Environment variables, all 1,448.

CLAUDE_GATEWAY_ALLOW_LOOPBACK ◇ The gateway can authenticate callers with Anthropic-issued JWTs verified against a remote key endpoint.v2.1.160 → v2.1.308 ◆in the build CLAUDE_GATEWAY_DRAIN_TIMEOUT_MS ◇ claude gateway now shuts down gracefully on SIGTERM/SIGINT instead of exiting abruptlyv2.1.274 → v2.1.308 ◆in the build CLAUDE_GATEWAY_LOG_LEVEL ◇ Minimum level for the gateway's [gateway] log lines written to stderr: debug, info, warn or error, case-insensitive. Default info; an unset or unrecognised value uses info.v2.1.162 → v2.1.308 ◆in the build CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY ◇ A new env var lets a gateway proxy handle outbound address safety checks itself instead of Claude Codev2.1.277 → v2.1.308 ◆in the build CLAUDE_IMPORT_CONVERSATIONS ◇ Boolean (1, true, yes or on) that enables the hidden claude import-conversations <exportPath> command, which imports conversations from an export into sessions anchored to the directory given by --cwd (--dry-run only parses and verifies the manifest). Without it the command exits with import-conversations is not enabled.v2.1.152 → v2.1.308 ◆in the build CLAUDE_INTERNAL_ASSISTANT_TEAM_NAME ◇ Set by Claude Code when it relaunches itself as the latest version, to carry the current agent team name into the new process. The new process reads it once, deletes it from its environment, and uses it as the team name when no existing team name is given; without it the name defaults to session- plus the first 8 characters of an id.v2.1.160 → v2.1.308 ◆in the build CLAUDE_INTERNAL_FC_OVERRIDES ◇ The CLAUDE_INTERNAL_FC_OVERRIDES env var no longer overrides feature flags.v2.1.160 → v2.1.308 ◆in the build CLAUDE_INTERNAL_WARM_RESUME_QA ◇ No description from anybody.v2.1.160 → v2.1.211 ○gone CLAUDE_JOB_DIR ◇ About 220 environment variables now parse and validate in one place, including bounds checks.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_LOCAL_OAUTH_API_BASE ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_LOCAL_OAUTH_APPS_BASE ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_LOCAL_OAUTH_CONSOLE_BASE ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_MEMORY_STORES ◇ The memory store picker nudges you to pick a specific project instead of leaving the legacy default.v2.1.143 → v2.1.308 ◆in the build CLAUDE_MOCK_HEADERLESS_429 ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_PREVIEW_CLASSIFIER_FLOOR ◇ An env var for the Chrome navigation classifier is gone; a server setting now controls it.v2.1.196 → v2.1.241 ○gone CLAUDE_PROJECT_TOOL ◇ When true (1, true, yes or on), Claude Code loads the Projects tool (search hint read and write the session's attached claude.ai project), which enables itself only when the allow_projects_tool policy permits it and CLAUDE_PROJECT_UUID is set, and fetches a context block for that claude.ai project into the session's user context as attachedProject. The fetch is skipped when the policy disallows it, the API provider is not supported, or Claude Code runs in essential-traffic-only mode. Unset, neither the tool nor the project context is loaded.v2.1.174 → v2.1.185 ○gone CLAUDE_PROJECT_UUID ◇ About 220 environment variables now parse and validate in one place, including bounds checks.v2.1.174 → v2.1.308 ◆in the build CLAUDE_PTY_HEARTBEAT_MS ◇ Interval in milliseconds of the background session PTY host's client heartbeat (not on Windows). Each tick, every armed client's missed count goes up; at 3 the connection is closed, otherwise a ping is sent. Default 60000; an unset, non-numeric or 0 value uses the default.v2.1.166 → v2.1.308 ◆in the build CLAUDE_PTY_HOST_EXEC ◇ Claude Code adds seven environment variable names for background sessions and terminal hosts to a list it keeps of such namesv2.1.144 → v2.1.308 ◆in the build CLAUDE_PTY_HOST_NO_STABLE_PATH ◇ On macOS, setting CLAUDE_CODE_DISABLE_BG_STABLE_PATH makes background sessions relaunch from the running program instead of the stable pathv2.1.295 → v2.1.308 ◆in the build CLAUDE_PTY_ORPHAN_CHECK_MS ◇ Interval in milliseconds of the background session PTY host's orphan watchdog (not on Windows). A tick counts when the host's parent process id has changed and no client is connected; after 30 consecutive such ticks it logs orphan watchdog, sends SIGTERM to the child process and SIGKILL 5 seconds later if it has not exited. Default 2000; an unset, non-numeric or 0 value uses the default.v2.1.154 → v2.1.308 ◆in the build CLAUDE_PTY_RECORD ◇ File path for a recording of a background session's terminal output, read by the --bg-pty-host process. The file is opened for writing (truncated), starts with the terminal's columns and rows as two big-endian 32-bit integers, then gets one record per output chunk: microseconds since start and the chunk length (both big-endian 32-bit) followed by the raw bytes. Unset or a file that cannot be opened records nothing.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_RELAUNCH_SESSION_ADD_DIRS ◇ Extra directories you add during a session now survive when Claude Code relaunches itselfv2.1.281 → v2.1.308 ◆in the build CLAUDE_REMOTE_CONTROL_SESSION_NAME_PREFIX ◇ Prefix for auto-generated Remote Control session names, which take the form <prefix>-<suffix>. The value is lowercased, each run of characters outside a-z0-9 becomes - and leading or trailing dashes are removed; if nothing is left the prefix is remote-control, and when the variable is unset a built-in fallback is used (the help text says the hostname). Claude Code sets it from --remote-control-session-name-prefix, both in claude remote-control and in the main CLI.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_REMOTE_TOOLS_BRIDGE_URL ◇ The device bridge now only connects to approved hosts, with local overrides limited to localhost.never mined ·never mined CLAUDE_REMOTE_WORKFLOW_ARGS ◇ JSON arguments for the workflow script given in CLAUDE_REMOTE_WORKFLOW_SCRIPT, read by the same hidden server-launch command. An unset or empty value runs the script with no arguments; a value over the size limit fails with CLAUDE_REMOTE_WORKFLOW_ARGS exceeds ... bytes and one that is not valid JSON fails with CLAUDE_REMOTE_WORKFLOW_ARGS is not valid JSON.v2.1.200 → v2.1.308 ◆in the build CLAUDE_REMOTE_WORKFLOW_SCRIPT ◇ The workflow script run by the hidden server-launch command (registered as __remote-workflow, "Run the workflow script delivered in this session environment"). The command requires CLAUDE_CODE_REMOTE; if this variable is unset or empty it returns CLAUDE_REMOTE_WORKFLOW_SCRIPT is not set instead of running anything, and otherwise it passes a policy check and runs the script.v2.1.200 → v2.1.308 ◆in the build CLAUDE_REPL_VARIANT ◇ String. When it contains trim<N>k (for example trim50k), the REPL tool's maximum result size becomes N * 1000 characters, and the text of REPL results that carry images or documents is truncated to that length. Unset, or without that pattern, the limit is 100000.v2.1.138 or earlier → v2.1.274 ○gone CLAUDE_RUNNER_ACTIVITY_FD ◇ Claude Code now recognises the CLAUDE_CODE_REMOTE_MCP_CONFIG environment variable and passes it on to processes it startsv2.1.196 → v2.1.308 ◆in the build CLAUDE_RUNNER_CLAUDE_BIN ◇ The runner writes a code-sign helper script, using CLAUDE_RUNNER_CLAUDE_BIN if you set it.never mined ·never mined CLAUDE_RUNNER_CLIENT_PLATFORM ◇ Odd platform labels from the runner are rejected or replaced with "unknown" instead of exported.v2.1.301 → v2.1.308 ◆in the build CLAUDE_RUNNER_DISABLE_AWAITING_ACTION_OVERRIDE ◇ Boolean (1, true, yes or on) read by the self-hosted runner when a child session's turn ends. Normally, if the turn cannot be treated as a plain turn-end and the child last reported waiting_on_user: true, the runner reports the session as awaiting-action ("child parked at a prompt"); when this is true that step is skipped and the turn end goes through the turn-end-deferred path instead.v2.1.200 → v2.1.308 ◆in the build CLAUDE_RUNNER_FAIL_FAST_FETCH ◇ Two env vars change how runner git fetches handle retries and stalled delta resolution.never mined ·never mined CLAUDE_RUNNER_FETCH_DEPTH ◇ Full git checkout machinery for self-hosted runners now ships, but you cannot reach it interactively.v2.1.172 → v2.1.308 ◆in the build CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS ◇ CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS sets a time cap on server-side progress when a runner fetches a repositoryv2.1.295 → v2.1.308 ◆in the build CLAUDE_RUNNER_SESSION_UUID ◇ Runner hooks now get CLAUDE_RUNNER_SESSION_UUID and CLAUDE_CODE_REMOTE_SESSION_UUID so your scripts can key on the session.never mined ·never mined CLAUDE_RUNNER_SKIP_GIT_VERIFY ◇ Set to exactly 1 in the self-hosted runner's environment to skip the check that a checkout hook left a .git entry in the checkout directory. Without it, a hook that exits 0 but leaves no .git fails with an error that suggests this variable for non-git SCMs; the checks that the checkout path exists and is a directory still run.v2.1.224 → v2.1.308 ◆in the build CLAUDE_RUNNER_UPDATE_BIN ◇ Claude Code now treats CLAUDE_RUNNER_UPDATE_BIN and CLAUDE_RUNNER_ACTIVITY_FD like its other remote session variablesv2.1.307 → v2.1.308 ◆in the build CLAUDE_RUNNER_USE_GIT_PROXY ◇ Environment form of the self-hosted runner's --use-anthropic-git-proxy flag; a truthy value turns on Anthropic-managed git. The runner then registers as opted in, exits with an error unless --capacity is 1, and installs its git proxy credential setup (the help text warns this replaces the account's home-level git config).v2.1.224 → v2.1.308 ◆in the build CLAUDE_SECURESTORAGE_CONFIG_DIR ◇ When CLAUDE_CODE_AUTO_MODE_TIER is set, Claude Code includes it in the list of environment settings it builds for commands it preparesv2.1.144 → v2.1.308 ◆in the build CLAUDE_SERVE_DRAIN_TIMEOUT_MS ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_SESSION_INGRESS_TOKEN_FILE ◇ Remote settings can now supply many more environment variables, including auth tokens and home directory paths.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_SLOW_FIRST_BYTE_MS ◇ Milliseconds after an API request is sent before Claude Code logs a Slow first byte warning and sends the tengu_api_slow_first_byte event if no stream chunk has arrived yet. Default 30000; an unset, 0 or unparseable value uses the default. It only adds the warning and the event; the request keeps running.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_SNIP ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_SSH_DAEMON_CHILD ◇ Claude Code treats a session as server- or SSH-started when CLAUDE_CODE_SPAWNED_BY_SERVER, CLAUDE_SSH_DAEMON_CHILD or CLAUDE_SSH_RUN_DIR is setv2.1.299 → v2.1.308 ◆in the build CLAUDE_SSH_LOCAL_BINARY ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_SSH_RUN_DIR ◇ Claude Code treats a session as server- or SSH-started when CLAUDE_CODE_SPAWNED_BY_SERVER, CLAUDE_SSH_DAEMON_CHILD or CLAUDE_SSH_RUN_DIR is setv2.1.299 → v2.1.308 ◆in the build CLAUDE_SSH_VERSION ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_STAGE_FILE_ROOT ◇ In remote mode the staging folder is marked as coming from chat, and an oversized git shared index is reported as its own casev2.1.174 → v2.1.308 ◆in the build CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS ◇ Hung requests are aborted when headers never arrive, and you can tune the window.v2.1.242 → v2.1.308 ◆in the build CLAUDE_STREAM_IDLE_TIMEOUT_MS ◇ A new admin key, inferenceStreamIdleTimeoutSec, sets Claude Code's stream idle timeout (300 to 1800 seconds) for sessions the app startsv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_SUBAGENT_BG_SHELL_MAX_MS ◇ Milliseconds a backgrounded shell command started by a subagent may keep running before Claude Code kills it (SIGTERM to its process tree). Default 3600000 (one hour); an unset or non-numeric value uses the default, and the value is declared as an integer with minimum 1. Shells with no subagent id get no cap, and on the start-in-background path monitor shells are exempt.v2.1.138 or earlier → v2.1.259 ○gone CLAUDE_TMPDIR ◇ Fallback for the TMPDIR that the sandbox gives sandboxed commands: Claude Code uses CLAUDE_CODE_TMPDIR first, then CLAUDE_TMPDIR, then /tmp/claude. When sandboxing computes its own temp directory and this variable is unset, Claude Code sets it to that directory.