linux sandbox.
/tag and /vim removed. Editor mode now in /config.

2.1.91

MCP tool results can be up to 500K chars via _meta.anthropic/maxResultSizeChars.
Edit tool uses shorter context anchors (token savings, no behavior change).
Plugins can ship executables in a bin/ directory.
disableSkillShellExecution setting blocks inline shell in skills/commands/plugins.

2.1.90

/powerup command: interactive lessons with animated terminal demos for CC features.
Auto mode respects explicit natural-language boundaries (“don’t push”, “only edit these files”).
Fixed Edit/Write failing when a PostToolUse hook reformats the file (Prettier, Black, rustfmt workflows).
Fixed full prompt-cache miss on the first request after --resume.

2.1.89

Autocompact thrash guard: after 3 consecutive compact-refill cycles, CC halts with an actionable error instead of looping.
Hook stdout/stderr over 50K chars saves to disk; context gets a file path and 2KB preview.
Edit tool works on files previously viewed via Bash cat/sed/head/tail without a prior Read call.
PreToolUse hooks can return {"defer": true} to pause headless -p sessions for later resume.
New settings: sandbox.failIfUnavailable, managed-settings.d/ drop-in directory, disableDeepLinkRegistration, showThinkingSummaries (now OFF by default).
New env vars: MCP_CONNECTION_NONBLOCKING, CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE.

2.1.87 — 2.1.88

2.1.88

Significant stability release. Highlights worth remembering:

Nested CLAUDE.md re-injection fix — in setups with global + per-project CLAUDE.md, nested files were re-injected dozens of times per long session, silently eating context. Now fixed.
Prompt cache fix — tool schema bytes changing mid-session no longer invalidate the prompt cache on every turn. Setups with many MCP servers benefit most.
Hook if compound-command fix — if: "Bash(git *)" now correctly matches ls && git push, FOO=bar git push, etc. Hooks may now fire more often than before (correct behavior).
PreToolUse/PostToolUse absolute file_path — Write/Edit/Read hooks now receive absolute paths, not relative.
CLAUDE_CODE_NO_FLICKER=1 env var opts into alt-screen rendering with virtualized scrollback (reduces flicker, especially on WSL2). Tradeoff: scrollback discarded on exit.
PermissionDenied hook event fires after auto-mode classifier denials. Return {retry: true} to allow retry.
Named subagents appear in @ mention typeahead.
LSP server auto-restarts after crash instead of staying zombie.
Fixed memory leak where large JSON stdin payloads were retained as LRU cache keys.
showThinkingSummaries now OFF by default.
Computer use (macOS only, Pro/Max) via computer-use MCP server.
Fixed StructuredOutput schema cache bug, Edit OOM on >1 GiB files, CJK/emoji prompt history truncation at 4KB, /stats dropping subagent/fork tokens.

2.1.87

Fixed Cowork Dispatch message delivery.

2.1.84 — 2.1.86

2.1.86

Skill description cap: 250 characters (raised to 1536 in 2.1.105). Descriptions over cap were silently truncated, cutting off “Use when…” trigger phrases. Audit with a wc -c loop over ~/.claude/skills/*/SKILL.md.
Read tool uses a compact line-number format and deduplicates unchanged re-reads.
@-mention content no longer JSON-escaped — token savings on @file references.
/skills sorted alphabetically.
.jj and .sl added to VCS exclusions (Jujutsu, Sapling).
X-Claude-Code-Session-Id header for proxy session aggregation.

2.1.85

Conditional if field for hooks on PreToolUse, PostToolUse, PostToolUseFailure, PermissionRequest. Uses permission-rule syntax (Bash(git *), Edit(*.ts)). Filters BEFORE process spawn — zero overhead for misses. Eliminates ~90% of wasted spawns in hooks that matched broadly and filtered internally.
MCP server env vars CLAUDE_CODE_MCP_SERVER_NAME and CLAUDE_CODE_MCP_SERVER_URL available in MCP headers.
Fixed /compact failing on very large sessions.

2.1.84

paths: frontmatter accepts YAML list on rules and skills. Domain-specific rules can now scope to file types (**/*.py, **/*.sh) so they only load when relevant.
System-prompt caching works with ENABLE_TOOL_SEARCH=true — previously the two were incompatible, forcing a choice between cache savings and deferred tools.
MCP tool descriptions and server instructions capped at 2KB.
Code intelligence LSP plugins published (typescript-lsp, pyright-lsp, gopls-lsp, rust-analyzer-lsp, clangd-lsp, ruby-lsp). Enable per-plugin in enabledPlugins.
New TaskCreated hook event.
Idle-return prompt nudges /clear after 75+ minutes idle.

