Environment variables, all 1,448.

CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS ◇ The release notes bundled inside Claude Code now open with a 2.1.292 section, listing an Agent tool effort setting and morev2.1.292 → v2.1.308 ◆in the build CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS ◇ No description from anybody.v2.1.296 → v2.1.308 ◆in the build CLAUDE_CODE_OVERRIDE_DATE ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_CODE_OWNERSHIP_FRAME ◇ Turns on the ownership framing of the system prompt; when unset or false, the tengu_walnut_prism flag (default false) decides. It only changes the lean system prompt that Claude Code uses for some models or when CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT is on: the intro says Claude works alongside the user and owns the outcome of what it takes on, and the confirm-first guidance drops the clause approval in one context doesn't extend to the next. The default system prompt is unchanged.v2.1.161 → v2.1.215 ○gone CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE ◇ When true and Claude Code was installed with Homebrew or WinGet, the package manager updater runs the upgrade itself (brew upgrade --cask or winget upgrade --id Anthropic.ClaudeCode, 5 minute timeout) when a newer version is available. Unset or false, it does not run the upgrade.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PARCHMENT_FERN ◇ An alternate prompt wording for two tools exists but stays off unless you set an environment variable.v2.1.223 → v2.1.284 ○gone CLAUDE_CODE_PARKED_PERMISSION_WAIT_MS ◇ Resuming can pick up an interrupted tool call, but only with two env vars set.v2.1.198 → v2.1.308 ◆in the build CLAUDE_CODE_PARKED_RUN_BEFORE_CLEAR ◇ In headless mode an approved parked tool call now runs only after its saved record is confirmed deleted; an env var skips the waitv2.1.282 → v2.1.308 ◆in the build CLAUDE_CODE_PARKED_STOP_RETIRES ◇ CLAUDE_CODE_PARKED_STOP_RETIRES lets you decide the fate of a set-aside permission prompt you never answered.v2.1.239 → v2.1.308 ◆in the build CLAUDE_CODE_PARSED_WILLOW ◇ New tengu_parsed_willow gate (defaulting on) added with CLAUDE_CODE_PARSED_WILLOW env overridev2.1.277 → v2.1.308 ◆in the build CLAUDE_CODE_PERFETTO_TRACE ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PERFETTO_WRITE_INTERVAL_S ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_CODE_PERFORCE_MODE ◇ Environment variables from your settings are now added before Claude Code filters the environment it passes to programs it startsv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PERSONAL_CONFIG_FIX ◇ Setting CLAUDE_CODE_SKIP_PLUGIN_LOCAL_MCP_SERVERS keeps only remote MCP servers from plugins, and several other environment variables are newv2.1.304 → v2.1.308 ◆in the build CLAUDE_CODE_PEWTER_OWL ◇ Tri-state override for both the pewter_owl_tool and pewter_owl_brief gates: a true value (1, true, yes, on) turns both on and a false value (0, false, no, off) turns both off, skipping the other checks (the tengu_pewter_owl_tool / tengu_pewter_owl_brief flags, a config key and a model check). Either gate on enables the SendUserMessage tool (alias Brief); CLAUDE_CODE_PEWTER_OWL_TOOL, when set, takes precedence for the tool gate.v2.1.159 → v2.1.308 ◆in the build CLAUDE_CODE_PEWTER_OWL_TOOL ◇ With CLAUDE_CODE_PEWTER_OWL_TOOL set, three tools can now leave out a response ID depending on a remotely controlled switchv2.1.169 → v2.1.308 ◆in the build CLAUDE_CODE_PLAN_ARTIFACTS ◇ Plan artifacts now have their own switch instead of riding along with an unrelated experiment.v2.1.223 → v2.1.237 ○gone CLAUDE_CODE_PLAN_MODE_INTERVIEW_PHASE ◇ No description from anybody.v2.1.138 or earlier → v2.1.211 ○gone CLAUDE_CODE_PLAN_MODE_REQUIRED ◇ Boolean fallback for whether an agent-team teammate needs plan approval, read only when the session has neither a teammate identity nor a dynamic team context (either of those supplies planModeRequired instead). When required, ExitPlanMode fails without a plan file and sends a plan_approval_request to the team lead, and the inbox poller ignores plan approval responses from anyone but the team lead.