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_TRUSTED_DEVICE_TOKEN ◇ Cloud sessions get told outright they cannot message peers that require a trusted device.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_WORKFLOW_NAME_ONLY ◇ When 1, true, yes or on, restricts workflows to named bundled ones: the Workflow tool rejects calls carrying script, scriptPath, resumeFromRunId or remote (invoke as {name, args} only), a nested workflow({scriptPath}) throws, and the workflow list is built from the bundled workflows without loading user or project workflow files.v2.1.200 → v2.1.308 ◆in the build CLAUDE_WORKSHOP_PROGRESS ◇ Streaming now feeds tool and partial-input events to the workshop progress display.never mined ·never mined CLAUDECODE Set to 1 in subprocesses Claude Code spawns (Bash and PowerShell tools, tmux sessions, hook commands, status line commands, stdio MCP server subprocesses). IDE extensions also set this in their integrated terminals. Use to detect when a script is running inside a subprocess spawned by Claude Code. To check whether the current process was spawned directly by a tool call or hook, rather than inside a stdio MCP server that Claude Code started, use CLAUDE_CODE_CHILD_SESSION insteadv2.1.160 → v2.1.308 ◆in the build CLI_WIDTH ◇ No description from anybody.v2.1.138 or earlier → v2.1.252 ○gone CLOUD_ML_REGION ◇ A malformed cloud region in your environment is now ignored in favour of the default.v2.1.138 or earlier → v2.1.308 ◆in the build CLOUD_RUN_JOB ◇ Read by the bundled Google Cloud auth libraries: any non-empty value counts as a serverless GCP environment in their GCP residency check, and their platform detection reports Cloud Run Jobs when it is set on a GCE-backed host that is not GKE or Cloud Run.v2.1.138 or earlier → v2.1.308 ◆in the build CLOUDSDK_ACTIVE_CONFIG_NAME ◇ Names the active gcloud configuration. It is part of the hash that names the lock and outcome files Claude Code uses when running the gcpAuthRefresh command. When building its GCP credential fingerprint, Claude Code reads configurations/config_<name> from the gcloud config directory using this name, else the contents of the active_config file, else default; no fingerprint is built when the active_config file does not exist, even if this variable is set.v2.1.285 → v2.1.308 ◆in the build CLOUDSDK_AUTH_ACCESS_TOKEN ◇ gh, aws, gcloud, bq and gsutil get placeholder credentials so they authenticate through the proxy instead of failing.v2.1.178 → v2.1.308 ◆in the build CLOUDSDK_CONFIG ◇ AWS and Google custom endpoint variables are now forwarded and scrubbed deliberately.v2.1.138 or earlier → v2.1.308 ◆in the build CLOUDSDK_CORE_CUSTOM_CA_CERTS_FILE ◇ No description from anybody.v2.1.251 → v2.1.308 ◆in the build CLOUDSDK_PROXY_USERNAME ◇ Concurrent sandboxes now get unique proxy usernames instead of sharing one identity.never mined ·never mined CODER ◇ When 1, true, yes or on, Claude Code's deployment environment detection reports coder (unless an earlier check such as CODESPACES or GITPOD_WORKSPACE_ID matched first). The detected environment is reported in telemetry as deployment_environment.v2.1.138 or earlier → v2.1.308 ◆in the build CODER_WORKSPACE_NAME ◇ Any non-empty value makes Claude Code's deployment environment detection report coder (unless an earlier check such as CODESPACES or GITPOD_WORKSPACE_ID matched first). The detected environment is reported in telemetry as deployment_environment.v2.1.138 or earlier → v2.1.308 ◆in the build CODESPACES ◇ When 1, true, yes or on, Claude Code's deployment environment detection reports codespaces; it is the first check, so it wins