Claude Code Version Tracker: Environment Variables, Feature Gates & Models#

Published March 7, 2026 by Alex Sorokoletov

We analyze every Claude Code release to document undocumented environment variables, feature gates, and model IDs. This page tracks all versions we've covered.

Looking for the full list? See The Complete List of Claude Code Environment Variables (500+) — every env var organized by category with practical tips.

Version History#

v2.1.303 — October 18, 2026#

CLAUDE_CODE_AUTO_MODE_MODEL · CLAUDE_CODE_ALLOW_OTHER_MODELS · fork subagents default on · idle compaction toggle · 128 MB image cap

v2.1.303 is a bug-fix build; the v2.1.302 release before it carries the user-visible changes since v2.1.301. Fork subagents now run on by default in claude -p, the Agent SDK, the desktop app and VS Code, with CLAUDE_CODE_FORK_SUBAGENT=0 to turn them off. /config gains an Idle compaction row, and CLAUDE_CODE_AUTO_MODE_MODEL chooses the client-side auto mode classifier model, a variable present in the build since at least v2.1.260 but named only now. The Read tool refuses image sources over 128 MB, and CLAUDE_CODE_ALLOW_OTHER_MODELS, CLAUDE_CODE_MCP_SERVE_WORKSPACE_ROOT and CLAUDE_BG_GATEWAY_WITHHELD first appear in v2.1.302.

v2.1.301 — October 15, 2026#

CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS · CLAUDE_CODE_ATTACHMENTS_IN_USER_DATA · safer model downgrades · classifier telemetry · vim % motion

Claude Code v2.1.301 adds CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS, which bounds how long unattended retry mode waits for a usage-limit reset: past the budget a 429 ends with an api_request_usage_limit_wait_too_long error instead of sleeping until the reset, alongside the overall CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS cap. Switching to a model with a smaller context window now summarizes an over-long conversation with the previous model first, no longer spends credits on that summary without asking, and stops the repeated compaction that happened behind an LLM gateway naming a different model than requested; the fable/best alias falls back once to the previous Fable model when the newest is unavailable. OpenTelemetry tool_decision events gain classifier_verdict, permission_mode and classifier_category to separate auto-mode classifier blocks from permission rules. A flag new to this release, CLAUDE_CODE_ATTACHMENTS_IN_USER_DATA, is wired but absent from the changelog: off unless set, it routes remote-session attachment reads through a user-data root behind a retention check. Quality-of-life fixes cover the vim % motion, VS Code LaTeX rendering, the Read tool refusing binary files, and Grep/Glob reporting search failures.

v2.1.300 — October 14, 2026#

CLAUDE_CODE_MESSAGES_PROMPT_CACHE_TTL · CLAUDE_CODE_FILE_READ_MAX_PDF_PAGES_PER_TURN · CLAUDE_CODE_DISABLE_ADVISOR_TOOL · gzip on by default · PDF read caps

Claude Code v2.1.300 is a large release weighted toward fixes, with a handful of new tuning knobs. CLAUDE_CODE_MESSAGES_PROMPT_CACHE_TTL set to 5m pins the messages prompt cache to the five-minute lifetime, matching a new telemetry split into cache_creation_5m_tokens and cache_creation_1h_tokens. CLAUDE_CODE_FILE_READ_MAX_PDF_PAGES_PER_MESSAGE and CLAUDE_CODE_FILE_READ_MAX_PDF_PAGES_PER_TURN bound how many PDF pages the Read tool loads. Large requests to api.anthropic.com now use gzip by default (opt out with CLAUDE_CODE_GZIP_REQUEST_BODIES=0), and the advisor tool arrives on Claude Platform on AWS, off until selected and disabled with CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1. The bulk of the release corrects worktree and branch handling in reftable repositories, session resume and compaction history, and cloud-session behavior.

v2.1.299 — October 14, 2026#

CLAUDE_CODE_DISABLE_FILE_MENTIONS · CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS · CLAUDE_CODE_STRICT_INLINE_SHELL_RM_PROMPT · idle compaction

Claude Code v2.1.299 documents idle compaction, running since v2.1.283: a conversation past 200K tokens, left idle for about 54 minutes, is compacted before its one-hour prompt cache expires, so the next message starts from the compacted state instead of reprocessing the whole history. "idleCompaction": false turns it off and CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS (present since v2.1.290, documented now) moves the 200K floor. CLAUDE_CODE_DISABLE_FILE_MENTIONS makes an @ mention hand Claude the file's path to read with its own tools rather than inlining the contents, keeping large files out of the prompt. The rm permission prompt no longer fires on shell scripts that never run rm, with CLAUDE_CODE_STRICT_INLINE_SHELL_RM_PROMPT=1 to keep the strict check, and MCP allow rules now reach only the whole server name they name.

v2.1.296 — October 14, 2026#

allow_large Read option · autoCompactWindow subagents · CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL · CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS · long fix list

v2.1.296 adds an allow_large option to the Read tool that reads a whole oversized text file in one call when the context window has room, and an autoCompactWindow field on subagent frontmatter and --agents that lets a subagent compact earlier than the main conversation without touching the parent's budget. CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL pins every workflow agent to one model while other subagents keep theirs, and CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS raises the backoff ceiling for 529 overloaded retries. The release lands after v2.1.295 with a long fix list: managed-settings hooks that block now end the turn, plugins survive reserved names like constructor, Edit refuses non-UTF-8 files instead of corrupting them, a BASH_ARGV0 permission gap now prompts, and Windows PowerShell commands up to 32 KB get allow-rule checks.

v2.1.295 — October 14, 2026#

CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS · CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS · CLAUDE_CODE_RESTRICT_PERSONAL_CONFIG · onFailure: "block" hooks · MCP reconnect backoff

Claude Code v2.1.295 adds onFailure: "block" for command and HTTP hooks, so a hook that cannot start, times out, or exits with an unexpected code blocks the action instead of letting it through. Remote MCP servers in headless and SDK sessions reconnect after outages longer than 15 seconds and back off up to 30 seconds when a server drops each connection, and MCP results in CSS, JavaScript, and XML keep their own extensions instead of being saved as .bin. CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS caps how long unattended retry mode waits through 429 and 529 responses, and CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS bounds a self-hosted runner's git fetch. Two more variables, CLAUDE_CODE_RESTRICT_PERSONAL_CONFIG and CLAUDE_CODE_WEBSEARCH_CITATIONS, ship in v2.1.295 without a release-note entry.

v2.1.294 — October 14, 2026#

prompt + agent hooks · rule vs condition judge · Stop/SubagentStop early-stop fix · hook-data injection guard

v2.1.294 fixes prompt and agent hooks written as plain-English instructions. In v2.1.293 the model that judged these hooks asked only whether a condition was met, so a hook phrased as a rule to block something could be read as satisfied and let the action through. The judge now runs under one contract that separates a rule to apply from a condition to test, and returns the same ok/reason shape, so existing hooks need no change. The same wording reaches the Stop and SubagentStop evaluator, where the old confusion made Claude stop too early, and a new line tells the judge to ignore any instruction embedded in the event data it reads.

v2.1.293 — October 7, 2026#

claude-haiku-5-5 (1M context, $0.10/$0.50) · subagentStatusLine agentType · $.tool.register isDeferred · compaction + MCP leak fixes

v2.1.293 adds Claude Haiku 5.5 (claude-haiku-5-5), a 1M-context model that costs $0.10 input and $0.50 output per million tokens ($0.50/$2.50 above 100K-token prompts) and is the default Haiku on the Anthropic API. Two scripting additions land with it: a subagent status-line script now receives agentType in its JSON payload, and a mod can set isDeferred on $.tool.register to list a registered tool's schema in the prompt from the start instead of behind tool search. The fix set stops Claude treating its own pre-compaction actions as done and retracting finished work, closes a memory leak where an HTTP MCP connection held every request until it closed, keeps a queued message from being lost when the left arrow backgrounds a session, stops the /model effort selector wrapping and saving Low by accident, and keeps /tui from disconnecting Claude in Chrome under --chrome.

