Feature flags, all 2,095.

tengu_juniper_bassoon ◇ Default false. When on, the usage-credits menu (the one offering Buy more and Adjust monthly limit) shows the soonest-expiring promotional credit grant on the balance, with its amount, optional name and expiry date. Under ten days before expiry the line reads <amount> of your balance expires in N days (or today) and is marked urgent.v2.1.216 → v2.1.309 ◆in the build never read tengu_juniper_sundial ◇ Default null, meaning 10. While ultra effort stays on, the number of user messages since the last ultra effort reminder after which Claude Code attaches a sparse ultra_effort_enter reminder again. A number is floored and raised to at least 1; a non-number uses the default, and CLAUDE_CODE_JUNIPER_SUNDIAL or one other override source take precedence.v2.1.218 → v2.1.309 ◆in the build AND v2.1.309 tengu_juniper_tide_wall ◇ Five analytics events cover the weekly limit-reset flow.never mined ·never mined never read tengu_juniper_vale ◇ Default {enabled: false, maxChars: 500, autoDismissAfterMs: 30000}. Controls the follow-up prompt shown after the user answers the in-session feedback survey with anything other than dismissing it: enabled: true turns it on, maxChars (a positive number, default 500) sets its character limit, and autoDismissAfterMs (at least 600, default 30000) sets when it is dismissed automatically. It is not shown while a response is loading, an overlay is open, or the prompt input has text, and another check can force it on with the default limits.v2.1.213 → v2.1.309 ◆in the build AND v2.1.309 tengu_kairos_brief ◇ Default false. With CLAUDE_CODE_BRIEF, one of the switches that turn on brief mode, subject to a further gate; in brief mode plain assistant text is hidden from the user and replies reach them through the SendUserMessage tool. The flag is also read directly when rendering user prompt messages and the spinner.v2.1.138 or earlier → v2.1.309 ◆in the build never read tengu_kairos_brief_config ◇ /brief is hidden by server config, but brief mode still works via an environment variable.v2.1.138 or earlier → v2.1.309 ◆in the build never read tengu_kairos_brief_stop_hook_text ◇ No description from anybody.v2.1.138 or earlier → v2.1.309 ◆in the build never read tengu_kairos_dream ◇ Default false. Enables the bundled dream skill (/dream consolidates memories; /dream nightly schedules a recurring /dream consolidate task at a random minute between 0:00 and 5:59 local time). The skill is enabled when this flag is on, auto memory is enabled and KAIROS mode is not active.never mined ·never mined never read tengu_keen_dolphin ◇ No description from anybody.v2.1.292 → v2.1.301 ○gone never read tengu_keybinding_fallback_used ◇ An analytics event, not a feature flag. Claude Code logs it when it looks up the shortcut for a keybinding action and finds none, so it uses the caller's fallback key text instead; the fields are action, context, fallback and reason (action_not_found, or no_context when no keybinding context is available).never mined ·never mined never read tengu_known_marketplaces_fallback_write ◇ Two sessions editing plugin marketplaces at once no longer clobber each other's entries.never mined ·never mined never read tengu_ladder_mq7 ◇ Default false. When on, print-mode (-p) startup connects the --mcp-config servers (except alwaysLoad ones) and the claude.ai connectors fully async without waiting for them. MCP_CONNECTION_NONBLOCKING set to a false value (0, false, no, off) turns this off regardless of the flag, and set to a true value turns it on.v2.1.138 or earlier → v2.1.141 ○gone never read tengu_lantern ◇ Reading someone else's shared artifact can now be auto-allowed via a remembered consent choicenever mined ·never mined never read tengu_lantern_wick ◇ Default false. When on, and the usage limit's grace window has just become active, the next main-thread follow-up request in a turn (not the first request of a user turn, and not one resuming incomplete thinking) gets a hidden message saying the usage limit is reached and the grace window is active, ending Wrap up: finish or checkpoint; don't start subagents or long work. A new user turn clears the pending notice without adding it.v2.1.217 → v2.1.226 ○gone AND v2.1.309 tengu_lantern_wick_braid ◇ A new 'wrap up your work' reminder can be injected near usage limits, currently gated offv2.1.277 → v2.1.309 ◆in the build never read tengu_lantern_wick_braid_release ◇ A block of text Claude Code inserts for the model can now be configured remotely per model, with separate text for subagentsv2.1.284 → v2.1.309 ◆in the build never read tengu_lantern_wick_braid_text ◇ A block of text Claude Code inserts for the model can now be configured remotely per model, with separate text for subagentsv2.1.284 → v2.1.309 ◆in the build never read tengu_lantern_wick_release_by_model ◇ A block of text Claude Code inserts for the model can now be configured