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLAN_V2_AGENT_COUNT ◇ Maximum number of Plan agents the plan-mode prompt tells the model it may launch in parallel during its design phase (You can launch up to N agent(s) in parallel.). An integer from 1 to 10 is used as is; otherwise the default is 3 on Max 20x, Enterprise and Team plans and 1 elsewhere.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLAN_V2_EXPLORE_AGENT_COUNT ◇ Maximum number of explore agents the plan-mode prompt tells the model it may launch in parallel during its initial understanding phase. An integer from 1 to 10 is used as is; anything else uses the default 3.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_ATTRIBUTION ◇ You can declare where a plugin came from via a JSON file or an environment variable.v2.1.246 → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_BINARY_ASSETS ◇ Boolean that turns on fetching of plugin binary assets for plugins from Anthropic's official marketplaces (such as claude-plugins-official), the same as the tengu_plugin_binary_assets flag; either one enables it. Off by default.v2.1.196 → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_CACHE_DIR ◇ A cleanup routine now removes set-aside and unreadable copies of installed_plugins in the plugins folder instead of zip files in the shares folderv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_DIR_WATCH ◇ CLAUDE_CODE_PLUGIN_DIR_WATCH now fully overrides whether plugin directory watching is enabled, not just disables itv2.1.259 → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_DIRS ◇ New CLAUDE_CODE_PLUGIN_DIRS environment variable lets you register extra plugin directories without --plugin-dir flagsv2.1.280 → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_GIT_TIMEOUT_MS ◇ Timeout in milliseconds for the git clone of a plugin or marketplace repository and for refreshing a marketplace cache. Default 120000; an unset, non-positive or non-numeric value uses the default. The clone timeout error names this variable.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE ◇ Marketplace refresh checks for the manifest through a storage handle instead of the raw filesystem.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_PREFER_HTTPS ◇ Boolean. When true, GitHub plugin and marketplace sources are cloned over https://github.com/<repo>.git instead of the SSH git@...:<repo>.git URL. CLAUDE_CODE_REMOTE has the same effect.v2.1.141 → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_SEED_DIR ◇ The self-hosted runner can now snapshot an operator's Claude config and seed it into every session, with a drain-marker mechanismv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PLUGIN_USE_ZIP_CACHE ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_POLISHED_DEWDROP ◇ A new hidden switch, settable via env var, can make certain responses silently drop or error instead.v2.1.263 → v2.1.308 ◆in the build CLAUDE_CODE_POLL_EVENT_DECLARATIONS ◇ Claude Code now lists a CLAUDE_CODE_POLL_EVENT_DECLARATIONS environment variable, though what it does is not knownv2.1.287 → v2.1.308 ◆in the build CLAUDE_CODE_POLL_EVENTS ◇ You can set CLAUDE_CODE_POLL_EVENTS in your shell, but settings files can no longer supply it.v2.1.237 → v2.1.308 ◆in the build CLAUDE_CODE_POST_FOR_SESSION_INGRESS_V2 ◇ No description from anybody.v2.1.138 or earlier → v2.1.220 ○gone CLAUDE_CODE_POST_TURN_MEMORY ◇ Three new env vars configure an unreleased 'post-turn memory' featurev2.1.269 → v2.1.308 ◆in the build CLAUDE_CODE_POST_TURN_MEMORY_CONFIG ◇ Three new env vars configure an unreleased 'post-turn memory' featurev2.1.269 → v2.1.308 ◆in the build CLAUDE_CODE_POST_TURN_MEMORY_SYNC ◇ Three new env vars configure an unreleased 'post-turn memory' featurev2.1.269 → v2.1.308 ◆in the build CLAUDE_CODE_POWERSHELL_RESPECT_EXECUTION_POLICY ◇ Boolean. By default Claude Code runs PowerShell tool commands and shell: 'powershell' hooks with -NoProfile -NonInteractive -ExecutionPolicy Bypass; when true, -ExecutionPolicy Bypass is left out for those so the system's execution policy applies. A few internal PowerShell helper calls always pass Bypass.v2.1.143 → v2.1.308 ◆in the build CLAUDE_CODE_POWERUP_ONBOARDING ◇ Overrides the tengu_birch_lantern flag (default off) when set to banner or step; other values are ignored. Either value shows the onboarding flow even when onboarding was already completed, and banner also activates the powerup-discovery notice New here? Type /powerup for a 5-minute tour... when its other condition holds.v2.1.181 → v2.1.308 ◆in the build CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS ◇ Fork snapshot files get a hard size cap and a background-wait timer becomes configurablev2.1.182 → v2.1.308 ◆in the build CLAUDE_CODE_PRINT_ENGINE_LOOP ◇ Set one variable and print mode keeps the model, tools and caches warm between prompts instead of rebuilding.v2.1.238 → v2.1.252 ○gone CLAUDE_CODE_PROACTIVE ◇ The main REPL and question dialog dropped their compiler memo caches, possibly re-rendering more.