v2.1.292 — October 6, 2026#

effort sub-agents · plugin install --marketplace · CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS · UNC read-prompt fix · CLAUDE_CODE_ARTIFACT_VERSIONS

Claude Code v2.1.292 adds an effort parameter to the Agent tool so a sub-agent runs at the level you ask for, and a --marketplace flag to claude plugin install that registers a marketplace and installs a plugin from it in one command. CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS lengthens the backoff on overloaded (529) requests. A security fix stops PreToolUse hook approvals and auto mode from skipping the permission prompt for file reads from network (UNC) paths, and three more fixes keep a read from reaching a file outside the approval through a link swapped in mid-read, the staged copies of /ultrareview uploads, or a managed read-deny path that changes mid-session. New CLAUDE_CODE_ARTIFACT_VERSIONS, CLAUDE_CODE_ARTIFACT_PREVIEW_EMULATOR, and CLAUDE_CODE_HOST_SKILL_CATALOG sit on the hosted artifact and skill surface, and plan mode, scheduled tasks, and a background /loop now survive a resume or restart.

v2.1.291 — October 6, 2026#

cloud permission-prompt fix · session-quit message-loss fix · 2 regression fixes

Claude Code v2.1.291 is a patch over v2.1.290. Its release notes list two fixes, both regressions: cloud sessions could drop your answers to permission prompts, and a session's last messages could be lost when you quit. Both are resolved in this build, which makes it a reliability-only patch for permission prompts and session shutdown.

v2.1.290 — October 6, 2026#

CLAUDE_CODE_WEB_SEARCH_REFILLS_PER_HOUR · CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS · CLAUDE_CODE_DISABLE_PROACTIVITY · WebFetch paging · scheduled tasks survive resume

Claude Code v2.1.290 hardens behavior on large inputs and makes scheduled tasks durable. WebFetch now reports how much of a page past 100,000 characters it left unread and takes an offset to read on, image-heavy sessions no longer stall on an unprocessable-request error, deeply nested replies no longer crash, and compaction stops failing intermittently. Scheduled tasks from /loop and reminders survive resume after compaction and a foreground-to-background hand-off. New environment variables set the hourly web-search budget refill rate (default 100, max 3600), the minimum token count before idle compaction runs (floored at 100,000), and an off switch for the proactivity selector. Plugin tool.check hooks can read an organization-required approval ceiling and a subagent agentId, turn.step hooks gain serverToolUses, and several fixes close reads that reached files outside an approved rule.

v2.1.289 — October 3, 2026#

agent.spawn hook · ui.fault · $.agent.list() idle/waiting · symlink-aware deny rules

v2.1.289 tightens where permission rules take hold and what a plugin runs inside. Deny and ask rules now apply to a nested part of a compound shell command, a Bash command behind an environment-variable prefix (TZ="$HOME" rm -rf build), and a file reached through a symlink, and on managed machines they hold against a user-installed mod or plugin. For plugin authors it adds an agent.spawn hook for teammates with one agent id across hook events, idle and waiting states in $.agent.list(), and a ui.fault that a failing Client raises so the mod can draw a replacement view instead of taking down the session.

v2.1.288 — October 2, 2026#

CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS · CLAUDE_CODE_DISABLE_INLINE_SHELL_RM_PROMPT · Ctrl+C draft recovery · dangerous-rm prompt

v2.1.288 restores a prompt you cleared with Ctrl+C when you press Up on the empty line, pasted text and images included, and now asks before a dangerous rm hidden in a bash -c or sh -c script runs under bypassPermissions or a shell allow rule. /code-review takes --max-findings, the agents view adds Ctrl+F and Alt+arrow group jumps, mods gain $.ui.selection(), and MCP OAuth scope escalation prompts to re-authenticate. Two new environment variables act as opt-outs: CLAUDE_CODE_DISABLE_STRUCTURED_OUTPUTS for gateways such as Mantle that reject structured outputs, and CLAUDE_CODE_DISABLE_INLINE_SHELL_RM_PROMPT to restore the old silent rm behavior. A large reliability cluster keeps turns alive through mid-response timeouts, zero-usage auto-compaction, and resume paths that hold on to restored context.

v2.1.287 — October 1, 2026#

Claude Mods · bareElicitationCapability · MCP URL prompts · CLAUDE_CODE_GZIP_REQUEST_BODY_BLOCKS · plugin directory

v2.1.287 extends plugins with Claude Mods, which register render sites, interfaces, and hooks that change deeper session behavior; the opt-in You-should-know plugin runs a side agent that flags things you or Claude might miss. A plugin directory backed by the version-pinned Anthropic Directory catalog adds discover, enable, disable, and uninstall screens. MCP servers on the 2025-11-25 protocol can send URL sign-in prompts, with bareElicitationCapability trimming the capability for servers on older Java MCP SDKs. The OpenTelemetry user_prompt event gains a redaction-aware prompt_text, a dangerous rm keeps its always-ask safeguard when it also redirects to a ~ or wildcard path, and CLAUDE_CODE_GZIP_REQUEST_BODY_BLOCKS selects a gzip mode for large request bodies.

v2.1.286 — September 30, 2026#

CLAUDE_CODE_DISABLE_AUTH_REFRESH_LOCK · same-tier model fallback · 2 of 5 permission counts · credential redaction

Claude Code v2.1.286 makes a refused model survivable: it retries once on the previous model of the same tier instead of failing every turn, runs slow fallbacks at standard speed, and caps a failing call at 14 requests. Stacked permission prompts now show a "2 of 5" count, and CLAUDE_CODE_DISABLE_AUTH_REFRESH_LOCK toggles a new lock that lets one process handle a GCP or AWS credential refresh so several processes no longer each open a login browser. A batch of redaction fixes keeps bearer tokens, invisible-character key names, and punctuation-laden URL passwords out of error messages and logs, and plugin installs now take only registry packages, refusing git repositories and folders.

v2.1.285 — September 29, 2026#

CLAUDE_CODE_DISABLE_WEB_FETCH · allowedProviders · claude plugin configure · CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES · reliability fixes

Claude Code v2.1.285 gives administrators more to pin per machine: CLAUDE_CODE_DISABLE_WEB_FETCH turns off the WebFetch tool without a policy file, and the allowedProviders managed setting limits which API providers a machine may use. claude plugin configure reads and sets a plugin's options from the command line, plugin install --config sets a bundled MCP server's settings at install time, and CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES caps re-sends of a timed-out non-streaming request. CLAUDE_CODE_DISABLE_MODEL_ACCESS_FALLBACK is present in the build but not named in the release notes. The rest is request-path and Remote Control fixes.

v2.1.284 — September 28, 2026#

claude-sonnet-5-5 · auto mode default · Ultracode toggle · outside-read ask-again

v2.1.284 adds Claude Sonnet 5.5 as the default Sonnet model on the Anthropic API, with a 1M-token context window and $2/$10 per Mtok pricing. Interactive terminal and VS Code sessions now start in auto mode on every plan and provider, extending the default-mode shift that v2.1.283 began on third-party providers, and auto mode's prompt before an outside-directory read gains a 'Yes, but ask again next time' answer that allows one read while keeping later ones gated. Ultracode splits into its own /effort toggle with rebindable slider keys instead of forcing xhigh effort.

v2.1.283 — September 26, 2026#

availableModelsMatch · deniedModels · CLAUDE_CODE_GATEWAY_HINT_HEADERS · /doctor prompt-audit · auto-mode default

v2.1.283 gives admins a fixed model set: availableModelsMatch "exact" allows only the model versions named in availableModels, so a new release stays blocked until listed, and deniedModels blocks specific models even when the allowlist would permit them. Outbound requests can carry an x-claude-code-prompt-id header so an LLM gateway groups the calls that serve one prompt; it is opt-in through CLAUDE_CODE_GATEWAY_HINT_HEADERS=1. /doctor prompt-audit audits CLAUDE.md files and prompts for prompting written for older models, and interactive sessions on third-party providers now start in auto mode when no permission mode is set.

