Feature flags, all 2,081.

tengu_account_plugins_sync_enabled ◇ Plugins you enabled on claude.ai can be pulled to your machine, but only your account server can enable it.never mined ·never mined AND v2.1.308 tengu_alder_compass ◇ Default false. When on, the spinner tip rotation can include New to Claude Code? Run /powerup for a quick interactive tutorial. The tip is offered only to a user with fewer than 10 startups who has not unlocked any powerups.v2.1.139 → v2.1.308 ◆in the build AND v2.1.308 tengu_amber_anchor ◇ Default false. Enables installing the background daemon as a launchctl/systemd service: when off, claude daemon install, start and restart fail as disabled in this version, the claude daemon usage text says the daemon runs on demand, and daemon messages call it background service in place of daemon. When on and the daemon cold-start mode is ask (CLAUDE_CODE_DAEMON_COLD_START, the daemonColdStart setting, or the tengu_quiet_harbor flag), starting a background session with no daemon running can offer claude daemon install instead of starting one on demand, unless that prompt was dismissed or CLAUDE_CONFIG_DIR is set.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_amber_prism ◇ Default false. When on and auto memory is enabled, the tool result Claude receives when the user rejects or cancels a tool call gets a note appended: the user's next message may contain a correction or preference, and Claude should consider saving it to memory for future sessions.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_amber_relay ◇ Default false. When on, session persistence is enabled for Remote Control connections from both the interactive session and the SDK: the local transcript is synced to the remote session and later internal events are written to it as they happen. The interactive session can also enable persistence through other conditions when the flag is off.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_auto_mode_classifier_queue ◇ Default false. When on, auto mode's permission-classifier calls are queued and run one at a time per agent (the main session is one queue); if the permission mode changed while a call waited so the auto decision no longer applies, the decision is dropped: the tool call is denied in dontAsk mode or where permission prompts are unavailable, and otherwise the user is asked. The env var CLAUDE_CODE_AUTO_MODE_CLASSIFIER_QUEUE, when set, takes precedence over the flag.v2.1.218 → v2.1.220 ○gone AND v2.1.308 tengu_auto_notice_once ◇ Default false. In auto mode, when on, the recurring Auto Mode Active reminders are replaced by a single meta message, The user has asked you to work without stopping for clarifying questions. ..., added only when no auto mode reminder is found since the last auto mode exit. When off, the full or sparse auto mode reminder repeats every few turns.v2.1.138 or earlier → v2.1.144 ○gone AND v2.1.308 tengu_birch_kettle ◇ Default false. Enables the bundled claude-code-docs skill, which answers questions about Claude Code's own commands, flags, settings, hooks and other features, can be invoked by the user with a [question] argument, and is allowed Read, Grep, Glob and WebFetch limited to code.claude.com and platform.claude.com. The skill is enabled only when this flag is on and its registration was not marked disabled.v2.1.154 → v2.1.308 ◆in the build AND v2.1.308 tengu_brass_sled ◇ Default false. When on, the PowerShell tool's prompt gets a "Developer tools verified on this machine's PATH" line listing the build and dev tools Claude Code detected, telling Claude that a tool missing from the list is likely not installed and should be checked with Get-Command first; on Windows it adds a note that the MSVC toolchain (cl, nmake, msbuild) is only on PATH inside a Visual Studio developer shell. The line is left out when no tools were detected.v2.1.198 → v2.1.308 ◆in the build AND v2.1.308 tengu_brick_follow ◇ Default false. When on, or when CLAUDE_CODE_VOICE_FORWARD_INTERIMS_TYPED is set, the voice dictation speech-to-text WebSocket is opened with forward_interims=typed added to its query string. The flag's value is also passed to the VS Code extension in its experiment_gates notification.v2.1.141 → v2.1.308 ◆in the build AND v2.1.308 tengu_cedar_sundial ◇ Default false. When on, an edit to a file that changed on disk since it was last read is allowed to proceed if the edit's old_string still matches the current contents exactly once (or matches at all with replace_all), instead of failing with File has been modified since read. A missing or ambiguous match still fails.v2.1.178 → v2.1.207 ○gone AND v2.1.308 tengu_cheerful_horizon ◇ Default false. When on, Claude Code stops requesting whether the claude.ai account has GitHub connected. It uses only a status it already holds, such as the one cached in config as githubWebConnectionStatusCache for the same account and organization, and a live check answers unknown.v2.1.259 → v2.1.308 ◆in the build AND v2.1.308 tengu_cobalt_heron ◇ Default false. Makes the context tip opus-on-pro-near-limit eligible: on a Pro plan, with an Opus model, after more than half of the usage limit is used, the tip suggests /model and says Sonnet 5 uses the weekly limit more slowly than Opus.v2.1.191 → v2.1.211 ○gone AND v2.1.308 tengu_cobalt_wren ◇ No description from anybody.