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PROFILE_QUERY ◇ Query profiling only runs when you set CLAUDE_CODE_PROFILE_QUERY, so ordinary sessions skip it.v2.1.160 → v2.1.247 ○gone CLAUDE_CODE_PROFILE_STARTUP ◇ Startup profiling now names specific phases including MCP connection and remote config setup.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PROJECT_DIR_NAME ◇ You can name the per-project history folder yourself, as long as you also set a custom config directory.v2.1.234 → v2.1.308 ◆in the build CLAUDE_CODE_PROJECTS_SESSION ◇ In some hosted sessions, WebFetch now returns an "Already fetched" note instead of fetching the same page with the same prompt againv2.1.282 → v2.1.308 ◆in the build CLAUDE_CODE_PROMPT_CACHE_TTL ◇ You can pin the prompt cache to five minutes or an hour, separately for subagents.v2.1.242 → v2.1.308 ◆in the build CLAUDE_CODE_PROPAGATE_TRACEPARENT ◇ Boolean. When true, Claude Code adds a W3C traceparent to API requests, to non-GET requests of two HTTP fetch wrappers, and as TRACEPARENT in the environment it builds for child processes, whenever a trace context exists. Without it, these are added only when the provider and base URL checks treat the API as first party.v2.1.152 → v2.1.308 ◆in the build CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST ◇ Setting CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST makes the parent process dictate your model settings, ignoring your settings file.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PROXY_AUTH_HELPER_TTL_MS ◇ How long in milliseconds a cached result of the proxy auth helper (enabled by CLAUDE_CODE_ENABLE_PROXY_AUTH_HELPER) is reused before the helper runs again. Default 300000; 0 or any other non-negative number is used as is, and a negative or non-numeric value uses the default.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PROXY_RESOLVES_HOSTS ◇ Plugin network calls are https-only and blocked from internal and metadata addresses.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_PWSH_PARSE_TIMEOUT_MS ◇ Timeout in milliseconds for each attempt of the pwsh process Claude Code runs to parse a PowerShell command (up to 2 attempts). Default 5000; an unset, non-positive or non-numeric value uses the default.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_QUESTION_EXTENDED ◇ An extended question mode is gated behind an env var and off by default.v2.1.260 → v2.1.308 ◆in the build CLAUDE_CODE_QUESTION_OPTIONAL_DESCRIPTIONS ◇ New CLAUDE_CODE_QUESTION_OPTIONAL_DESCRIPTIONS env var makes AskUserQuestion option descriptions optional, but only for certain entrypointsv2.1.268 → v2.1.308 ◆in the build CLAUDE_CODE_QUESTION_PREVIEW_FORMAT ◇ Sets the AskUserQuestion preview format when the value is markdown or html. Otherwise markdown is chosen for clients other than SDK, claude-desktop, local-agent, remote and some VS Code sessions. The SDK sets this variable for the CLI process from its askUserQuestion.previewFormat option.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_RATE_LIMIT_TIER ◇ Rate limit tier reported as rateLimitTier on the OAuth credentials Claude Code builds from an environment-supplied token such as CLAUDE_CODE_OAUTH_TOKEN; empty or unset gives null. Claude Code also sets it itself when a background session consumes an auth snapshot that carries a rateLimitTier.