remotely per model, with separate text for subagentsv2.1.284 → v2.1.309 ◆in the build never read tengu_lantern_wick_resume_by_model ◇ A block of text Claude Code inserts for the model can now be configured remotely per model, with separate text for subagentsv2.1.285 → v2.1.291 ○gone never read tengu_lantern_wick_text_by_model ◇ A block of text Claude Code inserts for the model can now be configured remotely per model, with separate text for subagentsv2.1.285 → v2.1.309 ◆in the build never read tengu_lapis_anchor_from_client_data ◇ Token-reminder settings can now arrive from the server, but your env var and setting still win.v2.1.246 ○gone never read tengu_lapis_anchor_threshold ◇ Claude Code now tracks token-budget usage per conversation and logs when it crosses key thresholdsnever mined ·never mined never read tengu_laptop_memory_context_wait ◇ Memory context fetching gains a new 'laptop' path alongside the existing cowork path, both currently inactivenever mined ·never mined never read tengu_larch_cistern ◇ Default false. When on, the system prompt gets an overcorrection section headed # Corrections, telling the model to avoid unnecessary self-correction, to correct an earlier statement only when the error would change the user's code, conclusions or decisions, and not to treat a follow-up question as a sign it was wrong. The section is also added when CLAUDE_CODE_LARCH_CISTERN is set, when the Opus 5 prompt bundle applies (unless tengu_fennel_godwit is on), or when the client data key of this name is true.v2.1.257 → v2.1.259 ○gone never read tengu_lean_cast ◇ No description from anybody.never mined ·never mined never read tengu_left_arrow_blocked ◇ Left-arrow no longer backgrounds agents while you have unsent text, so drafts survive.never mined ·never mined never read tengu_legacy_opus_migration ◇ An analytics event, not a feature flag. Claude Code logs it at startup after a first-party user's model in user settings is one of claude-opus-4-20250514, claude-opus-4-1-20250805, claude-opus-4-0 or claude-opus-4-1 and has been rewritten to opus; the field from_model holds the old value.never mined ·never mined never read tengu_lexical_crystal ◇ No description from anybody.v2.1.287 → v2.1.309 ◆in the build AND v2.1.309 tengu_lilac_loom ◇ Default null. A promotional "feature of the week" campaign payload: an object with feature, optional command, startsAt and endsAt (ISO 8601 with a timezone), and optional display fields such as titleLabel, tips, announcementLines and creditless. A payload that fails validation is logged and ignored, and the campaign is active only between startsAt and endsAt.v2.1.242 → v2.1.309 ◆in the build AND v2.1.309 tengu_lodestone_enabled ◇ No description from anybody.never mined ·never mined never read tengu_log_1p_events ◇ No description from anybody.never mined ·never mined never read tengu_log_segment_events ◇ No description from anybody.never mined ·never mined never read tengu_loggia_carousel ◇ Default false. Turns on the model refusal fallback: messages with subtype model_refusal_fallback are rendered, the /config toggle Switch models when a message is flagged (setting switchModelsOnFlag, default on) is shown, and with that setting on and a first-party API a fallback model can be requested from the server. CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK forces it off; the flag's value is also sent to the VS Code extension in its experiment_gates notification.v2.1.152 → v2.1.169 ○gone AND v2.1.309 tengu_loggia_denkbild ◇ Default false. When on and the session has a bridge permission-relay dialog that supports the refusal_fallback_prompt kind, the model-fallback prompt shown after an API refusal is sent through that bridge dialog. When off, or when the bridge dialog lacks that kind, an active REPL bridge session makes the client count as unable to show that prompt.v2.1.199 → v2.1.309 ◆in the build never read tengu_long_context_survey_probability ◇ Default "", which means 0.2. Chance that the long-context feedback survey opens once its token threshold is crossed: the value goes through Number, an empty or non-finite value uses 0.2, and the result is clamped to 0 through 1. The roll is made once per survey component, so a failed roll is not retried.v2.1.196 → v2.1.309 ◆in the build AND v2.1.309 tengu_long_context_survey_question_variant ◇ Default "". Selects the long-context survey question: instruction_following asks How well is Claude following the instructions you gave earlier in this conversation? (optional), and any other value uses the generic survey question.v2.1.191 → v2.1.309 ◆in the build AND v2.1.309 tengu_long_context_survey_threshold ◇ Default "". Token count at which the long-context feedback survey may appear; the value must be all digits and greater than 0, and anything else leaves the threshold unset so the survey never triggers. The count it is compared with is chosen by