over every other environment. The detected environment is reported in telemetry as deployment_environment.v2.1.138 or earlier → v2.1.308 ◆in the build CODEX_HOME ◇ The user-level OpenAI Codex directory that the Codex importer (claude migrate) reads: it detects Codex from config.toml, AGENTS.md or prompts there and imports user-scope config from it. Defaults to ~/.codex when unset.v2.1.182 ○gone COLORFGBG ◇ Fallback for telling whether the terminal background is dark or light when no system theme has been cached: the last ;-separated field is read as a color index, where 0 to 6 and 8 mean dark and 7 or 9 to 15 mean light. A missing, empty or out-of-range value gives no answer, and one caller then uses dark.v2.1.138 or earlier → v2.1.308 ◆in the build COLORTERM ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build COMPUTER_USE_SWIFT_NODE_PATH ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build COMPUTERNAME ◇ Claude Code can now read the computer name, using macOS's own tool, the Windows COMPUTERNAME value, or the hostname elsewherev2.1.285 → v2.1.308 ◆in the build COMSPEC ◇ Inline admin policy scripts run with a stripped environment and a PowerShell resolved from fixed paths only.v2.1.138 or earlier → v2.1.308 ◆in the build CONTAINER_SANDBOX_MOUNT_POINT ◇ Session cleanup now tracks which machine a process belongs to, so container ids stop colliding.v2.1.246 → v2.1.308 ◆in the build COREPACK_ENABLE_AUTO_PIN ◇ Claude Code sets COREPACK_ENABLE_AUTO_PIN=0 in its own environment at startup, overwriting any existing value, so child processes inherit it and Corepack does not auto-pin a packageManager field.v2.1.138 or earlier → v2.1.308 ◆in the build CURL_CA_BUNDLE ◇ Claude Code both reads and sets it. It reads it, along with SSL_CERT_FILE, NODE_EXTRA_CA_CERTS and the other CA-bundle variables, to find a custom CA bundle: in remote sessions the agent proxy folds that file's certificates into the bundle it serves, and the Bash sandbox resolves the file when computing its exclusions. It sets it when the Bash sandbox is given a CA certificate file, pointing each sandboxed command's CURL_CA_BUNDLE, SSL_CERT_FILE, NODE_EXTRA_CA_CERTS, REQUESTS_CA_BUNDLE, GIT_SSL_CAINFO and similar variables at it, and the agent proxy points it at its own bundle for child processes.v2.1.251 → v2.1.308 ◆in the build CURSOR_TRACE_ID ◇ Marks the session as running in Cursor: when it is set, the IDE and terminal detection returns cursor (one check needs a non-empty value, the others only that the variable exists), diagnostics label the terminal Cursor, and the terminal is treated as sending high-rate mouse wheel events (wheelFlood). Its presence is also reported in telemetry as hasCursorTraceId.v2.1.138 or earlier → v2.1.308 ◆in the build DATABASE_URL ◇ No description from anybody.v2.1.160 → v2.1.221 ○gone DAYTONA_WS_ID ◇ Any non-empty value makes Claude Code's deployment environment detection report daytona (unless an earlier check such as CODESPACES, GITPOD_WORKSPACE_ID, Coder or DevPod matched first). The detected environment is reported in telemetry as deployment_environment.v2.1.138 or earlier → v2.1.308 ◆in the build DEBUG ◇ When 1, true, yes or on, turns on Claude Code's debug mode from launch, as --debug does. Any non-empty value also makes the self-hosted-runner setup and doctor commands print the argv of the session they spawn to stderr. Bundled third-party libraries read DEBUG with their own meaning as well.v2.1.138 or earlier → v2.1.308 ◆in the build DEBUG_AUTH ◇ Read by the bundled Google Cloud metadata client when it checks whether a metadata server is reachable. When set to any non-empty value, the error from a failed probe is printed with console.info before it is classified.v2.1.138 or earlier → v2.1.308 ◆in the build DEBUG_CLAUDE_AGENT_SDK ◇ SDK debug logging tags a session id and can write through the storage layer when given one.v2.1.138 or earlier → v2.1.308 ◆in the build DEBUG_SDK ◇ When 1, true, yes or on, turns on Claude Code's debug mode