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REFUSAL_FALLBACK_CATCH_ALL ◇ Groundwork for spotting when the model refuses, with a banner slot that is still empty.v2.1.206 → v2.1.308 ◆in the build CLAUDE_CODE_RELAUNCH_HOME_TRUST ◇ If you already accepted the home-directory trust prompt, Claude Code no longer asks again when it relaunches itself during a resumev2.1.284 → v2.1.308 ◆in the build CLAUDE_CODE_RELAUNCH_PROACTIVITY_BASELINE ◇ Claude Code adds CLAUDE_CODE_DISABLE_PROACTIVITY and variables that pass proactivity state to a relaunched Claude Codev2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_RELAUNCH_PROACTIVITY_DECIDED ◇ Claude Code adds CLAUDE_CODE_DISABLE_PROACTIVITY and variables that pass proactivity state to a relaunched Claude Codev2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_RELAUNCH_PROACTIVITY_EVER_ON ◇ Claude Code adds CLAUDE_CODE_DISABLE_PROACTIVITY and variables that pass proactivity state to a relaunched Claude Codev2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_RELAUNCH_PROACTIVITY_LEVEL ◇ Claude Code adds CLAUDE_CODE_DISABLE_PROACTIVITY and variables that pass proactivity state to a relaunched Claude Codev2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_RELAUNCH_TERMINAL_SIZE ◇ Terminal size passed to a relaunched Claude Code process, as <columns>x<rows>. Claude Code sets it from the current stdout size when it re-executes or respawns itself; at startup the new process deletes it from its environment and, when stdout is a TTY and both numbers are 1 to 4 digits with no leading zero, uses them for process.stdout.columns and rows where those are unset or zero.v2.1.203 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE ◇ Remote sessions started from review can skip writing some state to disk.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_ENVIRONMENT_TYPE ◇ The skill-proposal tool now requires a remote environment and a specific entrypoint or env var, and can be vetoed per session.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_HERMETIC_MODE ◇ More environment variables get carried into remote sessions; none are set by default.v2.1.163 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_MCP_CONFIG ◇ Claude Code now recognises the CLAUDE_CODE_REMOTE_MCP_CONFIG environment variable and passes it on to processes it startsv2.1.299 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_MEMORY_DIR ◇ Claude Code now refuses paths in its temp folders or in folders background sessions write to freely, including for cloud credential settingsv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_RAW_EVENTS_FILE ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_SDK_URL ◇ When --sdk-url is not given in a remote session, Claude Code now reads the address from the CLAUDE_CODE_REMOTE_SDK_URL environment variablev2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_SEND_KEEPALIVES ◇ A new keepalive mechanism can keep remote hook connections open during slow hooks, off by defaultv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_SESSION_ID ◇ Commands blocked in cloud sessions now tell you the exact local command to run instead.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_SESSION_ORIGIN ◇ Remote sessions started from review can skip writing some state to disk.v2.1.217 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_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_CODE_REMOTE_SETTINGS_PATH ◇ Cached remote settings now come with a companion attestation file, and both must load cleanly.v2.1.160 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_SETTINGS_POLL_MS ◇ No description from anybody.v2.1.160 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_TOOLS_FORWARD ◇ With CLAUDE_CODE_REMOTE_TOOLS_FORWARD set to true, tools gain an optional field that runs the call on an attached machinev2.1.283 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_TOOLS_HOST_ALLOWS_UNATTENDED ◇ Claude Code's lists of allowed usage events and host environment variables gain entries about MCP, skills and hosted setupsv2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_TOOLS_PIN_STORED_LOGIN ◇ With CLAUDE_CODE_REMOTE_TOOLS_PIN_STORED_LOGIN set, the serve helper exits with code 78 if its API or sign-in addresses are not Anthropic's ownv2.1.283 → v2.1.308 ◆in the build CLAUDE_CODE_REMOTE_TOOLS_POLICY ◇ CLAUDE_CODE_REMOTE_TOOLS_POLICY picks between parity (the default) and interim remote-tool behaviour.never mined ·never mined CLAUDE_CODE_REMOTE_TOOLS_SESSION_CHANNEL ◇ Tools on another machine gain a third route over the session channel, gated off by default.never mined ·never mined CLAUDE_CODE_REMOVE_PROMPT_STRINGS ◇ Setting CLAUDE_CODE_REMOVE_PROMPT_STRINGS to a JSON list removes those strings from the prompts Claude Code sends to the modelv2.1.290 → v2.1.308 ◆in the build CLAUDE_CODE_REPL ◇ REPL contexts moved to a registry that releases timers even on error paths.