v2.1.269 — September 11, 2026#

CLAUDE_CODE_BASH_EDIT_DIFF · CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS · CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS · plugin eval reports · 36 new env vars

v2.1.269 lands claude plugin eval, which scores a plugin's own eval suite against Claude Code into a reproducible JSON and HTML report, and folds a diff of the files a Bash command changed into the Bash tool result (controlled by the bashEditDiffEnabled setting, with a CLAUDE_CODE_BASH_EDIT_DIFF override present in the build). CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS raises the Workflow tool's per-run concurrent-agent limit (1 to 256), CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS extends the gateway /v1/models discovery timeout, and /output-style now switches styles over Remote Control and in headless sessions. The 267-to-269 window also restores prompt-cache reuse that broke on resume after an interrupt and on the turn after an output-token cutoff.

v2.1.266 — September 9, 2026#

Fable 5.1 default · /diff panel · 128K output caps · --permission-prompts none · Containment Escape rule

Across v2.1.253 through v2.1.266, Claude Fable 5.1 becomes the default Fable model with a 1M-token window and $10/$50 per Mtok pricing, then gets its picker, prompt-caching, effort, and auto-compact behavior fixed over the following releases. A fullscreen /diff panel shows uncommitted changes as Claude edits, and /cost now names likely prompt-cache-miss causes. Output handling grew: bashOutputMaxChars and taskOutputMaxChars go up to 128K, and saved tool results are capped at 1 GB. Auto mode tightened: a Containment Escape rule and permissions.blockReadsOutsideWorkingDirectories stop cloud credential fetches and outside-directory reads from being auto-approved, and reading an artifact you don't own always asks. Version 2.1.266 itself is a single fix restoring CLAUDE_CODE_USE_GATEWAY to being ignored when set alone.

v2.1.252 — August 31, 2026#

CLAUDE_CODE_MODEL_CATALOG · CLAUDE_CODE_REMOTE_TOOLS_POLICY · CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND · PreModelSwitch/PostModelSwitch hooks · symlink permission fixes

v2.1.252 lands two new hook events, PreModelSwitch and PostModelSwitch, that can block, confirm, or annotate a model switch and gate fast-mode promotion, plus a batch of permission hardening: Read, Write, and Edit now resolve a symlink before acting instead of after the check, Grep and Glob apply Read deny rules through symlinked search paths, and plugin marketplace commands and Workflow scriptPaths can no longer reach outside their scope. /cost shows a per-session prompt-cache line and /usage a spend-limit bar, both with matching status-line fields. Four new environment variables surface underneath: CLAUDE_CODE_MODEL_CATALOG forces the account-scoped served model catalog off, CLAUDE_CODE_REMOTE_TOOLS_POLICY picks parity or interim serving to Remote Control clients, CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND keeps MCP tool calls in the foreground, and CLAUDE_CODE_AUTO_BACKGROUND_WORKER_CHECKIN_SECONDS sets how often a dispatched background worker reports progress.

v2.1.247 — August 27, 2026#

/claude-api cost-optimize · spinnerTipsOverride.tipsFile · feedbackDrafts · Admin API docs · Sonnet 5 1M compact

v2.1.247 lands operator tooling: a /claude-api cost-optimize skill that profiles an existing project's API spend and steps through caching, token hygiene, batching, effort and model choice, alongside Admin API documentation for usage, rate limits, workspaces, members and workload identity federation. spinnerTipsOverride gains a tipsFile field (absent in v2.1.245) so an admin can rotate an organization's tips from a JSON file through managed settings, with UNC paths rejected and per-entry id/text validation. Sessions on Sonnet 5's full 1M window now auto-compact near 967K tokens instead of 934K, and the Bash permission prompt points at auto mode with a one-keystroke switch. The changelog credits this release with the SendFeedback tool, but that tool, its ~/.claude/feedback/drafts store, per-session cap and transcript guards already shipped in the v2.1.245 baseline.

v2.1.245 — August 25, 2026#

CLAUDE_CODE_PROMPT_CACHE_TTL · CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS · CLAUDE_CODE_FEDERATION_CACHE_DIR · glibc 2.44 startup fix

v2.1.245 fixes a startup crash on Linux distributions that ship glibc 2.44 (Arch, CachyOS, Fedora Rawhide). Folded into this window from v2.1.243, two automatic behaviors become per-session controls: prompt-cache TTL now resolves through the CLAUDE_CODE_PROMPT_CACHE_TTL and CLAUDE_CODE_SUBAGENT_PROMPT_CACHE_TTL environment variables above the promptCacheTtl settings, letting a job force 5m or 1h regardless of account default; and a stalled API request that used to hang silently for ten-plus minutes now times out at the first byte (CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS, gated by tengu_api_no_response_timeout), retries once, then reports an error. CLAUDE_CODE_FEDERATION_CACHE_DIR points at the shared workload-identity-federation credential so CI processes in one job reuse the exchanged token instead of re-exchanging a single-use one.

v2.1.241 — August 24, 2026#

CLAUDE_CODE_ARTIFACT_DELETE · CLAUDE_CODE_THINKING_DISPLAY_UPDATES · lower-priority mode · effort-medium nudge · Bedrock double-billing fix

Three builds in four days, but only v2.1.239 carries a real changelog. Two user-facing features never made the notes: deleting an Artifact (CLAUDE_CODE_ARTIFACT_DELETE, owner-only, confirmation required, blocked in plan mode and remote sessions) and a lower-priority mode that keeps a Max session working after it hits the weekly limit, offered again after each reset. A one-time nudge to switch default effort to medium landed in v2.1.240 under 'bug fixes.' The named wins are a Bedrock streaming fix that stops proxies from silently doubling billed API calls, an HTTPS-proxy startup hang fix, and Chrome tab-group cleanup on /clear.

v2.1.238 — August 20, 2026#

CLAUDE_CODE_MARKETPLACE_URL · CLAUDE_CODE_PLUGIN_ARCHIVE_URL · headersHelper token minting · readline Ctrl+W · Remote Control resilience

v2.1.238 adds a headersHelper hook to plugin marketplaces: a command on a URL marketplace or catalog entry mints HTTP headers (such as a short-lived token) for catalog reads and same-origin archive downloads, running only at install or update after its command is shown, with a [y/N] consent prompt. Claude Code hands the helper a minimal environment (CLAUDE_CODE_MARKETPLACE_URL/_NAME for marketplaces, CLAUDE_CODE_PLUGIN_NAME/_ARCHIVE_URL for plugins), scrubs inherited credential env vars when the source is not operator-authored, and filters out request-routing and identity headers from untrusted sources. Alongside it: a fix for unbounded memory growth in long interactive sessions (subagent tool results are released past the display window), output styles no longer drifting back to the default voice, a wide Remote Control reliability pass, and a new keybindingFlavor setting whose readline mode makes Ctrl+W delete back to the previous whitespace. A new CLAUDE_CODE_ARTIFACT_ROOM variable and tengu_artifact_verify gate first appear here with no changelog mention, extending the artifact collaboration surface.

v2.1.237 — August 20, 2026#

ANTHROPIC_DEFAULT_MODEL · CLAUDE_CODE_SILENT_TURN_REMINDER · Concise output style · notify_when_idle

v2.1.237 rolls up three builds since v2.1.234. A built-in Concise output style trims preamble and narration while keeping error, test, and security output intact. ANTHROPIC_DEFAULT_MODEL sets the model new sessions start on, with a /model pick still overriding it and persisting across restarts, unlike ANTHROPIC_MODEL. SendMessage gains notify_when_idle for a one-shot cross-session idle ping on macOS and Linux, and an optional spellcheck setting underlines misspellings in the prompt. The build also carries CLAUDE_CODE_SILENT_TURN_REMINDER and its _TURNS threshold, present since v2.1.236 but absent from that release's notes, which inject a reminder for Claude to surface progress after a set number of silent turns (default 5).