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_copper_gazette ◇ No description from anybody.v2.1.292 → v2.1.308 ◆in the build AND v2.1.308 tengu_coral_fern ◇ Default false. When on, the auto memory prompt gets a ## Searching past context section telling the model to search the memory directory's .md topic files first and session transcript .jsonl logs as a last resort, with an example grep -rn command or Grep tool call for each, and to use narrow search terms.v2.1.138 or earlier → v2.1.146 ○gone AND v2.1.308 tengu_cosmic_crayon ◇ No description from anybody.v2.1.299 ○gone AND v2.1.308 tengu_crimson_vector ◇ Default false. When on, and requests go to Anthropic's first-party API (the firstParty provider with no ANTHROPIC_BASE_URL or an Anthropic one), the default maximum base64 size for an image sent to the model rises from 5242880 bytes (5 MB) to 10485760 bytes (10 MB). A model whose own image limits set a maximum size keeps that value.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_doorbell_agave ◇ No description from anybody.v2.1.138 or earlier → v2.1.294 ○gone AND v2.1.308 tengu_drifting_nova ◇ No description from anybody.v2.1.295 → v2.1.308 ◆in the build AND v2.1.308 tengu_dune_wren ◇ Default false. Turns on the subagent-fanout-nudge tip ("Say \"fan out subagents\" and Claude sends a team"), which has a 3-session cooldown, for users signed in with a claude.ai subscription (OAuth with the user:inference scope); for everyone else the tip is eligible whatever the flag says.v2.1.141 → v2.1.179 ○gone AND v2.1.308 tengu_expressive_church ◇ No description from anybody.v2.1.293 → v2.1.308 ◆in the build AND v2.1.308 tengu_garnet_finch ◇ Default false. Turns on the loop-command-nudge tip advertising /loop, which has a 3-session cooldown, for users signed in with a claude.ai subscription (OAuth with the user:inference scope); for everyone else the tip is eligible whatever the flag says. Either way the tip is hidden in a remote session (CLAUDE_CODE_REMOTE or a remote workspace) and needs cron enabled (CLAUDE_CODE_DISABLE_CRON unset and the tengu_kairos_cron flag on).v2.1.141 → v2.1.179 ○gone AND v2.1.308 tengu_gorse_fathom ◇ Default false. When on, the auto-memory instructions gain a ## Project skill upkeep section: after saving a feedback memory that corrects a repeatable step, Claude is told to fold the correction into the matching .claude/skills/<name>/SKILL.md. It may only edit existing skills, except verify, which it may create.v2.1.198 → v2.1.308 ◆in the build AND v2.1.308 tengu_hawthorn_steeple ◇ Default false. When on, a session keeps a record of tool-result replacements: when the tool results in one message together exceed a per-message size budget, the largest new ones are saved to files and replaced in the conversation, and the same replacements are reapplied on later turns and on resume. Tools without a finite maxResultSizeChars, or that opt out of the aggregate budget, are exempt.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_idle_amber_finch ◇ Default false. When on, Claude Code stops offering an upgrade to Max: the usage-limit menu drops its Upgrade to Max option, the shared upgrade-availability check returns false (so, for example, the five-hour limit message no longer suggests /upgrade), and the startup subscription check no longer reports an existing Claude Max subscription.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 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 v2.1.308 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 v2.1.308 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 v2.1.308 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 v2.1.308 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 v2.1.308 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 v2.1.308 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 v2.1.308 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 → v2.1.308 ◆in the build AND v2.1.308 tengu_post_compact_survey ◇ Default false. When true, Claude Code may show a feedback survey (survey_type: "post_compact") after a compaction: once a user or assistant message follows the newest compaction marker, it opens with probability 0.2. It stays hidden when another survey is open, product feedback is disallowed by policy, or CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY is set.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_quiet_dolphin ◇ The /diff panel is now prototyped internally as a plugin pane rather than built-in UInever mined ·never mined AND v2.1.308 tengu_quiet_slate_wren ◇ Default false. When on, Claude Code no longer shows the Use your existing Claude Pro plan with Claude Code · /login to activate notice to someone who is not signed in with a subscription but whose claude.ai account has Claude Pro. The equivalent notice for Max accounts is not affected.