tengu_long_context_survey_trigger_mode.v2.1.191 → v2.1.309 ◆in the build AND v2.1.309 tengu_long_context_survey_trigger_mode ◇ Default "". Which token count is compared with tengu_long_context_survey_threshold: virtual_context uses a context-token count computed from the current transcript, and any other value (cumulative) uses the session's cumulative token count.v2.1.196 → v2.1.309 ◆in the build AND v2.1.309 tengu_loud_sugary_rock2 ◇ Default false. Applies only when the main model is claude-opus-4-7. When on, the system prompt gets a # Thinking system reminder section, and each regular user prompt after the first assistant reply gets a meta reminder: Respond with just the action or changes and without a thinking block, unless this is a redesign or requires fresh reasoning. The reminder is not added with a custom system prompt or with thinking disabled, and a cached client-data value of "true" or "false" overrides the flag.v2.1.138 or earlier → v2.1.152 ○gone AND v2.1.309 tengu_lowpri_offer_withheld ◇ The lower-priority continuation offer is shorter and shows how much allowance is left.never mined ·never mined never read tengu_luminous_locket ◇ No description from anybody.v2.1.295 → v2.1.309 ◆in the build never read tengu_luminous_whistle ◇ Reasoning text from a non-Claude model can survive a model switch, off unless enabled.v2.1.251 → v2.1.278 ○gone AND v2.1.309 tengu_madrone_spindle_purlin ◇ No description from anybody.v2.1.260 → v2.1.309 ◆in the build never read tengu_malformed_tool_use_clean_retry ◇ Default false. When a response stops with tool_use but no tool call could be parsed, Claude Code retries once; with this flag on, the malformed assistant messages are tombstoned and dropped from the retry and the nudge reads The previous response failed to produce a valid tool call. Please retry the tool call now. With it off, the malformed messages stay in the history and the nudge is Your tool call was malformed and could not be parsed. Please retry.v2.1.174 → v2.1.250 ○gone AND v2.1.309 tengu_managed_config_ready ◇ New startup telemetry tracks managed-config readiness and an optional live identity checkv2.1.283 → v2.1.289 ○gone never read tengu_managed_settings_loaded ◇ Managed settings logging now records which admin config tiers applied and whether env union was opted out.never mined ·never mined never read tengu_managed_settings_os_read ◇ Reading administrator-set settings now records how the OS policy lookup went.never mined ·never mined never read tengu_managed_settings_security_dialog_shown ◇ The managed-settings security consent dialog retries on another surface before giving up.never mined ·never mined never read tengu_managed_settings_validation_errors ◇ Managed-settings startup output now separates non-error notices from actual errorsnever mined ·never mined never read tengu_maple_forge_w8k ◇ No description from anybody.never mined ·never mined never read tengu_maple_tide ◇ Default false. When true, enables the /goal command, which sets a condition for Claude to keep working toward, in interactive sessions and, in its non-interactive variant, in non-interactive or remote sessions; it also restores a goal from the transcript when a session is resumed.v2.1.138 or earlier ○gone AND v2.1.309 tengu_marble_lark ◇ Default false. When auto memory is enabled, turns on personal memory sync: the startup sync also pulls and pushes the user memory scope (still subject to the allow_memory_sync policy), and writes into the auto-memory directory are scanned for secrets with the error Memory is synced to your account. When it reads false, an in-flight user scope push stops with personal sync disabled.v2.1.147 → v2.1.209 ○gone AND v2.1.309 tengu_marketplace_background_install ◇ An analytics event, not a feature flag. Claude Code logs it after the background plugin marketplace install pass when at least one marketplace was installed, updated or failed; the fields are installed_count, updated_count, failed_count and up_to_date_count.never mined ·never mined never read tengu_marketplace_removed ◇ Marketplace names sent in telemetry are now hidden as 'third-party' unless the marketplace is an official onenever mined ·never mined never read tengu_marketplace_updated ◇ Marketplace names sent in telemetry are now hidden as 'third-party' unless the marketplace is an official onenever mined ·never mined never read tengu_mcp_auto_background ◇ Default true. When on, an MCP tool call still running after 120 seconds is moved into a background task so the conversation can continue, and Claude is notified when it finishes; when off, MCP calls are not moved to the background at all. IDE servers (sse-ide, ws-ide) and subagent calls are never moved, the 120-second move needs CLAUDE_AUTO_BACKGROUND_TASKS in non-interactive sessions (which otherwise can only move a call when a new message arrives), and CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS replaces the delay