v2.1.234 — August 17, 2026#

autoContinueAtUsageLimit · CLAUDE_CODE_GOAL_CHECKIN_MINUTES · CLAUDE_CODE_ENABLE_TODO_TOOLS · GitLab MR badge · +65 env vars

v2.1.234 continues a session automatically when a claude.ai usage limit resets, backed by a quota_auto_resume state machine and the autoContinueAtUsageLimit setting, with a /config toggle to turn it off. GitLab reaches parity with GitHub: a merge request badge in the footer and statusline (MR !N with draft/pending/green states), plus MR URLs in --worktree and the agents view. /goal now clears itself when a turn dies on an unrecoverable error and checks in when background work stalls 30+ minutes (CLAUDE_CODE_GOAL_CHECKIN_MINUTES). Todo and task tools are off by default on Opus 4.8, Sonnet 5, Fable 5, and Mythos 5, restorable with CLAUDE_CODE_ENABLE_TODO_TOOLS=1. Windows NT-namespace (??) paths are now rejected across remote reads, session restore, CLAUDE.md includes, workflow scripts, and uploads, and the built-in claude-api skill dropped from ~200k tokens to 25k.

v2.1.229 — August 12, 2026#

CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS · CLAUDE_CODE_DISABLE_DIR_SYNC · plugin command sources

v2.1.229 adds plugin marketplace command sources. A local program prints a plugin directory. Claude Code re-runs the command each session and applies the plugin with no restart (mode: "link", backed by the new tengu_plugin_command_source_refresh gate). An IDE can keep its own plugin current with no reinstall. CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS sets how long a workflow agent waits for a same-prefix sibling's first response, so it reads a warm prompt-cache prefix and sends fewer input tokens (0 disables). The rest is reliability work: long responses stop duplicating and vanishing mid-stream, non-string glob/file_path/command tool inputs stop crashing to the error screen, narrow terminals stop throwing RangeErrors, Windows handles extended-length paths, and SSE keepalive pings hold Vertex and Bedrock streams open through long thinking pauses.

v2.1.228 — August 12, 2026#

Write overwrites unread files · betaAgentToolset20260401 · frozen-session recovery · GCP/AWS auth fast-fail · CLAUDE_CODE_MESSAGING_TOKEN

The main change in v2.1.228 is a new Write tool rule: newer models can overwrite a file that Claude has not read this session, which matches the Edit tool. The old rule required a read first as a safety check, so the change removes one tool call and one guard. The rule applies only to newer models through a dated betaAgentToolset20260401 bundle (Bash, Read, Write, Edit, Glob, Grep), so older models still read first and one build can run both rules. The rest of the release is reliability work: a rare layout error that froze interactive sessions now recovers, expired Google Cloud and AWS logins fail in seconds instead of retrying for minutes, and two cleanup bugs that deleted a memory folder or a live plugin checkout are fixed. Two new environment variables ship: CLAUDE_CODE_MESSAGING_TOKEN controls access to the local cross-session messaging inbox, and NO_HEARTBEAT skips a worker session keep-alive signal.

v2.1.227 — August 11, 2026#

CLAUDE_CODE_ARTIFACT_COMMENT_RESPONDER · CLAUDE_CODE_MEMORY_API_TOKEN · bashCommandClamp · comment triage agent

The v2.1.227 notes list only fixes: a subscription-tier check that no longer prompts Max members to turn on Fable credits on an expired token, a claude-code-action Bash fix on GitHub runners, a /tui rewind fix, and clearer slash-command menu highlighting. The build holds three features the notes skip, all first appearing in this build. A built-in read-only comment-thread-analyst subagent sorts artifact comment threads into an edit request or a written reply, turned on with CLAUDE_CODE_ARTIFACT_COMMENT_RESPONDER behind the default-off gate tengu_bracken_sluice. CLAUDE_CODE_MEMORY_API_BASE_URL and CLAUDE_CODE_MEMORY_API_TOKEN point Claude Code's memory-sync client at the Managed Agents Memory Stores REST API (managed-agents-2026-04-01 beta header); the full endpoint reference, pseudocode, and enable-and-test env vars are in a companion post, Claude Code's Org Memory: the Memory Stores REST API. A per-spawn bashCommandClamp limits an agent's Bash use to a fixed set of command forms, refuses anything else, and fails closed if the permission check crashes.

v2.1.224 — August 7, 2026#

claude self-hosted-runner · CLAUDE_CODE_MESSAGING_SOCKET · ANTHROPIC_BEDROCK_REGION_PREFIX · cross-session SendMessage

v2.1.224 pushes Claude Code past a single local process. claude self-hosted-runner turns your own machines or containers into an execution pool for web, mobile, and desktop sessions on Team and Enterprise plans, installed as a launchctl/systemd service with a local health check and claude_code_self_hosted_runner_* Prometheus gauges. A cross-session SendMessage tool, paired with ListAgents, lets sessions message each other on macOS and Linux over a local-only Unix domain socket named by CLAUDE_CODE_MESSAGING_SOCKET and gated by agents_cross_session_inbox; the crossSessionInbound setting holds messages from permission-bypassing sessions for approval, and dialogExpiry bounds how long a forwarded dialog or held message waits. Bedrock gains ANTHROPIC_BEDROCK_REGION_PREFIX to prefer a cross-region inference profile, the 200-subagent-per-session cap is removed, and fixes close a long-path session cross-contamination bug and a trailing-slash sandbox deny bypass.

v2.1.222 — August 4, 2026#

tengu_worktree_resume_root_rejected · tengu_streaming_close_after_complete · family_step_down · worktree isolation in every session · SendMessage classifier

v2.1.222 is a hardening pass on session boundaries. Worktree-isolated sessions and their subagents can no longer run destructive git against the main checkout; isolation now covers file edits and Bash in every session type, with resume revalidating the recorded worktree and rejecting a missing, network-spelled, or unverifiable root (tengu_worktree_resume_root_rejected). PreToolUse auto-allow hooks stop bypassing tool restrictions in background tasks, and SendMessage to another agent runs through the permission classifier first. Connection fixes land alongside: the startup check goes proxy-aware, tengu_streaming_close_after_complete stops false 'connection closed mid-response' errors on finished turns, and stream idle timeouts no longer fire on keep-alive gateways. Org model: opus aliases now step down to the newest org-allowed model instead of the parent (family_step_down), and the ultraplan feature is removed.

v2.1.221 — August 4, 2026#

CLAUDE_CODE_ARTIFACT_COMMENTS · MCP_PROTOCOL_NEGOTIATION · CLAUDE_CODE_ARTIFACTS_API_TOKEN · sandbox credential masking · Bash bypass fix

v2.1.221 closes a Bash permission-check bypass where zsh could execute hidden commands inside [[ ]] regex conditionals, and adds a mode: "mask" for sandbox credential files on Linux and WSL that gives sandboxed commands a sentinel copy while the proxy substitutes the real value only on egress (macOS falls back to deny). MCP connections get a resilience pass: a new MCP_PROTOCOL_NEGOTIATION variable (legacy or auto) probes servers to negotiate a modern protocol revision per transport, print-mode servers connect before the first turn, and a wake-from-sleep double-refresh race is closed. Absent from the changelog, and confirmed to have landed in this build rather than the maintenance-only v2.1.220, is a full artifact comment-thread loop: CLAUDE_CODE_ARTIFACT_COMMENTS and CLAUDE_CODE_ARTIFACT_COMMENTS_AUTOREACT (over gates tengu_teal_corbel and tengu_sorrel_trellis) let Claude publish a prototype as a live artifact and then read, reply to, and edit it in response to human comments, authenticated by CLAUDE_CODE_ARTIFACTS_API_TOKEN. The org-memory path gained a matching write-consent step that asks before a durable project write.