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REPO_CHECKOUTS ◇ With CLAUDE_CODE_REPO_CHECKOUTS unset, Claude Code no longer assumes a checkout when it cannot work out the current folderv2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_REPORT_FINDINGS ◇ Boolean (1, true, yes or on). When true and the session has the ReportFindings tool, /code-review runs inline in the current conversation with a prompt built for that tool instead of running as a forked agent. It has no effect with --output-format text or json, or when the review is preloaded as a skill.v2.1.218 → v2.1.308 ◆in the build CLAUDE_CODE_RESTRICT_PERSONAL_CONFIG ◇ Setting CLAUDE_CODE_RESTRICT_PERSONAL_CONFIG stops your own skills, agents and hooks from widening permissions, and makes key hooks requiredv2.1.295 → v2.1.308 ◆in the build CLAUDE_CODE_RESTRICTED ◇ New --restricted flag starts a locked-down session with no shell tools and no settings files.v2.1.248 → v2.1.308 ◆in the build CLAUDE_CODE_RESULT_NONCE ◇ Claude Code now reads CLAUDE_CODE_RESULT_NONCE once at startup, keeps it, and removes it from the environmentv2.1.285 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_FROM_SESSION ◇ Child worker processes no longer inherit your session id, resume prompt or auth handles.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_INTERRUPTED_TURN ◇ Resuming a turn interrupted by shutdown now correctly sees the tool call as still pending instead of faked-answered.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS ◇ The staleness cutoff for auto-resuming an interrupted turn is now configurable, defaulting to 6 hoursv2.1.211 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_PROMPT ◇ Child worker processes no longer inherit your session id, resume prompt or auth handles.v2.1.140 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_REASON ◇ New CLAUDE_CODE_RESUME_REASON env var labels why an interrupted turn was resumedv2.1.268 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_SESSIONSTART_HOOKS_WAIT_MS ◇ Claude Code now reads three new environment variables: agent progress summaries, a sleep-compaction delay and a resume hook waitv2.1.296 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_SOURCE_ALIVE ◇ Set by Claude Code in the environment of a background session it launches, as <sessionId>|<boundary>[|<parentSessionId>] or a bare boundary value (1 when there is none), apparently to mark that the session it was forked from is still alive. On resume the child parses it and, when it is a bare value or names the resumed session or its parent, changes how the transcript is loaded; when the boundary is an ISO timestamp and a session ID is present, the job record gains forkSourceAlive, forkBoundaryAt and forkSessionId.v2.1.212 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_THRESHOLD_MINUTES ◇ Minimum age in minutes of the last user or assistant message older than one minute before Claude Code offers the resume-return prompt with the choices compact, continue, dismiss and never. Default 70 when unset or not a number; 0 is taken literally and removes the age requirement. The prompt also needs tengu_gleaming_fair or tengu_gleaming_fair_reuse on, the user not having picked never, and the conversation at or above CLAUDE_CODE_RESUME_TOKEN_THRESHOLD.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_TOKEN_THRESHOLD ◇ Minimum estimated token count of the conversation before Claude Code offers the resume-return prompt with the choices compact, continue, dismiss and never. Default 100000 when unset or not a number; 0 is taken literally and removes the token requirement. The prompt also needs tengu_gleaming_fair or tengu_gleaming_fair_reuse on and the last message older than CLAUDE_CODE_RESUME_THRESHOLD_MINUTES.v2.1.138 or earlier → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_TOLERATES_CONTEXT_APPENDS ◇ Resuming can pick up an interrupted tool call, but only with two env vars set.v2.1.242 → v2.1.308 ◆in the build CLAUDE_CODE_RESUME_TOLERATES_CONTEXT_SEEDS ◇ Setting CLAUDE_CODE_RESUME_TOLERATES_CONTEXT_SEEDS stops resume treating some final messages as an interrupted turnv2.1.290 → v2.1.308 ◆in the build