from launch, the same as DEBUG or --debug.v2.1.138 or earlier → v2.1.308 ◆in the build DEMO_VERSION ◇ When set to a non-empty string, the logo header shows this string as the version in place of the real version, shows the working directory as /code/claude, and the startup update summary is not computed.v2.1.138 or earlier → v2.1.308 ◆in the build DENO_CERT ◇ SSL certificate error message now lists the specific env var to check per toolv2.1.251 → v2.1.308 ◆in the build DENO_DEPLOYMENT_ID ◇ Claude Code's deployment-environment detection, which is sent as deployment_environment in telemetry event metadata, reports deno-deploy when this variable has any non-empty value. Platforms earlier in the detection order (such as Vercel, Netlify or Cloudflare Pages) take precedence.v2.1.138 or earlier → v2.1.308 ◆in the build DETECT_GCP_RETRIES ◇ Number of retries the bundled GCP metadata client uses when it probes the metadata server's instance endpoint to detect whether it is running on Google Cloud. Default 0; a set value is converted with Number().v2.1.138 or earlier → v2.1.308 ◆in the build DEVPOD ◇ Claude Code's deployment-environment detection, which is sent as deployment_environment in telemetry event metadata, reports devpod when DEVPOD is 1, true, yes or on (case-insensitive) or DEVPOD_WORKSPACE_UID is non-empty. Platforms earlier in the detection order (Codespaces, Gitpod, Coder) take precedence.v2.1.138 or earlier → v2.1.308 ◆in the build DEVPOD_WORKSPACE_UID ◇ Claude Code's deployment-environment detection, which is sent as deployment_environment in telemetry event metadata, reports devpod when this variable has any non-empty value or DEVPOD is 1, true, yes or on. Platforms earlier in the detection order (Codespaces, Gitpod, Coder) take precedence.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_AUTO_COMPACT ◇ Unknown model names now get an assumed context budget so auto-compact works instead of stalling.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_AUTOUPDATER ◇ DISABLE_AUTOUPDATER=0 or false no longer disables updates; only truthy values count now.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_BRIEF_MODE_STOP_HOOK ◇ When true, turns off brief mode's end-of-turn enforcement. Without it, in the main REPL thread or an SDK session with brief mode on and the brief tool available, Claude Code adds a hidden meta message carrying the brief enforcement text when the assistant made no brief tool call in the current turn and no such reminder was already added.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_BUG_COMMAND ◇ Temporary probes that checked whether old code paths still ran have been deleted; behaviour is unchanged.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_COMPACT ◇ Unknown model names now get an assumed context budget so auto-compact works instead of stalling.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_COST_WARNINGS ◇ When true, Claude Code does not print the session cost to stdout when the process exits and does not show the cost-threshold dialog. Without it, both depend on the account type and the organization or workspace role.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_DOCTOR_COMMAND ◇ When true, the /doctor command (alias /checkup) is disabled.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_ERROR_REPORTING ◇ Any non-empty value stops Claude Code from recording errors in its in-memory recent-error log and turns off the error-reporting path fed by logError and the uncaught-exception and unhandled-rejection handlers. Claude Code also copies the variable into the environment it builds for byoc sessions.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_EXTRA_USAGE_COMMAND ◇ Setting DISABLE_EXTRA_USAGE_COMMAND now also hides the buy-credits prompts, which point at /model instead.v2.1.138 or earlier → v2.1.308 ◆in the build DISABLE_FEEDBACK_COMMAND ◇ Feedback and bug-report commands are hidden when Claude Code runs through the SDK.v2.1.138 or earlier → v2.1.308 ◆in the build