v2.1.219 — July 24, 2026#

claude-opus-5 default · 1M context · $10/$50 per Mtok · claude -p fix · Vim + screen-reader QOL

v2.1.219 swaps the default Opus model: the opus alias default flips from claude-opus-4-8 to a new registry entry, claude-opus-5, carrying a native 1M-token context window, 64k/128k output, a tier_10_50 ($10/$50 per Mtok) tag, a May 2026 knowledge cutoff, and a full provider matrix (Foundry alone still defaults opus to 4.6). Around the model the changelog is fix-heavy in the way daily users notice: claude -p keeps the answer it already produced when a turn dies mid-stream, copy-on-select stops printing base64 inside GNU screen, Vim's left-arrow returns to the agent view, and screen-reader mode echoes only the typed character instead of rewriting the line. New capability lands too — a sandbox.network.strictAllowlist deny setting, a DirectoryAdded hook, mcp_server_errors in the headless stream-json init plus HTTP error text in claude mcp list, a workflowSizeGuideline default of under 15 agents, and nested subagents to depth 3 by default. The one undocumented addition, CLAUDE_CODE_DISPATCH_V2S, forces the request header anthropic-dispatch-id: v2s onto every first-party call; the header, value, and its server gate tengu_cedar_lattice all shipped in v2.1.218, so v2.1.219 adds only the per-machine override that opts into the new serving path without waiting for the percentage rollout.

v2.1.218 — July 23, 2026#

CLAUDE_CODE_AUTO_MODE_CLASSIFIER_QUEUE · CLAUDE_CODE_DISABLE_MEMORY_MASS_DELETE_HOLD · 4 new env vars

v2.1.218 is a cleanup release with real day-to-day payoff. /code-review moves to a background subagent so a review stops eating the conversation you were in, skills marked context: fork run in the background by default, the left-arrow key can no longer discard a session with no undo, and a retry loop that re-sent the same doomed request after a context overflow is closed (real money back on metered plans). Under the notes, the build adds four environment variables they never mention. CLAUDE_CODE_AUTO_MODE_CLASSIFIER_QUEUE puts auto mode's per-turn model classifier on a queue with a depth and wait bound (queueDepth, queueWaitMs) beside a cumulativeDroppedTokens counter, so a busy router lines requests up instead of stalling. CLAUDE_CODE_DISABLE_MEMORY_MASS_DELETE_HOLD controls a default brake on bulk deletion now that memory syncs to account and team stores: normally a large delete is held past a threshold, and setting the flag pushes that threshold to infinity so the hold never fires. Two codename gates, GORSE_PLOVER and JUNIPER_SUNDIAL, arrive as bare server flags while ALDER_WICKET is dropped. Both silent knobs guard a subsystem that recently went shared, which points at where the load is landing.

v2.1.217 — July 22, 2026#

CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS · CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH · CLAUDE_CODE_FLEETVIEW_SIMPLE · 12 new env vars

v2.1.217 turns the multi-agent fan-out down. A hard concurrency cap (CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS, default 20) limits how many background subagents run at once, and a new depth cap (CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH, default 1) stops subagents from spawning their own — both behind server gates, and, with the --max-budget-usd fix, the release bounds fan-out on the width, depth, and cost axes at once. They are the two additions the changelog names out of twelve new variables; a FleetView simple-render toggle (shipped a build early in v2.1.216), an Explore-agent model-inherit cap, a local transcript garbage-collector, a cloud-session origin flag, and five codename rollout gates all arrive undocumented. The build is a 239 MB Bun-compiled arm64 executable, up ~3 MB in two versions, with twelve genuinely new variables.

v2.1.215 — July 18, 2026#

CLAUDE_CODE_USE_ANTHROPIC_GOOGLE_CLOUD · CLAUDE_CODE_NO_MODEL_FALLBACK · CLAUDE_PID · 5 new env vars

Claude Code v2.1.215 adds five environment variables over the last analyzed build v2.1.212. The headline is CLAUDE_CODE_USE_ANTHROPIC_GOOGLE_CLOUD, a sixth first-party provider backend sitting beside USE_BEDROCK/USE_VERTEX/USE_ANTHROPIC_AWS with its own ANTHROPIC_GOOGLE_CLOUD_PROJECT config and a dedicated skip-auth flag — an Anthropic-operated inference plane on Google Cloud, not customer Vertex. CLAUDE_CODE_NO_MODEL_FALLBACK collapses the availability chain to the model you selected and throws a tripwire if any path attempts a fallback, even refusing compaction that would swap models. CLAUDE_CODE_OTEL_CONTENT_MAX_LENGTH caps telemetry content at 60 KB, and CLAUDE_PID powers a pkill shim that stops the agent from killing its own process. The changelog lands a permission-check hardening pass, a background-session lifecycle cleanup, reasoning effort in the subagent status line, and stops Claude from auto-running /verify and /code-review.

v2.1.212 — July 17, 2026#

CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION · CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION · CLAUDE_CODE_SEND_FEEDBACK · /bug · 11 new env vars

v2.1.212 adds eleven environment variables over v2.1.210. The two headliners are guardrails: CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION and CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION both default to 200 and both inject a stop-and-ask prompt when hit, containing the two operations an autonomous session can multiply without bound — delegation and search. The changelog confirms both caps and pairs them with a background-session rework: /fork now copies into a background session, the in-session subagent becomes /subtask, and the Task tool's mode parameter is deprecated. A silent /bug command lands beside them — a first-party feedback relay distinct from /feedback, fenced behind an allow_product_feedback capability, a firstParty account check, the server gate tengu_juniper_relay, and the CLAUDE_CODE_SEND_FEEDBACK toggle — none of it documented, and the same is true of a Windows-only sandbox gate (CLAUDE_CODE_NANKEEN_KESTREL / tengu_nankeen_kestrel). Two documented fixes matter on their own: plan mode no longer auto-runs file-modifying Bash without a permission prompt, and worktree creation no longer follows a committed .claude/worktrees symlink out of the repo.

v2.1.210 — July 15, 2026#

CLAUDE_CODE_DISABLE_ORG_MEMORY · CLAUDE_CODE_SYNC_SESSION_REFS · tengu_haze_glass · 4 new env vars

v2.1.210 adds four environment variables, three of them for an account-aware memory-sync subsystem that did not exist in v2.1.209. CLAUDE_CODE_DISABLE_ORG_MEMORY kills a new organization-scoped memory path fenced behind a fresh server gate (tengu_haze_glass, default off) and an allow_memory_sync permission; CLAUDE_CODE_DISABLE_MEMORY_STREAM_LIST toggles the incremental metadata-streaming path (gate tengu_memory_stream_list, with retry backoff). CLAUDE_CODE_SYNC_SESSION_REFS propagates session references into remote and background sessions alongside the account and organization UUIDs, mirroring how skills already sync. CLAUDE_CODE_FORCE_WINDOWS_CREDMAN pins credential storage to the Windows Credential Manager. Two new tengu_* gates confirm these are live subsystems, and — as usual — the wiring ships before the changelog names it.

v2.1.209 — July 14, 2026#

CLAUDE_CODE_AUTO_MODE_MODEL · CLAUDE_CODE_AUTO_MODE_EDIT_REMOVAL_CAP · CLAUDE_CODE_PROCESS_WRAPPER · CLAUDE_CODE_WORKFLOWS · Auto mode grows up

A two-month catch-up on Claude Code. The clearest story in the new environment-variable surface is Auto mode graduating from experiment to guarded, default-on product. A cluster of fourteen CLAUDE_CODE_AUTO_MODE_ variables reads as a rare, legible window into how autonomous editing works: a two-model design where an editor proposes changes and a separate cheaper classifier decides what's safe to auto-apply, wrapped in guardrails — an edit-removal cap, repo-visibility gating, external-permission hooks, git-status context limits, and structured outcome codes. The last ten changelogs confirm the direction: Auto mode is now default-on on Bedrock/Vertex/Foundry, the default permission mode became 'Manual', Opus 4.8 is the managed default, workflows gained a size setting and OTel attributes, and a genuine screen-reader mode shipped. Around Auto mode the surface hardens a distributed runtime — a bridge/remote-control layer so sessions can be driven from phone and web, centrally synced skills and plugins, and enterprise controls led by CLAUDE_CODE_PROCESS_WRAPPER, which forces every self-spawn through a required corporate launcher.