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_reactive_compact_remote ◇ Default false. Read only when CLAUDE_CODE_REMOTE is set, and latched for the rest of the process on first read. In remote sessions with the flag off, reactive compaction after a context-limit error and precomputed compaction are disabled, and proactive auto-compaction is no longer held back for reactive handling or routed through it. Outside remote sessions the flag is not consulted and these paths behave as if it were on.v2.1.172 → v2.1.308 ◆in the build AND v2.1.308 tengu_sequential_summit ◇ No description from anybody.v2.1.301 → v2.1.303 ○gone AND v2.1.308 tengu_shale_finch ◇ Default false. When on, subagents that are not teammates and are not spawned with an exact tool list lose the TodoWrite, TaskCreate, TaskGet, TaskUpdate and TaskList tools.v2.1.138 or earlier → v2.1.285 ○gone AND v2.1.308 tengu_slate_harbor_experiment ◇ Default false. When on, or when CLAUDE_CODE_NEW_INIT is set, /init sends the alternate multi-phase prompt instead of the original one that begins Please analyze this codebase and create a CLAUDE.md file. The /init command description follows CLAUDE_CODE_NEW_INIT alone.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_slate_moth ◇ Default false. Makes the context tip large-context-stale-files eligible: when the context is over 300,000 tokens and more than 100,000 of them are stale file tokens, the tip suggests /compact.v2.1.191 → v2.1.211 ○gone AND v2.1.308 tengu_sparrow_ledger ◇ Default false. When on, adds a Work step by step: section to the system prompt telling the model to reproduce the symptom before editing, edit the source, then re-observe the symptom until it is gone. The CLAUDE_CODE_VERIFY_PROMPT environment variable turns on the same section.v2.1.138 or earlier → v2.1.205 ○gone AND v2.1.308 tengu_spicy_pearl ◇ No description from anybody.v2.1.299 → v2.1.308 ◆in the build AND v2.1.308 tengu_sprightly_lagoon ◇ Default false. When on, the checks for per-turn effort and per-turn timing both report unsupported, so requests carry no per-turn effort or now values and those checks add no per-turn-control-2026-07-01 beta.v2.1.260 → v2.1.301 ○gone AND v2.1.308 tengu_subagent_cache_evict ◇ Default false. When on, model requests marked evictCacheOnComplete that carry sticky betas add the prompt-caching-evict-2026-05-12 beta. Provider checks come first: the request must go to the first-party API at a first-party base URL (or Claude Platform on AWS with no ANTHROPIC_AWS_BASE_URL), with neither CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS set nor HIPAA mode active, and under those conditions a truthy CLAUDE_CODE_SUBAGENT_CACHE_EVICT turns it on without consulting the flag.v2.1.198 → v2.1.308 ◆in the build AND v2.1.308 tengu_tallow_quoit_gate ◇ No description from anybody.v2.1.307 → v2.1.308 ◆in the build AND v2.1.308 tengu_twinkling_globe ◇ New remote gate tengu_twinkling_globe controls whether Claude Code treats account policy limits as absent instead of fetching themnever mined ·never mined AND v2.1.308 tengu_umber_petrel ◇ Default false. When on, auto memory is turned off for sessions whose model matches an entry in the tengu_sepia_cormorant model list, unless auto memory has been explicitly forced on or off.v2.1.138 or earlier → v2.1.308 ◆in the build AND v2.1.308 tengu_validated_clover ◇ A new internal check called tengu_validated_clover appeared in Claude Code's code, but what it does isn't known yetnever mined ·never mined AND v2.1.308 tengu_velvet_hammer ◇ Default false. When true, or when the per-model variant tengu_velvet_hammer_<bucket> is true for the current model (the bucket is a model family such as opus, sonnet or haiku; the base bucket uses the plain name), the Edit tool skips its "File has not been read yet. Read it first before writing to it." rejection, so a file that was never read, or was only partially read, can be edited.v2.1.166 → v2.1.195 ○gone AND v2.1.308 tengu_vscode_feedback_survey ◇ Default false. When on, the session's SDK initialize response carries a feedback_survey_config (the tengu_feedback_survey_config value, with its probability overridden by the feedbackSurveyRate setting when set, plus the last time a survey was shown) for VS Code extension sessions and certain other entrypoints. No config is sent when CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY is set or the allow_product_feedback policy is off.v2.1.162 → v2.1.308 ◆in the build AND v2.1.308 tengu_walnut_prism ◇ Default false. When on (or when CLAUDE_CODE_OWNERSHIP_FRAME is true), system prompt wording shifts to ownership framing in the prompt variants that carry these sections: the opening line becomes You work alongside the user on software engineering tasks and own the outcome of what you take on. (or an output-style variant with the same framing), the confirm-before-risky-actions sentence drops approval in one context doesn't extend to the next, and the When a task has been agreed, the approval covers it end to end paragraph is left out.v2.1.161 → v2.1.215 ○gone AND v2.1.308