2.1.82 — 2.1.83

2.1.83

New hook events: CwdChanged (fires on directory change, useful for direnv allow) and FileChanged (fires on disk changes, enables hot-reload patterns).
Background agent stability fix — agents now survive context compaction without becoming invisible or spawning duplicates.
TaskOutput tool deprecated — use Read on the task’s output file path instead.
MEMORY.md auto-capped at 25KB / 200 lines to prevent memory bloat.
Transcript search: Ctrl+O to enter transcript mode, / to search, n/N for next/previous.
Ctrl+L clears and redraws the UI.
CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 strips API credentials from all subprocesses.
CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS extends SessionEnd hook timeout beyond 1.5s default.
Agents can declare initialPrompt: in frontmatter for autonomous first-turn dispatch.
Fixed worktree names with / causing hangs.
sandbox.failIfUnavailable, managed-settings.d/, disableDeepLinkRegistration settings added.

2.1.82

Reactive-hooks groundwork shipped with 2.1.83.

2.1.76 — 2.1.81

2.1.81

--bare flag: skips hooks, LSP init, plugin sync, skill-directory walks. Requires ANTHROPIC_API_KEY. For CI pipelines and scripted -p calls.
--channels permission relay (GA) — routes approval prompts to a mobile notification channel.
Fixed concurrent-session OAuth re-auth storm.
MCP tool calls collapse into a single “Queried {server}” line in the transcript.

2.1.80

Statusline scripts receive a rate_limits field (5-hour and 7-day usage/limit).
Skills and slash commands accept effort: in YAML frontmatter.
source: 'settings' plugin marketplace declaration (groundwork for discovery).
Fixed --resume dropping parallel tool results.

2.1.79

claude auth login --console — terminal-only OAuth via paste-back code. For remote servers, Docker, CI.
Turn duration display toggle in /config.
/remote-control for VS Code.

2.1.78

StopFailure hook event fires when CC stops from API error (rate limit, network, auth). Fills the gap between Stop (normal) and SessionEnd (always).
${CLAUDE_PLUGIN_DATA} env var — per-plugin persistent state directory.
Plugin agent frontmatter accepts effort, maxTurns, disallowedTools.
Line-by-line response streaming (later disabled on Windows/WSL in 2.1.81).

2.1.77

Opus 4.6 default max output raised from 32k to 64k (upper bound 128k).
sandbox.allowRead — grant read-only sandbox access to specific paths.
/copy N copies the Nth-latest response.
/branch (renamed from /fork) for conversation branch points.

2.1.76

PostCompact hook event fires after compaction completes. Reload critical rules, CLAUDE.md, or state that compaction discarded. Pairs with PreCompact (save-before, restore-after).
/effort low|medium|high slash command (xhigh added in 2.1.111).
--name / -n flag and /rename command for session display names.
worktree.sparsePaths setting — git sparse-checkout for large monorepo worktrees.
MCP elicitation: MCP servers can request structured input mid-task. New Elicitation and ElicitationResult hook events.
1M context window default for Opus 4.6 on Max, Team, Enterprise plans.
autoMemoryDirectory setting redirects auto-memory storage to a custom path.
/context command shows actionable diagnostics, not just token counts.
Deferred tool schemas now survive compaction.
RTL text (Hebrew, Arabic) renders correctly in terminal output.

2.1.73 — 2.1.75

2.1.75: 1M context default for Opus 4.6 (see 2.1.76 entry).
2.1.74: Configurable SessionEnd hook timeout, autoMemoryDirectory, /context diagnostics.
2.1.73: Groundwork for 2.1.76 reactive-hook surface.

Superseded features

Features that shipped but have since been replaced or deprecated:

Shipped	Feature	Replaced by
Pre-2.1.88	Commands directory (`commands/`)	Skills directory (`skills/<name>/SKILL.md`)
2.1.83	`TaskOutput` tool	`Read` on the task’s output file path
Pre-2.1.92	`/tag`, `/vim` commands	Removed; editor mode via `/config`
2.1.86	250-char skill description cap	Raised to 1536 in 2.1.105
2.1.88	Ctrl+O = focus view	`/focus` command (2.1.110); Ctrl+O is verbose-only
2.1.76	`/fork`	`/branch`
2.1.118	`/cost`, `/stats` commands	`/usage` (both remain as typing shortcuts)
Pre-2.1.120	Git for Windows requirement on Windows	PowerShell fallback when Git Bash absent (2.1.120)