v2.1.143 — May 18, 2026#

CLAUDE_CODE_MAX_TURNS · CLAUDE_CODE_SUPERVISED · /goal · Agent View · Opus 4.7 Fast mode · 27 new env vars

Fifteen-release stretch (v2.1.129 → v2.1.143). /goal ships a real completion tracker for complex agentic work — declare a condition, Claude tracks progress and stops when it's met. Agent View (claude agents) becomes the user-facing surface for dispatched background sessions; CLAUDE_CODE_DISABLE_AGENTS_FLEET is removed and replaced by CLAUDE_CODE_DISABLE_AGENT_VIEW. Fast mode silently upgraded to Opus 4.7 at no extra cost (CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE pins back to 4.6). The agentic safety pair: CLAUDE_CODE_MAX_TURNS hard-caps conversation turns, CLAUDE_CODE_SUPERVISED tells Claude a human is watching. Hook args exec form lands — shell-less subprocess spawning for injection-safe hooks. MCP servers receive CLAUDE_PROJECT_DIR and survive /clear. Plugin marketplace shows context cost estimates before install. CLAUDE_CODE_STOP_HOOK_BLOCK_CAP ends the Ctrl+C hang. Three background session vars (CLAUDE_BG_AUTH_SNAPSHOT_PATH, CLAUDE_BG_STARTUP_WEDGE_MS, CLAUDE_BG_TCC_DISCLAIMED) document the BG failure modes. Rewind gains "Summarize up to here."

v2.1.128 — May 4, 2026#

CLAUDE_CODE_PROACTIVE · CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT · Sonnet 4.6 default

v2.1.127 was an internal build. v2.1.128 ships two env vars: CLAUDE_CODE_PROACTIVE turns idle detection into action — Claude prompts you on return from an idle session ("Start a new task?") instead of waiting silently. The idle-watching infrastructure (CLAUDE_CODE_IDLE_THRESHOLD_MINUTES, CLAUDE_CODE_IDLE_TOKEN_THRESHOLD) was already in v2.1.126; this is the first user-visible behavior built on top of it. CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT strips the system prompt down for background subagents — cuts the largest fixed per-turn token cost across parallel agent trees. Use it on leaf-node workers with a narrow job; leave it off on coordinators that need full tool and skill awareness. Sonnet 4.6 takes over as the embedded default for Claude Code Remote via undated alias, rolling forward automatically as Anthropic ships new checkpoints.

v2.1.126 — May 1, 2026#

CLAUDE_CODE_COLD_COMPACT · claude project purge · gateway model picker · 7 bug fixes

v2.1.124–v2.1.125 were internal builds. v2.1.126 ships claude project purge — surgical cleanup of transcripts, tasks, and config per project with --dry-run, --interactive, and --all flags. The /model picker now queries your gateway's /v1/models when ANTHROPIC_BASE_URL is set, showing what your environment actually serves instead of Anthropic's public catalog. --dangerously-skip-permissions now covers .claude/, .git/, .vscode/, and shell config writes — fully non-interactive CI at last. OAuth paste-in for WSL2, SSH, and containers. Security fix: managed sandbox allowManagedDomainsOnly restrictions were silently dropped when the org policy file had no sandbox block. Bug fixes: stream idle timeout after sleep, image paste breaking sessions (now auto-downscaled), Windows clipboard leaking to process args, CJK garbled in no-flicker mode, Ctrl+L erasing input, deferred tools missing in fork subagents on turn 1, Agent SDK hang on malformed tool names. Under the hood: CLAUDE_CODE_COLD_COMPACT enables the daemon to compact cold sessions on resume — solving the stale-context corruption that's affected sleep-resume and bridge-reattach scenarios since v2.1.62.

v2.1.123 — April 29, 2026#

290+ env vars (+30 net) · 17 models · 25 slash commands

Five-release stretch (v2.1.119 → v2.1.123). The background daemon — first flagged in v2.1.114 as a single CLAUDE_BG_BACKEND=daemon var — gets its complete IPC contract: rendezvous socket (CLAUDE_BG_RENDEZVOUS_SOCK), process isolation boundary (CLAUDE_BG_ISOLATION), job attribution (CLAUDE_BG_SOURCE), PTY recording (CLAUDE_PTY_RECORD), cold-start telemetry (CLAUDE_CODE_DAEMON_COLD_START), and a sequence-numbered bridge reconnection protocol (CLAUDE_BRIDGE_REATTACH_SEQ/SESSION). Sessions become first-class addressable objects via SESSION_ID/KIND/LOG/NAME — the plumbing behind v2.1.122's PR-URL→session /resume lookup. An agent fleet layer takes shape with selection (CLAUDE_AGENTS_SELECT), auto-relaunch tracking (CLAUDE_AGENTS_AUTO_RELAUNCHED_AT), and a kill switch (CLAUDE_CODE_DISABLE_AGENTS_FLEET). CoWork memory (CLAUDE_COWORK_MEMORY_GUIDELINES, CLAUDE_COWORK_MEMORY_INDEX_CONTENT) is the first inter-agent coordination primitive — a shared RAG layer with policy-governed writes. Daily wins: /config persists, alwaysLoad for MCP, PostToolUse hooks rewrite any tool output, claude ultrareview is a CLI subcommand, default effort raised to high for Pro/Max, three memory leaks fixed, Windows no longer requires Git Bash.

v2.1.118 — April 23, 2026#

260 env vars (+10 net) · 17 models · 25 slash commands (+1)

Four-release stretch (v2.1.115 → v2.1.118). The headline shift: forked subagents enabled on external builds via CLAUDE_CODE_FORK_SUBAGENT=1 — the Agent Teams multi-agent architecture leaves internal-only status. Subscription and rate-limit tier vars (CLAUDE_CODE_SUBSCRIPTION_TYPE, CLAUDE_CODE_RATE_LIMIT_TIER) describe a runtime moving toward per-plan capability differentiation. Named proxy vars (CLAUDE_CODE_HTTP_PROXY, CLAUDE_CODE_HTTPS_PROXY) simplify corporate proxy setup without the v2.1.114 auth helper protocol. Bridge reconnection handshake (CLAUDE_BRIDGE_REATTACH_SEQ, CLAUDE_BRIDGE_REATTACH_SESSION) and retry watchdog close the remaining "silent hang" surface alongside prior stall vars. CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT for token-constrained deployments. DISABLE_UPDATES kill-switch for externally managed environments. This window also coincides with Anthropic's April 23 postmortem on the March–April quality regression (reasoning effort default, caching bug, verbosity system prompt) and usage limit reset for all subscribers. v2.1.116 delivered 67% /resume speed improvement; v2.1.118 added vim visual mode, /usage (merged /cost+/stats), /theme JSON editing, hooks invoking MCP tools, and WSL→Windows settings inheritance.

v2.1.114 — April 20, 2026#

250 env vars (+20 net) · 17 models (+1) · 24 slash commands

Seven-release stretch (v2.1.108 → v2.1.114). Two headline shifts: the CLI now spawns a native per-platform binary instead of bundled JS (v2.1.113) — packaging, supply-chain, and signing posture all tighten. And /fewer-permission-prompts (v2.1.111; originally announced as /less-permission-prompts) scans transcripts to build a prioritized settings allowlist — the first feature that actually reduces Claude Code's biggest daily friction point. Around those: fullscreen TUI (/tui fullscreen) with DECSTBM scrolling, /ultrareview cloud PR review, Opus 4.7 xhigh effort + /effort slider, 1-hour prompt caching on API key/Bedrock/Vertex/Foundry, proxy auth helper subprocess protocol (5 vars) for Kerberos/SSO, Agent Teams scaffolding (team name propagation, subagent prompt append, stall timeout, daemon backend), experimental Advisor Tool, session recap for telemetry-disabled users, readline-compatible Ctrl+A/E, wrapped-URL click targets, Windows Ctrl+Backspace. ~80 bug fixes across the range.