and the flag.v2.1.212 → v2.1.309 ◆in the build never read tengu_mcp_ccr_prefetch ◇ No description from anybody.never mined ·never mined never read tengu_mcp_cgroup ◇ On Linux you can force MCP servers into their own memory cgroup with an environment variable.never mined ·never mined never read tengu_mcp_claudeai_eligibility_gate ◇ Default false. When on, a claude.ai connector (claudeai-proxy MCP server) whose config has eligible: false is not connected and is reported as needs-auth.v2.1.212 → v2.1.309 ◆in the build AND v2.1.309 tengu_mcp_claudeai_proxy_401 ◇ MCP proxy 401 telemetry now sanitizes the error code before logging itnever mined ·never mined never read tengu_mcp_cli_command_executed ◇ No description from anybody.never mined ·never mined never read tengu_mcp_degraded ◇ MCP events are forwarded for logging with identifying fields stripped, and skipped entirely if analytics are off.never mined ·never mined never read tengu_mcp_delete ◇ MCP CLI commands now hash server names in telemetry instead of sending them raw.never mined ·never mined never read tengu_mcp_directory_bff ◇ Default false. Chooses the endpoint for the list of official MCP server URLs: when on, /api/directory/servers (up to 500 per page, remote servers only); when off, the legacy /mcp-registry/v0/servers (version=latest, 100 per page). Either way at most 20 pages are fetched.v2.1.138 or earlier → v2.1.309 ◆in the build never read tengu_mcp_directory_visibility ◇ Default ["commercial", "gsuite", "enterprise", "health"]. Visibility values sent, comma-joined, as the visibility parameter when fetching the official MCP server directory. A non-array or an array with a non-string entry uses the default; empty strings are dropped, and an empty result skips the fetch and leaves the official URL set empty.v2.1.138 or earlier → v2.1.309 ◆in the build never read tengu_mcp_discovery_source ◇ Startup no longer dials cached HTTP or SSE MCP servers; they restore from cache and connect on first use.never mined ·never mined never read tengu_mcp_end_turn_overridden_by_failed_sibling ◇ No description from anybody.never mined ·never mined never read tengu_mcp_first_party_auto_auth ◇ Automatic MCP re-authentication using a stored design token now only applies to /v1/design/ server URLs, not other first-party serversnever mined ·never mined never read tengu_mcp_get ◇ MCP CLI commands now hash server names in telemetry instead of sending them raw.never mined ·never mined never read tengu_mcp_headers ◇ No description from anybody.never mined ·never mined never read tengu_mcp_instructions_pool_change ◇ Changed MCP server instructions would be re-sent instead of suppressed, but that needs the disabled discovery cache.never mined ·never mined never read tengu_mcp_list_changed ◇ Claude can now tell whether an MCP tool list refresh came from a reconnect or a server notification.never mined ·never mined never read tengu_mcp_list_paginated ◇ Unexpected telemetry values are replaced with a placeholder instead of being sent raw.never mined ·never mined never read tengu_mcp_listen_reopen ◇ MCP notification streams now reconnect after a server drop or redeploy instead of silently going dead.never mined ·never mined never read tengu_mcp_listen_server_denylist ◇ Remote config can force legacy MCP handshakes or suppress listening for named servers, matched by hostname.v2.1.259 ○gone never read tengu_mcp_listing_prior_rejected ◇ Two telemetry events were allowlisted, one for rejected MCP tool listings and one comparing served model catalogs.never mined ·never mined never read tengu_mcp_local_oauth_blocked_hosts ◇ Default {"hosts": ["microsoft365.mcp.claude.com", "gmail.mcp.claude.com", "gcal.mcp.claude.com"]}. An http or sse MCP server whose URL host (lowercased, trailing dot removed) is in hosts is classified as Anthropic-hosted for authentication, with a message that it doesn't support local OAuth and should be connected under Settings, Connectors on claude.ai. A missing or non-array hosts, or one with no string entries, uses the default list.v2.1.162 → v2.1.309 ◆in the build AND v2.1.309 tengu_mcp_multidialog_choice ◇ The new-MCP-server dialog now ignores stray Enter or Esc for a moment, and can be suppressed entirely.never mined ·never mined never read tengu_mcp_negotiation_server_denylist ◇ Remote config can force legacy MCP handshakes or suppress listening for named servers, matched by hostname.v2.1.259 ○gone AND v2.1.309 tengu_mcp_oauth_issuer_echo_mismatch ◇ MCP OAuth now checks a server's echoed issuer, but only records mismatches rather than blocking.never mined ·never mined never read tengu_mcp_pending_call ◇ The MCP wait debug line now shows cached servers as their own count.never mined ·never mined never read tengu_mcp_protocol_negotiation_ ◇ MCP servers using a newer protocol era no longer get custom notification handlers, and you see a warning.never mined ·never mined never read tengu_mcp_reconcile ◇ MCP server reconciliation now skips re-dialing servers that a newer server list already replacednever mined ·never mined never read tengu_mcp_retry_failed_remote ◇ Default true. When true, remote MCP servers (http, sse, claudeai-proxy) whose initial connect failed transiently (HTTP 500, 502, 503, 504, network errors such as ECONNREFUSED or ETIMEDOUT, or an sse failure with no error code) are retried. One connect path retries each such server up to 3 times with backoff starting at 1000 ms and capped at 30000 ms; the other retries the still-failed servers after 500, 1500 and 4000 ms.v2.1.138 or earlier ○gone AND v2.1.309 tengu_mcp_retry_on_tools_list_error ◇ Default false. When an MCP server connects but its tools/list call failed, and the server's transport is neither stdio nor sdk, Claude Code marks it pending and reconnects to retry, up to 3 times with exponential backoff starting at 1000 ms and capped at 30000 ms. Off, the server stays connected with the tools-list error.v2.1.152 → v2.1.153 ○gone never read tengu_mcp_sdk_generation ◇ MCP events are forwarded for logging with identifying fields stripped, and skipped entirely if analytics are off.never mined ·never mined never read tengu_mcp_server_connection_failed ◇ MCP debug output now shows which protocol version and path a server connection used.never mined ·never mined never read tengu_mcp_server_connection_succeeded ◇ MCP debug output now shows which protocol version and path a server connection used.never mined ·never mined never read tengu_mcp_skills_funnel ◇ Skills fetched from MCP servers now report extra usage stats back to Anthropic.never mined ·never mined never read tengu_mcp_stateless_skip_init ◇ Default true. When on, a claude.ai connector (claudeai-proxy) with stateless: true connects statelessly: MCP initialize is answered from its cached init response when that parses and carries a supported protocol version (otherwise a real initialize runs), and when tengu_mcp_discover_projection_prior is not off, server/discover can be resolved from a cached projection. It also turns on the anthropic-mcp-client-capabilities and MCP-Protocol-Version request headers even when tengu_mcp_discover_projection_prior is off.v2.1.169 → v2.1.309 ◆in the build AND v2.1.309 tengu_mcp_subagent_prompt ◇ Default false. Chooses the instructions Claude gets when an MCP tool result is too large and is saved to a file. Off gives the generic text (read it in chunks with offset and limit, then state how much was read); on gives format-specific steps (probe JSON with jq, slice long lines by character range, or read a stated number of lines per chunk) and suggests doing the reading in a subagent. MCP_TRUNCATION_PROMPT_OVERRIDE overrides it: legacy forces the generic text, any other non-empty value the new one.v2.1.138 or earlier → v2.1.309 ◆in the build AND v2.1.309 tengu_mcp_tool_call_auth_error ◇ A bug that could crash MCP auth-error logging with a stray exception has been fixednever mined ·never mined never read tengu_mcp_tool_search ◇ No description from anybody.never mined ·never mined never read tengu_mcp_tools_listed ◇ MCP tool-list telemetry now reports total, max, mean, and over-limit counts for tool description lengthsnever mined ·never mined never read tengu_mcp_tripwire ◇ MCP events are forwarded for logging with identifying fields stripped, and skipped entirely if analytics are off.never mined ·never mined never read tengu_media_byte_cap_stripped ◇ An analytics event, not a feature flag. Claude Code logs it when the image and document data in a request exceeds the media byte limit and it strips the oldest blocks: blocks inside tool results become [media removed: request limit], other blocks are dropped (a message left empty gets the placeholder). Fields are totalBytes, byteLimit, removedCount and removedBytes, and the removed totals also include blocks dropped by the media count limit in the same pass.never mined ·never mined never read tengu_mem_push_delete_mode ◇ Default corroborate. How a memory store sync push treats a previously synced file that is now missing locally: immediate deletes it remotely on that push, never never deletes it (and skips the mass-delete hold check), and corroborate deletes it only after it has been missing on at least 2 walks at least 30000 ms apart. Other flag values mean corroborate; a set CLAUDE_CODE_MEMORY_PUSH_DELETE_MODE overrides the flag.v2.1.212 → v2.1.309 ◆in the build AND v2.1.309 tengu_memdir_accessed ◇ Memory file access reporting now notes the file kind and how it reached the model.never mined ·never mined never read tengu_memdir_disabled ◇ Auto-memory loading is now skipped when a permissions setting fences reads outside the working directorynever mined ·never mined never read tengu_memdir_file_edit ◇ Editing files in your memory directory now clears stale caches right away, so changes take effect sooner.never mined ·never mined never read