v2.1.108 — April 14, 2026#

234 env vars (+4 net) · 4 new, 0 removed

Limits reckoning release. Prompt cache TTL now user-configurable (ENABLE_PROMPT_CACHING_1H, FORCE_PROMPT_CACHING_5M). Rate limits vs. plan limits get separate error messages. Startup warning if caching accidentally disabled. DISABLE_TELEMETRY → 5-minute TTL bug fixed. REPL experimentation returns: CLAUDE_CODE_REPL and CLAUDE_REPL_VARIANT for A/B testing stateful execution. Skill discovery via Skill tool (CLAUDE_API_SKILL_DESCRIPTION). Internal team routing (CLAUDE_INTERNAL_ASSISTANT_TEAM_NAME). /undo alias for /rewind. /model warns before mid-conversation switch. /recap away summary. Lazy-loaded language grammars reduce memory footprint. 15+ bug fixes.

v2.1.107 — April 14, 2026#

230 env vars (+6 net) · 6 new, 0 removed

Multi-worktree orchestration plumbing: CLAUDE_CODE_BASE_REFS for diff base overrides, CLAUDE_CODE_REPO_CHECKOUTS for worktree storage paths, CLAUDE_CODE_RESUME_FROM_SESSION for session chaining. Away mode summaries (CLAUDE_CODE_ENABLE_AWAY_SUMMARY) for async agent supervision. Byte-level stream watchdog (CLAUDE_ENABLE_BYTE_WATCHDOG) catches stalled API connections below SSE layer. UltraReview preflight test fixture. PreCompact hooks, background plugin monitors, worktree reuse via path parameter. Stalled stream abort-and-retry, SSH color fix. v2.1.106 skipped.

v2.1.104 — April 12, 2026#

224 env vars (+2 net) · 2 new, 0 removed

Enterprise TLS certificate store control (CLAUDE_CODE_CERT_STORE) — OS CA store now trusted by default, bundled reverts to Node.js CAs. SDK OAuth refresh coordination (CLAUDE_CODE_SDK_HAS_OAUTH_REFRESH) prevents double-refresh races in IDE extensions. /team-onboarding command generates teammate ramp-up guides. /ultraplan auto-provisioning. Command injection fix in LSP binary detection. Virtual scroller memory leak fix. Hardcoded 5-minute timeout removed — respects API_TIMEOUT_MS. v2.1.101 changelog covers this range; v2.1.102–103 skipped.

v2.1.100 — April 10, 2026#

222 env vars (0 net) · 3 new, 3 removed

Context token ceiling (CLAUDE_CODE_MAX_CONTEXT_TOKENS) with DISABLE_COMPACT interop. Perforce VCS mode (CLAUDE_CODE_PERFORCE_MODE) — fail on read-only files with p4 edit hint. Script capability caps (CLAUDE_CODE_SCRIPT_CAPS) paired with PID namespace sandboxing. Google Vertex AI setup wizard. Focus view toggle (Ctrl+O). Monitor tool for streaming background script events. 30+ Bash permission security fixes closing bypass vectors and reducing false prompts. MCP memory leak fix (~50 MB/hr). REPL experiment variables removed. v2.1.99 skipped.

v2.1.96 — April 8, 2026#

222 env vars (0 net) · 0 new, 0 removed

Bedrock auth hotfix. Fixed 403 "Authorization header is missing" regression from v2.1.94 when using AWS_BEARER_TOKEN_BEDROCK or CLAUDE_CODE_SKIP_BEDROCK_AUTH. Tracked surface identical to v2.1.94. v2.1.95 skipped.

v2.1.94 — April 7, 2026#

222 env vars (+5) · 5 new, 0 removed

Mantle auth proxy for Amazon Bedrock (CLAUDE_CODE_USE_MANTLE) with skip flag (CLAUDE_CODE_SKIP_MANTLE_AUTH). Sandbox state flag (CLAUDE_CODE_SANDBOXED). MCP server environment variable allowlist (CLAUDE_CODE_MCP_ALLOWLIST_ENV) — replaces all-or-nothing env inheritance. Team onboarding flows (CLAUDE_CODE_TEAM_ONBOARDING). --resume across worktrees. CJK text corruption fix. Write tool diff 60% faster. Default effort "high" for Team/Enterprise.

v2.1.92 — April 3, 2026#

217 env vars (+5) · 5 new, 0 removed

Proxy simulation without infrastructure (CLAUDE_CODE_SIMULATE_PROXY_USAGE). Fast mode org check bypass (CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK). Remote Control session name prefix for CI/CD observability. CLAUDE_CODE_EXECPATH for symlinked/wrapped binary overrides. CLAUDE_BASE shorthand for base URL config. Interactive Bedrock setup wizard. Per-model cost breakdowns for subscribers. Policy-enforced remote settings refresh.

v2.1.91 — April 2, 2026#

212 env vars (−1 net) · 1 new, 2 removed

MCP 500K character tool result persistence override (_meta["anthropic/maxResultSizeChars"]). Plugin bin/ executable support — plugins can now ship and invoke native binaries. disableSkillShellExecution setting to lock down shell access in skills. Multi-line deep links via claude-cli://open?q=. Edit tool token optimization with shorter old_string anchors. CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION replaces typo predecessor. Net −1 env var.

v2.1.90 — April 1, 2026#

213 env vars (+3 net) · 24 slash commands (+1)

New "Anthropic AWS" API provider (CLAUDE_CODE_USE_ANTHROPIC_AWS) — Anthropic's own AWS-hosted infrastructure, distinct from Bedrock. Companion CLAUDE_CODE_SKIP_ANTHROPIC_AWS_AUTH for IAM ambient credentials. Offline-resilient marketplace plugin caching (CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE). /powerup interactive lessons. Prompt-cache regression fix (since v2.1.69). Quadratic SSE transport and transcript write fixes. PowerShell security hardening. Auto mode now respects explicit user constraints.

v2.1.89 — March 31, 2026#

210 env vars (+2 net) · 4 new, 2 removed

Agent cost steering for subagent delegation (CLAUDE_CODE_AGENT_COST_STEER). Session resume thresholds by time and token count. Claude API skill disable flag. Deferred tool execution for headless sessions. StructuredOutput schema cache ~50% failure rate fixed. Autocompact thrash loop detection. Edit tool no longer requires prior Read. MCP non-blocking connections. Named subagents in typeahead. CLAUDE_CODE_DISABLE_COMMAND_INJECTION_CHECK removed — safety check now mandatory.

v2.1.88 — March 30, 2026#

208 env vars (+4 net) · 41 feature gates · 16 models · 23 slash commands · 29 dynamic configs

7 new env vars, 3 removed. Terminal UX overhaul: CLAUDE_CODE_NO_FLICKER for flicker-free alt-screen rendering, CLAUDE_CODE_DISABLE_MOUSE / CLAUDE_CODE_DISABLE_MOUSE_CLICKS for granular mouse control, CLAUDE_CODE_SCROLL_SPEED for tunable scrolling. REPL mode emergence: CLAUDE_CODE_REPL and CLAUDE_REPL_MODE for lightweight stateless interactions. CLAUDE_CODE_DISABLE_POLICY_SKILLS for suppressing org-pushed skills. v2.1.87 hotfix for Cowork Dispatch. Prompt cache miss fix, structured output schema bug, memory leaks for 1 GiB+ files. Thinking summaries off by default.

v2.1.87 — March 30, 2026#

202 env vars (0 net, +2/-2) · 16 models · 22 slash commands · 29 dynamic configs

OAuth console override (CLAUDE_LOCAL_OAUTH_CONSOLE_BASE) completes the auth triptych. Trusted device token (CLAUDE_TRUSTED_DEVICE_TOKEN) for persistent sessions. MCP server env vars (CLAUDE_CODE_MCP_SERVER_NAME, CLAUDE_CODE_MCP_SERVER_URL) removed after 2 versions. Cowork Dispatch message delivery fix. Dispatch model selection. Boris Cherny's 15 hidden features thread: mobile app, /loop, /schedule, hooks, --bare default flip coming.

v2.1.86 — March 27, 2026#

204 env vars (+2) · 41 feature gates · 16 models · 22 slash commands · 29 dynamic configs

Trusted device token (CLAUDE_TRUSTED_DEVICE_TOKEN) for enterprise SSO and CI/CD environments. OAuth console base URL override (CLAUDE_LOCAL_OAUTH_CONSOLE_BASE) completes the full OAuth endpoint decomposition across three releases. 20+ bug fixes: --resume tool_use ID errors, /feedback OOM on long sessions, masked input token leaks on narrow terminals, MCP server startup stalls. X-Claude-Code-Session-Id header for proxy observability.

v2.1.85 — March 27, 2026#

202 env vars (+5 net) · 41 feature gates · 16 models · 22 slash commands · 29 dynamic configs

MCP server multiplexing env vars (CLAUDE_CODE_MCP_SERVER_NAME, CLAUDE_CODE_MCP_SERVER_URL). OAuth endpoint overrides (CLAUDE_LOCAL_OAUTH_API_BASE, CLAUDE_LOCAL_OAUTH_APPS_BASE). PowerShell parse timeout. Commit logging. Executable shrank 7.15 MB (194 → 187 MB) by replacing WASM yoga-layout with pure TypeScript. Conditional if fields for hooks. Scheduled task timestamps. /compact context-exceeded fix. Memory leak patches for remote sessions. CLAUDE_CODE_SUBAGENT_MODELU typo removed.

v2.1.84 — March 26, 2026#

197 env vars (+6) · 41 feature gates · 16 models · 22 slash commands · 29 dynamic configs

Idle session management with configurable thresholds, CLAUDE_STREAM_IDLE_TIMEOUT_MS stream watchdog (default 90s). Bedrock/Vertex/Foundry capability override env vars. TaskCreated and WorktreeCreate HTTP hooks. Deep links open in preferred terminal. PowerShell tool preview for Windows. MCP 2KB description cap and server deduplication. Stats screenshot via Ctrl+S in /stats (16x faster). Drive root deletion protection on Windows.

v2.1.83 — March 25, 2026#

191 env vars (+8) · 41 feature gates · 16 models · 22 slash commands · 29 dynamic configs

managed-settings.d/ drop-in directories for enterprise policy management. 8 new env vars including CLAUDE_CODE_SUBPROCESS_ENV_SCRUB, CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK, CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST. Transcript search, plugin keychain storage, CwdChanged/FileChanged hook events. Largest release by fix count — 40+ fixes spanning macOS exit hang, startup latency, MCP policy bypass, sandbox piped commands. v2.1.82 skipped in changelog.

v2.1.81 — March 21, 2026#

183 env vars (+1) · 41 feature gates · 16 models · 22 slash commands · 29 dynamic configs (-1)

--bare flag for stripped-down scripted -p calls (skips hooks, LSP, plugins, auto-memory). Concurrent OAuth fix for multi-session re-auth race condition (#24317). --channels permission relay to phone. /btw pasted text fix, ! bash mode discoverability. Plan mode hides "clear context" by default. 17 bug fixes across voice mode, enterprise proxies, worktrees, Remote Control, tmux.

v2.1.80 — March 20, 2026#

182 env vars (no change) · 41 feature gates · 16 models · 22 slash commands · 30 dynamic configs (+1)

--channels research preview for messaging platform integration (Telegram, Discord), simplified /plugin install flow, source: 'settings' plugin declarations, 80MB startup memory reduction on large repos. Fine-grained tool streaming fix for Bedrock/Vertex/proxy 400 errors, --resume parallel tool result restoration, voice mode WebSocket fix, managed settings startup bug. 7 bug fixes total.

v2.1.79 — March 18, 2026#

182 env vars (+3, -2) · 41 feature gates · 16 models · 22 slash commands

Subprocess env scrubbing (CLAUDE_CODE_SUBPROCESS_ENV_SCRUB), extra API metadata (CLAUDE_CODE_EXTRA_METADATA), refactored init path (CLAUDE_CODE_NEW_INIT). 18MB startup memory reduction, --console auth for API billing, non-streaming fallback timeout. 2 vars removed.

v2.1.78 — March 17, 2026#

181 env vars (+1) · 41 feature gates · 16 models · 22 slash commands

Plugin persistent state (CLAUDE_PLUGIN_DATA), custom model picker entries (ANTHROPIC_CUSTOM_MODEL_OPTION). StopFailure hook event, plugin agent frontmatter expansion, line-by-line streaming. Security hardening: sandbox bypass fixes, MCP deny rule enforcement.

v2.1.77 — March 17, 2026#

180 env vars (+1) · 41 feature gates · 16 models · 22 slash commands (+1 via changelog)

Cowork memory extra guidelines, /branch (renamed from /fork), /copy N index. Stabilization release with 25+ bug fixes: terminal rendering, tmux, teammate sessions, memory management.

v2.1.76 — March 14, 2026#

179 env vars (+3) · 41 feature gates · 16 models · 22 slash commands (+1 via changelog)

Account tagged ID, fast-mode network error bypass, workspace host paths. MCP elicitation support, /effort command. 14 bug fixes.

v2.1.75 — March 13, 2026#

176 env vars (+2) · 41 feature gates · 16 models · 22 slash commands (+1 via changelog)

Auto-compact window threshold, plugin option prefix, /color command. 1M context windows for Opus 4.6.

v2.1.74 — March 12, 2026#

174 env vars (+2, -1) · 41 feature gates · 16 models · 22 slash commands

Frame timing log, session-end hooks timeout. Two-stage classifier flag removed (graduated to default).

v2.1.73 — March 11, 2026#

158 env vars · 40 feature gates · 16 models · 22 slash commands (first baseline)

Maintenance release. First tracking of the full slash command inventory.

v2.1.72 — March 10, 2026#

173 env vars (+6) · 41 feature gates · 16 models

Brief mode, cron control, debug log levels, two-stage classifier.

v2.1.70 — March 6, 2026#

167 env vars (+3, -1) · 41 feature gates · 16 models

CCR v2 bridge, effort level toggle, delta MCP instructions.

v2.1.69 — March 5, 2026#

165 env vars (+82) · 41 feature gates (+41) · 16 models (+16)

Plugin system, mTLS, agent teams, context compaction.

v2.1.19 — January 27, 2026#

83 env vars · ~16 feature gates

First tracked release. Agent swarms, Plan v2, debug flags.

What Each Post Covers#

Each release post documents:

• Environment variables — undocumented configuration knobs not in official docs
• Statsig feature gates — A/B test flags that control unreleased features
• Model IDs — all models the release supports
• Architecture changes — new subsystems, removed features, renamed internals

Results are diffed against the previous version to surface what actually changed.

Official Documentation#

For officially supported environment variables, see the Anthropic Claude Code Environment Variables Reference. Our tracker covers the undocumented variables missing from the official docs.

Related Guides#

• CLAUDE_CODE_EXPERIMENTAL_OBSERVER_AGENTS — the undocumented observer-agents flag: auto-spawned watcher agents, digests, ObserverReport
• Context Window Management — session strategies
• Productivity Tips — CLAUDE.md, permissions, custom commands
• Debugging Claude Code — env vars for troubleshooting
• Settings Schema with Autocomplete — IntelliSense for settings.json
• Autonomous Task Workflow — /task command deep dive

FAQ#

How do you extract this data?

We analyze publicly observable behavior of the installed Claude Code software to document configuration options and feature flags.

Are these officially supported?

No. These are undocumented internals. They may change or disappear without notice.

How often do you update?

We cover every significant release shortly after it becomes available.

This analysis is conducted for educational and research purposes under fair use principles. All trademarks and software referenced belong to their respective owners.